Subway operation decision optimization method based on knowledge graph and cross-modal association

Abstract

The application discloses a subway operation decision optimization method based on a knowledge graph and cross-modal correlation, relates to the technical field of subway operation, and comprises the following steps: deploying an edge computing node on an edge device of a subway and configuring a general weak classifier; performing calculation preference analysis to generate a preference database; performing reinforcement learning on the general weak classifier; identifying fault decision after receiving communication data, establishing an edge fault decision identification result; obtaining operation environment data and user feedback data of the subway, synchronizing the data to a subway knowledge graph, and generating a subway operation and maintenance decision optimization result. The method solves the technical problems that the prior art cannot efficiently process multi-modal big data, cannot fully mine the potential correlation of different modal data, and results in insufficient accuracy of subway fault decision identification and poor operation efficiency, realizes dynamic updating and multi-modal fusion of the subway operation knowledge graph, and achieves the technical effect of improving the accuracy of subway operation decision and operation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of subway operation, and particularly relates to a subway operation decision optimization method based on a knowledge graph and cross-modal association. BACKGROUND

[0002] With the continuous expansion of the city, as the core force of urban transportation, the efficiency and stability of subway operation are increasingly critical. The traditional subway operation decision method gradually shows its limitations in the face of complex and variable operating conditions. On the one hand, the subway system contains a large number of sensors, video monitoring devices, and inspection unmanned aerial vehicles, etc., and the data generated by them are large in scale and diverse in type, covering device status, environmental parameters, passenger behavior and other multi-modal information. How to efficiently process and analyze these data becomes a problem. On the other hand, the isolated decision-making method in the past cannot fully consider various factors in subway operation, and cannot fully mine the potential association between different modal data, resulting in insufficient accuracy and timeliness of decision-making, thereby affecting the overall operation efficiency and reliability of the subway.

[0003] In the related art at present, there are technical problems that it is difficult to efficiently process multi-modal big data, and the potential association between different modal data cannot be fully mined, resulting in insufficient accuracy of subway fault decision recognition and poor operation efficiency. SUMMARY

[0004] The present application provides a subway operation decision optimization method based on a knowledge graph and cross-modal association, which solves the technical problems in the prior art that it is difficult to efficiently process multi-modal big data, and the potential association between different modal data cannot be fully mined, resulting in insufficient accuracy of subway fault decision recognition and poor operation efficiency, and realizes the construction of a dynamically updated and multi-modal fused subway operation knowledge graph, thereby achieving the technical effect of improving the accuracy of subway operation decision and operation efficiency.

[0005] The application provides a subway operation decision optimization method based on a knowledge graph and cross-modal association, including: deploying an edge computing node at an edge device of a subway, configuring a general weak classifier for the edge computing node, the edge computing node being in communication connection with sensors, video monitoring, and inspection drones; performing computing preference analysis of the edge computing node to generate a preference database; using the preference database to perform reinforcement learning on the general weak classifier, integrating a reinforcement learning classifier with the general weak classifier into an edge decision classifier, and setting the edge decision classifier in the corresponding edge computing node after reinforcement learning, which includes using the preference database to build a reinforcement learning environment, loading the general weak classifier to perform learning iteration based on dynamic rewards, and building a reinforcement learning classifier; after the edge computing node receives communication data, using the edge decision classifier to perform fault decision identification, which includes respectively obtaining fault decision identification results of the reinforcement learning classifier and the general weak classifier, performing trust degree weighted calculation, and establishing an edge fault decision identification result; obtaining operation environment data and user feedback data of the subway, building cross-border data, synchronizing the cross-border data and the edge fault decision identification result to a subway knowledge graph, and generating a subway operation and maintenance decision optimization result.

[0006] In possible implementation manners, the subway operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: obtaining a monitoring coverage area of the edge computing node, and configuring a location sensitivity of the monitoring coverage area; reading a fault event in the monitoring coverage area, weighting and calculating an occurrence probability of the fault event through the location sensitivity, and establishing a first preference constraint; performing device aging analysis under regional environment fitting on the monitoring coverage area, and using a device aging analysis result to establish a second preference constraint; generating the preference database based on the first preference constraint and the second preference constraint.

[0007] In possible implementation manners, the subway operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: reading the preference database, performing reinforcement data matching based on the preference database, and building a reinforcement learning environment; loading the general weak classifier, using the reinforcement learning environment to perform parameter iteration of the general weak classifier, and setting a dynamic reward feedback in each round of parameter iteration; when a prediction result meets an expected threshold, outputting an optimized general weak classifier as a reinforcement learning classifier.

[0008] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: evaluating the current iteration stage of the general weak classifier, and establishing a stage influence coefficient; obtaining a prediction result of the general weak classifier in this round, identifying a prediction deviation of the prediction result based on the reinforcement learning environment, and generating a first error influence coefficient; establishing a linkage window, performing prediction backtracking analysis of the general weak classifier based on the linkage window, and generating a second error influence coefficient according to a prediction backtracking analysis result and a parameter iteration direction; and setting a dynamic reward feedback by using the stage influence coefficient, the first error influence coefficient, and the second error influence coefficient.

[0009] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: performing communication data-based fault decision identification by using the reinforcement learning classifier and the general weak classifier in the edge decision classifier respectively, establishing a first fault decision identification result and a second fault decision identification result, the first fault decision identification result being an identification result of the reinforcement learning classifier, and the second fault decision identification result being an identification result of the general weak classifier; performing identification trustworthiness analysis on the first fault decision identification result and the reinforcement learning classifier, and establishing a first trust factor; performing identification trustworthiness analysis on the second fault decision identification result and the reinforcement learning classifier, and establishing a second trust factor; and establishing an edge fault decision identification result based on the first fault decision identification result, the first trust factor, the second fault decision identification result, and the second trust factor.

[0010] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: performing trust weighting on the first fault decision identification result by using the first trust factor, and establishing a third fault decision identification result; performing trust weighting on the second fault decision identification result by using the second trust factor, and establishing a fourth fault decision identification result; performing trust conflict analysis on the third fault decision identification result and the fourth fault decision identification result, and establishing an edge fault decision identification result.

[0011] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: extracting voice data and complaint data of a user, taking the voice data and the complaint data as user feedback data, performing key information extraction on the user feedback data by using a natural language processing technology, and constructing user cross-border data; reading time sequence running noise data of a metro, taking the time sequence running noise data as running environment data, and constructing environment cross-border data according to the running environment data; and constructing cross-border data by using the user cross-border data and the environment cross-border data.

[0012] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: a set of emergency plans is established; plan matching of the set of emergency plans is performed based on the metro operation decision optimization result, a plan matching result is established; and emergency response management is performed by using the plan matching result.

[0013] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: abnormality identification of a metro knowledge graph is performed based on the cross-border data and the edge fault decision identification result, an abnormality identification result is established; association identification of the abnormality identification result is performed by using the metro knowledge graph, fault decision reconstruction is performed according to the association identification result, and a metro operation decision optimization result is generated.

[0014] In a possible implementation, the metro operation decision optimization method based on the knowledge graph and the cross-modal association further performs the following processing: data integrity verification is performed on the communication data, a data integrity verification result is established; and if the data integrity verification result is a verification failure result, data abnormality is reported.

[0015] The metro operation decision optimization method based on the knowledge graph and the cross-modal association provided in the present application can deploy an edge computing node on an edge device of a metro and configure a general weak classifier; perform computing preference analysis to generate a preference database; perform reinforcement learning on the general weak classifier and integrate the general weak classifier into an edge decision classifier which is then set on a corresponding edge computing node; receive communication data to perform fault decision identification and establish an edge fault decision identification result; obtain operation environment data and user feedback data of the metro, construct cross-border data, synchronize the cross-border data to a metro knowledge graph, and generate a metro operation decision optimization result. The technical problems of being difficult to efficiently process multi-modal big data, being unable to fully mine potential associations of different modal data, resulting in insufficient accuracy of metro fault decision identification and poor operation efficiency in the prior art are solved, dynamic updating and multi-modal fusion of a metro operation knowledge graph are implemented, and the technical effect of improving the accuracy of metro operation decision and operation efficiency is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below. In the present application, a flowchart is used to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in sequence. On the contrary, various steps can be processed in reverse order or simultaneously as needed. Meanwhile, other operations can be added to these processes, or one or more steps of operations can be removed from these processes.

[0017] Figure 1A flowchart of a subway operation decision optimization method based on a knowledge graph and cross-modal association provided by an embodiment of the present application is shown.

[0018] Figure 2 A flowchart of generating a preference database in the subway operation decision optimization method based on a knowledge graph and cross-modal association provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0019] The above description is only a summary of the technical solutions of the present application. In order to make the technical solutions of the present application more clear, the following will describe the specific embodiments of the present application in the light of the contents of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following will describe the specific embodiments of the present application.

[0020] In order to make the purposes, technical solutions and advantages of the present application more clear, the following will describe the specific embodiments of the present application in the light of the contents of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following will describe the specific embodiments of the present application.

[0021] In the following description, “some embodiments” are described, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict, and the term “first\second” referred to only distinguishes similar objects, and does not represent a specific order of the objects. The terms “include” and “have” and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, product or server including a series of steps does not have to be limited to those clearly listed, but can include other steps not clearly listed or inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art of the technology to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application.

[0022] The embodiments of the present application provide a subway operation decision optimization method based on a knowledge graph and cross-modal association, as shown in Figure 1 The method comprises the following steps:

[0023] In step S100, an edge computing node is deployed on an edge device of a subway, and a general weak classifier is configured for the edge computing node, and the edge computing node is in communication connection with a sensor, a video monitoring device and a patrol unmanned aerial vehicle.

[0024] Preferably, the edge device refers to various devices used for subway operation, such as ticket gates, ticket vending machines, ventilation equipment, power supply equipment in subway stations, and various control devices on subway trains, etc., which are distributed in various corners of the subway system and are directly related to the operation and service of the subway; then deploy edge computing nodes on the edge devices of the subway, wherein the edge computing nodes contain devices with certain computing capabilities, which can perform data processing and analysis close to the data source, reduce the bandwidth pressure and delay of data transmission to the central server, and realize real-time or near-real-time monitoring and response to the operation status of the subway. Specifically, the needs of different business scenarios in subway operation for edge computing are specified, such as real-time monitoring of device status, analysis of video images to detect abnormal behavior, etc., and the number, distribution location and computing capability requirements of the required edge computing nodes are determined according to the scale, layout, etc. of the subway; according to the physical structure and network layout of the subway, the network topology of the edge computing nodes is designed, and the edge computing nodes are ensured to be in efficient and stable communication connection with the existing sensors, video surveillance, inspection drones and other related devices of the subway, for example, a combination of wired networks (such as optical fiber, Ethernet) and wireless networks (such as Wi-Fi, 5G) is adopted to meet the connection needs of devices in different positions.

[0025] Preferably, according to the computing requirements, suitable edge computing devices are selected, which should have low power consumption, high reliability and certain computing capability, and be able to operate stably in harsh subway environment, which may include selecting industrial-grade servers, embedded computers, edge computing gateways, etc. For example, for scenarios that require processing of large amounts of video data, edge computing servers with powerful graphics processing capability (GPU) are selected; according to data storage requirements, appropriate storage devices are selected. Considering the importance and real-time nature of subway operation data, high-speed and large-capacity solid state drives (SSD) or disk arrays are usually selected to ensure fast read and write and safety of data; according to the communication needs with external devices, corresponding communication modules are selected, such as Ethernet interface module for wired network connection, Wi-Fi or 5G module for wireless network connection, to ensure that edge computing nodes can communicate seamlessly with various devices; then according to the planned position, install the edge computing node devices near the edge devices of the subway, for the equipment room of the subway station, the equipment box in the tunnel, etc., select appropriate installation supports or cabinets to ensure firm installation of the devices, facilitate maintenance and management, and at the same time, pay attention to the problems of heat dissipation, dust prevention and moisture prevention of the devices, provide a good operating environment for the devices; finally complete the deployment of edge computing nodes and perform performance test optimization.

[0026] Preferably, a general weak classifier is configured for the edge computing node, wherein the general weak classifier is a relatively simple, limited classification ability but with a certain generality of classification model, when processing data, it can usually only make a basic classification judgment based on part of the features, and its classification accuracy may not be high, but the calculation cost is low, the processing speed is fast, and it is suitable for running on the edge computing node with relatively limited resources, for example, preliminary classification of various types of data of subway operation (such as device state data collected by sensors, image data in video monitoring, etc.); Specifically, according to the data characteristics and business needs generated by subway operation, select appropriate general weak classifier models, such as decision tree stump and naive Bayes classifier, collect and organize multi-modal data from subway sensors, video monitoring, inspection drones and other equipment, and perform data cleaning and feature extraction, such as extracting the behavior characteristics of personnel from video monitoring data and extracting the temperature, pressure and other characteristics of equipment from sensor data; Then use these data to train the general weak classifier model, obtain the general weak classifier and deploy it to the edge computing node of the subway edge device, avoiding the transmission of a large amount of raw data, thereby reducing the network bandwidth pressure and reducing the data transmission cost.

[0027] Preferably, the edge computing node is connected with sensors (such as temperature sensors, pressure sensors, current sensors, etc.) to obtain the data collected by the sensors in real time and understand the running parameters and environmental conditions of the subway equipment; the edge computing node is connected with video monitoring to obtain video stream data (including monitoring passenger behavior, station environment and train operation), and analyze the image information in the video using the general weak classifier, such as detecting whether there is abnormal behavior, equipment failure signs, etc.; the edge computing node is connected with the inspection drone to receive the image and video data taken by the drone during inspection and the environmental data collected, and integrate the data from different data sources, thereby expanding the monitoring range and improving the safety and reliability of subway operation.

[0028] Step S200, perform computing preference analysis of the edge computing node to generate a preference database.

[0029] Preferably, the edge computing node performs a calculation preference analysis, which can include analyzing the number of failures and failure detection, aging prediction based on environmental data and equipment state detection. Specifically, the edge computing node collects information from various data sources such as sensors, video surveillance, inspection drones, etc. to statistically record the number of failures that have occurred in each device or area in the subway system. For example, the number of times a device fails can be recorded by a failure alarm sensor built into the device, or the number of times a device exhibits abnormal conditions can be analyzed from video surveillance. The number of failures reflects the reliability and stability of the device or area, and devices or areas with a high number of failures may have potential problems. Features related to failures are extracted from the data obtained from sensors and other data sources. For example, for motor equipment, features such as current, voltage, and speed can be extracted, and for track equipment, features such as track deformation and wear can be extracted. Machine learning, deep learning, and other technologies are used to analyze and process the extracted features to identify possible failure patterns. For example, by analyzing motor current data, it can be determined whether the motor has a short circuit, overload, or other failure.

[0030] Preferably, sensors are used to collect relevant data about the subway operating environment, such as temperature, humidity, dust concentration, vibration, etc. These environmental factors can affect the aging rate of subway equipment. For example, high temperatures and high humidity can accelerate the aging of electronic equipment, and frequent vibrations can cause mechanical parts to wear out more quickly. Based on the collected environmental data, the state of the equipment is evaluated. For example, the operating state of the equipment is divided into normal, warning, failure, and other levels. When a parameter of the equipment exceeds the normal range, a warning signal is sent in a timely manner. In combination with the design parameters and historical operation data of the equipment, an aging prediction model is established. Through this model, the aging degree and remaining service life of the equipment under the current environment can be predicted. For example, according to the temperature-aging curve of the equipment, it can be predicted how much the aging rate of the equipment will increase in a high-temperature environment. Finally, the data obtained from the calculation preference analysis is integrated, including the number of failures that have occurred, the results of the aging prediction, the results of the failure detection, and the data of the equipment state monitoring, etc. The data is stored in a certain format and structure to form a preference database, and the data in the database is labeled, including the source of the data, the collection time, the type of data, etc. At the same time, a data index is established to quickly query and retrieve the required data.

[0031] Further, as Figure 2As shown, step S200 further comprises step S210, obtaining the monitoring coverage area of the edge computing node and configuring the position sensitivity of the monitoring coverage area; step S220, reading the fault events in the monitoring coverage area and establishing the first preference constraint after weighting the occurrence probability of the fault events by the position sensitivity; step S230, performing device aging analysis under regional environment fitting on the monitoring coverage area and establishing the second preference constraint by using the device aging analysis result; and step S240, generating the preference database based on the first preference constraint and the second preference constraint.

[0032] Preferably, each edge computing node is connected with sensors, video monitoring, inspection drones and other devices in the subway and is distributed at different positions of the subway, and the range that can be monitored constitutes the monitoring coverage area of the edge computing node. For example, an edge computing node installed at a specific position of a subway platform can monitor the temperature, humidity, passenger flow and other data of the platform by the connected sensors and monitor the personnel activities and device status of the platform by the camera, and the platform related area is the monitoring coverage area of the edge computing node, which helps to determine the range monitored by the edge computing node. In the monitoring coverage area, the influence of faults occurring at different positions on the subway operation is different, and each position is configured with a corresponding sensitivity, wherein the sensitivity reflects the importance or attention of the position fault occurrence. For example, the sensitivity of the position where the key braking device of the subway train is located is set to be high, and the sensitivity of the position where the auxiliary lighting device is located is relatively low.

[0033] Preferably, in the monitoring coverage area, the possible fault events are read in real time by various monitoring devices, which can be device faults (such as motor faults, signal system faults), environmental abnormalities (such as excessive temperature, abnormal humidity) and the like. For example, if a sensor detects that the temperature of a device exceeds the normal range, it is recorded as a fault. The occurrence probability of the read fault events is weighted by the configured position sensitivity, that is, the weight of the fault events occurring at the position with high sensitivity is larger, and the proportion in the overall calculation is higher. The weight of the fault events occurring at the position with low sensitivity is smaller. For example, the occurrence probability of a fault event at a high sensitivity position is 10%, and the weight is 0.8. The occurrence probability of a fault event at a low sensitivity position is 20%, and the weight is 0.2. After weighting calculation, the fault occurrence condition considering the importance of the position is obtained. Based on the result of the weighting calculation, the first preference constraint is established, which reflects the relationship between the possibility and importance of the fault event occurrence in the case of considering the position sensitivity. For example, when performing resource allocation or decision making, the positions and events with higher fault occurrence probability after weighting are preferentially processed.

[0034] Preferably, the monitoring coverage area is regionally fitted with an environment, i.e. the relationship between the environmental factors (such as temperature, humidity, vibration, dust, etc.) of the region and the aging of the equipment is analyzed, specifically, by collecting historical data and real-time monitoring data, a mathematical model of the environmental factors and the aging of the equipment is established to predict the aging degree of the equipment, for example, in a high temperature and high humidity environment, the aging speed of electronic equipment may be accelerated, and by using this model, the aging conditions of equipment at different positions in the current environment are analyzed, and then according to the equipment aging analysis result, a second preference constraint is established, considering the influence of equipment aging on subway operation, paying attention to those equipment with high aging degree and possible failure, for example, for the equipment with serious aging degree, timely maintenance or replacement plan is arranged. Finally, based on the first preference constraint and the second preference constraint, a preference database is generated, which integrates the information of the location sensitivity of the fault occurrence and the aging of the equipment, and when making subway operation decision optimization, the information in the preference database can be used to reasonably arrange equipment maintenance plan and adjust resource allocation, so as to improve the safety and reliability of subway operation. Table 1 is an example data of the preference database:

[0035] Table 1: Example data of the preference database

[0036]

[0037] Step S300, using the preference database to strengthen the learning of the general weak classifier, and setting the reinforcement learning classifier and the general weak classifier as an edge decision classifier in the corresponding edge computing node, the reinforcement learning includes constructing a reinforcement learning environment using the preference database, loading the general weak classifier for dynamic reward-based learning iteration, and constructing a reinforcement learning classifier.

[0038] Preferably, the general weak classifier is reinforced learning using the preference database, wherein the reinforcement learning is a machine learning method, the agent (i.e. the general weak classifier) learns the optimal strategy by interacting with the environment according to the reward signal feedback from the environment to maximize the long-term cumulative reward, in the subway operation, the environment is the various running states and fault events of the subway, and the reward signal is related to the accuracy and timeliness of the fault decision; Specifically, the data in the preference database is taken as input, including fault event features in different positions, equipment aging features, etc. The general weak classifier classifies preliminarily according to these input features, and gives the general weak classifier corresponding rewards or punishments according to the comparison between the classification results and the actual situation, if the classification is accurate and can identify the fault event or accurately assess the equipment state in time, positive reward is given; otherwise, negative reward is given, for example, if the classifier correctly identifies the fault about to occur in the high sensitivity position, it will get higher positive reward; if the fault is misjudged or missed, it will be punished; then the general weak classifier adjusts its classification strategy according to the reward feedback, and gradually improves the accuracy and reliability of the classification through multiple iterations of learning, so that it can better adapt to the actual situation of subway operation; then the classifier after reinforcement learning (reinforcement learning classifier) is integrated with the original general weak classifier, such as using simple voting method, weighted average method, etc. to fuse the classification results of the reinforcement learning classifier and the general weak classifier, obtain the edge decision classifier, give full play to the advantages of the two, and improve the overall classification performance and decision-making ability; finally, the integrated edge decision classifier is deployed on the corresponding edge computing node, which is close to the data source (such as sensors, video monitoring, etc.), can obtain subway operation data in real time, and use the edge decision classifier to make rapid fault decision identification, so as to reduce data transmission delay, improve the timeliness and response speed of decision-making, and provide strong guarantee for the safety and efficiency of subway operation, for example, when the edge computing node receives the equipment state data from the sensor, the edge decision classifier can quickly judge whether the equipment fails, and timely issue a warning or take corresponding measures.

[0039] Further, step S300 further comprises step S310 of reading the preference database, step S320 of loading the general weak classifier, step S320 of using the reinforcement learning environment to perform parameter iteration of the general weak classifier, and step S320 of setting dynamic reward feedback in each round of parameter iteration; step S330 of outputting the optimized general weak classifier as a reinforcement learning classifier when the prediction result meets the expected threshold.

[0040] Preferably, various information related to the edge computing node in the preference database is read, including fault event data, device state data, location sensitivity information and device aging analysis results monitored by the edge computing node in the coverage area, etc. The data related to reinforcement learning is extracted from the preference database, and is sorted and matched, for example, the features of the fault event are associated with the corresponding location sensitivity, device aging degree, etc. to provide comprehensive input features for the general weak classifier. Through reinforcement data matching, the data in the preference database is converted into a format suitable for reinforcement learning algorithm processing; then based on the matched reinforcement data, an environment suitable for the general weak classifier to perform reinforcement learning is created, i.e. a reinforcement learning environment, which defines the action space (e.g. different classification decisions) that the agent (general weak classifier) can take, the state space of the environment (represented by the input reinforcement data) and the reward mechanism (used to feedback the good or bad of the classification decision), for example, the environment state can be the current monitored device state features and location information, the action of the agent is to classify and judge whether the device is faulty, and the reward is determined according to the accuracy and importance of the classification result.

[0041] Preferably, the pre-trained general weak classifier is loaded into the reinforcement learning environment, wherein the general weak classifier is a classification model with certain initial parameters, which can perform preliminary classification on the input data, but the classification performance may not be ideal; in the reinforcement learning environment, the general weak classifier makes classification decisions according to the current environment state (i.e. the input reinforcement data), and adjusts its parameters according to the reward signal feedback from the environment, through repeated iteration of this process, i.e. multiple iterations, to gradually optimize the parameters of the general weak classifier to improve its classification performance, for example, in each iteration, the parameters of the classifier are updated using gradient descent algorithm according to the difference between the classification result and the actual situation, so that the classifier can more accurately judge the fault event in subsequent classification; in each round of parameter iteration, the general weak classifier is given dynamic reward feedback according to the current classification result and environment state, and the reward is set according to the specific business requirements and targets, for example, if the classifier accurately identifies a fault event in a high sensitivity location, it is given a higher reward; if a normal device in a low risk location is misjudged as a fault event, it is given a certain punishment. The size and nature of the reward will dynamically change with different situations in order to guide the general weak classifier to learn the optimal classification strategy. As shown in Table 2, the reinforcement learning environment exemplary data is shown in Table 2:

[0042] Table 2 Reinforcement learning environment related data

[0043]

[0044] Preferably, during the parameter iteration process, the prediction results of the general weak classifier are constantly evaluated, and the evaluation indicators can be accuracy, recall rate, F1 value, etc. The current prediction results are compared with the pre-set expected threshold to determine whether the classifier has reached a satisfactory performance level. When the prediction results of the general weak classifier meet the expected threshold, it means that the classifier has been optimized enough through reinforcement learning and can perform well in the given task. The optimized general weak classifier is output as a reinforcement learning classifier for actual subway fault detection and equipment state monitoring tasks, etc. to ensure the accuracy and reliability of subway operation decision-making. As shown in Table 3, the following is an example of general weak classifier parameter iteration data:

[0045] Table 3: Example data for general weak classifier parameter iteration

[0046]

[0047] Further, step S320 further comprises step S321 of evaluating the current iteration stage of the general weak classifier and establishing a stage influence coefficient; step S322 of obtaining the prediction results of the general weak classifier in this round and identifying the prediction deviation of the prediction results based on the reinforcement learning environment to generate a first error influence coefficient; step S323 of establishing a linkage window and performing prediction backtracking analysis of the general weak classifier based on the linkage window to generate a second error influence coefficient according to the prediction backtracking analysis results and the parameter iteration direction; and step S324 of setting a dynamic reward feedback using the stage influence coefficient, the first error influence coefficient, and the second error influence coefficient.

[0048] Preferably, in the reinforcement learning process, different iteration stages have different significance for the performance improvement of the classifier. For example, in the early iteration stage, the classifier may be in a rapid learning and exploration stage, and has strong adaptability to new features and patterns, but the accuracy may be relatively low. In the later iteration stage, the classifier focuses more on fine-tuning the existing knowledge to improve the accuracy. By evaluating the current iteration stage, the progress of the classifier in the entire learning process can be understood, which may include evaluation based on the number of iterations, the trend of the loss function, the improvement speed of the classification accuracy, and the like. A stage influence coefficient is established to measure the relative importance of the current iteration stage to the final classifier performance. In each iteration, the general weak classifier makes a prediction result based on the input reinforcement data, i.e., classifies and judges the device state or fault event. Based on the known true labels or actual situation in the reinforcement learning environment, the prediction result can be compared with the true value to identify the prediction deviation. Then, by calculating the difference between the prediction result and the true value, a first error influence coefficient is generated, reflecting the influence of the accuracy of the current prediction result on the overall learning process. The greater the error, the greater the first error influence coefficient, indicating that the prediction result of this round has a more important role in improving the classifier, and stronger feedback is needed to adjust the parameters of the classifier.

[0049] Preferably, the linkage window is a time window for observing the prediction of the general weak classifier in a period of time or a plurality of consecutive iterations. By establishing a linkage window, the prediction result of the classifier can be analyzed retrospectively, i.e., reviewing the prediction results and their trends of the past several rounds. Then, by analyzing how the prediction result of the classifier changes under different parameter iteration directions, for example, when the parameters are adjusted in a certain direction, does the prediction result become more accurate or worse, the second error influence coefficient is generated by combining the prediction retrospective analysis result and the parameter iteration direction, which can more comprehensively reflect the learning dynamics and stability of the classifier. If the prediction result of the classifier gradually becomes better in the linkage window under a certain parameter iteration direction, the second error influence coefficient may be smaller, indicating that the current parameter adjustment direction is correct, otherwise it is larger.

[0050] Preferably, the comprehensive stage influence coefficient, the first error influence coefficient and the second error influence coefficient are used to set dynamic reward feedback for the general weak classifier to guide the classifier to learn in the correct direction, i.e. to improve the classification performance by adjusting the parameters. Specifically, according to different values of the three influence coefficients, a reward function is designed. For example, when the stage influence coefficient is large (in the key iteration stage), the first error influence coefficient is small (the current prediction is accurate), and the second error influence coefficient is also small (the parameter iteration direction is correct and the historical performance is stable), a higher reward is given to the classifier to encourage it to continue to maintain the current learning state and parameter adjustment direction. On the contrary, if a certain coefficient is large (such as the first error influence coefficient is large, indicating that the current prediction error is large), a lower reward or even a punishment is given to prompt the classifier to adjust the parameters to improve the prediction results. Through this dynamic reward feedback mechanism, the general weak classifier can effectively perform reinforcement learning and gradually optimize its performance to adapt to the requirements of subway fault detection and equipment state monitoring and other practical tasks.

[0051] Step S400, after receiving the communication data at the edge computing node, using the edge decision classifier to perform fault decision identification, and establishing an edge fault decision identification result. The fault decision identification includes respectively obtaining the fault decision identification results of the reinforcement learning classifier and the general weak classifier corresponding to the general weak classifier, and performing trust degree weighted calculation.

[0052] Preferably, the metro operation related data collected by monitoring devices such as sensors, video monitoring devices, and inspection drones, etc. are transmitted to the edge computing node through wired or wireless communication. The edge computing node serves as the center of data reception and preliminary processing, responsible for receiving communication data from different devices to ensure rapid monitoring and analysis of the metro operation status. Then, the communication data are input into the edge decision classifier, which extracts and analyzes the features of the data. For example, for device operation parameter data, the classifier determines whether the current parameter value is within the normal range, whether there is abnormal fluctuation or trend change. For video data, it analyzes whether there is abnormal behavior, equipment damage, etc. in the picture. Through comprehensive analysis of different types of data, the edge decision classifier can make decisions on whether there is a fault, the type of fault, and the severity of the fault, etc. to obtain the edge fault decision recognition result, which may include the type of fault (such as motor fault, signal fault, track fault, etc.), the location of the fault (such as a specific station, train carriage or track section), the severity of the fault (mild, moderate, severe), and possible causes of the fault, etc. On the one hand, it can be directly used for local fault handling and response, for example, the edge computing node can automatically trigger the corresponding alarm mechanism according to the severity of the fault to notify nearby maintenance personnel for inspection and maintenance. On the other hand, it can also be uploaded to the central management platform of the metro operation to optimize the operation strategy and ensure the safe and stable operation of the metro system.

[0053] Further, step S400 further comprises step S410 of performing data integrity check on the communication data to establish a data integrity check result; and step S420 of reporting data anomaly if the data integrity check result is a check failure result.

[0054] Preferably, the communication data is subjected to data integrity check, specifically, checking whether the data is lost, damaged or erroneous in the process of transmission, storage, etc., such as checking data integrity by cyclic redundancy check (CRC), parity check, hash check, etc. Taking the cyclic redundancy check as an example, the sender will calculate a CRC check code according to the data to be sent before sending the data, and send it together with the data. After receiving the data, the receiver (i.e. the edge computing node) will use the same algorithm to recalculate the CRC check code of the received data, and compare the calculation result with the received check code. In addition, it may also check whether the format of the data is correct, whether the fields of the data are complete, etc. For example, for the device status data sent by the sensor, it will check whether each data field (such as temperature, pressure, current, etc.) has a reasonable value, whether there is a missing or erroneous field.

[0055] Preferably, after completing the data integrity check of the communication data, a corresponding check result will be obtained. The check result is usually divided into two cases: check passed and check failed. If the data integrity check result is check failed, it means that the received communication data has a problem, which may be that part of the data is lost in the transmission process, or the data is tampered with or damaged, resulting in the inability to correctly reflect the actual running state of the subway system. The edge computing node will report data anomaly to avoid using incorrect or incomplete data for subsequent analysis and decision-making; for example, by recording abnormal information in the system log, including the time of abnormal occurrence, the data source involved (such as which sensor or device sends the data), the specific reason for the check failure, etc., and reminding the relevant operation and maintenance personnel to pay attention to the data anomaly situation, so that they can take timely measures, such as reacquiring data, checking communication lines or equipment, etc., to ensure that subsequent data can be accurately and completely received and processed.

[0056] Further, step S400 further comprises step S430, using the reinforcement learning classifier and the general weak classifier in the edge decision classifier to respectively perform fault decision identification based on the communication data, to establish a first fault decision identification result and a second fault decision identification result, the first fault decision identification result being the identification result of the reinforcement learning classifier, and the second fault decision identification result being the identification result of the general weak classifier; step S440, performing identification trust degree analysis on the first fault decision identification result and the reinforcement learning classifier to establish a first trust factor; step S450, performing identification trust degree analysis on the second fault decision identification result and the reinforcement learning classifier to establish a second trust factor; step S460, establishing an edge fault decision identification result based on the first fault decision identification result, the first trust factor, the second fault decision identification result and the second trust factor.

[0057] Preferably, the edge decision classifier comprises a reinforcement learning classifier and a general weak classifier. When the edge computing node receives the communication data, the two classifiers are used to process the data to identify whether there is a fault and the related situation of the fault. Specifically, the reinforcement learning classifier analyzes the communication data according to the learned strategy and pattern to obtain a decision recognition result about the fault, i.e., a first fault decision recognition result. The general weak classifier also analyzes the same communication data to obtain a second fault decision recognition result. Then, recognition trust degree analysis is performed to evaluate the reliability of the recognition results of the two classifiers. For the first fault decision recognition result, the trust degree of the result is determined by analyzing the performance of the reinforcement learning classifier in processing similar data in the past, the stability of its parameters, the degree of conformity with the actual situation, etc., and a first trust factor is used to represent the trust degree. If the reinforcement learning classifier has high accuracy and good stability on historical data, the first trust factor will be high, indicating that the recognition result of the reinforcement learning classifier is trusted. Similarly, for the second fault decision recognition result, the second trust factor is determined by analyzing various related factors of the general weak classifier when processing the communication data for fault recognition, such as its initial performance, adaptability in the current task, etc., to represent the trust degree of the recognition result of the general weak classifier. Finally, the recognition results of the two classifiers and the trust factors are considered comprehensively to determine the edge fault decision recognition result, i.e., the two results are weighted and fused according to the trust factors, etc. The reliability of each result is fully considered, and a more comprehensive, accurate and reliable edge fault decision recognition result is finally obtained, thereby ensuring the accuracy and reliability of the subway operation decision.

[0058] Further, step S460 further comprises step S461 of establishing a third fault decision recognition result by using the first trust factor to trust-weight the first fault decision recognition result; step S462 of establishing a fourth fault decision recognition result by using the second trust factor to trust-weight the second fault decision recognition result; and step S463 of performing trust conflict analysis on the third fault decision recognition result and the fourth fault decision recognition result to establish an edge fault decision recognition result.

[0059] Preferably, the first fault decision recognition result is trust-weighted by the first trust factor, that is, the first fault decision recognition result is adjusted according to the size of the first trust factor. Specifically, the first fault decision recognition result is multiplied by the first trust factor. If the first trust factor is close to 1, it means that the recognition result of the reinforcement learning classifier is very trusted, and the weighted result is not much different from the original first fault decision recognition result. If the first trust factor is small, such as 0.3, the weighted result will be greatly reduced on the basis of the original result. Then the third fault decision recognition result is established. Similarly, the second trust factor is used to trust-weight the recognition result of the general weak classifier (the second fault decision recognition result), that is, the second fault decision recognition result is multiplied by the second trust factor to obtain the fourth fault decision recognition result. Then, the third fault decision recognition result and the fourth fault decision recognition result are analyzed for trust conflict, that is, the difference between the two results is evaluated and processed, such as calculating the difference, relative error, and other evaluation conflicts between the two results. According to the degree and specific circumstances of the conflict, different processing strategies are adopted to determine the final edge fault decision recognition result, thereby ensuring the accuracy and reliability of the subway operation decision.

[0060] In step S500, the operation environment data and user feedback data of the subway are obtained, and the cross-border data is constructed. After synchronizing the cross-border data and the edge fault decision recognition result to the subway knowledge graph, the subway operation and maintenance decision optimization result is generated.

[0061] Preferably, the operation environment data of the subway is obtained, for example, through the deployment of temperature and humidity sensors, air quality monitoring equipment in subway stations and subway tunnels, which can collect real-time data such as environmental temperature and humidity, harmful gas concentration, etc.; with the help of track stress sensors and track geometry state monitors, the stress condition and track smoothness of the track can be obtained; using seismic monitoring equipment, the seismic activity information of the area along the subway can be monitored. User feedback data is collected from subway passengers, such as through social media platforms, setting up opinion feedback boxes in subway stations and carriages, developing feedback functions of mobile applications, and online surveys, etc. Passengers are encouraged to feedback their riding experience, which may include carriage crowdedness, air conditioning temperature comfort, train punctuality, station indication clarity, etc. Then the operation environment data and user feedback data are integrated and associated to obtain cross-border data, for example, the data of a station where users feedback that the carriages are hot (user feedback data) under high temperature weather (operation environment data) is associated to form a cross-border data record, thereby breaking the data field restrictions and mining the potential relationship between different types of data.

[0062] Preferably, the subway knowledge graph is a knowledge set stored in a graph structure, used to describe various entities (such as equipment, stations, lines, passengers, etc.) in subway operation and the relationships between them (such as the belonging relationship between equipment and stations, the connection relationship between lines and stations, the association relationship between passengers and riding behaviors, etc.), and then the cross-border data and edge fault decision recognition results are synchronized to the subway knowledge graph. For example, if the edge fault decision recognition result indicates that a certain equipment of a train has failed, the fault information is associated with the corresponding entity of the equipment in the knowledge graph, and the cross-border data at the time of the fault is added as additional information, so that the entity in the knowledge graph has more comprehensive and rich information, not only including its own inherent attributes, but also covering dynamic information related to the running environment, user feedback and fault conditions. Finally, the subway knowledge graph synchronized with cross-border data and edge fault decision recognition results is deeply mined and analyzed, for example, by analyzing the association relationship between equipment failure and user feedback in a specific running environment (such as high temperature and high humidity), it is found that a certain equipment fails frequently in a specific environment, and the user's complaint about the comfort of the car increases significantly when the fault occurs, and then a subway operation and maintenance decision optimization scheme is generated, for example, for the above problem, a decision is made to strengthen the inspection frequency and maintenance intensity of the related equipment in high temperature and high humidity weather; or adjust the running parameters of the car air conditioner to improve passenger comfort, thereby improving the accuracy of subway operation decision and the safety and reliability of subway operation.

[0063] Further, step S500 further includes step S510 of extracting voice data and complaint data of the user, taking the voice data and complaint data as user feedback data, and using natural language processing technology to extract key information of the user feedback data to construct user cross-border data; step S520 of reading time sequence running noise data of the subway, taking the time sequence running noise data as running environment data, and constructing environment cross-border data according to the running environment data; and step S530 of constructing cross-border data using the user cross-border data and the environment cross-border data.

[0064] Preferably, voice data and complaint data are extracted from the user's interaction with the subway, where the voice data can come from voice message devices installed in subway stations, call records of customer service centers, or voice feedback from passengers through mobile applications; complaint data includes complaint information submitted by passengers through various channels such as written complaints and online platform complaints; user feedback data is processed using natural language processing (NLP) technology, voice data is converted to text form using NLP for speech recognition, and then the text is segmented, tagged, and named entity recognized, etc. For example, from the passenger's complaint "Today I took the subway, the air conditioner at XX station is not cool at all, and it is hot and uncomfortable", the key information such as "XX station" (location entity) and "air conditioner not cool" (problem description) can be extracted through NLP technology; the extracted key information is integrated and structured to form user cross-border data.

[0065] Preferably, the time sequence running noise data of the subway is read as the running environment data, that is, the time sequence running noise data is collected in real time by noise sensors installed along the subway track, stations and other positions, reflecting the change of noise generated during the subway operation over time, and the environment cross-border data is constructed according to the collected time sequence running noise data. Specifically, the noise data is analyzed, such as calculating the average, maximum and minimum values of noise intensity at different times and different positions, and at the same time, combining other relevant information of the subway operation, such as train running speed, line type (underground, above ground), etc., the information is associated with the noise data to form environment cross-border data that can comprehensively describe the noise condition of the subway running environment. Finally, the user cross-border data and the environment cross-border data are integrated to form cross-border data, for example, it may be found that when the subway running noise exceeds a certain threshold, the complaints of users about the comfort of the car will increase, which helps to deeply understand the subway operation from multiple angles, so as to develop more reasonable operation strategies and optimization schemes, and improve the quality of subway operation.

[0066] Further, step S500 further comprises step S540 of establishing an emergency plan set; step S550 of performing plan matching of the emergency plan set based on the subway operation and maintenance decision optimization result to establish a plan matching result; and step S560 of performing emergency response management using the plan matching result.

[0067] Preferably, a set of emergency plans is formulated for various emergencies that may occur during subway operation, including but not limited to equipment failure, natural disasters, public health incidents, safety accidents, etc. For example, the emergency plan for train failure may include fault diagnosis process, emergency rescue measures, passenger evacuation scheme, etc. The emergency plan for fire may cover fire alarm mechanism, fire extinguishing measures, personnel escape route, etc. Each emergency plan specifies the specific actions, responsibility division and resource allocation under specific emergency conditions to ensure rapid and effective response to emergencies. The subway operation and maintenance decision optimization result is compared and analyzed with each plan in the emergency plan set, and the corresponding emergency plan is matched. For example, if the subway operation and maintenance decision optimization result shows that there is a potential failure risk of a key device, the emergency plan related to the device failure is found in the emergency plan set, i.e. the type, severity, and possible impact of the device failure are matched with the trigger conditions in the emergency plan to determine the most suitable plan, so that the correct emergency plan can be quickly started and effective measures can be taken to respond to emergencies.

[0068] Preferably, after finding the emergency plan suitable for the current situation, emergency response management is carried out according to the plan, including quickly organizing relevant personnel and resources to carry out emergency actions according to the procedures and responsibilities specified in the plan. For example, if the matched plan is the train failure emergency plan, the emergency response management may involve notifying maintenance personnel to rush to the scene for failure repair, organizing station staff to guide passengers to evacuate safely or take other trains, and informing passengers of the situation through broadcast, electronic display screen, etc. to calm passengers. During the emergency response process, the execution of the emergency action is monitored in real time, and the resources are adjusted and optimized according to the actual situation to ensure that the emergency action can proceed smoothly according to the plan, and finally the emergency situation is effectively controlled and properly handled to ensure the safety of subway operation.

[0069] Further, step S500 further comprises step S570 of identifying an abnormality in a subway knowledge graph based on the cross-border data and the edge fault decision identification result, and establishing an abnormality identification result; and step S580 of utilizing the subway knowledge graph to perform correlation identification on the abnormality identification result, reconstructing a fault decision based on the correlation identification result, and generating a subway operation and maintenance decision optimization result.

[0070] Preferably, the cross-boundary data and the edge fault decision recognition result are mapped into the subway knowledge graph, when an abnormal situation in the cross-boundary data (such as a large number of user complaints about elevator failure at a station for consecutive days, and the vibration parameter in the elevator operating environment data of the station exceeds the normal range) or the fault information in the edge fault decision recognition result is found, an abnormality is marked on the corresponding entity (such as the elevator entity of the station) in the knowledge graph, including adding an abnormal state attribute and recording detailed information related to the abnormality, such as abnormal occurrence time, abnormal description (specific description from cross-boundary data or fault recognition result), etc., and finally forming an abnormality identification result; then the subway knowledge graph uses semantic relationship to reason and identify other entities associated with the abnormality identification entity, for example, to find other signal machines on the train operating line that may be affected by the signal machine failure, the affected train numbers, etc., and then obtain the associated recognition result.

[0071] Preferably, finally, the fault decision is reconstructed based on the associated recognition result, that is, the original fault decision is adjusted, for example, originally only according to the edge fault decision recognition result, it is considered to be a single signal machine failure, but through the association recognition of the knowledge graph, it is found that the failure may have a chain effect on multiple train operations on multiple lines, so the fault decision cannot be limited to the maintenance of the signal machine, but also needs to consider how to adjust the train operation plan, and finally generate a subway operation and maintenance decision optimization scheme, which may include detailed equipment maintenance plan (maintenance personnel arrangement, maintenance tool preparation, maintenance time window determination), train operation adjustment scheme (adjust train number, change operating line, set temporary stop station) and passenger service optimization measures (increase broadcast notification, strengthen platform guidance, provide transfer suggestion), etc., which can more comprehensively and effectively cope with abnormal situations in subway operation, improve the efficiency and service quality of subway operation and maintenance, and ensure the safe and stable operation of subway operation.

[0072] The above detailed description does not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application. In some cases, the actions or steps described in the present application can be performed in a different order from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

Claims

1. A subway operation decision optimization method based on a knowledge graph and cross-modal association, characterized in that, The method comprises: Deploying an edge computing node at an edge device of a subway, and configuring a general weak classifier for the edge computing node, the edge computing node being in communication connection with sensors, video surveillance, and inspection drones; Performing computing preference analysis of the edge computing node to generate a preference database; Using the preference database to perform reinforcement learning on the general weak classifier, and setting a reinforcement learning classifier and the general weak classifier as an edge decision classifier at the corresponding edge computing node, the reinforcement learning comprising loading the general weak classifier to perform learning iteration based on dynamic rewards after constructing a reinforcement learning environment using the preference database, and constructing the reinforcement learning classifier; After receiving communication data at the edge computing node, using the edge decision classifier to perform fault decision identification, and establishing an edge fault decision identification result, the fault decision identification comprising respectively obtaining fault decision identification results of the reinforcement learning classifier and the general weak classifier, and performing trust degree weighted calculation; Obtaining operation environment data and user feedback data of the subway, constructing cross-border data, synchronizing the cross-border data and the edge fault decision identification result to a subway knowledge graph, and generating a subway operation and maintenance decision optimization result; Performing computing preference analysis of the edge computing node to generate a preference database, comprising: Obtaining a monitoring coverage area of the edge computing node, and configuring a position sensitivity of the monitoring coverage area; Reading fault events in the monitoring coverage area, and establishing a first preference constraint by weighting the occurrence probability of the fault events using the position sensitivity; Performing device aging analysis of the monitoring coverage area under regional environment fitting, and establishing a second preference constraint using the device aging analysis result; Generating the preference database based on the first preference constraint and the second preference constraint; Performing parameter iteration of the general weak classifier using the reinforcement learning environment, and setting a dynamic reward feedback at each round of parameter iteration, comprising: Evaluating the general weak classifier at a current iteration stage to establish a stage influence coefficient; Obtaining a prediction result of the general weak classifier at this round, and generating a first error influence coefficient based on prediction deviation identification of the prediction result using the reinforcement learning environment; Establishing a linkage window, performing prediction backtracking analysis of the general weak classifier based on the linkage window, and generating a second error influence coefficient according to the prediction backtracking analysis result and the parameter iteration direction; Setting a dynamic reward feedback using the stage influence coefficient, the first error influence coefficient, and the second error influence coefficient; Using the edge decision classifier to perform fault decision identification, and establishing an edge fault decision identification result, comprising: Using the reinforcement learning classifier and the general weak classifier in the edge decision classifier to respectively perform fault decision identification based on communication data, and establishing a first fault decision identification result and a second fault decision identification result, the first fault decision identification result being the identification result of the reinforcement learning classifier, and the second fault decision identification result being the identification result of the general weak classifier; Performing identification trust analysis on the first fault decision identification result and the reinforcement learning classifier to establish a first trust factor; Performing identification trust analysis on the second fault decision identification result and the reinforcement learning classifier to establish a second trust factor; Establishing an edge fault decision identification result based on the first fault decision identification result, the first trust factor, the second fault decision identification result, and the second trust factor, including: Establishing a third fault decision identification result by performing trust weighting on the first fault decision identification result using the first trust factor; Establishing a fourth fault decision identification result by performing trust weighting on the second fault decision identification result using the second trust factor; Performing trust conflict analysis on the third fault decision identification result and the fourth fault decision identification result to establish an edge fault decision identification result. The reinforcement learning of the general weak classifier using the preference database includes:

2. The subway operation decision optimization method based on a knowledge graph and cross-modal association according to claim 1, wherein, Reading the preference database, performing reinforcement data matching based on the preference database, and constructing a reinforcement learning environment; After loading the general weak classifier, performing parameter iteration of the general weak classifier using the reinforcement learning environment, and setting dynamic reward feedback at each round of parameter iteration; When the prediction result meets the expected threshold, output the optimized general weak classifier as the reinforcement learning classifier. The acquisition of the subway operation environment data and user feedback data, and the construction of cross-border data, includes:

3. The subway operation decision optimization method based on a knowledge graph and cross-modal association according to claim 1, wherein, Extracting user voice data and complaint data, taking the voice data and complaint data as user feedback data, and using natural language processing technology to extract key information from the user feedback data to construct user cross-border data; Reading the time-series operation noise data of the subway, taking the time-series operation noise data as the operation environment data, and constructing environment cross-border data based on the operation environment data; Constructing cross-border data from the user cross-border data and the environment cross-border data. After generating the subway operation decision optimization result, including:

4. The subway operation decision optimization method based on a knowledge graph and cross-modal association according to claim 1, wherein, Establishing an emergency plan set; Performing plan matching of the emergency plan set based on the subway operation decision optimization result to establish a plan matching result; Using the plan matching result for emergency response management. The synchronization of the cross-border data and the edge fault decision identification result to the subway knowledge graph includes:

5. The subway operation decision optimization method based on a knowledge graph and cross-modal association according to claim 1, wherein, Performing abnormality identification of the subway knowledge graph based on the cross-border data and the edge fault decision identification result to establish an abnormality identification result; Using the subway knowledge graph to perform associated identification of the abnormality identification result, reconstructing the fault decision based on the associated identification result, and generating a subway operation decision optimization result. After the edge computing node receives the communication data, including:

6. The subway operation decision optimization method based on a knowledge graph and cross-modal association according to claim 1, wherein, Performing data integrity verification on the communication data to establish a data integrity verification result; If the data integrity verification result is a verification failure result, report data anomaly. ​

Images