Search Results (789)

Arabic automatic speech recognition (ASR) faces distinct challenges due to its complex morphology, dialectal variations, and the presence of diacritical marks that strongly influence pronunciation and meaning. This study introduces a lightweight approach for diacritical Arabic ASR that employs a Transformer encoder architecture enhanced with Relative Positional Encoding (RPE) and Connectionist Temporal Classification (CTC) loss, eliminating the need for a conventional decoder. A two-stage training process was applied: initial pretraining on Modern Standard Arabic (MSA), followed by progressive three-phase fine-tuning on diacritical Arabic datasets. The proposed model achieves a WER of 22.01% on the SASSC dataset, improving over traditional systems (best 28.4% WER) while using only ≈14 M parameters. In comparison, XLSR-Large (300 M parameters) achieves a WER of 12.17% but requires over 20× more parameters and substantially higher training and inference costs. Although XLSR attains lower error rates, the proposed model is far more practical for resource-constrained environments, offering reduced complexity, faster training, and lower memory usage while maintaining competitive accuracy. These results show that encoder-only Transformers with RPE, combined with CTC training and systematic architectural optimization, can effectively model Arabic phonetic structure while maintaining computational efficiency. This work establishes a new benchmark for resource-efficient diacritical Arabic ASR, making the technology more accessible for real-world deployment. Full article

We propose State-DynAttn, a hybrid architecture for robust air traffic flow prediction under weather disruptions, which integrates state-space models (SSMs) with dynamic graph attention to address the challenges of long-range dependency modeling and adaptive spatial–temporal relationship learning. The increasing complexity of air traffic systems, exacerbated by unpredictable weather events, demands methods that can simultaneously capture global temporal patterns and localized disruptions; existing approaches often struggle to balance these requirements efficiently. The proposed method employs two parallel branches: an SSM branch for continuous-time recurrent modeling of long-range dependencies with linear complexity, and a dynamic graph attention branch that adaptively computes node-pair weights while incorporating weather severity features through sparsification strategies for scalability. These branches are fused via a data-dependent gating mechanism, enabling the model to dynamically prioritize either global temporal dynamics or localized spatial interactions based on input conditions. Moreover, the architecture leverages memory-efficient attention computation and HiPPO initialization to ensure stable training and inference. Experiments on real-world air traffic datasets demonstrate that State-DynAttn outperforms existing baselines in prediction accuracy and robustness, particularly under severe weather scenarios. The framework’s ability to handle both gradual traffic evolution and abrupt disruption-induced changes makes it suitable for real-world deployment in air traffic management systems. Furthermore, the design principles of State-DynAttn can be extended to other spatiotemporal prediction tasks where long-range dependencies and dynamic relational structures coexist. This work contributes a principled approach to hybridizing state-space models with graph-based attention, offering insights into the trade-offs between computational efficiency and modeling flexibility in complex dynamical systems. Full article

Reference-less evaluation of machine translation, or Quality Estimation (QE), is vital for low-resource language pairs where high-quality references are often unavailable. In this study, we investigate segment-level QE methods comparing encoder-based models such as MonoTransQuest, CometKiwi, and xCOMET with various decoder-based methods (Tower+, ALOPE, and other instruction-fine-tuned language models). Our work primarily focused on utilizing eight low-resource language pairs, involving both English on the source side and the target side of the translation. Results indicate that while fine-tuned encoder-based models remain strong performers across most low-resource language pairs, decoder-based Large Language Models (LLMs) show clear improvements when adapted through instruction tuning. Importantly, the ALOPE framework further enhances LLM performance beyond standard fine-tuning, demonstrating its effectiveness in narrowing the gap with encoder-based approaches and highlighting its potential as a viable strategy for low-resource QE. In addition, our experiments demonstrates that with adaptation techniques such as LoRA (Low Rank Adapters) and quantization, decoder-based QE models can be trained with competitive GPU memory efficiency, though they generally require substantially more disk space than encoder-based models. Our findings highlight the effectiveness of encoder-based models for low-resource QE and suggest that advances in cross-lingual modeling will be key to improving LLM-based QE in the future. Full article

Background: Football performance depends on the integration of physical, technical, and cognitive abilities under constantly changing conditions. In this context, dual-task training combining physical and cognitive demands has emerged as a promising approach to enhance decision-making and game intelligence in youth football players. Objective: The aim of this study was to determine the effects of an eight-week dual-task training programme on physical (speed, strength, and agility), cognitive (working memory, planning, processing speed, and response time), technical (dribbling and short passing), and dual-task performance in U16 football players. Methods: Thirty-two players (age: 14.88 ± 0.65 years; BMI: 20.98 ± 1.79 kg/m²) were randomly assigned to a control group (n = 14) and an experimental group (n = 18). The experimental group completed a dual cognitive–motor training (CMT) programme consisting of 24 sessions (3 sessions/week, 10–15 min each), integrated into regular football practice. Pre-intervention and post-intervention assessments included football skills (dribbling and passing tests), cognitive tests (Wom-Rest and Vismem-Plan), physical tests (countermovement jump, 20 m sprint, and 505 change-of-direction), and a dual-task test (soccer skills and cognitive aptitude test). Results: The experimental group showed significant improvements in all assessed variables, while the control group exhibited no changes or declines in performance. The most notable effects were observed in SoSCAT with visual interference, dual-task cost, and 505 change-of-direction. Conclusions: The findings suggest that integrating brief dual CMT programmes into regular football practice can simultaneously enhance physical, technical, and cognitive performance in youth players. This evidence supports the implementation of dual CMT as an effective and time-efficient tool in talent development programmes. Full article

Tennis serves heavily impact match outcomes, yet analysis by coaches is limited by human vision. The design of an automated tennis serve analysis system could facilitate enhanced performance analysis. As serve location and serve success are directly correlated, predicting the outcome of a serve could provide vital information for performance analysis. This article proposes a tennis serve analysis system powered by Machine Learning, which classifies the outcome of serves as “in”, “out” or “net”, and predicts the coordinate outcome of successful serves. Additionally, this work details the collection of three-dimensional spatio-temporal data on tennis serves, using marker-based optoelectronic motion capture. The classification uses a Stacked Bidirectional Long Short-Term Memory architecture, whilst a 3D Convolutional Neural Network architecture is harnessed for serve coordinate prediction. The proposed method achieves 89% accuracy for tennis serve classification, outperforming the current state-of-the-art whilst performing finer-grain classification. The results achieve an accuracy of 63% in predicting the serve coordinates, with a mean absolute error of 0.59 and a root mean squared error of 0.68, exceeding the current state-of-the-art with a new method. The system contributes towards the long-term goal of designing a non-invasive tennis serve analysis system that functions in training and match conditions. Full article

Accurately predicting the reception area of the Automatic Identification System (AIS) is critical for ship tracking and anomaly detection, as errors in signal interpretation may lead to incorrect vessel localization and behavior analysis. However, traditional propagation models, whether they are deterministic, empirical, or semi-empirical, face limitations when applied to dynamic environments due to their reliance on detailed atmospheric and terrain inputs. Therefore, to address these challenges, we propose a data-driven approach based on graph neural networks (GNNs) to model AIS reception as a function of environmental and geographic variables. Specifically, inspired by attention mechanisms that power transformers in large language models, our framework employs the SAmple and aggreGatE (GraphSAGE) framework convolutions to aggregate neighborhood features, then combines layer outputs through Jumping Knowledge (JK) with Bidirectional Long Short-Term Memory (BiLSTM)-derived attention coefficients and integrates an attentional pooling module at the graph-level readout. Moreover, trained on real-world AIS data enriched with terrain and meteorological features, the model captures both local and long-range reception patterns. As a result, it outperforms classical baselines—including ITU-R P.2001 and XGBoost in F1-score and accuracy. Ultimately, this work illustrates the value of deep learning and AIS sensor networks for the detection of positioning anomalies in ship tracking and highlights the potential of data-driven approaches in modeling sensor reception. Full article

Lower limb functional degeneration has become prevalent, notably reducing the core strength that drives motor control. Squats are frequently used in lower limb training, improving overall muscle strength. However, performing continuously with improper techniques can lead to dynamic knee instability. It worsens with little to no motivation to perform these power training motions. Hence, it is crucial to have a gaming-based exercise tracking system to adaptively enhance the user experience without causing injury or falls. In this work, 28 healthy subjects performed exergame-based squat training, and dynamic kinematic features were recorded. The five features acquired from a depth camera-based inertial measurement unit (IMU) (1—Knee shakiness, 2—Knee distance, and 3—Squat depth) and an Anima forceplate sensor (4—Sway velocity and 5—Sway area) were assessed using a Spearman correlation coefficient-based feature selection method. An input vector that defines knee instability is used to train and test the Long Short-Term Memory (LSTM) and Support Vector Machine (SVM) models for binary classification. The results showed that knee instability events can be successfully classified and achieved a high accuracy of 96% in both models with sets 1, 2, 3, 4, and 5 and 1, 2, and 3. The feature selection results indicate that the LSTM network with the proposed combination of input features from multimodal sensors can successfully perform real-time tracking of knee instability. Furthermore, the findings demonstrate that this multimodal approach yields improved classifier performance with enhanced accuracy compared to using features from a single modality during lower limb therapy. Full article

The increasing adoption of electric vehicles (EVs) within intelligent transportation systems (ITSs) has elevated the importance of cybersecurity, especially with the rise in Vehicle-to-Everything (V2X) communications. Traditional intrusion detection systems (IDSs) struggle to address the evolving and complex nature of cyberattacks in such dynamic environments. To address these challenges, this study introduces a novel deep learning-based IDS designed specifically for EV communication networks. We present a hybrid model that integrates convolutional neural networks (CNNs), long short-term memory (LSTM) layers, and adaptive learning strategies. The model was trained and validated using the VeReMi dataset, which simulates a wide range of attack scenarios in V2X networks. Additionally, an ablation study was conducted to isolate the contribution of each of its modules. The model demonstrated strong performance with 98.73% accuracy, 97.88% precision, 98.91% sensitivity, and 98.55% specificity, as well as an F1-score of 98.39%, an MCC of 0.964, a false-positive rate of 1.45%, and a false-negative rate of 1.09%, with a detection latency of 28 ms and an AUC-ROC of 0.994. Specifically, this work fills a clear gap in the existing V2X intrusion detection literature—namely, the lack of scalable, adaptive, and low-latency IDS solutions for hardware-constrained EV platforms—by proposing a hybrid CNN–LSTM architecture coupled with an elastic weight consolidation (EWC)-based adaptive learning module that enables online updates without full retraining. The proposed model provides a real-time, adaptive, and high-precision IDS for EV networks, supporting safer and more resilient ITS infrastructures. Full article

Early and accurate diagnosis of dementia is essential to improving patient outcomes and reducing societal burden. The Mini-Mental State Examination (MMSE) is widely used to assess cognitive function, yet traditional statistical and machine learning approaches often face limitations in capturing nonlinear interactions and subtle decline patterns. This study developed a novel deep learning-based dementia prediction model using MMSE data collected from domestic clinical settings and compared its performance with traditional machine learning models. A notable strength of this work lies in its use of item-level MMSE features combined with explainable AI (SHAP analysis), enabling both high predictive accuracy and clinical interpretability—an advancement over prior approaches that primarily relied on total scores or linear modeling. Data from 164 participants, classified into cognitively normal, mild cognitive impairment (MCI), and dementia groups, were analyzed. Individual MMSE items and total scores were used as input features, and the dataset was divided into training and validation sets (8:2 split). A fully connected neural network with regularization techniques was constructed and evaluated alongside Random Forest and support vector machine (SVM) classifiers. Model performance was assessed using accuracy, F1-score, confusion matrices, and receiver operating characteristic (ROC) curves. The deep learning model achieved the highest performance (accuracy 0.90, F1-score 0.90), surpassing Random Forest (0.86) and SVM (0.82). SHAP analysis identified Q11 (immediate memory), Q12 (calculation), and Q17 (drawing shapes) as the most influential variables, aligning with clinical diagnostic practices. These findings suggest that deep learning not only enhances predictive accuracy but also offers interpretable insights aligned with clinical reasoning, underscoring its potential utility as a reliable tool for early dementia diagnosis. However, the study is limited by the use of data from a single clinical site with a relatively small sample size, which may restrict generalizability. Future research should validate the model using larger, multi-institutional, and multimodal datasets to strengthen clinical applicability and robustness. Full article

Cognitive deficiencies are significant factors affecting aviation piloting capabilities. However, due to the limited stability resulting from the insufficient appeal of traditional attention or memory cognitive training, multi-task gamified cognitive training (MTGCT) may be more beneficial in generating far transfer effects in task performance. This study explores the enhancement effects of simulated flight operation capabilities based on visuo-spatial attention and working memory MTGCT. Additionally, we explore the neurophysiological impacts through changes in EEG power spectral density (PSD) characteristics and brain asymmetry, and whether these impacts exhibit a certain retention effect. This study designed a 28-day simulated flight operation capability enhancement experiment. In addition, the behavioral performance and EEG signal changes in 28 college students (divided into control and training groups) were analyzed. The results indicated that MTGCT significantly enhanced simulated flight operational capabilities, and the neural framework formed by physiological changes remains effective for at least two weeks. The physiological changes included a decrease in the $\theta$ band PSD and an increase in the $\alpha$ band PSD in the frontal and parietal lobes due to optimized cognitive resource allocation, as well as the frontal $\theta$ band leftward asymmetry and the frontoparietal $\alpha$ band rightward asymmetry due to the formation of neural activity patterns. These findings support, to some extent, the feasibility and effectiveness of using MTGCT as a periodic training method to enhance the operational and cognitive abilities of aviation personnel. Full article

Early train detection is vital for ensuring the safety of railway personnel, particularly in remote locations where fixed signaling infrastructure is unavailable. Unlike many existing solutions that rely on high-power, high-cost sensors and compute platforms, this work presents a lightweight, low-cost, and portable framework designed to run entirely on resource-constrained microcontrollers with just kilobytes of Random Access Memory (RAM). The proposed system uses vibration data from low-cost accelerometers and employs a simple yet effective Linear Regression (LR) model for almost real-time prediction of train arrival times. To ensure feasibility on low-end hardware, a parallel-processing framework is introduced, enabling continuous data collection, Machine Learning (ML) inference, and wireless communication with strict timing and energy constraints. The decision-making process, including data preprocessing and ML prediction, completes in under 10 ms, and alerts are transmitted via LoRa, enabling kilometer-range communication. Field tests on active railway lines confirm that the system detects approaching trains 15 s in advance with no false negatives and a small number of explainable false positives. Power characterization demonstrates that the system can operate for more than 6 days on a 10 Ah battery, with potential for months of operation using wake-on-vibration modes. Full article

Aviation safety remains a critical area of research, requiring accurate and efficient classification of incident reports to enhance risk assessment and accident prevention strategies. This study evaluates the performance of three deep learning models, BERT, Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) for classifying incidents based on injury severity levels: Nil, Minor, Serious, and Fatal. The dataset, drawn from ATSB records covering the years 2013 to 2023, consists of 53,273 records and was used. The models were trained using a standardized preprocessing pipeline, with hyperparameter tuning to optimize performance. Model performance was evaluated using metrics such as F1-score accuracy, recall, and precision. Results revealed that BERT outperformed both LSTM and CNN across all metrics, achieving near-perfect scores (1.00) for precision, recall, F1-score, and accuracy in all classes. In comparison, LSTM achieved an accuracy of 99.01%, with strong performance in the “Nil” class, but less favorable results for the “Minor” class. CNN, with an accuracy of 98.99%, excelled in the “Fatal” and “Serious” classes, though it showed moderate performance in the “Minor” class. BERT’s flawless performance highlights the strengths of transformer architecture in processing sophisticated text classification problems. These findings underscore the strengths and limitations of traditional deep learning models versus transformer-based approaches, providing valuable insights for future research in aviation safety analysis. Future work will explore integrating ensemble methods, domain-specific embeddings, and model interpretability to further improve classification performance and transparency in aviation safety prediction. Full article

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that significantly affects motor functions, including speech production. Voice analysis offers a less invasive, faster and more cost-effective approach for diagnosing and monitoring PD over time. This research introduces an automated system to distinguish between PD and non-PD individuals based on speech signals using state-of-the-art signal processing and machine learning (ML) methods. A publicly available voice dataset (Dataset 1, 81 samples) containing speech recordings from PD patients and non-PD individuals was used for model training and evaluation. Additionally, a small supplementary dataset (Dataset 2, 15 samples) was created although excluded from experiment, to illustrate potential future extensions of this work. Features such as Mel-frequency cepstral coefficients (MFCCs), spectrograms, Mel spectrograms and waveform representations were extracted to capture key vocal impairments related to PD, including diminished vocal range, weak harmonics, elevated spectral entropy and impaired formant structures. These extracted features were used to train and evaluate several ML models, including support vector machine (SVM), XGBoost and logistic regression, as well as deep learning (DL)architectures such as deep neural networks (DNN), convolutional neural networks (CNN) combined with long short-term memory (LSTM), CNN + gated recurrent unit (GRU) and bidirectional LSTM (BiLSTM). Experimental results show that DL models, particularly BiLSTM, outperform traditional ML models, achieving 97% accuracy and an AUC of 0.95. The comprehensive feature extraction from both datasets enabled robust classification of PD and non-PD speech signals. These findings highlight the potential of integrating acoustic features with DL methods for early diagnosis and monitoring of Parkinson’s Disease. Full article

This work presents the development of an intelligent system designed to support the predictive maintenance of the Colombian Navy’s maritime vessels through the estimation of remaining useful life and unsupervised anomaly detection, within the framework of the project called “Colombian Integrated Platform Supervision and Control System” (SISCP-C). This project seeks to guarantee the sustainability of the vessels over time, increase their operational availability, and optimize their life cycle cost, in accordance with the institution’s strategic direction established in the Naval Development Plan 2042. The system provides useful information to the crew, enabling informed decision-making for intelligent and efficient maintenance strategies. To address the limited availability of normal operating data, synthetic data generation techniques with seeding are implemented, including tabular variational autoencoders, conditional tabular generative adversarial networks, and Gaussian copulas. Among these, tabular variational autoencoders achieved the best performance and are used to generate synthetic datasets under normal conditions for the Wärtsilä 6L26 diesel engine (manufactured by Wärtsilä Italia S.p.A., Trieste, Italy). These datasets are used to train several unsupervised anomaly detection models, including one-class support vector machines, classical autoencoders, and long short-term memory-based autoencoders. The long short-term memory autoencoders outperformed the others in terms of detection metrics. Dedicated multivariate long short-term memory autoencoders are subsequently trained for each engine subsystem. By calculating the mean absolute error of the reconstructions, a subsystem-specific health index is computed, which is used to estimate the remaining useful life. Full article

Background/Objectives: This study evaluated the feasibility of predicting bone conduction (BC) thresholds and classifying air–bone gap (ABG) status using only air conduction (AC) data obtained from pure tone audiometry (PTA). Methods: A total of 60,718 PTA records from five tertiary hospitals in the Republic of Korea were utilized. Input features included AC thresholds (0.25–8 kHz), age, and sex, while outputs were BC thresholds (0.25–4 kHz) and ABG classification based on 10 dB and 15 dB criteria. Five machine learning models—deep neural network (DNN), long short-term memory (LSTM), bidirectional LSTM (BiLSTM), random forest (RF), and extreme gradient boosting (XGB)—were trained using 5-fold cross-validation with Synthetic Minority Over-sampling Technique (SMOTE). Model performance was evaluated based on accuracy, sensitivity, precision, and F1 score under ±5 dB and ±10 dB thresholds for BC prediction. Results: LSTM and BiLSTM outperformed DNN in predicting BC thresholds, achieving ~60% accuracy within ±5 dB and ~80% within ±10 dB. For ABG classification, all models performed better with the 10 dB criterion than the 15 dB. Tree-based models (RF, XGB) achieved the highest classification accuracy (up to 0.512) and precision (up to 0.827). Confidence intervals for all metrics were within ±0.01, indicating stable results. Conclusions: AI models can accurately predict BC thresholds and ABG status using AC data alone. These findings support the integration of AI-driven tools into clinical audiology and telemedicine, particularly for remote screening and diagnosis. Future work should focus on clinical validation and implementation to expand accessibility in hearing care. Full article