Algorithms doi: 10.3390/a19110970

Authors: Laura Antonelli Lucia Maddalena

Imaging problems, including image reconstruction, restoration, enhancement, segmentation,classification, object detection, and secure image transmission, have traditionallybeen addressed using analytical methods derived from mathematical models and domainspecificknowledge. Regularization, filtering, optimization, geometric modeling, andstatistical estimation remain important because their assumptions and behavior can oftenbe understood and controlled explicitly. More recently, Machine Learning (ML) and DeepLearning (DL) methods have achieved impressive results by learning complex image representationsdirectly from large datasets. Current research increasingly seeks to combinethese two perspectives. Instead of treating analytical and learning-based methods as competingalternatives, hybrid approaches incorporate physical constraints, signal-processingprinciples, or application-specific knowledge into trainable systems. This can improve dataefficiency, interpretability, robustness, and computational performance while retaining therepresentational power of modern neural networks. [...]

Algorithms doi: 10.3390/a19110969

Authors: Jonathan Josué Proaño Morales Karime Montes Escobar Purificación Vicente-Galindo

Keeping a quantitative review current in a fast-growing applied field is first a screening problem. We present two algorithms for modelling biofloc aquaculture. The first, Floccus, is an openly archived, PRISMA 2020-compliant pipeline that removes duplicates in three layers (deterministic, embedding-based and large language model (LLM)-judged) and screens each record against eleven criteria under an asymmetric decision rule with an auditable trail. Applied to 22,056 records from four databases, it yielded 16,965 unique and 4883 eligible studies in 25 thematic clusters. Agreement with blind human labels was low (κ=0.107), of the same order as between two human reviewers applying very different inclusion thresholds (κ=0.105; 65% versus 18% of records judged relevant); it saved more screening work than keyword and single-prompt baselines, and an aquatic-science database added under 1% of eligible studies. The second algorithm assembles the corpus’s six model families into one multi-scale hierarchy coupled through the aeration-driven dissipation field, with a proposed operator converting floc size into active biomass through the Thiele effectiveness factor. Manual screening confirmed formal floc modelling is rare in applied biofloc work (22 of 141 records). Corpus-level recall remains moderate (0.45), and the coupled model awaits validation against tank data.

Algorithms doi: 10.3390/a19110968

Authors: Li Liu Kaifeng Geng Shuangjun Liang

In distributed flexible job shops, preventive maintenance is required to manage machine degradation, but competition for limited maintenance resources can delay production. This study develops a bi-objective model to minimize makespan and total energy consumption by integrating factory assignment, operation sequencing, machine and processing speed selection, imperfect preventive maintenance, and maintenance resource allocation. To solve this problem, a multi-objective memetic algorithm with constrained maintenance resources (MRC-MOMA) is proposed. The algorithm combines four-layer encoding with a decoder that generates maintenance requests according to machine degradation and schedules them under maintenance resource constraints. Computational experiments show that MRC-MOMA achieves better overall performance than the comparison algorithms in terms of hypervolume (HV) and inverted generational distance (IGD), with sensitivity analysis supporting this advantage within the tested parameter ranges. For the representative instances, increasing maintenance resource capacity substantially reduces maintenance waiting time and shortens makespan, but does not necessarily reduce total energy consumption.

Algorithms doi: 10.3390/a19110967

Authors: Chi-Yi Hsieh Caroline Hsiu-Fang Hsu Chi-Wen Hsieh Yi-Chi Huang I-Yu Chen Yu-Han Wu

Large language models (LLMs) generate fluent Chinese, but controlling lexical difficulty remains challenging. This study examined a Python-orchestrated external-control workflow using one GPT model, two sequential prompting conditions (no-word-list and lexicon-aware), three post-generation revision strategies (word-level substitution, sentence-level rewriting, and downstream regeneration), and four Test of Chinese as a Foreign Language (TOCFL) levels (A1–B2). Automatic indicators covered 96 initial drafts and 288 revision outputs. For 24 selected tasks, two experts directly rated 180 text instances; 12 unchanged outputs inherited initial-draft ratings, yielding 192 condition-specific expert-evaluation records. Task-matched comparisons used 20,000 TOCFL-stratified task-cluster bootstrap resamples and Holm-adjusted tests; automatic–expert associations were analyzed using Spearman correlations with task-cluster confidence intervals. Lexicon-aware prompting was associated with fewer initial out-of-level words and less revision, although the prompting batches were sequential. Word-level substitution and sentence-level rewriting generally showed greater source-relative embedding resemblance, whereas downstream regeneration received higher expert ratings for several language-quality dimensions but not consistently for level appropriateness. Residual out-of-level word density showed a weak, inconclusive pooled association with level appropriateness. Whole-text cosine similarity was negatively associated with semantic plausibility overall but positively associated within word-level substitution. Automatic indicators and source-blind expert judgments thus provide complementary, non-interchangeable evidence.

Algorithms doi: 10.3390/a19110966

Authors: Dimitrios Galiatsatos Panagiota Galiatsatou

Unsupervised feature selection is difficult when relevance must be inferred without labels, and redundancy, outliers, and sample-localized geometry can distort marginal criteria. This study presents Geometry-Supported Residual Feature Selection (GSR-FS), a fully label-blind algorithm and open-source Python library that represents each feature by its contribution to pairwise sample geometry. Structural relevance is quantified by Bhattacharyya affinity to the aggregate geometry of the remaining features, calibrated against within-feature permutations, and downweighted by an entropy-based Sample-Geometric Coverage term. A non-negative residual reconstruction criterion then penalizes candidates whose geometry is already represented by previously selected features. Automatic selection uses a feature-wise max-null permutation distribution, whereas fixed-cardinality selection follows the same greedy score without using target information. The released library additionally provides balanced pair subsampling, numerically equivalent residual solvers, source-integrity audits, publication guardrails, and diagnostic structural-group outputs. In a source-audited 30-dataset evidence panel, 29 datasets supported complete five-method comparisons. Mean balanced accuracy was 0.6031 for GSR-FS, compared with 0.6019 for SPEC, 0.5970 for MCFS, 0.5892 for MaxVariance, and 0.5810 for Laplacian Score. The Friedman test was not significant (p = 0.488), and no Holm-corrected pairwise comparison was significant. Additional robustness analyses showed that automatic GSR-FS achieved mean balanced accuracy of 0.6216 compared with 0.6069 for fixed-cardinality GSR-FS on the 30-dataset panel, without a significant paired difference (p = 0.1215). Further validation showed high-precision independent-null calibration at α = 0.05, with 25 of 500 datasets producing a non-empty subset (empirical rate of 0.050; exact 95% CI of 0.0326–0.0729). Controlled dependent-feature experiments showed that GSR-FS combined higher latent-group coverage with lower redundancy than Laplacian Score and MCFS in the examined structural-recovery scenarios, while a recent FGMRW-UFS comparator showed limited recovery under this specific criterion. An engineering stress test was completed successfully through 20,000 candidate features under explicitly capped pair and permutation budgets. Controlled nuisance simulations further confirmed that structurally coherent but target-independent variation can be selected, emphasizing that GSR-FS estimates internal structural relevance rather than guaranteed target-specific relevance. These results support GSR-FS as a reproducible, structurally motivated alternative rather than a statistically superior replacement.

Algorithms doi: 10.3390/a19110965

Authors: Alaa A. K. Ismaeel Mahmoud Abuelneel Ali M. El-Rifaie Tamer Mohamed Barakat Mokhtar Said Mohamed Barakat

Load frequency control (LFC) is essential for maintaining frequency stability and regulating tie-line power in interconnected power systems. Conventional fixed-order PI and PID controllers offer limited tuning flexibility for complex multi-area systems, whereas fractional-order controllers provide additional tuning parameters at the cost of increased optimization complexity. Furthermore, the performance of optimization-based LFC schemes depends strongly on the selected objective function, as conventional integral criteria, such as the integral of time-weighted absolute error (ITAE), may not adequately distinguish between candidate responses with different error magnitudes and persistence. To address this research gap, this study investigates an exponential time-weighted cost function, termed the integral exponential time-weighted error saturation (IEMTES) criterion. Its principal contribution is the exponential transformation of the time-weighted error before integration. This preserves an ITAE-related first-order contribution while introducing higher-order nonlinear error terms; a saturation mechanism limits the exponential argument. The objective is to evaluate the effectiveness of IEMTES, in combination with the Glider Snake Optimizer (GSO), for tuning four controller structures: PI, PID, FOPID, and FOIID. Two benchmark systems are considered: a two-area non-reheat thermal system and a three-area thermal–thermal–hydro system incorporating generation rate constraints (GRCs). For the investigated GSO–FOIID configuration in the three-area system, the ITAE decreases from 76.20 with ITAE-based optimization to 69.69 with IEMTES-based optimization, corresponding to an improvement of approximately 8.5%. Additional evaluations consider random step-load patterns, larger load perturbations, and parameter variations. The findings are limited to the investigated systems, controller structures, disturbances, parameter ranges, and optimization settings.

Algorithms doi: 10.3390/a19110964

Authors: Guanjian Zhu Shuozheng Si Yang Guo Ke Wang Xiangyu Liu Jianyu Chen Wei Li Zujun Ding Baolian Liu Jie Ji

This study examines model based interpretation of HVDC grounding responses in horizontally layered soil. Grounding impedance denotes the real voltage to current ratio in the steady state DC conduction model. A finite element forward model is combined with a surrogate assisted constrained robust optimization framework, denoted SACRO, to fit equivalent soil parameter combinations subject to physical bounds and prior information. The observation model distinguishes response fitting from the separate identification of layer resistivities and thicknesses. For a fixed geometry, varying current amplitude in the linear model does not create independent normalized observations; consequently, a small impedance residual alone does not establish unique soil parameter recovery. The ten reported response pairs have relative errors of 1.56% to 1.62% and an arithmetic mean of approximately 1.59%, calculated directly from the displayed values. For SACRO, the results showed 415 FEM evaluations, a runtime of 98.2 min and a post-perturbation mean response error of 2.55%. The algorithm and perturbation summaries are interpreted as case specific descriptive results. The framework supports layered soil response analysis and the interpretation of grounding measurements within the stated model assumptions.

Algorithms doi: 10.3390/a19110963

Authors: Jorge Fernández-Rodríguez Lidia M. Belmonte Rafael Morales Javier de las Morenas

In motor current signature analysis, bearing-related components are weak and their spectral location varies with operating conditions and fault morphology. This paper proposes KESK, the kurtosis of the Hilbert-envelope spectrum of the most impulsive wavelet-packet subband in each current window, and evaluates it as one additional feature to a 38-feature baseline using one phase current from 20 Paderborn bearings (11,173 windows). With supervised energy selection inside training folds, new recordings, unseen operating conditions and unseen bearings were evaluated separately. On the primary recording-level partition, XGBoost macro-F1 increased by 0.002 with KESK, 0.003 with a training-selected fixed node and 0.003 with a DC-free KESK variant; over ten additional partitions, the mean increments were 0.002, 0.003 and 0.007, respectively. With an operating condition held out, the fixed node reduced macro-F1 (−0.014; 95% CI −0.024 to −0.005) and the DC-free variant increased it (+0.016; 0.008 to 0.023). With bearings held out, macro-F1 was 0.48–0.49 for every representation. On nine bearings excluded from all design decisions, a pre-specified test did not confirm the DC-free advantage (−0.003; −0.012 to 0.005). Envelope-spectrum kurtosis therefore provides a small but reproducible amount of complementary information at negligible additional computational cost, although baseline composition and validation level have a substantially greater influence on estimated performance.

Algorithms doi: 10.3390/a19110962

Authors: Kongkidakhon Worasan Karn Moonsri

A buffer requirement inferred from one optimized schedule can depend on the solver’s choice among equally good schedules. This study develops a mixed-integer model of batching and picking-packing scheduling with effective order deadlines and instantaneous whole-batch admission and a target-constrained model that minimizes waiting space over all schedules attaining the certified relaxed lexicographic optimum; equivalent event states reduce repeated matrix coefficients. Generated experiments used eight exact demand families, 24 search families, and a scalability extension with 8 to 192 orders. Minimum-buffer boundaries were verified for 30 of 32 exact configurations, requiring 0–3 positions, and the selected relaxed schedule exceeded this minimum in 16 cases. One position below the minimum, the component lost was the late-order count in 3, tardiness in 7, and picking effort in 10 of 20 configurations; raising a positive buffer from one to two positions never reduced late orders. Certification extended to 10 orders within 180 s per call. The sparse formulation reduced nonzeros by 70.19–82.31%. Guided search won 56 and lost 148 comparisons against ILS/VND, the compact-repair package won 66 and lost 132 against guided search, and at 96 and 192 orders with 30 s budgets, guided search won 17 and lost 16. Under one-tick duration deviations, the certified minimum was exceeded in 38.8% of re-timed samples, and one additional position reduced this to 4.9%. The conclusions are limited to the tested generated configurations.

Algorithms doi: 10.3390/a19110961

Authors: Muhammad Asim Xiaoheng Deng Muhammad Ejaz

Edge computing (EC) has emerged as a critical paradigm for enabling low-latency, real-time applications, yet the distributed and resource-constrained nature of edge infrastructure poses significant challenges for energy-efficient task allocation. Traditional heuristic and reinforcement learning (RL) approaches may not explicitly represent the changing relationships among edge servers, potentially limiting their effectiveness in energy-aware task allocation and load balancing. This paper proposes Graph-Enhanced Adaptive Reinforcement Learning (GEA-RL), a framework that integrates Graph Neural Networks (GNNs) with Deep Q-Networks (DQN) for task allocation and migration in EC systems. Unlike conventional approaches that treat servers as independent entities, GEA-RL explicitly models server relationships and resource dependencies through dynamic graph structures, enabling more intelligent allocation decisions. Our method formulates the optimization problem as a Markov Decision Process with realistic power consumption models incorporating both static and dynamic components while considering resource constraints and migration costs. We evaluate GEA-RL through simulations under multiple workload scenarios and compare its performance with heuristic and RL baselines in terms of power consumption, load balancing, and response time. Simulation results across multiple workload scenarios reveal that GEA-RL achieves an average response time of 99 ms, compared with 121 ms for the Worst-Fit baseline, representing an 18.2% reduction. The framework also achieves a service availability of 99.82%, compared with 99.41% for Worst-Fit, and reduces power consumption by 3.0% under the evaluated settings. These results indicate that incorporating graph-structured information into RL-based resource management can improve response time, service availability, and energy efficiency in the simulated EC environment.

Algorithms doi: 10.3390/a19110960

Authors: Antonios Parasyris Vassiliki Metheniti Georgios V. Kozyrakis Milen Todorov Nikolaos A. Kampanis George Alexandrakis

The water level at ports governs under-keel clearance, pilotage windows, cargo operations and local flood risk, yet regional ocean forecasting systems resolve it at kilometre scales that may lie outside the harbour basin. This study evaluates a station-only machine learning surrogate that forecasts a single low-cost hydromast water-level sensor at the port of Varna (Bulgarian Black Sea) 1 to 24 h ahead, using only the sensor’s own recent history and deterministic harmonic (tidal and diurnal) cyclic features. The target is framed as a residual relative to the current observation, which keeps it bounded regardless of season or tidal phase. Persistence, ridge regression, random forest, XGBoost, a multilayer perceptron, a long short-term memory network and PySR symbolic regression are compared at horizons of 1, 3, 6, 12 and 24 h on a chronological hold-out period. Ridge regression is the most accurate model at 1–6 h, random forest is marginally best at 12 h, and PySR is best at 24 h, where all methods lie close to persistence. PySR reduces the residual problem to short closed-form expressions dominated by harmonic phase terms, at four of five horizons without any lagged water-level input. Grouping the methods into four interpretability tiers shows that, at this site, the intrinsically interpretable models are competitive with or better than the black-box alternatives at four of the five horizons.

Algorithms doi: 10.3390/a19110959

Authors: Narantuya Vandantseren Erdenebayar Lamjav Dae-ki Hong Otgonbayar Bataa Chuluunbandi Naimannaran

Resource allocation that is both energy-efficient and fair remains a persistent challenge in 5G, beyond-5G, and emerging 6G networks, where users with heterogeneous service requirements compete for limited radio resources. Although rate-splitting multiple access (RSMA) is commonly investigated through physical-layer models, service-level datasets often contain demand indicators without channel matrices, beamforming variables, or stream-level power parameters. We develop and evaluate a QoS-driven allocation framework based on three allocation-inspired models: an RSMA-inspired balanced model, a NOMA-inspired signal-strength-aware model, and an OFDMA proportional-fair reference model. Four metaheuristic optimizers—Genetic algorithm(GA), simulated annealing (SA), particle swarm optimization (PSO), and ant colony optimization (ACO)—are applied to the same QoS demand profile, resource budget, constraints, feasibility-repair procedure, and performance measures, producing 12 model–optimizer combinations. The input comprises 400 user records, including 393 records with positive bandwidth requirements and seven records with zero bandwidth. Each method was evaluated over 30 independent runs. Under this QoS-level abstraction, RSMA-PSO produced the highest mean fitness and energy-efficiency values, whereas RSMA-GA achieved the strongest composite balance across fitness, energy efficiency, fairness, and satisfaction. The mean fitness of RSMA-GA was 9.32% higher than that of NOMA-GA and 18.00% higher than that of OFDMA-PF-GA. These results characterize allocation behavior under a shared QoS profile and should not be interpreted as full physical-layer validation of RSMA, NOMA, or OFDMA transmission.

Algorithms doi: 10.3390/a19110957

Authors: Guanqiu Qi

It is with pleasure that we announce the launch of a new section of Algorithms, named “Signal Processing, Image Processing and Computer Vision Algorithms” [...]

Algorithms doi: 10.3390/a19110958

Authors: William J. Marin-Rodriguez Víctor H. Guadalupe-Mori Rubén A. Rayme-Serrano Jenny M. Zavaleta-Oliver Sandy S. Altamirano-Cáceres Alexia F. Geldres-Chávez Marcelo R. Bedoya-Ostolaza

Background: Machine-learning models for classifying or predicting overweight and obesity in children vary widely in design, validation, and reporting. Methods: We registered a protocol (OSF, doi:10.17605/OSF.IO/EP47Z) and searched PubMed, Scopus, and Web of Science on 15 October 2026 for original studies of machine-learning models for overweight or obesity in people under 19 years. One reviewer screened all records, and a second reviewer screened a 15% sample; one reviewer assessed risk of bias with PROBAST+AI. Results: Of 105 full texts sought, 36 (34%) could not be retrieved; 37 studies were included. None was judged at low risk of bias. External, temporal, or geographic validation was reported in 6 studies (16%) and calibration in 6 (16%); up to 4 (11%) used potentially outcome-defining inputs, and preprocessing preceded data partitioning in at least 11 (30%). In the 2021 Peruvian VIANEV survey, only 185 children, all in Lima Metropolitana, had measured weight and height, too few for multivariable model development; national implementation requires data with comprehensive anthropometric coverage. Conclusions: Reported accuracy and AUC values should be interpreted cautiously; studies should define the outcome with a named reference, avoid outcome-defining inputs, apply preprocessing inside the training data, and report calibration and external validation.

Algorithms doi: 10.3390/a19110956

Authors: Aymen Abu-Errub Ashraf Odeh Tariq Bishtawi Salam Fraihat

The increasing interconnection of industrial devices, programmable logic controllers, edge platforms, and networked control systems has expanded the attack surface of Industrial Internet of Things (IIoT) environments. Machine-learning-based intrusion detection can improve the identification of malicious activities, but centralized training requires the collection of distributed industrial data and may provide an unrealistic estimate of detector performance when evaluation relies on randomly partitioned traffic. Federated learning offers a data-local alternative; however, its effectiveness can be affected by non-independent and identically distributed (non-IID) client data and temporal changes in attack and network behavior. This study presents FedIDS-IIoT, a federated intrusion detection framework evaluated using the CICAPT-IIoT dataset, with particular emphasis on non-IID attack distributions and temporal distribution shift. Sample-size-weighted federated averaging (FedAvg) is evaluated across controlled client configurations in which attack observations are progressively concentrated among five simulated clients. A training-only feature-selection procedure is used to derive a compact 20-feature representation, while model selection and operating-threshold determination are completed before evaluation on a chronologically later and previously untouched traffic block. The federated detector is evaluated against a centralized XGBoost reference and two matched federated optimizers, FedProx and FedAdam, under the same development protocol and feature representation. In the matched federated comparison, FedAvg achieves the highest F1-score in the mild and moderate development partitions, whereas FedProx is marginally higher in the severe partition (0.9643 vs. 0.9633). This result is interpreted as an optimizer-dependent operating trade-off rather than evidence of universal FedAvg superiority. Under the frozen operating threshold, FedAvg also produces substantially fewer false positives on the temporal test set. The non-IID experiments indicate that moderate differences in attack distribution have limited impact, whereas severe concentration of attack evidence can reduce detection performance. A paired McNemar test confirms that the difference in temporal error behavior between the two detectors is statistically significant, and a cost analysis shows that the federated detector needs only a few thousand parameters and a few megabytes of communication over the entire training run. The MADQN component is treated as a secondary exploratory simulation and is not presented as real-life network-response validation. These findings demonstrate the importance of evaluating federated IIoT intrusion detection under both client-level distribution heterogeneity and temporal distribution shift rather than relying solely on conventional random train–test evaluation.

Algorithms doi: 10.3390/a19110955

Authors: Fu-Shiung Hsieh

Although the Cyber–Physical Production Systems (CPPSs) paradigm offers significant potential for enhancing the economic and environmental performance of production systems and contributing to the achievement of the Sustainable Development Goals (SDGs) in the manufacturing sector, it also presents challenges for developing sustainable CPPSs. Sustainable CPPSs consist of collaborative networked entities in cyber space and physical space through ICT. Optimizing sustainable CPPSs is a challenging task due to the discrete and nonconvex solution space involved, making the development of effective problem solvers essential for improving their performance. Metaheuristic algorithms provide an approach to addressing the challenges of optimizing sustainable Cyber–Physical Production Systems (CPPSs). Cooperative Coevolution (CC) provides a divide-and-conquer approach to improving the performance. However, studies on the applying CC to sustainable CPPSs planning problems remain limited. To bridge this gap, we combine CC with a single Differential Evolution (DE) strategy to develop several solution algorithms for planning sustainable CPPSs. The results show that the proposed CC-based DE algorithms outperform most single strategy metaheuristic algorithms and are comparable to the self-adaptive algorithms with two DE strategies in terms of performance and robustness. The results indicate that the proposed CC-based algorithms provide an effective alternative approach to the optimization of sustainable CPPSs.

Algorithms doi: 10.3390/a19110954

Authors: Chenhao Ying Lexing Xiong

Financial-news studies commonly aggregate articles by day or use broad event windows, obscuring the elapsed time between publication and a tradable market response. We ask whether timestamped news and minute-level prices can identify how information persistence varies by release session and volatility. We propose Differentiable Session-Aware Rapid Attention Decay (DiffSARAD), which jointly learns a continuous global half-life, session- and volatility-conditioned adjustments to it, and a three-class post-publication abnormal-return classifier. The representation combines conditional exponential weights with both weighted FinBERT sentiment composition and decayed news intensity, preserving half-life identifiability for single-news histories. Using 23,294 analyst-rating items for 23 equities from 2 January 2018 to 30 June 2020, executable 30-, 60-, and 120-min labels, and 18 expanding purged folds with strictly separated validation, the estimated base half-life is 17.4–23.0 h. Intraday coefficients are positive and post-market coefficients negative in all 18 folds at every horizon, indicating slower intraday and faster post-market decay, and the pattern recurs under a second processing version. Synthetic recovery experiments on the archived event-history design confirm that these parameters are recoverable and that near-zero coefficients are returned under a zero-effect generating process. News and decay features lower daily log loss versus price-and-market features under serial-dependence-robust inference, and verified nested ablations attribute small but significant loss reductions to each conditioning component. The news-count baseline remains the strongest classifier, and a two-basis-point cost stress turns net returns negative.

Algorithms doi: 10.3390/a19110953

Authors: Yang Ou Wenjun Duan Qun Zhang Xinyi Zhou Liulin Ou Tian Lan Yiming Zhou

Riverine navigation-oriented linear cultural and tourism corridors often face challenges such as the unidirectional diffusion of upstream–downstream risks, inconsistent spatiotemporal standards for monitoring data across multiple departments, and difficulties in tracing the logic underlying the assessment of composite disasters. Existing long-series forecasting networks and basic retrieval-enhanced generation frameworks are often ill-suited to the unique chain-like risk evolution patterns inherent in linear waterways. Taking the Pinglu Canal Cultural and Tourism Integration Belt as a case study, this paper proposes a knowledge-driven, multi-source data fusion intelligent early-warning system named KE-MFEWF. The paper introduces an MSTF-Net spatiotemporal feature extraction network incorporating asymmetric corridor attention mechanisms, which automatically adjusts feature weights based on the differing patterns of risk propagation along the upstream and downstream sections of the waterway; it also develops a large-scale model simulation module leveraging graph-based and multi-vector database collaborative retrieval techniques, and establishes a dynamic early-warning determination criterion by integrating two types of correction coefficients—those accounting for seasonal variations and holiday passenger flow. To compensate for the lack of long-term observational samples for the canal under construction, this study employs a combined dataset obtained from real-world measurements, numerical simulations, and cross-basin data migration; the model’s generalization performance was tested on the navigable tourism section of the Xijiang River. When compared with mainstream benchmark models—including Informer, TimesNet, and single-channel RAG—the proposed method achieves an improvement of 6.8% in the AUC metric for comprehensive composite risk prediction; it can issue risk warnings up to 45 min in advance, while reducing the rates of false alarms and missed alerts by 7.2% and 6.5%, respectively; after undergoing INT lightweight compression, the model’s storage footprint is reduced by 65%, and the inference time for a single batch of time-series data on edge devices is less than 0.3 s, making it suitable for real-time on-site deployment. The comprehensive algorithmic framework proposed in this paper can serve as a reusable technical solution for the digital security management and control of similar inland waterway linear cultural tourism corridors in China.

Algorithms doi: 10.3390/a19110952

Authors: Qinlan Wu Jiaoli Xu Xiaochao Xiong

To address multi-UAV cooperative path planning in complex three-dimensional dynamic environments, this paper proposes an Improved Spider Wasp Optimization (ISWO) algorithm to overcome the original SWO’s shortcomings: poor population diversity, low convergence accuracy, premature convergence, and weak dynamic obstacle avoidance. Three key optimizations are introduced: a Logistic–Tent hybrid chaotic map initializes the population to expand the search range and improve diversity; modified chasing and escaping update strategies enhance adaptive search and global exploration; an optimal neighborhood perturbation mechanism balances exploration and exploitation, improving local optima escape and convergence stability. A multi-constraint 3D path planning cost model is constructed, considering path length, smoothness, altitude, energy consumption, and threat avoidance to meet practical UAV flight requirements. Experiments on static, dynamic threat, and multi-UAV cooperative scenarios (5 and 7 UAVs) compare ISWO with the original SWO and mainstream algorithms, with ablation experiments verifying each improvement. Results show that ISWO achieves better accuracy and stability in static scenarios, reduces path cost and collision rates in dynamic environments, and generates shorter, smoother, safer trajectories, demonstrating its excellent robustness and adaptive planning capability in complex 3D environments.

Algorithms doi: 10.3390/a19110951

Authors: Wei Li Shuangyu Shi Siying Xie Xiaoyan Dong Yang Zhao

Objects in UAV imagery are typically small and densely distributed, while complex backgrounds, occlusion, and progressive downsampling further complicate detection. To improve the balance between detection accuracy and computational cost, this paper proposes an improved model based on RT-DETR-R18. First, partial convolution (PConv) is introduced into the backbone to reduce computational overhead. Second, efficient multi-scale attention (EMA) is embedded within the backbone residual blocks to enhance feature interaction. Finally, Wise-IoUv3 (WIoUv3) with its non-monotonic focusing mechanism is used to optimize bounding box regression during training. Experiments on VisDrone2019 show that the improved model increases mAP50 from 47.2% to 49.6%, mAP50:95 from 28.1% to 30.2%, and small-object average precision (AP_small) from 17.8% to 19.6%, while reducing parameters from 20.2 M to 18.5 M and computation from 60.1 to 49.2 GFLOPs; measured inference speed improves from 67 to 72.3 FPS. Ablation results show that the effect of each module depends on the combination configuration, and the full combination achieves the highest accuracy in this set of experiments. In addition, when trained and evaluated independently on DOTA v1.0, the improved model also yields higher overall detection accuracy. These results indicate that the proposed method improves overall and small-object detection performance while reducing model size and computation, and maintains real-time inference capability.

Algorithms doi: 10.3390/a19110950

Authors: Lishuang Shen

Background: Rare disease molecular diagnosis remains bottlenecked by the need to manually review numerous candidate variants per case—up to millions per case in whole-genome sequencing (WGS) despite increasingly rich annotation resources and phenotype-aware prioritization tools. To address this operational hurdle, we developed the Phenotype-Centric Variant Classifier (PCVC), a machine-learning framework designed to convert candidate variant prioritization into a binary triage task for streamlined rapid identification of reportable pathogenic variants. Methods: The training dataset comprised 108,614 independent variant-by-case observations from 915 pediatric cases evaluated for suspected rare genetic disorders who underwent clinical whole-exome or whole-genome sequencing (WES/WGS). The dataset contained 610 clinically reported pathogenic or likely pathogenic (P/LP) variants (positive class), while variants of uncertain significance (VUSs), benign/likely benign (B/LB) variants, and non-diagnostic background variants were coded as the negative class. VUSs were grouped with B/LB variants for operational purposes, but this is not an assertion that these variants are biologically benign. Each observation was annotated across 254 features, integrating gene-level metrics, functional consequences, in silico effect predictions, population frequencies, ClinVar evidence, and Exomiser-derived pathogenicity and phenotype-matching scores. Models were trained using H2O.ai AutoML across multiple positive-class weighting strategies. External benchmarking was conducted on a GA4GH Phenopacket-derived simulated spike-in dataset (N = 60,000 observations, 279 positive variants) processed through an identical annotation pipeline. Results: Systematic evaluation of AutoML configurations established that tree-based algorithms consistently outperformed deep learning and linear models, with Extreme Gradient Boosting (XGBoost) dominating the top ranks. A leading single model (t2_weight_1_v2_no_clinvar) achieved an area under the precision–recall curve (AUCPR) of 0.872 on an internal holdout dataset (N = 21,785 observations) with a sensitivity of 0.83 (recall confidence interval/CI 0.758–0.897), 0.825 positive predictive value (PPV with 95% CI 0.745–0.895), and an F1-score of 0.828. On the external GA4GH dataset, it achieved 0.8817 sensitivity (95% CI: 0.8387–0.9145) and 0.6685 PPV. A multi-model decision-tree meta-learner identifies a high-confidence tier-1 node whose variants reached 0.9855 and 0.955 node-level PPV on the internal holdout and external GA4GH datasets respectively. Based on a previous estimation on negative cases’ variants and assuming approximately 20% positive diagnosis rate for rare disease cohort, approximately 0.5 predicted candidate variants, including 0.42 false-positive calls, would be called per case. So, integration of the PCVC into the clinical interpretation workflow may reduce candidate review burden to approximately 0.5 predicted candidate variants and 0.42 false-positive calls per case, while capturing approximately 84–96% of clinically reportable pathogenic variants in the first pass. SHapley Additive exPlanations (SHAP) analysis demonstrated that phenotype-matching metrics (EXOMISER_VARIANT_SCORE and HUMAN_PHENO_SCORE) and Maverick deleteriousness scores were primary drivers of model predictions. Conclusions: The PCVC provides a robust, phenotype-centric pathogenicity triage framework trained on expert-curated clinical labels. These retrospective results support projected workflow acceleration by the PCVC in generating actionable binary triage tiers instead of ranked candidate lists. The PCVC is thus projected to reduce manual interpretation burden while maintaining high sensitivity. Prospective validation and workflow time studies are required to confirm turnaround-time impact on rare disease clinical sequencing workflows.

Algorithms doi: 10.3390/a19110949

Authors: Mohsin Shahzad Muhammad Amjad Muhammad Bux Alvi Serosh Karim Noon Kashif Hussain Memon Fareeha Zahoor

k-nearest-neighbor (k-NN) methods remain fundamental to pattern recognition, classification, and image analysis because they are simple, non-parametric, and effective for nonlinear decision structures. However, k-NN methods are constrained by two related limitations: explicit pairwise distance computation and reliance on fixed neighborhood parameters, most notably the parameter k in k-NN and the fixed window-size parameter Diff in Fast Geometrical Extraction of Nearest Neighbors (FGENN). FGENN introduced a distance-free method by transforming data into rank space and extracting nearest neighbors through a geometric hyper-window rather than metric-distance ordering, but its fixed window size does not adapt to local density variation, causing over-selection in dense regions and under-selection in sparse or boundary regions. In this paper, we propose Density-Adaptive Rank-Space Neighbor Extraction using Boundary-Corrected Hyper-windows (DARNE), an adaptive rank-space neighbor-extraction and geometric neighborhood-filtering mechanism that generalizes FGENN, where its fixed window and exhaustive sweep are replaced using an adaptive window size governed by a two-stage local-to-global occupancy-density control law. DARNE outputs ordinal rank neighborhoods in two steps: Stage 1 checks the local rank-lattice occupancy of the query neighborhood by using boundary-corrected hyper-window volumes; Stage 2 analytically scales the window, based on the local-to-global density ratio, without calculating pairwise Euclidean distances. DARNE also outperforms fixed-window FGENN in query latency for all benchmarks and brute-force k-NN variants on the IRIS, SONAR and Covertype datasets, achieving speedups of 2.78×, 11.55× and 21.14×. DARNE achieved accuracies of 97.78%, 94.09%, and 69.12% on the IRIS, SONAR and Covertype datasets, respectively, at a lower computational cost for queries relative to existing methods. Together, these results demonstrate that two-stage density-adaptive windowing in rank space is a reproducible and computationally scalable neighborhood-filtering framework for a wide range of multi-dimensional classification and hashing benchmarks.

Algorithms doi: 10.3390/a19110948

Authors: Duván Petro-Hernández Arcadio A. Pérez-Solano Victor A. Morales-Pérez Oscar E. Coronado-Hernández Modesto Pérez-Sánchez Rafael D. Méndez-Anillo Helena M. Ramos

Non-Revenue Water (NRW) remains a major operational challenge for water utilities, particularly in small-sized distribution systems where limited monitoring and data availability constrain the development of efficient loss-management strategies. This study proposes a data-driven surrogate modelling framework based on deep learning (DL) algorithms to support NRW prediction within a telemetry-enabled operational control strategy. The proposed methodology framework integrates hydraulic modelling, pressure-dependent leakage representation, District Metered Areas (DMAs), Pressure-Reducing Valves (PRVs), and telemetry for the analysis of alternative operating conditions in the drinking water distribution system of Arboletes, Antioquia, Colombia. Real water losses were represented using the Germanopoulos pressure-dependent formulation with a pressure exponent of γ = 1, consistent with the predominantly high-density polyethylene network. Three operational scenarios were evaluated through 24 h extended-period simulations: current operation without sectorisation, DMA-based sectorisation, and DMA-based sectorisation combined with PRV-based pressure control. The resulting hydraulic simulations were used to develop and evaluate deep learning surrogate models—including Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Multilayer Perceptron (MLP) architectures—for NRW prediction. The DL models successfully reproduced the temporal dynamics of NRW across all scenarios, achieving RMSE values of 0.551%, 0.482%, and 0.223% for the no-DMA, DMA, and DMA+PRV scenarios, respectively. Both the Gated Recurrent Unit (GRU) and Multilayer Perceptron (MLP) architectures demonstrated their applicability to NRW prediction under the simulated operating conditions, as part of an exploratory study. Hydraulic results showed that DMA implementation alone produced limited reductions in water losses, whereas the combined DMA+PRV strategy reduced the average daily NRW from 43.04% to 30.99% (with the daily volumetric NRW decreasing from 42.39% to 27.18%). The results demonstrate the potential of deep learning-based surrogate modelling to accurately reproduce NRW dynamics and provide a data-driven tool that, once trained, can reproduce NRW dynamics without repeatedly executing the hydraulic model, supporting the operational analysis of drinking water distribution systems. These results are based on a single 24 h simulation scenario for one case-study network and should be regarded as proof of concept rather than generalised to other seasons or system types without further validation.

Algorithms doi: 10.3390/a19110947

Authors: R. Raja Subramanian Sri Ram Ganesh M M. Ali Fauzi Deshinta Arrova Dewi

Diseases of mango leaves have the potential to greatly diminish agricultural yields, and the task of identifying these diseases from pictures captured in field is complicated by varying light levels, complex backgrounds, appearance of the disease, and image quality. While modern deep learning techniques have achieved good results in terms of accuracy of classification, most current techniques are either computationally intensive or offer a limited view into the relationship between compactness of the models, representation of local and global features, robustness, and interpretability. This paper presents MangoXNet, a novel lightweight hybrid CNN–Transformer framework for optimal classification of mango leaf diseases. The hybrid framework features a compact convolutional backbone: MobileNetV3Small, depthwise separable refinement, squeeze-and-excitation-based channel re-calibration, compact Transformer-based contextual encoding, and CNN–Transformer feature fusion. The main contribution of this study lies not in the introduction of separate components but in task-related integration of local and global feature representations into a compact model. The model achieves a classification accuracy of 99.83% on the MangoLeafBD dataset, together with macro-F1 and MCC scores of 99.80% and 99.78%, respectively. The model comprises 1.17 million parameters and occupies a size of 6.90 MB in storage, and the average inference time is estimated at 11.70 ms per image in the experiment. Moreover, performance of the model during fivefold cross-validation is found to be 99.44 ± 0.19%, a trait that proves stability of model performance across different validation folds. Explainability analysis using Grad-CAM and Grad-CAM along with LayerCAM methods helps to confirm that model predictions are mainly linked to leaf regions relevant to disease. Results achieved prove that it is possible to build a compact local–global representation model that has high classification ability without major increase in computational processing effort. However, the model’s validation on held-out test sets, different agricultural environments, and proper edge hardware should be performed in order to make sure the model can fit a broader scope of conditions.

Algorithms doi: 10.3390/a19110946

Authors: Zainab A. Hameed Sadiq H. Abdulhussain

Video surveillance systems play a vital role in modern security, transportation, and public safety infrastructure. Most of the analytical functions of an intelligent surveillance system, from detection and tracking to ROI-aware compression, begin by isolating a Region of Interest (ROI), yet the algorithms that perform this step are scattered across the computer-vision, video-coding and sensor-network literatures, use the term “ROI” in different senses and report results on metrics that cannot be compared directly. This narrative review classifies and critically analyzes ROI extraction algorithms for video surveillance published between 2020 and 2026. A comparison with survey papers published between 2021 and 2026 shows that none is centered on the extraction step itself. The reviewed methods are organized by explicit criteria (hand-crafted or learned features, learned or online-estimated decision rules, annotation dependence and output granularity) into background-subtraction, classical machine-learning, deep-learning and hybrid families, and the deep-learning family is further divided into six groups. Five semantic variants of the ROI are distinguished, and reported results are compared only within the metric family appropriate to each variant. Widely used video surveillance datasets are compiled, described and compared. The analysis shows that the evaluation base of the field is narrow, with fourteen of the compared studies reporting their principal results on CDnet 2014 and none reporting run-to-run variance. Because the compared studies differ in dataset, subset, training split and averaging, the reported values indicate operating ranges and do not show that any method family, deep learning included, is generally superior to another. Research directions are derived from these limitations, including cross-dataset evaluation with temporal-stability measures, codec-aligned ROI extraction for edge devices, motion-gated promptable segmentation, extraction from moving cameras, and the joint design of ROI extraction and privacy protection.

Algorithms doi: 10.3390/a19110945

Authors: Wenquan Jiang Yuehao Yan Jiaqiu Zheng Ping Deng Jiaxi Dong

Aiming at the operational challenges of low-altitude UAV path planning in complex 3D environments—where rapid feasibility, trajectory smoothness, and safety margins must be balanced under practical computational budgets—this paper proposes a two-layer collaborative framework (GA-B-A*) integrating bidirectional improved A with a customized genetic algorithm. The key technical innovations are threefold: (i) a bidirectional A* variant with a composite cost function that encodes terrain-aware safety margins, reducing search space by approximately 50% compared with unidirectional search; (ii) a customized genetic algorithm with domain-specific operators (path-shaping mutation, altitude-correction mutation) that preserve path continuity during evolution; and (iii) a two-stage smoothing-and-verification pipeline that enforces kinematic feasibility (pitch angle ≤ 30°, curvature constraints) on the final trajectory. Across six scenarios with 0/100/300 obstacles, GA-B-A* shortens path length by 20%~30% (Cohen’s d = 1.90, 95% CI: [2.87, 3.97]), reduces maximum turning angle by ~50% (Cohen’s d = 4.52), and lowers the defined path cost by 15%~18% compared with A*, with all improvements statistically significant at p < 0.001. All paths satisfy pitch angle and curvature constraints with 100% compliance.

Algorithms doi: 10.3390/a19110944

Authors: Sadia Sahar Muhammad Younas Shazia Riaz Rafiullah Khan Mohib Ullah Kashif Saleem

Cross-project defect prediction (CPDP) is a significant solution in software defect prediction (SDP) when historical data is unavailable. In CPDP, the defect data of another project (called the source project) is used to train the model. Different training and test projects may have different feature distributions. Divergence in data distribution between source and target projects adversely affects prediction performance, and the semantic information of the tokens cannot be captured adequately. However, most of the existing deep learning approaches assume that these two projects have a uniform distribution, so the features extracted from these deep learning models are not optimal. Moreover, these approaches neglected the information about the class to which the sample belonged and were unable to discriminate the samples according to the class information. To address this issue, we proposed a framework, TC3, that exploits the combination of contrastive and transfer learning integrated with the CNN-CodeBERT model to learn class-discriminative information and obtain more transferable features for the target project. The proposed model was evaluated on PROMISE datasets. The proposed model achieved an absolute F1-score improvement of 0.047 to 0.107, corresponding to a relative improvement of 7.89% to 19.96%, and an absolute AUC improvement of 0.014 to 0.148, corresponding to a relative improvement of 2.18% to 29.19% as compared to the baseline methods.

Algorithms doi: 10.3390/a19110942

Authors: Teerath Kumar Raja Vavekanand Muhammad Javed Shevta Kumari Sajjad Hussain Bhutto Edward Curry Muhammad Umair Kislay Raj Mamoona Naveed Asghar

Deepfake generation now combines autoencoders, generative adversarial networks, diffusion models, transformers, neural rendering, and audio-conditioned synthesis. This survey presents a structured narrative synthesis of image, video, audio, and multimodal generation and detection, organised around the relationship between synthesis mechanisms and forensic evidence. Literature selection follows explicit source, search-term, eligibility, deduplication, and evidence-appraisal criteria. Generation methods are compared by conditioning requirements, identity preservation, temporal coherence, and computational cost. Detection methods are analysed through spatial, frequency, temporal, physiological, device, audio–visual, and provenance evidence, with attention to the assumptions that make each cue informative. Evaluation definitions distinguish reconstruction quality, distributional realism, identity and speech fidelity, classification performance, calibration, robustness, localisation, and efficiency. Results are interpreted within their dataset and protocol rather than pooled into an architecture-wide ranking. The comparative analysis separates demonstrated benchmark behaviour from proposed directions in open-set detection and multimodal synthesis. Practical recommendations prioritise held-out generators and datasets, realistic processing pipelines, explicit operating thresholds, calibrated confidence, faithful explanations, and provenance validation. Together, these requirements connect algorithmic performance to the evidence needed for trustworthy synthetic-media assessment.

Algorithms doi: 10.3390/a19110943

Authors: Nina Golyandina Pavel Dudnik Egor Poteshkin

Automatic decomposition of a time series into trend, periodicities, and noise remains open in singular spectrum analysis (SSA): trend extraction is reliably automated, but automatic identification of periodic components is still challenging, especially for weakly separated or non-dominant components. We propose SSA-Aid, a modular framework inferring the number and structure of periodic components from the data. It includes separability improvement and two complementary periodicity-identification strategies with an empirical stopping rule against over-extraction. Its most general variant, SSA-AidMC, extracts dominant and non-dominant components using Monte Carlo SSA and refines them jointly. In the tested configurations, both strategies perform well on four synthetic examples, while SSA-AidMC provides consistent gains for non-dominant components, including those with modulated amplitudes. In forecasting, no single method dominates; standalone SSA variants are more accurate in some signal structures, conventional baselines or hybrid variants (SSA decomposition with an external trend forecaster) in others. The framework uses generic tuning parameters with recommended defaults, and its main computational cost comes from repeated MC-SSA assessments, each requiring a fixed number of surrogate series. This cost is the price of its central advantage: no seasonal periods or functional forms must be specified within the general SSA signal model.

Algorithms doi: 10.3390/a19110941

Authors: Mohammadreza Nehzati

We ask whether a variational autoencoder whose latent space is given the differentiable structure of the 3-sphere can, under curvature-driven regularization, learn a metric of constant sectional curvature, and we report a study designed to make that question operationally well posed. Three ingredients are developed. First, a fully differentiable Riemannian pipeline (Christoffel symbols, Riemann tensor, Ricci tensor, Ricci scalar, and their scale-invariant combinations) together with a numerical test suite that validates it against the eight Thurston model geometries, for which the exact scalar curvature, Ricci endomorphism spectrum, and Einstein defect are recovered to or better. Second, a diagnostic hierarchy: we prove that the coefficient of variation of the Ricci scalar, the quantity used as the primary criterion in earlier versions of this programmer, vanishes identically for all eight homogeneous geometries and therefore cannot identify any of them; we replace it by the spectrum of the Ricci endomorphism, which is a complete, scale-invariant signature of a model geometry, and by the dimensionless Einstein defect, which vanishes if and only if the metric has constant sectional curvature (Schur’s lemma in three dimensions). Third, a well-posedness analysis: on the stereographic chart that the latent space actually occupies, a complete metric of constant curvature can only be or ; requires an explicit compactification. The implemented model imposes no such condition, and we show that its metric network does not enter the likelihood at all, so the geometry is not identified by the data. Recomputing all geometric quantities from the released checkpoints in float64 with validated derivatives changes the earlier conclusions qualitatively: the median Einstein defect of the learned metric is (baseline) and (Hessian penalty), comparable to the anisotropic model geometries (1.22 to 4.90) and far from the space forms (); the signature matches none of the eight, whatever the tolerance, fewer than 1% of latent points lie within 0.25 of a model signature, and the Hessian penalty has no statistically resolvable effect (its confidence interval for the median defect overlaps the baseline’s). The Ricci-variance regulariser diverges within four epochs; we identify the mechanism as a scale degeneracy of the functional combined with third-order derivative amplification, and show that gradient clipping cannot prevent it. We conclude that the hypothesis, as stated for a fixed topology, is falsifiable but was not adequately tested by this architecture, and we specify the conditions under which a decisive test becomes possible. We also implement the first three of those conditions in a pilot study: in the round conformal class, with a scale-invariant loss and a volume coupling, the same latent cloud admits a metric that satisfies the geometrisation criterion to ten per cent at a cost of nats per point of density fit, whereas the freely trained metric satisfies no tolerance and is anti-correlated with the density of the data. The obstruction documented here therefore lies in the formulation rather than in the spherical topology.

Algorithms doi: 10.3390/a19110940

Authors: Guadalupe Jasmin Muñoz-Salazar Heriberto Adamas-Pérez Jesús Dario Mina-Antonio Jesus Aguayo-Alquicira Mario Ponce-Silva Guillermo Ramírez-Zúñiga Susana Estefany De León-Aldaco

This article presents a hybrid (offline–online) algorithm for controlling a single-phase full-bridge inverter connected to the grid using unipolar SPWM modulation and an LCL filter. The methodology could reduce the real-time computational load by shifting the solution of transcendental equations to a preliminary stage. Using Bolzano’s Theorem, the approximate switching instants are precalculated and stored in a lookup table (LUT) with high temporal resolution. During online execution, the algorithm synchronizes power injection by indexing the switching times in a lookup table (LUT) based on the phase extracted by a discrete phase-locked loop (PLL). The main contribution lies in the integration of an analytical feedforward angular compensation, which cancels the intrinsic phase shift of the LCL filter and allows operation at unity power factor in open-loop, eliminating the need for complex current controllers such as Proportional-Resonant (PR) schemes.

Algorithms doi: 10.3390/a19110939

Authors: Anton Pljonkin Alexander Shelupanov

Measurement-device-independent quantum key distribution (MDI-QKD) eliminates detector side-channel attacks by relocating all measurements to an untrusted intermediate node. However, its practical implementation critically relies on picosecond-level temporal synchronization between spatially separated users. In this work, we present a physically motivated autonomous synchronization algorithm for fiber-based MDI-QKD networks that does not require auxiliary optical channels or shared clock references. The method exploits round-trip optical pulse propagation and statistical signal detection in the presence of Gaussian noise. We derive analytical expressions for false-alarm probabilities, quantify detection reliability, and demonstrate through numerical modeling that a scanning resolution of 10 ps is achievable for channel lengths of up to 100 km with realistic optical power levels. The actual timing accuracy is limited by detector, source, and electronic jitter and is not quantified in this work. The proposed approach improves the scalability and robustness of MDI-QKD architectures and is directly applicable to metropolitan and backbone quantum networks.

Algorithms doi: 10.3390/a19110938

Authors: Velin Kralev Radoslava Kraleva

This study investigates exact algorithms for solving the Travelling Salesman Problem (TSP), formulated as the problem of finding a minimum Hamiltonian cycle in a complete weighted undirected graph. Two recursive algorithms are presented and analyzed. The first algorithm, denoted RecursiveHamiltonianCycle (RHC), performs an exhaustive recursive search of the solution space, while the second algorithm, denoted PrunedHamiltonianCycleSearch (PHCS), applies an early branch-pruning mechanism to eliminate non-promising partial solutions. The main objective of this research is to evaluate the influence of recursive branch pruning and initial upper-bound selection on the computational efficiency of exact TSP search procedures. A series of computational experiments was conducted using complete graphs containing between 12 and 21 vertices. The obtained results show that both algorithms always produce identical optimal Hamiltonian cycles. However, the PHCS algorithm requires significantly fewer recursive calls and substantially less execution time. Additional experiments demonstrate that the quality of the initial upper bound has a substantial impact on pruning efficiency. A nearest-neighbor heuristic was therefore used to generate high-quality initial solutions, resulting in further reductions in the search space. The results confirm that combining heuristic initialization with exact branch-pruning search significantly improves the practical applicability of exact TSP algorithms while preserving solution optimality.

Algorithms doi: 10.3390/a19110937

Authors: Xiaoquan Chen Jin Xie Shoujun Xu

A geometric algorithm is developed for quadratic uniform trigonometric B-spline interpolation with segment-specific shape parameters. The interpolation conditions generate all control points recursively from an initial vertex. For prescribed parameters, the complete derivative calculations give a weighted initial vertex that uniquely minimizes each of the three piecewise derivative penalties. The segment-dependent weights are retained in the subsequent constrained parameter optimization. Individual and normalized weighted objectives define different choices of curve, while a consistent squared graph-error functional is used for function approximation. The geometric interpretation of the derivative penalties is distinguished from arc length, bending energy, and curvature continuity. Recomputed spatial, closed planar, and function-graph examples report objective values, interpolation and graph errors, and join diagnostics, with comparisons against polynomial constructions. The results exhibit objective-dependent trade-offs and do not establish uniform superiority or general shape preservation.

Algorithms doi: 10.3390/a19110936

Authors: Jianfeng Li Junhao Chen Ding Jiang Zhenfan Zhang Ziyang Hu

The quadratic assignment problem (QAP) is a fundamental NP-hard combinatorial optimization problem characterized by a quadratic objective and discrete permutation constraints. Continuous relaxation makes neurodynamic optimization applicable to QAP, but a single search trajectory may converge to a poor local solution because the relaxed problem is nonconvex. To address these challenges, a collaborative neurodynamic optimization (CNO) framework integrating majorization-minimization (MM) and a momentum recurrent neural network (MRNN), termed CNO-MM-MRNN, is proposed. MM constructs quadratic surrogate subproblems for a penalized doubly stochastic relaxation, and the MRNN approximately solves these subproblems to refine each candidate. Multiple candidates exchange personal-best and global-best information to explore different attraction regions. Feasible permutations decoded from relaxed matrices are evaluated using the original QAP objective, which primarily guides the selection of personal-best and global-best records. The selected assignments are fed back into the continuous candidate states to guide subsequent updates. The best, average, and worst percentage deviations (BPD, APD, and WPD), averaged over fifteen QAPLIB instances, are 0.144%, 0.192%, and 0.245%, respectively. The inner MM and global-attainment results depend on assumptions that are not fully established for the finite-step implementation. They therefore do not guarantee finite-budget global optimality.

Algorithms doi: 10.3390/a19110935

Authors: Qing Ye Gaoxian Peng Xinyan Huang Shan Liu Di Hou

Learning-based routing for the Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) must let a policy visit infeasible states in training, yet deploy routes that provably respect every time window under stated assumptions. In existing deep reinforcement learning (DRL) methods, however, time information serves a single constraint-handling mode, fixed for training and inference. We propose ETHC-VRC, a DRL construction method whose explicit time state is shared by the decoder context and the hard time-window mask, so that a policy trained with soft penalties can be decoded under this mask without retraining; its hybrid constraint strategy decodes each instance in both modes and keeps the lower-cost solution. On 20- to 100-node instances, greedy decoding solves 10,000 instances within seconds on a GPU at a lower hybrid cost than the greedily decoded DRL baselines (not significantly against MVMoE at 100 nodes). Sampling 1280 solutions per mode lowers it by another 3.8–4.9% at 0.07–0.32 s per instance, to within 2.1–4.3% of OR-Tools run for 10–120 s per instance and below OR-Tools at equal time. In zero-shot generalization across five scales, greedy decoding stays ahead of TPA-Reward on all 15 training–test combinations and of MVMoE, which has 5.3 times the parameters, on 8. A single model thus offers an explicit quality–latency trade-off, which we expect to suit both latency-critical and quality-critical logistics applications.

Algorithms doi: 10.3390/a19110934

Authors: Néstor A. García-Rojas Miguel A. Jiménez-Domínguez Saúl Zapotecas-Martínez Raquel Díaz-Hernández Leopoldo Altamirano-Robles

Particle Swarm Optimization (PSO) frequently suffers from premature convergence in complex, multimodal landscapes. While local neighborhood topologies help mitigate stagnation, they typically compromise convergence speed. To resolve this trade-off, this paper introduces Particle Swarm Optimization with Wave Agitation and Covariance Matrix Adaptation (PSO-WA-CMA). Inspired by the collective evasion and wave propagation dynamics observed in biological flocks, the proposed framework injects structured, spatio-temporal perturbations rather than isotropic, random mutations. The main contributions of this work are threefold: (1) a scale-invariant adaptive stagnation detection mechanism that accurately triggers interventions independent of the objective function’s magnitude; (2) a distance-dependent wave propagation model that spatially regulates perturbation intensity; and (3) a covariance-guided directional search that aligns the wave agitation with the swarm’s historically promising search trajectories. The method was rigorously evaluated on 40 benchmarks—23 classical functions, 10 CEC-C06 2019 functions, and 7 engineering design problems—against 18 reference algorithms. PSO-WA-CMA demonstrated robust competitiveness, achieving a higher number of statistically significant wins than losses against all eight evaluated PSO variants. Compared to canonical PSO, the proposed method secured a favorable record of 12 wins and eight losses, maintaining competitive parity in the remaining 20 cases where no significant difference was detected. While excelling particularly on classical and complex real-world engineering problems, the broader evaluations reinforce the problem-dependent nature of metaheuristics. Overall, these findings confirm PSO-WA-CMA as a highly effective extension of PSO that successfully balances exploration and exploitation without permanently disrupting convergence.

Algorithms doi: 10.3390/a19110933

Authors: Mohammad Rajabi Nasab Muhammad Muzammal Islam Rodrigo Fiorotti Sergio Bruno Hamid Yaghobi Massimo La Scala

Power system stability can be threatened by severe transient disturbances. With the growing penetration of wind turbines in modern power systems, the development of effective real-time transient stability assessment methods for wind-dominated grids is becoming essential. Among the available wind turbine technologies in wind farms, Doubly-Fed Induction Generator (DFIG) is one of the most widely used solutions. Thus, this paper proposes a new method for assessing the transient stability of a single DFIG following large disturbances in a multi-machine power system. The presented method continuously monitors system stability by estimating the critical rotor-speed limit from the power–speed (Pe–ωr) characteristic of the grid-integrated DFIG and subsequently using this limit to quantify the available stability margin. The measured rotor speed is then continuously compared with this computed critical speed. As soon as the speed exceeds this critical value, an unstable state is detected. A key advantage of this method is that it does not require predefined relay settings, detailed network parameters, or fault-clearing time information, making its deployment rather straightforward for protection applications. The proposed method is validated through comprehensive simulations using both a Single-Machine Infinite Bus (SMIB) system and a WSCC 9-bus multi-machine network under a wide range of scenarios. The simulation results demonstrate that the proposed method accurately estimates the critical rotor speed and reliably detects unstable conditions, even when the grid topology changes.

Algorithms doi: 10.3390/a19110932

Authors: Zhi Li Zhehao Liu Wenjuan Xie Hongguang Sun Jing Fu

Deep Q-Network (DQN) has demonstrated significant application value in path planning. However, several challenges remain in UAV path planning tasks, such as difficulty in perceiving dependencies among the components of the input state, imbalanced experience sampling, and the lack of constraints on action selection. To address these issues, this paper proposes a DQN framework based on a semantic-group self-attention mechanism (SGSA-DQN). First, a semantic-group self-attention Q-network (SGSAQ-network) is designed to improve the accuracy of the network’s Q-value output by capturing the dependencies among semantic component groups in the state vector. Second, a guidance- and perception-based action selection algorithm (GPAS) is proposed to reduce unnecessary exploration in the early training stage and accelerate model convergence. Then, a coverage-constrained experience replay mechanism (CCER) is proposed to balance experience sampling probabilities across different state spaces and enhance model robustness. Finally, a Lyapunov safety constraint (LSC) is introduced to mask unsafe actions and improve the task success rate. Experimental results demonstrate that the proposed method achieves an average success rate of 79.17% in UAV path planning tasks.

Algorithms doi: 10.3390/a19110931

Authors: Wang Yao Weiyuan Huang Chong Wang Mingjun Li Yang Shen Xintai Man

Dual tower-crane crossing operations require timely identification of collision risks as the spatial relationship between the two jibs changes. Conventional systems based on crane-state sensors and static safety zones provide limited information about the surrounding three-dimensional workspace, while sparse point clouds, occlusion, and class imbalance complicate reliable jib recognition. This study develops a LiDAR-based method that converts point-cloud segmentation into geometric models for short-term collision-risk prediction. A geometric-temporal dynamic graph convolutional neural network (GT-DGCNN) combines geometric priors, class-balanced learning, and temporal consistency to identify the jibs and tower structures of both cranes. The segmented jibs are represented by capsule-based envelopes and evaluated using swept-volume prediction, dynamic safety distance, and time to collision. The complete method achieves a mean intersection over union of 87.3% and a jib recall of 91.7%, exceeding the DGCNN baseline by 6.4 and 7.4 percentage points, respectively. The missed-alarm rate decreases from 7.8% to 2.4%, with a processing rate of 17.2 frames per second. These results support the use of geometric-temporal point-cloud perception for jib recognition and collision warning in the evaluated dual-crane scenarios.

Algorithms doi: 10.3390/a19110930

Authors: Thisarani Atulugama Udara Jayasinghe Anil Fernando

This paper presents a secure image transmission framework for superposition-based quantum communication systems that integrates quantum key distribution (QKD) with one-time-pad encryption. Existing superposition-based image transmission systems achieve reliable image reconstruction over noisy channels but do not provide confidentiality: an eavesdropper intercepting the transmission can recover the image. To address this limitation, we introduce a key-based security layer and analyse its interaction with the quantum modulation and channel coding components. In the proposed approach, a secret key is generated using the BB84 quantum key distribution protocol and then used to encrypt the image data before Hadamard modulation. The system is evaluated both with and without polar channel coding. Simulation results show that encryption does not degrade image quality at the legitimate receiver: its reconstruction performance is indistinguishable from that of the corresponding unsecured baseline under the evaluated conditions. In contrast, the eavesdropper obtains only noise-like plaintext reconstructions, with an empirical PSNR floor of approximately 8.1 dB, even under more favourable channel conditions. Attempts to intercept the BB84 key-generation process remain detectable through the quantum bit error rate. Polar coding provides approximately 8 dB of performance gain for the legitimate receiver; although it can improve ciphertext recovery, it does not improve the eavesdropper’s plaintext reconstruction under the evaluated conditions. These results demonstrate confidentiality without an additional reconstruction-fidelity penalty, while leaving key-rate and protocol costs as separate practical considerations.

Algorithms doi: 10.3390/a19110929

Authors: Sergio A. Pérez-Rodríguez José M. Álvarez-Alvarado Diana-Margarita Córdova-Esparza Julio-Alejandro Romero-González Perla Yazmín Sevilla-Camacho José Billerman Robles-Ocampo Juvenal Rodríguez-Reséndiz

Deep learning solar forecasters depend strongly on their hyperparameters, yet comparisons of hyperparameter optimization methods usually rely on few runs per optimizer, confounding the optimizer with the random seed. A protocol is proposed that tests whether optimizers are practically equivalent and how reliable published comparison designs are. Seven methods (four metaheuristics, two model-based methods, one of them multi-fidelity, and Random Search) tuned three deep networks for one-hour-ahead forecasting of global horizontal irradiance in Querétaro, Mexico, with public reanalysis data, a shared search space, one hundred evaluations, and thirty paired seeds, in nine hundred and thirty runs across all experiments. The tuned models improved on smart persistence by 4–11% in mean absolute error. With the 10-epoch search objective, no pair of optimizers differed after Holm correction, 62 of 63 pairs were equivalent within 1 W/m2, and the optimizer explained 0.8% of the error variance, compared with 7.9% for the seed. In a higher-fidelity check with 50 epochs, two guided methods improved on Random Search by 0.74 W/m2, significantly for one of them, while remaining equivalent to each other, suggesting that, in this setting, objective fidelity mattered more than the choice among guided optimizers. In emulations of published designs, even with 30 runs, 2 studies agreed on the winner in at most 55% of cases. A validated surrogate benchmark, code, data, and predictions are released.

Algorithms doi: 10.3390/a19110928

Authors: Robert-Emanuel Brezoaie Beatrice-Iuliana Uta Bianca-Gabriela Dobre Maria-Iuliana Dascalu

Gamified e-learning platforms and Artificial Intelligence (AI)-driven adaptive learning systems can improve learner engagement, but existing solutions rarely combine the two, and few explain their adaptive decisions to learners. To address this gap, we propose LearnSpace, a hybrid AI–gamification framework for responsive e-learning that adds an Explanation Layer, based on Explainable Artificial Intelligence (XAI), to a deterministic reward architecture. The framework merges classical gamification mechanics with an AI personalization module in which experience points, rewards, and difficulty adjustments follow explicit weighted formulas. A Thompson-sampling bandit learns which feedback works best after a mistake. For each adaptive decision, the Explanation Layer names only the factors that drive it and turns them into a short justification for the learner. The design draws on a qualitative overview of 14 platforms, and the LearnSpace mobile prototype implements the gamification components and the bandit. In a pre-registered simulation with 500 synthetic learners per run, the first explanation rule named a factor that had no effect on the decision in 95.3% of difficulty adjustments, while the decision-specific rule never did. The bandit had lower regret than uniformly random feedback allocation. Effects on real learners remain to be tested.

Algorithms doi: 10.3390/a19110927

Authors: Yi Pan Yufei Lei Runjie Li

The rapid growth of the digital economy has made the accurate identification of Core Data Assets (CDAs) increasingly important for enterprise data governance. Existing approaches, however, often rely on static, homogeneous dependency modeling or node-centric analysis, limiting their ability to capture complex value propagation patterns and the structural importance of data in enterprise business processes. To address these limitations, this study proposes a graph learning framework with dynamic (structure-adaptive) edge weighting for automated CDA discovery using data lineage graphs. The framework first introduces a dynamic edge-weighting mechanism that integrates node influence and relationship scarcity to represent heterogeneous value propagation between data entities. The framework then constructs a comprehensive feature representation by combining global influence, local topological characteristics, and functional connectivity patterns, enabling the model to capture different aspects of data-table importance. Based on these representations, we develop an ensemble learning model with group-based cross-validation and automated hyperparameter optimization for robust CDA classification. Experiments on five real-world lineage graphs from a single enterprise data warehouse demonstrate the framework’s effectiveness. The model achieves average AUC and F1 scores of 0.926 and 0.753, respectively, and outperforms the evaluated baseline methods in terms of F1-score and Precision; the strongest competing baseline attains higher AUC and Recall but substantially lower Precision and F1-score. The results show that the multidimensional structural features provide the main share of the improvement, that the dynamic weighting mechanism contributes a modest additional benefit, and that the approach is effective for lineage graphs of moderate scale in enterprise data governance.

Algorithms doi: 10.3390/a19110926

Authors: Emily Weng Georg Hahn

We consider the computation of the network modularity, a common measure of the structure of networks or graphs that quantifies the division of a network into modules or clusters. The computation of the network modularity is an NP-hard problem, thus making its exact computation often infeasible in practice. This has resulted in the development of several heuristics in the literature. In this work, we consider the exact computation of the network modularity value with the help of a branch-and-bound algorithm. Our algorithm is guaranteed to find the optimal clustering that maximizes the network modularity; however it achieves this without a full exploration of the search space. This is accomplished by bounding partial solutions and discarding them if it can be foreseen that a partial solution will not yield an improvement over an existing solution. We assess our algorithm with respect to accuracy and runtime and compare it to a brute-force approach as well as two benchmark algorithms, simulated annealing and the Louvain method.

Algorithms doi: 10.3390/a19110925

Authors: Long Huang Zongshuai Cui Jianzhong Zhang Hao Wang Rizhao Fan Xiaoyu Yang

Intelligent applications in the coal industry require models to read professional images such as safety tables and engineering drawings. Small multimodal models lack sufficient domain-specific knowledge, whereas large multimodal models are expensive to deploy. Black-box distillation with a commercial API as the teacher provides a feasible way to transfer the image–text understanding capability of large models to a locally deployable small model. However, conventional distillation usually optimizes only the average performance, which hardly meets the requirements of high-risk coal-mining scenarios in terms of the performance floor and the concentration of the score distribution. This paper, therefore, proposes two improvements that act only during training. The first is difficulty-weighted distillation (Diff), which uses the cross-entropy loss of the base student on the teacher-generated answer as a computable proxy for sample difficulty and oversamples hard samples so that the limited gradient budget is preferentially assigned to weak samples. The second is semantic centroid regularization (SCR), which aligns the last-token representations of the student toward the semantic centroids of their categories. The two components are combined to form the Diff-SCR training framework. The weights and the centroids are computed once and fixed before training, so the model structure is unchanged and no inference overhead is added. We distill Qwen3.5-0.8B using 1800 industry-specific images. The 223 test images are then independently evaluated by three human raters, and the mean of their scores for each image is used as the final score. Compared with ordinary distillation, Diff-SCR raises the average score from 73.37 to 80.82 and the average score among the lowest-scoring 25% of samples from 29.7 to 42.2, and reduces the share of complete failures from 2.2% to 0.4%. A 2×2 ablation shows that the two improvements are complementary: SCR alone already pushes the share of complete failures down to 0.9% and the average score among the lowest-scoring 5% of samples up from 7.8 to 9.1. The results show that, without increasing the inference cost, reallocating the training gradient resources improves the average performance while markedly raising the performance floor and compressing the dispersion of the score distribution.

Algorithms doi: 10.3390/a19110924

Authors: Haijun Zhao Yangshuo Zhang Hong Zhang Huaihao Zheng Pengfei Zhao Bin Zhang Zhaowen Chen Yiran Peng

Electroencephalogram (EEG)-based motor imagery (MI) decoding is vital in brain–computer interface (BCI) systems. Although convolutional neural networks (CNNs) have achieved promising results, existing methods still face limitations in multi-scale feature extraction. Specifically, some CNN architectures focus primarily on local and global temporal features, while medium temporal information remains underexplored. Moreover, conventional feature integration strategies often overlook the distinct characteristics of features at different scales, limiting their effective utilization. To address these limitations, we propose a multi-perspective attention-based multi-scale CNN, called MPAMCNN. The model extracts local, medium, and global EEG features simultaneously by applying convolution kernels with different receptive fields. In addition, we incorporate different attention mechanisms into the corresponding branches to emphasize discriminative information and reduce feature redundancy before fusion. Extensive experiments on two benchmark datasets, BCI Competition IV-2a and IV-2b, demonstrate that MPAMCNN achieves classification accuracies of 79.71% and 88.02%, respectively, validating its efficacy for EEG motor imagery classification. Visualization analysis also reveals improved feature separability and clearer decision boundaries.

Algorithms doi: 10.3390/a19110923

Authors: Gaofeng Che Jinchun Deng

Welding robot path planning has drawn increasing attention due to its broad applications in intelligent manufacturing. This paper proposes an integrated optimization framework that combines an improved ant colony optimization (IACO) algorithm for shortest-distance planning and an improved genetic algorithm (IGA) for shortest-time fixed-point trajectory generation. In the distance-planning layer, IACO is enhanced with parameter optimization and a dynamic random-disturbance factor to overcome the premature convergence and limited global search capability of traditional ACO. This achieves a high-quality welding point sequence with reduced computational overhead. In the time-planning layer, cubic B-spline interpolation is adopted to ensure path smoothness, and IGA is designed with a penalty-augmented fitness function and sinusoidal adaptive crossover and mutation operators to minimize execution time while satisfying kinematic constraints. Simulation experiments on a 24-point welding task demonstrate that the proposed IACO converges within approximately 100 iterations and reduces the best-found path length to 822 cm, which is 12.9% shorter than that of traditional ACO. Furthermore, the IGA reduces the total welding task time from 129 s to 96.30 s, representing a 24.7% improvement in time efficiency. Compared with conventional PSO and A* algorithms, the proposed method saves 17.60 s and 11 s, respectively. Additional comparisons with state-of-the-art hybrid algorithms confirm that the IACO-IGA framework achieves superior performance in terms of path length, convergence speed, task duration, and joint motion smoothness, with maximum joint-angle deviation and angular-velocity fluctuation reduced to 3.73 rad and 2.22 rad/s. Simulation-based Gaussian-noise robustness simulations confirm that the hybrid IACO-IGA framework exhibits better anti-interference capability under sensor measurement noise. The results verify the effectiveness and practicability of the proposed method for complex multi-welding-point industrial scenarios (simulation-based theoretical results; physical robot validation remains future work).

Algorithms doi: 10.3390/a19110922

Authors: Jinliang Wang Wenjuan Wang Guobao Chen Siwei Zhi

Parallax images often suffer from insufficient and unstable point features, making accurate alignment in overlapping regions extremely challenging. To address this issue, this paper proposes a robust image stitching framework based on content-preserving warping and quality-aware seam-cutting. The framework first estimates a global homography by combining point and line features to achieve reliable initial alignment. Subsequently, an energy function composed of alignment, distortion, and saliency terms is optimized to reduce geometric deformation while preserving image structures. After obtaining an initial seam using conventional seam-cutting, a structural similarity (SSIM)-based quality evaluation mechanism is introduced to identify misaligned regions and extract corresponding patches. To refine local alignment, the scale-invariant feature transform flow (SIFT flow) is employed to estimate dense correspondences, and a sigmoid-based smoothing constraint is adopted to avoid boundary artifacts. Finally, the optimized patches are integrated into the final result using a revised seam-cutting strategy with enhanced data penalty terms. Experimental results demonstrate that the proposed method achieves competitive performance, with favorable artifact suppression and structural preservation compared with mainstream stitching methods.

Algorithms doi: 10.3390/a19110920

Authors: Henrik Schwarz Zheng Grace Ma Bo Nørregaard Jørgensen

Agentic coding systems increasingly automate repository inspection, tool use, code generation, testing, and repair, but the human commonly remains responsible for decomposing the task, directing workflow progression, and judging completion. This limits their use in specialised domains where users understand the problem but not the development procedure required to solve it. This study introduces inversion of initiative (IoI) as a human–Agentic AI interaction paradigm in which procedural and modelling initiative shifts from the user to a procedure-aware Agentic AI, while human authority over domain meaning, consequential assumptions, validity boundaries, and progression is retained. The paradigm is instantiated for agent-based simulation development. Starting from an ordinary-language account of the problem, intended use, system, and investigative questions, the Agentic AI constructs and evaluates a structured natural-language simulation specification, resolves modelling and technical gaps through encoded skills, documentation, data, and authorised retrieval, and invokes bounded coding and assurance tools without requiring the user to orchestrate development. A browser platform operationalises IoI through an eight-stage deterministic workflow, versioned artefacts, readiness gates, traceability, semantic checks, controlled returns, and remote execution. In one end-to-end session, the platform transformed a neighbourhood energy-balancing brief containing neither an agent-based model nor a development plan into an AI-authored neighbourhood energy-balancing simulation specification, a verified configurable simulation package, and executed scenario comparison. The results support the feasibility of IoI in this specific case and illustrate an interaction pattern for agentic systems in which procedural and modelling initiative can reside with the AI while accountable domain authority remains human.

Algorithms doi: 10.3390/a19110921

Authors: Iván Neftalí Chávez-Flores Fabián García-Vázquez Jesús Antonio Nava-Pintor Miriam A. Carlos-Mancilla Luis F. Luque-Vega Juvenal Rodríguez-Resendiz Héctor A. Guerrero-Osuna

Residential water end-use monitoring is rapidly advancing from passive metering to AI-driven digital telemetry; however, systemic operational trade-offs remain poorly quantified. This systematic review synthesizes N = 100 peer-reviewed studies to critically evaluate algorithm architectures, sensing topologies, and deployment constraints in smart water disaggregation. Quantitative analysis reveals an overwhelming prevalence of non-intrusive paradigms (89.0%, N=89), with 45.0% (N=45) specifically targeting multi-fixture concurrent usage events. While high-frequency setups (≥1Hz) utilizing deep learning architectures (e.g., Seq2Point, 1D-CNNs) achieve high disaggregation accuracy (F1>0.88) in laboratory environments, performance degrades severely under commercial Automated Metering Infrastructure (AMI) supplying coarse 15–60min interval data. Four critical technical bottlenecks hindering utility-scale scalability are systematized: (1) severe temporal resolution mismatches between high-frequency training models and utility-scale AMI, (2) non-linear hydraulic superposition during simultaneous fixture activations, (3) domain transferability loss across heterogeneous plumbing topologies, and (4) memory and power constraints hindering edge TinyML execution (≤512KB SRAM). To address these challenges, an empirically validated three-layer conceptual analytics framework is integrated alongside five evidence-based future research directions, providing a structured roadmap to transition smart water analytics from experimental setups to robust, utility-scale deployments.

Algorithms doi: 10.3390/a19110919

Authors: Ahmed Kammoun Moulay A. Akhloufi

Glioma is the most common primary malignant brain tumor in adults, and accurate determination of isocitrate dehydrogenase (IDH) mutation status is important for diagnosis, treatment planning, and prognostic assessment. Currently, definitive IDH status determination relies on molecular analysis of tumor tissue obtained through biopsy or surgical resection, motivating the development of accurate, interpretable, and non-invasive imaging biomarkers. This study proposes an explainable multimodal generative framework for IDH mutation prediction using multiparametric magnetic resonance imaging (MRI), radiomics, and clinical information. The framework employs a self-supervised 3D convolutional neural network (CNN) to learn imaging representations, a dedicated encoder for radiomic features, and late fusion to construct a unified multimodal representation. A Conditional Variational Autoencoder (CVAE) is used for minority-class feature-space augmentation, while a Generalized Additive Model with Interactions (GAMI-Net) provides intrinsically interpretable classification without relying on post hoc explanation methods. Experiments were conducted on 622 patients with diffuse glioma from the University of Texas Southwestern Medical Center, comprising a training cohort of 380 patients, a validation cohort of 96 patients, and a held-out temporal test cohort of 146 patients. On the temporal test cohort, the proposed framework achieved an area under the receiver operating characteristic curve (AUC) of 0.9519 (95% CI: 0.9158–0.9802), an accuracy of 87.7%, and a macro F1-score of 0.847, with higher discrimination than the evaluated DenseNet169-3D and ConvNeXt imaging-only baselines. Intrinsic explainability analysis identified contributions from imaging, radiomic, and clinical information, while global effect functions revealed nonlinear relationships between model features and predicted IDH mutation probability. External evaluation on an independent UCSD cohort showed lower discrimination (AUC = 0.675, 95% CI: 0.513–0.822), highlighting a substantial cross-institutional generalization gap. Overall, the findings demonstrate that the proposed framework combines strong internal temporal discrimination with intrinsic and generative interpretability, while the external results emphasize the need for further multi-center validation and improved robustness to domain shift before clinical deployment.

Algorithms doi: 10.3390/a19110918

Authors: Xiaodan Dong Xiaolong Deng

In this paper, an adaptive disturbance rejection controller (A-DRC) is proposed for tunnel inspection robots equipped with a mobile platform, which system is modeled as a flexible system by partial differential equations (PDEs). First, the system dynamics are formulated as PDEs subject to boundary conditions. Furthermore, an A-DRC is proposed, which relies solely on boundary state feedback to suppress vibration while handling unknown system parameters and external disturbances. Then, the asymptotic stability of the closed-loop system is rigorously established via the Lyapunov method. Finally, compared with the existing control method, the proposed controller guarantees that the unknown parameters tension can be well estimated and vibrations can be well suppressed while eliminating the influence that external disturbance caused observed by the numerical simulation, verifying the effectiveness and advantages of the proposed scheme. Finally, physical experiments are carried out to confirm the method’s practical viability.

Algorithms doi: 10.3390/a19110917

Authors: Gulzhan Uskenbayeva Salimzhan Tassanbayev Gulnara Abitova Guldana Taganova Aliya Shukirova Ardak Mukhamedrakhimova Korlan Kulniyazova Lili Nurliyana Abdullah

This study evaluates a causal, six-target, 30 min forecasting component for diesel hydrotreating using campaign-separated synthetic data, not a deployed digital twin. An empirical, engineering-constrained simulator generated 30 original campaigns; the recovered code reproduced their datasets byte-for-byte but the simulator did not implement mechanistic balances. The archived AURA-SSM hybrid combines a frozen Ridge anchor, a released measurement observer, multiscale context, bounded corrections, and conditional empirical risk weights; despite its identifier, it has no learned state-space layer. On two previously inspected evaluation domains, frozen NRMSE was 0.16996 and 0.18085, versus ten-degradation means of 0.19170 and 0.20490. These retrospective improvements over leading frozen controls were below the prespecified 3% practical threshold; adaptive Ridge was more accurate on average. Additional matched controls, CRISP/TimeFormer adaptations, and a prespecified 16-campaign synthetic evaluation clarified limitations. On new common/modified simulator structures, AURA-SSM achieved ten-degradation NRMSE of 0.18804/0.20332, outperforming NLinear but not a fixed convex blend; context LightGBM was more accurate with clean measurements. A separately corrected observer passed causal tests and improved prolonged-outage prediction without replacing archived results. Interval undercoverage, lack of mechanistic validation, and absent independent industrial data constrain applicability.

Algorithms doi: 10.3390/a19110916

Authors: Takafumi Matsuura

The covering salesman problem with nodes and segments (CSPNS) extends the traveling salesman problem by allowing targets to be covered from visited nodes or route segments. Motivated by drone-based monitoring and information collection after disasters, this study investigates a destroy-and-repair (DR) heuristic for CSPNS. The heuristic reconstructs both the visited-node set and the tour through partial destruction and repair, restores feasibility using node and segment coverage, and applies CSPNS local search. The performance of the proposed method is compared with local search, multi-start local search (MSLS), and tabu search (TS). The destroy–repair combination and destruction rate are selected using one training instance per size–radius setting, and the selected configuration is evaluated on four separate instances using 100 trials per instance. MSLS, TS, and DR use equal per-trial CPU-time limits. For 100-node and 150-node problems, DR obtains smaller setting-level average reference gaps than TS. Two-sided exact Wilcoxon signed-rank tests with the Holm correction show significant differences in all 16 DR–TS comparisons, favoring DR in 14 cases and TS in two. DR also significantly outperforms MSLS in all 16 evaluation instances. For 200-node and 300-node problems, DR obtains lower mean route costs than both MSLS and TS when averaged over the evaluation instances in each size–radius setting. Additional experiments compare uniform and adaptive operator selection under the same training and evaluation protocol. These comparisons find no consistent advantage of adaptive operator weighting over uniform selection under the tested conditions. These results support the effectiveness of the proposed DR heuristic for CSPNS.

Algorithms doi: 10.3390/a19110915

Authors: Assem Shayakhmetova Nurbolat Tasbolatuly Guldana Taganova Nurbay Dossanov Marat Shurenov Aigul Bissarinova Nurdaulet Issimov Gulnur Karakhanova

Reliable milling monitoring requires calibrated intervention probabilities and evaluation that respects experimental clustering. This methods-and-evaluation study examines Hierarchical Bayesian Adaptive Fusion (HBAF), combining calibrated TabM and Gaussian Naive Bayes experts. MSM provides 927 source-machine cutting paths and 1585 from a historically inspected second machine. Nested development assessment uses 563 paths from 11 experiments; frozen evaluation additionally uses 172 internal-test paths and the second-machine cohort. The development of HBAF achieves MCC 0.6383, PR-AUC 0.8552, and Brier score 0.0773, but equal fusion and a nonhierarchical control remain stronger on several measures. Added component, prefix, scale-sensitivity, and NUTS analyses do not establish general hierarchical-fusion superiority or uniformly valid Laplace uncertainty. Frozen HBAF performance decreases to MCC 0.4587 and PR-AUC 0.5820 internally, and MCC 0.1040 and PR-AUC 0.2809 across machines. The experiment-aware inference also leaves asymmetric-cost superiority unestablished. The contribution is a leakage-safe evaluation framework and a reproducible case study identifying limits of Bayesian adaptive fusion. Historically inspected transfer data, conditional posterior uncertainty, and hypothetical error costs do not support claims of independent cross-machine confirmation or industrial deployment validation.

Algorithms doi: 10.3390/a19110914

Authors: Yun Gao Xinyue Ji Lipu Zhang

Resampling agreement can select a coarse partition that is reproducible under perturbation but changes substantially when the candidate cluster count is increased. This study introduces an adjacent-partition-agreement refinement for cluster-number selection. The refinement supplements resampling stability with the normalized mutual information between adjacent full-data partitions and overrides the stability argmax when that agreement falls sufficiently below the median of the agreement curve. The two margins are fixed for the primary experiments and varied only in the reported sensitivity grid. The rule is evaluated in Consensus-Neighborhood Clustering, which provides a sparse neighborhood-consensus affinity, and with an independent radial-basis-function spectral carrier. On Blobs, the reported protocol changes the selected count from 2 to 4 and the adjusted Rand index (ARI) increases from 0.32 to 0.83; an embedding-protocol ablation shows that this exact count is protocol-sensitive. On Digits, the count changes from 2 to 9 and ARI increases from 0.05 to 0.74. The selection is unchanged in the other tested scenarios, including two hyperspectral cases. A candidate upper bound Kmax remains user-specified, and the Indian Pines range is label-informed. The results support correction of a specific coarse-stability failure mode, not generally superior or range-free cluster-number estimation.

Algorithms doi: 10.3390/a19110913

Authors: Essa Q. Shahra Rasheed Mohammad Waheb A. Jabbar

The growing prevalence of mental health disorders, combined with limited access to traditional therapeutic services, has accelerated the demand for scalable and empathetic AI-based interventions. This study explores the development and evaluation of an AI-powered chatbot designed to provide mental health support using advanced Large Language Models (LLMs), specifically LLaMA 2 and LLaMA 3. Using a mental health counseling dataset, the models were fine-tuned on counseling context–response pairs and evaluated for semantic and lexical similarity to reference responses. The training pipeline incorporated quantization, LoRA adaptation and a structured prompt template. The models were evaluated using BERTScore, BLEU, ROUGE, and METEOR, offering a multi-dimensional view of linguistic and semantic quality. Results show that LLaMA 3 achieved a higher mean BERTScore F1 than LLaMA 2 (0.860 versus 0.835). The results characterize the fine-tuned models’ performance under automatic text-similarity evaluation. Human evaluation is required before conclusions can be drawn about empathy, safety or suitability for mental health support. Ethical considerations and future directions, including user validation and multimodal expansion, are also discussed.

Algorithms doi: 10.3390/a19110912

Authors: Wei Chen Linna Wei

Three-dimensional wireless sensor networks (3D WSNs) often face low coverage, uneven node distribution, and obstacle interference in real monitoring scenarios. To solve these problems, the multi-objective coverage optimization problem is first converted into a single-objective one via the weighted-sum method, and then solved by a hybrid-strategy artificial lemming algorithm (HS-ALA) that aims to improve coverage, node distribution uniformity, and network connectivity in 3D environments. The HS-ALA adopts an elite-guided adaptive differential mutation strategy that adjusts the mutation factor and crossover probability to enhance global exploration and maintain population diversity. The HS-ALA also embeds a neighborhood search operator to refine local search accuracy. A hybrid method combining penalty functions and hard constraint repair is applied to handle obstacle constraints in 3D deployment scenarios. In the CEC2017 benchmark tests, the HS-ALA ranks first in Friedman test at both 50 and 100 dimensions (1.34 and 1.55, respectively). It performs better than ALA, PSO, GWO, and other algorithms. In a 60 m × 60 m × 60 m simulation environment with 70 sensor nodes, the HS-ALA improves coverage by 12.83 percentage points and increases uniformity from 0.8656 to 0.9116 compared with the original ALA in the ideal scenario. In both obstacle scenarios, it maintains high coverage and uniformity, outperforming all compared algorithms. In all tested scenarios, the network maintains 100% global connectivity. It also performs well across different node scales, demonstrating good robustness and practical value.

Algorithms doi: 10.3390/a19110911

Authors: Hugo Mainguy Bruce Golden Luca Bertazzi Debdatta Sinha Roy

In the Close Enough Traveling Salesman Problem (CETSP), one must find the shortest tour that travels through every given customer-centered neighborhood, which is usually a disk centered on each customer or target. This problem has applications in various fields, such as radio frequency identification (RFID), mission planning for Unmanned Aerial Vehicles (UAVs), multi-visit drone routing, and aerial route reconnaissance planning. In this study, we combine prior work on CETSP estimation models and on using results from a randomized heuristic on combinatorial optimization problems such as the TSP, VRP, Split Delivery VRP (SDVRP), and the Knapsack Problem in order to obtain high-quality optimal solution value estimates efficiently. Here, we use discretization heuristics based on the Generalized TSP (GTSP) and merging nodes in Steiner zones. With our regression-based approach, we are able to obtain estimates on instances that would otherwise take hours to solve to optimality. We obtain an MAPE of 0.56% on a set of 780 CETSP instances, handily improving on similar predictor-based approaches that did not use heuristics, and we compute accurate estimates for instances that take much longer to solve or even approximate. We are motivated to contribute to this research stream in order to show that we can obtain excellent optimal solution value estimates to a wide variety of hard combinatorial optimization problems.

Algorithms doi: 10.3390/a19110910

Authors: Rafał Michalski Jerzy Grobelny

Background: Logistics layouts are commonly evaluated using flow-distance criteria, but layouts with similar economic performance may differ in spatial coverage and pairwise-distance variability. This study examines whether measures of these concepts exhibit distinguishable sensitivity to layout-generation conditions and whether their empirical association supports treating them as complementary rather than redundant descriptors. Methods: Fuzzy LP-Alinks generated layout configurations in a full-factorial simulation comprising 108 conditions and 30 replications per condition; the minimum-flow-distance solution was retained from each run. Coverage was measured with the Normalized Coverage Score (NCS), and distance regularity with the Distance Variability Index (DVI), defined as the coefficient of variation of pairwise distances. Five-factor ANOVAs examined main and interaction effects of object count, link density, virtual-force magnitude, and fuzzy membership function settings, while joint-distribution analyses assessed the association between NCS and DVI. Results: The indicators exhibited different main-effect and interaction patterns. In particular, object count and link density had strong main effects on DVI but nonsignificant main effects on NCS, although their interaction substantially affected NCS. The joint distributions further showed that similar NCS values could coexist with different DVI values. Conclusions: Within the tested experimental setting, the results provide evidence that NCS and DVI behave as complementary rather than directly redundant descriptors of spatial configuration. This conclusion does not imply complete separability or formal discriminant validity.

Algorithms doi: 10.3390/a19110909

Authors: Yingkang Yao Mengfei Yu Chunping Wu Yongsheng Jia Xinnan Cui Nan Jiang Jiarun Huang Chuang Liu

Accurate prediction for Peak Particle Velocity (PPV) induced by blasting vibrations is crucial for controlling engineering risks and optimizing blasting parameter design. In order to address the practical challenges of numerous influencing factors and limited available research samples for PPV, a PPV prediction model based on a Backpropagation neural network (BP) optimized by the Improved Grey Wolf Optimizer (I-GWO), termed the I-GWO-BP network, is developed. The I-GWO algorithm was employed to efficiently search for the optimal combination of weights and bias parameters that need to be optimized in the BP neural network. The results indicate that the proposed integrated prediction model forecasts PPV with greater accuracy and reliability than comparative models including the Grey Wolf Optimizer-enhanced BP model (GWO-BP) and Extreme Learning models (ELM). Beyond demonstrating optimal prediction performance, explainability analysis used the SHapley Additive exPlanations (SHAP)method further enhances the I-GWO-BP model’s interpretability, revealing that the distance from the blasting center is the most significant influencing factor. The application potential of the I-GWO-enhanced neural network for predicting PPV under similar geological and operational conditions is highlighted, providing a methodological reference for related prediction work.

Algorithms doi: 10.3390/a19110908

Authors: Xia Wu Liang Yu Xinyu Han Xun Gong Xinrong Li Chaofan Yang

Rain and fog significantly degrade visibility in rail transit monitoring systems due to their non-uniform spatial distribution and dynamic temporal variations. Traditional image restoration methods struggle to balance computational efficiency and edge preservation under extreme weather conditions. We propose a Frequency-Adaptive Multi-Scale Network (FAMNet) that jointly optimizes wavelet-based spectral decomposition and similarity-driven feature aggregation. Our framework dynamically models degradation patterns through adaptive spectral kernels while capturing cross-region dependencies via similarity-driven aggregation. Evaluated on 8500 synthetic rail transit images derived from the RailSem19 benchmark, the method achieves 32.6 dB PSNR/0.95 SSIM for deraining (+1.0 dB/+4% vs. Restormer). Our method, evaluated on an NVIDIA RTX 3080 GPU with 256 × 256 inputs (batch size 1, FP32 precision), achieves an end-to-end inference speed approximately 40× faster than Restormer—a leading baseline in restoration quality—while maintaining competitive performance. This efficiency shows strong potential for cost-effective deployment in all-weather rail transit inspection systems, with possible applications in safety-critical scenarios such as track obstacle detection.

Algorithms doi: 10.3390/a19110907

Authors: Jesús Gil Ruiz Hernán Díaz Rodríguez David Ariza Ruiz

When symbolic regression (SR) is run on data pooled from several facilities, a compact expression can fit well and still describe where the measurements happen to lie rather than a physical relationship. We present an audit framework for this structure-level domain confounding. A design-identifiability screen (Stage 0) checks, from design metadata alone, which dimensionless groups the experiments can separate; honesty tests, discriminant tests, few-cluster inference and normalized conformal prediction then assign separate verdicts to the fitted law and to any structure selected by model selection. On 14,400 simulated designs the rank/condition-number screen is correct for 79% against 66% for a pairwise-correlation screen applied to the same design. On a factorial benchmark of 2880 simulated corpora in nine regimes, the discriminant tests recognized a genuine transition at the reference threshold in 100% of the main regime (320/320) and flagged a structure in 1.4% of the 1600 corpora without one; after the conformal check, 74% of the genuine transitions of the main regime (37.5% across the four genuine-transition regimes) and 0.06% of the others were reported as physics. A genuine transition on sparse support was never recovered (0/320), because the minority-side rule that withholds attribution from the artefact rejects it too, and model selection alone presented a structure as physical for up to 100% of the support-induced pseudo-transitions. As an illustration we audit a public corpus of 114 windage measurements from four facilities spanning 70× in scale. Bayesian model selection retains a two-regime law with a nominal Bayes factor of 5.8×105; the discriminant tests classify it as statistically supported but physically unattributed, and the fitted law does not transport to a withheld facility.

Algorithms doi: 10.3390/a19110906

Authors: Bao Zhou Sha Gao Ji Zhang Shuang Liu

In large-scale UAV photogrammetry, redundant and weakly correlated image relationships can substantially increase computational costs and reduce matching reliability during SfM reconstruction. To address this issue, this study proposes SC-DFS, a spatial constraint-based image relationship optimization method for efficient UAV image organization and reconstruction. The contribution lies in integrating POS-derived spatial constraints with image relationship graph optimization, where spatial overlap information, image similarity, and F-RANSAC-guided geometric verification are combined to construct weighted image relationships. Subsequently, a depth-first search strategy is employed to extract the connected component containing the largest number of images and optimize the image set for subsequent SfM reconstruction. Experimental results on a large-scale UAV dataset demonstrate that SC-DFS reduces the number of retained images from 2556 to 2274 with visually consistent overall 3D structure. The optimized dataset reduces the sum of the three reported processing-stage times by 2 h and 17 min. The proposed method effectively reduces redundant image relationships, improves image matching efficiency, and provides an efficient solution for large-scale UAV SfM reconstruction in the evaluated acquisition.

Algorithms doi: 10.3390/a19110905

Authors: Sergio Adrian Prieto Flores Jesús Aguayo Alquicira Susana Estefany De León-Aldaco Mario Ponce-Silva Víctor Hugo Olivares Peregrino Heriberto Adamas-Pérez Pedro Javier García-Ramírez Ricardo Eliu Lozoya-Ponce

Differential Evolution (DE), Teaching–Learning-Based Optimization (TLBO), Grey Wolf Optimizer (GWO), and the Ali Baba and the Forty Thieves Algorithm (AFTA) are compared for switching angle calculation in an ideal seven-level single-phase cascaded H-bridge inverter. Each method uses 30 paired seeds, 1000 candidates, and 201,000 objective-function evaluations. The common objective combines voltage THD through the 49th harmonic with absolute fundamental voltage error relative to 250 V peak. DE and AFTA attain THD49 = 11.6691079% and meet the empirical target in all 30 runs; TLBO succeeds in 2 runs and GWO in none. DE requires a median of 29,361 evaluations to reach the target, compared with 53,822.5 for AFTA. AFTA has the lowest mean full-budget runtime, 1.2354 s, versus 2.0742 s for DE. Paired tests with multiplicity correction distinguish DE/AFTA from TLBO/GWO in final objective value, but do not distinguish DE from AFTA at the declared numerical resolution. The maximum separation between the best THD values is only 0.0004657 percentage points. The results separate numerical precision, search effort, and runtime within one analytical operating condition; they do not establish global optimality or a practically significant electrical advantage.

Algorithms doi: 10.3390/a19100904

Authors: Mohammed Hassan Nawaf Alshdaifat Hamza Abu Owida Tareq Hamadneh Hamza A. Mashagba Azlan B. Abd Aziz Ibrahim Abd El-Fattah Suhaila Abuowaida

Super-resolution (SR) facilitates high-fidelity mapping beyond the spatial resolution capacity of clinical neuro-oncology MRI. While state-of-the-art deep learning algorithms have augmented this domain, such methods exclusively learn from spatial or spectral domains and are often compounded by tumor boundary over-smoothing and geometrically inconsistent pathologies. In this paper, we introduce Hy-DTGL, a hybrid-domain transformer with a global–local attention modal super-resolution (SR) framework that learns a joint reconstruction across four inter-dependent knowledge domains: local spatial, global spatial, spectral and wavelet. Using windowed self-attention, Hy-DTGL learns highly localized pathologic textures; pooled global attention constrains anatomical plausibility; Fourier-learnable modulation provides spectral energy per semblance; and multi-scale discrete wavelet tiling preserves multi-scale tumor morphology. Optimization occurs via a hybrid objective, combining Charbonnier loss, gradient edge-loss, and spectral domain loss. The proposed framework was trained on BraTS 2019 (BT_19) and evaluated on BT_19, BraTS 2020 (BT_20), BraTS 2023 (BT_23), and IXI under both in-domain and strict zero-shot settings, and achieves competitive performance relative to the evaluated baselines under the adopted degradation and evaluation protocol at ×2, ×4, and ×8 SR factors. In the BraTS 2020 ×4 test resolution, 37.46 dB PSNR and 0.986 SSIM are attained, outperforming deep learning CNN-based SR (ECFNet) by 2.47 dB and transformer-based (WavTrans) by 4.05 dB in PSNR. Ablative studies reveal the spatial, spectral and multi-scale constraints notably converge to a faithfully mapped image space with restored tumor boundary delineation and global-structure consistency. We also show the Hy-DTGL generalizability by how it improves downstream tumor segmentation metrics: the super-resolved scans improve the segmentation Dice coefficient by 0.063 and significant reductions are observed in the Hausdorff Distance from 19.05 mm to 13.01 mm relative to the original low-resolution images, suggesting potential utility for downstream neuro-oncological image analysis.

Algorithms doi: 10.3390/a19100903

Authors: María Fernanda Molina-Miranda Angel Cuenca-Ortega Juan Fernandez-Olivares Raúl Pérez

We evaluate supervised classification models on European Union public tenders, pairing random forest and gradient boosting with explainable artificial intelligence (XAI) and a rule-based system. Using TED data processed with Apache Spark, we modeled the main activity of the contracting authority among 21 classes (Experiment A) and whether an SME wins the contract (Experiment B). Random forest outperforms gradient boosting and a decision tree baseline in every fold of both experiments (one-sided Wilcoxon test, p = 0.001). Its F1-macro in Experiment A is 0.3550, far above the majority-class baseline (0.0228), yet only three of the 21 classes exceed F1 = 0.6, and the score falls to 0.2203 on later years and 0.1443 on countries absent from training. Experiment B is easier: F1-macro reaches 0.5863 (AUC = 0.7100) against a baseline of 0.4500 and 0.5515 when the model sees only the features that are known before bidding. That restricted model, however, has no discriminative ability on unseen countries (AUC = 0.51). National markets explain much of this. In the seven largest, F1-macro averages 0.59 when the model is trained within the country and 0.50 when the country is left out. The if–then rules derived from LIME have low fidelity (R2 = 0.43 and 0.16), and, on held-out data, a third of the Experiment B rules are no more precise than an empty rule; TreeSHAP supports the global feature ranking but not these local explanations. Any use for SME decision support would be limited to markets that are represented in training and would first need validation with end users.

Algorithms doi: 10.3390/a19100902

Authors: Jie Chen Mengli Wang Yajuan Zhang Guisheng Li

Automatic modulation classification (AMC) is a core task in cognitive radio and spectrum monitoring. Mainstream deep-learning-based AMC models typically contain hundreds of thousands to millions of parameters, making them difficult to deploy on resource-constrained edge devices. This work proposes MicroSSM, a lightweight state-space model for automatic modulation classification. MicroSSM decomposes signal processing into three functionally independent stages: a learnable soft-thresholding module performs adaptive denoising at the input stage; a four-layer small-kernel one-dimensional convolutional neural network (1D CNN) extracts local time–frequency features with a minimal parameter count; and a hierarchical state-space dual-channel module (HSMSSD) performs global temporal modeling with linear complexity via a softmax-normalized discrete state selection mechanism. Unlike Mamba-family models, which rely on continuous-time discretization, HSMSSD selects states through this softmax-normalized discrete choice, thereby avoiding the mapping overhead of continuous-time parameterization on short sequences. The capacities of the three modules are jointly regulated by a unified scaling factor, enabling model variants spanning two orders of magnitude in parameter count. On RML2016.10B, the 2.3K-parameter variant achieves 61.63% accuracy, and scaling to 133.5K parameters raises it to 64.13%; consistent results are observed on RML2016.10A, while a 2D grid extension of MicroSSM reaches 87.86% on the long-sequence, multi-class HisarMod2019.1 dataset. Ablation experiments confirm the effectiveness of each module and show that the marginal benefit of model scaling decays sharply: MicroSSM realizes its accuracy gain within the low-parameter window, and the larger variants are not cost-effective for edge deployment. Finally, MicroSSM is deployed on a Raspberry Pi 4B and achieves sub-millisecond per-sample inference latency over a real 433.92 MHz RF link, demonstrating the feasibility of the model on resource-constrained edge devices.

Algorithms doi: 10.3390/a19100901

Authors: Andrei Novikov Peihua Zhu Konstantin Koltsov Ming Dong Ruxing Wang Dmitry Lipis Elena Gebel

This study develops a digital-thread framework for predictive assessment of digital transformation in heavy equipment manufacturing and evaluates it through resource-constrained discrete-event simulation (DES) of the VFP-1600 route. The model represents nine stages, shared engineering capacity, Poisson order releases, data-readiness lags, and first-come-first-served queues. Seven expert rankings yield Kendall’s W = 0.893 and a normalized Digital Maturity Index of 0.264. Maturity priorities guide the activation of stage-level intervention assumptions. Five specified configurations are compared in 10,000 independent annual-cohort replications per load level, using paired random inputs. At six expected releases per year, simulated mean cycle time is 163.35 days for the baseline, 126.70 days for integrated PLM + MES + QMS/IoT, and 119.46 days with APS-associated duration assumptions. At twelve releases, the corresponding means are 213.02, 159.23, and 149.58 days. Directly measured time-average work-in-process is checked against an exact finite-horizon residence-time identity. A joint sensitivity experiment varies nine intervention-realization factors, resource occupation by delays, and arrival intensity. Its 5th–95th percentile ranges of estimated mean cycle-time benefit are 9.0–25.3% and 11.0–28.7%. The computational experiment shows how resource contention changes both the magnitude and uncertainty of projected benefits. The framework supports ex ante comparison of explicitly specified production and transformation scenarios.

Algorithms doi: 10.3390/a19100900

Authors: Attila Farkas Mátyás Nagy Gábor Kertész

Distributed graph neural network training usually starts with an offline partitioner such as METIS, which fixes the layout before any workload is seen. SoftPart GNN is a distributed GraphSAGE runtime on PyTorch DistributedDataParallel (DDP) without one. A short online warm-up turns the vertices that the training sampler touches into an exclusive soft ownership, and a budgeted cache replicates the context that it omits. Training samples only from this local view, so gradient synchronisation is the only per-step cross-worker traffic. A layout-independent full-neighbour evaluator exchanges hidden representations in place of features and, on the four graphs where we verified it, agrees with a single-machine reference to rounding level. We measured a two-layer GraphSAGE of hidden width 16 on OGBN-Arxiv, Reddit, OGBN-Products, IGB-hom-small and IGB-hom-medium at four to 16 nodes with one GPU each, and on four graphs at 32. At four to 16 nodes, SoftPart GNN completes training, placement included, in 22 to 62 percent of the time of a DistDGL baseline on the Deep Graph Library’s (DGL) original distributed sampler, and 21 to 52 percent with the baseline’s partitioning counted. Its test accuracy there stays within 0.006 of DistDGL’s on four of the five graphs. At 32 nodes, the ratio is 0.40 to 0.91. On Reddit, the accuracy is 0.0042 lower, and OGBN-Arxiv is left open. A paired control with modulo ownership and the same cache shows no detected accuracy difference. At four to 16 nodes, the warm-up shortens the training epoch on four graphs and gives no end-to-end gain.

Algorithms doi: 10.3390/a19100899

Authors: Sebastian-Alexandru Drǎguşin Nicu Bizon Gabriel-Vasile Iana

Recent advances in neural speech synthesis and voice conversion have increased the risk of logical-access spoofing attacks against voice-authentication systems. This study presents an interpretable, enrollment-conditioned framework that separates speaker-identity verification from subsequent speech-authenticity assessment. The first stage uses Emphasized Channel Attention, Propagation and Aggregation Time Delay Neural Network (ECAPA-TDNN) speaker embeddings and cosine similarity, while the second stage evaluates speaker-conditioned Mel-Frequency Cepstral Coefficient (MFCC), pitch, and local-jitter anomaly scores. These scores are normalized using development target bona fide distributions and combined through a fixed, unweighted score-level fusion; Principal Component Analysis (PCA) waveform reconstruction is retained as an ablation component. The framework was evaluated on the ASVspoof 2019 Logical Access (LA) benchmark using strict separation between development-based calibration and independent evaluation. All feature formulations, normalization references, fusion components, and decision thresholds were frozen before evaluation. On 102,579 evaluation trials, the ECAPA-TDNN stage alone achieved a target False Rejection Rate (FRR) of 0.5773%, a bona fide non-target False Acceptance Rate (FAR) of 1.0532%, and a spoof acceptance rate of 66.1892%. The complete hierarchical system achieved an observed target FRR of 4.8045%, a non-target FAR of 0.7681%, and a spoof acceptance rate of 58.7239%, corresponding to an absolute spoof-acceptance reduction of 7.4653 percentage points. MFCC provided the strongest standalone second-stage discrimination, with a Receiver Operating Characteristic–Area Under the Curve (ROC-AUC) of 0.653055, while the frozen MFCC-pitch-jitter fusion achieved an ROC-AUC of 0.604293 and was selected prospectively under a 5% development target-FRR constraint. Attack-specific evaluation revealed substantial variability across the A07–A19 evaluation conditions, including systematic failure modes for A12 and A16. Within the ASVspoof 2019 LA evaluation setting, the results demonstrate a reproducible and interpretable hierarchical authentication framework while also showing that fixed acoustic descriptors do not provide attack-invariant spoof detection. Because development and evaluation originate from the same benchmark, these results should not be interpreted as evidence of cross-corpus, cross-channel, physical-access, or real-world deployment generalization.

Algorithms doi: 10.3390/a19100898

Authors: Shizheng Zhang Bing Zhou Yanbu Guo Min Huang Wanwei Huang Dongyuan Lei Tianbao Dan

Single-cell multi-omics clustering requires integrated representations that preserve shared biological states and modality-specific signals. Graph-based generative methods often couple clustering and reconstruction through a common latent pathway and a single neighborhood graph. We propose SP-LSNTA-GeDGC, which combines task-oriented shared–private (SP) routing and Latent Spectral Neighborhood–Transport Alignment (LSNTA) within a graph-embedded deep generative clustering (GeDGC) framework. Modality-specific residuals support reconstruction, while clustering uses the shared representation without introducing an independent private posterior. LSNTA integrates shared-latent spectral learning, feature-gated mutual nearest neighbors and unbalanced optimal transport to refine graph sample centers and cross-modal relations. The refined centers also serve as reconstruction targets during joint training. Consensus neighborhoods and averaged pairwise transport extend the framework to more than two inputs. Across five bimodal benchmarks, five-run comparisons on identical processed inputs yield a mean clustering accuracy (ACC) increase of 3.08 percentage points over GeDGC. Holm-adjusted paired t tests support ACC and adjusted Rand index (ARI) improvements on Pbmc10k, while gains on other datasets remain descriptive. Metric means decrease on the transcripts, epitopes and accessibility sequencing (TEA-seq) dataset, whereas the compatible four-input dyngen simulation shows small gains. Component ablations reveal dataset-dependent effects without establishing an additional SP–LSNTA interaction. These findings support conditional clustering benefits under known-K evaluation, while broader multimodal generalization and computational scalability remain open challenges.

Algorithms doi: 10.3390/a19100897

Authors: Sergio Bonini Sergio Da Silva

This study examines the computational scaling of mixed-integer linear programming (MILP) for 12-bar blues guitar-solo generation. A Python/PuLP implementation uses CBC to select and order licks under duration, turnaround, repetition, pause, and connectivity constraints. Transitions receive deterministic rule-based scores. The 7 baseline configurations contain 100 sampled instances each, yielding 693 proven-optimal completions, 6 timeouts, and 1 error. Among completed runs, larger candidate sets have lower mean and median transition scores and higher optimization times. Standard errors, bootstrap confidence intervals, distributional diagnostics, and nonparametric comparisons characterize these results. A paired sensitivity study uses 100 candidate instances per pool to vary repetition and pause limits. Relaxing the repetition limit substantially reduces transition scores across all pools, while runtime decreases in Moderate and increases in Fast. Relaxing the pause limit produces smaller score improvements. The released implementation exports MusicXML outputs and records the information needed to reproduce sampled instance.

Algorithms doi: 10.3390/a19100896

Authors: Jorge Arturo Pinedo Gaucin Bárbara Alexandra Anaya Sánchez Luis Asunción Pérez-Domínguez David Luviano-Cruz Roberto Romero López Nelly Rigaud Téllez Diana Yaziel Ortiz-Muñoz Luis Alberto Rodríguez-Picón José Luis López Galván

Digital twins are expected to improve supply chain decisions by linking physical operations with models that simulate and optimize them. This literature review, guided by PRISMA 2020, analyzes 104 studies (2016 to 2026) on the enabling technologies of supply chain digital twins, the barriers to their adoption and the algorithmic evidence behind them. The studies were classified into evidence tiers by application domain; twelve supply chain studies and six algorithm studies from other sectors were coded by decision problem, algorithm family, integration level and validation, and the methodological quality of fifteen of them was appraised. The Internet of Things, artificial intelligence, cloud and edge computing, extended reality and blockchain are the main enablers, but in the analyzed set the supply chain evidence is largely conceptual: only two supply chain studies apply an algorithm, none implements a bidirectional digital twin, and only one reports performance measures from a model run. State estimation, reduced-order models and reinforcement learning, which would allow real-time operation, have been evaluated only in other sectors. Technical, organizational and economic barriers appear to reinforce each other. We propose a three-stage implementation pathway and an agenda of seven research gaps, including closed-loop integration, shared benchmarks and multi-company data governance.

Algorithms doi: 10.3390/a19100895

Authors: Mohadese Khosravi Bahram Sadeghi Bigham Mazyar Zahedi-Seresht

Background: Choosing the number of clusters is still one of the least settled steps in a K-means analysis. The within-cluster sum of squares (WCSS) elbow plot is easy to understand, yet visual inspection is difficult to reproduce, and different validity indices often favor different values because they measure different aspects of the partition. Methods: We propose and evaluate Straw-K, a geometric cluster-number selection method that scores local bends on the normalized WCSS curve. Straw-K is a coined method name, not an acronym. Its construction adapts the local-window and chord-to-path geometry used in ShortStraw corner detection to discrete cluster-number selection; the contribution is the WCSS-specific criterion and its minimization over candidate values of K, rather than a claim that the chord-to-path relation itself is new. Twelve empirical datasets were analyzed without a general sample-size cap. Seven selectors were compared over eight K-means random-state seeds: Straw-K, the official Kneedle implementation, Silhouette, Calinski–Harabasz, Davies–Bouldin, the Gap statistic with B = 50, and the L-method. Agreement with the recorded class count, absolute error, seed stability, runtime, the Friedman test, Holm-adjusted paired Wilcoxon tests, an exploratory ambiguity margin, and a separate Digits/PCA analysis were examined. Results: Across the 96 dataset–seed evaluations available for each method, Straw-K matched the recorded class count in 44.8% of cases and produced a mean absolute error of 1.54, the highest exact-match rate and lowest mean error in the descriptive table. The Friedman test detected an overall difference among the seven selectors (χ2=13.51, p = 0.036), but none of the 21 paired Wilcoxon comparisons remained significant after Holm correction (minimum adjusted p = 0.123). Gap was the slowest method in the independent uncached timing analysis. The ambiguity index showed only a weak, non-significant association with seed instability (Spearman ρ = 0.33, p = 0.303) and is therefore treated as descriptive rather than calibrated uncertainty. Conclusions: Taken together, the experiments suggest a fairly specific use for Straw-K: it offers an inspectable ranking of bends once the short WCSS support curve has been computed. The method did not recover a unique cluster count in every dataset, and Digits remained particularly difficult. The class labels used in the benchmark are best understood as external comparison points; they do not by themselves define the geometry that K-means will uncover.

Algorithms doi: 10.3390/a19100894

Authors: Maksym Saliutin Yuriy Kondratenko Ievgen Sidenko Galyna Kondratenko Dmytro Chumachenko

The spread of infectious diseases under wartime and resource-constrained conditions creates a need for analytical tools that support both population-level epidemiological monitoring and preliminary patient-level image analysis. This study develops and evaluates a web-based analytical framework incorporating two complementary modules: infectious-disease forecasting and chest X-ray image analysis. For epidemiological forecasting, ARIMA, deep neural network (DNN), and long short-term memory (LSTM) models were evaluated to compare statistical, feed-forward neural-network, and recurrent neural-network approaches to one-month-ahead prediction using Ukrainian epidemiological data from 2020 to 2023. For chest X-ray analysis, YOLOv8 and YOLOv11 nano and small configurations were compared in a proof-of-concept experiment involving five pulmonary categories. The two analytical modules were integrated at the application level within a common web-based interface providing data upload, model selection, analysis, and result visualization. Under the adopted one-month-ahead hold-out evaluation, DNN and LSTM generally provided lower forecasting errors than the ARIMA baseline. In the image-analysis experiment, YOLOv11s produced the highest validation mAP@0.5:0.95 among the evaluated configurations, reaching 0.74. However, the public chest X-ray dataset had been augmented by its provider before acquisition, and the present image-analysis evaluation used training and validation partitions without an independent test subset for final performance assessment. Therefore, the reported image-analysis metrics characterize only the adopted validation partition, may overestimate performance on independent data, and should not be interpreted as estimates of clinical or external generalization. In addition, the manually prepared bounding-box annotations were not independently verified by a qualified clinical expert. Accordingly, the image-analysis component should be regarded as a proof-of-concept computational evaluation rather than a clinically validated diagnostic system. The developed prototype demonstrates the technical feasibility of providing epidemiological forecasting and preliminary chest X-ray image analysis within a common application environment for emergency and resource-constrained settings.

Algorithms doi: 10.3390/a19100893

Authors: Wenjie Wang Huiying Xu Hanjiang Lin Hongdan Gu Longfei Wang Yaxing Yang Mengyang Li Zhendong Chen Wenhao Li Xinzhong Zhu

Lightweight single-image super-resolution (SISR) requires effective local structural representation and broader contextual modeling under a restricted computational budget. We propose DPALSR, a Dual Prior-Aware Lightweight Super-Resolution Network that conditions successive attention-query representations with complementary directional cues. Learnable Constrained Laplacian Attention (LCLA) derives local directions from the gradient of a learnable zero-sum structural response, whereas Fourier Neural Operator Attention (FNOA) derives coarser within-window contextual directions through compact spectral processing. These cues provide structured inductive biases for learned feature interaction. On five standard SISR benchmarks, DPALSR achieves a competitive reconstruction accuracy–complexity trade-off. At ×4, it obtains 27.01 dB/0.8128 on Urban100 and 31.52 dB/0.9203 on Manga109 using 887K parameters and 59.8G reported FLOPs. Relative to the published MambaIRv2-Light operating point, these PSNR values are higher by 0.19 and 0.28 dB, respectively, with approximately 20.9% fewer FLOPs.

Algorithms doi: 10.3390/a19100892

Authors: Nyurgun P. Lazarev Victor A. Kovtunenko

Inspired by bio-medical applications, a new mathematical model of the equilibrium of a heterogeneous Timoshenko plate with two thin rigid inclusions connected by a hinge joint is studied. It is assumed that one inclusion is not delaminated from the plate’s elastic matrix, while the second inclusion is delaminated on both sides, thus generating two cracks. Nonlinear conditions of the non-penetration type are imposed on the delaminated cracks. The existence of a variational solution to the delamination problem is established based on the Weierstrass theorem. Optimality conditions characterizing the solution in the domain, as well as on the curves describing the cracks and the inclusion, are derived in the form of a primal-dual system. For its solution, a semi-smooth Newton algorithm is provided rigorously and tested numerically.

Algorithms doi: 10.3390/a19100891

Authors: Ziyi Chen Yongsheng Yao Chen Liu Yongdong Lu

Transparent detection of underground space is crucial to urban safety and resource development, and ground-penetrating radar (GPR) inversion is the core technology to achieve this goal. In recent years, GPR inversion has undergone major changes, shifting from traditional physics-driven methods to deep-learning-driven methods and their fusion strategies. This article systematically reviews the technological evolution of GPR inversion technology, taking 257 documents screened by bibliometric analysis of the Web of Science Core Collection from 2018 to 2025 as the main sample, focusing on the ground-penetrating radar electromagnetic parameter inversion algorithm and on the reconstruction of parameters such as dielectric constant and conductivity. First, traditional physics-driven methods, including ray tomography, hyperbola fitting inversion, and full waveform inversion (FWI), are reviewed; their advantages in physical interpretability are discussed, while their limitations, such as computational inefficiency and strong dependence on initial models, are also pointed out. Subsequently, pure data-driven architectures were classified and sorted according to the three dimensions of A-scan, B-scan and C-scan, covering convolutional neural networks (CNNs), recurrent neural networks (RNNs) and generative adversarial networks (GANs). Additionally discussed is the fact that, although these have the advantage of millisecond-level real-time reasoning, they still lack physical consistency, rely on annotation or simulation data and have limited cross-scenario generalization capabilities. In addition, the article focuses on fusion methods, including model parameter prior regularization, physical information neural networks (PINNs), closed-loop architecture and multi-task collaborative guidance. The article also clarifies the mechanism by which to improve physical consistency through model parameter prior constraints, Maxwell equation residual embedding, differentiable forward closed-loop or network structure enhancement. Finally, three development directions are proposed: enhancing physical interpretability, optimizing real-time inversion calculations, and integrating multi-source heterogeneous data. This review aims to provide a comprehensive reference for the theoretical research and engineering application of GPR inversion methods.

Algorithms doi: 10.3390/a19100890

Authors: Lin Li Tao Wang Yongqiang Chen Wenjuan Wang Dan Yang

To address the challenges posed by the structural complexity and frequent resource conflicts in flexible job shop scheduling (FJSP) at practical power grid inspection bases, this paper formulates a makespan-minimization optimization model and proposes a dual-loop multi-strategy migrating birds optimization algorithm (DMMBO). The algorithm first generates a high-quality initial population by integrating the MOR operation priority rule with a weighted shortest processing time (SPT). Then, a four-phase perturbation mechanism for leader birds is designed to accelerate convergence toward optimal solution. To handle the critical path structure, four targeted strategies—time window relaxation, entropy-driven migration, path topology reconfiguration, and temporal shifting of key operations—are introduced to effectively alleviate bottlenecks. Within the algorithmic framework, DMMBO adopts a dual-loop optimization structure: the inner loop conducts local search based on simulated annealing criteria and an adaptive perturbation acceptance mechanism to enhance convergence accuracy, while the outer loop incorporates global perturbations and a restart strategy to maintain population diversity and prevent premature convergence. To evaluate the effectiveness of the proposed DMMBO algorithm, simulation experiments were conducted on both real-world dry-type transformer inspection tasks and Brandimarte instances. The results indicate that DMMBO outperforms other comparative algorithms in terms of solution quality and result stability, demonstrating its practical value in solving complex power grid scheduling problems.

Algorithms doi: 10.3390/a19100889

Authors: Katerine Marí Sadie Johan Adam du Preez Willie Brink

Probabilistic graphical models (PGMs) provide a powerful framework for modelling complex systems, but inference over loopy graphs requires approximate methods whose accuracy depends on how factors are clustered in the graphical representation. Existing factor clustering methods rely on the number of variables in a cluster to bound its memory footprint, or on the maximal possible entropy as a proxy for how informative a merged cluster would be—both loose approximations that can lead to suboptimal merging decisions. We address this limitation by proposing an efficient algorithm for estimating the joint entropy of a group of clusters directly, without explicitly multiplying out the constituent factors, thereby avoiding the exponential computational cost that makes exact computation prohibitive. The algorithm integrates naturally with both static and dynamic graph restructuring methods, and reduces to a specific case of the Kikuchi entropy approximation when applied to the complete graph. Building on the entropy estimation algorithm, we further propose an algorithm for estimating the number of non-zero probabilities in a factor product, again without explicit multiplication. Experiments on models with up to 24 variables demonstrate that the algorithm produces accurate entropy estimates (when compared with ideal junction tree performance) across diverse model types, with errors remaining within tight bounds. Scalability is further validated on a substantially larger model defined over 2640 random variables. Our results confirm that accurate entropy estimation is achievable wherever reliable probabilistic inference is possible, and that the proposed estimation algorithm yields approximations objectively close to the true entropy values, thereby supporting improved clustering decisions in PGM structuring algorithms.

Algorithms doi: 10.3390/a19100888

Authors: Irmak Hatıpoglu Nuray Akar Mariappan Kadarkarainadar Marichelvam Ömür Tosun

This systematic review examines the evolution of Artificial Bee Colony (ABC)-based scheduling research between 2011 and 2025, based on 197 peer-reviewed journal articles systematically identified through a search of the Web of Science Core Collection (up to December 2025). No risk-of-bias assessment was conducted at the study level; coding reliability was confirmed using Cohen’s Kappa (κ = 0.82). The findings were synthesized using bibliometric mapping (VOSviewer), binary logistic regression, and Monte Carlo permutation-based chi-square tests. Hybrid and memetic ABC variants dominate the literature (78.7%), while standalone applications are rare (4.1%). Publication activity has increased significantly since 2020. The likelihood of adopting sustainability-focused objective functions has increased significantly over time (OR = 1.78; 95% CI [1.34; 2.39]; p < 0.001). Learning-based mechanisms, particularly Q-learning adaptations, have emerged as a new phenomenon specific to the period after 2021. The findings indicate that hybridization remains the backbone of the field, while learning-based and sustainability-integrated approaches point toward the future. Two limitations stand out: the heterogeneous “Others/Specialized” category and the limited sample size for the post-2021 period. This research received no external funding and was not registered in a public registry.

Algorithms doi: 10.3390/a19100887

Authors: Hunter Young Steve Wainaina Cam Peters John Quinn Chad A. Oian Taufiquar Khan

Physics-informed neural networks (PINNs) have recently emerged as an effective surrogate to classical numerical methods for solving partial differential equations (PDEs) in a wide range of physical systems. When augmented with supervised data, PINN performance can depend strongly on the quality and characteristics of the training data; however, the influence of the numerical method used to generate that data remains largely unexplored. In this study, we investigate how training data generated by finite difference method (FDM) and finite element method (FEM) solvers affects PINN training and predictive accuracy, particularly in solving the one-dimensional heat equation under conditions relevant to laser–tissue interactions. A supervised data loss term is incorporated, and models trained on FDM- and FEM-generated temperature fields under identical physical conditions are compared to assess differences in convergence and accuracy. The results demonstrate that, under controlled one-dimensional diffusion settings, the numerical solver used for data generation produces a quantifiable, resolution-dependent effect on hybrid PINN performance: FDM-generated training data yields the lowest error across most tested resolutions, an advantage that narrows and becomes indistinguishable only at the finest resolution considered, providing practical guidance for selecting a data-generating solver in biological heat transfer applications.

Algorithms doi: 10.3390/a19100885

Authors: Bagdat Yagaliyeva Olga Ivashchuk Dmitry Goncharov Yergali Baibolatov Kaiyrbek Makulov Atkeldy Ogan Arshyn Nauryzbayeva

This paper presents a set of algorithms and models for the comprehensive assessment of the environmental situation in urbanised areas subject to anthropogenic impacts from industrial and transport facilities. The proposed mechanism combining fuzzy logic and neural network predictive models, with subsequent spatial visualization of the results using GIS technologies, enabling a shift from traditional component-by-component assessments to a holistic spatio-temporal characterisation of the environmental situation, with the possibility of formulating well-founded management measures. An algorithm for generating management scenarios has been developed, ensuring a closed-loop cycle: monitoring—integrated assessment—forecasting—development of environmental protection measures. The algorithm was tested using field environmental monitoring data from urban areas, collected at 82 road sections (2023–2025), for which motor vehicle traffic was identified as the primary source of anthropogenic pressure. Forty-three mathematical models were developed based on an artificial neural network, which achieved a coefficient of determination of 0.75–0.82 on a test data sample. Spatial visualisation of the results was implemented in the ArcMap 10.8 GIS system, whilst modelling was carried out in the MatLab environment. The objective of this study is to develop a set of tools for a comprehensive assessment of the environmental conditions in urbanized areas. The proposed set of tools can be expanded to include other components, however, testing is currently limited to two components (CO and noise) in a single city (Oryol, Russia).

Algorithms doi: 10.3390/a19100886

Authors: Magzhan Kairanbay Anar Fazylzhan Akbayan Bekarystankyzy Zhanara Zhumabayeva Nurlan Kairanbayev

Text-to-speech (TTS) synthesis for low-resource languages such as Kazakh is still difficult: the language is agglutinative, uses an extended Cyrillic alphabet and has few annotated speech corpora. We present Kazakh TTS systems trained on a new corpus of about 3000 pages of educational and scientific text from 23 domains, read by two professional speakers. A character-level Tacotron 2 model with guided attention and a Parallel WaveGAN vocoder was trained in ESPnet2 on the female voice, and end-to-end VITS models were trained at 24 kHz for both voices. Guided attention produced a sharp monotonic alignment within 21 epochs. On a new test set of 138 sentences, transcribed with a Kazakh-specific automatic speech recognition (ASR) model, VITS reached a character error rate (CER) of 1.94% (male) and 2.08% (female), against 8.79% for Tacotron 2, whose attention failed on 6.2% of the utterances. On 60 sentences with formulas, numbers and abbreviations, normalised by the system’s text front-end, Tacotron 2 failed on 11.7% of the utterances, while VITS had no failures and lowered the CER from 19.7% to 7.2%. The multilingual Whisper model reported about ten times higher error rates on the same audio, which shows that Whisper is not a reliable intelligibility measure for Kazakh. In a listening test with seven native speakers, Tacotron 2 with Parallel WaveGAN obtained a mean opinion score of 4.5 out of 5 for both naturalness and intelligibility.

Algorithms doi: 10.3390/a19100884

Authors: Amina Shoukat Muhammad Safyan Muhammad Waseem Asghar

The process of image denoising is very crucial in the field of image processing, since noise tends to adversely affect both the appearance and structure of images that are applied in different fields of study such as science and engineering. In this research paper, we present a new hybrid image-denoising scheme called Modified Total Generalized Variation (MTGV) for the restoration of noisy images. The proposed methodology makes use of the BM3D model along with a Modified Total Generalized Variation model to achieve better noise reduction while retaining image features, edges and other details. In the proposed scheme, the BM3D model is first applied to reduce the noise in the observed image. The mean of the observed image and the one obtained after applying the BM3D model is computed, which is further used in the restoration process using the proposed MTGV model. In order to measure the effectiveness of the suggested technique, the computational experiments were performed under different noise levels. The proposed approach is quantitatively analyzed in comparison with the BM3D and TGV approaches based on the metrics of PSNR and SSIM as well as computation time. Moreover, visual comparisons and intensity profiles along the horizontal and vertical directions are provided for selected examples of both low-noisy and high-noisy images. From the results of the experiment one can see that using the new MTGV technique results in noise suppression, preserving structural and intensity characteristics of original images.

Algorithms doi: 10.3390/a19100882

Authors: Morteza Fani Tazekand Kahlan Mohammad Mahdi Rasoul Amini Sina Samadi Gharehveran Kimia Shirini Arya Abdollahi Ahad Shahhoseini

Cross-context accuracy in electroencephalography (EEG)-based emotion recognition is usually reported without the model-selection rule, chance-level test, or reliability reference needed to interpret it. We quantify the consequences in simulations with known ground truth and on real EEG features with randomly assigned labels. Selecting a linear classifier’s regularization parameter on the test set gave mean cross-context accuracies of 0.706 (two classes; chance 0.500) and 0.212 (seven classes; chance 0.143) on data without signal, while nested cross-validation stayed at chance. Real EEG features from 89 participants gave 0.640 and 0.207. Because the offset is shared by all participants, group-level t-tests were falsely significant in 95–100% of groups of 10–60 participants. The offset was nearly constant across true effect sizes and appeared for every model class tested, for pooled feature selection, and for repeated test-set reuse. Representational similarity analysis was less sensitive than decoding when conditions shared mean patterns across contexts, and it alone detected structure that was shared relationally but carried by different features. A single condition carrying a power artifact produced a spurious cross-context geometry. An exact permutation test was falsely significant in about one-third of realizations, whereas a conditional permutation test and condition dropping remained calibrated. We derive five reporting items from these results.

Algorithms doi: 10.3390/a19100883

Authors: Volodymyr Sokolov Pavlo Skladannyi

In the 2.4 GHz band, the channel of a short-range link is usually selected at one end; in Bluetooth Low Energy, a peripheral’s channel classification is advisory. This paper examines, for one instrumented case, whether the two ends of a link perceive interference similarly. A purpose-built protocol on two nRF52840 devices reported each end’s noise floor once per adaptation window. On a 4.2 m indoor link with one swept narrowband interferer, 656 time-paired windows were obtained. In the 614 quiet windows, the transmitter read a standing +6.30 dB more than the receiver, whereas a co-location control gave a mean offset of 0.66 dB. In the 42 windows within fifteen interference episodes, the difference had a standard deviation of 19.2 dB, and the two ends classified the channel differently in 38. Co-located devices also differed by a median of 9 dB under a directed source, so these differences cannot be attributed wholly to position. With two simultaneously sampled receive chains 60 cm apart, the levels received from a moving source were uncorrelated. A negotiation in which the receiver selects among the transmitter’s candidate channels completed 17 of 17 changes on a quiet bench and 43 of 44 under interference; in all 43, however, it selected the channel that transmitter-only selection would have used. Endpoint perception therefore differed substantially in this configuration, whereas an advantage of the method over simpler alternatives was not demonstrated.

Algorithms doi: 10.3390/a19100881

Authors: Alaa A. K. Ismaeel Taher Anwar Ali M. El-Rifaie Ahmed A. F. Youssef Mokhtar Said

Accurate identification of transformer insulation faults is essential for reliable power-system operation, and dissolved gas analysis (DGA) remains an important diagnostic tool. However, conventional Rogers’ Ratio interpretation relies on fixed decision boundaries that may not represent overlapping gas patterns and measurement variability. This study proposes an explainable Hybrid FGO–TLBO framework that adaptively calibrates four Rogers’ Ratio decision boundaries while preserving the original rule-based diagnostic logic. The study uses 500 real DGA records representing six classes: Normal, Partial Discharge (PD), Low-Energy Discharge (D1), High-Energy Discharge (D2), Low-Temperature Thermal Fault (T1), and Medium-Temperature Thermal Fault (T2). A stratified 80/20 split was used, with 400 records for threshold calibration and 100 untouched records for held-out evaluation. The optimization employed 50 independent runs, 100 iterations, a population size of 50, 2% Gaussian perturbation during calibration, and a penalty coefficient of λ = 0.05. The proposed Hybrid FGO–TLBO achieved a mean calibration fitness of 0.70892 and a standard deviation of 2.41 × 10−4 over 50 runs. On the held-out test set, the adaptive framework achieved 83% accuracy and a macro-F1 score of 0.7701, compared with 67% and 0.6064, respectively, for conventional Rogers’ Ratio. The optimized thresholds were x1 = 0.227385, x2 = 1.337761, x3 = 0.147130, and x4 = 0.556188. The results support hybrid metaheuristic calibration as an interpretable threshold-optimization approach, while the held-out results are presented as an internal validation and not as evidence of universal superiority or guaranteed optimality.

Algorithms doi: 10.3390/a19100880

Authors: Guodong Duan Bing He Xiaowei Shen Weijie Kang

Distributed navigation interference is a critical technique for degrading the performance of multi-element adaptive nulling antennas, where optimal spatial deployment of interference sources is pivotal to maximizing overall interference effectiveness. However, interference signal propagation is constrained by transmit power, path loss, and terrain occlusion, rendering optimal cooperative source deployment in complex terrain a core challenge in navigation spectrum analysis and interference management. To address this gap, this paper proposes DDL-NSGA-II, an improved NSGA-II algorithm for distributed navigation interference source deployment. By integrating the free-space propagation model with reference-plane-based viewshed analysis, the effective coverage of interference signals in 3D terrain is estimated, and a bi-objective model is constructed with multi-weight coverage rate and effective dispersion as optimization objectives. Three well-established mechanisms are adapted and integrated into standard NSGA-II for this specific problem: differential evolution mutation with binomial crossover, stage-wise dynamic deduplication, and external archive-assisted local search refinement. The contribution lies in their problem-specific integration and adaptation to integer-encoded multi-source deployment, which significantly enhances global exploration and convergence accuracy. Simulation results show that DDL-NSGA-II delivers high-coverage deployment schemes with favorable dispersion under varying terrain resolutions, source quantities, and single-station interference ranges. It outperforms three baseline algorithms in convergence, Pareto front coverage, and solution uniformity, providing robust algorithmic support for intelligent multi-source cooperative deployment in complex terrain.

Algorithms doi: 10.3390/a19100879

Authors: Suresh Rasappan Mani Sankar Praveen Kumar Poonia

This paper develops a reliability and availability model for a repairable two-tier k-out-of-n computer network system. Four analytical components are combined: event-triggered repair intensification, Gumbel–Hougaard copula-based restoration, a semi-Markovian state-transition formulation, and a reinforcement Q-learning layer for adaptive threshold selection. Subsystem 1 consists of n identical web servers operating under a k-out-of-n: G policy. Subsystem 2 contains three identical database replicas under a one-out-of-three: G policy. The full system is available only when both subsystems and both load balancers are operational. State probabilities are derived via the supplementary variable technique and Laplace transforms. The semi-Markovian extension decouples embedded-chain transition probabilities from state-specific holding-time distributions. Stationary availability is expressed as the weighted ratio of mean up-state sojourn times to total mean sojourn times, accommodating non-exponential failure and repair time laws. The event-triggered policy fires repair intensification when the system availability A(t) falls; specifically, when the degradation indicator Ξ(t)=1−A(t) crosses a threshold subject to a minimum inter-event time. A Q-learning agent selects the threshold–gain action from a finite set, updating a Q-table by the Bellman rule. Numerical results from MATLAB 2026 show baseline availability converging to 0.8581, 0.8537, and 0.8496 for k=5, k=15, and k=25. Without repair, reliability reaches 0.0720, 0.0521, and 0.0398 at t=25. The fixed event-triggered policy generates two trigger events per configuration, achieving area-under-the-curve (AUC) gains between 0.5412% and 0.5652%. The Q-learning policy generates three trigger events and achieves larger AUC gains for the stricter configurations: 0.7154% for k=15 and 1.0066% for k=25. These gains are validated by 30 Monte Carlo replications with non-overlapping 95% confidence intervals for k=15 and k=25. Under Weibull aging with shape βw=1.8, semi-Markovian stationary availability is lower than the exponential Markov baseline for all values of k.

Algorithms doi: 10.3390/a19100878

Authors: Guangyu Pan Yao Chen Bo Hou Wende Yang Yuhui Chen

Reinforcement learning is increasingly deployed in autonomous systems whose reward functions encode operational priorities and decision-making objectives. However, observable behavior can enable external observers to infer sensitive reward parameters through inverse reinforcement learning. Existing deployment-time defenses mainly rely on action randomization or deception, without explicitly accounting for how previously released behavior affects the additional reward-sensitive information contributed by subsequent actions. To address this gap, we propose FisherShield, a deployment-time mechanism that balances task value against marginal reward-information exposure. Under a specified reference observer model, FisherShield computes the sensitivity of action log-likelihoods to the private reward and accumulates the sensitivity directions of released actions in an empirical information matrix. For each candidate action, the increase in the matrix’s log-determinant measures its marginal reference-score exposure and is converted into a weighted penalty subtracted from the task value. Updating the matrix after each release makes the penalty depend on the observation history, enabling sequential exposure control without retraining the task policy. Experiments show that FisherShield reduces cumulative reference-score log-det exposure by 26.4% in FourRooms and 34.3% in LavaGrid relative to Expert while preserving near-expert utility. The results are limited to the specified reference-model Bayesian observer and do not establish robustness to model misspecification or a defense-aware attacker.

Algorithms doi: 10.3390/a19100877

Authors: Omneya Attallah Adem Maman Ishak Pacal

Distinguishing malignant pulmonary lesions from benign nodules, infection-related abnormalities, and normal findings remains difficult because structural and metabolic imaging patterns overlap. This study presents a multi-textural feature-fusion framework for four-class classification using independent computed tomography (CT) and positron emission tomography (PET) cohorts comprising 1213 CT and 1013 PET images. Most patients contributed one representative image and some contributed two; known same-patient images were kept together when linkage was available, but complete linkage was not retained after de-identification, so the protocol is described as hybrid and partially patient-aware. Original images were complemented by DCT-based illumination-normalized representations and one-level stationary-wavelet representations, and the two transformed branches were combined through DWT-based multi-textural fusion before NNMF, multi-backbone integration, mRMR selection, and multiclass SVM classification. The best PET model achieved 0.930 accuracy, 0.885 macro F1-score, and 0.868 MCC; the corresponding CT values were 0.878, 0.781 macro F1-score, and 0.759 MCC. Benign findings remained the least stable category. Higher PET performance reflects these independent institutional cohorts and display-image inputs and should not be interpreted as intrinsic superiority of PET over CT. The study remains limited to representative two-dimensional images and internal validation.

Algorithms doi: 10.3390/a19100875

Authors: Danyang Jiang Minghui Lyu Zhaopeng Liu Mi Wen

Simulation model parameter optimization requires selecting optimizers from task knowledge and translating those decisions into execution configurations that satisfy interface, authorization, and budget requirements. We propose ParaOpt-Agent, a knowledge-driven pre-execution agent that combines task and optimizer descriptions, historical cases, and large language model scoring. Optimizers are selected by equally combining model ranking with historical winner-coverage ranking, and deterministic checks then produce call configurations. The portfolio contains 10 optimizers, and the combined ranking uses 224 historical records. Across 2 batches totaling 64 new instances, the post hoc pooled mean utility difference relative to a same-information history rule is +0.001909; the individual batch differences are +0.013045 and −0.009227. On a further 320 new tasks, the Agent produces 319 valid actions and meets a five-percentage-point non-inferiority criterion against a model without history, the same-information rule, and an ExtraTrees selector. A same-input control ablation retains acceptance of all 16 legal inputs and reduces optimizer-entry arrivals among 528 constructed faults from 448 to zero. Public-observation cases using the 20-task history selector demonstrate parameter search and temporal validation. This study separately quantifies the selection utility of combined knowledge, action validity, and the contribution of execution checks, providing empirical evidence for optimizer decisions in simulation modeling.

Algorithms doi: 10.3390/a19100876

Authors: Zhiyue Wu Zhengyan Ding Yuantong Chen Li Qi Yanfeng Shang

Open-vocabulary text-based person retrieval must match natural-language descriptions containing concepts that may be absent from task-specific training data while controlling unreliable generated evidence and inference cost. This paper formulates heterogeneous evidence reconstruction and masked branch fusion as a trainable retrieval procedure around frozen foundation backbones. EmbedTPR contains two interconnected learnable modules. Unified Heterogeneous Evidence Reconstruction (UHER) aligns gallery-side generated semantics with visual evidence and estimates attribute reliability; for a text-only query, the semantic path operates without visual tokens and uses generation agreement to calibrate query attributes. Query-Adaptive Coherent Evidence Fusion (QACEF) predicts query-specific weights over the three branches available to a text query through a retrieval–consistency–calibration objective. The optional image-to-image branch is part of the general formulation but is not evaluated or claimed in this study. Only 15.2 M of 5015 M parameters are trainable. EmbedTPR obtains 49.66% Rank-1 and 64.45% mAP on CUHK-PEDES, 46.20% Rank-1 and 58.30% mAP on ICFG-PEDES, and 54.30% Rank-1 and 68.10% mAP on RSTPReid. Its CUHK-PEDES Rank-1 is lower than that of CLIP ViT-L/14, whereas Rank-5, Rank-10, and mAP are higher. Cross-dataset transfer from CUHK-PEDES to ICFG-PEDES reaches 18.20% Rank-1. On the sequestered FineGrained-1282 query/label evaluation manifest, learnable fusion improves mAP by 4.04 points over fixed weights. Component removal, evidence perturbation, and 1k–100k gallery tests characterize the evaluated accuracy, robustness, and computational scope. These results support reliability-calibrated query-adaptive fusion for text-query person retrieval; broader multimodal-query and open-world generalization remain to be validated. The source code, configurations, prompts, and evaluation scripts will be released publicly upon acceptance.

Algorithms doi: 10.3390/a19100874

Authors: Lyudmila Kovalchuk Andrii Nastenko Oleksii Bespalov Volodymyr Artemchuk

Preferential voting allows participants to rank multiple alternatives according to their preferences rather than being restricted to selecting a single option. Such mechanisms are used in national and local elections and, in decentralized governance, can be applied to committee selection, the prioritization of protocol upgrades or funding applications, and the choice of an electricity supplier or energy source. Of particular practical interest are systems in which participants have different stakes and their voting weights are proportional to those stakes. In the model considered in this work, each voter submits a complete ranking of the alternatives. The key technical challenge is to verify that an encrypted ballot encodes a valid ranking, that is, that exactly one rank is assigned to each alternative and each rank is used exactly once, without revealing the voter’s preferences. We propose a UC-secure protocol for privacy-preserving preferential voting with verifiable encrypted ballots. A ranking is represented as a permutation matrix whose entries are encrypted using an additively homomorphic cryptosystem. To verify ballot validity, each encrypted row of the matrix is accompanied by a non-interactive zero-knowledge proof that it is a unit vector. The encrypted rows are then homomorphically summed, and the resulting vector is checked against the all-ones vector. Together, these checks ensure that the encrypted matrix is a permutation matrix without revealing the ranking itself. Using an O(logm)-size unit-vector proof for each of the m rows yields O(mlogm) ballot-validity proof data, while the overall ballot size remains O(m2) because the encrypted matrix contains m2 ciphertexts. The protocol supports stake-weighted tallying and threshold decryption, so that only the final aggregate result is revealed. The UC security of the protocol is established in the random-oracle model under the Decisional Diffie–Hellman (DDH) assumption. The theoretical analysis shows that the computational complexity of generating and verifying a single ballot is O(m2) and that the ballot size is also O(m2), where m is the number of alternatives. The experimental results are consistent with this quadratic scaling: for 32 alternatives, the mean ballot-generation time is 87 ms, the mean verification time is 36 ms, and the reported cryptographic ballot payload is 106 KiB. These measurements indicate practical ballot-construction and verification costs for elections with a moderate number of alternatives.

Algorithms doi: 10.3390/a19100873

Authors: Muhammad Waseem Asghar Mujahid Abbas

In this paper, we introduce a new class of mappings called enriched Cλg-nonexpansive mappings, which extends the class of g-nonexpansive mappings employing an enriched Cλ-type condition. We discuss the existence of its fixed point and approximation of generalized g-attractive points associated with the proposed class of mappings. In particular, we study several structural properties of the corresponding attractive point set and develop a g-averaged viscosity-type iterative scheme to approximate its elements. Under suitable assumptions for the control sequences, we prove the strong convergence of the generated sequence to a generalized g-attractive point. Fixed-point and coincidence-point results are thus obtained as special cases of the proposed approach. As an application, we apply our result to find the existence of the best approximations. We also illustrate the applicability of the proposed iteration method through a numerical experiment. The numerical results demonstrate the effect of the parameter λ on the convergence behavior and show that larger values of λ provide better numerical performance for the considered example.

Algorithms doi: 10.3390/a19100872

Authors: Osman Yildiz Muhammed Enes Subasi Abdulhamit Subasi

Depression is a widespread mental health disorder whose early diagnosis is often hindered by the reliance on subjective clinical assessments. This study proposes a subject-aware machine learning framework for detecting depression using wrist-worn actigraphy data, enabling continuous and objective monitoring of daily activity patterns. Unlike many previous studies that evaluate models at the day or window level, the proposed approach employs strict leave-one-subject-out (LOSO) cross-validation, ensuring that no data from the same individual appear in both the training and testing sets and thereby preventing information leakage. The framework extracts time-series features with the Time-Series Feature Extraction Library (TSFEL) from daily actigraphy signals and aggregates them into subject-level representations using distributional statistics. Classical circadian rhythm biomarkers, including interdaily stability (IS), intradaily variability (IV), relative amplitude (RA), cosinor parameters, and activity fragmentation measures, were also investigated to improve interpretability. Seven feature selection techniques and fifteen machine learning classifiers were systematically evaluated. Model selection is performed entirely inside each training partition. The feature selector, the classifier and the number of retained features are chosen together by inner cross-validation over 840 candidate configurations, so the held-out subject contributes nothing to the choice of pipeline. Averaged over five random seeds the nested protocol classifies 45.4 of 55 subjects correctly (82.5%; range 80.0–85.5% across seeds; mean Cohen’s κ = 0.637; and mean sensitivity 74.8%, mean specificity 88.1%). For the median run, 45 of 55, the Wilson 95% confidence interval is 69.7–89.8%. The estimate is above the 58.2% majority-class rate (exact McNemar, p = 0.003), and above the 73% reported by the previously published subject-disjoint evaluation of the same 55-subject cohort; per-subject predictions are not available for that study, so the latter comparison is unpaired. Recording length is a confound in this cohort—used alone as a single feature under the same protocol it reaches 72.7%—and at n = 55 the framework is not separable from that baseline (exact McNemar p = 0.09 to 0.42 across seeds; achieved power 0.40). The single best configuration, Information Gain feature selection with a Bagging classifier, reaches 92.7%. That value is the maximum over a grid of 105 selector–classifier pairs and we report it as such, not as an out-of-sample estimate. Although incorporating circadian biomarkers did not improve classification accuracy (92.7% vs. 90.9%), one bout-fragmentation biomarker, the within-subject mean of daily bout count, was retained in every cross-validation fold; it is, however, a rescaling of the daily Zero Crossing Rate already present in the signature, so the retained features summarize a single daily quantity rather than several independent ones. On this benchmark, subject-level machine learning on wearable actigraphy classifies well above chance under an evaluation protocol that matches the clinical screening setting, but this cohort is too small to establish that the learned representation adds to recording length alone. All results come from one single-site cohort of 55 subjects, and the framework has not been evaluated on an independent dataset. The findings are therefore a reference point for Depresjon, not evidence of generalization to other populations.

Algorithms doi: 10.3390/a19100871

Authors: Qinghua Cui Yipeng Geng Huazhen Xu Jianlin Qiu Jiajun Zhu

Monitoring forest diseases is essential for sustainable forestry management and ecological security. However, unmanned aerial vehicle (UAV)-based forest inspection remains challenging because diseased pine canopies are often small, densely distributed, partially occluded, and easily confused with complex background patterns caused by illumination variations, shadows, and surrounding healthy vegetation. To address these challenges, this paper proposes a lightweight detector, termed Lightweight Feature Transfer YOLO (LFT-YOLO), based on the YOLO26n architecture. The central design of LFT-YOLO is a lightweight small-target information preservation–reconstruction pathway. First, a Local Feature Injection (LFI) module is introduced to establish an independent shallow-feature transfer path, enabling high-resolution spatial and texture information to be directly aligned with and injected into the P3 feature space, thereby alleviating the loss of fine-grained information for small targets during progressive downsampling and multi-scale fusion. Second, conventional fixed upsampling in the neck is replaced with DySample-based adaptive upsampling to improve spatial reconstruction and cross-scale feature alignment. Experiments on the public Pests and Diseases Tree-Low-Level (PDT-LL) dataset show that LFT-YOLO achieves a Precision of 87.75%, a Recall of 87.02%, and an mAP@0.5 of 93.31%, corresponding to a 0.51-percentage-point improvement over the YOLO26n baseline. More importantly, the size-specific evaluation demonstrates that the performance gains are concentrated on small targets: compared with YOLO26n, small-object AP@0.5, AP@0.75, AP@0.5:0.95, Precision, and Recall improve by 2.15, 2.55, 2.15, 2.10, and 2.30 percentage points, respectively. These results indicate that the proposed architecture provides more pronounced benefits for the dominant small-object detection regime while largely preserving the performance of medium- and large-scale targets. LFT-YOLO contains only 2.43 M parameters and requires 5.80 GFLOPs. On an NVIDIA Jetson Orin Nano Super 8 GB platform using TensorRT FP16 under a fixed 15 W power mode, the model achieves 90.39 ± 0.40 FPS, with a P95 inference latency of 11.37 ms and an average power consumption of approximately 8.53 W. These results demonstrate that LFT-YOLO provides a favorable accuracy–efficiency trade-off and is suitable for real-time UAV-based detection of small and densely distributed pine wilt disease targets under resource-constrained edge-computing conditions.