Search Results (1,395)

Wind turbine gearbox condition monitoring increasingly combines Supervisory Control and Data Acquisition (SCADA) data with Explainable Artificial Intelligence (XAI) for predictive maintenance. However, current XAI applications report attributions as static or globally aggregated feature-importance results. Such representations do not reveal when fault-related variables emerge, how dominance shifts between features, or how the explanatory structure evolves as degradation progresses. This limits their value for time-resolved diagnostic interpretation. To address this gap, this study proposes the Temporal Attribution Matrix (TAM), a temporal interpretability framework that tracks the evolution of XAI-derived feature importance across degradation periods. The central hypothesis is that temporal attribution patterns contain diagnostic information not captured by static feature-importance summaries. TAM was applied to a three-year SCADA dataset from Fuhrländer FL2500 wind turbines using XGBoost-SHAP and 1D-CNN Grad-CAM within sliding weekly windows. Four temporal measures were derived: feature onset time, dominance transition, attribution entropy, and cross-model consistency. Both XAI methods independently identified gearbox bearing temperatures 451 and 152 as the most influential features. TAM further revealed a synchronized thermal-feature onset on 23 October 2012, 14 SHAP dominance transitions compared with 70 Grad-CAM transitions, and a moderate cross-model Spearman correlation of 0.488. Secondary validation using WT82 confirmed TAM’s applicability beyond a single turbine. These results demonstrate that TAM extends static XAI by producing time-resolved degradation narratives for SCADA-based wind turbine predictive maintenance. Full article

Potato production has become a major global food crop, generating high profitability while reducing the pressure on the world’s food supply chain. Therefore, potato tuber disease inspection for quality control and food security is of utmost importance. With the emergence of Artificial Intelligence (AI), Deep Convolutional Neural Networks (DCNNs) have become the most widely employed approach for this task. However, DCNNs mainly require large annotated datasets and often generalize poorly to unseen conditions or potato tuber diseases. In this paper, Potato-Free is introduced as a training-free framework for potato disease detection and localization. In detail, the framework extracts potato tuber masks, estimates the average color of healthy tubers, and generates colored potato masks. These masks are compared with the potato tuber images to produce MSE maps, from which optimal thresholds for potato tuber disease detection and localization are determined. Experimental results show that the proposed framework achieves a macro F1-Score of 63.20% in a cross-dataset evaluation on completely unseen potato tuber diseases, providing comparable performance to State-of-the-Art DCNN classification models (macro F1-Score of 63.40%) while also outperforming a reconstruction-based baseline (macro F1-Score of 42.92%), without requiring training. An extensibility study also illustrates that Potato-Free can separate different disease types with a macro F1-Score close to the best-performing DCNN model (5.02% difference in macro F1-Score) despite being a threshold-based framework that does not require training. Full article

The rapid growth of sophisticated malware, including polymorphic, metamorphic, and zero-day threats, has made traditional signature-based and heuristic detection methods increasingly insufficient in modern desktop computing environments. As cyber threats continue to evolve in both complexity and scale, the demand for intelligent and adaptive malware detection mechanisms capable of identifying previously unseen attacks has become more critical than ever. In this study, we propose a novel deep learning framework for three-dimensional malware image classification that utilizes visual representation learning to improve malware detection performance. The proposed framework converts raw malware binaries into three-dimensional grayscale and RGB image representations, allowing hidden structural and spatial patterns within malware samples to be analyzed more effectively. By transforming malware data into multi-dimensional visual forms, the proposed system facilitates the process of automatically learning hierarchical features by CNN through multi-dimensional visualization of malware binary codes. In addition, an optimization technique using Genetic Algorithms is implemented within this architecture to improve classification performance and stability. The proposed evolutionary algorithm performs an effective search process within the large parameter space of 3D-CNN, leading to the identification of models that facilitate learning. It is shown that multi-dimensional visualization of malware achieves improved classification performance. It can be concluded that the combination of three-dimensional malware visualization, deep learning, and genetic optimization is promising for the development of future intelligent malware detection tools. Full article

Background/Objectives: Reliable detection of epileptic seizures using electroencephalography (EEG) is crucial for clinical diagnosis and for alleviating clinicians’ workload. However, existing studies still make insufficient use of phase information, and the synergy between local time–frequency pattern extraction and global dependency modeling remains limited. Methods: We propose a seizure detection framework based on the continuous wavelet transform (CWT), a three-dimensional convolutional neural network (3D-CNN), and a vision transformer (ViT). First, multichannel EEG segments are preprocessed, after which CWT is used to generate power spectrograms and phase spectrograms. These representations are then fused along the depth dimension into a unified power-phase volume and fed into a hybrid network composed of a 3D-CNN feature extractor and a single-layer ViT encoder to jointly learn local time–frequency–channel coupling patterns and higher-level global dependencies. Finally, seizure detection is completed by combining moving-average filtering, thresholding, and collar correction. Results: On the public CHB-MIT dataset and the clinical SH-SDU dataset, the proposed method achieved average segment-level sensitivities of 98.68% and 92.05%, specificities of 98.33% and 97.53%, accuracies of 98.49% and 96.37%, and AUC values of 97.26% and 92.89%, respectively. In event-level evaluation, the average sensitivities were 99.13% and 96.08%, with false detection rates of 0.88/h and 0.69/h, respectively. Further multi-stage ablation experiments together with t-SNE and Grad-CAM visualizations provided qualitative and experimental support for the design rationale of the joint power-phase input and the hybrid 3D-CNN-ViT architecture. Conclusions: The proposed framework effectively exploits the complementary discriminative value of power and phase information in epileptic EEG and demonstrates strong detection performance under patient-specific evaluation on both public and clinically collected datasets. Full article

Early detection of Alzheimer’s disease (AD) and mild cognitive impairment (MCI) remains critically important, yet conventional neuroimaging and biomarker-based approaches are costly, invasive, and poorly scalable for population screening. Speech offers a non-invasive, cost-effective alternative cognitive biomarker, but existing systems rarely integrate its multiple linguistic dimensions. We present Multi-Spec MCI-Net, a multimodal framework for HC/MCI classification that jointly models three complementary speech representations: token-level semantics via dVAE and BERT operating on Mel spectrograms; temporal prosodic dynamics via a 1D-CNN with attention; and discourse-level semantic coherence via a graph convolutional network. A gated fusion mechanism adaptively weights these modalities, yielding clinically interpretable predictions tailored to individual phenotypic profiles. Evaluated on the Chinese NCMMSC2021_AD challenge dataset and the DementiaBank Mandarin subset, the model achieves 89.29% accuracy and 0.9584 ROC AUC on NCMMSC2021_AD, with 92.31% MCI recall—critical for minimizing false negatives in screening contexts. Evaluation on the combined NCMMSC2021_AD and DementiaBank Mandarin dataset attains 77.46% accuracy and 0.8280 AUC, demonstrating robustness across spontaneous dialog and picture description tasks. Ablation studies confirm that multimodal fusion outperforms the semantic-only baseline by 5.16 percentage points, with each branch contributing non-redundant diagnostic information. These results establish an effective, interpretable approach for scalable, speech-based early MCI screening. Full article

This paper proposes a novel fault-identification method for micro-coils in relays with forcibly guided contacts, a type of electromechanical elementary component (EMEC), combining a composite analytical model, a 2D thermal model, and a 1D-CNN. A low-order thermal circuit with one central node and four boundary nodes is established, while a two-dimensional anisotropic Poisson equation is used as a high-order calibration model. The two models are coupled through iterative correction of reusable thermal resistances. For thermal aging, enamel-film delamination, and inter-turn short-circuit faults, thermal-conductivity attenuation, asymmetric branch-resistance perturbation, and localized abnormal heat-source injection are introduced to generate physically constrained temperature sequences. Orthogonal centerline temperature distributions are extracted as one-dimensional feature vectors for 1D-CNN classification. Simulation results show that the hybrid model has an error of approximately 1.7% compared with finite-element results, and the trained 1D-CNN achieves 98.13% accuracy on 160 test samples. Experimental reconstruction and deep-feature visualization further verify its ability to distinguish normal, aging, delamination, and local short-circuit states. Full article

Recidivism prediction is increasingly embedded in criminal justice decision-making, yet most deployed systems remain opaque and have been shown to exhibit discriminatory behavior against certain demographic groups. This paper presents a fairness-aware interpretable framework for recidivism prediction applied to three real-world datasets from Bulgaria, Greece, and Portugal. The classification core relies on a 1-Dimensional Convolutional Neural Network (1D-CNN), trained by a custom objective function that embeds the Equalized Odds fairness criterion as an L1-regularized penalty reflecting on gender-based disparities in false positive and false negative rates. Model-level interpretability is provided through Kernel SHAP, which decomposes individual predictions into additive feature attributions grounded in cooperative game theory. Experiments across prediction tasks, each evaluated over randomized runs, demonstrate that the baseline model exhibits statistically significant bias against the female group in all datasets. The fairness-constrained model substantially reduces these disparities across all tasks at a moderate and expected cost to classification accuracy. Kernel SHAP analysis reveals the relative contribution of static and dynamic offenders’ attributes to individual risk scores, supporting auditability and contestability. The proposed framework advances the integration of predictive performance, algorithmic fairness, and structural interpretability in criminal justice analytics. Full article

Sleep-related respiratory disturbances are difficult to monitor continuously outside specialized laboratories because conventional polysomnography is resource-intensive and intrusive. This study presents a contactless edge-AI engineering prototype for detecting controlled voluntary respiratory-motion suppression and motion artifacts using a 60 GHz frequency-modulated continuous-wave radar. The system integrates a 60 GHz radar front end, lightweight local preprocessing, an INT8 one-dimensional convolutional neural network deployed on the Analog Devices MAX78000 CNN accelerator (Analog Devices Thailand, Chon Buri, Thailand), and an event-driven Raspberry Pi Zero 2W gateway for alert transmission. Evaluation was performed using a controlled healthy-volunteer dataset consisting of normal breathing, voluntary breath-holding-induced respiratory suppression, and deliberate motion artifact. The final valid test set contained 270 technically valid 30 s windows balanced across the three classes. The INT8 model achieved an overall accuracy of 92.6% (95% confidence interval: 88.8–95.2%), with a macro-averaged precision, recall, and F1-score of 92.6%, 92.6%, and 92.5%, respectively. Active CNN inference on the MAX78000 consumed 0.152 ± 0.011 mJ and was completed in 5.20 ± 0.11 ms, corresponding to approximately 280-fold lower active inference energy than Python 3.14.6/TensorFlow Lite 2.21.0-based execution on the Raspberry Pi Zero 2W. These results demonstrate the feasibility of privacy-aware, low-power respiratory-pattern classification at the edge. However, the study should be interpreted strictly as an engineering proof-of-concept based on controlled voluntary breathing and movement tasks in healthy volunteers. It is not a clinically validated apnea or obstructive sleep apnea detection system and did not include polysomnography, oxygen saturation measurement, airflow sensing, sleep staging, or diagnosed patient cohorts. Full article

This study demonstrates single-channel fiber Bragg grating (FBG) sensing for relative vibration-state monitoring of a motor–support system under angle-dependent boundary conditions. A packaged FBG accelerometer-type sensing unit was mounted on the motor–support structure, and the reflected Bragg wavelength was recorded as a one-dimensional optical vibration response. Because the sensor was installed away from the rotating shaft, the measured wavelength fluctuation was interpreted as a coupled vibration-sensitive response of the motor, fixture, sensor package, bonding condition, and changing boundary state, rather than as a calibrated shaft speed or absolute acceleration signal. Adaptive variational mode decomposition (AVMD) was applied to track the time-varying narrowband spectral-response trajectory of the Bragg-wavelength signal. In parallel, raw wavelength windows were supplied to LSTM, 1D-CNN, and CNN–LSTM autoencoders to evaluate waveform departures from learned nominal fixed-angle behavior. The fixed-angle results showed stable but distinguishable optical vibration responses under different boundary states, whereas the dynamic angle-transition records produced local trajectory changes and alarm-candidate intervals. Baseline and autoencoder comparisons further clarified the trade-off between transition coverage and false-alarm tendency. The RMS threshold baseline was more sensitive to transition-related amplitude changes but produced more false alarms, whereas the CNN–LSTM autoencoder provided the most selective response among the tested autoencoder branches. The results are interpreted as task-specific evidence for relative vibration-state transition monitoring rather than as general motor fault diagnosis. Overall, the framework demonstrates a compact FBG-based route for relative vibration-state transition monitoring when speed references, dense sensor layouts, and labeled fault data are unavailable. Full article

Non-Intrusive Load Monitoring (NILM) enables estimation of the energy use of individual appliances in smart buildings from a single aggregate meter. In practice, however, this task is not straightforward. Signals from different appliances can overlap, and the measured data may also include distortions such as spikes, drops, and noise. To address these issues, this study presents a multi-task triple-hybrid deep learning framework that handles appliance classification and anomaly detection together. The model brings together 1D-CNN, BiLSTM, and Transformer Attention so that local patterns, temporal dependencies, and wider contextual information can be learned within the same structure. It also uses a dual-output design to classify appliance categories and detect anomaly types simultaneously. Experiments were carried out on Building 1 of the UK-DALE dataset with four appliances: kettle, microwave, washer dryer, and fridge freezer. For the anomaly task, synthetic disturbances were added to segmented signal windows and grouped as normal, spike, drop, and noise. To check how well the proposed framework handled different scenarios, it was tested on both the UK-DALE and REDD datasets. Looking at the main UK-DALE results, the model correctly identified appliances 99.48% of the time and spotted anomalies with 98.80% accuracy. A secondary test on the REDD dataset yielded an 86.44% classification score. This proves the architecture can adjust to completely new power grid environments without losing its edge. On top of that, when pitted against standard benchmark models like Seq2Point, this triple-hybrid design clearly does a better job of mapping out complex signal changes. As a result, it yields much stronger anomaly detection metrics. Full article

Accurate waveform classification in noisy environments is an important task in modern communications, radar signal analysis, biomedical signal interpretation, industrial monitoring and other signal processing systems. This paper investigates the performance of three neural network architectures: Multilayer Perceptron (MLP), one-dimensional Convolutional Neural Network (1D-CNN), and Long Short-Term Memory (LSTM) for multiclass waveform classification in the presence of Additive White Gaussian Noise (AWGN). A time series dataset consisting of multiple waveform classes is generated and corrupted with AWGN across a wide range of signal-to-noise ratio (SNR) levels to simulate noisy signal distortion conditions. The three models are trained and evaluated under identical conditions to ensure a fair comparison. Their classification performance is evaluated in terms of accuracy, Confusion Matrix (CM), Receiver Operating Characteristic (ROC) curve and the Area Under the ROC curve (AUC) across varying SNR values. Simulation results demonstrate that the 1D-CNN effectively captures local temporal patterns and achieves superior robustness in classification at moderate and high SNR levels. The LSTM model demonstrates the ability to capture temporal dependencies in sequential waveform data but exhibits sensitivity to waveform variations due to amplitude, phase and frequency changes and noise at lower SNR values. The MLP, although computationally simpler, shows comparatively limited performance in low-SNR conditions due to its lack of temporal feature extraction capability. For the case of multiclass deterministic waveforms, the accuracy of classification for the 1D-CNN and LSTM is nearly 100% at SNR = 5 dB showing their robustness in classification, whereas the accuracy of MLP is approximately 70% that shows poor classification in noisy conditions. When there is random amplitude, frequency and phase variations in the waveforms, the accuracy of the 1D-CNN and MLP increases with SNR, and 1D-CNN superior to MLP. However, the LSTM accuracy fails to improve with SNR, resulting in poor classification performance in such a scenario. The results provide an insight into the suitability of different neural architectures for waveform classification tasks in noisy communication or other time series applications and highlight the advantages of convolutional feature extraction for robust signal recognition. Full article

Background/Objectives: Spontaneous intracranial hypotension (SIH) is caused by spinal cerebrospinal fluid (CSF) leakage and is typically diagnosed by clinical presentation and characteristic MRI signs; however, objective tools for monitoring physiological changes and treatment response remain limited. Cine phase-contrast MRI (PC-MRI) enables noninvasive quantification of aqueductal CSF dynamics, yet reliable analysis is challenging since the cerebral aqueduct is extremely small and susceptible to low contrast, partial volume effects, and ROI-dependent measurement variability—particularly in SIH where CSF pulsatility is often reduced. Methods: We propose an end-to-end automated framework that integrates (1) a cascade localization–segmentation strategy, consisting of Tiny YOLOv4 detection followed by MultiResUNet segmentation on a YOLOv4-derived cropped ROI; (2) physiology-informed pulsatility-based segmentation (PUBS) to refine anatomical masks into functional flow ROIs; and (3) one-dimensional convolutional neural networks (1D-CNNs) to extract exploratory waveform morphology features from 32-phase cardiac-cycle velocity waveforms. The study includes 39 participants, yielding 59 cine PC-MRI examinations: 11 controls, 28 Pre-treatment SIH scans and 20 Post-treatment Recovery scans. Results: The cascade model significantly improves segmentation robustness compared with a full-image baseline, achieving higher Dice scores and markedly lower boundary errors across cohorts (e.g., Pre-treatment SIH HD95: 1.66 ± 0.74 px vs. 15.37 ± 44.98 px). PUBS refinement reduces quantification deviation from expert manual references in SIH (mean relative error: 7.4% to 5.6%) and improves diagnostic performance for multiple hemodynamic parameters (e.g., downward mean flow AUC: 0.747 to 0.792). For waveform morphology analysis, the end-to-end 1D-CNN classifier was evaluated using repeated-seed participant-level grouped LOOCV. The repeated-seed ensemble prediction showed modest out-of-sample discrimination between Normal controls and Pre-treatment SIH scans, with an AUC of 0.646, a bootstrap 95% confidence interval of 0.455–0.826, and a permutation-test p-value of 0.072. Separately, exploratory analysis of the final baseline-trained 1D-CNN latent space showed marked, apparent Normal-versus-SIH separability and an intermediate recovery distribution in PCA space, suggesting that aqueductal waveform morphology may encode SIH-related physiological information. Conclusions: These findings suggest that SIH-related information may be reflected not only in flow magnitude but also in aqueductal CSF waveform morphology. However, the modest and statistically non-significant out-of-sample performance of the end-to-end 1D-CNN classifier indicates that morphology-based AI features should currently be regarded as exploratory biomarker candidates rather than validated stand-alone diagnostic tools. Larger independent cohorts are required to confirm their reproducibility, physiological meaning, and clinical utility. Full article

Driving behavior recognition plays a crucial role in intelligent driving systems and road traffic safety. Due to insufficient driving experience and limited ability to allocate visual attention, novice drivers are considered a high-risk group for traffic accidents. Existing approaches primarily focus on experienced drivers and rely on single-modal eye-tracking data, making it difficult to model spatial attention distributions and long-term temporal dependencies simultaneously. Moreover, these methods are often affected by modality asynchrony during multimodal fusion, further limiting performance gains. To address these challenges, this study proposes a novice driver behavior recognition method based on bimodal eye-tracking features and a gated cross-modal attention fusion (GCMAF) mechanism. The model adopts a spatial–temporal dual-branch architecture. The spatial branch employs ResNet34 to extract eye-tracking heatmap features to represent the visual attention distribution. In contrast, the temporal branch integrates a 1D-CNN with the Mamba model to capture local dynamic patterns and long-range temporal dependencies. In the fusion stage, the GCMAF module is introduced to enhance cross-modal interactions, and a gating mechanism is further used to adaptively adjust modality weights, thereby mitigating the adverse effects of modality asynchrony. To validate the effectiveness and generalization ability of the proposed method, repeated experiments and five-fold cross-validation are conducted. The results demonstrate that the model achieves an average classification accuracy of 93.86% across four driving behavior categories, with standard deviations below 0.3%. Compared with baseline methods, paired t-test results show that the performance improvement is statistically significant (p < 0.01). Ablation studies further confirm the independent contribution of each component. Overall, the proposed method outperforms existing approaches in terms of accuracy and stability, providing effective support for driving behavior assessment and proactive safety warning systems. Full article

Intrusion detection serves as a core technology for securing heterogeneous edge networks, including IoT, industrial edges, and 5G networks. However, existing federated learning-based intrusion detection systems suffer from environmental heterogeneity, limited sample availability, and severe class imbalance—issues that result in inefficient resource allocation and compromised detection performance against rare attacks. In this paper, we propose a novel lightweight intrusion detection model for heterogeneous edge networks, named FedNIDS-CNN, which is based on dynamic distillation-aided federated learning with a CNN backbone. In the data preprocessing phase, a two-level class balancing strategy integrating nearest-neighbor interpolation augmentation and adaptive synthetic sampling is employed to ensure distortion-free sample synthesis. For feature and model optimization, principal component analysis (PCA) is used to reduce the dimensionality of traffic features, while a lightweight 1D-CNN is adopted as the base model to alleviate computational overhead on edge devices. During federated training and knowledge aggregation, a dynamic weight distillation loss mechanism is designed to enhance the model’s ability to recognize minority-class attacks. Meanwhile, the federated framework supports client-side local training and server-side weighted soft-label aggregation, enabling effective knowledge fusion across heterogeneous models. Experimental results on the CICIDS2017 dataset demonstrate that the proposed method achieves an accuracy of 98.55% and an F1-score of 98.40%. Benefiting from the soft-label transmission and parameter-free aggregation design, the framework gets rid of the constraint of homogeneous model architecture and natively supports heterogeneous network models and edge devices with different computing capabilities. It also significantly reduces communication traffic and per-round training latency, confirming its excellent real-time performance and applicability in resource-constrained edge environments. Full article

Rolling bearing fault diagnosis remains challenging due to environmental noise, insufficient information sharing between diagnosis and prediction tasks, and poor model generalization ability. To address these issues, this paper proposes a fault diagnosis and severity prediction method integrating probabilistic principal component analysis (PPCA) and empirical mode decomposition (EMD) with a one-dimensional convolutional neural network (1DCNN) and bidirectional gated recurrent unit (BiGRU). The proposed model consists of two parallel branches for fault diagnosis and fault severity prediction. A self-attention mechanism is integrated into both branches to enhance feature extraction via adaptive feature weighting. In addition, parameter sharing and weighted loss functions are adopted to improve the training efficiency and collaborative learning between the two tasks. PPCA and EMD are employed for signal denoising and reconstruction while preserving fault-related features. Experiments on public datasets and industrial production-line data show that the proposed method improves the fault classification accuracy from 92.43% to 99.71% under different load conditions, while achieving 98.99% accuracy in fault severity prediction. Noise interference tests further demonstrate the effectiveness of the model. A production-line case study further illustrates the feasibility of applying the proposed method to real monitoring signals. These results confirm the effectiveness and practical potential of the proposed method for rolling bearing fault diagnosis and health assessment. Full article