Search Results (4,448)

Compact face verifiers at access-control and surveillance terminals must operate under specular reflection, off-frontal pose, and mask occlusion, disturbances that frequently co-occur in practice. Existing objectives optimize identity discrimination and score-threshold reliability separately. As a result, disturbance cues leak into the identity embedding and inflate the impostor-score tail near the operating false-acceptance rate (FAR)—a structural weakness that backbone scaling alone does not resolve within a fixed parameter budget. We propose RCF-Face, a 1.12 M-parameter verifier for the sub-2 M-parameter, sub-0.2 GFLOP budget of Ampere-class edge GPUs, trained under a unified objective that combines representation-level leakage suppression (HSIC-based identity–disturbance factorization with augmentation-based counterfactual consistency) and decision-level score shaping via a softplus tail-risk penalty anchored to the target FAR. Across four disturbance-focused protocols (CFP-FP, IJB-C, RMFRD, and a reflection-augmented subset), RCF-Face reaches a Composite-Disturbance Score (CDS, the mean verification accuracy over the four protocols) of 89.05 at 0.166 GFLOPs, improving on the capacity-matched AdaFace-Lite baseline by 4.67 percentage points, with both per-protocol gains over this baseline significant under Bonferroni correction. Zero-shot evaluation on two open-source reflection-related datasets (SoF and MeGlass) shows that the robustness gains transfer to real-world images, although these datasets isolate single disturbance axes and do not constitute a composite-disturbance benchmark. On a Jetson Orin NX 16 GB, batch-1 latency is 2.41 ms (FP16) and 1.66 ms (INT8), supporting Ampere-class edge-GPU deployment. Full article

High-fidelity ship dynamics models are essential for the reliable operation of maritime autonomous systems. However, existing Markov-based maneuvering models and purely data-driven predictors struggle to capture hydrodynamic memory and degrade under non-ideal sensing. To address these challenges, this paper proposes a novel approach for robust ship motion prediction, the Non-Markovian Memory-Augmented Environment-Perceived and Physics-Informed Network (NMA-EPIN). This method explicitly models long-term hydrodynamic dependencies through a memory-augmented architecture. Within NMA-EPIN, a Control-Physics-Informed Neural Network (CPINN) paradigm enforces velocity–position kinematic consistency and control-logic alignment as soft constraints, suppressing cumulative drift under degraded observations. Experiments on a high-fidelity simulated dataset show that NMA-EPIN attains an average coefficient of determination $R^2=0.977$ under nominal conditions, effectively eliminating the position drift observed in baselines. Under extreme compound perturbations (50% sensor noise, packet loss, and delays), NMA-EPIN retains $R^2\approx 0.91$, which significantly outperforms the baselines. Full article

The trustworthiness of AI-powered network security auditing depends not only on detection accuracy but on the faithfulness of the explanations that support compliance verdicts. In IoT network security, Large Language Models (LLMs) are increasingly utilized to produce natural-language security assessments from raw network traffic, yet the extent to which these explanations are grounded in retrieved evidence is rarely measured. This paper presents the Retrieval-Augmented Generation Assessment Suite (RAGAS) as an evaluation framework that compares three retrieval paradigms—rule-based heuristic scoring, dense vector retrieval, and knowledge graph traversal—on the task of explaining network compliance against ETSI EN 303 645 IoT cybersecurity provisions. Using 30 human expert-validated compliance scenarios derived from the CIC-IoT2023 dataset and three LLMs (DeepSeek-R1, Qwen-2.5, Llama-3.2), we find that graph-based retrieval achieves the highest faithfulness (0.570), outperforming rule-based (0.524) and vector retrieval (0.509). All methods, however, exhibit low context recall (≤22.4%), and we highlight that high detection F1 scores do not guarantee faithful explanations; over 40% of statements in compliance answers are unsupported by retrieved evidence. A proof-of-concept prototype, Security Audit Compliance Agent (SACA), demonstrates how knowledge graph traversal can be integrated with interactive visualization to support human auditor oversight. We argue that, in adherence to responsible AI principles, faithfulness measurement should become a standard complement to accuracy reporting for an AI-driven network audit or forensic analysis. Full article

In large-scale state-owned enterprise audit and compliance tasks, ensuring that similar violations receive consistent disciplinary decisions is essential for procedural fairness and institutional credibility. However, existing retrieval methods face three major challenges: lexical matching methods fail to recognize semantically equivalent violation descriptions, general-purpose semantic encoders lack knowledge of inspection-specific terminology and regulatory distinctions, and retrieved precedents are often not directly transformed into actionable disciplinary references. To address these problems, this paper proposes InspectCL, a domain-enhanced contrastive learning and Retrieval-Augmented Generation framework for similar case retrieval, validated on audit data from a provincial power grid company. First, to provide task-specific supervision that is unavailable in existing benchmarks, we construct InspectCase, a de-identified dataset of 4200 audit and compliance cases across 12 violation categories, with expert-validated positive pairs and hard negative pairs. Second, to overcome the weak domain awareness of generic encoders, we design a domain-enhanced contrastive learning model. Specifically, terminology-masking augmentation improves robustness to specialized inspection expressions, regulatory semantic injection incorporates disciplinary rules to distinguish factually similar but legally different cases, and hierarchical contrastive optimization strengthens both case-level similarity learning and category-level boundary separation. Third, to convert retrieved precedents into practical decision support, the Top-K similar cases are used as evidence for a large language model to generate structured disciplinary recommendation summaries, including violation classification, penalty references, applicable regulations, and rectification measures. Experimental results on InspectCase show that InspectCL substantially outperforms BM25, BERT-base, SimCSE, and Legal-BERT baselines, achieving 56.9% ± 0.7% Recall@5 and an 87.6% ± 0.4% Penalty Consistency Score (PCS). These results demonstrate that the proposed problem-driven modules jointly improve semantic retrieval accuracy and disciplinary decision consistency, offering a practical reference for similar power-grid audit scenarios, with broader applicability to be validated in future cross-domain studies. Full article

Smart manufacturing relies on programmable logic controllers (PLCs) that translate sensor inputs into actuator commands. Generating PLC programs in legacy textual languages such as Mitsubishi FX-series Instruction List (IL) remains an expert-only task, and IL’s deprecation in IEC 61131-3 Edition 3.0 leaves it under-represented in the corpora that train modern large language models (LLMs). We benchmark ten open-source LLMs (five vendors, 7B–122B parameters) in both LLM-only and Retrieval-Augmented Generation (RAG) configurations on a frozen 285-question dataset; the pipeline uses ChromaDB with all-MiniLM-L6-v2 embeddings and Maximal Marginal Relevance (MMR) retrieval ($k=3$, $\lambda =0.5$). To move beyond lexical similarity we introduce a three-tier static syntax checker (Lexical/Syntactic/Semantic) calibrated against a $93.3\%$ ground-truth pass rate. RAG raises the syntactic pass rate by $+6.7$ to $+61.1$ percentage points across all ten models; the best configuration, qwen3.5:122b with RAG, reaches $95.8\%$, exceeding the ground-truth baseline. Two outliers (llama3.3:70b at $+6.7$ pp, gpt-oss:120b at $+25.6$ pp) are reported rather than excluded. The results indicate that for deprecated-but-deployed industrial languages a curated dialect corpus paired with a locally-hosted open-source LLM is more effective than scaling raw model size, supporting reproducible, on-premise industrial-monitoring and code-generation tooling for sustainable smart manufacturing. Full article

Named entity recognition in natural disaster emergencies is a critical foundational task for emergency management. However, existing methods face challenges including complex entity types, frequent emergence of new terminology, model knowledge obsolescence, and poor adaptability to dynamic knowledge updates, resulting in limited accuracy and generalization in real-world disaster scenarios. To address these issues, this paper proposes a named entity recognition method for natural disaster emergencies based on instruction tuning and knowledge graph retrieval-augmented generation. We first construct a dedicated instruction-tuning dataset, EM-InstructNER, and a domain-specific knowledge graph, EmergencyKG, tailored to natural disasters. Then, LoRA is employed for parameter-efficient fine-tuning of the Qwen2-7B-Instruct base model, while KG-based RAG dynamically retrieves subgraphs from the knowledge graph to generate semantically enriched augmented prompts, providing external structured knowledge support for generative NER. Experimental results demonstrate that the proposed method achieves a macro F1 score of 0.9205 on the EM-InstructNER test set, representing a 36.6% relative improvement over the best-performing zero-shot baseline (DeepSeek-R1:14B), while remaining competitive with strong supervised sequence labeling approaches (e.g., BERT + CRF). The framework provides knowledge graph update flexibility and significantly reduces training computational cost and GPU memory consumption through LoRA-based parameter-efficient fine-tuning. Cross-domain evaluation on the public CLUENER2020 benchmark further demonstrates its generalization capability. Full article

Background/Objectives: Brain tumor analysis using magnetic resonance imaging (MRI) remains a challenging task due to tumor heterogeneity, complex anatomical structures, and reliance on expert interpretation. Although deep learning approaches have shown promising results in medical image analysis, many existing studies focus on either tumor classification or segmentation independently, limiting their applicability in comprehensive automated brain tumor analysis workflows. This study proposes an integrated dual-task deep learning framework for automated brain tumor classification and segmentation using MRI scans. The framework aims to provide complementary diagnostic support by combining tumor-type prediction and tumor boundary delineation within an integrated workflow. Methods: The proposed framework utilizes EfficientNet-based convolutional neural networks for multi-class brain tumor classification and U-Net++ architectures with EfficientNet encoders for tumor segmentation. Experiments were conducted using the BRISC2025 dataset, consisting primarily of 6000 T1-weighted 2D MRI slices collected from axial, coronal, and sagittal planes. Standard preprocessing, augmentation, transfer learning, and selective fine-tuning strategies were applied. Multiple architectures were systematically evaluated using evaluation metrics. Results: EfficientNet-B1 achieved a classification accuracy of 99.70% with near-perfect precision, recall, and F1-scores across glioma, meningioma, pituitary tumor, and no-tumor classes. For segmentation, U-Net++ with an EfficientNet-B1 encoder achieved a Dice score of 0.9055, an IoU score of 0.8442, and an HD95 value of 12.21 pixels on the held-out test set. The proposed framework demonstrated robust performance in detecting small and low-contrast tumor regions while maintaining strong generalization performance across diverse MRI samples. Conclusions: The proposed integrated framework demonstrated strong performance in both brain tumor classification and segmentation tasks, effectively detecting small and low-contrast tumor regions while maintaining good generalization across diverse MRI samples. These findings suggest that the framework may serve as a reliable decision-support tool for automated brain tumor analysis in clinical practice. Full article

This study investigates data augmentation as a strategy for addressing dataset scarcity in Internet of Medical Things (IoMT) cybersecurity and improving intrusion-detection system performance. Four augmentation methods—Rule-Based, Tabular Variational Autoencoder (TVAE), Conditional Tabular Generative Adversarial Network (CTGAN), and Gaussian Copula—were applied to two publicly available IoMT datasets (ECU-IoHT and WUSTL-EHMS) to generate augmented training data with differing class distributions and feature characteristics. Eleven machine learning algorithms were evaluated using Matthews Correlation Coefficient (MCC), F1-score, accuracy, and error-based metrics. Results showed consistent performance improvements across all evaluated models relative to the baseline datasets. The Rule-Based method produced the strongest overall results, achieving the highest MCC (0.9757), F1-score (99.19%), and accuracy (99.18%) with LightGBM, alongside low false-positive and false-negative rates. Among the generative approaches, TVAE delivered the strongest overall practical performance (F1-score = 96.94%, accuracy = 96.92%), while CTGAN achieved a marginally higher MCC (0.9047) and also produced competitive results with balanced class representation. Gaussian Copula generated the weakest overall outcomes, primarily due to highly skewed class distributions. Traditional models, such as Logistic Regression and Naive Bayes, recorded the largest relative gains, indicating that augmentation can substantially improve simpler classifiers in data-scarce environments. Overall, the findings demonstrate that augmentation quality depends not only on dataset expansion, but also on preserving class balance, feature diversity, and realistic traffic relationships. These results provide practical guidance for strengthening IoMT intrusion-detection systems in healthcare environments. Full article

Automated defect detection in precision agriculture serves as a critical technology for enhancing the quality of agricultural products. Although supervised-only semantic segmentation has demonstrated remarkable performance in citrus surface defect detection, it relies heavily on training with large-scale labeled data, which results in prohibitive acquisition costs. Semi-supervised learning mitigates reliance on labeled data by generating pseudo-labels. However, existing semi-supervised segmentation methods still face challenges. On the one hand, the instability of pseudo-labels and the propagation of noise can mislead the training of semi-supervised models. On the other hand, due to the lack of semantic constraints in feature learning, models often suffer from insufficient feature discriminability when handling complex samples, such as citrus surface defects characterized by similar textures and blurred boundaries. Therefore, this study proposes UP-ETS, a dual-guided semi-supervised semantic segmentation model based on the Mean Teacher–Student framework, specifically designed for the segmentation of complex citrus surface defects. UP-ETS employs Uncertainty Estimation (UE) based on Kullback–Leibler (KL) divergence to quantify the prediction discrepancy between the teacher and student models on blurred and ambiguous pixels. This mechanism guides the model to dynamically adjust weights, thereby reducing noise propagation and enhancing pseudo-label stability under complex citrus surface textures. Prototype Contrastive Learning (PCL) is utilized to align pixel-level features of difficult samples with class prototypes, optimizing the feature discriminability for complex citrus surfaces. Experimental results demonstrate that the UP-ETS model exhibits superior semi-supervised segmentation performance. Notably, at a labeled data ratio of only 1/16, the dice improved from 85.57% to 87.76% compared to the supervised-only baseline. Furthermore, the model shows significant performance enhancements in segmenting difficult samples, such as small targets, complex boundaries, and blurred regions. The results of ablation studies and t-SNE visualization prove the effectiveness of the proposed UE and PCL. These two methods synergistically guide the model to construct a feature space that is better structured and highly discriminative. Furthermore, UP-ETS outperforms various representative semi-supervised segmentation models in terms of segmentation performance, parameters, and inference speed. In cross-dataset validation, the model exhibits robust generalization capabilities, achieving performance comparable to supervised-only methods trained on the full augmented dataset. Consequently, the framework introduced in this study effectively mitigates the heavy dependency on annotated datasets, providing significant practical value for agricultural deployment. Full article

The proliferation of digital financial news offers unprecedented opportunities for automated analysis of market-moving events. This paper presents a framework for event-based sentiment analysis of financial news that leverages Large Language Models (LLMs). The approach brings together three complementary ideas: Retrieval-Augmented Generation (RAG) for contextual enhancement, Graph Neural Networks (GNNs) for modeling relationships between events, and a multi-agent ensemble for orchestrated reasoning. The methodology targets well-known difficulties in financial text processing, including domain-specific terminology, implicit event detection, and temporal reasoning, and it combines transformer-based event extraction with sentiment classification enhanced by external knowledge retrieval. We evaluate six model configurations on an aggregated corpus of 14,851 financial news samples. On the event-detection task, every configuration reaches a weighted F1-score of 100%; we show that this is a ceiling effect produced by a binary event/no-event formulation over a highly imbalanced dataset rather than evidence of a difficult problem being solved, and we discuss what it implies for how such systems should be evaluated. On three-way sentiment classification, the strongest configuration—the multi-agent ensemble—reaches 87.4% accuracy, narrowly ahead of a RoBERTa (Robustly Optimized BERT Pretraining Approach) baseline at 87.2%; however, because the gaps reported between models are small and we did not run significance testing, we report them as indicative rather than definitive. The GNN component is described as part of the proposed design, but it has not yet been validated experimentally, and we state this limitation explicitly. The framework produces interpretable, structured outputs suited to downstream use in algorithmic trading, risk assessment, and investment decision support, and the paper contributes a reusable financial NLP pipeline together with a candid account of where the current evidence is, and is not, convincing. Full article

Laser beam welding (LBW) of aluminium busbar-to-terminal connections for electric-vehicle battery packs requires precise in-process monitoring. Membrane-free optical microphones provide a high-bandwidth (DC–MHz) acoustic channel that captures keyhole, melt-pool, and plume dynamics. This study proposes Acoustic Signal Intelligence (ASI), a deep learning framework for LBW state classification from a single optical microphone, evaluated on an open dataset (183 AA1050 welds, $f_s$ = 2.5 MHz) under a five-class taxonomy: lack of fusion, lack of connection, sound, marginal, and piercing. The contributions are: (i) a compact 1-D CNN encoder on a mel-scale STFT spectrogram, reaching the highest macro-F1 (0.72 mean across three-fold replicate-out cross-validation) and 100% piercing recall in every fold—a multi-representation fusion variant adding a wavelet-packet decomposition and a 24-feature library targeting the 8, 63 and 110 kHz keyhole-resonance peaks was evaluated as an ablation arm and did not survive cross-validation, so the proposed model is mel-only; (ii) a systematic benchmark against six classical-ML and four deep learning baselines in which Transformer-hybrid ablations and ACGAN-style augmentation underperform compared to the compact CNN on the 122-sample training set, with the Transformer underperformance confirmed by a 30-configuration grid search over learning rate, weight decay, and dropout (best tuned macro-F1 = 0.441 vs. CNN 0.724); and (iii) a Grad-CAM analysis that recovers the keyhole-resonance bands without prior knowledge. A single optical microphone is thus a viable real-time alternative to multi-sensor stacks for battery-pack laser welding. Full article

Economic diversification across the Gulf region, including Kuwait Vision 2035, positions smart manufacturing as a key enabler of sustainable growth. Yet, industrial datasets in the region are typically small, heterogeneous, and incomplete, limiting the performance and trustworthiness of conventional AI models. This paper introduces a Scalable Random Forest (SRF) framework enhanced with a Decision Path Search (DPS) mechanism to address these challenges through both technical robustness and practical interpretability. The SRF pipeline incorporates leakage-safe preprocessing, mixed-type imputation, and small-data augmentation to improve prediction stability under real industrial constraints, while DPS transforms model internals into actionable operational causal knowledge identifying optimal and avoidance parameter ranges. Case studies, including investment casting, demonstrate that SRF + DPS not only outperforms established baselines such as Random Forest (RF), XGBoost, LightGBM, and CatBoost but also deliver transparent insights that engineers can directly apply to reduce defects and enhance process control. The findings highlight how interpretable AI frameworks can accelerate industrial modernization, strengthen regional manufacturing competitiveness, and support national economic diversification strategies. Full article

Complex agricultural question answering often requires multi-step reasoning over domain-specific knowledge and reliable retrieval of heterogeneous evidence. However, existing retrieval-augmented generation (RAG) methods usually rely on one-shot retrieval and provide limited control over whether the retrieved evidence is sufficient, accurate, and consistent for answering complex agricultural questions. To address this limitation, this paper proposes an agent-driven question-answering framework for the agricultural planting domain based on a Planning–Execution–Feedback (PEF) closed-loop mechanism. The framework decomposes complex queries into executable subtasks, performs knowledge acquisition through a knowledge-graph-guided hybrid retrieval module, and iteratively refines the reasoning process according to retrieval-quality feedback. Specifically, in the retrieval stage, a two-stage strategy is introduced to first localize candidate entities in the knowledge graph and then conduct context-enhanced dense retrieval with entity-consistency reranking, thereby reducing semantic drift and improving domain alignment. In the feedback stage, the agent evaluates the adequacy of the retrieved evidence and determines whether to continue execution, re-retrieve evidence, or replan the workflow. Experimental results on the AgroQA dataset show that the proposed method achieves 88.9%, 79.1%, and 92.6% on the Answer-C, Answer-R, and CR metrics, respectively, outperforming traditional retrieval-augmented and general large language model baselines. In addition, a three-dimensional digital human interface is implemented as an application prototype to demonstrate the feasibility of integrating the proposed framework into interactive agricultural knowledge services. Full article

To address the issues in existing PCB defect detection models, including insufficient capability for capturing small defects, weaker global feature modeling, and inadequate multi-scale feature fusion, this paper proposes a C2f-FPN-PAN++-Mamba model based on an improved YOLOv8n. The Mamba state–space model is embedded into the C2f module to construct a C2f-Mamba feature extraction unit, which, while retaining the local perception capability of convolution, enhances long-range dependency modeling, accurately capturing global semantic information of subtle defects in complex backgrounds and significantly improving the model’s feature representation ability for small defects. Meanwhile, an FPN-PAN++ enhanced feature fusion structure is introduced, achieving efficient complementary interaction between high and low-level features through bidirectional cross-scale feature aggregation and path augmentation, thereby strengthening the model’s robustness in identifying multi-scale and multi-form defects. Finally, the C2f-Mamba and FPN-PAN++ are organically integrated, improving global modeling and multi-scale fusion capabilities while maintaining lightweight computational efficiency, effectively reducing the miss and false detection rates of small defects. Experimental results indicate that, compared with the original YOLOv8n model, the proposed method achieves significant performance improvements in PCB defect detection tasks. On the PCB defect dataset, the model’s precision increased from 96.4% to 98.5%, recall from 94.6% to 98.4%, and mAP@0.5 from 97.2% to 98.8%, with the mAP@0.5:0.95 metric, reflecting multi-scale detection performance, rising dramatically from 57.5% to 62.5%. Experiments demonstrate that this method effectively enhances detection capability for small and complex defects while preserving the advantages of a lightweight model and high inference speed, providing a reliable technical solution for high-precision, real-time PCB defect detection in industrial scenarios. Full article

The increasing expansion of ubiquitous sensing systems has created large streams of time-series data that are difficult for non-technical users to interpret. Large Language Models (LLMs) offer a promising interface for transforming sensor data into natural language insights, particularly in distributed environments where users may lack familiarity with data analysis. However, models optimized for text generation often struggle to interpret raw time-series signals, producing responses that are generic, inaccurate, or poorly grounded in the data. This study evaluates a prompt structure based on the Retrieval-Augmented Generation (RAG) framework for interpreting sensor-derived time-series data from water-consumption monitoring systems installed in household storage tanks. The prompt integrates statistical summaries, sensor metadata, and contextual information about household water-use practices. Performance is evaluated using synthetic datasets representing a year of tank water-consumption measurements and a rubric-based evaluation framework applied by three independent language-model evaluators. Results show that augmenting prompts with structured contextual information improves the clarity and grounding of language model responses to sensor time-series data, increasing evaluation scores and reducing failure modes such as hallucination, contradiction with the data, and misuse of contextual information, as assessed by independent evaluator models. These findings highlight the potential of structured contextual prompting to support locally deployed language models that produce reliable and actionable interpretations of sensor time-series data. Full article