Abstract:To address the challenge of balancing accuracy and detection speed in processing large-scale visual image data of railway fasteners on embedded devices in real time,a lightweight real-time detec-tion model based on attention divide-and-conquer and feature fusion is proposed.First,a hybrid divide-and-conquer attention module,leveraging both spatial and channel features,is introduced to en-hance the model's feature extraction capability and reduce interference from complex backgrounds in the image.Second,a dual divide-and-conquer feature fusion method is designed to improve detection performance across targets of varying sizes.In the construction of the cost volume detection head(YOLO Head),a Varifocal Loss(VFL)function is employed to replace the binary cross-entropy loss used in YOLOX-Nano,thereby enhancing the accuracy of lightweight real-time detection.Further-more,a Random Alpha-IoU(RAL)loss function is adopted to dynamically adjust parameters,slow down convergence,and optimize the training curve,thus preventing the model from falling into local optima.Finally,a dataset of 10,233 annotated fastener targets categorized into six types is used for evaluation.Comparative experiments are conducted using mainstream object detection models,includ-ing YOLOX-Nano,Faster R-CNN,and YOLOv8n.The research results indicate that the proposed model achieves a frame rate of 60.24 Frames Per Second(FPS)and an Average Precision(AP)of 83.40％,representing a 3.24％improvement over the baseline.The parameter count is 2.31 M,which is 54.08％fewer than YOLOX Tiny,and the floating-point operations is 1.99 G,a 69.15％de-crease compared to YOLOX Tiny.These findings provide valuable insights for the development of lightweight real-time detection models and embedded computing systems.

Abstract:Aiming at the low estimation accuracy of the observer as well as the strong coupling,exter-nal perturbation,and time-varying parameters of the high-speed train system,a Fractional Order Non-singular Fast Terminal Sliding Mode Control(FONFTSMC)based on the Compensating Function Observer(CFO)is put forward to improve the robustness and control accuracy of the high-speed train speed control.Firstly,a longitudinal multi-mass dynamics model of the high-speed train is estab-lished,and a high-precision compensation function observer is designed to estimate and compensate the total perturbation of the system in real time;Secondly,a fractional-order non-singular fast terminal sliding mode control algorithm with state-negative exponential control law is designed for tracking and controlling the train's running curve,and the system's convergence is proved to be in finite time by the Lyapunov stability theory;Finally,the parameters of CRH3 high-speed train and the actual line of Hefei Station-Bengbu South Station are used as examples to track the ideal operation curve and the energy-saving optimized operation curve for experimental verification,respectively.The simulation results show that the average error of the proposed algorithm in tracking the ideal operating speed curve is 0.0137 7 km/h,and the average error in tracking the energy-saving optimized operating speed curve with interference is 0.036 4 km/h.Compared with the sliding mode and non-singular fast terminal sliding mode control methods based on dilated state observer,the proposed method has the smallest tracking error and higher tracking accuracy,which verifies its effectiveness and feasibility,and can provide a reference for the research in the field of train speed tracking control.

Abstract:In the modeling of bearing currents in variable-frequency AC motors,the direct measure-ment of stray capacitance often yields unstable results due to fluctuations in terminal capacitance data.To address this issue,this paper proposes a short-circuit method for measuring stray capacitance.First,the generation mechanism of bearing currents in variable-frequency drive systems is analyzed,along with a description of the lumped-parameter bearing current model and the associated stray capaci-tance parameters.Second,the direct measurement method is employed to determine motor stray ca-pacitance,and the impact of terminal capacitance fluctuations on the stability of both the measured ca-pacitance and its derivative are assessed.Finally,the short-circuit method is introduced for extracting stray capacitance,on which the influence of terminal capacitance fluctuations is examined.Experimen-tal results indicate that in the direct measurement method,terminal capacitance fluctuations signifi-cantly impair the stability of the measured stray capacitance and its derivative,resulting in discontinui-ties and high variability of stray capacitance.In contrast,the short-circuit method yields stray capaci-tance values and derivatives that are unaffected by sensitivity intervals and show no data fluctuations.These findings demonstrate that the short-circuit method is more suitable for measurement of stray ca-pacitance in AC motors.

Abstract:Regular inspection of insulator conditions is a crucial part for ensuring the safe operation of power grids.To address the challenges of unsatisfactory insulator defect detection performance in aerial images,caused by defect regions occupying a small proportion of the image and target sizes vary-ing inconsistently,a defect detection algorithm for insulators based on a feature refinement network is proposed.First,the algorithm uses the Focal Modulation Network(FocalNet)to encode spatial con-text at multiple granularity levels and integrates it with Spatial Pyramid Pooling Faster Cross Stage Partial Channel(SPPFCSPC)to construct the feature extraction backbone,enhancing the network's feature extraction capabilities.Next,an enhanced feature-adaptive fusion pyramid is designed,incor-porating a positioning information supplementation branch to mitigate the loss of defect features.Addi-tionally,Efficient Multi-Scale Attention(EMA)is introduced to generate rich semantic feature maps at different resolutions.Finally,the Feature Refinement Detection Head extracts and aggregates multi-scale feature information of insulators and defects,producing more discriminative features for detecting targets of varying scales.Experimental results demonstrate that proposed method achieves an mAP of 98.2％,effectively identifying multi-scale insulators and defects,thereby providing references for multi-scale detection in insulator aerial images.

Abstract:To address the issue of high-speed railway subgrade settlement influenced by factors such as temperature and humidity,this study proposes a combined prediction model integrating an improved Dung Beetle Optimization(My Improved Dung Beetle Optimization,MIDBO)algorithm,Back Propagation(BP)neural network,and Adaptive Boosting(Adaboost).First,to overcome the limita-tions of the conventional Dung Beetle Optimization(DBO)algorithm,specifically its tendency to con-verge to local optima and poor performance in complex engineering applications,a novel MIDBO algo-rithm is developed.This algorithm integrates composite chaotic mapping,simulated annealing,and a nonlinear exponential dynamic weighting strategy.Then,the MIDBO algorithm is used to optimize the BP neural network,which is then combined with the Adaboost algorithm to establish the MIDBO-BP-Adaboost model.Finally,the proposed model and comparative models are applied to predict sub-grade settlement on the Lanzhou-Urumqi High-Speed Railway.Experimental results demonstrate that the MIDBO algorithm effectively optimizes the BP neural network,significantly improving prediction accuracy of the model.The Adaboost algorithm further enhances the model's robustness and general-ization capability.Compared with the BP model,the MIDBO-BP-Adaboost model reduces the mean absolute error,root mean square error,and mean absolute percentage error by 63.81％,63.84％,and 62.26％,respectively,while increasing the coeffi-cient of determination by 18.82％.These findings offer a valuable reference for predicting subgrade settlement in Lanzhou-Urumqi High-speed Railway.

Abstract:To address the limitations of existing urban rail transit energy consumption anomaly detec-tion methods in terms of accuracy and adaptability,this study proposes a subway traction energy con-sumption anomaly detection framework based on time series characteristics.First,a hybrid prediction model is constructed by integrating Convolutional Neural Networks(CNN),Long Short-Term Memory(LSTM)networks,and Extreme Gradient Boosting(XGBoost)with an attention mechanism.This approach enhances the prediction accuracy of typical energy consumption values through feature extraction and a nonlinear ensemble strategy.Second,a joint anomaly detection method combining a Density-Based Spatial Clustering of Applications with Noise algorithm and the Local Outlier Factor is designed to enable dynamic threshold calibration and quantitative evaluation of anomaly factors.Fi-nally,the model's performance is validated using real-world data from a subway line.The results indicate that proposed hybrid model reduces prediction error by 47.3％compared to the traditional LSTM,with the Mean Absolute Percentage Error stabilized at 1.02％.The proposed anomaly detec-tion module achieves an accuracy of 96.8％,outperforming the Isolation Forest algorithm by 12.5％.Moreover,the proposed method demonstrates robustness against sudden fluctuations in streaming data.The research results offer a practical reference for anomaly localization and energy-saving optimi-zation in subway energy management,effectively supporting refined energy efficiency monitoring and control in rail transit systems.

Abstract:To address the challenge of extrinsic parameter calibration for multi-camera automotive surround-view systems,this paper proposes an extrinsic parameter calibration algorithm based on a cross-attention mechanism.First,multi-scale features from multi-view images are independently extracted using residual convolutional modules to capture fine-grained image details.Then,a cross-attention module is introduced to learn global features of each camera image as well as the inter-camera feature relationships with surrounding cameras,thereby enhancing the overall feature representation capability.These features are subsequently integrated via a feature fusion module,which combines outputs from both the residual convolutional and cross-attention modules to regress the extrinsic parameters.Finally,the proposed model is validated on two datasets through perfor-mance evaluation and ablation studies.Experimental results demonstrate that,compared with existing extrinsic parameter calibration algorithms based on lane lines and textures,the algorithm proposed in this paper has better generalization and robustness in different environments,with significant improve-ments in performance metrics and bird's-eye view stitching visualization results.Compared with exist-ing extrinsic parameter calibration algorithms based on lane markings and texture cues,the proposed algorithm exhibits superior generalization and robustness across diverse environments,with notable improvements in quantitative performance metrics and bird's-eye view stitching quality.Specifically,the algorithm achieves absolute reprojection and photometric errors of 3.1 and 16.7,respectively,representing improvements of 8.82％and 8.74％over the current state-of-the-art weakly-supervised extrinsic self-calibration network(WESNet).The research findings provide technical support for online extrinsic parameter calibration in automotive surround-view systems.

Abstract:In response to the problem that the on-board personnel monitoring the rail end of the steel rail transportation train have difficulty in judging whether the rail has detached from the fastening de-vice in a timely manner,a rail end detection algorithm based on the improved YOLOv5s is proposed.Firstly,the lightweight GhostNet backbone network is adopted to replace the original Cross Stage Par-tial Network(CSPNet),reducing the high requirements of the model for hardware resources;Sec-ondly,BiFomer and Receptive-Field Attention(RFA)attention mechanism are added to weaken the irrelevant background regions while improving the positioning ability of the rail end;Thirdly,the loss function SIoU is used to replace the original CIoU,enhancing the generalization ability of the model,and enabling the model to converge faster.Finally,the algorithm is verified and evaluated from the as-pects of detection accuracy and detection speed,and compared with algorithms such as Single Shot MultiBox Detector(SSD)and YOLOv8.The research results show that the improved detection algo-rithm achieves an average detection accuracy of 91.7％for the rail end detection,with an average de-tection time of 24 ms,which is 5.3％higher than the original YOLOv5s model.The missed detection and false detection situations have been significantly improved.The improved algorithm can achieve precise detection of the rail end in different environments,has good adaptability in adverse environ-ments such as dim lighting,and has a lower floating-point operation per second,which can be de-ployed in the embedded device RK3399 to better meet the real-time detection requirements of the rail end.

Abstract:Addressing the challenges of high communication latency,low reliability,and poor stability in current remote-control signal transmission,this study proposes a multi-link communication architec-ture with low correlation among links.First,the sensing signals and control signals are spectrally sepa-rated according to the characteristics of radio wave propagation,and the attenuation laws of radio waves in different frequency bands under the tunnel scenario are analyzed.Second,for control signal transmission,leveraging finite blocklength coding theory and Little's theorem,the relationships among signal-to-noise ratio,achievable rate,and information frame length under a fixed bit error rate are derived.A low-latency transmission communication model for tunnels is established,and a frame structure tailored for control signal transmission is designed to enable a low-latency communication transceiver.Third,regarding sensing signal transmission,the propagation characteristics of ultra-high-frequency radio waves in tunnels and the correlation between signal strength and communication band-width are used.By referencing a lookup table based on bandwidth requirements,an appropriate signal strength range for sensing signal transmission is determined.Finally,experiments simulating remote-control operations in a factory environment are conducted with a focus on low-latency remote-control signal transmission.Experimental results indicate that the average communication delay of control sig-nal transmission is 0.75 ms,while the communication delay for sensing signal transmission is 200 ms,thereby validating the effectiveness and reliability of the proposed communication architecture and associated technologies for remote-control signal transmission in tunnel construction.

Abstract:The performance of existing environmental sensing technologies under limited millimeter-wave(mmWave)signal bandwidth has not been thoroughly investigated.Moreover,these technolo-gies often face challenges such as high computational complexity and insufficient real-time capability in complex scenarios.To address these issues,this paper proposes an active environmental map construc-tion algorithm based on limited-bandwidth mm Wave communication signals,aiming to enable real-time environment mapping and enhance the overall performance of communication systems.First,the proposed method begins by actively transmitting and receiving mm Wave communication signals at the mobile terminal to extract the propagation delay and Angle Of Arrival(AOA)information of echoes.Combined with the terminal's pose information,these signals are used to preliminarily estimate ob-stacle coordinates.Second,considering the impact of limited signal bandwidth on map resolution,an occupancy grid mapping algorithm is employed to characterize the spatial positions of obstacles.Bresenham's algorithm is then used to compute free grid cells,facilitating rapid and accurate environ-mental map construction.Subsequently,simulation experiments are conducted under varying map resolutions to analyze the mapping performance acorss different settings.Furthermore,the AOA distri-bution in the sensed environment is statistically analyzed and fitted to validate the applicability of the von Mises distribution.Finally,the constructed maps are compared against lidar-based benchmark maps.Accuracy is evaluated using Root Mean Square Error(RMSE),obstacle shape similarity is as-sessed via the Jaccard index,and the algorithm's efficiency and system performance are measured by code execution time.The results demonstrate that the proposed algorithm achieves optimal perfor-mance under a 3 GHz bandwidth and a resolution of 25 grids per meter,yielding an RMSE of 9.294 3,a Jaccard index of 0.625 4,and a code execution time of 14.774 6 minutes,satisfying real-time envi-ronment sensing requirements.Compared to results obtained with a 500 MHz bandwidth at the same resolution,the RMSE,Jaccard index,and execution time are improved by 56.1％,394.4％,and 70.6％,respectively.These findings provide valuable insights for the development of future high-dynamic communication systems.

Abstract:This paper proposes an adaptive non-singular fast terminal sliding mode controller based on a Radial Basis Function(RBF)neural network to solve the speed tracking control problem of high-speed trains,considering the effects of unknown model parameters,uncertain additional resistance,unknown inter-car forces,and external disturbances.First,a multi-mass dynamic model of high-speed trains is established,incorporating nonlinear resistance and inter-car coupling forces between ad-jacent carriages.Second,a finite-time speed tracking control strategy for high-speed trains based on a novel saturation function is designed.The non-singular fast terminal sliding mode control method is in-troduced to achieve finite-time convergence of the system state,enhancing both steady-state accuracy and transient performance in speed tracking.Furthermore,an adaptive non-singular fast terminal sliding mode control strategy based on RBF neural network is developed.Adaptive estimation techniques are employed to estimate train model parameters and the upper limit of uncertainty terms,including addi-tional resistance and inter-vehicle forces,in real time.To mitigate chattering caused by discontinuous switching control,an RBF neural network is utilized to remap the switching control term.Additionally,an adaptive update law for weight coefficients is designed to ensure continuous switching,effectively sup-pressing chattering effects.Finally,the stability of the high-speed train speed tracking control system and the finite-time convergence of system states are proven using Lyapunov stability theory.The pro-posed approach is validated through simulations using the CRH380B high-speed train as the control ob-ject.Simulation results demonstrate that the proposed controller enables high-speed trains to converge and track the desired trajectory within a finite time,with a 49％reduction in tracking error and achieving high tracking accuracy.These findings provide valuable insights for high-speed train tracking control.

Abstract:To address the challenges of long-term OD passenger flow forecasting for high-speed rail trains,characterized by large data volumes and high prediction complexity,this study proposes a pre-diction method based on the Informer model.First,the connotation of the OD passenger flow forecast-ing problem is defined,and a research framework encompassing data acquisition,data processing,and flow prediction is developed.Second,historical data from the passenger train timetable system and the passenger transportation big data platform are collected to extract key features influencing OD passen-ger flow.Then,an Informer-based prediction model is constructed,leveraging a decoder structure em-bedded with a probabilistic sparse self-attention mechanism to capture the long-range dependencies within OD flow data across different trains.The model assigns higher attention weights to critical time points and generates predicted passenger flow trends through the decoder.Finally,a case study using high-speed rail trains on the Shanghai Hongqiao-Beijing South segment of the Beijing-Shanghai HSR line validates the effectiveness of the proposed method.Results demonstrate that the Informer model achieves superior prediction accuracy compared to the Transformer,Gated Recurrent Unit(GRU),and Long Short-Term Memory(LSTM)models,improving training set accuracy by 2.11％,6.97％,and 6.79％,and test set accuracy by 1.42％,7.19％,and 8.24％,respectively.The results of this study provide a robust data-driven basis for refined passenger transport planning in high-speed rail op-erations and offer valuable reference for optimizing high-speed rail passenger service strategies.

Abstract:To address the limitations of current turnout track inspection methods,such as heavy reli-ance on manual labor,low detection efficiency,and the lack of depth information in 2D visual ap-proaches,this paper proposes a turnout point cloud segmentation method based on a multi-scale fusion strategy,named Point-Bidirectional Encoder Representations from Transformers-Turnout(Point-BERT-T).First,in the local point cloud encoding stage,spherical groupings with varying radii are employed to extract and fuse features from points within each sphere,generating a spatially hierarchi-cal hybrid feature representation.This multi-scale fused feature captures information at different spatial levels of the turnout,improving the efficiency and accuracy of railway infrastructure recognition and segmentation.It significantly enhances the capability of 3D point cloud recognition for turnouts and supports downstream tasks such as defect and deformation detection.Next,a random rotation-translation and non-uniform slicing strategy is introduced during data preprocessing to simulate real-world scanning variability,thereby improving the model's robustness under diverse data acquisition conditions.Finally,to validate the effectiveness of the proposed method,comparative experiments are conducted against existing approaches.The research results demonstrate that,compared to Point-BERT,the Point-BERT-T method improves the overall turnout point cloud segmentation perfor-mance by 1.9％.Furthermore,for the more challenging frog and wing rail components,the Intersec-tion over Union(IoU)increases by 4.7％and 5.6％,respectively,demonstrating the method's accu-racy and robustness in semantic segmentation of 3D turnout point cloud data.