Search Results (218)

Rolling bearings are indispensable elements in mechanical equipment, and their condition is closely related to system reliability and operational safety. To enhance diagnostic performance with limited samples and reduce the dependence on manual parameter selection, this study proposes a fault diagnosis approach that combines Continuous Wavelet Transform (CWT), Convolutional Neural Network (CNN), Black-winged Kite Algorithm (BKA), and Least Squares Support Vector Machine (LSSVM). The original 1-D vibration signals are first processed by CWT to obtain 2-D time–frequency representations. CNN is then employed to learn deep fault-sensitive features, which are subsequently fed into LSSVM for state classification. To further improve classification performance, BKA is used to automatically search for the optimal LSSVM parameters, with validation accuracy adopted as the fitness criterion. Experiments conducted on three public bearing datasets, namely CWRU, JNU, and SEU, indicate that the proposed method outperforms CNN, CNN-SVM, and CNN-BiGRU under small-sample conditions. In addition, t-SNE results show more distinct feature clusters, while BKA exhibits faster convergence and better global search capability than Particle Swarm Optimization (PSO) and Genetic Algorithm (GA). Full article

The Internet of Medical Things (IoMT) is a promising technology for enabling smart and efficient healthcare systems through continuous physiological monitoring, early anomaly detection, and proactive patient management. However, IoMT sensors are vulnerable to False Data Injection Attacks (FDIAs), in which adversaries can manipulate sensor measurements to compromise diagnostic accuracy, mislead clinical decision-making, and threaten patient safety. Existing detection approaches often rely on population-level statistical models that may not fully capture individual physiological variations or residual-based thresholds designed for relatively simple attack scenarios, limiting their ability to exploit the spatio-temporal dependencies of multi-sensor physiological streams and detect stealthy or adversarial FDIAs. This paper proposes a patient-specific FDIA detection framework based on a Graph Convolutional Network–Gated Recurrent Unit (GCN–GRU) digital twin that learns an individual patient’s normal physiological behaviour from clean baseline telemetry. The trained digital twin is integrated into a Kalman filter as the state prediction model, and the resulting standardised innovation residuals are used as detection features. To characterise stealthy attack behaviours, four complementary window-based feature groups are extracted from the innovation sequence: innovation statistics, sensor correlation drift, temporal smoothness, and uncertainty mismatch. A CNN-1D classifier is then trained to learn discriminative temporal attack patterns from these features for accurate detection. A structured attack taxonomy comprising five stealthy and adversarial FDIA scenarios is developed, where attacks are injected as smooth gradual or abrupt coordinated modifications to sensor measurements while remaining within plausible physiological ranges. Experiments conducted on the WUSTL-EHMS-2020 benchmark dataset demonstrate that the proposed framework achieves an F1-score of 94.3%, outperforming Isolation Forest and PCA Reconstruction by 34 percentage points. Furthermore, the proposed framework reduces the false alarm rate to 3.6%, compared with 35.1% and 9.2% achieved by Isolation Forest and PCA Reconstruction, respectively. These results demonstrate the effectiveness of the proposed framework for reliable detection of stealthy FDIAs in IoMT-based healthcare systems. Full article

Parkinson’s disease (PD) diagnosis and severity assessment increasingly rely on gait analysis. To capture complex gait dynamics from limited sensor data, we propose a novel multi-branch framework integrating a one-dimensional convolutional neural network (1D-CNN), a bidirectional gated recurrent unit (Bi-GRU), and Kolmogorov–Arnold networks (KANs) (CNN-GRU-KAN) utilizing 18-channel vertical ground reaction force (VGRF) signals. Each module serves a distinct clinical purpose: the 1D-CNN branch with squeeze-and-excitation (SE) attention extracts localized spatial plantar pressure patterns, while the Bi-GRU branch with temporal attention captures long-range rhythm abnormalities. Crucially, the KAN serves as the classification head. By utilizing learnable B-spline functions instead of traditional fixed activations, KAN adaptively models the highly non-linear boundaries between healthy controls and varying PD severities, effectively mitigating overfitting. Under rigorous subject-independent cross-validation, our model achieves 98.43% accuracy for binary PD detection and 93.46% for five-class UPDRS severity grading. These results highlight the framework’s strong potential for low-cost, unobtrusive clinical tracking and home monitoring. Full article

Occupancy detection is fundamental to the operational intelligence of smart buildings, driving critical functions in energy management, HVAC automation, and physical security. While modern Deep Learning (DL) models have achieved high accuracy in parsing complex environmental sensor data, they remain highly vulnerable to adversarial examples, imperceptibly perturbed inputs designed to deceive neural networks. These vulnerabilities pose severe real-world risks, ranging from energy sabotage, in which systems heat empty rooms, to critical security breaches in which intruders go undetected. To address this security gap, we propose ADS-Guard, a novel Adversarial Detection and Sanitization (ADS) framework rooted in sequence-to-sequence autoencoder purification. Unlike standard denoising techniques, ADS-Guard incorporates a latent consistency regularization mechanism that encourages alignment between clean and adversarial representations in the latent feature space. We evaluated ADS-Guard using a comprehensive experimental pipeline comprising five distinct DL architectures (LSTM, GRU, 1D-CNN, MLP, and Transformer) across three diverse datasets: (1) The UCI Occupancy dataset (20,699 samples) for standard binary detection; (2) Building59 dataset (7200 samples) for three-class occupancy-level classification (Low, Medium, High); and (3) Room Occupancy dataset (10,129 samples), representing a highly imbalanced binary occupancy-detection task. We evaluate ADS-Guard against both Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks across diverse occupancy datasets and model architectures. We further assess the framework under adaptive white-box attacks and compare its performance with FGSM-based and PGD-based adversarial training baselines. Our results demonstrate that adversarial attacks can substantially degrade occupancy-detection performance across datasets and model architectures. ADS-Guard consistently improves robustness relative to undefended models against both FGSM and PGD attacks, recovering a substantial portion of the lost performance in binary occupancy tasks and providing meaningful gains in the more challenging multi-class setting. Furthermore, ADS-Guard remains effective under stronger adaptive threat models while providing a practical retraining-free defense that can be integrated with existing occupancy-detection systems without modifying downstream classifiers. Full article

Permanent magnet synchronous motors (PMSMs) are widely used in new energy vehicles, electric drive systems, and industrial servo applications. Excessive permanent magnet temperature may lead to magnetic performance degradation or even irreversible demagnetization; therefore, accurate estimation of permanent magnet temperature is of considerable importance. However, existing data-driven methods generally rely heavily on high-frequency measurements, and their prediction accuracy tends to deteriorate under low-frequency sampling conditions. Moreover, purely data-driven models lack explicit physical constraints, which limits their interpretability and generalization capability. To address these issues, this study proposes a physics-informed long short-term memory model for permanent magnet temperature prediction. A physics-based loss function is formulated using the PMSM d–q-axis voltage balance equations, while the d- and q-axis inductances are treated as trainable parameters during network optimization. This design enables the temperature prediction task and the electromagnetic constraints to be optimized jointly. Multi-operating-condition experiments are conducted using a publicly available electric motor temperature dataset, and the proposed model is compared with CNN, GRU, MLP-PINN and TNN models. In addition, experiments involving different downsampling ratios, errors in the high-temperature region, parameter sensitivity, physical parameter identification, and input-feature effects are performed to comprehensively evaluate the proposed model. The results show that the PINN-LSTM model achieves the best overall prediction performance, with an MAE of 1.6048 °C, an RMSE of 2.1890 °C, and an R² of 0.9861, outperforming all comparison models. The model also maintains high prediction accuracy in the high-temperature region, with an MAE of 1.363 °C and an RMSE of 1.896 °C. Furthermore, the parameters learned by the model can effectively reconstruct the variation trends of the d- and q-axis voltages under the test operating conditions. Sensitivity analysis of the temperature coefficients further demonstrates that the model is robust to deviations in key physical parameters. These results indicate that the proposed method can achieve accurate and robust permanent magnet temperature prediction under low-frequency sampling conditions, providing an effective solution for motor thermal-state monitoring and health management. Full article

Accurate vehicle trajectory prediction is essential for autonomous driving safety. However, existing data-driven models often ignore kinematic constraints, causing lateral jitter and trajectory distortion, while purely kinematics-based models lack flexibility in complex interactions. To address this, this paper presents a multimodal trajectory prediction method combining kinematics with dynamic interaction features. Operating in the Frenet coordinate system, the proposed model extracts historical features via a Bidirectional Gated Recurrent Unit (Bi-GRU) and utilizes an Adaptive Social Gating Network (ASGN) with multi-head attention to filter irrelevant interaction noise. This paper introduces a Stochastic Gating Decoder for multimodal latent variable sampling, adaptively fusing kinematics and data-driven paths to capture driver intention uncertainty while maintaining kinematic consistency. The model is trained using a composite loss function (Focal Loss and Best-of-K) to mitigate dataset long-tail distribution and trajectory divergence. Experiments on the HighD dataset show the proposed model achieves a minADE of 0.425 m and a minFDE of 0.955 m, outperforming baselines and reducing Lat-ADE by 53.9% compared to Social-GAN. These results confirm the model generates smoother, kinematically interpretable trajectories with higher accuracy in long-tail lane-changing scenarios. Full article

Iced conductor galloping poses a critical threat to high-voltage transmission line safety and stability, yet existing monitoring systems are constrained by single-sensor dependency, inadequate signal denoising, and limited early warning accuracy. This paper presents a field-validated framework integrating multi-source sensor fusion, adaptive signal denoising, and deep learning-based early warning for iced conductor galloping. A five-layer Internet of Things (IoT) monitoring architecture is designed, fusing fiber Bragg grating (FBG) arrays, MEMS inertial sensors, and micro-meteorological stations, with dual-spectrum cameras providing auxiliary visual verification. Self-powered MEMS nodes utilizing electromagnetic induction energy harvesting are shown to have achieved year-round zero-external-power maintenance. An improved hummingbird local optimization algorithm (IHLOA) adaptively optimizes variational mode decomposition (VMD) parameters, combined with wavelet threshold denoising (WTD) for joint signal preprocessing, achieving a 17.78 dB signal-to-noise ratio improvement [95% CI: 17.2–18.3 dB]. A 26-dimensional multi-domain feature vector is constructed and reduced to 12 discriminative features via ReliefF selection. A Temporal Adaptation Gated Recurrent Unit with Attention (TA-GRU-Attention) model incorporating an adaptive irregular time series perception module is developed for galloping early warning, with all models evaluated exclusively on real-event test samples. Experimental validation through 1:20 scale wind tunnel aeroelastic tests and a 15-month field demonstration on an operating 220 kV transmission line at 2800–3200 m elevation demonstrates a galloping prediction accuracy of 91.3% [95% CI: 81.5–97.2%] under stratified time series split, a missed alarm rate of 12.2%, and a median advance warning time of 38.5 min (range: 18–65 min, IQR: 26–52 min), with 97.3% system availability over the deployment period. Full article

Touchless interfaces for embedded systems require models that are both compact and evaluated under protocols that reflect realistic generalization. Public low-resolution gesture datasets can obscure this requirement because repeated samples, temporal neighbors, or subject identity may leak across train and test partitions. This study presents a leakage-aware multi-corpus benchmark in which four public corpora are evaluated within their own dataset-specific partitions after conversion to a common 8 × 8 depth representation when needed: a public IEEE DataPort ToF CSV with 8400 frames from four gestures, DS.GestureRecognition_TOF.1.0 with 3109 frames extracted from 40 low-resolution ToF sequences and three gestures, the depth component of the public Kinect+Leap dataset with 1400 samples from 14 subjects and 10 gesture classes, and the public Senz3D depth archive with 1320 samples from 4 subjects and 11 gesture classes. No blurred RGB image dataset, deblurring target, or synthetic blur augmentation is produced in the experiments; the only spatial reduction applied to external depth corpora is deterministic block averaging for a shared low-resolution depth input. The term multi-corpus is used here to clarify scope: the experiments do not train on one dataset and test on another, but instead compare optimistic random splits against structure-aware alternatives matched to each corpus: chronological block-aware, sequence-aware, and subject-wise evaluation. On the original IEEE ToF benchmark, the best macro F1-score decreases from 1.0000 under random splitting to 0.8494 under block-aware evaluation. On DS.GestureRecognition_TOF.1.0, random splitting again reaches 1.0000, whereas the best sequence-aware result falls to 0.6697. On the downsampled Kinect+Leap depth benchmark, random frame classification reaches 0.7639 macro F1, while subject-wise evaluation drops to 0.0566. On Senz3D reduced to 8 × 8, random frame classification reaches 0.9068 macro F1, while subject-wise evaluation drops to 0.0717. Device-side validation on an NVIDIA Jetson Nano Developer Kit preserved a macro F1-score of 0.8667 for the selected GRU on the primary IEEE ToF dataset with 14.48 ms/window steady-state CPU latency. The results show that low-resolution gesture recognition conclusions are strongly protocol- and corpus-dependent: compact temporal models remain viable for genuine ToF streams, but naive random splitting can substantially overstate generalization, and subject-wise recognition after aggressive 8 × 8 conversion remains difficult for the evaluated datasets and architectures. Full article

The proliferation of space debris in Low Earth Orbit (LEO) poses a growing threat to operational satellites, requiring robust surveillance and tracking systems. Radar is a primary technology for monitoring these objects; however, standard tracking algorithms often degrade when measurements are sparse, corrupted by non-Gaussian noise, or available only as short tracklets. Traditional methods, such as the Unscented Kalman Filter (UKF), rely on explicit physical and statistical models that may struggle to converge under highly nonlinear dynamics, uncertain initialization, and non-ideal sensor perturbations. In this study, we propose a hybrid deep learning framework for learned one-step tracklet filtering of radar measurements. The architecture consists of a stateful Gated Recurrent Unit (GRU) layer followed by one-dimensional Convolutional Neural Network (1D-CNN) layers, complemented by variable-specific preprocessing strategies, including residual learning for range and relative pivoting for azimuth, to handle the scale disparities and heterogeneous behavior of radar observables. This design combines the ability of GRUs to model temporal dependencies with the effectiveness of CNNs in extracting local features for signal denoising. The method is validated on synthetically generated LEO trajectories with realistic orbital perturbations and tunable radar noise profiles, including Gaussian noise, impulsive spikes, transient degradation, and state-dependent perturbations. Compared with EKF- and UKF-based analytical baselines, the proposed model achieves lower filtering error and improved robustness under severe non-Gaussian disturbances. Stress testing further shows that the network can reject non-physical sensor anomalies without requiring long initialization warm-up phases, making it suitable for sparse short-tracklet processing in synthetic SST benchmark scenarios. Full article

With the increasing complexity of the electromagnetic environment, traditional radar target recognition methods face severe challenges. High-resolution range profile (HRRP) and Radar Cross Section (RCS), as two important radar features, each has its own advantages in target recognition but also exhibits limitations. To enhance radar target recognition performance in complex scenarios such as low signal-to-noise ratio (SNR), this paper proposes a recognition method based on heterogeneous multi-modal feature fusion. The proposed method constructs a three-channel parallel encoding network, which utilizes Convolutional Long Short-Term Memory (ConvLSTM), One-Dimensional Convolutional Gated Recurrent Unit (Conv1D-GRU), and Gated Recurrent Unit (GRU) to extract deep discriminative features from raw HRRP sequences, RCS sequences, and HRRP statistical features, respectively. Furthermore, it innovatively designs a dual-path cooperative fusion mechanism, achieving explicit inter-modal correlation modeling through a cross-attention module and dynamically learning the importance of each modality through an adaptive weight fusion layer, thereby realizing deep complementarity and enhancement of multi-modal information. Experimental results demonstrate that under various signal-to-noise ratios and polarization conditions, the proposed method achieves a maximum average recognition accuracy of over 99% for 6 ship targets. Compared with the traditional three-channel fixed-weight fusion method, the recognition accuracy of the proposed method increases from 90.01% to 99.42% under co-polarization, and from 81.14% to 98.15% under cross-polarization, fully validating the effectiveness and superiority of the dual-path fusion mechanism. Full article

3D point cloud understanding plays a vital role in remote sensing, robotic perception and intelligent scene analysis. However, real-world point cloud data are often affected by sensing noise, incomplete geometry, occlusion, and irregular sampling, posing significant challenges to reliable geometric representation learning and long-range contextual modeling. Existing methods typically rely on fixed neighborhood aggregation or computationally expensive global interaction mechanisms, leaving considerable room for improvement in terms of robustness and efficiency under complex sensing conditions. To address these challenges, we propose GeoSeqNet, a geometry-aware contextual learning framework for robust 3D point cloud analysis. Specifically, an Enhanced Local Operator (ELO) is introduced to strengthen local geometric representation, while a Geometric Encoding Module (GEM) is employed to preserve spatial geometric information during long-range feature interactions. In addition, an Adaptive Gate Fusion (AGF) module is designed to effectively integrate Gate-Scaled LSTM and GRU branches, enabling efficient long-range contextual modeling. By jointly exploiting local geometric cues and long-range contextual information, GeoSeqNet achieves robust feature learning with low computational overhead. Extensive experiments on ModelNet40, ScanObjectNN, and ShapeNetPart demonstrate the effectiveness of GeoSeqNet. The proposed method achieves competitive performance while maintaining a favorable efficiency–accuracy trade-off and exhibits strong robustness in complex real-world scenarios. These results indicate that GeoSeqNet provides an effective and reliable solution for point cloud understanding in challenging sensing environments. Full article

Accurate structural damage identification under limited data availability and measurement noise remains a persistent challenge in structural health monitoring (SHM). This study proposes TLCA-RATNet, a transfer learning-enhanced residual attention temporal network for vibration-based damage state classification under noisy and small sample conditions. RATNet integrates adaptive threshold residual denoising, residual attention, and bidirectional gated recurrent unit (BiGRU)-based temporal modeling to suppress noise, emphasize damage-sensitive features, and capture global temporal dependencies. The current implementation is formulated as a single-task classifier, in which local feature refinement and global temporal representation are jointly optimized end-to-end through a unified damage classification objective. Transfer learning further initializes the target domain model using knowledge learned from a data-rich source structure, while regularized fine-tuning reduces overfitting on limited target samples. Experiments were conducted on a six-story lumped-mass shear structure, a three-story physical frame, and the IASC-ASCE SHM benchmark structure, using 10–145 training samples per damage class and additive-noise conditions ranging from 0 to 30 dB signal-to-noise ratio. On Dataset 2 at 5 dB, TLCA-RATNet achieved an accuracy of 89.86%, exceeding LSTM and CNN-BiGRU by 9.94 and 13.10 percentage points, respectively. On Dataset 3 at 0 dB, it achieved 86.00% accuracy, outperforming CNN-BiGRU by 10.72 percentage points. In the limited sample transfer experiment on Dataset 2, transfer learning increased the accuracy from 93.05% to 100.00%, representing a gain of 6.95 percentage points over training from scratch. These results indicate that TLCA-RATNet provides a data-efficient and noise-robust approach for damage state screening and rapid model adaptation in SHM applications with scarce labeled data and noisy measurements. Full article

The knee adduction moment (KAM) is an important biomechanical marker of medial knee joint loading, but conventional assessment requires laboratory-based motion analysis. This study aimed to develop and evaluate an improved inertial measurement unit (IMU)-based KAM estimation system that incorporates a one-dimensional U-Net (1D U-Net) with a gated recurrent unit (GRU) bottleneck. Gait data from 50 participants were used for model development and internal evaluation, and an independent holdout dataset collected using the same study protocol included 45 participants. Simultaneously recorded IMU signals and reference KAM obtained from motion capture-based inverse dynamics were used for training and evaluation. The model used a single IMU attached to the tibial tuberosity to reconstruct the KAM waveform during the stance phase. For peak KAM, Pearson’s r values were 0.819, 0.602, and 0.621 for the training, internal validation, and holdout datasets, respectively. For KAM impulse, the corresponding values were 0.941, 0.744, and 0.745. These findings indicate that a simplified wearable system combining a single tibial IMU with a 1D U-Net incorporating a GRU bottleneck can reconstruct the KAM waveform during walking and derive clinically relevant parameters from it. Although further improvement in prediction accuracy is required, this framework may provide a practical basis for simplified screening or monitoring of medial knee loading in settings closer to routine clinical practice. Full article

Modern distribution networks operate under increasingly demanding conditions, characterized by the integration of distributed energy resources, unbalanced three-phase operation, low measurement redundancy, variable topologies, and data uncertainty. In this context, distribution system state estimation (DSSE) is a key tool for operational monitoring; however, its practical deployment is often hindered by topological inconsistencies and gross measurement errors. This paper proposes an integrated and modular deep learning-based methodological framework that combines active topology identification (ATI), gross error detection (GED), error-type identification (ETI), error-location identification (ELI), measurement reconstruction and correction (MRC), and DSSE. The ATI module is formulated as a global multiclass classifier, whereas the subsequent modules are trained as topology-specific models. Compromised measurements are handled through an iterative GED–ETI–ELI–MRC loop that detects, identifies, locates, and corrects one anomalous measurement per iteration before re-evaluating the input vector. The proposed methodology was validated by using simulation-based scenarios generated in OpenDSS for a real unbalanced three-phase 240-node distribution feeder. The results show that no single architecture is dominant across all subproblems: WaveNet1D achieved the best relative performance in ATI, GED, and ETI; EncDec-CNN in ELI; NBEATS1D in MRC; and EncDec-GRU in DSSE. Additionally, WLS estimators based on both nodal voltages and branch currents failed to achieve numerical convergence on the 240-node test system under the evaluated conditions, a finding consistent with recent literature reporting analogous convergence failures in distribution networks of similar or smaller scale. Furthermore, the integrated evaluation shows that omitting ATI increases the voltage-magnitude MAE by a factor of 12.3 and the voltage-angle MAE by a factor of 8.1 with respect to the complete framework, whereas omitting only the compromised-measurement treatment increases these errors by factors of 1.8 and 1.9, respectively. The total offline computational cost was approximately 1587.9 h (66.2 GPU-days), while the online inference latency was approximately 0.45 ms per sample, making the framework compatible with AMI- and SCADA-based monitoring cycles. These findings confirm that topological consistency is the dominant factor in DSSE accuracy and that iterative measurement correction meaningfully improves estimator robustness under anomalous measurement conditions. Full article

Forecasting the deformation and movement of cumulus clouds provides an important basis for ultra-short-term solar irradiance nowcasting in photovoltaic (PV) power generation. Existing methods mainly use two-dimensional (2D) ground-based sky images for forecasting, which have limited ability to represent the three-dimensional (3D) spatial structure of cumulus clouds and the influence of wind on cloud motion. In this study, we propose a wind-aware ultra-short-term spatiotemporal forecasting model for 3D cumulus clouds, termed three-dimensional Cloud Long Short-Term Memory with Wind Gate Recurrent Unit (3dCLSTM + WindGRU). The model uses 3dCLSTM to learn the spatial structure and temporal evolution of 3D voxel cumulus cloud sequences, and embeds a WindGRU unit between 3dCLSTM layers to introduce wind speed and wind direction information for wind-driven transient motion modeling. Experiments were conducted on 1-min and 10-min 3D cumulus cloud datasets reconstructed from ground-based sky image datasets collected at sites in California and Colorado, USA. All voxel sequences were resampled to 64 × 64 × 64, with five-step prediction for the 1-min dataset and three-step prediction for the 10-min dataset. The results show that 3dCLSTM achieved a structural similarity index measure (SSIM) of 0.7913 on the 1-min dataset, while 3dCLSTM + WindGRU achieved the best performance on the 10-min dataset, with an SSIM of 0.3512 and a peak signal-to-noise ratio (PSNR) of 18.3625. Compared with 3dCLSTM, introducing WindGRU improved the SSIM by 4.8% on the 10-min dataset, with a more evident improvement under relatively high wind-speed conditions. These results indicate that wind-aware volumetric spatiotemporal modeling can support ultra-short-term 3D cumulus cloud forecasting and provide a useful technical basis for solar irradiance nowcasting. Full article