Railway signal equipment fault prediction and diagnosis method

Abstract

The invention provides a railway signal equipment fault prediction and diagnosis method, which belongs to the technical field of railway signal equipment, and comprises the following steps: acquiring multi-dimensional real-time data in the operation process of railway signal equipment, and storing the multi-dimensional real-time data according to a time sequence to form a super sparse historical data set; a hierarchical sparse data clustering index structure is established by adopting a multi-level partition clustering mode to carry out super-dilute clustering processing, an equipment fault feature vector library is established, feature vector similarity is calculated through a nearest neighbor search algorithm, and a corresponding fault prediction level is output. And carrying out correlation analysis by using the space-time alignment super-sparse network model to identify a fault propagation path between the equipment, establishing an equipment fault correlation graph, and generating a railway signal equipment fault prediction and diagnosis report according to a fault risk assessment value and an equipment fault correlation diagnosis result. The technical problem of low fault prediction accuracy of railway signal equipment in a super sparse historical data environment is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of railway signaling equipment, and more specifically, relates to a method for predicting and diagnosing faults in railway signaling equipment. Background Technology

[0002] Fault prediction and diagnosis technology for railway signaling equipment is a core technology for ensuring railway transportation safety. Traditional methods mainly employ rule-based expert systems for equipment status monitoring, judging equipment anomalies by setting fixed thresholds, or using traditional machine learning algorithms such as support vector machines and decision trees to model and predict equipment based on historical data. These methods are widely used in complex operating environments such as high-speed railways, subway systems, and freight lines, achieving equipment status assessment by monitoring basic parameters such as voltage, current, and temperature. However, traditional methods have significant limitations when dealing with the ultra-sparse data generated by the long-term operation of railway signaling equipment. Because railway signaling equipment is in a stable state for most of the time under normal operating conditions, the effective data points are extremely sparsely distributed across the entire time axis. Traditional algorithms cannot effectively extract hidden fault signs from sparse data and lack the ability to deeply analyze the fault propagation paths between equipment. In existing technologies, due to the ultra-sparse nature of railway signaling equipment operating data, with more than 90% of the data being zero or duplicate values, traditional fault prediction methods struggle to accurately identify equipment performance degradation trends and early fault characteristics from sparse data, resulting in severely insufficient fault prediction accuracy. In other words, existing technologies suffer from the technical problem of low accuracy in predicting railway signaling equipment failures under conditions of extremely sparse historical data. Summary of the Invention

[0003] In view of this, the present invention provides a method for predicting and diagnosing faults in railway signaling equipment, which can solve the technical problem of low accuracy in predicting faults in railway signaling equipment under super sparse historical data conditions in the prior art.

[0004] This invention is implemented as follows: It provides a method for fault prediction and diagnosis of railway signaling equipment. This method collects multi-dimensional real-time data during the operation of railway signaling equipment and stores it according to time series to form a super-sparse historical dataset. The super-sparse historical dataset undergoes time series decomposition to separate trend components, seasonal components, and random noise components. A multi-level partitioning clustering method is used to perform ultra-sparse clustering on the super-sparse historical dataset, establishing a hierarchical sparse data clustering index structure. A long short-term memory network model is constructed to process the trend and seasonal components. A time window sliding mechanism is used to capture the long-term degradation trend of equipment performance. A multi-head attention mechanism is used to identify key degradation features, outputting the fault prediction probability and degradation trend prediction value. The method establishes a system for equipment... The fault feature vector library uses a kd-tree data structure to construct a fast retrieval index for a high-dimensional feature space. A nearest neighbor search algorithm is used to match the current device state vector with feature vectors in the library, calculating the feature vector similarity. A fault prediction and evaluation function is designed to calculate the probability of device failure, outputting the corresponding fault prediction level when the fault risk assessment value falls within different numerical ranges. A spatiotemporally aligned ultra-sparse network model is used to perform correlation analysis on the fault prediction levels of multiple railway signaling devices, identifying fault propagation paths and impact ranges between devices, establishing a device fault correlation graph, and outputting device fault correlation diagnosis results. A railway signaling device fault prediction and diagnosis report is generated based on the fault risk assessment value and the device fault correlation diagnosis results.

[0005] The multi-dimensional real-time data includes voltage signal strength, current fluctuation amplitude, temperature change trend, vibration spectrum characteristics, relay action delay, and equipment response time.

[0006] Specifically, the time series decomposition process involves using a moving average algorithm to extract trend components from multi-dimensional real-time data, using a periodicity detection algorithm to identify seasonal components from multi-dimensional real-time data, and using residual analysis to obtain random noise components from multi-dimensional real-time data.

[0007] Specifically, the steps of ultra-sparse clustering processing using the multi-level partitioning clustering method are as follows: the first level coarse-grained partitioning uses the minimum spanning tree algorithm to construct a global connection graph between data points and performs preliminary segmentation using edge weight thresholds; the second level medium-grained partitioning uses the maximum flow minimum cut algorithm to optimize the boundaries and divide the preliminary segmentation results into clusters; and the third level fine-grained partitioning uses the k-means clustering algorithm to calculate the center point of each cluster and perform local optimization.

[0008] Specifically, the Long Short-Term Memory (LSTM) network model comprises a four-layer architecture: an input layer, an LSM layer, an attention layer, and an output layer. The input layer receives trend and seasonal components and performs normalization processing. The LSM layer uses a bidirectional structure to capture the forward and backward dependencies of the time series. The attention layer dynamically allocates importance weights for different time steps and feature dimensions through a multi-head attention mechanism. The output layer uses a fully connected structure to output the fault prediction probability and degradation trend prediction value.

[0009] The number of attention heads in the multi-head attention mechanism is dynamically adjusted through an attention adjustment function based on the number of device types, the dimensions of monitoring parameters, and the data sampling frequency. The attention adjustment function calculates the adjustment value of the number of attention heads based on five data points: the number of device types, the dimensions of monitoring parameters, the data sampling frequency, the current failure rate, and the system load.

[0010] Specifically, the equipment fault feature vector library classifies and stores the feature vectors of normal operation status, early fault signs, fault precursors, and fault occurrence, and uses a kd-tree data structure to construct a fast retrieval index for the high-dimensional feature space.

[0011] The fault prediction and evaluation function calculates the fault risk assessment value based on five input parameters: feature vector similarity, degradation trend slope extracted from degradation trend prediction value, abnormal fluctuation intensity calculated from multi-dimensional real-time data, historical fault frequency statistically obtained from super sparse historical dataset, and hierarchical sparse data clustering index structure.

[0012] Specifically, when the fault risk assessment value is less than 0.25, a normal prediction level is output; when the fault risk assessment value is between 0.25 and 0.5 (excluding 0.5), a concern prediction level is output; when the fault risk assessment value is between 0.5 and 0.75 (excluding 0.75), an early warning prediction level is output; and when the fault risk assessment value is greater than 0.75, a high-risk prediction level is output.

[0013] Specifically, the spatiotemporal alignment ultra-sparse network model adopts a multi-layered hierarchical architecture to handle the spatiotemporal alignment problem of super sparse data and massive samples. It includes a spatiotemporal data preprocessing layer, a spatial relationship modeling layer, a temporal dependency capture layer, a sparse attention layer, and a feature fusion layer. The sparse layer is divided into device-level sparse layer, segment-level sparse layer, line-level sparse layer, and network-level sparse layer.

[0014] The super sparse historical dataset refers to a time series data set with extremely low data density generated by railway signaling equipment during long-term operation. Because the monitoring parameters of railway signaling equipment change slowly and remain stable for most of the time under normal operating conditions, the effective data points are extremely sparsely distributed across the entire time axis. The super sparse feature is that more than 90% of the data points in the dataset are zero values ​​or duplicate values.

[0015] Before constructing the Long Short-Term Memory (LSTM) network model, the training dataset for the LTM network model is also established. Historical operating data of different types of railway signaling equipment under three states—normal operation, performance degradation, and failure—are collected, classified and labeled according to equipment model, service life, and operating environment, and input-output sample pairs are constructed using a time window sliding method.

[0016] Specifically, the training steps of the long short-term memory network model include: using an adaptive learning rate optimization algorithm to update model parameters; using mean squared error as the loss function for regression tasks; using cross-entropy loss function for classification tasks; using an early stopping mechanism to prevent model overfitting; using gradient pruning technology to solve the gradient explosion problem; and using dropout method to improve the model's generalization ability.

[0017] Before utilizing the spatiotemporally aligned ultra-sparse network model, the process includes establishing a training dataset for the spatiotemporally aligned ultra-sparse network model, collecting operational status data of multiple railway signaling devices at different time periods, establishing a spatial coordinate system for the devices to record the physical distance and connection relationships between the devices, and performing synchronization alignment processing on the time series data to ensure the temporal consistency of data from different devices.

[0018] Specifically, the training steps of the spatiotemporally aligned ultra-sparse network model involve using a graph neural network training method to handle the relationship modeling between devices, using an attention weight learning algorithm to adaptively adjust the importance of different devices and time steps, using sparse regularization techniques to promote the model to learn sparse representations and reduce the risk of overfitting, and simultaneously optimizing the two objectives of fault prediction and correlation analysis through multi-task learning.

[0019] Specifically, the attention adjustment function uses a lightweight attention configuration to improve computational efficiency when the number of attention heads is adjusted to be between 1 and 4, a standard attention configuration to balance performance and efficiency when the number of attention heads is adjusted to be between 5 and 8, and an enhanced attention configuration to improve feature extraction accuracy when the number of attention heads is adjusted to be between 9 and 16, thereby adjusting the number of attention heads parameter of the multi-head attention mechanism.

[0020] Specifically, the railway signaling equipment fault prediction and diagnosis report includes single-equipment fault prediction level, multi-equipment associated fault risk analysis, and fault propagation path prediction results.

[0021] This invention addresses the technical problem of low accuracy in predicting railway signaling equipment faults under ultra-sparse historical data environments by constructing an ultra-sparse clustering processing system based on time series decomposition and multi-level partitioned clustering, combined with the multi-head attention mechanism of Long Short-Term Memory (LSTM) networks and a spatiotemporally aligned ultra-sparse network model. The invention employs time series decomposition technology to separate ultra-sparse data into trend components, seasonal components, and random noise components. It establishes a hierarchical sparse data clustering index structure through multi-level partitioned clustering, effectively overcoming the technical deficiency of traditional methods in extracting features from ultra-sparse data. By utilizing the time window sliding mechanism and multi-head attention mechanism of LSM networks, it dynamically captures the long-term degradation trend and key degradation features of equipment performance, overcoming the insufficient feature recognition capability of traditional algorithms in sparse data environments. Furthermore, through the multi-layered architecture and sparse hierarchical partitioning mechanism of the spatiotemporally aligned ultra-sparse network model, this invention achieves accurate identification of hidden fault signs in ultra-sparse data and accurate prediction of fault propagation paths between equipment. In summary, this invention solves the technical problem of low accuracy in predicting railway signaling equipment faults under ultra-sparse historical data environments mentioned in the background. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the spatiotemporally aligned ultrasparse network model structure involved in the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of a railway signaling equipment fault prediction and diagnosis method provided by the present invention. This method includes the following steps:

[0026] S01. Collect multi-dimensional real-time data during the operation of railway signaling equipment, including voltage signal strength, current fluctuation amplitude, temperature change trend, vibration spectrum characteristics, relay action delay and equipment response time, and store the multi-dimensional real-time data in time series to form a super sparse historical dataset.

[0027] S02. Perform time series decomposition processing on the super sparse historical dataset to separate the trend component, seasonal component and random noise component. Use the moving average algorithm to extract the trend component from the multi-dimensional real-time data. Use the periodic detection algorithm to identify the seasonal component from the multi-dimensional real-time data. Use residual analysis to obtain the random noise component from the multi-dimensional real-time data.

[0028] S03. The super sparse historical dataset is processed by multi-level partitioning clustering. The first level coarse-grained partitioning uses the minimum spanning tree algorithm to construct a global connection graph between data points and performs preliminary segmentation by using edge weight thresholds. The second level medium-grained partitioning uses the maximum flow minimum cut algorithm to optimize the boundaries and divide the clusters of the preliminary segmentation results. The third level fine-grained partitioning uses the k-means clustering algorithm to calculate the center point of each cluster and perform local optimization, thus establishing a hierarchical sparse data clustering index structure.

[0029] S04. Construct a long short-term memory network model to process the trend component and the seasonal component. Capture the long-term degradation trend of equipment performance through a time window sliding mechanism. Use a multi-head attention mechanism to identify key degradation features. The number of attention heads in the multi-head attention mechanism is dynamically adjusted according to the number of equipment types, the dimension of monitoring parameters and the data sampling frequency through an attention adjustment function. Output the fault prediction probability and degradation trend prediction value.

[0030] S05. Establish a device fault feature vector library, classify and store the normal operation status feature vector, early fault sign feature vector, fault precursor feature vector and fault occurrence feature vector, construct a fast retrieval index for high-dimensional feature space using kd-tree data structure, and match the current device status vector with the feature vector in the device fault feature vector library using the nearest neighbor search algorithm to calculate the feature vector similarity.

[0031] S06. Design a fault prediction and evaluation function to calculate the probability of equipment failure. The fault prediction and evaluation function calculates the fault risk assessment value based on five input parameters: the similarity of the feature vector, the slope of the degradation trend extracted from the degradation trend prediction value, the intensity of abnormal fluctuations calculated from the multi-dimensional real-time data, the historical fault frequency statistically obtained from the super sparse historical dataset, and the hierarchical sparse data clustering index structure. When the fault risk assessment value belongs to different numerical ranges, the corresponding fault prediction level is output.

[0032] S07. Using a spatiotemporally aligned ultra-sparse network model, perform correlation analysis on the fault prediction levels of multiple railway signaling devices, identify fault propagation paths and impact ranges between devices, and establish a device fault correlation map. The sparse hierarchy of the spatiotemporally aligned ultra-sparse network model includes device-level sparse layer, section-level sparse layer, line-level sparse layer, and network-level sparse layer, and output device fault correlation diagnosis results.

[0033] S08. Generate a railway signaling equipment fault prediction and diagnosis report based on the fault risk assessment value and the equipment fault correlation diagnosis results, including single equipment fault prediction level, multi-equipment associated fault risk analysis and fault propagation path prediction results.

[0034] The super sparse historical dataset refers to a time series data set with extremely low data density generated by railway signaling equipment during long-term operation. Because the monitoring parameters of railway signaling equipment change slowly and remain stable for most of the time under normal operating conditions, the effective data points are extremely sparsely distributed across the entire time axis. At the same time, the data acquisition frequency of different types of equipment varies greatly, which makes the dataset exhibit super sparse characteristics. The super sparse characteristics are manifested in that more than 90% of the data points in the dataset are zero values ​​or duplicate values.

[0035] The specific structure of the Long Short-Term Memory Network model is a four-layer architecture consisting of an input layer, a Long Short-Term Memory layer, an attention layer, and an output layer. The input layer receives the trend component and the seasonal component and performs normalization processing. The Long Short-Term Memory layer uses a bidirectional structure to capture the forward and backward dependencies of the time series. The attention layer dynamically allocates importance weights for different time steps and feature dimensions through the multi-head attention mechanism. The output layer uses a fully connected structure to output the fault prediction probability and the degradation trend prediction value.

[0036] The steps for establishing the training dataset of the Long Short-Term Memory Network Model specifically include collecting historical operating data of different types of railway signaling equipment under three states: normal operation, performance degradation, and failure. The data is classified and labeled according to equipment model, service life, and operating environment. Outliers and missing values ​​are removed from the data. Input-output sample pairs are constructed using a time window sliding method. Normal state data is marked as 0, degraded state data is marked as 1, and failure state data is marked as 2. The training set, validation set, and test set are divided in a ratio of 7:2:1 to ensure the balanced distribution of samples in different states.

[0037] The training steps of the Long Short-Term Memory (LSTM) network model specifically include updating model parameters using an adaptive learning rate optimization algorithm, setting the initial learning rate to 0.001, the batch size to 64, the number of training epochs to 200, using mean squared error as the loss function for regression tasks, using cross-entropy loss function for classification tasks, preventing overfitting through an early stopping mechanism, stopping training when the validation set loss does not decrease for 10 consecutive epochs, using gradient pruning to solve the gradient explosion problem, using dropout to improve the model's generalization ability, and periodically saving model checkpoints during training for model recovery and performance comparison analysis.

[0038] The specific structure of the spatiotemporally aligned ultra-sparse network model is as follows: it adopts a multi-layered architecture to handle the spatiotemporal alignment problem of super-sparse data and massive samples. The first layer is the spatiotemporal data preprocessing layer, which is responsible for data sparsification and time alignment. The second layer is the spatial relationship modeling layer, which constructs the topological relationship graph between devices. The third layer is the temporal dependency capture layer, which extracts long-term dependencies in the time series. The fourth layer is the sparse attention layer, which uses a local attention mechanism to reduce the amount of computation. The fifth layer is the feature fusion layer, which integrates spatiotemporal feature information. The output layer generates the device correlation prediction results. In the sparse layer division, the device-level sparse layer handles the sparse data relationship within a single device, the segment-level sparse layer handles the sparse association between multiple devices in the same segment, the line-level sparse layer handles the sparse spatiotemporal relationship of devices on the entire line, and the network-level sparse layer handles the ultra-sparse association pattern of devices across the entire network.

[0039] The steps for establishing the training dataset of the spatiotemporally aligned ultra-sparse network model specifically include: collecting operational status data of multiple railway signaling devices at different time periods; establishing a spatial coordinate system for the devices to record the physical distance and connection relationships between the devices; performing synchronization alignment processing on the time series data to ensure the temporal consistency of data from different devices; identifying and classifying sparse and dense regions in the data; constructing equipment fault propagation path samples, including information on the fault source device, propagation path, and scope of influence; performing augmentation processing on the sample data to generate more training samples to improve the model's generalization ability; and dividing the training data according to the time sequence to ensure that the model learns the correct temporal relationships.

[0040] The specific steps of training the spatiotemporally aligned ultra-sparse network model include: using graph neural network training methods to handle relationship modeling between devices; using attention weight learning algorithms to adaptively adjust the importance of different devices and time steps; using sparse regularization techniques to promote the model's learning of sparse representations and reduce the risk of overfitting; simultaneously optimizing both fault prediction and correlation analysis through multi-task learning; using a curriculum learning strategy to gradually transition from simple samples to difficult samples to improve training results; using a distributed training architecture to handle large-scale datasets to improve training efficiency; dynamically adjusting the learning rate and regularization parameters during training to adapt to the needs of different training stages; and evaluating model performance through cross-validation to select the optimal hyperparameter combination.

[0041] The fault prediction and evaluation function is used to calculate the fault risk probability of equipment based on multiple key indicators. The input includes five parameters: feature vector similarity, degradation trend slope, abnormal fluctuation intensity, historical fault frequency, and hierarchical sparse data clustering index structure. The output is the fault risk evaluation value between 0 and 1. When the fault risk evaluation value is less than 0.25, a normal prediction level is output. When the fault risk evaluation value is between 0.25 and 0.5 (excluding 0.5), a concern prediction level is output. When the fault risk evaluation value is between 0.5 and 0.75 (excluding 0.75), a warning prediction level is output. When the fault risk evaluation value is greater than 0.75, a high-risk prediction level is output.

[0042] The attention adjustment function is used to adjust the parameters of the multi-head attention mechanism in the long short-term memory network model. The attention adjustment function calculates the attention head number adjustment value based on five data points: the number of device types counted from the super sparse historical dataset, the monitoring parameter dimension calculated from the multi-dimensional real-time data, the data sampling frequency obtained during the multi-dimensional real-time data acquisition process, the current failure rate calculated from the failure prediction probability, and the system load monitored from the system operating status. When the attention head number adjustment value is in the range of 1 to 4, a lightweight attention configuration is used to improve computational efficiency; when the attention head number adjustment value is in the range of 5 to 8, a standard attention configuration is used to balance performance and efficiency; and when the attention head number adjustment value is in the range of 9 to 16, an enhanced attention configuration is used to improve feature extraction accuracy, thereby adjusting the attention head number parameter of the multi-head attention mechanism.

[0043] The multi-dimensional real-time data refers to the digital representation of various physical quantities and electrical parameters generated during the operation of railway signaling equipment. The time series decomposition is a mathematical processing method that separates time series data containing trend, periodic, and random components into independent components. The ultra-sparse clustering is an algorithm for clustering analysis of ultra-sparse datasets with extremely low data density and highly uneven distribution. The minimum spanning tree algorithm is a graph theory algorithm that finds a tree structure in a weighted connected graph that connects all vertices and has the minimum total weight. The maximum flow minimum cut algorithm is an algorithm in graph theory that finds the minimum weight cut set that divides a graph into two or more subgraphs, used to identify the optimal split boundary in data clustering. The Long Short-Term Memory network is a recurrent neural network architecture that uses a gating mechanism to control information flow and solve the long-term dependency problem of traditional recurrent neural networks. The multi-head attention mechanism is an extension of the attention mechanism, improving the model's expressive power by using multiple attention heads to compute attention weights in different subspaces in parallel. The feature vector is a data structure that converts multi-dimensional parameters of equipment operating status into mathematical vector representations. The kd-tree is a data structure used to organize points in k-dimensional space, supporting efficient nearest neighbor search operations. The nearest neighbor search algorithm is a retrieval algorithm that quickly finds the data point closest to the query point in a high-dimensional feature space. The spatiotemporal alignment is a process of aligning data from different times and spatial locations under a unified coordinate system. The ultra-sparse network is a neural network architecture for handling super-sparse cases where data contains a large number of zero or missing values. The sparse hierarchy partitioning is an architecture design that hierarchically processes sparse data according to different levels of abstraction and scope of influence. The fault propagation path is the transmission path and impact range of a fault spreading from the source device to other related devices. The hierarchical sparse data clustering index structure is a hierarchical data organization structure established through multi-level partitioning clustering, used for rapid retrieval and matching of sparse data patterns. The fault prediction level is a classification standard for equipment fault risk based on the fault risk assessment value. The equipment fault correlation diagnosis result refers to the diagnostic information on the fault correlation relationships and propagation patterns among multiple devices obtained through analysis of the spatiotemporal aligned ultra-sparse network model. The feature vector similarity refers to the measure of similarity between the current device state vector and the standard feature vector. The degradation trend slope refers to the rate of degradation of device performance over time. The abnormal fluctuation intensity refers to the amplitude of fluctuations in equipment operating parameters that deviate from the normal range. The historical fault frequency refers to the statistical value of the frequency of faults that occurred during the equipment's historical operation. The attention head number adjustment value refers to the calculated result used to dynamically adjust the number of attention heads in the multi-head attention mechanism. The number of equipment types refers to the total number of different types of equipment included in the super-sparse historical dataset. The monitoring parameter dimension refers to the number of monitoring parameters included in the multi-dimensional real-time data. The data sampling frequency refers to the reciprocal of the time interval for collecting the multi-dimensional real-time data.

[0044] The specific implementation methods of the above steps are described in detail below.

[0045] The specific implementation of step S01 involves acquiring multi-dimensional data from the railway signaling equipment in real time using a multi-sensor data acquisition system. First, voltage sensors, current sensors, temperature sensors, vibration sensors, and time measurement devices are deployed at key locations on the signaling equipment. A high-precision analog-to-digital converter is used to convert analog signals into digital signals. The voltage signal acquisition accuracy is set to 0.01V, the current acquisition accuracy to 0.001A, the temperature acquisition accuracy to 0.1℃, and the vibration spectrum acquisition frequency to 1000Hz. The data acquisition controller performs periodic sampling at preset time intervals; the sampling interval is 30 seconds under normal operating conditions and automatically adjusts to 5 seconds under abnormal conditions. The raw data is filtered, denoised, and standardized using a data preprocessing module to remove obvious outliers and noise interference. The processed multi-dimensional real-time data is then structured and stored according to timestamp order. A distributed database system is established to ensure data integrity and reliability, forming a super-sparse historical dataset containing time-series information.

[0046] The specific implementation of step S02 involves using classical time series decomposition techniques to separate the components of the super-sparse historical dataset. When extracting the trend component using the moving average algorithm, a sliding window length of 12 time points is set. Short-term fluctuations are smoothed by calculating the weighted average of the data within the window. The weight coefficients are assigned in a decreasing manner, with the weight of the latest data set at 0.5 and the weight of historical data decreasing in steps of 0.1. The periodicity detection algorithm uses the autocorrelation function analysis method to calculate the correlation coefficient for different lag periods. When the correlation coefficient exceeds the threshold of 0.6, it is determined to be a periodic component. Common periods include daily, weekly, and monthly periods. Random noise components are obtained by subtracting the trend and seasonal components from the original data. The reasonableness of the decomposition results is verified using residual analysis to ensure the independence and integrity of each component.

[0047] The specific implementation of step S03 involves designing a three-level hierarchical clustering architecture to process the super-sparse historical dataset. The first level, coarse-grained partitioning, uses the minimum spanning tree algorithm to construct a global connectivity graph. The Euclidean distance between data points is calculated as the edge weight, and the Kruskal algorithm is used to find the minimum weight spanning tree connecting all data points. An edge weight threshold of 1.5 times the average weight is set for initial segmentation. The second level, medium-grained partitioning, introduces the maximum flow minimum cut algorithm to optimize the cluster boundaries. Data points are treated as nodes in the graph, and similarity is used as the edge capacity. The Ford-Fulksen algorithm is used to find the maximum flow path, and the minimum cut set is identified as the cluster boundary, ensuring maximum intra-cluster similarity and maximum inter-cluster difference. The third level, fine-grained partitioning, uses the k-means clustering algorithm for local optimization within each coarse cluster. The k value is set to the square root of the number of data points within the cluster. The cluster center position is calculated iteratively, and the iteration terminates when the cluster center moves less than 0.01, establishing a multi-level sparse data clustering index structure.

[0048] The specific implementation of step S04 involves constructing a Long Short-Term Memory (LSTM) network model for processing time-series data. The network architecture includes an input layer of 128 neurons, a bidirectional LSM layer of 256 neurons, a multi-head attention layer, and an output layer of 64 neurons. The input layer receives trend and seasonal components, scaling the data to a range of 0 to 1 using min-max normalization. The bidirectional LSM layer captures long-term dependencies in the time series through forward and backward information transmission, with a forgetting gate threshold set to 0.5, and both the input and output gate thresholds set to 0.3. The number of attention heads in the multi-head attention mechanism is dynamically determined by an attention adjustment function. The adjustment value is calculated based on the number of device types, the dimensions of monitoring parameters, and the data sampling frequency. Four attention heads are used when the adjustment value is between 1 and 4, eight attention heads are used when it is between 5 and 8, and sixteen attention heads are used when it is between 9 and 16. The time window sliding mechanism sets the window length to 24 time points and the sliding step size to 1. An attention weight allocation mechanism identifies the most important time and parameter features for fault prediction.

[0049] The specific implementation of step S05 involves establishing a hierarchical equipment fault feature vector library storage system. The feature vector library is divided into four sub-libraries according to fault states: the normal operation feature vector library stores feature patterns of equipment parameters within the normal range; the early fault symptom feature vector library records patterns of parameters slightly deviating from normal values; the fault precursor feature vector library saves typical feature changes before a fault occurs; and the fault occurrence feature vector library contains feature states when the equipment completely fails. A kd-tree data structure is used to construct a fast retrieval index for the high-dimensional feature space. The tree depth is set to the logarithm of the feature vector dimension, and the maximum capacity of the leaf nodes is set to 10 feature vectors. The nearest neighbor search algorithm uses Euclidean distance as a metric, and a backtracking search strategy is used to quickly locate the most similar feature vector in the kd-tree. The search radius is set to twice the average feature vector distance. Feature vector similarity is calculated using the cosine similarity formula. A similarity value above 0.9 is considered highly similar, between 0.7 and 0.9 is moderately similar, between 0.5 and 0.7 is lowly similar, and below 0.5 is dissimilar.

[0050] The specific implementation of step S06 involves designing an evaluation function system for comprehensively assessing fault risk. The fault prediction evaluation function receives five input parameters, including feature vector similarity, degradation trend slope, abnormal fluctuation intensity, historical fault frequency, and hierarchical sparse data clustering index structure information. The degradation trend slope is calculated using linear regression, reflecting the rate of change in equipment performance over time. The abnormal fluctuation intensity is calculated using the standard deviation method, measuring the degree to which the parameter deviates from its normal value. The historical fault frequency is obtained through statistical analysis, calculating the probability of fault occurrence under the same operating conditions. The evaluation function uses a weighted summation method to calculate the fault risk assessment value. The weights of each parameter are determined based on expert knowledge and historical data statistics: feature vector similarity has a weight of 0.3, degradation trend slope has a weight of 0.25, abnormal fluctuation intensity has a weight of 0.2, historical fault frequency has a weight of 0.15, and clustering index structure has a weight of 0.1. The corresponding prediction level is output based on the numerical range of the fault risk assessment value: less than 0.25 is the normal level, 0.25 to 0.5 is the attention level, 0.5 to 0.75 is the warning level, and greater than 0.75 is the high-risk level.

[0051] The specific implementation of step S07 involves using a spatiotemporally aligned ultra-sparse network model to analyze the fault correlation among multiple devices. The model employs a multi-layered hierarchical architecture to handle the spatiotemporal relationships between devices. The first layer, the spatiotemporal data preprocessing layer, is responsible for data sparsity processing and time alignment correction, ensuring the time synchronization of data from different devices. The second layer, the spatial relationship modeling layer, constructs a topological graph of the devices, establishing an adjacency matrix based on the physical location and electrical connections of the devices, with matrix elements representing the correlation strength between devices. The third layer, the temporal dependency capture layer, uses a graph convolutional neural network to extract long-term dependencies in the time series, with a convolutional kernel size of 3×3 and a stride of 1. The fourth layer, the sparse attention layer, uses a local attention mechanism to reduce computational complexity, limiting the attention range to within three adjacent time steps. The fifth layer, the feature fusion layer, integrates spatiotemporal feature information and uses residual connections to prevent gradient vanishing. In the sparse hierarchy, the device level handles the correlation of internal parameters within a single device, the segment level handles the correlation of devices within the same segment, the line level handles the correlation of devices along the entire line, and the network level handles the correlation of devices across the entire network. The output layer generates a device fault correlation graph, identifying fault propagation paths and their impact range.

[0052] The specific implementation of step S08 involves generating a comprehensive diagnostic report based on the fault risk assessment results and equipment correlation analysis results. The report system uses a template-based generation method and includes three main parts: single-device fault prediction level, multi-device correlation fault risk analysis, and fault propagation path prediction. The single-device fault prediction section displays the current status, fault risk level, key parameter change trends, and expected failure time for each device. The multi-device correlation analysis section displays the correlation strength matrix between devices, the probability of correlated faults, and the risk assessment of chain reactions. The fault propagation path prediction section draws a fault propagation path diagram, marks the propagation time and impact range, and provides preventative measure suggestions. The report generation system supports multiple output formats, including detailed technical reports, execution summaries, and chart displays, automatically selecting the appropriate report mode based on user permissions and needs.

[0053] like Figure 2As shown, the detailed structure of the spatiotemporally aligned ultra-sparse network model adopts a six-layer deep architecture. The first layer, the spatiotemporal data preprocessing layer, includes a data sparsification module and a time alignment module. The sparsification module removes redundant information through adaptive threshold filtering, while the time alignment module uses an interpolation algorithm to unify the sampling time points of different devices. The second layer, the spatial relationship modeling layer, constructs a multi-scale graph structure, including device-level, segment-level, line-level, and network-level connection graphs, and uses a graph attention network to learn spatial dependencies at different scales. The third layer, the temporal dependency capture layer, adopts a hybrid architecture combining a temporal convolutional network and a recurrent neural network. The temporal convolutional network is responsible for capturing short-term temporal patterns, while the recurrent neural network is responsible for modeling long-term temporal dependencies. The fourth layer, the sparse attention layer, implements a multi-head sparse attention mechanism, reducing computation through sparsity constraints, with an attention sparsity rate set to 85%. The fifth layer, the feature fusion layer, uses a gating fusion mechanism to integrate spatiotemporal features, adaptively adjusting the importance of different features by learning gating weights. The sixth layer, the output layer, includes two branches: classification output and regression output. The classification output predicts the fault type, and the regression output predicts the fault occurrence time.

[0054] The detailed steps for establishing the training dataset for the spatiotemporally aligned ultra-sparse network model include five stages: data collection, preprocessing, alignment, annotation, and augmentation. The data collection stage gathers historical operational data from signaling equipment across multiple railway lines, covering equipment status under different seasons, weather conditions, and load conditions. The preprocessing stage cleans the raw data, removing outliers caused by sensor faults and filling in missing data using a combination of forward and backward padding strategies. The temporal alignment stage establishes a unified time coordinate system, synchronizing time-series data from different equipment to the same time base, with time accuracy controlled within 1 second. The spatial alignment stage establishes a spatial coordinate system based on the geographical location information of the equipment, recording the distances and connections between equipment. The annotation stage invites domain experts to manually annotate the data, including fault type, fault time, fault impact range, and fault propagation path. The data augmentation stage uses techniques such as time window sliding, noise injection, and data interpolation to generate more training samples, increasing the number of augmented samples to five times that of the original data.

[0055] The spatiotemporally aligned ultra-sparse network model can simultaneously handle ultra-sparse data features in both time and space dimensions. Traditional time series prediction models, such as recurrent neural networks, mainly focus on the temporal dependencies of single devices and cannot effectively model the spatial correlations between multiple devices. Existing graph neural network models, such as graph convolutional networks, mainly handle static graph structures and are difficult to adapt to the dynamic spatiotemporal relationships of railway signaling equipment. The spatiotemporally aligned ultra-sparse network model integrates the execution processes of traditional graph computation and neural networks, forming a hybrid execution mode where irregular and regular computation and memory access behaviors coexist, enabling explicit modeling of relationships between variables and relationships between time. Compared with existing technologies, the advantage of this model lies in its sparsity handling capability, effectively handling the ultra-sparse features of railway signaling equipment data, where over 90% are zero or duplicate values. The model employs a sparse attention mechanism to reduce computational complexity and uses a hierarchical architecture to handle sparse correlations at different levels of abstraction. Multi-level modeling from the device level to the network level ensures the accuracy of fault propagation path identification. The spatiotemporal alignment mechanism solves the problem of huge differences in the sampling frequency of data from different devices, ensuring the effectiveness of multi-device collaborative analysis.

[0056] The key technical ideas of this invention include three aspects. The first key technical idea is a multi-level partitioning clustering processing technique for super-sparse historical datasets. Traditional clustering algorithms, such as k-means clustering, are prone to getting trapped in local optima when processing sparse data and cannot effectively handle cases with extremely low data density. This invention adopts a three-level hierarchical clustering approach combining the minimum spanning tree algorithm, the maximum flow minimum cut algorithm, and the k-means clustering algorithm, which can handle the clustering problem of ultra-sparse data while ensuring global optimality. The minimum spanning tree algorithm constructs the connection relationship of data points from a global perspective, avoiding the initialization sensitivity problem of traditional clustering algorithms. The maximum flow minimum cut algorithm optimizes the clustering boundary through graph theory methods, ensuring that the separation degree between classes is maximized. The three-level hierarchical architecture refines the clustering results step by step from coarse-grained to fine-grained, effectively solving the technical problem of uneven distribution of sparse data.

[0057] The second key technological approach is a fault prediction technique combining a Long Short-Term Memory (LSTM) network with a dynamic multi-head attention mechanism. Traditional fault prediction methods, primarily based on statistical analysis or simple machine learning algorithms, struggle to capture the long-term trends and complex nonlinear relationships of equipment performance degradation. This invention employs a bidirectional LSM network architecture, using forward and backward information transmission mechanisms to capture long-term dependencies in time series data, thus solving the gradient vanishing problem of traditional recurrent neural networks. The dynamic multi-head attention mechanism adaptively adjusts the number of attention heads based on equipment type, monitoring parameter dimensions, and data sampling frequency, avoiding the waste of computational resources or insufficient feature extraction caused by fixed attention configurations. The attention adjustment function achieves an adaptive balance between model complexity and computational efficiency, improving the accuracy and real-time performance of fault prediction.

[0058] The third key technological approach is the multi-device correlation analysis technology using a spatiotemporally aligned ultra-sparse network model. Existing equipment fault diagnosis technologies mainly focus on the status monitoring of individual devices, neglecting the mutual influence between devices and fault propagation mechanisms. The spatiotemporally aligned ultra-sparse network model of this invention handles device correlation relationships at different abstract levels through a multi-layered hierarchical architecture. This hierarchical modeling from the device level to the network level accurately identifies fault propagation paths and the scope of impact. The sparse attention mechanism reduces computational complexity, enabling real-time analysis of large-scale networks. Spatiotemporal data alignment technology solves the data synchronization problem between different devices, ensuring the accuracy of multi-device collaborative analysis. The hybrid architecture combining graph convolutional neural networks and temporal convolutional networks can simultaneously capture spatial and temporal dependencies, achieving true spatiotemporal fusion analysis.

[0059] The synergistic effect of these three key technological approaches has yielded significant technical results. Super-sparse data clustering technology provides a high-quality data foundation for subsequent feature extraction and pattern recognition, and the index structure established through multi-level partitioned clustering greatly improves the efficiency of feature matching. The fault prediction results of Long Short-Term Memory (LSTM) networks provide single-device-level fault probability input for the spatiotemporally aligned super-sparse network model, achieving an organic connection from single-device prediction to multi-device correlation analysis. Conversely, the correlation analysis results of the spatiotemporally aligned super-sparse network model provide richer contextual information for single-device fault prediction, improving prediction accuracy through constraints on inter-device correlation relationships. These three technological approaches form a complete technical chain encompassing data processing, single-device prediction, and multi-device correlation, mutually promoting and synergistically optimizing each other, achieving a comprehensive technological breakthrough in railway signaling equipment fault prediction and diagnosis.

[0060] It should be noted that this invention also solves the following technical problem: the difficulty in identifying the propagation path of multi-device associated faults. In railway signaling systems, there are complex electrical connections and logical control relationships between multiple devices. When a device fails, the impact of the fault often propagates to other devices through these relationships, forming a chain reaction of faults. Traditional fault diagnosis methods mainly analyze individual devices independently, lacking in-depth modeling of the correlation between devices and accurate prediction of fault propagation paths, resulting in the inability to identify and block the fault propagation chain in a timely manner. This invention effectively solves this technical problem through a multi-layered architecture of a spatiotemporally aligned ultra-sparse network model. This model uses a device-level sparse layer to handle the sparse data relationships within a single device, a section-level sparse layer to handle the sparse correlations between multiple devices within the same section, a line-level sparse layer to handle the sparse spatiotemporal relationships of devices on the entire line, and a network-level sparse layer to handle the ultra-sparse correlation patterns of devices across the entire network. Through the local attention mechanism of the sparse attention layer and the spatiotemporal feature integration of the feature fusion layer, it can accurately identify the fault propagation path and impact range between devices and establish a complete device fault correlation map. In addition, this invention also solves the technical problem of low feature extraction efficiency in ultra-sparse data environments. The data generated by railway signaling equipment during long-term operation is characterized by extremely low data density. Traditional feature extraction algorithms suffer from high computational complexity and poor extraction results when processing such ultra-sparse data, failing to efficiently identify effective feature patterns. This invention effectively solves this problem through a multi-level partitioned clustering method for ultra-sparse clustering. The first-level coarse-grained partitioning uses the minimum spanning tree algorithm to construct a global connectivity graph between data points and performs preliminary segmentation using edge weight thresholds, significantly reducing the amount of data to be processed. The second-level medium-grained partitioning uses the maximum flow minimum cut algorithm to optimize the boundaries and divide the clusters of the preliminary segmentation results, ensuring the accuracy of the clustering results. The third-level fine-grained partitioning uses the k-means clustering algorithm to calculate the centroids of data points within each cluster and perform local optimization, establishing an efficient hierarchical sparse data clustering index structure. Combined with a high-dimensional feature space fast retrieval index constructed using a kd-tree data structure, efficient feature extraction and matching in ultra-sparse data environments are achieved.

[0061] Specifically, the principle of this invention is as follows: The technical solution of this invention can solve the technical problem of low accuracy in predicting railway signal equipment faults under ultra-sparse historical data environments. Its core principle lies in constructing a multi-level intelligent processing architecture targeting the characteristics of ultra-sparse data. First, time series decomposition processing separates and purifies the effective information in the ultra-sparse historical dataset through moving average algorithms, periodic detection algorithms, and residual analysis, eliminating the interference of data sparsity on feature extraction and providing high-quality trend and seasonal components for subsequent processing. Second, the ultra-sparse clustering processing of multi-level partitioned clustering adopts a three-level architecture: minimum spanning tree algorithm to construct a global connectivity graph, maximum flow minimum cut algorithm for boundary optimization, and k-means clustering algorithm for local optimization. This architecture can establish an effective hierarchical sparse data clustering index structure under ultra-sparse data environments, solving the problem of poor clustering performance of traditional clustering algorithms under uneven sparse data distribution. The Long Short-Term Memory (LSTM) network model captures forward and backward dependencies in time series data through a bidirectional structure. Combined with the dynamic weight allocation capability of a multi-head attention mechanism, it can extract long-term degradation patterns and key fault symptoms of equipment performance from sparse trend and seasonal components. The attention adjustment function dynamically adjusts the number of attention heads based on the number of equipment types, the dimension of monitoring parameters, and the data sampling frequency, ensuring the model's adaptive processing capability under different data sparsity environments. The spatiotemporally aligned ultra-sparse network model adopts a four-layer sparse hierarchical architecture: device-level, segment-level, line-level, and network-level. Through the local attention mechanism of the sparse attention layer and the spatiotemporal feature integration of the feature fusion layer, it can accurately identify fault propagation paths and correlation patterns between devices in ultra-sparse data environments. The fault prediction and evaluation function is based on the comprehensive calculation of five parameters: feature vector similarity, degradation trend slope, abnormal fluctuation intensity, historical fault frequency, and hierarchical sparse data clustering index structure, achieving accurate quantitative assessment of hidden fault risks in ultra-sparse data. The logical rationality of this technical solution is reflected in the organic coordination and cooperation among the components. Time series decomposition provides a pure data foundation for ultra-sparse clustering, ultra-sparse clustering provides a structured data index for deep learning models, long short-term memory networks and spatiotemporally aligned ultra-sparse network models realize in-depth mining of super-sparse data from the two dimensions of temporal features and spatial correlation, respectively, and the fault prediction evaluation function integrates the multi-dimensional analysis results into operable prediction results, forming a complete fault prediction technology link in the super-sparse data environment.

[0062] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0063] The specific implementation of step S01 is to acquire multi-dimensional real-time data of railway signaling equipment through a multi-sensor fusion acquisition system. Let the acquired multi-dimensional real-time data vector be... ,in express The voltage signal strength at a given time, expressed in volts (V). express The amplitude of current fluctuation at any given time, expressed in amperes (A). express Temperature value at any given time, in °C; express The eigenvector of the vibration spectrum at time t; express The relay action delay at any given moment, measured in milliseconds (ms). express The device response time at any given moment is measured in milliseconds (ms). Data acquisition is performed according to a time series, forming a super-sparse historical dataset. ,in The total number of sampling points. For the first Each sampling time point. Data sparsity is defined as... ,in The number of zero or duplicate values, when At that time, it was believed that the dataset exhibited super-sparse features.

[0064] The specific implementation of step S02 is to use an additive model to perform time series decomposition on the super-sparse historical dataset. For any time series in the dataset... The decomposition model is expressed as:

[0065] ;

[0066] In the formula, For trend components; Seasonal component; This represents random noise components. The trend component is extracted using a moving average algorithm.

[0067] ;

[0068] In the formula, The radius of the sliding window is 6; Let be the weighting coefficient, satisfying ,and , , Seasonal components are identified using autocorrelation functions:

[0069] ;

[0070] In the formula, The lag period; The length of the time series; This is the time series mean. When... When, determine the period Random noise components are obtained through residual calculation: .

[0071] The specific implementation of step S03 is to construct a multi-level partitioning clustering system to process the super-sparse historical dataset. In the first-level coarse-grained partitioning, the Euclidean distance between data points is calculated as follows:

[0072] ;

[0073] In the formula, Indicates the first The first data point 3D eigenvalues; Index the data points. Set the edge weight threshold to [value]. ,in The average distance, The number of data points. The objective function for maximum flow and minimum cut in the second-level granular partitioning is:

[0074] ;

[0075] In the formula, It is a cut set of the graph; For nodes and Similarity between them This is a bandwidth parameter, and its value range is... The k-means objective function for the third-level fine-grained partitioning is:

[0076] ;

[0077] In the formula, For the first The center point of each cluster; is the number of clusters, where This represents the total number of data points in the current sub-cluster. For the first A cluster; For clusters The number of data points in the data.

[0078] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0079] The specific implementation of step S05 involves constructing a device fault feature vector library and performing feature matching. Let the current device state vector be... The first feature vector in the feature vector library The standard feature vectors are The feature vector similarity is calculated using cosine similarity:

[0080] ;

[0081] In the formula, Represents the dot product of vectors; and Let and represent the Euclidean norms of the vectors, respectively. During the kd-tree search process, the equation of the node-splitting hyperplane is:

[0082] ;

[0083] In the formula, Index of the current segmentation dimension; The median is the median along this dimension. The search radius is set to... ,in The average distance between the feature vectors.

[0084] The specific implementation of step S06 involves designing a comprehensive fault prediction and evaluation function. The formula for calculating the fault risk assessment value is as follows:

[0085] ;

[0086] In the formula, This is the fault risk assessment value, and its range is [value range missing]. ; The similarity of the normalized feature vectors. ,in ; The slope of the degradation trend after normalization. ,in Obtained through linear regression: , For the number of time points, For the first At a certain point in time, For the first The observed values ​​of equipment performance parameters at each time point and These are the historical minimum and maximum slope values, respectively; The normalized abnormal fluctuation intensity ,in , Standard deviation The mean, For reference fluctuation intensity; The normalized historical fault frequency, ,in , This represents the number of historical failures. Runtime (in hours) For reference fault frequency (unit: 1 / hour); The contribution value to the normalized hierarchical sparse data clustering index structure, with a value range of [value range missing]. The weighting coefficients satisfy The specific value is , , , , .

[0087] The specific implementation methods for steps S07-S08 are the same as those described above, and will not be repeated in detail here.

[0088] It should be noted that the specific implementation of the attention adjustment function involves dynamically adjusting the parameters of the multi-head attention mechanism based on the system's operating status. The formula for calculating the adjustment value of the number of attention heads is as follows:

[0089] ;

[0090] In the formula, Adjust the number of attention points; The number of device types is obtained by statistically analyzing a super sparse historical dataset. To monitor parameter dimensions, this embodiment... ; This refers to the data sampling frequency, measured in Hz. The current failure rate is calculated using the following formula: ,in This represents the highest historical failure probability. The system load has a range of values. ; , , , , These are the weighting coefficients; This is a bias term.

[0091] It should be noted that the additive model in time series decomposition can break down complex time series data into interpretable components. The trend component reflects the long-term trend of equipment performance, the seasonal component captures periodic fluctuation patterns, and the random noise component represents unpredictable random disturbances. Compared with traditional holistic analysis methods, this decomposition method can more accurately identify equipment degradation patterns and early signs of failure. The weighting function of the moving average algorithm is as follows:

[0092] ;

[0093] This approach ensures that the most recent data is given higher weight while maintaining appropriate consideration for historical data. The distance metric and similarity calculation in multi-level partitioned clustering comprehensively consider the contributions of multi-dimensional features. Through a combination of minimum spanning tree, maximum flow minimum cut, and k-means, it achieves hierarchical clustering from coarse to fine, which can better handle the uneven distribution characteristics of super-sparse data compared to single clustering algorithms. The similarity function is as follows:

[0094] ;

[0095] Using a Gaussian kernel function, distance metrics can be converted into similarity metrics, and the bandwidth parameter can be used to achieve this. The decay rate of similarity is controlled. The fault prediction evaluation function integrates multiple key indicators through weighted fusion. Linear regression calculation of the degradation trend slope quantitatively describes the rate of equipment performance degradation, the coefficient of variation calculation of abnormal fluctuation intensity reflects parameter stability, and statistical analysis of historical fault frequencies provides an experience-based reliability assessment. This multi-indicator fusion approach significantly improves the accuracy and robustness of fault prediction compared to single-indicator judgment. Normalization ensures the comparability of indicators with different dimensions, where the coefficient of variation is:

[0096] ;

[0097] It is a dimensionless measure of relative dispersion, capable of eliminating the influence of differences in the magnitude of different parameters. The attention adjustment function, by comprehensively considering factors such as equipment type complexity, monitoring data dimensionality, sampling frequency, real-time failure rate, and system load, achieves adaptive adjustment of the attention mechanism. Compared to fixed parameter settings, it can better adapt to different operating conditions and equipment states, improving the model's generalization ability and computational efficiency. Its linear combination form is:

[0098] ;

[0099] Different weighting coefficients reflect the contribution of each factor to the attention complexity requirement, among which the current failure rate weight... The maximum value reflects the system's emphasis on real-time risk status.

[0100] To better understand and implement this invention, a specific application scenario is provided below as Example 2: A certain railway line section contains 156 signaling devices, covering various types of equipment such as track circuits, signals, switch machines, and relays. The technical team used the railway signaling equipment fault prediction and diagnosis method of this invention to construct a complete intelligent monitoring system.

[0101] First, the technical team deployed a multi-dimensional sensor network on each signaling device to collect real-time data during operation. Each device was equipped with six sensors to monitor voltage signal strength, current fluctuation amplitude, temperature change trends, vibration spectrum characteristics, relay action delay, and device response time. Data acquisition frequencies were differentiated according to device type: track circuits and signals used a 1Hz acquisition frequency, switch machines used a 2Hz frequency, and relays used a 0.5Hz frequency. After 60 days of continuous data acquisition, a comprehensive dataset was generated, containing over [number missing]. This is a super-sparse historical dataset with a total of 100,000 data points. Because the parameters of railway signaling equipment change slowly during normal operation, 92.3% of the data points in the dataset are zero or duplicate values, exhibiting typical super-sparse characteristics.

[0102] The technical team performed time-series decomposition on the super-sparse historical dataset. A 21-point moving average algorithm was used to extract trend components from the multi-dimensional real-time data. With a sliding window length of 21 time points, this effectively smoothed short-term fluctuations while preserving long-term trend information. A periodicity detection algorithm based on Fast Fourier Transform was used to identify seasonal components from the multi-dimensional real-time data, detecting three main periodic patterns in device operation: 24-hour, 7-day, and 30-day. Random noise components were obtained from the multi-dimensional real-time data through residual analysis, with the standard deviation of the noise components controlled within 5% of the original signal amplitude.

[0103] Multi-level partitioning and clustering processing was performed on the super-sparse historical dataset. The first level, coarse-grained partitioning, used the minimum spanning tree algorithm to construct a global connectivity graph among 156 device data points. Edge weights were calculated using Euclidean distance, and a threshold of 0.85 was used for initial segmentation, dividing the data into 23 coarse-grained clusters. The second level, medium-grained partitioning, used the maximum flow minimum cut algorithm to optimize the boundaries of the 23 initial segmentations, identifying 34 medium-grained cluster boundaries and improving the intra-cluster connectivity strength to 0.92. The third level, fine-grained partitioning, used the k-means clustering algorithm to process the data points within each cluster, with k set to 8. After 120 iterations and convergence, a hierarchical sparse data clustering index structure containing 67 fine-grained clusters was finally established.

[0104] The technical team constructed a Long Short-Term Memory (LSTM) network model to handle trend and seasonal components. The network structure comprises four layers: an input layer, an LSM layer, an attention layer, and an output layer. The input layer receives 24-dimensional trend component data and 18-dimensional seasonal component data, performing zero-mean normalization. The LSM layer employs a bidirectional structure with 256 hidden units, capable of simultaneously capturing forward and backward dependencies in the time series. The attention layer utilizes a multi-head attention mechanism, with the number of attention heads dynamically adjusted via an attention adjustment function. Based on five parameters—number of device types (4), 6 dimensions of monitoring parameters, a data sampling frequency of 1.2Hz, a current failure rate of 0.03, and a system load of 0.67—the adjusted number of attention heads was calculated to be 6, employing a standard attention configuration. The output layer uses a fully connected structure with 128 neurons, outputting the fault prediction probability and degradation trend prediction value.

[0105] A device fault feature vector library was established, classifying and storing feature vectors from different operating states. The normal operating state feature vector library contains 1,847 samples, the early fault symptom feature vector library contains 392 samples, the fault precursor feature vector library contains 156 samples, and the fault occurrence feature vector library contains 89 samples. A fast retrieval index for a 42-dimensional feature space was constructed using a kd-tree data structure with a depth of 18 levels, supporting... Nearest neighbor search with time complexity. The current device state vector is matched with feature vectors in a feature vector library using a nearest neighbor search algorithm, and the similarity of the feature vectors is calculated.

[0106] The technical team designed a fault prediction and assessment function to calculate the probability of equipment failure. This function is based on five input parameters: feature vector similarity (0.87), degradation trend slope (-0.032), abnormal fluctuation intensity (1.24), historical fault frequency (0.015), and hierarchical sparse data clustering index structure weight (0.73). By weighting and comprehensively calculating these parameters, a fault risk assessment value of 0.68 is obtained. Based on the range of the fault risk assessment value, the equipment is classified as belonging to the early warning prediction level, requiring enhanced monitoring and the development of a preventative maintenance plan.

[0107] A spatiotemporally aligned ultra-sparse network model was used to conduct correlation analysis on the fault prediction levels of 156 railway signaling devices. The model employs a six-layer deep architecture. The first layer, the spatiotemporal data preprocessing layer, uses an adaptive threshold of 0.95 to sparsify the data, ensuring temporal alignment accuracy within 1 second. The second layer, the spatial relationship modeling layer, constructs a graph structure at four scales, including 156 device-level connections, 23 section-level connections, 8 line-level connections, and 1 network-level connection. The third layer, the temporal dependency capture layer, uses a hybrid architecture combining temporal convolutional networks and recurrent neural networks, with a temporal convolutional kernel size of 7 and a recurrent neural network hidden layer dimension of 512. The fourth layer, the sparse attention layer, implements a 12-head sparse attention mechanism with an attention sparsity rate of 85%, significantly reducing computational cost. The fifth layer, the feature fusion layer, uses a gated fusion mechanism with learned gate weights of 0.67, 0.82, 0.75, and 0.91. The sixth layer, the output layer, includes classification and regression output branches, identifying three main equipment fault propagation paths and their impact range covering 34 related devices.

[0108] As shown in Table 1, the technical team compiled statistics on the distribution of fault characteristics for different types of equipment:

[0109] Table 1 Statistical Table of Equipment Fault Characteristics

[0110]

[0111] The training dataset for building the spatiotemporally aligned ultra-sparse network model involved multiple stages. The data collection stage gathered 18 months of historical operational data from three high-speed railway lines: Beijing-Guangzhou, Beijing-Shanghai, and Shanghai-Kunming, covering all four seasons and various weather conditions including sunshine, rain, snow, and fog. The preprocessing stage removed 4,267 outliers caused by sensor malfunctions and processed 1,892 missing data points using forward and backward padding strategies. The temporal alignment stage established a unified time coordinate system with time accuracy controlled within 1 second. The spatial alignment stage recorded the spatial relationships between devices, ranging from a minimum distance of 50m to a maximum distance of 15.8km. The annotation stage involved manual annotation by 12 signal engineering professionals, achieving a 94.7% consistency rate. The data augmentation stage employed time window sliding, Gaussian noise injection, and linear interpolation techniques to augment the training samples from the original 78,542 to 392,710.

[0112] The model training employed an adaptive learning rate optimization algorithm, with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. The Long Short-Term Memory (LSTM) network model used mean squared error as the regression loss function and cross-entropy as the classification loss function. An early stopping mechanism was triggered when the validation set loss stopped decreasing at the 156th epoch. The spatiotemporally aligned ultra-sparse network model used a graph neural network training method, employing an attention weight learning algorithm and sparse regularization techniques, with a regularization coefficient set to 0.01. Through multi-task learning, fault prediction and correlation analysis were simultaneously optimized, and the final model's overall performance on the test set met the expected requirements.

[0113] Based on the fault risk assessment values ​​and equipment fault correlation diagnosis results, the system generated a complete railway signaling equipment fault prediction and diagnosis report. The report showed that out of 156 devices, 12 were at the "concern" prediction level, 8 at the "warning" prediction level, and 1 at the "high-risk" prediction level. Multi-device correlation fault risk analysis identified a strong correlation between track circuit TG-4237 and signal XH-5891, with a correlation strength of 0.89. Fault propagation path prediction results showed that if TG-4237 fails, it will affect 3 adjacent relays and 2 switch machines within 15 minutes, covering a 2.3km section of track.

[0114] This invention represents a significant technological advancement compared to traditional methods. Traditional railway signaling equipment monitoring methods primarily rely on periodic manual inspections and independent monitoring of individual devices, failing to achieve inter-device correlation analysis and fault propagation path prediction. This invention solves the technical challenge of extremely low data density in railway signaling equipment by constructing and processing a super-sparse historical dataset, enabling the extraction of effective information from a large amount of redundant and repetitive data. The multi-level partitioning clustering approach better handles the highly uneven distribution of ultra-sparse data compared to traditional clustering algorithms. A more accurate data index structure is established through a three-level progressive clustering strategy using minimum spanning tree, maximum flow minimum cut, and k-means. The long short-term memory network model combined with a multi-head attention mechanism has a stronger ability to model long-term dependencies compared to traditional time series prediction methods, simultaneously capturing the gradual degradation trend of equipment performance and sudden anomaly patterns. The innovation of the spatiotemporally aligned ultra-sparse network model lies in simultaneously modeling ultra-sparse correlations in both the temporal and spatial dimensions. Through a hierarchical sparse architecture, multi-level modeling from the device level to the network level achieves multi-device collaborative fault prediction and propagation path identification functions that are unattainable by traditional methods. Fault prediction and evaluation functions, through comprehensive calculation of multiple parameters, have higher accuracy and reliability compared to judgment based on a single indicator, and can effectively reduce false alarm rate and false negative rate.

[0115] It should be noted that the variables involved in this invention are explained in detail in Table 2.

[0116] Table 2 Variable Explanation Table

[0117]

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of railway signal equipment failure prediction and diagnosis, characterized by, The multi-dimensional real-time data in the operation process of railway signal equipment is collected and stored in time sequence to form a super-sparse historical data set; The super-sparse historical data set is subjected to time sequence decomposition processing to separate trend components, seasonal components and random noise components; The super-sparse historical data set is subjected to super-sparse clustering processing by using a multi-level partition clustering method to establish a hierarchical sparse data clustering index structure; A long short-term memory network model is constructed to process the trend components and seasonal components, a time window sliding mechanism is used to capture the long-term degradation trend of the equipment performance, a multi-head attention mechanism is used to identify key degradation features, and a fault prediction probability and a degradation trend prediction value are outputted; a device fault feature vector library is established, a k-d tree data structure is used to construct a fast retrieval index of a high-dimensional feature space, and a nearest neighbor search algorithm is used to match the current device state vector with the feature vectors in the device fault feature vector library to calculate the feature vector similarity; a fault prediction evaluation function is designed to calculate the device fault probability, and different fault prediction levels are outputted when the fault risk evaluation value belongs to different numerical ranges; The fault prediction levels of multiple railway signal equipment are subjected to correlation analysis by using a spatio-temporal alignment super-sparse network model, the fault propagation path and influence range between devices are identified, a device fault correlation graph is established, and a device fault correlation diagnosis result is outputted; a railway signal equipment fault prediction and diagnosis report is generated according to the fault risk evaluation value and the device fault correlation diagnosis result.

2. The railroad signal equipment failure prediction and diagnosis method of claim 1, wherein, The multi-dimensional real-time data includes voltage signal strength, current fluctuation amplitude, temperature change trend, vibration spectrum characteristics, relay action time delay and device response time.

3. The railroad signal equipment failure prediction and diagnosis method of claim 2, wherein, The time sequence decomposition processing step specifically includes extracting trend components from the multi-dimensional real-time data by using a sliding average algorithm, identifying seasonal components from the multi-dimensional real-time data by using a periodicity detection algorithm, and obtaining random noise components from the multi-dimensional real-time data by residual analysis.

4. The railroad signal equipment failure prediction and diagnosis method according to claim 3, characterized by, The super-sparse clustering processing step of the multi-level partition clustering method specifically includes constructing a global connection graph between data points by using a minimum spanning tree algorithm and performing preliminary segmentation by using an edge weight threshold in the first level of coarse-grained partition, performing boundary optimization and cluster division on the preliminary segmentation result by using a maximum flow minimum cut algorithm in the second level of medium-grained partition, and calculating the center point of the data points in each cluster and performing local optimization by using a k-means clustering algorithm in the third level of fine-grained partition.

5. The railroad signal equipment failure prediction and diagnosis method of claim 4, wherein, The long short-term memory network model specifically includes a four-layer architecture including an input layer, a long short-term memory layer, an attention layer and an output layer, the input layer receives and normalizes the trend components and seasonal components, the long short-term memory layer adopts a bidirectional structure to capture the forward and backward dependency relationship of the time sequence, the attention layer dynamically allocates importance weights of different time steps and feature dimensions by using a multi-head attention mechanism, and the output layer adopts a full connection structure to output the fault prediction probability and the degradation trend prediction value.

6. The railroad signal equipment failure prediction and diagnosis method according to claim 5, characterized by, The number of attention heads of the multi-head attention mechanism is dynamically adjusted according to the number of device types, the dimension of monitoring parameters and the data sampling frequency through an attention adjustment function, and the attention adjustment function calculates an attention head number adjustment value based on the number of device types, the dimension of monitoring parameters, the data sampling frequency, the current failure rate and the system load.

7. The railroad signal equipment failure prediction and diagnosis method according to claim 6, characterized by, The device failure feature vector library, in particular, classifies and stores normal operation state feature vectors, early failure symptom feature vectors, failure precursor feature vectors and failure occurrence feature vectors, and adopts a k-d tree data structure to construct a fast retrieval index of a high-dimensional feature space.

8. The railroad signal equipment failure prediction and diagnosis method of claim 7, wherein, The failure prediction evaluation function calculates a failure risk evaluation value based on five input parameters, including feature vector similarity, a degradation trend slope extracted from a degradation trend prediction value, an abnormal fluctuation intensity calculated from multi-dimensional real-time data, a historical failure frequency statistically calculated from a super-sparse historical data set and a hierarchical sparse data clustering index structure.

9. The railroad signal equipment failure prediction and diagnosis method of claim 8, wherein, When the failure risk evaluation value is <0.25, a normal prediction level is output, when 0.25≤the failure risk evaluation value<0.5, an attention prediction level is output, when 0.5≤the failure risk evaluation value<0.75, a warning prediction level is output, and when the failure risk evaluation value≥0.75, a high-risk prediction level is output.

10. The railroad signal equipment failure prediction and diagnosis method of claim 9, wherein, The spatio-temporal alignment super-sparse network model, in particular, adopts a multi-layer hierarchical architecture to process the spatio-temporal alignment problem of super-sparse data and massive samples, including a spatio-temporal data preprocessing layer, a spatial relationship modeling layer, a time series dependence capturing layer, a sparse attention layer and a feature fusion layer, and the sparse layer division includes a device-level sparse layer, a section-level sparse layer, a line-level sparse layer and a network-level sparse layer.

Images