Abstract Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is essential for the reliability and safety of modern energy systems. However, the capacity regeneration (CR) phenomenon, a temporary recovery of capacity during cycling or rest, introduces non-monotonic fluctuations in degradation trajectories, posing significant challenges to existing prediction models. Many current approaches treat CR as noise or overlook its physical significance, limiting interpretability and accuracy. To address this, we propose a hybrid framework that explicitly models both regenerative and degenerative battery behaviors. The method uses Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose capacity sequences into high- and low-frequency components. A graph convolutional network-long short-term memory (GCN-LSTM) branch captures high-frequency local fluctuations containing regeneration-related information, while a deep neural network (DNN) branch learns the long-term degradation trend. These are fused to reconstruct the full degradation path and predict RUL. The CEEMDAN-GCN-LSTM-DNN hybrid model achieves a mean absolute percentage error below 0.15%, outperforming several state-of-the-art baselines. It also demonstrates strong robustness under data-limited conditions and effectively captures complex CR patterns often missed by conventional methods. This study offers a new perspective for handling non-monotonic battery degradation and provides a useful tool for battery health management and predictive maintenance.
Abstract The concurrent demand for lightweight structural materials and industrial waste valorisation has intensified interest in hybrid natural fibre composites. This work presents a computational framework built on a response surface methodology (RSM) surrogate base with two parallel optimisation approaches for the design and multi-objective optimisation of hybrid abaca fibre /red mud/ epoxy composites targeting simultaneous maximisation of tensile, flexural, and impact strengths. This three-stage framework comprises an RSM surrogate base (Stage 1) feeding two parallel optimisation routes as a multi-objective genetic algorithm (MOGA)-driven multi-objective search (Stage 2) and a Bayesian-optimisation (BO) benchmark for evaluation-efficiency comparison (Stage 3), with a SHapley Additive exPlanations (SHAP) explainability layer applied across the surrogate models. At Stage 1, RSM was applied to a 3 3 full-factorial experimental dataset (ASTM D638/D790/D256) to develop second-order quadratic surrogate models achieving R 2 ⩾ 88.77% across all mechanical responses. At Stage 2, a within three iterations (186 design evaluations) to a 70% Pareto front, yielding three candidate compositions, all at the upper abaca bound ( A w = 7.9 wt.%). Candidate Point 3 ( R w = 11.79%, R p = 82.07 μ m) achieved tensile strength 47.88 MPa, flexural strength 42.22 MPa, and impact strength 42.93 J m −1 , validated experimentally within 6.00% error. Stage 3 (benchmarking): BO with Gaussian process surrogates and expected improvement acquisition identified a model-predicted candidate optimum in only 50 evaluations, reducing computational cost by 73%, demonstrating the feasibility of rapid-screening campaigns for related fibre systems. The BO result (+29.0% tensile strength relative to MOGA Candidate Point 3) is a Random Forest (RF) surrogate prediction and has not been experimentally validated. Complementing the optimisation, SHAP analysis of RF surrogate models trained on the RSM-MOGA surrogate dataset (CV R 2 ⩾ 0.88) provides quantitative per-variable attribution of surrogate model predictions: red mud content ( R w ) is the dominant tensile driver (mean |SHAP| = 6.40 MPa), while abaca content ( A w ) governs both flexural (|SHAP| = 2.92 MPa) and impact (|SHAP| = 3.22 J m −1 ) responses, overturning the ranking produced by conventional linear sensitivity analysis and providing surrogate-model-level validation for the MOGA-optimal composition.
Abstract Efficient lossless compression is essential for reducing storage and transmission requirements while preserving data exactly. Traditional dictionary-based and statistical compressors can be limited in their ability to exploit complex, long-range dependencies. In this paper, we propose a lossless compression framework that combines a T5-small architecture with Advantage Actor-Critic reinforcement learning to generate a variable-length sequence of discrete compressed tokens. Rather than relying on a continuous autoencoder bottleneck, the proposed framework directly optimizes the length of this discrete representation while preserving the information required for exact reconstruction. The method operates without external grammatical rules or world knowledge. On the enwik8 benchmark, it achieves a compression ratio of 4.14, improving upon XZ (4.05) by approximately 2.3% and GZIP (2.74) by 51.0%. Although this result remains below NNCP v3.2 (6.70), it demonstrates that reinforcement learning can learn a compact, lossless discrete representation using two T5-small networks evaluated on a single 12 GB GPU device.
Abstract Most of the existing spatial-layout studies focus on geometry optimisation or floor plan generation and fail to consider changes in operating pressure, occupant experience, engineering constraints and lifecycle management. Therefore, this study will consider architectural spatial form perception and layout optimisation as a human-centred intelligent-building decision-support problem. All of the above can be used to build a spatial topology graph: BIM models, laser point clouds, floor plans, passenger flow records, indoor environmental sensing data and anonymised service feedback. Spatial units are expressed as geometric, functional, adjacency, temporal, indoor environmental quality (IEQ), and user-oriented attributes, and an NSGA-II model evaluates accessibility, congestion risk, functional adjacency, daylight availability, usable area, comfort exposure and engineering compliance. Operational data from a full-scale public building in March 2025 to February 2026 were collected for verification. Multiple-source method: F1 score of 0.931 for spatial-unit identification and adjacency-recognition accuracy of 0.946. Compared with the baseline, the recommended solution has reduced the mean travel distance to 62.8m, lowered the congestion-risk index to 0.43, decreased adjacency conflicts by 42.4%, and increased daylight coverage by 12.4 percentage points. The results offer architects, BIM specialists and facility managers an explainable basis for design revision and building operation planning; however, extended occupant satisfaction and cross-building mobility still need to be verified.
Abstract To address feature dilution and overlap in small and dense target coexistence scenarios, this paper proposes an efficient improved framework, Enhanced Feature-Aware YOLO (EFA-YOLO). Based on YOLO11, EFA-YOLO integrates targeted feature enhancement, adaptive fusion, and dedicated detection head design, combining frequency-domain enhancement, collaborative attention, and adaptive fusion to preserve fine-grained details and global context.A Multi-Scale Frequency-Aware Attention (MSFA) module adopts multi-directional wavelet decomposition to strengthen high-frequency cues for small targets, while maintaining coherent semantic representations in dense scenes via coordinated channel-spatial attention. To eliminate redundancy from static feature concatenation, an Adaptive Weighted Concatenation Fusion (AWCF) mechanism dynamically balances cross-scale contributions for cleaner fused features.Additionally, the small-target detection head is restructured with shallow high-resolution features to retain micro-scale information, and redundant branches are pruned for efficiency. These designs enable EFA-YOLO to robustly handle small and dense targets without introducing excessive architectural complexity. Experiments on the visdrone2019 and coco datasets show that compared with the baseline model yolo11 mAP50 It has increased by 3.7% and 2.8% respectively, which can effectively solve the core pain points of small target and dense target detection, and provide technical support for actual scenes such as autonomous driving and UAV patrol inspection.
Abstract Deep underground mining is increasingly constrained by severe thermal hazards, particularly high-temperature and high-humidity conditions, which affect worker health, equipment reliability and energy performance. This review examines the causes, control strategies and utilisation pathways of thermal hazards in deep underground mines. Available field evidence is weighted towards coal-mine applications, whereas metal-mine studies mainly address ventilation and localised or mobile cooling; consequently, transferability across mine types requires validation against site-specific thermal loads and operating constraints. Heat sources are analysed in terms of their coupled sensible and latent heat contributions, while control technologies are organised into passive thermal-load mitigation and active thermal-regulation systems. Active regulation is further considered in terms of cooling-capacity generation, distribution and release, with localised cooling highlighted as a supplementary strategy for high-exposure working zones. Thermal-energy recovery is reviewed from mine water, return airflow and heat rejected by cooling and dehumidification systems. On this basis, a Reduce–Regulate–Recover framework is proposed to connect source reduction, environmental regulation and waste-heat recovery. Personal cooling is also considered as a worker-centred supplementary pathway. Overall, this review provides a structured synthesis to support safer, more energy-efficient and more sustainable thermal management in deep underground mines.
Abstract During the operation of high-speed trains, the pantograph-catenary system is responsible for continuously delivering traction power from the catenary to the locomotive. The occurrence of pantograph-catenary arcing not only indicates the deterioration of dynamic power transmission quality but also severely ablates contact components and causes electromagnetic interference, thereby affecting train operation safety. Therefore, detecting pantograph-catenary arcing is of great significance. Considering the high uncertainty in the morphology and scale of arcing behavior in complex scenes, achieving reliable, fast, and real-time detection remains a challenge. To this end, we propose a real-time detection method for pantograph-catenary arcing based on morphology-aware and efficient feature modeling (Morphology-Aware Efficient Feature Modeling RT-DETR, MEFM-RTDETR). This method introduces a dynamic adaptive convolution kernel weight generation mechanism to model arcing features of varying scales, orientations, and morphologies, thereby enhancing the network's morphology-aware capability. Second, a multi-scale feature enhancement and fusion network (MEFusion) is constructed to fully integrate shallow detail information with deep semantic information, addressing the issue of the loss of tiny arcing features. Finally, an efficient additive attention mechanism is introduced into the intra-scale feature interaction module (AIFI) to enhance the modeling capability of global contextual information. Experimental results show that MEFM-RTDETR achieves a 2.63% improvement in mAP50, a 2.04% improvement in mAP50-95, a 4.9 GFLOPs reduction in computational complexity, and approximately a 6 M reduction in the number of parameters compared to the baseline model on the arcing detection task.
Abstract With the development of the low-altitude economy, industries such as general aviation, electric vertical takeoff and landing (eVTOL), and electric vehicle manufacturing are developing rapidly. Consequently, the requirements for forming processes of thin-walled components have become more stringent. As a process-optimization study aimed at industrial applications, this paper investigates a novel spatiotemporal energy control scheme for the electrohydraulic forming of complex-shaped sheet metal components. A multi-contour discharge block and impulse current generator system is developed to achieve precise temporal and spatial regulation of energy flow, verifying the feasibility of implementing spatiotemporal control of energy flow on plastically deformed sheets. It is confirmed that energy flow can be concentrated in specific regions of the sheet; the energy concentration in the central region is significantly enhanced compared with traditional single-contour systems, and can be redirected to address hard-to-form edge regions via liquid jet interaction. Regulating the pre-discharge stage and discharge delays between multiple discharge chambers enables synchronous or differentiated energy release. The optimal discharge delay of 150–300 μs maximizes energy release efficiency, with the central region achieving a plastic strain rate of 35 s⁻¹. Compared with traditional single-contour electrohydraulic forming equipment, the proposed technology reduces energy consumption via optimized energy distribution, improves forming efficiency by over 30%, and shortens the process cycle. This research provides a feasible technical solution for the efficient manufacturing of complex thin-walled sheet metal components, addressing the issues of energy loss and poor deformation uniformity in complex thin-walled sheet part forming and supporting the industrial upgrading of aviation and automotive manufacturing.
Abstract This paper details the analysis of a Triple-Metal Junctionless Dual-Gate Gate-All-Around MOSFET (7nm channel length) to investigate switching and electrostatic properties. This research examines varying the metal gate work-function between 4.4 eV and 5.0 eV, where 4.6 eV is the metal gate work-function to provide balanced switching characteristics with a subthreshold slope of about 74-75 mV/dec and an Ion/Ioff ratio of the order of 10¹¹. It has been found that using the TiO2 arrangement permits low drain-induced barrier lowering (DIBL) of approximately 18.7 mV along with high intrinsic gain (made available by better electrostatic confinement of the channel, 1.45×10-2S/µm). Electrical properties such as Id-Vgs, Id-Vds, transconductance, Intrinsic gain and capacitances of gates are examined coupled with internal electrostatic properties with band energy profiles, electrostatic distribution potential as well as quasi -fermi level to understand the behavior of carrier transport. The results are presented and analyzed by means of Silvaco TCAD and the results obtained point to a better switching efficiency and the improvement of the electrostatic stability, which sets the perspectives of the TM-JLDGAA MOSFET in the context of low-power digital, high-speed switching and analog/RF mixed-signal circuit design applications.
Abstract The heat dissipation efficiency and temperature uniformity of battery systems serve as critical determinants governing battery safety and electrochemical performance. In this study, a corner-reinforced hybrid liquid cooling-phase change material (PCM) strategy with multi-objective optimization for 18650 Li-ion battery thermal management was developed. The parallel aluminum micro-conduction inserts with internal coolant channels are embedded within the PCM matrix to strengthen heat conduction pathways toward corner regions while forming shortened and parallel coolant flow paths. Through single-factor, orthogonal, and Latin hypercube designs with response surface methodology, NSGA-II, MOPSO, and MODE were compared against the weighting coefficient method. The evaluated results demonstrate that the NSGA-II algorithm achieves the optimal trade-off between minimizing the maximum temperature and the maximum temperature difference, yielding a maximum temperature of approximately 300.599 K and reducing the temperature difference to around 2.581 K, thereby effectively enhancing the thermal uniformity, dissipation efficiency, and safety of the module. This study provides a theoretical foundation and optimization approach for the thermal management design of battery systems in electric vehicles and large-scale energy storage applications.
Abstract As a typical unsupervised representation learning model, autoencoders (AEs) can extract low-dimensional latent features from high-dimensional monitoring signals through an encoding–decoding structure. They are valuable for fault diagnosis scenarios, including rotating machinery, power equipment, and industrial processes. First, this review examines the basic AEs and their typical variants, analyzes their working mechanisms, and summarizes the general paradigm of AE-based fault diagnosis. It further focuses on the current research status of AEs in transfer learning, cross-condition generalization, and few-shot data augmentation. Finally, it points out the remaining problems in current research and discusses future directions.
Abstract This paper proposes a centralized control architecture for small-scale vertical axis wind turbine (WT) clusters. In the proposed system, a single power converter is employed to regulate the energy harvested from multiple WTs, thereby reducing converter count and overall system complexity. To demonstrate the feasibility of the proposed architecture, a perturb and observe (P&O)-based maximum power point tracking (MPPT) method using direct inductor current control is investigated to coordinate the multi-turbine system through the shared converter. The proposed architecture is validated in MATLAB/Simulink under various wind speed conditions and different P&O-based MPPT methods.
Abstract This paper presents an improved multi-target tracking method based on YOLOv11 and adaptive DeepSORT, named RFCB-KF Tracker, aimed at addressing challenges such as small target loss and occlusion in multi-target tracking. The main contributions include the introduction of a novel attention mechanism, RFA, during the detection phase. This mechanism combines the convolutional block attention module and coordinated attention and features lightweight convolution operations, receptive field channel attention convolution and receptive field convolution with block attention modules, to enhance detection accuracy and efficiency. In the tracking phase, we improved the Kalman filter by expanding the state space from two dimensions to eight, optimizing the state transition and measurement matrices to effectively manage variations in target size and shape, thereby reducing estimation errors. In addition, we implemented an adaptive noise weight mechanism that dynamically adjusts noise parameters to enhance adaptability in diverse scenarios. Cholesky decomposition was utilized to optimize Kalman gain calculations, addressing numerical instability in complex motion contexts, and the Mahalanobis distance was introduced to effectively handle measurement errors. Experimental results demonstrate that RFCB-KF Tracker achieves improvements of 2.3% in MOTA, 0.9% in MOTP, 0.3% in IDF1, and a 5.6 frames per second increase in FPS on the MOT17 dataset, indicating a significant enhancement in multi-target pedestrian tracking performance. However, it is important to note that, while accuracy has improved, the increased algorithmic complexity may lead to slower computational speeds. Future research will focus on finding the optimal balance between accuracy and processing speed to meet real-time application requirements.
Abstract Insulators are critical components that directly determine the secure operation of power systems. Prolonged exposure to harsh outdoor environments renders them susceptible to typical defects, including breakage and missing components. Insulator inspection images captured in field scenarios are commonly characterized by cluttered backgrounds and tiny target sizes, which pose substantial obstacles to accurate defect identification. These obstacles include excessive model parameter counts, unsatisfactory detection precision, and elevated rates of missed detections and false alarms. To solve these problems, a lightweight small-target detection model is proposed tailored for insulator defects, which is improved based on YOLOv5s. First, several C3 modules in the backbone of the baseline YOLOv5s are replaced by HAT (Hybrid Attention Transformer) Stages. These blocks exploit the complementary strengths of window-based self-attention and channel-wise attention, effectively enhancing the model's capacity to extract discriminative features for small-target insulator defects. Second, the Slim-Neck architecture is employed to guarantee the model’s ability to capture multi-scale features of insulator defects. Third, the directionally sensitive SIoU loss function is utilized to improve the precision of bounding box localization. Experimental results illustrate that Small-Target Insulator Defect YOLO (STID-YOLO) yields a 12.2% absolute gain in mAP@0.5 while cutting model parameters by 48.8% and GFLOPs by 56.3% relative to the baseline YOLOv5s. Consequently, the proposed model offers a promising solution for accurate insulator defect detection in complex environments with favorable detection efficiency.
Abstract The wide-range application of new energy vehicle batteries puts forward higher requirements for the performance of vehicle battery fault detection. However, the high dimensionality, intense noise interference, and scarcity of labeled samples in battery data degrade the performance of fault detection methods. This study proposes a fault detection framework integrating autoencoders (AE) and semi-supervised self-training for limited labeled data with abundant unlabeled data. First, Z-score normalization computes sample outlier degrees, and the 3σ criterion filters valid data to finish raw battery data preprocessing. Next, autoencoders are built and dimensionality reduction is adopted to suppress noise interference and extract low-dimensional features that reflect battery states. Subsequently, the low-dimensional features obtained by encoding pseudo-labeled battery data with the autoencoder are fed into the support vector machine (SVM), and the kernel parameters of the SVM are optimized through Bayesian optimization. Then, high-confidence samples are selected to generate pseudo labels, and semi-supervised learning is implemented in an iterative manner. Experimental results show that the proposed method not only has excellent detection performance but also displays prominent robustness in noisy data situations.
Abstract Ship detection in synthetic aperture radar images is difficult because ships vary greatly in scale, sea and land clutter can resemble targets, and speckle weakens object boundaries. This paper presents AMFF-DETR, an RT-DETR-based detector with four task-oriented modifications. MEIE-Net strengthens multi-scale contour features, dynamic histogram self-attention-based inter-scale feature interaction module groups features by intensity before attention, adaptive multi-branch cross-scale feature fusion adaptively fuses spatial and semantic information, and Inner-MPDIoU adds scaled-box overlap to minimum-point-distance regression. On SSDD, AMFF-DETR achieves 98.8% mAP 50 and 91.3% mAP 75 ; on HRSID, it achieves 93.4% and 78.1%, respectively. The model requires only 16.54 M parameters and 54.4 GFLOPs, while the loss-function and module-control experiments consistently demonstrate the effectiveness of the integrated design.
Abstract Laser bending is a die-less process suitable for hard-to-form 7075 aluminum alloy sheets, but its multi-pass deformation behavior and associated property changes remain insufficiently quantified. This study combined a three-dimensional thermo-mechanically coupled finite element model with experiments on 3 mm-thick 7075-T6 sheets to determine the effects of laser power, scanning speed, spot diameter, and scanning passes on bending angle, microstructure, and hardness. After experimental validation, microstructural evolution and property changes in the bending zone were examined using optical microscopy, Vickers microhardness testing, and x-ray diffraction (XRD). Validation against measured bending angles and residual bending-height profiles yielded a mean bending-angle error of 5.64%, a mean relative profile difference of 7.78%, a root-mean-square error of 0.332 mm, and a coefficient of determination of 0.969 after six passes. The results indicated that the bending angle was governed by heat input and the through-thickness temperature gradient, with the latter playing the dominant role. The angle peaked at a line energy density of 26.00 J·mm −2 , and parameter influence was ranked as number of passes > spot diameter > laser power > scanning speed. The maximum angle of 7.649° occurred at 750 W, a 5 mm spot, twelve passes, and 5 mm s −1 . For practical processing, 650 W, 5 mm s −1 , a 5 mm spot, and six passes were preferred, producing 3.900° without surface ablation. XRD revealed no new phases but a reduced mean peak width, while elongated rolled grains transformed into fine equiaxed grains. The mean Vickers hardness decreased from approximately 165 HV to 128.62–139.36 HV because of precipitate dissolution and coarsening, although grain refinement and short-duration re-precipitation at 650 W partially mitigated the softening. This study provides guidance for precision laser forming of 7075 aluminum alloy sheets.
Abstract Power grid maintenance involves diverse abnormal targets, such as unsafe worker behavior, abnormal meter readings, cabinet-door faults, foreign objects, oil leakage, and equipment damage. These targets often appear with scale variation, occlusion, and complex backgrounds, which challenges real-time inspection systems. This paper proposes PG-YOLO, a lightweight detector based on YOLOv12 for multi-class anomaly detection in power grid maintenance. A dataset with 28 categories and 188 353 annotated instances is built to support fine-grained recognition of normal and abnormal states. In PG-YOLO, wConv2D is introduced to reduce redundant computation and enhance local responses. A2C2F_CGLU is used to strengthen spatial–channel feature interaction, and SPPF&C2PSA is added to improve high-level semantic representation. Experiments on the constructed dataset show that PG-YOLO achieves 83.7 precision, 73.9 recall, 79.1 mAP 50 , and 56.3 mAP 50 : 95 with 2.98 M parameters, 5.0 GFLOPs, and 227.28 FPS. The ablation study verifies the contribution of each module, and Grad-CAM visualizations show more compact activation on anomaly-related regions.
Abstract Titanium alloys such as Ti–6Al–4V and Ti–3Al–2.5V are commonly utilized in the aerospace industry and medical sectors, but their low surface hardness and poor wear resistance may reduce their service life. In this study, CrN thin-film coatings were deposited on both alloys using RF magnetron sputtering to investigate the influence of substrate composition on coating microstructure, crystallinity, nano-mechanical behavior, and scratch resistance. The coatings were characterized using 3D profilometry, field emission scanning electron microscopy/energy dispersive x-ray spectroscopy, x-ray diffraction (XRD), nanoindentation, and scratch testing. Dense and uniform CrN coatings with an average thickness of approximately 460 nm were successfully obtained. XRD confirmed the formation of single-phase crystalline CrN, while the coating on Ti–3Al–2.5V exhibited broader diffraction peaks, indicating substrate-dependent differences in crystallinity. Nanoindentation showed a significant increase in hardness from 6.01 to 12.57 GPa for Ti–6Al–4V (109.2% increase) and from 6.77 to 16.49 GPa for Ti–3Al–2.5V (143.6% increase). The residual indentation depth decreased by 36.3% for CrN G5 and 61.3% for CrN G9, indicating improved resistance to permanent deformation and enhanced elastic recovery. Scratch testing demonstrated good adhesion for both coatings, with critical loads of 372 and 532 mN for CrN G5 and 428 and 476 mN for CrN G9. The coating on Ti–3Al–2.5V showed better resistance to crack initiation, whereas the coating on Ti–6Al–4V exhibited greater resistance to crack propagation. Overall, RF-sputtered CrN coatings significantly improved the nano-mechanical performance and scratch resistance of both titanium alloys, highlighting the important role of substrate composition in determining coating behavior.
