Hand-held terminal for detecting railway electric service vehicle-mounted equipment and detection method

Abstract

The invention discloses a handheld terminal for detecting railway electric service vehicle-mounted equipment and a detection method, and the method comprises the following steps: S1, collecting operation data, and carrying out the preprocessing of the operation data; s2, constructing a graph neural network model, mapping the operation data into a node feature graph, and setting an attention weight distribution rule; s3, graph convolution operation is executed, features of adjacent nodes are aggregated, and attention weight is dynamically adjusted in combination with a node feature graph; s4, introducing a cross-scale fusion structure, performing feature compression on the embedded feature graph, and retaining an association relationship between nodes and structural feature information; s5, inputting the fusion feature matrix into the residual aggregation sub-network, and executing multi-layer feature extraction and jump connection operation; s6, generating a structured detection report, and visualizing the structured detection report on the handheld terminal; and S7, continuously optimizing network model parameters and attention weight distribution rules. According to the invention, equipment state identification and abnormity positioning are realized, and high efficiency, accuracy and visualization are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of embedded intelligent terminals and target detection, and particularly relates to a handheld terminal for detecting railway electric service on-board equipment and a detection method. BACKGROUND

[0002] In a railway transportation system, electric service on-board equipment, as a key basis for train operation safety and dispatching instruction transmission, its real-time monitoring and abnormal early warning are of great significance to the stability and safety of railway operation. Electric service equipment usually includes train control system communication units, signal acquisition modules, power distribution devices, data transmission equipment, etc., which are distributed in multiple key parts of locomotives and rolling stocks, and have large quantities, complex types, diverse communication protocols and harsh operating environments. Therefore, efficient and reliable detection and fault diagnosis of these devices is a highly technically challenging task.

[0003] The existing detection method of railway electric service on-board equipment mainly relies on manual detection or traditional handheld test instruments for periodic maintenance. Such traditional detection methods have problems such as low detection efficiency, data processing lag, poor device interface adaptation, and coarse detection granularity. The operator needs to access the device through a physical interface and rely on manual experience to determine whether there is an abnormality. This way not only takes time and effort, but also easily leads to a decrease in detection accuracy due to environmental interference, human error, etc., and is difficult to meet the needs of modern railway transportation for intelligent and digital device detection.

[0004] On the other hand, with the increasing complexity of the railway system, the interaction behavior and network structure between devices exhibit graph structure characteristics. The existing detection methods lack the ability to model the communication relationship between devices and cannot capture the potential dependencies and fault propagation chains between devices. In addition, some systems attempt to use traditional machine learning-based detection algorithms to analyze on-board data, but most models ignore the topological association between devices and only perform static evaluation of the operating state of individual devices, making it difficult to achieve cross-device and cross-node abnormal association recognition.

[0005] In terms of data processing, the existing detection technology usually uses static feature extraction methods, which cannot dynamically respond to real-time changes in device operating state. At the same time, the data preprocessing process is not standardized, and some methods fail to effectively handle missing values, outliers, and repeated values, resulting in insufficient data quality for input models and affecting detection accuracy. Although some detection tools have visualization functions, the data display dimension is limited and cannot fully present the state distribution and time evolution process between multiple nodes, which is not conducive to comprehensive judgment by operation and maintenance personnel.

[0006] In addition, the traditional detection tools generally have poor interface compatibility, low software scalability, slow model updating and other problems, which are difficult to adapt to the equipment needs of different types and different manufacturers in the railway field. Due to the failure to adopt a unified equipment detection process and nested model training mechanism, the promotion of the detection system is difficult and the maintenance cost is high, which is not conducive to the construction of a standardized intelligent detection system.

[0007] Based on the above problems, it is necessary to propose a new intelligent detection method for railway electric vehicle-mounted equipment, which can model the communication topology between equipment from the graph structure level, combine embedded feature maps, attention mechanisms, and various deep learning technologies such as cross-scale feature fusion to effectively improve detection accuracy, stability and real-time performance. The superior performance of graph neural network model in graph structure data modeling makes it a powerful tool for electric equipment detection. Through graph convolution operation, node adjacency information can be effectively aggregated, and the spatial diffusion trend of fault influence can be dynamically perceived. The attention mechanism can achieve importance weighting between equipment nodes, effectively improving the sensitivity and relevance of anomaly detection.

[0008] At the same time, the residual aggregation structure and the skip connection mechanism can enhance the feature expression ability of the deep network, avoid the gradient vanishing problem in the deep neural network, and the cross-scale fusion structure can process node feature distribution under different spatial scales, capture small but critical structural disturbances, and effectively improve the simultaneous perception ability of local anomalies and large-scale faults. By constructing an end-to-end model training and reasoning process, combined with structured report generation, anomaly visualization display, and equipment number association mapping functions, the practicality and on-site response efficiency of detection information are further improved.

[0009] Therefore, how to provide a handheld terminal and detection method for railway electric vehicle-mounted equipment detection is a problem that those skilled in the art need to solve. SUMMARY

[0010] One object of the present application is to provide a handheld terminal and detection method for railway electric vehicle-mounted equipment detection, which fully utilizes graph neural networks, multi-scale feature fusion, residual aggregation subnetworks and attention mechanisms, and describes in detail the intelligent detection process based on node feature modeling and communication relationship topology analysis, which has the advantages of strong adaptability, high detection accuracy, high abnormality recognition sensitivity and strong visual interaction capability.

[0011] According to the handheld terminal and detection method for railway electric vehicle-mounted equipment detection of the embodiment of the present application, the following steps are included:

[0012] S1, connect the vehicle-mounted equipment through the handheld terminal, collect the running data, and perform preprocessing;

[0013] S2. Build a graph neural network model, map the preprocessed running data into a node feature graph, set the attention weight distribution rule, and initialize the weight coefficients of each node state and edge in the graph;

[0014] S3. Perform graph convolution operations in the graph neural network model, aggregate the features of adjacent nodes based on the edge weight coefficients, and dynamically adjust the attention weights based on the node feature graph to generate an embedded feature graph;

[0015] S4. Introducing a cross-scale fusion structure into the graph neural network model, compressing multi-scale node features of the embedded feature graph, retaining the correlation relationship and structural feature information between nodes, and generating a fusion feature matrix;

[0016] S5. Input the fused feature matrix into the residual aggregation sub-network, perform multi-layer feature extraction and skip connection operations, and output the operating status label of the device and the index information of the node with communication anomaly;

[0017] S6. Generate a structured detection report based on the operation status tag and the communication abnormal node index, and visualize it on the handheld terminal. The visualization information includes the device status, the abnormal node location and the corresponding timestamp;

[0018] S7. Continue to optimize the network model parameters and attention weight distribution rules, and store the structured detection report in the local buffer.

[0019] Optionally, the operating data includes communication protocol data, voltage and current data, and status signal data.

[0020] Optionally, the preprocessing includes data deduplication, outlier removal and data standardization.

[0021] Optionally, the S2 specifically includes:

[0022] S21, based on the pre-processed running data, define three feature vector sequences: communication protocol data sequence D c ={d c1 ,d c2 ,…,d cn}, voltage and current data sequence D e ={d e1 ,d e2 ,…,d en}, status signal data sequence D s ={d s1 ,d s2 ,…,d sn}, where n represents the number of data points in the time window, d ci represents the characteristic value of the i-th communication protocol, d ei represents the characteristic value of the i-th electrical signal, dsi represents the i-th state signal feature value;

[0023] S22, a graph neural network model is constructed, three feature vector sequences are spliced, and a node feature graph G=(V, E, A) is generated through a mapping function, wherein V represents a node set, E represents an edge set, and A represents an adjacency matrix, a ij =1 represents that there is a direct connection between node i and node j, and a ij =0 represents that there is no direct connection between the two nodes.

[0024] S23, the state value of each node in the graph and the weight coefficient of each edge are initialized:

[0025]

[0026] wherein s i is the state value of node i, are the k-th communication, electrical and state signal features of node i, respectively, and a1, a2 and a3 are node state weighting coefficients, w ij is the edge weight between node i and node j, d ij is the physical distance between node i and node j, q ij is the number of successful communications between node i and node j within a given time window, Q is the maximum number of communications within the time window, and Δt ij is the time difference of the last communication, and b1, b2 and b3 are edge weight weighting coefficients.

[0027] Optionally, the graph neural network model constructs node feature input based on the three types of feature vector sequences, fuses the three types of feature vector sequences through a multi-channel splicing operation, maps the spliced feature vectors to a low-dimensional embedding representation using a differentiable mapping function, and introduces a gated attention mechanism to assign feature channel weights, to generate a node feature graph for graph convolution operation.

[0028] Optionally, the S3 specifically includes:

[0029] S31, performing graph convolution operation in the graph neural network model, and performing adjacency feature aggregation on the feature vector x i of each node i, wherein the convolution calculation formula of the l-th layer of the adjacency feature aggregation is:

[0030]

[0031] wherein, is the embedding feature vector of node i at the l-th layer, and s is a ReLU activation function, represents the direct adjacent node set of node i, is the attention weight of node i and node j at the l-th layer, and w ijis a weight coefficient of edge (i, j), is an embedding feature vector of node j in the (l-1)-th layer, a ij represents a connection relationship in an adjacency matrix, W (l-1) is a weight matrix of the (l-1)-th layer, b (l-1) is a bias vector of the (l-1)-th layer.

[0032] S32, after performing L-layer graph convolution, the embedding feature vectors of all nodes are stacked, and the attention weight is dynamically adjusted in combination with the node feature map to generate an embedding feature map H, the embedding feature map is a matrix structure formed by stacking the embedding feature vectors of all nodes according to the node index order, reflecting the communication relationship strength between the vehicle-mounted devices and the state association.

[0033] Optionally, the attention weight calculation formula of the l-th layer is:

[0034]

[0035] wherein a is a trainable attention vector, a T is a transpose of a, and || represents a vector splicing operation are embedding feature vectors of node i and node k in the (l-1)-th layer respectively, exp(·) represents a natural exponential function with e as the base, and LeakyReLU(·) is a nonlinear activation function.

[0036] Optionally, the S4 specifically includes:

[0037] S41, a cross-scale fusion structure is introduced in the graph neural network model, and the embedding feature map H is input into the cross-scale fusion structure, and a multi-scale processing window set is set for each scale performing an intra-scale node feature compression operation to obtain a node feature matrix, the node feature matrix retains the association relationship and structure feature information between nodes:

[0038]

[0039] wherein, is a node feature matrix under scale s p is an activation function under scale s p is a feature compression matrix under scale s p is a bias vector under scale s p

[0040] S42, performing cross-scale fusion on all scale feature matrices to construct a fusion feature matrix:

[0041] ​​​​

[0042] Among them, F is the fusion feature matrix, Scale s p The fusion weight under ⊙ is the Hadamard element-wise product operation, A scale mask matrix, where each element in the scale mask matrix is ​​defined as:

[0043]

[0044] in, is the distance between node i and adjacent node j at scale s p The structural difference response under and For nodes i and j at scale s p The eigenvector row under represents the absolute difference of the corresponding vector elements, and tanh(·) represents the hyperbolic tangent function, which is used for normalization. Represents the set of directly adjacent nodes of node i.

[0045] Optionally, the S5 specifically includes:

[0046] S51. Input the fusion feature matrix F into the residual aggregation sub-network. The residual aggregation sub-network consists of two layers of nonlinear mapping layers and corresponding skip connection structures. A nonlinear activation function is used at the output of each layer to perform multi-layer nonlinear mapping operations to construct the l-th layer residual feature:

[0047]

[0048] in, is the feature vector of node i in the lth layer of the residual network, F i is the fused feature vector of node i, is the weight matrix of the l-th layer residual network, is the bias vector, ξ is the Swish activation function, ψ is the linear mapping function, and the output of the previous layer is Perform dimension alignment;

[0049] S52, in the last layer L of the residual aggregation sub-network r Perform a jump connection operation and output the device's operating status and the node's communication score:

[0050]

[0051] Among them, y i is the operating status of the device corresponding to node i, softmax(·) represents the nonlinear activation function, δi communication score of node i, W o output weight matrix, b o output bias vector, d is feature dimension of node i, jth feature value of node i, μ j average of jth feature of all nodes, normalized variance of jth feature of all nodes, epsilon is a constant stabilizing term to prevent the denominator from being zero, n is the number of nodes;

[0052] S53, according to y i determine the running state label of the device, and attach the corresponding device identifier, and delta i is compared with the set threshold theta, if delta i > theta, it is determined that node i is a communication abnormal node, and the index information of the communication abnormal node is output.

[0053] Optionally, the S6 specifically comprises:

[0054] S61, generating a structured detection report according to the running state label and the communication abnormal node index, the structured detection report arranges the on-board devices in order of node number, and inserts the label number, label name and classification confidence information correspondingly;

[0055] S62, for each number identified as a communication abnormal node, associating the corresponding node with the timestamp sequence in the original running data, generating a time record corresponding to the abnormal node, and dividing the abnormal level into three levels according to the communication abnormal score, and marking them as general, serious and urgent respectively;

[0056] S63, associating the running state label, the communication abnormal node number, the abnormal level, the timestamp and the device number, further supplementing the information amount of the structured detection report, and uploading the detection report to the handheld terminal visualization interface, displaying the device state, the abnormal node position and the corresponding timestamp on the terminal interface in real time, and displaying the normal indicators and various fault information according to the node state classification and partition.

[0057] The beneficial effects of the present application are:

[0058] Firstly, the present application introduces a graph neural network model, establishes a communication relationship-based on-board device topology graph, and realizes effective modeling of the complex interaction relationship between devices. Compared with the traditional single-node independent detection method, this method can analyze the device running state from the overall system level, capture the propagation path of abnormal signals in the network, and improve the systematization and accuracy of fault detection. At the same time, the introduction of node feature aggregation and attention weight mechanism enables the model to dynamically focus on key nodes and high-risk connections, effectively enhancing the identification ability of communication abnormal nodes.

[0059] Secondly, the application constructs a cross-scale fusion structure and a residual aggregation sub-network, realizes the unified alignment and deep expression of node features at different spatial scales at the feature extraction level, so that the detection model has good robustness when dealing with local small disturbances and wide area structure abnormalities. Combined with multi-layer jump connection and residual learning mechanism, the gradient attenuation problem in deep network training is effectively alleviated, and the stability and generalization ability of the model in large-scale data input scene are improved. The construction of the fusion feature matrix not only preserves the structural dependence information between nodes, but also improves the diversity of feature expression and the sensitivity of the detection model.

[0060] Finally, the application generates a structured detection report and visualizes the terminal display, significantly improving the usability and on-site adaptability of the device detection results. The system can display the device running state, communication abnormal node position and corresponding timestamp in real time, and the operation and maintenance personnel can intuitively obtain key information to realize rapid positioning and intervention. At the same time, the detection results are stored in association with the original device number and running data, providing data support for subsequent state tracking and maintenance strategy optimization. The handheld terminal has good compatibility, expandability and on-site deployment convenience, and can be widely used for various electric railway equipment detection needs in railway operation guarantee, and has significant practical application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0061] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0062] Fig. 1 A flowchart of a handheld terminal and detection method for railway electric railway vehicle-mounted equipment detection proposed by the application;

[0063] Fig. 2 A data processing and graph structure construction schematic diagram of a handheld terminal and detection method for railway electric railway vehicle-mounted equipment detection proposed by the application;

[0064] Fig. 3 A feature embedding and fusion schematic diagram of a handheld terminal and detection method for railway electric railway vehicle-mounted equipment detection proposed by the application. DETAILED DESCRIPTION

[0065] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0066] Reference Figs. 1-3 A handheld terminal and detection method for railway electric railway vehicle-mounted equipment detection, comprising the following steps:

[0067] S1, connect the vehicle-mounted equipment through the handheld terminal, collect the running data, and perform preprocessing;

[0068] S2, construct a graph neural network model, map the preprocessed running data to a node feature graph, set an attention weight distribution rule, and initialize the state of each node and the weight coefficient of the edge in the graph;

[0069] S3, perform a graph convolution operation in the graph neural network model, aggregate the features of adjacent nodes based on the weight coefficient of the edge, and dynamically adjust the attention weight in combination with the node feature graph to generate an embedded feature graph;

[0070] S4, introduce a cross-scale fusion structure into the graph neural network model, perform multi-scale node feature compression on the embedded feature graph, and retain the correlation relationship and structural feature information between nodes to generate a fusion feature matrix;

[0071] S5, input the fusion feature matrix into a residual aggregation subnetwork, perform multi-layer feature extraction and jump connection operations, and output the running state label of the device and the index information of the node with communication abnormalities;

[0072] S6, generate a structured detection report according to the running state label and the index of the communication abnormal node, and visualize it on the handheld terminal, wherein the visualization information includes the device state, the abnormal node position and the corresponding timestamp;

[0073] S7, continuously optimize the network model parameters and the attention weight distribution rule, and store the structured detection report in the local buffer.

[0074] The present application constructs a detection process based on a graph neural network, models the running data of the railway electric service vehicle-mounted equipment as a graph structure, combines the attention mechanism and embedded feature generation, can accurately identify the communication abnormalities between multiple nodes and comprehensively evaluate the running state, and improves the intelligent level of detection and system reliability.

[0075] In the embodiment, the running data includes communication protocol data, voltage and current data, and state signal data.

[0076] The present application defines the running data types as communication protocol data, voltage and current data, and state signal data, ensures the comprehensiveness and structure of the data source, provides multi-dimensional and complementary input basis for subsequent graph model construction, and enhances the coverage and accuracy of fault identification.

[0077] In the embodiment, the preprocessing includes data deduplication, outlier rejection, and data standardization.

[0078] The application effectively improves the quality and stability of input data by performing data deduplication, outlier removal and standardization operation in the preprocessing stage, reduces the influence of noise on model inference results, and enhances the robustness of the model in complex environments.

[0079] In the embodiment, S2 specifically includes:

[0080] S21, based on the running data of preprocessing, three feature vector sequences are defined respectively: a communication protocol data sequence D c ={d c1 ,d c2 ,…,d cn}, a voltage and current data sequence D e ={d e1 ,d e2 ,…,d en}, and a state signal data sequence D s ={d s1 ,d s2 ,…,d sn}, wherein n represents the number of data points in the time window, d ci represents the i-th communication protocol feature value, d ei represents the i-th electrical signal feature value, and d si represents the i-th state signal feature value.

[0081] S22, a graph neural network model is constructed, the three feature vector sequences are spliced, and a node feature map G=(V,E,A) is generated through a mapping function, wherein V represents a node set, E represents an edge set, and A represents an adjacency matrix, a ij =1 indicates that node i and node j have a direct connection, and a ij =0 indicates that there is no direct connection between the two nodes.

[0082] S23, the state value of each node in the graph and the weight coefficient of each edge are initialized:

[0083]

[0084] wherein s i is the state value of node i, are the k-th communication, electrical and state signal features of node i, respectively, a1, a2 and a3 are node state weighting coefficients, w ij is the edge weight between node i and node j, d ij is the physical distance between node i and node j, q ij is the number of communication successes between node i and node j in a given time window, Q is the maximum number of communications in the time window, and Δt ijis the time difference of the last communication, and β1, β2, β3 are edge weight coefficients.

[0085] The application fully mines the structural dependence and communication characteristics between devices by constructing a multi-source feature vector sequence and realizing state initialization of nodes and edges in a graph neural network model, provides high-quality structural input for graph model learning, and improves the perception and representation capabilities of the model.

[0086] In the embodiment, the graph neural network model constructs node feature input based on three types of feature vector sequences, fuses the three types of feature vector sequences through a multi-channel splicing operation, maps the spliced feature vector to a low-dimensional embedding representation using a differentiable mapping function, introduces a gated attention mechanism to assign feature channel weights, and generates a node feature map for graph convolution operation.

[0087] The application fuses three types of feature vector sequences in a multi-channel manner, introduces a differentiable mapping function and a gated attention mechanism, strengthens the selective expression of feature channels, improves the integrated analysis capability of the graph model for multi-modal information, and effectively improves the sensitivity of anomaly recognition.

[0088] In the embodiment, the S3 specifically comprises:

[0089] S31, performing graph convolution operation in the graph neural network model, and performing adjacent feature aggregation on the feature vector x i of each node i, the convolution calculation formula of the lth layer of the adjacent feature aggregation is:

[0090]

[0091] wherein, is the embedding feature vector of node i at the lth layer, and σ is a ReLU activation function, denotes a set of direct adjacent nodes of node i, is the attention weight of node i and node j at the lth layer, w ij is the weight coefficient of edge (i, j), is the embedding feature vector of node j at the (l-1)th layer, a ij denotes the connection relationship in the adjacent matrix, W (l-1) is the weight matrix of the (l-1)th layer, b (l-1) is the bias vector of the (l-1)th layer;

[0092] S32, after performing L-layer graph convolution, stacking the embedding feature vectors of all nodes, and dynamically adjusting the attention weight combined with the node feature map to generate an embedding feature map H, the embedding feature map is a matrix structure formed by stacking all node embedding feature vectors in node index order, reflecting the communication relationship strength and state association between vehicle-mounted devices.

[0093] The application realizes hierarchical aggregation of node features and generates an embedded feature graph by implementing a graph convolution operation in a graph neural network in combination with edge weights and a dynamic attention mechanism, accurately depicts state conduction and communication strength between devices, and enhances the response capability of the model to state changes of key nodes.

[0094] In the embodiment, the attention weight calculation formula of the lth layer is:

[0095]

[0096] wherein a is a trainable attention vector, a T is the transpose of a, and || represents a vector splicing operation and are embedded feature vectors of the (l-1)th layer node i and node k respectively, exp(·) represents a natural exponential function with e as the base, and LeakyReLU(·) is a nonlinear activation function.

[0097] The application introduces a trainable attention vector and realizes attention weight calculation based on node feature splicing, effectively improves the information weighting accuracy in the feature aggregation process, so that the model can focus on key fault signals in a multi-node distribution detection scene and improve the overall detection accuracy.

[0098] In the embodiment, the S4 specifically comprises:

[0099] S41, introducing a cross-scale fusion structure in the graph neural network model, inputting the embedded feature graph H into the cross-scale fusion structure, setting a multi-scale processing window set for each scale performing an intra-scale node feature compression operation to obtain a node feature matrix, the node feature matrix retaining the correlation relationship and structure feature information between nodes:

[0100]

[0101] wherein, is a node feature matrix under scale s p , is an activation function under scale s p , a GELU function is used, is a feature compression matrix under scale s p , is a bias vector under scale s p .

[0102] S42, performing cross-scale fusion on all scale feature matrices to construct a fusion feature matrix:

[0103]

[0104] wherein F is a fusion feature matrix, is a scale s p under a fusion weight, satisfying is a Hadamard element-by-element multiplication operation, is a scale mask matrix, each element in the scale mask matrix is defined as:

[0105]

[0106] wherein is a structural difference response of node i and adjacent node j at scale s p under a fusion weight, and is a feature vector row of node i and node j at scale s p under a fusion weight, denotes the absolute difference value of the corresponding vector elements, tanh(·) is a hyperbolic tangent function for normalization, denotes a set of direct adjacent nodes of node i.

[0107] The application effectively enhances the model's ability to capture multi-scale structural disturbances and abnormal patterns, and improves the model's adaptability by constructing a fusion feature matrix through node feature compression and cross-scale fusion operations under multi-scale windows, preserving the structural differences between nodes and weighting their response degrees.

[0108] In the embodiment, the S5 specifically includes:

[0109] S51, input the fusion feature matrix F into the residual aggregation subnetwork, the residual aggregation subnetwork is composed of two layers of nonlinear mapping layers and corresponding jump connection structures, and adopts a nonlinear activation function at each layer output, performs a multi-layer nonlinear mapping operation, and constructs the lth layer residual feature as:

[0110]

[0111] wherein is a feature vector of node i at the lth layer residual network, F i is a fusion feature vector of node i, is a weight matrix of the lth layer residual network, is a bias vector, ξ is a Swish activation function, and ψ is a linear mapping function, which performs dimension alignment on the output of the previous layer .

[0112] S52, perform a jump connection operation at the last layer L r of the residual aggregation subnetwork, and output the running state of the device and the communication score of the node:

[0113]

[0114] wherein y i is the running state of the device corresponding to node i, softmax(·) represents a nonlinear activation function, δ i is the communication score of node i, W o is an output weight matrix, b o is an output bias vector, and d is the feature dimension of is the jth feature value of node i, μ j is the mean of the jth feature of all nodes, and σ is the normalized variance of the jth feature of all nodes, ε is a constant stabilization term to prevent the denominator from being zero, and n is the number of nodes.

[0115] S53, determining the running state label of the device according to y i , and attaching the corresponding device identifier, comparing δ i with a set threshold θ, and if δ i > θ, determining that node i is a communication abnormal node, and outputting the index information of the communication abnormal node.

[0116] The application adopts a residual aggregation structure to realize deep feature extraction and skip connection, and combines a communication score function to realize abnormal level determination, so that the model has high resolution and interpretability in identifying the running state of the device and communication abnormality, facilitating result visualization and operation and maintenance decision.

[0117] In the embodiment, the S6 specifically includes:

[0118] S61, generating a structured detection report according to the running state label and the index of the communication abnormal node, and arranging the onboard device in order of node number, and inserting the label number, label name and classification confidence information correspondingly;

[0119] S62, for each number identified as a communication abnormal node, associating the corresponding node with the timestamp sequence in the original running data, generating a time record corresponding to the abnormal node, and dividing the abnormal level according to the communication abnormal score into three levels, and marking them as general, serious and urgent respectively;

[0120] S63, associating the running state label, the communication abnormal node number, the abnormal level, the timestamp and the device number, further supplementing the information amount of the structured detection report, and uploading the detection report to a handheld terminal visualization interface, displaying the device state, the abnormal node position and the corresponding timestamp in real time on the terminal interface, and displaying the normal indicators and various fault information according to the node state classification and partition.

[0121] ​The application constructs a structured detection report and realizes visual display on a handheld terminal, presents the equipment operation state, abnormal level and time information in real time, improves the on-site fault judgment efficiency and response speed of the operation and maintenance personnel, and has good practicality and popularization value.

[0122] Embodiment 1

[0123] In order to verify the feasibility of the application in implementation, the application is applied to the intelligent detection scene of the on-board signal equipment in the operation guarantee of a certain type of train electric system. The train contains multiple signal acquisition units, communication controllers, power conversion modules and state feedback devices, which are distributed in multiple carriages to complete the control, communication and data feedback tasks during operation in a mixed structure of series and star topology. Due to the variety of equipment models and differences in communication protocols, the traditional detection method adopts manual handheld table testing and serial port monitoring, which is tedious and inefficient in diagnosis, especially in the face of sudden communication failures, it is difficult to accurately determine the source of the fault, and it is easy to cause the expansion of the fault range and the extension of the maintenance time.

[0124] In actual application process, the detection personnel carries the handheld terminal proposed in the application, and comprehensively detects the on-board equipment of the whole train during train stopping. First, the handheld terminal connects the detection port of the signal equipment through the high-speed interface, the system automatically collects the operation data including the communication protocol data, voltage and current data and state feedback signal data, and performs real-time deduplication, abnormal elimination and standardization processing. After data processing, the node feature map is automatically generated, and the communication relationship graph is constructed. Through graph neural network graph convolution operation, the attention weight is calculated combined with the node state and communication strength, realizing the dynamic perception of the running state between devices.

[0125] In many detection tasks, we selected 15 representative trains, each train was equipped with an average of 60 signal devices, and the total number of detection devices was 900. In the detection task, the system intelligently labeled the device running state and communication abnormal nodes, and generated a structured detection report. Under the traditional method, the average time of manual detection per train is 72 minutes, while using the handheld terminal detection method of the application, the average time is 22 minutes, and the efficiency is improved by nearly 3.3 times.

[0126] In terms of detection accuracy, the total number of abnormal devices confirmed by manual confirmation after multiple detections using traditional methods is 74, of which 18 are false positives and 12 are missed; while using the handheld terminal detection device of the application, the number of abnormal devices detected is 78, and the number of effective abnormal nodes confirmed by manual comparison is 76, there are only 2 false positives, and there is no missed detection, the detection accuracy is 97.4%, which is improved by nearly 20% compared with the traditional method.

[0127] In terms of communication abnormal node identification, the traditional method cannot provide communication strength indicators and state propagation path analysis, while the present application outputs communication abnormal score through graph neural network and accurately identifies fault propagation nodes based on threshold. In a typical fault scenario, the original system cannot track the abnormal source when signal link interference occurs, and only indicates signal loss, while the handheld terminal detection structure successfully locates the signal connection intermittent problem occurring between car 3 and car 4, and identifies 5 associated nodes. After manual troubleshooting, it is found that the device connection port is caused by poor contact due to aging, the model reasoning path is consistent with the actual situation, which greatly reduces the troubleshooting time and human error rate.

[0128] The visual detection report shows the node running state, abnormal position and corresponding timestamp, which assists the operation and maintenance personnel to realize point maintenance and early warning backtracking. In terms of detection stability, in the continuous detection of 15 trains, the device data transmission stability rate is 99.2%, the average online updating time of the detection model is less than 6 seconds, which ensures that the model can adapt to different train structures and device topologies in real time.

[0129] Table 1 Comparison of railway electric power car-mounted equipment detection results

[0130]

[0131] In summary, the railway electric power car-mounted equipment detection handheld terminal of the present application has significant technical advantages in improving detection efficiency, improving accuracy, realizing multi-node communication topology perception and visual identification, and can effectively replace traditional inefficient detection methods, is suitable for modern railway operation and maintenance support tasks, and has wide application prospect and promotion value.

[0132] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A handheld terminal and detection method for railway electric vehicle equipment detection, characterized in that: The steps include: S1. Connect the vehicle-mounted equipment through the handheld terminal to collect operation data and perform pre-processing; S2. Build a graph neural network model, map the preprocessed running data into a node feature graph, set the attention weight distribution rule, and initialize the weight coefficients of each node state and edge in the graph; S3. Perform graph convolution operations in the graph neural network model, aggregate the features of adjacent nodes based on the edge weight coefficients, and dynamically adjust the attention weights based on the node feature graph to generate an embedded feature graph; S4. Introducing a cross-scale fusion structure into the graph neural network model, compressing multi-scale node features of the embedded feature graph, retaining the correlation relationship and structural feature information between nodes, and generating a fusion feature matrix; S5. Input the fused feature matrix into the residual aggregation sub-network, perform multi-layer feature extraction and skip connection operations, and output the operating status label of the device and the index information of the node with communication anomaly; S6. Generate a structured detection report based on the operation status tag and the communication abnormal node index, and visualize it on the handheld terminal. The visualization information includes the device status, the abnormal node location and the corresponding timestamp; S7. Continue to optimize the network model parameters and attention weight distribution rules, and store the structured detection report in the local buffer.

2. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The operation data includes communication protocol data, voltage and current data, and status signal data.

3. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The preprocessing includes data deduplication, outlier removal and data standardization.

4. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The S2 specifically includes: S21, based on the pre-processed running data, define three feature vector sequences: communication protocol data sequence D c ={d c1 ,d c2 ,…,d cn }, voltage and current data sequence D e ={d e1 ,d e2 ,…,d en }, status signal data sequence D s ={d s1 ,d s2 ,…,d sn }, where n represents the number of data points in the time window, d ci represents the characteristic value of the i-th communication protocol, d ei represents the characteristic value of the i-th electrical signal, d si represents the characteristic value of the i-th state signal; S22. Construct a graph neural network model, concatenate the three feature vector sequences, and generate a node feature graph G = (V, E, A) through a mapping function, where V represents the node set, E represents the edge set, A represents the adjacency matrix, and a ij =1 means that there is a direct connection between node i and node j, a ij =0 means there is no direct connection between the two nodes; S23. Initialize the state value of each node and the weight coefficient of each edge in the graph: Among them, s i is the state value of node i, are the kth communication, electrical, and state signal characteristics of node i, α1, α2, and α3 are the node state weighting coefficients, and w ij is the edge weight between node i and node j, d ij is the physical distance between node i and node j, q ij is the number of successful communications between node i and node j in a given time window, Q is the maximum number of communications in the time window, Δt ij is the time difference of the most recent communication, β1, β2, and β3 are the edge weight coefficients.

5. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 4, characterized in that: The graph neural network model constructs node feature inputs based on three types of feature vector sequences, fuses the three types of feature vector sequences through multi-channel splicing operations, maps the spliced ​​feature vectors into low-dimensional embedding representations using a differentiable mapping function, and introduces a gated attention mechanism to assign feature channel weights to generate a node feature map for graph convolution operations.

6. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The S3 specifically includes: S31. Perform graph convolution operations in the graph neural network model to compute the feature vector x of each node i. i Perform adjacent feature aggregation, and the convolution calculation formula of the first layer of adjacent feature aggregation is: in, is the embedded feature vector of node i in layer l, σ is the ReLU activation function, represents the set of directly adjacent nodes of node i, is the attention weight of node i and node j in layer l, w ij is the weight coefficient of edge (i, j), is the embedded feature vector of node j in the l-1th layer, a ij Represents the connection relationship in the adjacency matrix, W (l-1) is the weight matrix of the l-1 layer, b (l-1) is the bias vector of the l-1th layer; S32. After performing L layers of graph convolution, the embedded feature vectors of all nodes are stacked, and the attention weights are dynamically adjusted in combination with the node feature graph to generate an embedded feature graph H. The embedded feature graph is a matrix structure formed by stacking the embedded feature vectors of all nodes in the order of node index, reflecting the communication relationship strength and state correlation between on-board devices.

7. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 6, characterized in that: The calculation formula for the attention weight of the lth layer is: Among them, a is the trainable attention vector, a T is the transpose of a, ‖ represents the vector concatenation operation are the embedded feature vectors of node i and node k in the l-1th layer respectively, exp(·) represents the natural exponential function with e as the base, and LeakyReLU(·) is the nonlinear activation function.

8. The handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The S4 specifically includes: S41. Introduce a cross-scale fusion structure into the graph neural network model, input the embedded feature map H into the cross-scale fusion structure, and set a multi-scale processing window set For each scale Perform the node feature compression operation within the scale to obtain the node feature matrix, which retains the association relationship and structural feature information between nodes: in, Scale s p The node feature matrix under Scale s p The activation function below uses the GELU function. Scale s p The feature compression matrix under Scale s p The bias vector under ; S42. Perform cross-scale fusion on all scale feature matrices to construct a fusion feature matrix: Among them, F is the fusion feature matrix, Scale s p The fusion weight under ⊙ is the Hadamard element-wise product operation, A scale mask matrix, where each element in the scale mask matrix is ​​defined as: in, is the distance between node i and adjacent node j at scale s p The structural difference response under and For nodes i and j at scale s p The eigenvector row under represents the absolute difference of the corresponding vector elements, and tanh(·) represents the hyperbolic tangent function, which is used for normalization. Represents the set of directly adjacent nodes of node i.

9. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The S5 specifically includes: S51. Input the fusion feature matrix F into the residual aggregation sub-network. The residual aggregation sub-network consists of two layers of nonlinear mapping layers and corresponding skip connection structures. A nonlinear activation function is used at the output of each layer to perform multi-layer nonlinear mapping operations to construct the l-th layer residual feature: in, is the feature vector of node i in the lth layer of the residual network, F i is the fused feature vector of node i, is the weight matrix of the l-th layer residual network, is the bias vector, ξ is the Swish activation function, ψ is the linear mapping function, and the output of the previous layer is Perform dimension alignment; S52, in the last layer L of the residual aggregation sub-network r Perform a jump connection operation and output the device's operating status and the node's communication score: Among them, y i is the operating status of the device corresponding to node i, softmax(·) represents the nonlinear activation function, δ i is the communication score of node i, W o is the output weight matrix, b o is the output bias vector, d is The characteristic dimension of is the jth eigenvalue of node i, μ j is the mean of the j-th dimension features of all nodes, is the normalized variance of the j-th dimension features of all nodes, ε is a constant stabilizing term to prevent the denominator from being zero, and n is the number of nodes; S53, according to y i Determine the operating status label of the equipment and attach the corresponding equipment identification. i Compared with the set threshold θ, if δ i >θ, node i is determined to be a communication abnormal node, and the node index information of the communication abnormality is output.

10. A handheld terminal and detection method for railway electric vehicle equipment detection according to claim 1, characterized in that: The S6 specifically includes: S61. Generate a structured detection report based on the operation status label and the communication abnormality node index. The structured detection report arranges the vehicle-mounted devices in order of node number and inserts the label number, label name, and classification confidence information accordingly. S62. For each node identified as having communication anomaly, associate the corresponding node with the timestamp sequence in the original operation data, generate a time record corresponding to the abnormal node, and classify the anomaly level into three levels based on the communication anomaly score, marking them as general, severe, and urgent respectively; S63. Associate the operation status label, communication abnormal node number, abnormal level, timestamp and device number to further supplement the information content of the structured detection report, and upload the detection report to the handheld terminal visualization interface. The device status, abnormal node location and corresponding timestamp are displayed in real time on the terminal interface, and normal indicators and various types of fault information are displayed according to node status classification and partition.

Images