Railway outdoor signal equipment integrated operation and maintenance monitoring system based on machine learning

Abstract

The application discloses a railway outdoor signal equipment integrated operation and maintenance monitoring system based on machine learning and relates to the technical field of industrial equipment state monitoring.The system comprises a monitoring configuration module, a multi-source acquisition and processing module, a feature attribute division module, a cause-effect domain generation module, a cause-effect relationship learning module, a root cause verification and confirmation module and a diagnosis result output module.The monitoring configuration module is used for establishing equipment monitoring objects and monitoring points and configuring monitoring identification information.The multi-source acquisition and processing module is used for generating a multi-source state feature matrix.The feature attribute division module is used for extracting railway outdoor signal equipment state features.The cause-effect domain generation module is used for generating a local cause-effect learning matrix.The cause-effect relationship learning module is used for inputting the local cause-effect learning matrix into a DirectLiNGAM cause-effect discovery model to generate a local cause-effect adjacency matrix.The root cause verification and confirmation module is used for determining a fault root cause object.The diagnosis result output module is used for generating an operation and maintenance diagnosis result according to the fault root cause object.The application improves the accuracy and efficiency of outdoor signal equipment fault positioning and operation and maintenance disposal.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring technology, and in particular to an integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning. Background Technology

[0002] With the continuous increase in railway transport density and operating speed, the operational stability of railway outdoor signaling equipment has a significant impact on train operation safety. Existing railway outdoor signaling equipment typically includes track circuit-related equipment, signal lighting units, traction return current-related equipment, and field equipment such as cable boxes. To monitor the operational status of this equipment, current technologies typically install sensors and data acquisition units beside the track or inside the boxes to collect status data such as voltage, current, return current, temperature and humidity, water immersion, opening of the box, and vibration. This collected data is then transmitted to indoor monitoring equipment or a centralized monitoring platform to achieve status display, alarm functions, and maintenance assistance.

[0003] However, existing monitoring methods for railway outdoor signaling equipment still have certain shortcomings. On the one hand, the monitoring data sources for different types of outdoor signaling equipment are scattered, and there is a lack of a unified data organization method among track circuits, signals, traction return current, and enclosure environment. The monitoring results are mostly limited to single-point data display or threshold alarms, making it difficult to comprehensively judge the correlation changes between multi-source status data. On the other hand, existing alarm methods usually rely on fixed thresholds or human experience. When multiple monitoring points show abnormalities at the same time, it is difficult to distinguish the source of the fault, the abnormal propagation path, and the downstream response results, which can easily lead to false alarms, missed alarms, or an excessively large investigation scope.

[0004] In addition, railway outdoor signaling equipment is affected by a combination of factors such as the enclosure environment, traction return current, line connection status, and the equipment's own operating status. A single abnormal parameter may not necessarily reflect the true cause of the fault.

[0005] Therefore, how to provide an integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning. This invention uses multi-source status data acquisition, feature attribute division, minimum fault causal domain construction, DirectLiNGAM causal relationship learning, and residual independence verification to reverse-locate abnormal performance of railway outdoor signaling equipment to the root cause of the fault. This reduces misjudgments caused by single-point threshold alarms and manual experience judgment, and improves the accuracy of fault identification, the efficiency of root cause location, and the targeted nature of operation and maintenance.

[0007] The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to an embodiment of the present invention includes: The monitoring configuration module is used to establish the monitoring objects and monitoring points of railway outdoor signal equipment, and to configure monitoring identification information for each monitoring object. The multi-source acquisition and processing module is used to acquire multi-source status data of railway outdoor signaling equipment, perform data preprocessing, and generate a multi-source status feature matrix corresponding to the monitoring identification information. The feature attribute segmentation module is used to extract the status features of railway outdoor signaling equipment from the multi-source status feature matrix and segment them into fault exogenous source anchor point features and equipment response features. The causal domain generation module is used to determine the corresponding monitoring center location based on the device response characteristics, construct the minimum fault causal domain, extract the state features belonging to the minimum fault causal domain from the multi-source state feature matrix, and generate a local causal learning matrix. The causal relationship learning module is used to input the local causal learning matrix into the DirectLiNGAM causal discovery model, learn the causal order and causal connection strength between state features in the minimum fault causal domain, and generate a local causal adjacency matrix. The root cause verification module is used to trace upstream candidate root cause features from the response features of devices in an abnormal state back to the local causal adjacency matrix, and perform residual independence verification to determine the root cause object of the fault. The diagnostic results output module is used to generate operation and maintenance diagnostic results based on the root cause of the fault and output them to the railway outdoor signal equipment monitoring and management platform.

[0008] Optionally, the monitoring configuration module includes: Read the basic information and monitoring deployment information of railway outdoor signaling equipment, identify and classify the railway outdoor signaling equipment included in the monitoring scope, and form a set of monitoring objects; Based on the installation location and monitoring requirements of each monitoring object in the monitoring object set, establish monitoring points corresponding to each monitoring object to form a monitoring point set; Associate each monitoring object in the monitoring object set with the corresponding monitoring point in the monitoring point set, and configure unique monitoring identification information for each monitoring object.

[0009] Optionally, the multi-source acquisition and processing module includes: Based on the set of monitoring objects, the set of monitoring points, and the monitoring identification information, multi-source status data are collected from railway outdoor signaling equipment to form an original multi-source status dataset corresponding to the monitoring identification information. The original multi-source state dataset is sorted, merged, and aligned according to monitoring identification information and sampling time information, and data preprocessing is performed to form a window state dataset corresponding to each monitoring identification information; Window state features are extracted from the window state dataset and organized according to monitoring identification information, time window, and feature category to generate a multi-source state feature matrix corresponding to the monitoring identification information.

[0010] Optionally, the feature attribute segmentation module includes: According to the monitoring identification information, the window status features belonging to the same monitoring object are read from the multi-source status feature matrix, and the window status features are merged and organized to form a railway outdoor signal equipment status feature set; The system performs feature attribute determination on the status feature set of railway outdoor signaling equipment, and generates feature attribute determination results based on the leading nature, persistence and correlation of the status features of each railway outdoor signaling equipment in the anomaly formation process. Based on the characteristic attribute determination results, the status characteristics of railway outdoor signal equipment with source attributes are classified as fault exogenous source anchor point characteristics, and the status characteristics of railway outdoor signal equipment with response attributes are classified as equipment response characteristics.

[0011] Optionally, the causal domain generation module includes: The system identifies abnormal changes in equipment response characteristics, identifies abnormally changing equipment response characteristics as response end characteristics, and determines the corresponding monitoring objects and monitoring points based on the monitoring identification information corresponding to the response end characteristics. The current monitoring point is then designated as the monitoring center point. Starting from the monitoring center point, candidate status features are read along the field cause link of railway outdoor signal equipment to form a candidate cause feature set; Perform causal closure judgment on the candidate causal feature set, retain the candidate state features that have formed an abnormal influence transmission relationship between the fault exogenous source anchor point feature and the response end feature, exclude the candidate state features that cannot form an abnormal influence transmission relationship with the response end feature, and determine the local feature range formed by the retained candidate state features as the minimum fault causal domain. State features belonging to the minimum fault causal domain are extracted from the multi-source state feature matrix, and the extracted state features are organized according to monitoring identification information, time window, response end features and causal closure relationship to generate a local causal learning matrix.

[0012] Optionally, the causal relationship learning module includes: Read the local causal learning matrix, sample the state features in the local causal learning matrix according to the minimum fault causal domain, organize the state feature values ​​belonging to the same minimum fault causal domain within the same time window into causal learning samples, and gather the causal learning samples corresponding to multiple time windows into the input sample set. The input sample set is fed into the DirectLiNGAM causal discovery model, which performs role labeling. Sample columns from fault exogenous source anchor point features are labeled as root cause candidate columns, and sample columns from equipment response features are labeled as response columns. Role labeling results are generated, and the correspondence between each sample column and monitoring identification information, monitoring center location, and minimum fault causal domain is maintained. Causal order learning is performed based on the input sample set and role labeling results. In the unsorted state features, state features that meet the upstream candidate conditions are selected round by round and written into the causal order list. After each round of selection, the linear influence of the selected state features on the remaining state features is eliminated until the state features in the minimum fault causal domain are causally sorted. The DirectLiNGAM causal discovery model determines the upstream and downstream relationships between state features based on a causal order list. State features ranked first in the causal order list are selected as candidate sources of influence, and state features ranked last are selected as candidate objects of influence. For each object of influence, the candidate sources of influence ranked first are read, and the state feature values ​​of the candidate sources of influence are used to estimate the linear influence on the state feature values ​​of the object of influence. The causal connection strength between the corresponding state features is determined based on the linear influence estimation results. A local causal adjacency matrix is ​​generated based on the causal order list and causal connection strength. The rows and columns of the local causal adjacency matrix correspond to the state features within the minimum fault causal domain, and the matrix elements in the local causal adjacency matrix represent the causal direction and causal connection strength between the corresponding two state features.

[0013] Optionally, the root cause verification module includes: Identify the abnormal response characteristics from the device response characteristics, and use these response characteristics as the objects for reverse tracing. In the local causal adjacency matrix, starting from the matrix position corresponding to the response end feature, upstream state features that have a directed connection relationship with the response end feature are read in the opposite direction of the causal direction, and upstream state features that belong to the fault exogenous source anchor point features and can reach the response end feature through the local causal adjacency matrix are identified as upstream candidate root cause features. Based on the causal direction and causal connection strength recorded in the local causal adjacency matrix, the causal path between each upstream candidate root cause feature and the response end feature is extracted. The residual independence check is performed on each causal path. According to the causal connection strength in the local causal adjacency matrix, the linear influence of the upstream candidate root cause feature on the response end feature is subtracted from the value sequence of the response end feature to obtain the residual sequence. Then, it is determined whether there is still a statistical correlation between the residual sequence and the value sequence of the upstream candidate root cause feature. When there is no statistical correlation between the residual sequence and the value sequence of the upstream candidate root cause features that represents the residual causal dependence, the railway outdoor signal equipment object corresponding to the upstream candidate root cause features is identified as the fault root cause object.

[0014] Optionally, the diagnostic result output module includes: Read the root cause object, monitoring identification information and local causal adjacency matrix, match the root cause object with the corresponding monitoring object and monitoring point, and generate root cause location record; Based on the root cause localization records and the local causal adjacency matrix, the causal path from the fault root cause object to the device response characteristics is extracted, and the operation and maintenance diagnosis results are generated based on the causal path. The operation and maintenance diagnosis results are bound with the monitoring identification information and output to the railway outdoor signal equipment monitoring and management platform.

[0015] The beneficial effects of this invention are: This invention, through a monitoring configuration module, a multi-source acquisition and processing module, and a feature attribute division module, unifies the acquisition, preprocessing, and feature organization of multi-source status data from scattered track circuits, signal lighting units, traction return current, and enclosure environments in railway outdoor signaling equipment. This enables different monitoring objects, different monitoring points, and different acquired data to form corresponding relationships through monitoring identification information, avoiding the problems of scattered data sources and difficulty in unifying data correlation in existing technologies. This provides a complete, continuous, and traceable data foundation for subsequent fault analysis.

[0016] This invention determines the monitoring center location based on the device response characteristics through a causal domain generation module and constructs a minimum fault causal domain. This allows fault analysis to no longer blindly calculate all monitoring data, but instead extracts state features that are causally related to the current anomaly by focusing on the monitoring object and monitoring point where an abnormal response occurs, and generates a local causal learning matrix. This reduces the interference of irrelevant state features on fault judgment and improves the pertinence and stability of causal relationship learning.

[0017] This invention uses the DirectLiNGAM causal discovery model to learn the causal order and causal connection strength between state features within the minimum fault causal domain, and generates a local causal adjacency matrix. This enables the system to trace upstream candidate root cause features from device response anomalies, not only identifying whether an anomaly has occurred, but also determining the direction of anomaly propagation and the source of the root cause. This overcomes the problem that traditional threshold alarm methods have difficulty distinguishing between the source of the fault and the downstream response results.

[0018] This invention verifies the upstream candidate root cause features through residual independence verification. Only when the candidate root cause features can explain the abnormal changes at the response end are the corresponding railway outdoor signal equipment objects identified as the root cause objects of the fault. This reduces the risk of misjudgment caused by single-point anomalies, environmental disturbances, traction return fluctuations, or synchronous changes at multiple monitoring points, and improves the accuracy and reliability of fault root cause location.

[0019] This invention can transform the root cause of a fault into an operation and maintenance diagnosis result and output it to the railway outdoor signal equipment monitoring and management platform. This enables maintenance personnel to obtain operation and maintenance information such as fault type, fault location, abnormal propagation path and handling priority. This helps to narrow the scope of on-site investigation, improve maintenance efficiency, reduce reliance on manual experience, and enhance the intelligent level of integrated operation and maintenance monitoring of railway outdoor signal equipment. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning proposed in this invention; Figure 2 This is a schematic diagram of the core algorithm module structure of the integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-2 An integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning includes: The monitoring configuration module is used to establish the monitoring objects and monitoring points of railway outdoor signal equipment, and to configure monitoring identification information for each monitoring object. The multi-source acquisition and processing module is used to acquire multi-source status data of railway outdoor signaling equipment, perform data preprocessing, and generate a multi-source status feature matrix corresponding to the monitoring identification information. The feature attribute segmentation module is used to extract the status features of railway outdoor signaling equipment from the multi-source status feature matrix and segment them into fault exogenous source anchor point features and equipment response features. The causal domain generation module is used to determine the corresponding monitoring center location based on the device response characteristics, construct the minimum fault causal domain, extract the state features belonging to the minimum fault causal domain from the multi-source state feature matrix, and generate a local causal learning matrix. The causal relationship learning module is used to input the local causal learning matrix into the DirectLiNGAM causal discovery model, learn the causal order and causal connection strength between state features in the minimum fault causal domain, and generate a local causal adjacency matrix. The root cause verification module is used to trace upstream candidate root cause features from the response features of devices in an abnormal state back to the local causal adjacency matrix, and perform residual independence verification to determine the root cause object of the fault. The diagnostic results output module is used to generate operation and maintenance diagnostic results based on the root cause of the fault and output them to the railway outdoor signal equipment monitoring and management platform.

[0023] In this embodiment, the monitoring configuration module includes: The system reads the basic information and monitoring deployment information of railway outdoor signaling equipment, identifies and classifies the railway outdoor signaling equipment included in the monitoring scope, and forms a set of monitoring objects. The railway outdoor signaling equipment includes outdoor signaling equipment corresponding to 25Hz phase-sensitive track circuits, signal lighting units, traction return current, and cable box environments. The basic information includes equipment name, equipment number, station, section, and box information. The monitoring deployment information includes sensor deployment information, data acquisition unit deployment information, and communication access information. Based on the installation location and monitoring requirements of each monitoring object in the monitoring object set, monitoring points corresponding to each monitoring object are established to form a monitoring point set. The monitoring points are used to correspond to the monitoring deployment locations of track circuits, signals, traction return current, and enclosure environment. Among them, the track circuit monitoring points correspond to the relevant acquisition locations of track circuits, the signal monitoring points correspond to the relevant acquisition locations of lighting units, the traction return current monitoring points correspond to the return current acquisition locations, and the enclosure environment monitoring points correspond to the acquisition locations of the enclosure interior and outdoor environment. Each monitoring object in the monitoring object set is associated with the corresponding monitoring point in the monitoring point set, and a unique monitoring identification information is configured for each monitoring object. The monitoring identification information is used to correspond one-to-one between the monitoring object and the monitoring point.

[0024] In this embodiment, the multi-source acquisition and processing module includes: Based on the set of monitoring objects, the set of monitoring points, and the monitoring identification information, multi-source status data is collected from railway outdoor signaling equipment to form an original multi-source status dataset corresponding to the monitoring identification information. The multi-source status data is obtained by sensors and acquisition units installed in railway outdoor boxes, trackside equipment, and related connection locations. The multi-source status data corresponds to the operating status of track circuits, signal lighting units, traction return current, and box environment, including voltage and current data related to 25Hz phase-sensitive track circuits, primary and secondary status data of signal lighting units, traction return current data of traction line and rail, and temperature and humidity data inside the box, temperature and humidity data of outdoor ambient temperature and humidity, water immersion, opening, and vibration data. Each sampled data in the original multi-source status dataset corresponds to unique monitoring identification information and sampling time information. The original multi-source state dataset is sorted, merged, and aligned according to monitoring identification information and sampling time information, and data preprocessing is performed to form a window state dataset corresponding to each monitoring identification information. The data preprocessing includes data integrity verification, duplicate data removal, missing data completion, abnormal sampling data screening, noise suppression processing, time synchronization processing, and unit unification processing. Window state features are extracted from the window state dataset and organized according to monitoring identification information, time window, and feature category to generate a multi-source state feature matrix corresponding to the monitoring identification information. The window state features are used to characterize the state changes of the corresponding monitoring object within the time window. The multi-source state feature matrix contains multiple state feature values ​​corresponding to each time window under the same monitoring identification information.

[0025] In this embodiment, the feature attribute segmentation module includes: According to the monitoring identification information, the window state features belonging to the same monitoring object are read from the multi-source state feature matrix, and the window state features are merged and organized to form a railway outdoor signal equipment state feature set. The railway outdoor signal equipment state feature set is used to characterize the state changes of the corresponding monitoring object within a continuous time window. The system performs feature attribute determination on the status feature set of railway outdoor signaling equipment, and generates feature attribute determination results based on the leading nature, persistence, and correlation orientation of each railway outdoor signaling equipment status feature in the anomaly formation process. The leading nature is used to characterize whether the status feature changes abnormally first relative to other status features. The persistence is used to characterize the maintenance state of the abnormal change of the status feature within a continuous time window. The correlation orientation is used to characterize the direction of anomaly transmission between the status feature and other status features. The feature attribute determination results are used to characterize the attribute attribution of each railway outdoor signaling equipment status feature and the corresponding determination basis. Based on the characteristic attribute determination results, the status characteristics of railway outdoor signal equipment with source attributes are classified as fault exogenous source anchor point characteristics, and the status characteristics of railway outdoor signal equipment with response attributes are classified as equipment response characteristics. The source attribute is used to characterize that the corresponding railway outdoor signal equipment status characteristics change abnormally before the equipment response characteristics and point to subsequent abnormal changes. The response attribute is used to characterize that the corresponding status characteristics change in conjunction with the status characteristics corresponding to the source attribute.

[0026] In this embodiment, the causal domain generation module includes: The system identifies abnormal changes in equipment response characteristics, identifies abnormally changing equipment response characteristics as response end characteristics, and determines the corresponding monitoring objects and monitoring points based on the monitoring identification information corresponding to the response end characteristics. The current monitoring point is then designated as the monitoring center point. Starting from the monitoring center point, candidate state features are read along the field cause link of the railway outdoor signal equipment to form a candidate cause feature set. The field cause link includes the abnormal influence transmission path extending from the monitoring center point to its corresponding box, corresponding acquisition unit, corresponding acquisition channel, corresponding track circuit section, corresponding signal lighting circuit or corresponding traction return circuit. The candidate state features are the state features of the railway outdoor signal equipment that have a field cause link with the monitoring center point and have not yet undergone cause closure determination. A causal closure determination is performed on the candidate causal feature set. Candidate state features that have formed an abnormal influence transmission relationship between the fault exogenous source anchor feature and the response end feature are retained, while candidate state features that cannot form an abnormal influence transmission relationship with the response end feature are excluded. The local feature range formed by the retained candidate state features is determined as the minimum fault causal domain. The minimum fault causal domain is used to limit the range of state features that have a causal closure relationship with the current response end feature. State features belonging to the minimum fault causal domain are extracted from the multi-source state feature matrix, and the extracted state features are organized according to monitoring identification information, time window, response end features and causal closure relationship to generate a local causal learning matrix.

[0027] This invention determines the monitoring center location by identifying equipment response characteristics and converges candidate state features along the field causal chain. This ensures that the minimum fault causal domain retains only the state features that can form a causal closure relationship with the response end features, avoiding interference from irrelevant features caused by directly inputting all monitoring data into the model. This improves the matching between the local causal learning matrix and the current abnormal scenario, enhances the accuracy of expressing the fault propagation relationship of railway outdoor signal equipment, and makes subsequent causal relationship learning and root cause localization more focused, stable, and interpretable.

[0028] In this embodiment, the causal relationship learning module includes: Read the local causal learning matrix, sample the state features in the local causal learning matrix according to the minimum fault causal domain, organize the state feature values ​​belonging to the same minimum fault causal domain within the same time window into causal learning samples, and gather the causal learning samples corresponding to multiple time windows into the input sample set. The input sample set is fed into the DirectLiNGAM causal discovery model, which performs role labeling. Sample columns from fault exogenous source anchor point features are labeled as root cause candidate columns, and sample columns from equipment response features are labeled as response columns. Role labeling results are generated, and the correspondence between each sample column and monitoring identification information, monitoring center location, and minimum fault causal domain is maintained. Causal order learning is performed based on the input sample set and role labeling results. In each round of unsorted state features, state features that meet the upstream candidate conditions are selected and written into the causal order list. After each round of selection, the linear influence of the selected state features on the remaining state features is eliminated until the state features in the minimum fault causal domain are causally ordered. The upstream candidate conditions refer to the process in each round of causal order learning. First, state features marked as root cause candidates are selected from the input sample set. Then, the influence relationship between the current state feature and other unsorted state features is judged. Specifically, the linear influence of the current state feature on other unsorted state features is removed from the corresponding state features to obtain the remaining data after the influence is removed. The statistical correlation between the current state feature and the remaining data is used as the judgment criterion. The lower the statistical correlation, the closer the current state feature is to the causal upstream position of the current round. The DirectLiNGAM causal discovery model writes the state feature with the lowest statistical correlation into the causal order list. The DirectLiNGAM causal discovery model determines the upstream and downstream relationships between state features based on a causal order list. State features ranked first in the causal order list are selected as candidate sources of influence, and state features ranked later are selected as candidate objects of influence. For each object of influence, the candidate sources of influence ranked first are read, and the state feature values ​​of the candidate sources of influence are used to estimate the linear influence on the state feature values ​​of the object of influence. The causal connection strength between corresponding state features is determined based on the linear influence estimation results. Specifically, for any object of influence, one or more candidate sources of influence ranked before the current object of influence are selected in the causal order list. The state feature values ​​of the candidate sources of influence within each time window are used as input, and the state feature values ​​of the object of influence within the same time window are used as output. The linear influence degree of the candidate sources of influence on the object of influence is estimated, and the current linear influence degree is used as the causal connection strength of the corresponding directed connection. A local causal adjacency matrix is ​​generated based on the causal order list and the causal connection strength. The rows and columns of the local causal adjacency matrix correspond to the state features within the minimum fault causal domain, and the matrix elements in the local causal adjacency matrix represent the causal direction and causal connection strength between the corresponding two state features. The DirectLiNGAM causal discovery model is used to learn causal relationships among state features within the minimum fault causal domain. Its input is an input sample set obtained by organizing the local causal learning matrix. Each row in the input sample set corresponds to the state feature value within a time window, and each column corresponds to a class of state features within the minimum fault causal domain. The DirectLiNGAM causal discovery model does not rely on manual fault labels for training. Instead, it determines the causal order of each state feature in the anomaly formation process step by step based on the non-Gaussianity, linear influence relationship, and residual independence among the state features. After determining the causal order, it estimates the causal connection strength between upstream state features and downstream state features, generating a local causal adjacency matrix. The local causal adjacency matrix is ​​used to record the causal direction and causal connection strength among the state features within the minimum fault causal domain.

[0029] This invention organizes the state features within the minimum fault causal domain into an input sample set and uses the DirectLiNGAM causal discovery model to learn the causal order and causal connection strength between fault exogenous source anchor features and equipment response features. This enables the multi-source state data of railway outdoor signaling equipment to move beyond correlation analysis or threshold alarms and form a local causal adjacency matrix with a clear causal direction. This improves the accuracy of anomaly propagation relationship identification, enhances the interpretability of fault root cause tracing, reduces the interference of irrelevant state features on diagnostic results, and provides a reliable basis for subsequent fault root cause confirmation and operation and maintenance diagnosis output.

[0030] In this embodiment, the root cause verification module includes: Identify the abnormal response characteristics from the device response characteristics, and use these response characteristics as the objects for reverse tracing. In the local causal adjacency matrix, starting from the matrix position corresponding to the response end feature, upstream state features that have a directed connection relationship with the response end feature are read in the opposite direction of the causal direction, and upstream state features that belong to the fault exogenous source anchor point features and can reach the response end feature through the local causal adjacency matrix are identified as upstream candidate root cause features. Based on the causal direction and causal connection strength recorded in the local causal adjacency matrix, the causal path between each upstream candidate root cause feature and the response end feature is extracted. The residual independence check is performed on each causal path. According to the causal connection strength in the local causal adjacency matrix, the linear influence of the upstream candidate root cause feature on the response end feature is subtracted from the value sequence of the response end feature to obtain the residual sequence. Then, it is determined whether there is still a statistical correlation between the residual sequence and the value sequence of the upstream candidate root cause feature. When there is no statistical correlation between the residual sequence and the value sequence of the upstream candidate root cause features that represents the residual causal dependence, the railway outdoor signal equipment object corresponding to the upstream candidate root cause features is identified as the fault root cause object.

[0031] This invention uses a local causal adjacency matrix to trace the response characteristics of equipment in an abnormal state in reverse. It identifies the exogenous fault anchor point features that can reach the response end features as candidate root causes, and combines causal connection strength and residual independence checks to determine the explanatory power of the candidate root causes for the response anomalies. This makes the identification of fault root causes no longer dependent on single-point threshold alarms or manual experience judgment, effectively distinguishing the fault source from the downstream response results, reducing the risk of misjudgment caused by synchronous anomalies at multiple monitoring points, and improving the accuracy, reliability, and maintenance efficiency of fault location for railway outdoor signaling equipment.

[0032] In this embodiment, the diagnostic result output module includes: Read the fault root cause object, monitoring identification information and local causal adjacency matrix, match the fault root cause object with the corresponding monitoring object and monitoring point, and generate a root cause location record. The root cause location record is used to characterize the equipment ownership, point ownership and data ownership corresponding to the fault root cause object. Based on the root cause localization record and the local causal adjacency matrix, the causal path from the fault root cause object to the equipment response characteristics is extracted, and the operation and maintenance diagnosis results are generated based on the causal path. The operation and maintenance diagnosis results include fault type, fault location, abnormal propagation path, equipment status and handling priority, which are used to indicate the fault type, fault location and abnormal propagation path corresponding to the current abnormality, so that maintenance personnel can quickly determine which box, acquisition unit, track circuit point, traction return point or signal lighting unit should be checked first. The operation and maintenance diagnosis results are bound with the monitoring identification information and output to the railway outdoor signal equipment monitoring and management platform. The railway outdoor signal equipment monitoring and management platform receives the operation and maintenance diagnosis results output by the diagnostic host. The diagnostic host is used to store, analyze, display and alarm the returned data, local causal adjacency matrix and fault root cause objects.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to the operation and maintenance monitoring scenario of outdoor signal equipment in a railway section. This scenario included multiple cable boxes, several 25Hz phase-sensitive track circuit sections, signal lighting units, traction return current monitoring points, and box environment monitoring points. The original operation and maintenance method for this section mainly relied on electrical parameter alarms from a centralized monitoring system, on-site personnel inspection records, and manual judgment based on experience. When track circuit current fluctuations, signal lighting current abnormalities, increased box humidity, or traction return current fluctuations occurred simultaneously, maintenance personnel typically needed to investigate box by box, circuit by circuit, and data collection point by data collection point. This made it difficult to quickly determine whether the source of the anomaly was a box environment problem, traction return current disturbance, lighting unit malfunction, or a change in the track circuit's own condition. This resulted in a large investigation scope, weak alarm interpretation, and delayed confirmation of the root cause of the fault.

[0034] In this embodiment, sensors and data acquisition units are installed at the railway outdoor enclosures and trackside equipment to collect data on the voltage and current status of the 25Hz phase-sensitive track circuit, the primary and secondary status of the signal lighting unit, the traction return current of the traction line and rail, as well as the temperature and humidity inside the enclosure, the ambient temperature and humidity outside, water immersion, opening, and vibration status. Each sampled data is bound to monitoring identification information and sampling time information, enabling the collected data to be mapped to a specific monitoring object, monitoring point, enclosure, and data acquisition unit. The collected multi-source status data is processed through sorting, merging, alignment, integrity verification, missing data completion, abnormal sampling removal, noise suppression, time synchronization, and unit unification to form a multi-source status feature matrix.

[0035] During operation, the system extracts the status features of railway outdoor signaling equipment from a multi-source status feature matrix. Based on the leading nature, persistence, and correlation of these status features in the anomaly formation process, the system categorizes them into exogenous fault anchor point features and equipment response features. When a track circuit section experiences a drop in receiving end current, fluctuations in rail side current, and an imbalance in rail lead-in current, the system does not directly identify that track circuit section as a fault point. Instead, it first determines these manifestations as equipment response features, then determines the corresponding monitoring center point based on the equipment response features, and establishes a field cause link along the box, acquisition unit, acquisition channel, track circuit section, and traction return circuit to that point.

[0036] When constructing the minimum fault causal domain, the system reads candidate state features from the multi-source state feature matrix that have a field causal link with the monitoring center point. It then performs a causal closure judgment on these candidate state features, retaining only those that can form an abnormal influence transmission relationship with the response end features. For example, in a trial run, the humidity inside a certain box continuously increased from the normal range of 58% to 86%, and the box's water immersion state changed from non-triggered to triggered. Subsequently, the current at the receiving end of the track circuit associated with this box decreased over several consecutive time windows, and the imbalance of the rail lead-in current increased. The system uses the box's humidity and water immersion state as exogenous fault anchor features, and the track circuit current change and lead-in current imbalance as equipment response features. It then forms the minimum fault causal domain around this causal closure relationship, avoiding including unrelated signal light current fluctuations or other box environmental disturbances within the same section into the current causal learning scope.

[0037] Subsequently, the system organizes the state features within the minimum fault causal domain into a local causal learning matrix and inputs it into the DirectLiNGAM causal discovery model. The model performs causal order learning on the input sample set obtained by organizing the local causal learning matrix, generating a local causal adjacency matrix to represent the causal direction and causal connection strength between the fault exogenous source anchor features and the equipment response features. For equipment response features in an abnormal state, the system traces upstream candidate root cause features in the reverse direction of the local causal adjacency matrix and uses residual independence verification to determine whether the candidate root cause features can explain the abnormal changes at the response end. If, after removing the linear influence of the upstream candidate root cause features on the response end features, there is no longer a statistical correlation between the residual sequence and the upstream candidate root cause features that characterizes the remaining causal dependence, then the railway outdoor signaling equipment object corresponding to the upstream candidate root cause feature is identified as the fault root cause object.

[0038] During continuous trial operation, the system of this invention was compared with the original threshold alarm plus manual troubleshooting method. The original threshold alarm plus manual troubleshooting method is a common handling method in the operation and maintenance of existing railway outdoor signal equipment. In this method, the system first sets alarm thresholds for monitoring data such as track circuit voltage and current, signal light current, traction return current, box temperature and humidity, water immersion, box opening, and vibration. When a certain monitored value exceeds or falls below the threshold, the monitoring platform issues an alarm. Then, maintenance personnel check the cause of the fault item by item based on the alarm point, historical experience, and on-site inspection results. The specific comparison results are shown in the table below: Table 1 Comparative Verification Results of Integrated Operation and Maintenance Monitoring Methods for Railway Outdoor Signaling Equipment

[0039] As shown in the table, under the same conditions of the number of effective monitoring windows and the number of abnormal event records, the number of abnormal events correctly identified by the system of the present invention increased from 35 to 44, the anomaly identification accuracy increased from 74.47% to 93.62%, and the root cause localization accuracy increased from 61.70% to 89.36%. This indicates that the present invention, through the minimum fault causal domain, the DirectLiNGAM causal discovery model, and residual independence verification, can more accurately distinguish between the source of the fault and the downstream response results. At the same time, the average root cause confirmation time was shortened from 52 minutes to 14 minutes, the average number of on-site inspection points decreased from 7.8 to 2.6, and the number of misjudgments of multi-point synchronous anomalies decreased from 11 to 3. This shows that the present invention can effectively reduce the workload of manual inspection and improve the efficiency and reliability of operation and maintenance of railway outdoor signal equipment.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based integrated operation and maintenance monitoring system for railway outdoor signaling equipment, characterized in that, include: The monitoring configuration module is used to establish the monitoring objects and monitoring points of railway outdoor signal equipment, and to configure monitoring identification information for each monitoring object. The multi-source acquisition and processing module is used to acquire multi-source status data of railway outdoor signaling equipment, perform data preprocessing, and generate a multi-source status feature matrix corresponding to the monitoring identification information. The feature attribute segmentation module is used to extract the status features of railway outdoor signaling equipment from the multi-source status feature matrix and segment them into fault exogenous source anchor point features and equipment response features. The causal domain generation module is used to determine the corresponding monitoring center location based on the device response characteristics, construct the minimum fault causal domain, extract the state features belonging to the minimum fault causal domain from the multi-source state feature matrix, and generate a local causal learning matrix. The causal relationship learning module is used to input the local causal learning matrix into the DirectLiNGAM causal discovery model, learn the causal order and causal connection strength between state features in the minimum fault causal domain, and generate a local causal adjacency matrix. The root cause verification module is used to trace upstream candidate root cause features from the response features of devices in an abnormal state back to the local causal adjacency matrix, and perform residual independence verification to determine the root cause object of the fault. The diagnostic results output module is used to generate operation and maintenance diagnostic results based on the root cause of the fault and output them to the railway outdoor signal equipment monitoring and management platform.

2. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The monitoring configuration module includes: Read the basic information and monitoring deployment information of railway outdoor signaling equipment, identify and classify the railway outdoor signaling equipment included in the monitoring scope, and form a set of monitoring objects; Based on the installation location and monitoring requirements of each monitoring object in the monitoring object set, establish monitoring points corresponding to each monitoring object to form a monitoring point set; Associate each monitoring object in the monitoring object set with the corresponding monitoring point in the monitoring point set, and configure unique monitoring identification information for each monitoring object.

3. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The multi-source acquisition and processing module includes: Based on the set of monitoring objects, the set of monitoring points, and the monitoring identification information, multi-source status data are collected from railway outdoor signaling equipment to form an original multi-source status dataset corresponding to the monitoring identification information. The original multi-source state dataset is sorted, merged, and aligned according to monitoring identification information and sampling time information, and data preprocessing is performed to form a window state dataset corresponding to each monitoring identification information; Window state features are extracted from the window state dataset and organized according to monitoring identification information, time window, and feature category to generate a multi-source state feature matrix corresponding to the monitoring identification information.

4. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The feature attribute segmentation module includes: According to the monitoring identification information, the window status features belonging to the same monitoring object are read from the multi-source status feature matrix, and the window status features are merged and organized to form a railway outdoor signal equipment status feature set; The system performs feature attribute determination on the status feature set of railway outdoor signaling equipment, and generates feature attribute determination results based on the leading nature, persistence and correlation of the status features of each railway outdoor signaling equipment in the anomaly formation process. Based on the characteristic attribute determination results, the status characteristics of railway outdoor signal equipment with source attributes are classified as fault exogenous source anchor point characteristics, and the status characteristics of railway outdoor signal equipment with response attributes are classified as equipment response characteristics.

5. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The causal domain generation module includes: The system identifies abnormal changes in equipment response characteristics, identifies abnormally changing equipment response characteristics as response end characteristics, and determines the corresponding monitoring objects and monitoring points based on the monitoring identification information corresponding to the response end characteristics. The current monitoring point is then designated as the monitoring center point. Starting from the monitoring center point, candidate status features are read along the field cause link of railway outdoor signal equipment to form a candidate cause feature set; Perform causal closure judgment on the candidate causal feature set, retain the candidate state features that have formed an abnormal influence transmission relationship between the fault exogenous source anchor point feature and the response end feature, exclude the candidate state features that cannot form an abnormal influence transmission relationship with the response end feature, and determine the local feature range formed by the retained candidate state features as the minimum fault causal domain. State features belonging to the minimum fault causal domain are extracted from the multi-source state feature matrix, and the extracted state features are organized according to monitoring identification information, time window, response end features and causal closure relationship to generate a local causal learning matrix.

6. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The causal relationship learning module includes: Read the local causal learning matrix, sample the state features in the local causal learning matrix according to the minimum fault causal domain, organize the state feature values ​​belonging to the same minimum fault causal domain within the same time window into causal learning samples, and gather the causal learning samples corresponding to multiple time windows into the input sample set. The input sample set is fed into the DirectLiNGAM causal discovery model, which performs role labeling. Sample columns from fault exogenous source anchor point features are labeled as root cause candidate columns, and sample columns from equipment response features are labeled as response columns. Role labeling results are generated, and the correspondence between each sample column and monitoring identification information, monitoring center location, and minimum fault causal domain is maintained. Causal order learning is performed based on the input sample set and role labeling results. In the unsorted state features, state features that meet the upstream candidate conditions are selected round by round and written into the causal order list. After each round of selection, the linear influence of the selected state features on the remaining state features is eliminated until the state features in the minimum fault causal domain are causally sorted. The DirectLiNGAM causal discovery model determines the upstream and downstream relationships between state features based on a causal order list. State features ranked first in the causal order list are selected as candidate sources of influence, and state features ranked last are selected as candidate objects of influence. For each object of influence, the candidate sources of influence ranked first are read, and the state feature values ​​of the candidate sources of influence are used to estimate the linear influence on the state feature values ​​of the object of influence. The causal connection strength between the corresponding state features is determined based on the linear influence estimation results. A local causal adjacency matrix is ​​generated based on the causal order list and causal connection strength. The rows and columns of the local causal adjacency matrix correspond to the state features within the minimum fault causal domain, and the matrix elements in the local causal adjacency matrix represent the causal direction and causal connection strength between the corresponding two state features.

7. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The root cause verification module includes: Identify the abnormal response characteristics from the device response characteristics, and use these response characteristics as the objects for reverse tracing. In the local causal adjacency matrix, starting from the matrix position corresponding to the response end feature, upstream state features that have a directed connection relationship with the response end feature are read in the opposite direction of the causal direction, and upstream state features that belong to the fault exogenous source anchor point features and can reach the response end feature through the local causal adjacency matrix are identified as upstream candidate root cause features. Based on the causal direction and causal connection strength recorded in the local causal adjacency matrix, the causal path between each upstream candidate root cause feature and the response end feature is extracted. The residual independence check is performed on each causal path. According to the causal connection strength in the local causal adjacency matrix, the linear influence of the upstream candidate root cause feature on the response end feature is subtracted from the value sequence of the response end feature to obtain the residual sequence. Then, it is determined whether there is still a statistical correlation between the residual sequence and the value sequence of the upstream candidate root cause feature. When there is no statistical correlation between the residual sequence and the value sequence of the upstream candidate root cause features that represents the residual causal dependence, the railway outdoor signal equipment object corresponding to the upstream candidate root cause features is identified as the fault root cause object.

8. The integrated operation and maintenance monitoring system for railway outdoor signaling equipment based on machine learning according to claim 1, characterized in that, The diagnostic result output module includes: Read the root cause object, monitoring identification information and local causal adjacency matrix, match the root cause object with the corresponding monitoring object and monitoring point, and generate root cause location record; Based on the root cause localization records and the local causal adjacency matrix, the causal path from the fault root cause object to the device response characteristics is extracted, and the operation and maintenance diagnosis results are generated based on the causal path. The operation and maintenance diagnosis results are bound with the monitoring identification information and output to the railway outdoor signal equipment monitoring and management platform.

Images