Crack detection and measurement in slab tracks are critical for maintenance decision-making. Pre-trained deep learning segmentation models often struggle with cracking instances due to domain adaptation and data scarcity. This study proposes an instance segmentation framework incorporating dynamic snake convolution (DSConv) modules, combined with an unsupervised style transfer-based boundary augmentation strategy. The DSConvenhanced architecture prioritizes linear crack features in cluttered backgrounds, while the augmentation introduces controlled perturbations to global pixels and local crack boundaries, generating structurally consistent diversified training samples. The results demonstrate that the deformable DSConv enhanced architecture achieves optimal mean average precision (mAP), improving segmentation performance by nearly 13 % compared to "fine-tuned Segment Anything Model". Its segmentation capability surpasses eight state-of-the-art models, especially for multiple intermittent microcracks. Furthermore, unsupervised style transfer-generated augmented data enhances crack instance segmentation performance by 10 % compared to non-augmented baselines, surpassing conventional methods including horizontal flipping and color jittering. Quantitative crack width distributions from segmentation-quantification analysis provide more comprehensive structural health insights than manual discrete-point measurements, facilitating precise maintenance decisions for railway infrastructure.

In this paper, we study Dirichlet problem on bounded open sets in R-d for non-local operators of the following form Lu = div(A(x)del u(x)) + b(x) . del u(x) + integral(d)(R) (u(y) - u(x))J(x,dy), where A(x) = (a(ij)(x))(1 <= i,j <= d) is a measurable d x d matrix-valued function on R-d that is uniformly elliptic and bounded, b is an R-d-valued function so that |b|(2) is in some Kato class Kd, J(x, dy) is a finite measure on R-d for each x is an element of R-d so that x bar right arrow J(x, R-d) is in the Kato class K-d. We show there is a unique Feller process X having strong Feller property associated with L, which can be obtained from the diffusion process having generator div(A(x)del u(x)) + b(x) . del u(x) through redistribution. We further show that for any bounded connected open subset D subset of R-d that is regular with respect to the Laplace operator Delta and for any bounded continuous function phi on D (c), the Dirichlet problem Lu = 0 in D with u = phi on D-c has a unique bounded continuous weak solution on R-d. This unique weak solution can be represented in terms of the Feller process associated with L.

In this paper, the authors study the homogenization of the third boundary value problem for semilinear parabolic PDEs with rapidly oscillating periodic coefficients in the weak sense. Their method is entirely probabilistic, and builds upon the work of Tanaka (2020) and Buckdahn (1999). Backward stochastic differential equations with singular coefficients play an important role in this approach.

Railway fastener systems necessitate regular inspections to uphold the safety standards of high-speed trains. Previously, the capture of geometry characteristics and evaluation of fastener tightness relied on costly structured light cameras, falling short of meeting the growing demand for rapid and cost-effective detection. This study introduces a novel approach that amalgamates instance segmentation and monocular depth estimation, enabling fastener tightness inspection using a standard camera. The proposed method entails the following steps: Firstly, leveraging an enhanced ZoeDepth model, absolute depth is inferred from a single railway structure image to extract the vertical spatial features of the fastener system. Secondly, the YOLOv8 network is deployed to delineate the fastener elastic clip and bolt in the railway structure images, producing masks that facilitate depth distribution computation. Thirdly, by fusing the absolute depth maps and masks, apparent depth distribution features are computed utilizing the proposed metrics. These features undergo analysis and comparison with an online updated threshold library, facilitating the identification of loose fasteners. In this study, the collected Railway Structure Intensity-Depth dataset was used for model training, while on-site experiments were conducted to evaluate the accuracy of the proposed method. The experimental findings demonstrate that this method adeptly identifies loose fasteners, achieving a detection rate of 86.2 %.

Currently, object detection-based rail fastener defect detection methods still face challenges such as limited detection categories, insufficient accuracy, and high computational complexity. To this end, the YOLOv8n-FDD, an advanced multi-category fastener defect detection model designed upon the YOLOv8n with comprehensive optimizations is developed in this paper. Concretely, by introducing the CUT-based style transfer model to generate diverse defect samples, the concern due to imbalanced distribution of sample categories is effectively alleviated. The CA mechanism is incorporated to enhance the feature extraction capability, and the bounding box loss function is further upgraded to improve the model’s generalization performance. With respect to efficiency, the Conv and c2f modules of the YOLOv8n model are, respectively, replaced with the GSConv and VoVGSPCP modules, accordingly achieving a lightweight design. Comparative experimental results demonstrate that the presented YOLOv8n-FDD model outperforms several classic object detection models in terms of detection accuracy, detection speed, model size, and computational complexity.

Cooperative control for virtual coupling systems of multiple heavy-haul trains can improve the safety and efficiency of heavy-haul railway transportation. However, the false data injection attack for the virtual coupling system is a serious obstacle, which will lead to imprecise train operation control. To address this issue, a deep learning-based false data injection attack (FDIA) detection for virtual coupling systems of heavy-haul trains is proposed. First, the cyber-physical model of the virtual coupling system is established. Second, a cooperative control law is designed for the virtual coupling system, and the effects of the FDIA on the virtual coupling system is analyzed. Then, the unsupervised autoencoder method is introduced to achieve the false data injection attack detection. The autoencoder network model is trained with normal operation data and tested with abnormal operation data. The performance of the proposed method is verified in four different simulation scenarios: normal case, velocity attack case, position attack case, and joint attack case. Simulation results show that the proposed method can effectively increase the detection accuracy and reduce the error rate with other supervised methods.

Deep reinforcement learning has become a promising method for the energy management of electric vehicles. However, deep reinforcement learning relies on a large amount of trial-and-error training to acquire near- optimal performance. An adversarial imitation reinforcement learning energy management strategy is proposed for electric vehicles with hybrid energy storage system to minimize the cost of battery capacity loss. Firstly, the reinforcement learning exploration is guided by expert knowledge, which is generated by dynamic programming under various standard driving conditions. The expert knowledge is represented as the optimal power allocation mapping. Secondly, at the early imitation stage, the action of the reinforcement learning agent approaches the optimal power allocation mapping rapidly by using adversarial networks. Thirdly, a dynamic imitation weight is developed according to the Discriminator of adversarial networks, making the agent transit to self-explore the near-optimal power allocation under online driving conditions. Results demonstrate that the proposed strategy can accelerate the training by 42.60% while enhancing the reward by 15.79% compared with traditional reinforcement learning. Under different test driving cycles, the proposed method can further reduce the battery capacity loss cost by 5.1%-12.4%.

Accurate detection and quantification of damage to railway fasteners are crucial for ensuring railway safety. The spatial damage defects caused by the complex shape of fasteners and the problem of data imbalance in actual scenarios are significant challenges faced by deep learning models. This study innovatively proposes a railway-fastener point cloud analysis method based on deep learning as follows: (1) use four cameras to capture three-dimensional point cloud data and construct a virtual negative sample supplementary data set, (2) develop Rail-Swin3D models for precise segmentation of fastener components, and (3) introduce quantitative indicators to objectively evaluate the damage situation. A data set containing 120 real and virtual damaged fasteners was ultimately constructed, achieving up to 99.35% mean intersection over union (mIoU) in point cloud segmentation tasks. This study not only improves the efficiency of railway safety detection, but also opens new paths for the application of point cloud data in the field of railway maintenance, with profound theoretical and practical value.

The hysteresis effect represents the difference in open circuit voltage (OCV) between the charge and discharge processes of batteries. An accurate estimation of open circuit voltage considering hysteresis is critical for precise modeling of LiFePO4 batteries. However, the intricate influence of state-of-charge (SOC), temperature, and battery aging have posed significant challenges for hysteresis modeling, which have not been comprehensively considered in existing studies. This paper proposes a data-driven approach with adversarial learning to model hysteresis under diverse conditions, addressing the intricate dependencies on SOC, temperature, and battery aging. First, a comprehensive experimental scheme is designed to collect hysteresis dataset under diverse SOC paths, temperatures and aging states. Second, the proposed data-driven model integrates a conditional generative adversarial network with long short-term memory networks to enhance the model’s accuracy and adaptability. The generator and discriminator are designed based on LSTM networks to capture the dependency of hysteresis on historical SOC and conditional information. Third, the conditional matrix, incorporating temperature, health state, and historical paths, is constructed to provide the scenario-specific information for the adversarial network, thereby enhancing the model’s adaptability. Experimental results demonstrate that the proposed model achieves a voltage error of less than 3.8 mV across various conditions, with accuracy improvements of 31.3 – 48.7% compared to three state-of-the-art models.

In this paper, we study a new type of SPDEs with reflection (called mean reflected stochastic partial differential equations (SPDEs)), where the compensating reflection part depends not on the paths but on the law of the solution. Focusing on solutions (u, K) with deterministic K, we obtain the well-posedness of such SPDEs. Utilizing the weak convergence approach, we then establish large deviation principles for this type of SPDEs.

In next-generation networks, distributed clients collaborate to generate an aggregated global model tailored for various vertical applications. However, this convenience comes at the cost of potential privacy risks, as personal information may be exposed within the global model aggregation process. In response, the right to be forgotten was introduced, granting individuals the right to withdraw their consent for the processing of their personal information. To address this challenge, machine unlearning has been developed, enabling models to erase any memory of private data. Previous approaches, such as retraining or incremental learning, often require additional storage or are difficult to implement in neural networks. Our method, by contrast, introduces a small perturbation to the model's weights, guiding it to iteratively move towards a model trained only on the remaining data subset until the contribution of the unlearned data is completely removed. In our approach, machine unlearning is conceptualized as a process that iteratively adjusts the initial model to remove any trace of the forgotten data. Our key contribution is the introduction of a reference model, trained on a subset of the remaining data, which guides the target unlearning model toward successfully forgetting the data. Additionally, we discuss two evaluation methods-membership inference and backdoor evaluation-that effectively assess the success of our machine unlearning approach. These methods verify whether the private data has truly been forgotten by the target unlearning model. Through experiments on five datasets, we demonstrate the effectiveness of our approach, which is 15x faster than the traditional retraining method.

Accurate detection and estimation of railway fastener tightness are vital for rail infrastructure safety and reliability. Traditional methods depend on manual annotation tools like Label Me, which are error-prone, laborintensive, and costly. Additionally, monocular depth estimation and instance segmentation involve complex computations that challenge real-time implementation, particularly on resource-constrained platforms. This study introduces a novel three-phase solution using the Multimodal Geometric Autoencoder (MGAE) for fastener tightness detection, integrating point clouds with monocular-depth-guided multimodal data. Our approach utilizes a hybrid autoencoder for high-quality feature extraction, enabling precise tightness estimation. Employing unsupervised learning, MGAE eliminates the need for labeled data, thus reducing labor and costs. The framework integrates point clouds, mesh, monocular depth, and 2D images, with various fusion blocks enhancing feature extraction accuracy and computational efficiency. Post-feature extraction, classical techniques such as isolation forest, stress-strain, and elastic potential energy methods assess fastener tightness.

Crack detection at the pixel level across complex scenarios (structural interference and adverse working conditions) is a critical consideration in the maintenance of ballastless track slab (BTS). Although existing deep learning models achieve acceptable accuracy on cracks with a monotonous background, the ground truth with high labor cost is inevitable and their performance in complex scenarios may fall far below their theoretical bounds or even cause “all-black images.” A hybrid algorithm based on synthetic data from digital twin model and weakly supervised style transfer is proposed in this paper for addressing the above challenges. The algorithm uses a region-attention strategy to enable the uncontrolled generative adversarial network (GAN) focusing its attention on weak labels containing crack regions, directly obtaining segmentation results with the same style as the ground truth of the crack forest dataset. In addition, a digital twin model that can simulate the real inspection working conditions is established to generate a synthetic crack dataset, enabling the hybrid algorithm to extract the most discriminative features. The results show that the performance of the hybrid algorithm on inspection images across complex scenarios is nearly 25% higher than that of the DeepLabv3+ network, while the time cost consumed is only 0.5% of the latter. The deployment of the region attention strategy also enables the hybrid algorithm to achieve a mean intersection of union (MIoU) of 79.38%, which is nearly twice as much as that of GAN. It not only eliminates the oversegmentation caused by structures such as rails and fastener systems but also overcomes “all-black images.” In addition, synthetic data can greatly enlarge the range, type, and number of discriminative crack features compared with data augmentation based on limited real data, thus enhancing the performance of the hybrid algorithm for uncertain inspection data. Particularly, the fully trained hybrid algorithm based on the synthetic dataset shows good adaptability and generalization to adverse working conditions such as uneven lighting, noise, and blur.

This paper studies stability conditions for neural networks with stochastic impulsive intensity and impulsive density. Firstly, the pth moment exponential stability conditions are given, where the impulsive intensities of impulsive neural networks are independent and identically distributed, and the impulsive sequence is defined by impulsive density. Then, taking a linear time-varying impulsive system as an example, it is shown that impulsive sequence defined by average impulsive interval is sometime insufficient in terms of stabilizing such linear time-varying systems. Finally, two examples are presented to verify that impulsive density method is more effective and universal than average impulsive interval method when dealing with time-varying systems.

The cooperative control of the virtual coupling system for heavy-haul trains relies on information interaction with each train, resulting in various forms of network attacks posing a huge threat to the reliable transmission of information and safe and efficient operation of the virtual coupling system for heavy-haul trains. To address this problem, this research introduces an reservoir computing-Based method to investigate the strategy for detecting false data injection attacks (FDIAs) in the heavy-haul train virtual coupling system. Initially, a cyber-physical model of the virtual coupling system is established. Subsequently, a cooperative control law is devised for the virtual coupling system, and the potential impact of FDIAs on the system is analyzed. The proposed method learns to reconstruct normal signal data and identifies anomalies according to reconstruction errors. Its effectiveness in identifying attacks on the virtual coupling system for heavy-haul trains is then validated through simulations.

The distributed control-based DC microgrid relies on communication networks for information exchange, rendering it susceptible to cyber-attacks and intrusions. Attack detection is crucial to ensure the safety of microgrids. However, it remains a challenge for existing detection methods to distinguish state changes caused by load changes or attacks. To address the issue, this paper proposes a distributed attack detection method that combines consistency and stationarity features to assist a resilient recovery strategy. Firstly, the layered architecture of an interconnected DC microgrid is introduced, and the impact of false data injection attacks on microgrids is analyzed. Then, the attack detection is decomposed into two phases, anomaly detection and attack identification. In the first phase, consistency and stationarity features are extracted to indicate whether the microgrid is abnormal. In the second phase, the correlation between voltage and current characteristics is analyzed to determine the attack types. Furthermore, an event-triggered resilient recovery learning control protocol is proposed, which estimates the normal states and reconstructs normal control inputs through two trained neural networks to ensure the operation safety of the DC microgrid. The experimental results validate the effectiveness of the proposed attack detection and resilient recovery strategy.

Thermal runaway of lithium-ion batteries is one of the key challenges hindering the development of electric vehicles. Realizing timely fault detection in battery systems is of great significance for preventing thermal runaways and safeguarding people’s lives and properties. As it is difficult to obtain fault battery datasets in the real world, there is a need to develop novel fault detection methods that can operate with normal data. In this paper, we propose an optimized Gated Recurrent Unit autoencoder architecture that integrates the sparse representation technique to detect battery faults in electric vehicles. Firstly, the Gated Recurrent Unit is employed to efficiently learn the information in battery data from normal electric vehicles. Then, the autoencoder utilizes the sparse representation technique to improve its ability to recognize abnormal data by learning a set of basis vectors that can sparsely represent normal data. Finally, the reconstruction errors between the original and reconstructed vectors are calculated in a sliding window and compared to the threshold to detect the fault. The effectiveness of the proposed method is verified on a real operating dataset including two normal electric vehicles and two faulty electric vehicles. The results show that it can provide early alarm time and reduce the probability of false alarms.

Fastener damage detection is an integral component of track safety inspections. The lack of balanced dataset caused by insufficient data on defective fasteners poses a significant challenge to the current development of robust fastener detection models. This study proposes the YOLOv8-FAM detection model algorithm, which combines the enhanced capabilities of YOLOv8, and generates realistic images of defective fasteners by performing style transfer on masked images of non-defective fasteners, thereby creating a balanced dataset for training the detection model. Experimental results show that the defect detection accuracy of YOLOv8-FAM improves by 8% compared to the original model, while also reducing the data acquisition cost by 40%. The YOLOv8-FAM model surpasses existing models in detecting defective fasteners while minimizing inference costs to the maximum extent. The proposed style transfer data synthesis method drives the practical deployment of deep learning, offering an efficient and cost-effective solution for the transportation infrastructure industry.

Batteries play a crucial role in electric vehicles since they are the only power sources of the vehicles. In order to guarantee that batteries can work efficiently and safely, battery management systems (BMS) are employed to measure, estimate and regulate battery states during the operation of electric vehicles. To do that, numerous voltage, current, and temperature sensors are required to be installed in the BMS. However, a large number of sensors may lead to some problems, e.g., high cost, reduced space, low efficiency, and high failure rates. To address these challenges, in this paper, a digital twin paradigm is proposed for the BMS to estimate and predict the battery states with only a voltage sensor. A multi-linear regression algorithm is utilized to build the regression model between battery voltage and the other variables. Experiment results show that the proposed digital twin model achieves over 90% prediction accuracy in practical applications.

Battery degradation is a main hinder to extend the persistent lifespan of the portable heterogeneous computing device. Excessive energy consumption and prominent current fluctuations can lead to a sharp decline of battery endurance. To address this issue, a battery-aware workflow scheduling algorithm is proposed to maximize the battery lifetime and release the computing potential of the device fully. Firstly, a dynamic optimal budget strategy is developed to select the highest cost-effectiveness processors to meet the deadline of each task, accelerating the budget optimization by incorporating deep neural network. Second, an integer-programming greedy strategy is utilized to determine the start time of each task, minimizing the fluctuation of the battery supply current to mitigate the battery degradation. Finally, a long-term operation experiment and Monte Carlo experiments are performed on the battery simulator, SLIDE. The experimental results under real operating conditions for more than 1800 hours validate that the proposed scheduling algorithm can effectively extend the battery life by 7.31%-8.23%. The results on various parallel workflows illustrate that the proposed algorithm has comparable performance with speed improvement over the integer programming method.