ISSN 1673-0291 CN 11-5258/U
The stable operation of switch machines is a key guarantee for the safety of High-Speed Railways (HSRs). With the increasing demand for intelligent railway systems, higher requirements are imposed on the precise perception and autonomous diagnosis of switch machine working conditions. To overcome the limitations of traditional methods in terms of diagnostic accuracy, computational real-time performance, and anti-interference capability, this paper constructs an intelligent diagnosis model that integrates cross-modal feature attention mechanism and Transformer-XL recursive memory mechanism. The proposed model enhances fault recognition and environmental adaptability under complex operating conditions. By introducing a cross-modal attention mechanism, it enables dynamic interaction between power curve signals and switch rail vibration signals, mitigating the judgment bias caused by missing single-modal information. The Transformer-XL with recursive memory mechanism dynamically adjusts the model’s perception of historical information, allowing it to extract state information across time windows. Additionally, 1D-CNN is incorporated for short-term dynamic feature extraction, optimizing global sequential modeling while improving noise robustness and reducing computational complexity. Experimental results demonstrate that the proposed model exhibits significant advantages in cross-modal feature representation, long-term dependency modeling, noise robustness, and computational efficiency. This study provides an intelligent diagnostic solution for HSR maintenance, featuring high efficiency, low resource consumption, and strong environmental adaptability. It facilitates the transition from reactive maintenance to predictive maintenance, thereby improving operational safety margins and robustness.
To address the current limitation that switch machine health monitoring predominantly relies on power signals while vibration signals remain underutilized for fault diagnosis, this study proposes a fault diagnosis model that integrates Variational Mode Decomposition (VMD) optimized by Tuna Swarm Optimization (TSO) and Least Squares Support Vector Machine (LSSVM) optimized by the Crested Porcupine Optimizer (CPO). First, vibration data from eight typical operating conditions of the ZD6 switch machine are collected through experiments. TSO is employed to optimize VMD and determine the optimal number of decomposition layers k and the penalty factor α, after which the optimized VMD decomposed the vibration signals into several Intrinsic Mode Functions (IMF). Second, IMFs are selected using a dual screening criterion based on envelope entropy and kurtosis, and the signals are reconstructed. The Refined Composite Multiscale Diversity Entropy (RCMDE) of the reconstructed signals is then extracted. Third, the RCMDE features are divided into training and testing data and used as feature vectors for LSSVM, which is configured with the optimal combination of penalty factor γ and kernel function parameter σ obtained via CPO optimization. Finally, accuracy, macro-precision, macro-recall and macro-F1 are adopted as evaluation metrics for comparative analysis across multiple models. The results show that the proposed TSO-VMD-RCMDE-CPO-LSSVM, as the fault diagnosis model of switch machine, achieves an average runtime of 20.4 s over 10 iterations. The average accuracy of the training data is 99.68% with a standard deviation of 0.12. The average accuracy of the testing data is 99.25% with a standard deviation of 0.07. Compared with other models, it attains the highest mean macro-precision, macro-recall and macro-F1 scores, along with the lowest standard deviations. These findings demonstrate the superior performance and feasibility of the proposed model for switch machine fault diagnosis.
To address the widespread challenges of limited data samples and low diagnostic accuracy in existing switch machine fault diagnosis, this study takes the action power curve, a critical time-series signal of the S700K switch machine, as the research object and proposes a fault diagnosis model based on GAN-BO-BiGRU. First, a small number of power data samples collected by the Centralized Signalling Monitoring system (CSM) are fed into a Generative Adversarial Network (GAN). Through adversarial training between the generator and the discriminator, more sample data are generated to address the issue of data scarcity. Second, a BO-BiGRU fault diagnosis algorithm model is established. The Bayesian Optimization (BO) algorithm is used to determine the optimal values of key hyperparameters for the Bidirectional Gated Recurrent Unit (BiGRU) model, including the number of hidden-layer neurons, the initial learning rate, and the L2 regularization parameter, thereby obtaining the optimal hyperparameter combination. By exploiting BiGRU’s capability to capture information bidirectionally, the proposed model more comprehensively mines patterns from the time-series power data of the switch machine. Finally, simulations are conducted using both the generated data and the original data as samples. The simulation results demonstrate that the data generated by GAN exhibits minimal difference from the original data and can effectively serve as an augmented dataset for fault diagnosis. Moreover, compared to the Long Short-Term Memory (LSTM) model, the BO-BiGRU fault diagnosis model improves the F1 score by 1.77%, indicating its superior ability to extract fault features and its effectiveness in enhancing the accuracy of switch machine fault diagnosis.
To address the challenge of extracting fault features from switch machine vibration signals under varying noise conditions, this study proposes a denoising method combining Crested Porcupine Optimization (CPO)-optimized Variational Mode Decomposition (VMD) with Stationary Wavelet Transform Total Variation (SWTTV). First, the vibration signal is decomposed into multiple Intrinsic Mode Functions (IMF) using CPO-optimized VMD. Second, a hybrid criterion combining correlation coefficient and kurtosis is employed to select relevant IMFs. The selected IMFs are then denoised and reconstructed using the SWTTV algorithm. Finally, the method’s performance is evaluated using both simulated test signals and field-collected switch machine vibration signals. Experimental results demonstrate that under different Signal Noise Ratios (SNR), compared to the VMD-WT algorithm, the proposed method achieves an improvement in output SNR of approximately 1 to 6 dB, a reduction in Root Mean Square Error (RMSE) by about 0.03, and an increase in the correlation coefficient with the original signal by approximately 0.01. Furthermore, the proposed algorithm effectively preserves signal features under various operating conditions, avoiding signal distortion. The proposed algorithm exhibits strong generalization capability and robustness, providing a theoretical basis for feature extraction and fault diagnosis of switch machine vibration signals.
To address the low efficiency of manual inspection and the difficulty of processing massive data when detecting leaky cable clamps in high-speed railway tunnels, this study proposes a detection method based on an improved YOLOv5 model. First, a Ghost Spatial Convolutional Spatial Pyramid (GSCSP) module is designed, where Ghost convolution replaces standard convolution to reduce feature redundancy and achieve model lightweighting. Second, a lightweight Efficient Channel Attention Network (ECANet) is integrated to strengthen the model’s ability to distinguish clamp features from complex tunnel backgrounds and improve the detection accuracy of small objects. Then, a structured channel pruning strategy is applied, in which redundant channels are pruned based on the scaling factors of Batch Normalization (BN) layers, achieving a lightweight architecture while maintaining model accuracy. Finally, a dataset covering various states of leaky cable clamps under real high-speed railway operation scenarios is constructed. Data diversity is enhanced through methods such as Gaussian noise addition and data stitching, providing a richer and more robust sample foundation for subsequent model training. Experimental results show that the improved model reduces the number of parameters by 72.1% while preserving detection accuracy and real-time performance. The findings provide a reference for the intelligent operation and maintenance of railway communication equipment.
To address the complex and diverse characteristics of track fastener defects, as well as the low efficiency and high missed-detection rates of traditional detection methods, this study proposes a lightweight detection model, FPSI-YOLOv8s, based on the YOLOv8s framework. First, to reduce model complexity, FasterNet, featuring higher processing speed and fewer parameters, is adopted to replace the CSPDarkNet53 backbone in YOLOv8s for defect feature extraction. Second, the C2f module in the YOLOv8s neck is redesigned using Position-aware Recurrent Convolution(ParConv) to form a new FasterBlock module, enabling multi-scale feature fusion and further model lightweighting. Third, a Spatial Group-wise Enhance (SGE) attention mechanism is integrated after the SPPF layer to enhance the model’s sensitivity to defect features and mitigate accuracy degradation. Finally, the Inner-IoU loss function replaces CIoU to improve detection performance for objects of varying scales and shapes, while refined quality evaluation and gradient-gain strategies further enhance model robustness. Experimental results show that the improved model reduces model size by 29.78%, and decreases computational cost and parameter count by 29.93% and 30.46%, respectively, with only a 0.7% decrease in detection accuracy. These results demonstrate that the proposed model achieves significant lightweighting and improved operational efficiency while maintaining high accuracy, indicating strong application potential for rapid inspection of track fasteners.
To address the prevalent false positives and missed detections of railway fasteners in complex scenarios such as switch machines and turnouts for railway fastener inspection tasks, this paper proposes a railway fastener condition detection method based on an improved YOLOv9. First, to overcome the challenge of extracting features from fastener regions in complex scenes, the Large Selective Kernel (LSK) attention mechanism is integrated with the RepNCSPELAN4 module. This optimization enhances the performance of the feature extraction module, enabling more effective capture of critical feature information in fastener regions and improving the model’s adaptability to diverse scenes and targets. Second, to better distinguish subtle details of confusing fastener damage states, a feature fusion network based on Space to Depth (SPD) convolution is developed, thereby increasing accuracy in low-resolution and small-object detection and ensuring precise identification of fastener damage states even against complex backgrounds. Third, Shape IoU is adopted as the new loss function to more accurately measure the overlap between predicted and ground-truth bounding boxes, endowing the model with greater robustness and superior target localization precision. Finally, to validate the effectiveness of the proposed method, a real-world railway fastener dataset encompassing complex operating conditions is collected and constructed, and comprehensive comparative experiments are conducted. Experimental results demonstrate that the proposed method effectively detects railway fastener conditions, achieving a 1.3% improvement in detection accuracy over the baseline model, while reducing the false detection rate by 0.7% and the missed detection rate by 1.4%. This enhances the reliability and stability of railway fastener condition detection in complex scenarios.
To address the ill-posedness of transfer matrices and the poor noise robustness encountered in load identification and response reconstruction, this study proposes a joint regularization method that integrates Modified Truncated Randomized Singular Value Decomposition (MTRSVD) with Joint Bidiagonalization QR (JBDQR). By alleviating matrix ill-posedness and reducing the impact of measurement noise on identification results, the method can identify unknown loads and reconstruct responses at unmeasured locations based solely on limited measurement information. First, the structural dynamic equations are derived, and a state-space model and transfer matrix are established to formulate the load identification and response reconstruction problem. Then, MTRSVD is used to preprocess the transfer matrix, and a randomized projection technique is used to reduce matrix dimensionality, while an adaptive truncation criterion preserves the dominant features, thereby alleviating the ill-posedness of the transfer matrix and reducing the influence of measurement noise. Subsequently, the JBDQR algorithm is introduced for load identification, in which unknown loads are solved via iterative regularization through a joint bidiagonalization process, and responses at unmeasured locations are reconstructed using the corresponding transfer matrices. Finally, the proposed method is validated through the numerical example of a 3 kW small wind turbine blade and experimental tests on a simply supported beam. The results indicate that the proposed method can still effectively achieve load identification and the response reconstruction at unmeasured locations under a noise level of 15%.
To address the susceptibility of ground-to-train information transmission in track circuits to interference from traction currents in electrified railway traction power supply systems, this study employs finite element simulation software to develop a ground-to-train information transmission model and investigates the electromagnetic field characteristics of the information transfer process between the track circuit and the Track Circuit Reader (TCR) antenna. First, a simulation model of ground-to-train information transmission is established based on the operational principles and installation configuration of the track circuit system. Second, the model’s validity is confirmed by using track circuit signals as the excitation source and comparing key parameters of the TCR antenna’s induced voltage, including peak value, carrier frequency, and low-frequency modulation. Finally, measured traction currents are applied as excitation sources to analyze the electromagnetic interference effects on ground-to-train transmission under both balanced and unbalanced traction current conditions. Simulation results indicate that with measured data as the excitation source, when traction current in the rails is balanced, the peak induced voltage at the TCR antenna is approximately 1 mV, and traction harmonic currents cause no interference to the track circuit system. However, when the peak value of the unbalanced current in the rails reaches 1.1 A, the maximum peak induced voltage of the TCR antenna reaches 220 mV, generating interference signals within the operating frequency band of the track circuit. These findings provide valuable theoretical guidance and practical reference for electromagnetic compatibility design of electrified railway signaling systems and for enhancing the anti-interference performance of TCR equipment.
To address the issue of reduced control accuracy in high-speed train speed tracking caused by system susceptibility to internal and external disturbances, this study proposes an Active Disturbance Rejection Control (ADRC) scheme for high-speed train speed tracking based on fractional-order integral sliding mode. The scheme introduces improvements to both the Linear Extended State Observer (LESO) and the nonlinear error feedback control law in ADRC. First, in the LESO design, a differential state variable of the total disturbance is incorporated to enhance the disturbance estimation capability of the observer. Second, Fractional Order Integral Sliding Mode Control (FOISMC) is employed to improve the nonlinear error feedback control law, thereby mitigating chattering in sliding-mode control while enhancing tracking accuracy. Finally, a composite fractional-order integral sliding mode ADRC scheme is developed and applied to track a desired speed profile in simulations using CRH3 train parameters. The proposed scheme is compared with other traditional control methods to verify the tracking performance of the control scheme. The results demonstrate that, under identical conditions and external disturbances, the proposed control scheme achieves higher tracking accuracy and stronger disturbance rejection capability than other control schemes, with a maximum speed-tracking error of 0.000 05 m/s.
To address the insufficient adaptability of fixed signal timing in dynamic traffic flow environments, this study proposes an improved Gradient-Based Meta-Learning Deep Q-Network (GBML-DQN) algorithm integrating a meta-learning mechanism for adaptive signal control at isolated intersections. First, the state space is constructed based on lane density, the signal phase action space is defined, and a multi-objective weighted reward function is designed. Second, using DQN as the base architecture, the discount factor γ is dynamically adjusted via a meta-gradient strategy. Third, a Dueling structure is used to decouple state values from action advantages, and NoisyNet is employed to replace the traditional ε-greedy strategy. Consequently, two improved algorithms are constructed: Improved GBML-DQN-Dueling (I-GD-D) and Improved GBML-DQN-Noisy (I-GD-N). Finally, experimental validation is conducted on the SUMO simulation platform across high, medium, and low traffic volume scenarios. The experimental results indicate that I-GD-N exhibits superior robustness and adaptability across different traffic scenarios. Specifically, under medium traffic conditions using the Stochastic Gradient Descent (SGD) optimizer, the average delay is reduced to 20.55 s, representing an improvement of approximately 20% compared to DQN. While DQN exhibits higher stability than GBML-DQN in medium and low traffic scenarios due to its simpler structure, it suffers from significant policy degradation in high traffic scenarios or when using the Root Mean Square Propagation (RMSprop) optimizer, performing worse than even fixed signal timing. The dynamic γ adjustment mechanism adaptively optimizes based on traffic intensity; in high traffic scenarios, it significantly outperforms fixed γ strategies by reducing γ to rapidly respond to congestion. The research results provide a valuable reference for urban intersection signal control.
Highway traffic speed is jointly influenced by multiple external factors, including adjacent land-use types, weather conditions, and traffic volume, and exhibits nonlinear spatiotemporal variations. To address this, this paper proposes WP-STGCN-GRU, a multi-step short-term traffic speed prediction model that integrates external attributes with spatiotemporal features. First, a spatiotemporal prediction framework is constructed using the Spatial-Temporal Graph Convolutional Network (STGCN) and Gated Recurrent Unit (GRU). Adaptive weighted adjacency matrices are employed to encode spatial relationships among highway segments, capturing the spatiotemporal dependencies of traffic speed. Second, weather conditions and Points of Interest (POI) features are incorporated into a speed attribute enhancement unit, which expands the feature dimensionality and strengthens the representation of external factors in speed variation patterns, enabling more accurate traffic speed predictions. Finally, model performance is evaluated through baseline comparisons and ablation studies. Experimental results indicate that dynamically integrating external attributes such as weather and POIs significantly improves both single-step and multi-step traffic speed predictions. Compared with baseline models, the proposed model reduces Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) by over 14.1% and 14.8%, respectively, for single-step prediction, and by over 4% and 14%, respectively, for multi-step prediction. In a 15-minute forecasting task, fusing both external attributes reduces MAE and RMSE by 23.0% and 17.8%, respectively, with weather information contributing more prominently and demonstrating clear complementarity between two factors. These findings provide valuable insights for highway speed prediction and offer guidance for improving traffic safety and intelligent highway management.
Pedestrian intrusion into railway perimeters poses a serious threat to train safety. To address the high false-negative rate of single-image pedestrian detection algorithms based on deep learning in complex railway environments, this paper proposes a railway perimeter intrusion detection algorithm that combines YOLOv5x single-frame detection results with pedestrian trajectory correlation in image sequences. First, high-confidence pedestrian targets from single-frame images are integrated with three types of features, including spatial relationships of trajectories, trajectory irregularity information, and matching results between detections and trajectory predictions, to achieve reliable pedestrian trajectory updating and filtering. Second, the trajectory-correlated prediction results are used to reset the confidence scores of low-confidence pedestrian targets in single-frame images, thereby improving detection performance in challenging railway environments such as nighttime and adverse lighting conditions. Furthermore, a pedestrian intrusion dataset covering five typical railway scenarios is constructed, including cases prone to missed detections, such as occluded objects, blurred targets, and distant small objects under low-light or glare conditions. Finally, to further validate the effectiveness of the proposed algorithm, Faster R-CNN and YOLOv3 are employed as detectors to assess the performance improvement. Experimental results demonstrate that, compared to the single-frame detection algorithm YOLOv5x, the proposed method improves recall by 4.5% and reduces the log-average miss rate by 4.0% without significant loss in detection precision, proving the efficacy of the trajectory updating, filtering, and confidence reset modules. Additionally, compared to Faster R-CNN and YOLOv3, the proposed algorithm achieves recall improvements of 2.3% and 3.6%, respectively, and reduces log-average miss rates by 1.3% and 3.5%, confirming its adaptability to different detection models.
To address the limitations of existing road defect detection algorithms, such as low accuracy in complex backgrounds, limited generalization capability, and frequent missed detections of small objects, this study proposes an improved YOLOv8-based detection algorithm. First, a Coordinate Attention (CA) mechanism is integrated into the backbone network layer to introduce positional information, enabling the model to better capture spatial dependencies and enhancing its feature discrimination ability under complex background conditions. Second, the Path Aggregation Network (PANet) in the neck network layer is replaced with a weighted Bi-directional Feature Pyramid Network (BiFPN). By incorporating bidirectional connections and learnable weights, the network facilitates bidirectional information flow across different resolution levels, leading to more effective fusion of low-level positional features with high-level semantic features and improving multi-scale feature representation. Finally, small-object feature maps are introduced to more accurately capture the small-object characteristics and reduce missed detections, thereby improving detection precision. Experimental results show that on the RDD2022 road defect dataset, the improved algorithm increases mean Average Precision (mAP) by 3.1% compared to the original version, while reducing model parameters by 2.3%, achieving more accurate and rapid road defect detection.
To address the challenges in time series anomaly detection tasks, including label scarcity, high false positive and false negative rates, and model performance vulnerability to seasonal and long-term trend variations, this paper proposes a Time-Frequency Cross-Fusion-Guided Dual-Stream Clustering Anomaly Detection Method (TFCDC). First, time-domain and frequency-domain feature vectors are extracted separately, and a cross-fusion mechanism is employed to enable complementary modeling of multi-scale information, yielding a joint time-frequency feature representation. Second, the time-frequency features are encoded using a Long Short-Term Memory (LSTM) network and a linear transformation layer to produce two latent variables. Third, a two-stage clustering strategy is applied in the latent spaces to aggregate normal samples while distinguishing anomalous ones, which effectively reduces false positive and false negative rates. Finally, comparative experiments are conducted against state-of-the-art baseline models on six benchmark anomaly detection datasets. The experimental results demonstrate that the proposed TFCDC model outperforms mainstream baselines on multiple dataset, achieving an average F1-score improvement of 14.5% across the six datasets and a notable 32.8% improvement on the ultra-high-dimensional, ultra-long-sequence WADI dataset compared to the baseline. These findings confirm that TFCDC exhibits superior accuracy and robustness, effectively mitigating the interference from long-term trend variations and anomalous data.
Under complex rainy conditions, U-shaped encoder-decoder networks struggle to remove similar rain streaks appearing at multiple scales and often suffer from the loss of fine-grained details. To address these issues, this paper proposes a Skip Fusion Connection Network (SFCNet) designed to enhance multi-scale feature transmission between the encoder and decoder. First, Skip Fusion Connection Block (SFCB) is designed with multi-scale inputs and outputs. By cascading Cross-scale Features Fusion Blocks (CFFB), this module globally fuses output features across different encoder levels. Second, a gating mechanism is integrated to dynamically optimize the attention weights of multi-scale features as they are transmitted to the decoder. This enables each decoder layer to learn rain-context information that matches its specific resolution, significantly improving the removal of dense, overlapping, and multi-scale rain streaks. Finally, to better restore structural details heavily occluded by complex rain patterns, a Detail Enhancement Block (DEB) is proposed to enhance and fuse original image details at different scales back into the network. Experimental results on standard datasets such as Rain200L and DID-Data demonstrate that SFCNet achieves an average PSNR improvement of 0.24 dB over the IDT method. These findings provide an effective approach for single-image deraining under complex rain conditions.
In Integrated Sensing and Communication (ISAC) networks, network topology design plays a decisive role in both communication and sensing performance. To provide general guidance for system design, Stochastic Geometry (SG) is employed to mathematically model and analyze network deployment and performance from a network-layer perspective. The spatial distribution of ISAC base stations is modeled using Matern Hard Core Point Processes (MHCPP). A time-division multiplexing mechanism between communication and sensing is considered. Taking a typical user as the object of study, the user is associated with the nearest base station. Performance metrics for communication and sensing are established, and lower bounds on the average interference power and average throughput experienced by the typical user and the associated base station are derived. The accuracy of the derived results is further verified through simulations. Simulation results demonstrate, compared with Poisson Point Processes (PPP), MHCPPs significantly reduce the average interference power: by approximately 30%-43% in communication scenarios and 91%-93% in sensing scenarios. As the minimum inter-base-station distance increases, the average communication throughput increases, whereas the average sensing throughput first increases and then decreases. In practical scenarios, ISAC network parameters should be designed in accordance with specific system requirements. The research results can provide valuable references for system design and scalability analysis.
The Industrial Internet of Things (IIoT) has been widely adopted in the intelligent transformation of factories due to its flexibility and cost-effectiveness. To ensure delay determinacy for Time-Sensitive (TS) services and high throughput for Real-Time (RT) services in industrial 5th Generation Mobile Communication (5G) heterogeneous networks, this study investigates wireless resource allocation methods for industrial 5G heterogeneous networks. First, a downlink transmission scenario for 5G heterogeneous network with coexisting TS and RT services is constructed, and a variable blocklength transmission scheme based on finite blocklength theory is designed. Then, an optimization problem is established to ensure the deterministic delay of TS services and high throughput of RT services. Finally, to solve the Mixed Integer Nonlinear Programming (MINLP) problem, it is decoupled into three sub-problems: channel allocation, power allocation, and transmission delay design. An algorithm based on Alternative Optimization (AO) is proposed to obtain a suboptimal solution to the original problem. Research results indicate that compared to the 5G New Radio (NR) scheme, the variable blocklength transmission scheme improves the transmission sum-rate of RT devices by about 13% through more flexible resource allocation, while ensuring the delay determinacy of TS services.
