Train door system abnormality monitoring method based on component timing unit association screening

Abstract

This invention discloses a method for anomaly monitoring of train door systems based on component temporal unit association filtering, relating to the field of urban rail transit technology. The method includes: acquiring train door system monitoring data; dividing the monitoring data into component temporal units and converting it into a structured temporal association graph; decomposing the structured temporal association graph into a central graph; constructing parsing components adapted to different association scenarios based on the central graph; generating a sparse graph by filtering key association edges from the central graph using a noisy gating mechanism; and learning the anomaly features of component temporal unit nodes using a graph neural network based on the filtered sparse graph, ultimately outputting a binary classification result for train door system anomalies. This method effectively solves the problems of coarse granularity, insufficient utilization of association information, and scarcity of anomaly samples in traditional methods for continuous temporal data modeling, achieving accurate identification of early anomaly states in the train door system.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit technology, and more specifically to a method for anomaly monitoring of train door systems based on component time sequence unit correlation screening. Background Technology

[0002] Among the various key equipment in rail transit trains, the door system undertakes high-frequency, human-machine direct interaction operations, and its operational stability has a significant impact on normal train operation and passenger safety. Because the door system needs to repeatedly perform opening and closing actions throughout its service life, its internal transmission mechanisms, motors, and control components are inevitably subjected to continuous stress. Coupled with external factors such as environmental temperature changes and vibration shocks, the system's performance gradually changes, potentially leading to future failures. Currently, the maintenance and management of door systems mainly relies on post-fault handling or periodic inspections, methods that typically depend on obvious abnormal signals or human experience. However, before the equipment's function completely fails, the door system often enters a transitional state where performance degrades but it is still operational. Anomalies in this stage are often subtle and difficult to detect accurately using traditional monitoring methods, easily overlooked, thus missing the optimal intervention opportunity. If this state persists without effective measures, system reliability will continue to decrease, potentially leading to sudden failures. Therefore, introducing anomaly monitoring methods for the operational status of the door system is of practical significance.

[0003] Existing methods for anomaly detection in gate systems attempt to statically cluster channels based on the overall similarity of sensor signals, and then model the correlations within each cluster. However, this approach has the following shortcomings: 1) Lack of time-varying nature: The dependencies between sensors change dynamically with the gate's operation. Sometimes, although some channels are correlated overall, they may be completely independent at certain times. Forcing modeling based on the entire process signal in this case will introduce noise. Sometimes, some channels are uncorrelated overall, but may be highly correlated in local time periods. Ignoring this will result in the loss of key correlations. 2) Ignoring inter-cluster relationships: Some methods only focus on intra-cluster relationships, ignoring key correlations that may exist between different clusters in local time periods, leading to incomplete anomaly pattern recognition. Current research mostly focuses on dependencies within a single stage, while ignoring potential cross-stage connections.

[0004] Therefore, how to dynamically monitor the dependencies between sensors and consider potential cross-stage connections to improve the accuracy of anomaly monitoring in train door systems has become a key technical problem that urgently needs to be solved in anomaly monitoring of train door systems. Summary of the Invention

[0005] In view of the above problems, this invention is proposed to provide a train door system anomaly monitoring method based on component-based temporal unit correlation filtering to overcome or at least partially solve the above problems. The method divides the continuous monitoring data of the door system into fine-grained units, achieving correlation modeling at the unit level, thus overcoming the limitations of traditional global modeling. Through multiple parsers and a noisy gating routing strategy, it adaptively filters the correlation information crucial for anomaly detection, effectively eliminating redundant noise. Combined with a customized loss function, it addresses the problems of positive and negative sample imbalance and insufficient utilization of correlation information in door system anomaly detection.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for anomaly monitoring of a train door system based on component timing unit association filtering, comprising the following steps: S1. Obtain monitoring data of the train door system; S2. Divide the monitoring data of the train door system into component time sequence units and convert them into a structured time sequence correlation graph; S3. Decompose the structured temporal correlation graph into a central graph, construct a parsing component adapted to different correlation scenarios based on the central graph, and generate a sparse graph by filtering key correlation edges from the central graph through a noisy gating mechanism. S4. Based on the filtered sparse graph, the abnormal features of the temporal units of the components are learned through a graph neural network, and finally the abnormal binary classification result of the train door system is output.

[0008] Preferably, in S1, the monitoring data of the train door system is defined as follows: ,in Number of sensor types To monitor the length of the time window, For the first Time-series data monitored by each sensor.

[0009] Preferably, S2 includes: S21. The time-series data monitored by each of the sensors According to the length of the division unit Perform non-overlapping partitioning; the resulting data is: ,in, The number of timing units in a component that monitors data from a single sensor. , It is a rounding function. To divide the unit length; Perform a linear mapping on the partitioned data to obtain the embedded representation:

[0010] In the formula, The data after linear mapping, , which represents the total number of timing units in the system components. For embedding operations, For the embedding dimension, used to convert the original length to... The sequence is compressed into a latent representation; S22, the data after linear mapping By number of heads Divided into multiple sub-feature matrix groups , Using the floor function, calculate the similarity matrix for each sub-feature matrix group:

[0011] In the formula, The weight matrix is ​​a learnable matrix. This is a similarity matrix. , It is the transpose matrix. For projection dimensions; The initial association matrix is ​​calculated based on the similarity matrix using the k-Nearest Neighbor method. The calculation formula is:

[0012] In the formula, For the k-Nearest Neighbor method, It is a non-linear activation function. Neighborhood adjustment parameters are used to balance sparsity and information preservation, preserving information for each node. The nearest neighbors, among which For the initial correlation matrix By applying different masks, the correlation matrix of each subgraph is obtained; S23. Define a structured temporal correlation graph. ,in It is a set of nodes, with each node corresponding to a component timing unit; The edges between component time-series units are defined by the initial association matrix. describe, Indicates the first The timing unit of the first component and the first The correlation strength between the timing units of neighboring components; S24. Define the central graph and three types of subgraphs of the structured temporal correlation graph.

[0013] Preferably, the three types of subgraphs include: 1) Same type of time series subgraph The calculation formula is:

[0014] 2) Subgraphs of the same time type The calculation formula is:

[0015] 3) Cross-domain time series subgraph The calculation formula is:

[0016] In the formula, For time edge sets, For spatial edge sets, Indicates one of the edges. It is a set of nodes, and each node corresponds to a component timing unit.

[0017] Preferably, S3 includes: S31. Structure the temporal correlation map Decomposed into Central map of each component's timing unit Each central map is further divided into three sub-maps. Three types of parsing components are constructed to adapt to different related scenarios: temporal parser, spatial parser, and spatiotemporal fusion parser; Among them, the time domain parser When the time-domain resolver is enabled, retain the time side set. Spatial resolver When the spatial domain resolver is enabled, the spatial edge set is preserved. Spatiotemporal fusion parser When the spatiotemporal fusion resolver is enabled, the condition that satisfies the following is retained. edge set, Let be the set of edges; S32. Based on a sparse gating expert hybrid architecture and graph structure sparsification method, a noisy gating mechanism is introduced to calculate the confidence weights of each parser by the temporal unit of the computation component. :

[0018] In the formula, For the softmax function, For the softplus function, For a clean confidence mapping layer, For noise confidence mapping layer, Indicates standard Gaussian noise; S33, The confidence weight Sort in descending order and select those with a cumulative probability exceeding the threshold. The minimum set of parsers:

[0019] In the formula, The number of parsers to be calculated. For the parser's index, For the set threshold, The number of relations selected for a node, i.e., the number of resolvers filtered out, retains the confidence weight of the filtered resolvers, and sets the weight of unselected resolvers to 0. Therefore, the output route is calculated as follows:

[0020] In the formula, For the parser's index, Given the total number of parsers, select the set of parsers that meet the requirements according to the above formula. ,in This is the list of selected parser indices. Indicates the first The selected parsers that meet the requirements are individual parser instances in the parser set, used to participate in the selection of associated edges and the generation of sparse graphs. Indicates the first parser in the selected set. One parser, This is a list of indices for the selected parser; Based on the selected set of minimum parsers, key related edges are selected from the central graph, redundant related edges are removed, and a sparse graph is generated.

[0021] Preferably, S4 includes: S41. The filtered sparse graph Reconstructed into a global relational graph And aggregate to obtain the global sparse adjacency matrix. The calculation formula is:

[0022]

[0023] In the formula, For component timing unit nodes The neighborhood group, This represents the total number of timing units in the system components. It is a set of nodes, where each node corresponds to a component timing unit. It is the first The local sparse adjacency matrix corresponding to the timing unit of each component; S42. A graph neural network consists of an input layer, an improved graph convolutional layer, and an output layer: In the input layer, the initial features of the receiving nodes are... and global sparse adjacency matrix ; Improved graph convolutional layer layer, in the Layers, the network is based on the global sparse adjacency matrix Get Nodes Neighbor set and utilize As a weight, Global sparse adjacency matrix The line, number Column elements represent the target node. with neighboring nodes The correlation strength weight between them; Features of the previous layer of neighboring nodes Perform weighted summation aggregation:

[0024] In the formula, , For the first The neighboring node Layer node representation, For weighted summation aggregate functions, This is the index of the time sequence unit node of the neighboring component; The aggregated neighborhood features are combined with the node's previous layer features. Perform residual connection, and then through The layer performs feature transformation, and finally introduces non-linearity using the ReLU activation function to obtain the node features of the current layer:

[0025] In the formula, For the first The node Layer node representation, Introducing non-linearity into the ReLU activation function maps the aggregated features to a unified dimension. As the layer graph convolution iterates, the node features will continuously fuse with the correlation information of the global neighborhood, and the features of abnormal nodes will be continuously strengthened. Finally, in the output layer, the features of abnormal nodes will be significantly distinguished from normal nodes, thereby completing the anomaly detection. S43. By calculating the average value of all node features, the scattered node-level anomaly information is integrated into global graph-level features, representing the operating state of the entire gate system. The node features output by the last layer of graph convolution are averaged and pooled to obtain global graph features. Then, the anomaly probability is output through two layers of MLP. And based on the threshold Obtain binary classification results :

[0026] In the formula, and For linear layers of MLP, Map the output to a probability interval, if If it is normal, it is considered abnormal; otherwise, it is considered normal.

[0027] Preferably, it also includes: during the training process, designing a triple loss function to address the problems of imbalanced positive and negative samples in gate system anomaly detection and excessive reliance on the parser.

[0028] Preferably, the design of the triple loss function includes: S51. Constructing a binary classification cross-entropy loss. :

[0029] In the formula, For training batch size, For the first The true label of each sample For the first The predicted probability of anomalies for each sample. For the first The class weights of each sample are calculated using the following formula:

[0030] S52, Constructing Dynamic Routing Entropy Loss :

[0031] In the formula, The total number of component time-series units in a single sample. for In the nth sample The central map of each component's timing unit contains information about the relationship between that component's timing unit and other units. No. In the nth sample The timing unit center map of the first component is related to the... Confidence level of each associated expert, Corresponding to time domain experts, spatial experts, and spatiotemporal fusion experts respectively, satisfying ; S53, Constructing a parser to balance the loss :

[0032] In the formula, For the first In the nth sample The standard deviation of the resolver confidence of the temporal unit center map of each component. For the first In the nth sample The mean resolver confidence of the temporal unit center map of each component. For small constants, avoid due to Approaching 0 results in a denominator of 0, ensuring numerical stability; S54. The total loss function is obtained by fusing the binary cross-entropy loss, dynamic routing entropy loss, and parser equilibrium loss through weights:

[0033] In the formula, The weights are the dynamic routing entropy loss. The weights are used to balance the loss of the parser.

[0034] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects: This invention addresses the anomaly detection needs of train door system operation data, effectively solving the problems of coarse granularity, insufficient utilization of correlation information, and scarcity of abnormal samples in traditional methods for continuous time-series data modeling. It achieves accurate identification of early abnormal states in the door system and has the following technical features: 1) Introduce the concept of "component time series unit" to divide continuous monitoring data into fine-grained units and characterize the time series correlation between components at the unit level, thereby enhancing the ability to model local abnormal features.

[0035] 2) Design a dynamic association filtering mechanism that combines multiple parsers and gated routing strategies to adaptively retain key information related to anomaly detection and suppress redundancy and noise interference.

[0036] 3) Construct an integrated framework of "unit construction - association screening - classification learning" and alleviate the sample imbalance problem by using a customized loss function to improve the accuracy and stability of anomaly detection. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 The flowchart of the train door system anomaly monitoring method based on component timing unit association filtering provided by the present invention is shown below. Figure 2 This is an overall architecture diagram of the train door system anomaly monitoring method based on component timing unit association filtering provided by the present invention. Figure 3 This is a basic structural diagram of the gate system anomaly monitoring network of the present invention; Figure 4 This is a schematic diagram comparing the performance of the present invention with existing methods. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] This invention discloses a method for anomaly monitoring of train door systems based on component timing unit correlation filtering, such as... Figure 1 As shown, it includes the following steps: S1. Obtain monitoring data of the train door system; S2. Divide the monitoring data of the train door system into component time sequence units and transform them into a structured time sequence correlation graph. Decompose the structured time sequence correlation graph into a central graph. S3. Based on the central graph, construct a parsing component that adapts to different association scenarios, and use a noisy gating mechanism to filter key association edges from the central graph to generate a sparse graph; S4. Based on the filtered sparse graph, the abnormal features of the temporal unit nodes of the component are learned through graph neural network, and the final output of the gate system abnormal binary classification result is output.

[0041] The following provides further explanation of each of the above steps, with the specific process as follows: Figure 1 and Figure 2 As shown.

[0042] S1. Acquire monitoring data of the train door system, including data monitored by the motor speed sensor, displacement sensor and current sensor.

[0043] Define the monitoring data of the train door system as ,in Number of sensor types To monitor the length of the time window, For the first Time-series data monitored by each sensor.

[0044] S2, Module for constructing the temporal correlation graph of components.

[0045] The core function of the component temporal correlation map is to "transform multi-source heterogeneous monitoring data into a structured correlation map," solving the problems of "heterogeneous data formats and ambiguous correlation information," and providing high-quality input for subsequent correlation filtering and classification. To achieve fine-grained correlation modeling and avoid the problem of "masking local anomalies" in global data modeling, component temporal units are introduced. The data monitored by each sensor is divided into non-overlapping component temporal units. The specific steps are as follows: S21. Divide the monitored time-series data and perform linear mapping. For each sensor's monitored time-series data... According to the length of the division unit Perform non-overlapping partitioning; the resulting data is: The number of timing units of the component that monitors data by a single sensor is , , It is a rounding function. To divide the unit length, the total number of timing units in the system components The partitioned data is linearly mapped through the nonlinear projection layer #1 in the temporal feature preprocessing stage to obtain the embedded representation:

[0046] In the formula, The data after linear mapping, , For embedding operations, For the embedding dimension, used to convert the original length to... The sequence is compressed into a latent representation.

[0047] S22. Perform similarity measurement. Building upon S21, construct the association matrix of the temporal association graph. Calculate the similarity measure using a multi-head mechanism (parallel extraction of different feature representations) to achieve a more diverse expression of similarity. This parallel processing of multiple sets of sub-feature matrices allows for the calculation of the association strength between temporal unit nodes of the component from multiple different perspectives and dimensions, thereby more comprehensively and meticulously exploring the association patterns between nodes and avoiding the limitations of calculating associations using only a single method.

[0048] Through the nonlinear projection layer #2 and batch normalization layer #1 in the temporal feature preprocessing stage, By number of heads Divided into multiple sub-feature matrix groups , As a floor function, each sub-feature matrix group can independently capture association information of a specific dimension or pattern. The similarity metric calculation formula is as follows:

[0049] In the formula, The weight matrix is ​​a learnable matrix. This is a similarity matrix. , It is the transpose matrix. For the projection dimension.

[0050] By using the nonlinear projection layer #3 in the temporal feature preprocessing stage, k-Nearest Neighbor is used to retain the local key neighbors of each node, balancing the sparsity of the association matrix with information preservation (without losing important associations), and obtaining the initial association matrix. The calculation formula is:

[0051] In the formula, For the k-Nearest Neighbor method, It is a non-linear activation function. Neighborhood adjustment parameters are used to balance sparsity and information preservation, preserving information for each node. The nearest neighbors, among which , This is the floor function. For the initial incidence matrix... By applying different masks, the correlation matrix of each subgraph can be obtained.

[0052] S23. Define a structured temporal correlation graph. ,in It is a set of nodes, with each node corresponding to a component timing unit; The edges connecting the time-series units of the components are defined by the initial correlation matrix. describe, Indicates the first The timing unit of the first component and the first The correlation strength between timing units of each component.

[0053] S24. Define the central graph and three types of subgraphs of the structured temporal correlation graph. To avoid the influence of interfering information, the structured temporal correlation graph will be used in subsequent operations. Decomposed into Central map of each component's timing unit Each central map is further subdivided into three types of submaps. The definitions of each type of submap are as follows: 1) Same type of time series subgraph The correlation between timing units of components of the same type at different time points (such as the correlation between the timing units of the first and second time points of a motor displacement signal component) is used to capture timing anomalies in a single component. The calculation formula is as follows:

[0054] 2) Subgraphs of the same time type The correlation between timing units of different types of data components at the same time point (such as the correlation between the timing unit of the motor displacement signal at the first time point and the timing unit of the motor current at the first time point) is used to capture timing anomalies in different components. The calculation formula is as follows:

[0055] 3) Cross-domain time series subgraph The correlation between timing units of different types and different time windows (such as the correlation between the timing unit of the motor displacement signal at the first time point and the timing unit of the motor speed signal at the second time point) is used to capture delay anomalies. The calculation formula is as follows:

[0056] In the formula, For time edge sets, For spatial edge sets, Indicates one of the edges. It is a set of edges, and the edges of different dimensions (time / space) and different subgraphs are distinguished by superscript and subscript.

[0057] S3, Construct a component timing unit specific parsing module.

[0058] Cross-domain time series data often exhibit differentiated intrinsic characteristics and structural forms, thus requiring targeted modeling of multi-variable correlation patterns.

[0059] S31. Structure the temporal correlation map Decomposed into Central map of each component's timing unit Each central map is further divided into three sub-maps. Three types of parsing components were constructed to adapt to different association scenarios: temporal parsers, spatial parsers, and spatiotemporal fusion parsers. The allocation of parsers is determined by the router's decisions; specifically, based on the routing decisions, temporal parsers, spatial parsers, and spatiotemporal fusion parsers corresponding to different association regions are assigned to different component temporal units. Different parsers retain corresponding edges to handle different types of association relationships. Through specialized division of labor, the parsers provide adaptive modeling capabilities for each association pattern. Temporal parser When the time-domain resolver is enabled, the system will retain time edges. Spatial resolver When the spatial domain resolver is enabled, the system will preserve spatial edges. Spatiotemporal fusion parser When the spatiotemporal fusion resolver is enabled, the system will retain the condition that satisfies... The edge set.

[0060] S32. Drawing on the sparse gating expert hybrid architecture and graph structure sparsification method, a noisy gating mechanism is introduced, with the timing unit of the computational component assigning confidence weights to each parser. :

[0061] In the formula, For the softmax function, For the softplus function, For a clean confidence mapping layer, For noise confidence mapping layer, This represents standard Gaussian noise.

[0062] S33. The classic Top-K routing expert hybrid architecture assumes that the same number of experts are assigned to each input token, but this approach ignores the differences between inputs. Specifically, in dependency modeling, some time patches require different numbers of dependencies. To adaptively customize the dependencies for each component's temporal unit, a dynamic expert allocation module is proposed. This module selects a resolver based on the component's own confidence level. This method allows the model to evaluate whether the currently selected dependencies are sufficient; if not, it continues to assign more dependencies. Confidence weights are then used to assign these dependencies. Sort in descending order and select those with a cumulative probability exceeding the threshold. The minimum set of parsers:

[0063] In the formula, The number of parsers to be calculated. The number of relations selected for a node, i.e., the number of resolvers filtered out, retains the confidence weight of the filtered resolvers, and sets the weight of unselected resolvers to 0. Therefore, the output route is calculated as follows:

[0064] In the formula, Given the total number of parsers, select the set of parsers that meet the requirements according to the above formula. ,in For indexed lists, Indicates the first The selected parsers that meet the requirements are individual parser instances in the parser set, used to participate in the selection of associated edges and the generation of sparse graphs. Indicates the first parser in the selected set. The last (i.e., the last) parser, This is the list of indices for the selected parser.

[0065] Based on the selected minimal parser set From the central graph, key association edges are selected, redundant associations are removed, and a sparse graph is generated. That is, for each central graph... Based on dynamic output routing Based on the selected minimum set of parsers The analysis process involves filtering key related edges from the central graph, removing redundant connections, and obtaining the association matrix. , corresponding edge set And the sparse graph after parsing out redundant dependencies This corresponds to the temporal feature update layer (embedding end) in the temporal feature preprocessing stage, and its inverse process corresponds to the temporal feature update layer (restoration end). For example, if the minimum parser set ={temporal resolver, spatiotemporal fusion resolver}, then the resolved central graph Preserving time edges and cross-domain edges, we have:

[0066] Based on the layer correspondence between the encoding and decoding ends, the subsequent splitting and feature restoration operations of the sparse graph are executed sequentially on the decoding end: After the temporal feature update layer (restoration end) completes the feature reconstruction and information backhaul of the sparse graph, the nonlinear inverse projection layer #1 realizes the dimensional reconstruction and feature mapping of the sparse graph features to the central graph features. The batch normalization layer #2 performs standardization processing on the reconstructed central graph features to eliminate feature distribution offsets and dimensional differences, providing a standardized feature basis for the splitting of the central graph. Subsequently, the nonlinear inverse projection layer #2 completes the splitting operation of the central graph into three types of subgraphs based on the subgraph definition of S24. Finally, the nonlinear inverse projection layer #3 restores the features of the split subgraphs to the feature dimensions that match the temporal units of the original components, completing the graph screening and splitting process of the entire temporal feature preprocessing stage. Ultimately, each central graph completes the edge screening in this way, generating a sparse graph containing only key associations. It is worth noting that, for The parsing operations of each central map can be performed in parallel without adding extra computational complexity.

[0067] S4. Output the binary classification results of train door system anomalies.

[0068] Based on the filtered sparse graph, this module learns the abnormal features of the temporal unit nodes of the components through a graph neural network, and finally outputs the abnormal binary classification results of the train door system.

[0069] S41. The filtered sparse center map Reconstructed into a global relational graph And aggregate to obtain the global sparse adjacency matrix. The calculation formula is:

[0070] In the formula, For component timing unit nodes The neighbor set of all nodes is merged to ensure that the global graph contains the key relationships between the temporal units of all components of the gate system, thus avoiding the omission of important relationships in the local graph. By aggregating elements by taking the maximum value, the key associations of the temporal unit center graphs of all components are preserved. This maintains matrix sparsity to reduce computational cost without losing key association information, providing accurate association weights for subsequent feature learning. Global sparse adjacency matrix. It is a global relational graph Core mathematical expression and topology-driven, global correlation graph It provides the topology of the graph neural network, defines the neighbor relationships of nodes, and determines the range of feature aggregation; a global sparse adjacency matrix. It provides weight information for graph neural networks, offering a weighting basis for neighborhood feature aggregation and ensuring that strongly correlated anomalies are captured first. Each local sparse subgraph... Corresponding to a local sparse adjacency matrix This matrix defines the association weights of nodes within a subgraph; when all subgraphs are aggregated into a global association graph... At that time, the global sparse adjacency matrix The weights of all local matrices are aggregated synchronously, and the maximum value is taken to ensure that the correlation strength between global nodes is the strongest value under all local views.

[0071] S42, convert the global sparse adjacency matrix As the core input of the improved graph convolutional network, it defines the global relational topology between all component temporal unit nodes. Through an improved structure of "neighborhood feature aggregation + self-feature residual fusion," the network adaptively learns latent features of nodes from the global sparse graph, prioritizing and enhancing the representation of anomalous features, ultimately highlighting anomalous information in the global features. The network adopts a structure of "input layer → improved graph convolutional layer (multi-layer iteration) → output layer," as follows... Figure 3 As shown: In the input layer, the initial features of the receiving nodes are... (From embedded component time-series unit data), and global sparse adjacency matrix .

[0072] Improved graph convolutional layers (total) Each layer (layers) contains two core modules: "neighborhood feature aggregation" and "self-feature residual fusion," which are the core improvement points of the network. Layers, the network is based on the global sparse adjacency matrix Get Nodes Neighbor set and utilize As a weight, the previous layer features of neighboring nodes Perform weighted summation aggregation:

[0073] In the formula, , The number of layers in a graph neural network. For the first The neighboring node Layer node representation, For weighted summation aggregate functions, This is the index of neighboring nodes. The weighted aggregation here gives higher weight to strongly correlated neighbor features, thus prioritizing the capture of strong correlation information of abnormal nodes. The aggregated neighborhood features are then compared with the node's own previous-layer features. Perform residual connection, and then through The layer performs feature transformation, and finally introduces non-linearity using the ReLU activation function to obtain the node features of the current layer:

[0074] In the formula, ( (Number of layers in a graph neural network) No. Layer node representation, , For weighted summation aggregate functions, By weighting neighbor features based on association weights, strong association anomalies can be highlighted. Introducing nonlinearity into the ReLU activation function enhances its ability to express anomalous features. This is a feature transformation layer that maps the aggregated features to a unified dimension.

[0075] S43. Node-level features are transformed into graph-level features using "global mean pooling + multilayer perceptron (MLP)," outputting the anomaly binary classification probability and result. By calculating the average of all node features, the scattered node-level anomaly information is integrated into global graph-level features, representing the overall operating state of the gate system (e.g., "most node features normal → global features normal," "multiple associated node features abnormal → global features abnormal"). Mean pooling is applied to the node features output by the last graph convolution layer to obtain global graph features, which are then output as anomaly probabilities through two MLP layers. And based on the threshold Obtain binary classification results :

[0076] In the formula, and For linear layers of MLP, Map the output to a probability interval, if If it is normal, it is considered abnormal; otherwise, it is considered normal.

[0077] S5. Design a triple loss function.

[0078] To address the severe imbalance in the number of positive and negative samples (where normal samples constitute the vast majority and abnormal samples are relatively scarce) that is common in gate system anomaly detection tasks, and the problem that models tend to over-rely on a few parsers and ignore information from other parsers during training and inference, a triple loss function is designed to impose multi-dimensional constraints on the model learning process.

[0079] S51. Constructing a binary classification cross-entropy loss. To measure the prediction error of binary classification, sample weights are introduced. To address the imbalance between positive and negative samples, weight adjustment is used to give abnormal samples a greater "voice" in loss calculation. During model training, the model will prioritize learning the features of abnormal samples, effectively reducing the false negative rate, while avoiding an increase in the false positive rate of normal samples due to excessive weight.

[0080]

[0081] In the formula, For training batch size, For the first The true label of each sample For the first The predicted probability of anomalies for each sample. For the first The class weights of each sample are calculated using the following formula:

[0082] S52, Constructing Dynamic Routing Entropy Loss This approach encourages each component's temporal unit center graph to activate more associated resolvers, avoiding "resolver idleness" caused by the model relying on only a few resolvers and enhancing the model's adaptability to different types of anomalies. The associated resolvers have different functions (temporal experts capture progressive anomalies, spatial experts capture synchronicity anomalies, and spatiotemporal fusion experts capture latency anomalies). If the model only activates one resolver, other types of anomalies may be missed. Therefore, entropy loss is introduced. Entropy loss maximizes the entropy of the confidence distribution, making expert activation more dispersed. The dynamic routing entropy loss calculation formula is:

[0083] In the formula, The total number of component time-series units in a single sample. for In the nth sample The central map of each component's timing unit contains information about the relationship between that component's timing unit and other units. No. In the nth sample The timing unit center map of the first component is related to the... Confidence level of each associated expert, Corresponding to time domain experts, spatial experts, and spatiotemporal fusion experts respectively, satisfying .

[0084] S53, Parser Equalization Loss The "standard deviation / mean ratio" quantifies the balance of parser usage, preventing the model from over-relying on a particular type of parser (e.g., always prioritizing cross-domain parsers) and ensuring that all three types of parsers are used evenly throughout the full training run. While dynamic routing entropy loss encourages dispersed parser activation, it may lead to "local equilibrium but global imbalance" (e.g., one type of component's temporal units always activate the temporal parser, while another type always activates the spatial parser). Therefore, the "standard deviation / mean ratio" is introduced; a smaller ratio indicates more balanced parser activation. By minimizing the reciprocal of this ratio, balanced global parser usage is ensured.

[0085]

[0086] In the formula, For the first In the nth sample The standard deviation of the resolver confidence of the temporal unit center map of each component. For the first In the nth sample The mean resolver confidence of the temporal unit center map of each component. For small constants, avoid due to Approaching 0 results in a denominator of 0, ensuring numerical stability.

[0087] S54. Finally, the classification loss, routing entropy loss, and parser equalization loss are fused by weights to obtain the total loss function, ensuring that the three types of losses work together—ensuring classification accuracy while achieving the dispersion and balance of expert activation, and avoiding model bias caused by a single loss.

[0088]

[0089] In the formula, The weights for dynamic routing entropy loss are used to ensure that the contribution of routing entropy loss is moderate, neither suppressing the classification loss nor failing to effectively encourage the parser to activate the dispersion. The weights are used to balance the parser's loss, thus balancing the global parser's balance with local classification accuracy.

[0090] Next, the effectiveness of the method of this invention is verified using a rail vehicle door system based on the door system operation data obtained through actual vehicle simulation. This simulation simulated eight states, including normal, alignment anomaly, V-shaped dimension anomaly, sealing dimension anomaly, buffer head wear, door opening / closing time anomaly, overall resistance anomaly, and local resistance anomaly. The door system simulation method is shown in Table 1. Data acquisition included three types: motor displacement, motor speed, and motor current signals. Data was collected every time the door opened and closed, with a sampling frequency of 100Hz. The door opening / closing time was 3.5 seconds, so the collected data length was generally 350 data points. However, to ensure data completeness, the collected data length was 380. Since it was difficult to guarantee that the collected data length was completely consistent each time, to ensure data quality, the input data length was set to 400. That is, when the data length was greater than 400, the data was discarded; when the data length was less than 400, the last data point was used to pad the data length to 400. To verify the robustness and superiority of the proposed gate system anomaly detection method, this invention is compared with five more advanced gate system temporal anomaly detection methods, including: an anomaly detection model based on attention decoupling mechanism (AnomalyTransformer, AT), a graph attention network model for multivariate time-series anomaly detection (Multivariate Time-series Anomaly Detection with Graph Attention, MTAD-GAT), an end-to-end temporal anomaly detection model based on generative adversarial networks (TadGAN), a transformer anomaly detection model that integrates prediction and reconstruction mechanisms (TranAD), and a multivariate anomaly detection model based on probabilistic generative modeling (OmniAnomaly), corresponding to methods 1 to 5 respectively. All methods use the same feature extraction network and hyperparameter configuration. Accuracy, recall, and macro-average F1 score are used as performance evaluation metrics. The basic structure of the gate system anomaly detection network of the proposed method is shown in Table 2, the hyperparameter settings are shown in Table 3, and the experimental results are summarized in Table 4. Figure 4 The results show that the method of the present invention is significantly better than the comparison method in all three evaluation metrics. Specifically, the accuracy and recall metrics are improved by an average of about 4%–7% compared with the best comparison method, and the macro F1 value is improved by more than 4%, which fully verifies the advantages of the proposed method in anomaly detection of gate systems.

[0091] Table 1. Gate System Simulation Methods

[0092] Table 2. Basic Structure of Gate System Anomaly Monitoring Network

[0093] Table 3. Summary of Hyperparameters

[0094] Table 4. Performance Comparison of Six Methods

[0095] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0096] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for anomaly monitoring of train door systems based on component timing unit correlation filtering, characterized in that, Includes the following steps: S1. Obtain monitoring data of the train door system; S2. Divide the monitoring data of the train door system into component time sequence units and convert them into a structured time sequence correlation graph; S3. Decompose the structured temporal correlation graph into a central graph, construct a parsing component adapted to different correlation scenarios based on the central graph, and generate a sparse graph by filtering key correlation edges from the central graph through a noisy gating mechanism. S4. Based on the filtered sparse graph, the abnormal features of the temporal units of the components are learned through a graph neural network, and finally the abnormal binary classification result of the train door system is output.

2. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 1, characterized in that, In S1, the monitoring data of the train door system is defined as follows: ,in Number of sensor types To monitor the length of the time window, For the first Time-series data monitored by each sensor.

3. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 2, characterized in that, S2 include: S21. The time-series data monitored by each of the sensors According to the length of the division unit Perform non-overlapping partitioning; the resulting data is: ,in, The number of timing units in a component that monitors data from a single sensor. , It is a rounding function. To divide the unit length; Perform a linear mapping on the partitioned data to obtain the embedded representation: In the formula, The data after linear mapping, , which represents the total number of timing units in the system components. For embedding operations, For the embedding dimension, used to convert the original length to... The sequence is compressed into a latent representation; S22, the data after linear mapping By number of heads Divided into multiple sub-feature matrix groups , Using the floor function, calculate the similarity matrix for each sub-feature matrix group: In the formula, The weight matrix is ​​a learnable matrix. This is a similarity matrix. , It is the transpose matrix. For projection dimensions; The initial association matrix is ​​calculated based on the similarity matrix using the k-Nearest Neighbor method. The calculation formula is: In the formula, For the k-Nearest Neighbor method, It is a non-linear activation function. Neighborhood adjustment parameters are used to balance sparsity and information preservation, preserving information for each node. The nearest neighbors, among which For the initial correlation matrix By applying different masks, the correlation matrix of each subgraph is obtained; S23. Define a structured temporal correlation graph. ,in It is a set of nodes, with each node corresponding to a component timing unit; The edges between component time-series units are defined by the initial association matrix. describe, Indicates the first The timing unit of the first component and the first The correlation strength between the timing units of neighboring components; S24. Define the central graph and three types of subgraphs of the structured temporal correlation graph.

4. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 3, characterized in that, The three types of subgraphs include: 1) Same type of time series subgraph The calculation formula is: 2) Subgraphs of the same time type The calculation formula is: 3) Cross-domain time series subgraph The calculation formula is: In the formula, For time edge sets, For spatial edge sets, Indicates one of the edges. It is a set of nodes, and each node corresponds to a component timing unit.

5. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 4, characterized in that, S3 include: S31. Structure the temporal correlation map Decomposed into Central map of each component's timing unit Each central map is further divided into three sub-maps. Three types of parsing components are constructed to adapt to different related scenarios: temporal parser, spatial parser, and spatiotemporal fusion parser; Among them, the time domain parser When the time-domain resolver is enabled, retain the time side set. Spatial resolver When the spatial domain resolver is enabled, the spatial edge set is preserved. Spatiotemporal fusion parser When the spatiotemporal fusion resolver is enabled, the condition that satisfies the following is retained. edge set, Let be the set of edges; S32. Based on a sparse gating expert hybrid architecture and graph structure sparsification method, a noisy gating mechanism is introduced to calculate the confidence weights of each parser by the temporal unit of the computation component. : In the formula, For the softmax function, For the softplus function, For a clean confidence mapping layer, For noise confidence mapping layer, Indicates standard Gaussian noise; S33, The confidence weight Sort in descending order and select those with a cumulative probability exceeding the threshold. The minimum set of parsers: In the formula, The number of parsers to be calculated. For the parser's index, For the set threshold, The number of relations selected for a node, i.e., the number of resolvers filtered out, retains the confidence weight of the filtered resolvers, and sets the weight of unselected resolvers to 0. Therefore, the output route is calculated as follows: In the formula, For the parser's index, Given the total number of parsers, select the set of parsers that meet the requirements according to the above formula. ,in This is the list of selected parser indices. Indicates the first The selected parsers that meet the requirements are individual parser instances in the parser set, used to participate in the selection of associated edges and the generation of sparse graphs. Indicates the first parser in the selected set. One parser, This is a list of indices for the selected parser; Based on the selected set of minimum parsers, key related edges are selected from the central graph, redundant related edges are removed, and a sparse graph is generated.

6. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 5, characterized in that, S4 include: S41. The filtered sparse graph Reconstructed into a global relational graph And aggregate to obtain the global sparse adjacency matrix. The calculation formula is: In the formula, For component timing unit nodes The neighborhood group, This represents the total number of timing units in the system components. It is a set of nodes, where each node corresponds to a component timing unit. It is the first The local sparse adjacency matrix corresponding to the timing unit of each component; S42. A graph neural network consists of an input layer, an improved graph convolutional layer, and an output layer: In the input layer, the initial features of the receiving nodes are... and global sparse adjacency matrix ; Improved graph convolutional layer layer, in the Layers, the network is based on the global sparse adjacency matrix Get Nodes Neighbor set and utilize As a weight, Global sparse adjacency matrix The line, number Column elements represent the target node. with neighboring nodes The correlation strength weight between them; Features of the previous layer of neighboring nodes Perform weighted summation aggregation: In the formula, , For the first The neighboring node Layer node representation, For weighted summation aggregate functions, This is the index of the time sequence unit node of the neighboring component; The aggregated neighborhood features are combined with the node's previous layer features. Perform residual connection, and then through The layer performs feature transformation, and finally introduces non-linearity using the ReLU activation function to obtain the node features of the current layer: In the formula, For the first The node Layer node representation, Introducing non-linearity into the ReLU activation function maps the aggregated features to a unified dimension. As the layer graph convolution iterates, the node features will continuously fuse with the correlation information of the global neighborhood, and the features of abnormal nodes will be continuously strengthened. Finally, in the output layer, the features of abnormal nodes will be significantly distinguished from normal nodes, thereby completing the anomaly detection. S43. By calculating the average value of all node features, the scattered node-level anomaly information is integrated into global graph-level features, representing the operating state of the entire gate system. The node features output by the last layer of graph convolution are averaged and pooled to obtain global graph features. Then, the anomaly probability is output through two layers of MLP. And based on the threshold Obtain binary classification results : In the formula, and For linear layers of MLP, Map the output to a probability interval, if If it is normal, it is considered abnormal; otherwise, it is considered normal.

7. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 1, characterized in that, Also includes: During training, a triple loss function was designed to address the issues of imbalanced positive and negative samples in gate system anomaly detection and excessive reliance on the parser.

8. The train door system anomaly monitoring method based on component timing unit association filtering according to claim 7, characterized in that, The design of the triple loss function includes: S51. Constructing a binary classification cross-entropy loss. : In the formula, For training batch size, For the first The true label of each sample For the first The predicted probability of anomalies for each sample. For the first The class weights of each sample are calculated using the following formula: S52, Constructing Dynamic Routing Entropy Loss : In the formula, The total number of component time-series units in a single sample. for In the nth sample The central map of each component's timing unit contains information about the relationship between that component's timing unit and other units. No. In the nth sample The timing unit center map of the first component is related to the... Confidence level of each associated expert, Corresponding to time domain experts, spatial experts, and spatiotemporal fusion experts respectively, satisfying ; S53, Constructing a parser to balance the loss : In the formula, For the first In the nth sample The standard deviation of the resolver confidence of the temporal unit center map of each component. For the first In the nth sample The mean resolver confidence of the temporal unit center map of each component. For small constants, avoid due to Approaching 0 results in a denominator of 0, ensuring numerical stability; S54. The total loss function is obtained by fusing the binary cross-entropy loss, dynamic routing entropy loss, and parser equilibrium loss through weights: In the formula, The weights are the dynamic routing entropy loss. The weights are used to balance the loss of the parser.

Images