Intelligent monitoring and predicting system for rail transit equipment

Abstract

The invention discloses an intelligent monitoring and predicting system for rail transit equipment, which relates to the technical field of rail transit engineering, and comprises the following steps: acquiring a historical fault operation log of a rail system, learning a fault representation independent space, constructing a dynamic knowledge graph of historical faults of the rail, identifying a high-incidence area of the rail faults, and establishing a fault predicting model of the rail system. Based on a sensor, the coupling situation of the vehicle and the track is calculated in real time, a dynamic reference model of operation characteristics in a healthy state is constructed, and the real-time drift deviation between the vehicle and the track is calculated; establishing a drift deviation-environment correlation model, comparing the reference deviation with the actual drift deviation after environment compensation, and setting a dynamic safe operation boundary; and if the drift deviation exceeds a safe operation boundary, judging that the drift deviation is abnormal, immediately triggering graded alarm, associating and pairing a track system fault prediction model, quantifying the future fault occurrence probability and risk level of the region, and realizing intelligent monitoring and prediction of the track. The beneficial effect of the invention is that the fault occurrence rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of rail transit engineering technology, specifically to an intelligent monitoring and prediction system for rail transit equipment. Background Technology

[0002] Intelligent monitoring and prediction systems for rail transit equipment, both domestically and internationally, primarily rely on traditional sensors and rule-based fault diagnosis methods. These systems typically provide only single pieces of equipment status information, failing to effectively integrate data from multiple heterogeneous sensors and lacking dynamic monitoring of the complex coupling between the track system and the vehicle. Existing fault prediction technologies largely depend on static fault characterization models, failing to fully consider the impact of environmental factors on sensors and the dynamic changes in equipment drift deviations. This results in low accuracy of fault warnings, making it difficult to detect potential fault risks in a timely manner, thus affecting the safety and reliability of rail transit systems. Traditional methods have significant shortcomings in terms of accuracy, real-time performance, and adaptability in fault prediction, necessitating more intelligent and comprehensive monitoring and prediction technologies. Summary of the Invention

[0003] To address the aforementioned technical problems, an intelligent monitoring and prediction system for rail transit equipment is provided. This technical solution solves the problems of more intelligent and comprehensive monitoring and prediction technologies.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A method for intelligent monitoring and prediction of rail transit equipment, comprising:

[0006] Obtain historical fault operation logs of the track system, learn the independent space of fault representation, construct a dynamic knowledge graph of historical faults of the track system, identify high-incidence areas of track system faults, and establish a track system fault prediction model.

[0007] Based on multi-source heterogeneous sensors, the coupling status between the vehicle and the track is calculated in real time, a dynamic benchmark model of the operating characteristics of the vehicle and the track under healthy conditions is constructed, and the real-time drift deviation between the two is calculated.

[0008] To address the external effects of the environment on the sensors, a vehicle-track drift deviation-environment correlation model is established. By comparing the baseline deviation with the actual drift deviation after environmental compensation, a dynamic safe operating boundary is set. If the drift deviation is within the safe operating boundary, it is considered normal; if the drift deviation exceeds the safe operating boundary, it is considered abnormal, and a graded alarm is immediately triggered.

[0009] If the drift deviation exceeds the safe operating boundary, it is judged as abnormal, and a graded alarm is immediately triggered. The associated track system fault prediction model is then used to quantitatively output the probability and risk level of fault occurrence in a specific time window in the future, thereby realizing intelligent monitoring and prediction of the track system.

[0010] Preferably, based on the acquisition of historical fault operation logs of the track system, maintenance records, work orders, component replacement records and event reports of the track system are obtained, and data preprocessing is performed to generate a historical fault data table of the track system;

[0011] Using the historical fault data table of the orbital system as input, the SBERT model is trained, and the fixed-dimensional vector of the historical fault description text is used as output to obtain the text representation of the historical fault.

[0012] Create a unique integer ID for each equipment type in the track system to encode the equipment type data;

[0013] Based on the historical fault data table of the track system, if two historical fault events have the same fault code, it proves that they occurred on the same equipment and are judged as a similar pair of historical fault events;

[0014] Create an embedding layer for each device type, and input its unique integer ID into the corresponding embedding layer to obtain a low-dimensional dense vector for each device type;

[0015] By concatenating the historical fault text representation with the low-dimensional dense vector of each device type, a comprehensive fault event representation vector is obtained.

[0016] Based on the dual-tower encoder model, two fault events are input into the dual-tower encoder model respectively. A fully connected layer is designed to learn the feature interaction of the dual-tower encoder model. The cosine similarity of the comprehensive feature vectors of the two fault events is calculated. The minimum contrast loss is used as the fault event similarity loss function. The error between the prediction and the actual similarity of the dual-tower encoder model is calculated. The backpropagation method is introduced to optimize the dual-tower encoder model. The optimized dual-tower encoder model is obtained. It is input into the historical fault data table of the track system to generate the corresponding embedding vector and construct a fault event vector database.

[0017] A new fault event is randomly selected from the historical fault data table of the orbital system. The corresponding new fault event vector is calculated. The fault event vector database is associated and paired. The cosine similarity between the new fault event vector and all vectors in the fault event vector database is calculated. The vectors are sorted from high to low similarity to obtain the most similar fault event.

[0018] The dual-tower encoder model is evaluated based on the normalized loss cumulative gain.

[0019] Preferably, based on the acquisition of historical fault operation logs of the track system, the historical fault equipment, components, fault modes, locations, maintenance lists and time slices of the track system are obtained as entity types, the semantic relationships between entity types are defined, the semantic relationship between any two entity types is used as the edge type, the BERT context information of the fault event is obtained according to the dual-tower encoder model, and the average pooling operation is introduced to obtain the relationship between entity types.

[0020] Create an independent timestamp graph node for each fault event, connect fault events that occur in chronological order using PRECEDES relationships, capture the dynamic temporal relationships of fault events, and construct a dynamic knowledge graph of historical faults in the track system.

[0021] Based on the dynamic knowledge graph of historical faults in the track system, the number of entity connections is used as the degree centrality index, the number of times an entity appears on all shortest paths is used as the betweenness centrality index, and the reciprocal of the average distance from an entity to all other entities in the dynamic knowledge graph of historical faults in the track system is used as the proximity centrality index.

[0022] Using the PageRank algorithm, the entities in the dynamic knowledge graph of historical faults of the track system are initialized as web pages, the PageRank value of each entity is calculated, and the above process is repeated until the PageRank values ​​converge, thus obtaining the PageRank value ranking of the entities.

[0023] Using the Louvain community discovery algorithm, each entity in the dynamic knowledge graph of historical faults of the orbital system is initially regarded as an independent community, and local optimization of entities is performed.

[0024] Aggregate each entity to form a new graph structure, with each community as an entity, retain the edge weights between previous communities, continuously update and optimize the merged communities, and identify high-incidence areas of track system failures.

[0025] Using the Node2Vec algorithm, a random walk is used to sample the dynamic knowledge graph of historical faults in the track system, generating multiple entity node sequences. Based on the Word2Vec technique, the entity node embedding representation of fault events is learned, capturing the structural features and semantic information of entity nodes in the dynamic knowledge graph of historical faults in the track system. The learned entity node embedding is used as the feature representation of the dynamic knowledge graph of historical faults in the track system.

[0026] Based on the historical fault data table of the orbital system, it is divided into training set, validation set and test set. The feature representation of the dynamic knowledge graph of historical faults of the orbital system is used as input to train the GNN graph neural network model. The cross-validation strategy is introduced. Through the multi-layer propagation mechanism of the graph neural network, the information of each entity node and its neighboring entity nodes is aggregated. The training is continuously iterated, the model parameters are adjusted and the model performance is optimized, the dependency relationship of fault events in the graph structure is captured, and global and local fault event embedding representations are generated.

[0027] Based on global and local fault event embedding representations, and using time series analysis methods, combined with the historical fault occurrence time patterns of the track system, we predict the future occurrence trend of fault events and construct a fault prediction model for the track system.

[0028] Preferably, based on multi-source heterogeneous sensors, vibration, acoustic, temperature, stress and image data of vehicle and track are acquired, and Kalman filtering algorithm is used to preprocess them to filter out noise interference, extract effective feature signals, and generate time series of vehicle and track coupling status.

[0029] Based on the maintenance records of the historical fault operation log of the track system, the data of the stable operation period of the vehicle and track after maintenance is selected as the health operation data of the vehicle and track.

[0030] Based on vehicle and track health operation data, the data is divided into training set, validation set and test set.

[0031] Using vehicle and track health operation data within a time window as the input layer, an LSTM long short-term memory network is trained, a fully connected layer is designed, and the predicted sensor value at the current time step is used as the output layer.

[0032] Using the sliding window technique, vehicle and track health operation data are converted into health operation data sample pairs. The mean square error is used as the loss function, and early stop is set to prevent data overfitting.

[0033] Based on the validation set of vehicle and track health operation data, the prediction error of the health operation data at each time point is calculated, and the model parameter with the smallest prediction error is selected as the optimal parameter to generate a dynamic benchmark model of the operation characteristics of vehicle and track under health conditions.

[0034] Based on the time series of vehicle-track coupling status acquired in real time by multi-source heterogeneous sensors, and using the dynamic benchmark model of the vehicle and track operating characteristics under healthy conditions as a reference standard, the numerical difference between the two at each time point is calculated to obtain the real-time drift deviation between the vehicle and track.

[0035] Preferably, the loss function and the formula for the real-time drift deviation between the vehicle and the track are as follows:

[0036]

[0037] in, Let θ be the loss function, θ be the model parameters, N be the number of vehicle and track health operation data converted into health operation data sample pairs, and y be the loss function. i This represents the true sensor reading vector at the next time point after i samples within a time window, under healthy operating conditions. This is the predicted value of the sensor reading vector at the next time point after i samples in a time window, under healthy operating conditions.

[0038] δ t Let be the drift deviation vector at time t, where t is the current time point. This represents the actual state of the vehicle-track coupling system at the current moment, obtained after real-time acquisition by multi-source heterogeneous sensors and preprocessing with Kalman filtering. A dynamic benchmark model for the operational characteristics of vehicles and tracks under healthy conditions. A real-time historical data window for a dynamic benchmark model of the operational characteristics of vehicles and tracks under healthy conditions.

[0039] Preferably, a dynamic benchmark model based on the operating characteristics of the vehicle and track under healthy conditions is used to obtain vehicle and track drift deviation data;

[0040] Based on environmental sensors, data on vehicle and track temperature, humidity, and air pressure are acquired.

[0041] The vehicle and track drift deviation data were timestamped with temperature, humidity, and air pressure data, and outliers and missing values ​​were processed.

[0042] Based on the maintenance records of the historical fault operation log of the track system, we filter the known data and the data during the two fault maintenance cycles to obtain the drift deviation-environment dataset under the healthy state;

[0043] Based on the drift bias-environment dataset under healthy conditions, it is divided into training set, validation set and test set;

[0044] A training subset of the drift bias-environment dataset under healthy conditions is extracted using Bootstrap random sampling to train the decision tree model;

[0045] Using vehicle-track drift deviation data as feature variables and temperature, humidity, and air pressure data as environmental variables, a decision tree regression model is constructed to predict the expected drift deviation between the vehicle and the track under different environmental conditions.

[0046] The hyperparameters of the decision tree model are tuned using grid search and cross-validation. The model parameters with the smallest prediction error on the validation set are selected as the optimal parameters to generate a vehicle and track drift deviation-environment correlation model.

[0047] The hyperparameters include: the number of trees, the maximum depth of the trees, the minimum number of leaf node samples, and the minimum number of sample splits.

[0048] Preferably, based on the vehicle-track drift deviation-environment correlation model, the expected drift deviation between the vehicle and the track under the current environmental conditions is calculated, and compared with the actual vehicle-track drift deviation data to obtain the actual drift deviation after environmental compensation.

[0049] Based on the actual drift deviation after environmental compensation, a dynamic safe operating boundary is set.

[0050] The dynamic safety boundary formula is as follows:

[0051] B(t1)=μ e (t1)±k·σ e (t1)

[0052] Where B(t1) is the dynamic safety boundary at any time t1, μ e (t1) represents the expected drift deviation at any time t1 in the vehicle-track drift deviation-environment correlation model, k is the safety factor, and σ e (t1) represents the drift deviation at any time t1 in a healthy state - the standard deviation of the drift deviation of all historical data points in the environmental dataset that are similar to the current environmental conditions.

[0053] Based on dynamic safety boundaries, if the drift deviation is within the safe operating boundary, it is judged as normal; if the drift deviation exceeds the safe operating boundary, it is judged as abnormal and a graded alarm is immediately triggered.

[0054] The tiered early warning system includes:

[0055] If the drift deviation exceeds 1-2 times the safe operating boundary, a warning will be issued;

[0056] If the drift deviation exceeds 2-3 times the safe operating boundary, an alarm will be triggered;

[0057] If the drift deviation exceeds three times or more of the safe operating boundary, a serious alarm will be issued.

[0058] Preferably, if the drift deviation exceeds the safe operation boundary, it is judged as abnormal, and a graded alarm is immediately triggered. The paired track system fault prediction model is associated to predict the probability of fault occurrence in a specific time window in the future as the prior probability.

[0059] Based on real-time drift deviation data between vehicles and tracks, the Bayesian update algorithm is used to dynamically correct the prior probability in the track system fault prediction model. The abnormal data before the occurrence of historical faults is analyzed as the likelihood function to obtain the posterior probability as the quantitative output of the probability of fault occurrence in a specific time window in the future in this area.

[0060] Based on the posterior probability, the probability of failure occurring in the area is obtained, and different risk levels are divided, including low risk, medium risk and high risk, to achieve a quantitative assessment of the risk level.

[0061] Low risk: posterior probability < 0.3;

[0062] Medium risk: 0.3 ≤ posterior probability < 0.7;

[0063] High risk: posterior probability ≥ 0.7;

[0064] By combining the probability of failure and the risk level of failure in the area, a failure early warning report for the area is generated, realizing intelligent monitoring and prediction of the track system.

[0065] Furthermore, a smart monitoring and prediction system for rail transit equipment is provided to implement a smart monitoring and prediction method for rail transit equipment, comprising:

[0066] Track system fault prediction module, drift deviation module, safe operation boundary module, and intelligent monitoring and prediction module;

[0067] The track system fault prediction is used to obtain historical fault operation logs of the track system, learn the fault representation independent space, construct a dynamic knowledge graph of historical faults of the track system, identify high-incidence areas of track system faults, and establish a track system fault prediction model.

[0068] The drift deviation is used to calculate the coupling status between the vehicle and the track in real time based on multi-source heterogeneous sensors, construct a dynamic benchmark model of the operating characteristics of the vehicle and track under healthy conditions, and calculate the real-time drift deviation between the two.

[0069] The safe operation boundary module is electrically connected to the drift deviation module. It is used to establish a vehicle and track drift deviation-environment correlation model based on the external effects of the environment on the sensor. It compares the baseline deviation with the actual drift deviation after environmental compensation and sets a dynamic safe operation boundary. If the drift deviation is within the safe operation boundary, it is judged as normal. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal and a graded alarm is immediately triggered.

[0070] The intelligent monitoring and prediction module is electrically connected to the safe operation boundary module. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal, and a graded alarm is immediately triggered. It is associated with the track system fault prediction model and quantitatively outputs the probability and risk level of fault occurrence in a specific time window in the future, so as to realize the intelligent monitoring and prediction of the track system.

[0071] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0072] This invention proposes an intelligent monitoring and prediction system for rail transit equipment. By constructing a dynamic fault knowledge graph and health status model of the track system, it achieves real-time monitoring and fault prediction based on multi-source heterogeneous sensors. By calculating the coupling state and drift deviation between the vehicle and the track in real time and compensating for environmental factors, it can effectively identify and predict abnormal states of the track system. When the drift deviation exceeds the safe operating boundary, a graded alarm is immediately triggered, and the future fault risk in the area is quantitatively analyzed in conjunction with the track system fault prediction model, allowing for early prevention and maintenance. This intelligent monitoring and prediction method can improve the safety and reliability of rail transit equipment, reduce the failure rate, optimize operation and maintenance costs, and improve the overall system's operating efficiency and stability. Attached Figure Description

[0073] Figure 1 A flowchart of a method for intelligent monitoring and prediction of rail transit equipment;

[0074] Figure 2 This is a framework diagram of an intelligent monitoring and prediction system for rail transit equipment. Detailed Implementation

[0075] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0076] Reference Figure 1 As shown, a method for intelligent monitoring and prediction of rail transit equipment includes:

[0077] S1. Obtain historical fault operation logs of the track system, learn the fault representation independent space, construct a dynamic knowledge graph of historical faults of the track system, identify high-incidence areas of track system faults, and establish a track system fault prediction model.

[0078] Step S1 includes the following:

[0079] Based on the historical fault operation logs of the track system, maintenance records, work orders, component replacement records and event reports of the track system are obtained, and data preprocessing is performed to generate a historical fault data table of the track system.

[0080] Using the historical fault data table of the orbital system as input, the SBERT model is trained, and the fixed-dimensional vector of the historical fault description text is used as output to obtain the text representation of the historical fault.

[0081] Create a unique integer ID for each equipment type in the track system to encode the equipment type data;

[0082] Based on the historical fault data table of the track system, if two historical fault events have the same fault code, it proves that they occurred on the same equipment and are judged as a similar pair of historical fault events;

[0083] Create an embedding layer for each device type, and input its unique integer ID into the corresponding embedding layer to obtain a low-dimensional dense vector for each device type;

[0084] By concatenating the historical fault text representation with the low-dimensional dense vector of each device type, a comprehensive fault event representation vector is obtained.

[0085] Based on the dual-tower encoder model, two fault events are input into the dual-tower encoder model respectively. A fully connected layer is designed to learn the feature interaction of the dual-tower encoder model. The cosine similarity of the comprehensive feature vectors of the two fault events is calculated. The minimum contrast loss is used as the fault event similarity loss function. The error between the prediction and the actual similarity of the dual-tower encoder model is calculated. The backpropagation method is introduced to optimize the dual-tower encoder model. The optimized dual-tower encoder model is obtained. It is input into the historical fault data table of the track system to generate the corresponding embedding vector and construct a fault event vector database.

[0086] A new fault event is randomly selected from the historical fault data table of the orbital system. The corresponding new fault event vector is calculated. The fault event vector database is associated and paired. The cosine similarity between the new fault event vector and all vectors in the fault event vector database is calculated. The vectors are sorted from high to low similarity to obtain the most similar fault event.

[0087] The dual-tower encoder model is evaluated based on the normalized loss cumulative gain.

[0088] Step S1 also includes the following:

[0089] Based on the acquisition of historical fault operation logs of the track system, the historical fault equipment, components, fault modes, locations, maintenance lists and time slices of the track system are obtained as entity types. Semantic relationships between entity types are defined, and the semantic relationship between any two entity types is used as edge type. According to the dual-tower encoder model, the BERT context information of the fault event is obtained, and average pooling operation is introduced to obtain the relationship between entity types.

[0090] Create an independent timestamp graph node for each fault event, connect fault events that occur in chronological order using PRECEDES relationships, capture the dynamic temporal relationships of fault events, and construct a dynamic knowledge graph of historical faults in the track system.

[0091] Based on the dynamic knowledge graph of historical faults in the track system, the number of entity connections is used as the degree centrality index, the number of times an entity appears on all shortest paths is used as the betweenness centrality index, and the reciprocal of the average distance from an entity to all other entities in the dynamic knowledge graph of historical faults in the track system is used as the proximity centrality index.

[0092] Using the PageRank algorithm, the entities in the dynamic knowledge graph of historical faults of the track system are initialized as web pages, the PageRank value of each entity is calculated, and the above process is repeated until the PageRank values ​​converge, thus obtaining the PageRank value ranking of the entities.

[0093] Using the Louvain community discovery algorithm, each entity in the dynamic knowledge graph of historical faults of the orbital system is initially regarded as an independent community, and local optimization of entities is performed.

[0094] Aggregate each entity to form a new graph structure, with each community as an entity, retain the edge weights between previous communities, continuously update and optimize the merged communities, and identify high-incidence areas of track system failures.

[0095] Using the Node2Vec algorithm, a random walk is used to sample the dynamic knowledge graph of historical faults in the track system, generating multiple entity node sequences. Based on the Word2Vec technique, the entity node embedding representation of fault events is learned, capturing the structural features and semantic information of entity nodes in the dynamic knowledge graph of historical faults in the track system. The learned entity node embedding is used as the feature representation of the dynamic knowledge graph of historical faults in the track system.

[0096] Based on the historical fault data table of the orbital system, it is divided into training set, validation set and test set. The feature representation of the dynamic knowledge graph of historical faults of the orbital system is used as input to train the GNN graph neural network model. The cross-validation strategy is introduced. Through the multi-layer propagation mechanism of the graph neural network, the information of each entity node and its neighboring entity nodes is aggregated. The training is continuously iterated, the model parameters are adjusted and the model performance is optimized, the dependency relationship of fault events in the graph structure is captured, and global and local fault event embedding representations are generated.

[0097] Based on global and local fault event embedding representations, and using time series analysis methods, combined with the historical fault occurrence time patterns of the track system, we predict the future occurrence trend of fault events and construct a fault prediction model for the track system.

[0098] When using it, refer to the content of step S1 above:

[0099] In the current field of fault prediction for rail systems, existing technologies both domestically and internationally tend to focus on single-dimensional data analysis, relying on statistical fault frequency statistics or simple time-series model predictions. They lack deep integration of multi-source heterogeneous fault logs and the construction of dynamic knowledge graphs, failing to effectively capture semantic relationships between fault events, equipment topology relationships, and temporal evolution patterns. This results in incomplete fault representation and limited accuracy in identifying high-incidence areas. This step, by fusing textual representations and equipment embedding vectors to construct a dynamic knowledge graph, combined with graph neural networks and community detection algorithms, achieves multi-dimensional fault feature extraction, accurate high-incidence area location, and trend prediction, improving the accuracy of fault warnings and the efficiency of systemic operation and maintenance.

[0100] S2. Based on multi-source heterogeneous sensors, calculate the coupling status between the vehicle and the track in real time, construct a dynamic benchmark model of the operating characteristics of the vehicle and the track under healthy conditions, and calculate the real-time drift deviation between the two.

[0101] Step S2 includes the following:

[0102] Based on multi-source heterogeneous sensors, vibration, acoustic, temperature, stress and image data of vehicles and tracks are acquired. Kalman filtering algorithm is used to preprocess the data, filter out noise interference, extract effective feature signals, and generate time series of vehicle-track coupling status.

[0103] Based on the maintenance records of the historical fault operation log of the track system, the data of the stable operation period of the vehicle and track after maintenance is selected as the health operation data of the vehicle and track.

[0104] Based on vehicle and track health operation data, the data is divided into training set, validation set and test set.

[0105] Using vehicle and track health operation data within a time window as the input layer, an LSTM long short-term memory network is trained, a fully connected layer is designed, and the predicted sensor value at the current time step is used as the output layer.

[0106] Using the sliding window technique, vehicle and track health operation data are converted into health operation data sample pairs. The mean square error is used as the loss function, and early stop is set to prevent data overfitting.

[0107] Based on the validation set of vehicle and track health operation data, the prediction error of the health operation data at each time point is calculated, and the model parameter with the smallest prediction error is selected as the optimal parameter to generate a dynamic benchmark model of the operation characteristics of vehicle and track under health conditions.

[0108] Based on the time series of vehicle-track coupling status acquired in real time by multi-source heterogeneous sensors, and using a dynamic benchmark model of the vehicle and track's operational characteristics under healthy conditions as a reference standard, the numerical difference between the two at each time point is calculated to obtain the real-time drift deviation between the vehicle and track, as shown in the following formula:

[0109]

[0110] in, Let θ be the loss function, θ be the model parameters, N be the number of vehicle and track health operation data converted into health operation data sample pairs, and y be the loss function. i This represents the true sensor reading vector at the next time point after i samples within a time window, under healthy operating conditions. This is the predicted value of the sensor reading vector at the next time point after i samples in a time window, under healthy operating conditions.

[0111] δ t Let be the drift deviation vector at time t, where t is the current time point. This represents the actual state of the vehicle-track coupling system at the current moment, obtained after real-time acquisition by multi-source heterogeneous sensors and preprocessing with Kalman filtering. A dynamic benchmark model for the operational characteristics of vehicles and tracks under healthy conditions. A real-time historical data window for a dynamic benchmark model of the operational characteristics of vehicles and tracks under healthy conditions.

[0112] When using it, refer to the content of step S2 above:

[0113] Vehicle-track coupling situation analysis based on multi-source heterogeneous sensors often focuses on single-type data or offline diagnosis, lacking the ability to integrate multi-source data such as vibration, acoustics, temperature, stress, and images for real-time dynamic modeling. Health benchmark models often rely on static thresholds or fixed historical data, making it difficult to adapt to the gradual changes in the track system during long-term operation. This step constructs a dynamically updated health benchmark model by combining Kalman filtering and LSTM networks, realizing real-time drift deviation calculation of vehicle and track operating status, and improving the sensitivity and adaptability of early fault warning.

[0114] S3. To address the external effects of the environment on the sensors, establish a vehicle-track drift deviation-environment correlation model, compare the baseline deviation with the actual drift deviation after environmental compensation, and set dynamic safe operating boundaries. If the drift deviation is within the safe operating boundaries, it is considered normal; if the drift deviation exceeds the safe operating boundaries, it is considered abnormal, and a graded alarm is immediately triggered.

[0115] Step S3 includes the following:

[0116] Based on a dynamic benchmark model of the operational characteristics of vehicles and tracks under healthy conditions, drift deviation data of vehicles and tracks are obtained.

[0117] Based on environmental sensors, data on vehicle and track temperature, humidity, and air pressure are acquired.

[0118] The vehicle and track drift deviation data were timestamped with temperature, humidity, and air pressure data, and outliers and missing values ​​were processed.

[0119] Based on the maintenance records of the historical fault operation log of the track system, we filter the known data and the data during the two fault maintenance cycles to obtain the drift deviation-environment dataset under the healthy state;

[0120] Based on the drift bias-environment dataset under healthy conditions, it is divided into training set, validation set and test set;

[0121] A training subset of the drift bias-environment dataset under healthy conditions is extracted using Bootstrap random sampling to train the decision tree model;

[0122] Using vehicle-track drift deviation data as feature variables and temperature, humidity, and air pressure data as environmental variables, a decision tree regression model is constructed to predict the expected drift deviation between the vehicle and the track under different environmental conditions.

[0123] The hyperparameters of the decision tree model are tuned using grid search and cross-validation. The model parameters with the smallest prediction error on the validation set are selected as the optimal parameters to generate a vehicle and track drift deviation-environment correlation model.

[0124] The hyperparameters include: the number of trees, the maximum depth of the trees, the minimum number of leaf node samples, and the minimum number of sample splits.

[0125] Step S3 also includes the following:

[0126] Based on the vehicle-track drift deviation-environment correlation model, the expected drift deviation of the vehicle and track under the current environmental conditions is calculated, and compared with the actual vehicle-track drift deviation data, to obtain the actual drift deviation after environmental compensation.

[0127] Based on the actual drift deviation after environmental compensation, a dynamic safe operating boundary is set.

[0128] The dynamic safety boundary formula is as follows:

[0129] B(t1)=μ e (t1)±k·σ e (t1)

[0130] Where B(t1) is the dynamic safety boundary at any time t1, μ e(t1) represents the expected drift deviation at any time t1 in the vehicle-track drift deviation-environment correlation model, k is the safety factor, and σ e (t1) represents the drift deviation at any time t1 in a healthy state - the standard deviation of the drift deviation of all historical data points in the environmental dataset that are similar to the current environmental conditions.

[0131] Based on dynamic safety boundaries, if the drift deviation is within the safe operating boundary, it is judged as normal; if the drift deviation exceeds the safe operating boundary, it is judged as abnormal and a graded alarm is immediately triggered.

[0132] The tiered early warning system includes:

[0133] If the drift deviation exceeds 1-2 times the safe operating boundary, a warning will be issued;

[0134] If the drift deviation exceeds 2-3 times the safe operating boundary, an alarm will be triggered;

[0135] If the drift deviation exceeds three times or more of the safe operating boundary, a serious alarm will be issued.

[0136] When using it, refer to the content of step S3 above:

[0137] In the current field of rail transit condition monitoring, existing technologies mostly focus on static threshold alarms for single vehicle or track conditions, lacking intelligent methods that can deeply integrate environmental variables (such as temperature, humidity, and air pressure) for dynamic benchmark modeling and deviation compensation, resulting in high false alarm rates and delayed warnings. This step, by constructing a vehicle and track drift deviation-environment correlation model, achieves quantitative compensation for environmental impacts and precise setting of dynamic safety boundaries, improving the adaptability and accuracy of condition monitoring and reducing interference from non-fault factors.

[0138] S4. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal and a graded alarm is immediately triggered. The paired track system fault prediction model is associated and output in a quantitative manner the probability and risk level of fault occurrence in a specific time window in the future, so as to realize intelligent monitoring and prediction of the track system.

[0139] Step S4 includes the following:

[0140] If the drift deviation exceeds the safe operating boundary, it is judged as abnormal and a graded alarm is immediately triggered. The paired track system fault prediction model is associated to predict the probability of fault occurrence in a specific time window in the future as the prior probability.

[0141] Based on real-time drift deviation data between vehicles and tracks, the Bayesian update algorithm is used to dynamically correct the prior probability in the track system fault prediction model. The abnormal data before the occurrence of historical faults is analyzed as the likelihood function to obtain the posterior probability as the quantitative output of the probability of fault occurrence in a specific time window in the future in this area.

[0142] Based on the posterior probability, the probability of failure occurring in the area is obtained, and different risk levels are divided, including low risk, medium risk and high risk, to achieve a quantitative assessment of the risk level.

[0143] Low risk: posterior probability < 0.3;

[0144] Medium risk: 0.3 ≤ posterior probability < 0.7;

[0145] High risk: posterior probability ≥ 0.7;

[0146] By combining the probability of failure and the risk level of failure in the area, a failure early warning report for the area is generated, realizing intelligent monitoring and prediction of the track system.

[0147] When using it, refer to the content of step S4 above:

[0148] Domestic and international track system monitoring generally suffers from a disconnect between "monitoring" and "prediction." Alarms often rely on simple threshold-based judgments, lacking quantitative predictions of fault evolution trends after alarms. Risk level classification often depends on static experience and cannot be dynamically updated with system status. This step introduces a Bayesian update algorithm to deeply integrate real-time monitoring data with fault prediction models, achieving a leap from passive alarms to proactive prediction. It can immediately alarm when anomalies are detected and can dynamically correct and output precisely quantified fault probabilities within specific future time windows. This establishes risk level assessment on continuously updated objective data, providing crucial data support for accurate operation and maintenance decisions.

[0149] Reference Figure 2 As shown, an intelligent monitoring and prediction system for rail transit equipment includes:

[0150] Track system fault prediction module, drift deviation module, safe operation boundary module, and intelligent monitoring and prediction module;

[0151] The track system fault prediction is used to obtain historical fault operation logs of the track system, learn the fault representation independent space, construct a dynamic knowledge graph of historical faults of the track system, identify high-incidence areas of track system faults, and establish a track system fault prediction model.

[0152] The drift deviation is used to calculate the coupling status between the vehicle and the track in real time based on multi-source heterogeneous sensors, construct a dynamic benchmark model of the operating characteristics of the vehicle and track under healthy conditions, and calculate the real-time drift deviation between the two.

[0153] The safe operation boundary module is electrically connected to the drift deviation module. It is used to establish a vehicle and track drift deviation-environment correlation model based on the external effects of the environment on the sensor. It compares the baseline deviation with the actual drift deviation after environmental compensation and sets a dynamic safe operation boundary. If the drift deviation is within the safe operation boundary, it is judged as normal. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal and a graded alarm is immediately triggered.

[0154] The intelligent monitoring and prediction module is electrically connected to the safe operation boundary module. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal, and a graded alarm is immediately triggered. It is associated with the track system fault prediction model and quantitatively outputs the probability and risk level of fault occurrence in a specific time window in the future, so as to realize the intelligent monitoring and prediction of the track system.

[0155] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring and prediction of rail transit equipment, characterized in that, include: S1. Obtain historical fault operation logs of the track system, learn the fault representation independent space, construct a dynamic knowledge graph of historical faults of the track system, identify high-incidence areas of track system faults, and establish a track system fault prediction model. S2. Based on multi-source heterogeneous sensors, calculate the coupling status between the vehicle and the track in real time, construct a dynamic benchmark model of the operating characteristics of the vehicle and the track under healthy conditions, and calculate the real-time drift deviation between the two. S3. To address the external effects of the environment on the sensors, establish a vehicle-track drift deviation-environment correlation model, compare the baseline deviation with the actual drift deviation after environmental compensation, and set dynamic safe operating boundaries. If the drift deviation is within the safe operating boundaries, it is considered normal; if the drift deviation exceeds the safe operating boundaries, it is considered abnormal, and a graded alarm is immediately triggered. S4. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal, and a graded alarm is immediately triggered. The track system fault prediction model is associated with the model, and the probability and risk level of fault occurrence in a specific time window in the future are quantitatively output to realize intelligent monitoring and prediction of the track system.

2. The intelligent monitoring and prediction method for rail transit equipment according to claim 1, characterized in that, S1 includes: Based on the historical fault operation logs of the track system, maintenance records, work orders, component replacement records and event reports of the track system are obtained, and data preprocessing is performed to generate a historical fault data table of the track system. Using the historical fault data table of the orbital system as input, the SBERT model is trained, and the fixed-dimensional vector of the historical fault description text is used as output to obtain the text representation of the historical fault. Create a unique integer ID for each equipment type in the track system to encode the equipment type data; Based on the historical fault data table of the track system, if two historical fault events have the same fault code, it proves that they occurred on the same equipment and are judged as a similar pair of historical fault events; Create an embedding layer for each device type, and input its unique integer ID into the corresponding embedding layer to obtain a low-dimensional dense vector for each device type; By concatenating the historical fault text representation with the low-dimensional dense vector of each device type, a comprehensive fault event representation vector is obtained. Based on the dual-tower encoder model, two fault events are input into the dual-tower encoder model respectively. A fully connected layer is designed to learn the feature interaction of the dual-tower encoder model. The cosine similarity of the comprehensive feature vectors of the two fault events is calculated. The minimum contrast loss is used as the fault event similarity loss function. The error between the prediction and the actual similarity of the dual-tower encoder model is calculated. The backpropagation method is introduced to optimize the dual-tower encoder model. The optimized dual-tower encoder model is obtained. It is input into the historical fault data table of the track system to generate the corresponding embedding vector and construct a fault event vector database. A new fault event is randomly selected from the historical fault data table of the orbital system. The corresponding new fault event vector is calculated. The fault event vector database is associated and paired. The cosine similarity between the new fault event vector and all vectors in the fault event vector database is calculated. The vectors are sorted from high to low similarity to obtain the most similar fault event. The dual-tower encoder model is evaluated based on the normalized loss cumulative gain.

3. The intelligent monitoring and prediction method for rail transit equipment according to claim 2, characterized in that, S1 further includes: Based on the acquisition of historical fault operation logs of the track system, the historical fault equipment, components, fault modes, locations, maintenance lists and time slices of the track system are obtained as entity types. Semantic relationships between entity types are defined, and the semantic relationship between any two entity types is used as edge type. According to the dual-tower encoder model, the BERT context information of the fault event is obtained, and average pooling operation is introduced to obtain the relationship between entity types. Create an independent timestamp graph node for each fault event, connect fault events that occur in chronological order using PRECEDES relationships, capture the dynamic temporal relationships of fault events, and construct a dynamic knowledge graph of historical faults in the track system. Based on the dynamic knowledge graph of historical faults in the track system, the number of entity connections is used as the degree centrality index, the number of times an entity appears on all shortest paths is used as the betweenness centrality index, and the reciprocal of the average distance from an entity to all other entities in the dynamic knowledge graph of historical faults in the track system is used as the proximity centrality index. Using the PageRank algorithm, the entities in the dynamic knowledge graph of historical faults of the track system are initialized as web pages, the PageRank value of each entity is calculated, and the above process is repeated until the PageRank values ​​converge, thus obtaining the PageRank value ranking of the entities. Using the Louvain community discovery algorithm, each entity in the dynamic knowledge graph of historical faults of the orbital system is initially regarded as an independent community, and local optimization of entities is performed. Aggregate each entity to form a new graph structure, with each community as an entity, retain the edge weights between previous communities, continuously update and optimize the merged communities, and identify high-incidence areas of track system failures. Using the Node2Vec algorithm, a random walk is used to sample the dynamic knowledge graph of historical faults in the track system, generating multiple entity node sequences. Based on the Word2Vec technique, the entity node embedding representation of fault events is learned, capturing the structural features and semantic information of entity nodes in the dynamic knowledge graph of historical faults in the track system. The learned entity node embedding is used as the feature representation of the dynamic knowledge graph of historical faults in the track system. Based on the historical fault data table of the orbital system, it is divided into training set, validation set and test set. The feature representation of the dynamic knowledge graph of historical faults of the orbital system is used as input to train the GNN graph neural network model. The cross-validation strategy is introduced. Through the multi-layer propagation mechanism of the graph neural network, the information of each entity node and its neighboring entity nodes is aggregated. The training is continuously iterated, the model parameters are adjusted and the model performance is optimized, the dependency relationship of fault events in the graph structure is captured, and global and local fault event embedding representations are generated. Based on global and local fault event embedding representations, and using time series analysis methods, combined with the historical fault occurrence time patterns of the track system, we predict the future occurrence trend of fault events and construct a fault prediction model for the track system.

4. The intelligent monitoring and prediction method for rail transit equipment according to claim 1, characterized in that, The 2 includes: Based on multi-source heterogeneous sensors, vibration, acoustic, temperature, stress and image data of vehicles and tracks are acquired. Kalman filtering algorithm is used to preprocess the data, filter out noise interference, extract effective feature signals, and generate time series of vehicle-track coupling status. Based on the maintenance records of the historical fault operation log of the track system, the data of the stable operation period of the vehicle and track after maintenance is selected as the health operation data of the vehicle and track. Based on vehicle and track health operation data, the data is divided into training set, validation set and test set. Using vehicle and track health operation data within a time window as the input layer, an LSTM long short-term memory network is trained, a fully connected layer is designed, and the predicted sensor value at the current time step is used as the output layer. Using the sliding window technique, vehicle and track health operation data are converted into health operation data sample pairs. The mean squared error is used as the loss function, and early stop is set to prevent data overfitting. Based on the validation set of vehicle and track health operation data, the prediction error of the health operation data at each time point is calculated, and the model parameter with the smallest prediction error is selected as the optimal parameter to generate a dynamic benchmark model of the operation characteristics of vehicle and track under health conditions. Based on the time series of vehicle-track coupling status acquired in real time by multi-source heterogeneous sensors, and using the dynamic benchmark model of the vehicle and track operating characteristics under healthy conditions as a reference standard, the numerical difference between the two at each time point is calculated to obtain the real-time drift deviation between the vehicle and track.

5. The intelligent monitoring and prediction method for rail transit equipment according to claim 4, characterized in that, The loss function and the formula for the real-time drift deviation between the vehicle and the track are as follows: in, Let θ be the loss function, θ be the model parameters, N be the number of vehicle and track health operation data converted into health operation data sample pairs, and y be the loss function. i This represents the true sensor reading vector at the next time point after i samples within a time window, under healthy operating conditions. This is the predicted value of the sensor reading vector at the next time point after i samples in a time window, under healthy operating conditions. δ t Let be the drift deviation vector at time t, where t is the current time point. This represents the actual state of the vehicle-track coupling system at the current moment, obtained after real-time acquisition by multi-source heterogeneous sensors and preprocessing with Kalman filtering. A dynamic benchmark model for the operational characteristics of vehicles and tracks under healthy conditions. A real-time historical data window for a dynamic benchmark model of the operational characteristics of vehicles and tracks under healthy conditions.

6. The intelligent monitoring and prediction method for rail transit equipment according to claim 4, characterized in that, S3 includes: Based on a dynamic benchmark model of the operating characteristics of vehicles and tracks under healthy conditions, vehicle and track drift deviation data are obtained. Based on environmental sensors, data on vehicle and track temperature, humidity, and air pressure are acquired. The vehicle and track drift deviation data were timestamped with temperature, humidity, and air pressure data, and outliers and missing values ​​were processed. Based on the maintenance records of the historical fault operation log of the track system, we filter the known data and the data during the two fault maintenance cycles to obtain the drift deviation-environment dataset under the healthy state; Based on the drift bias-environment dataset under healthy conditions, it is divided into training set, validation set and test set; A training subset of the drift bias-environment dataset under healthy conditions is extracted using Bootstrap random sampling to train the decision tree model; Using vehicle-track drift deviation data as feature variables and temperature, humidity, and air pressure data as environmental variables, a decision tree regression model is constructed to predict the expected drift deviation between the vehicle and the track under different environmental conditions. The hyperparameters of the decision tree model are tuned using grid search and cross-validation. The model parameters with the smallest prediction error on the validation set are selected as the optimal parameters to generate a vehicle and track drift deviation-environment correlation model. The hyperparameters include: the number of trees, the maximum depth of the trees, the minimum number of leaf node samples, and the minimum number of sample splits.

7. The intelligent monitoring and prediction method for rail transit equipment according to claim 6, characterized in that, S3 further includes: Based on the vehicle-track drift deviation-environment correlation model, the expected drift deviation of the vehicle and track under the current environmental conditions is calculated, and compared with the actual vehicle-track drift deviation data, to obtain the actual drift deviation after environmental compensation. Based on the actual drift deviation after environmental compensation, a dynamic safe operating boundary is set. The dynamic safety boundary formula is as follows: B(t1)=μ e (t1)±k·σ e (t1) Where B(t1) is the dynamic safety boundary at any time t1, μ e (t1) represents the expected drift deviation at any time t1 in the vehicle-track drift deviation-environment correlation model, k is the safety factor, and σ e (t1) represents the drift deviation at any time t1 in a healthy state - the standard deviation of the drift deviation of all historical data points in the environmental dataset that are similar to the current environmental conditions. Based on dynamic safety boundaries, if the drift deviation is within the safe operating boundary, it is judged as normal; if the drift deviation exceeds the safe operating boundary, it is judged as abnormal and a graded alarm is immediately triggered. The tiered early warning system includes: If the drift deviation exceeds 1-2 times the safe operating boundary, a warning will be issued; If the drift deviation exceeds 2-3 times the safe operating boundary, an alarm will be triggered; If the drift deviation exceeds three times or more of the safe operating boundary, a serious alarm will be issued.

8. The intelligent monitoring and prediction method for rail transit equipment according to claim 7, characterized in that, S4 includes: If the drift deviation exceeds the safe operating boundary, it is judged as abnormal and a graded alarm is immediately triggered. The paired track system fault prediction model is associated to predict the probability of fault occurrence in a specific time window in the future as the prior probability. Based on real-time drift deviation data between vehicles and tracks, the Bayesian update algorithm is used to dynamically correct the prior probability in the track system fault prediction model. The abnormal data before the occurrence of historical faults is analyzed as the likelihood function to obtain the posterior probability as the quantitative output of the probability of fault occurrence in a specific time window in the future in this area. Based on the posterior probability, the probability of failure occurring in the area is obtained, and different risk levels are divided, including low risk, medium risk and high risk, to achieve a quantitative assessment of the risk level. Low risk: posterior probability < 0.3; Medium risk: 0.3 ≤ posterior probability < 0.7; High risk: posterior probability ≥ 0.7; By combining the probability of failure and the risk level of failure in the area, a failure early warning report for the area is generated, realizing intelligent monitoring and prediction of the track system.

9. A smart monitoring and prediction system for rail transit equipment, characterized in that, The method for intelligent monitoring and prediction of rail transit equipment according to any one of claims 1-8 includes: Track system fault prediction module, drift deviation module, safe operation boundary module, and intelligent monitoring and prediction module; The track system fault prediction is used to obtain historical fault operation logs of the track system, learn the fault representation independent space, construct a dynamic knowledge graph of historical faults of the track system, identify high-incidence areas of track system faults, and establish a track system fault prediction model. The drift deviation is used to calculate the coupling status between the vehicle and the track in real time based on multi-source heterogeneous sensors, construct a dynamic benchmark model of the operating characteristics of the vehicle and track under healthy conditions, and calculate the real-time drift deviation between the two. The safe operation boundary module is electrically connected to the drift deviation module. It is used to establish a vehicle and track drift deviation-environment correlation model based on the external effects of the environment on the sensor. It compares the baseline deviation with the actual drift deviation after environmental compensation and sets a dynamic safe operation boundary. If the drift deviation is within the safe operation boundary, it is judged as normal. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal and a graded alarm is immediately triggered. The intelligent monitoring and prediction module is electrically connected to the safe operation boundary module. If the drift deviation exceeds the safe operation boundary, it is judged as abnormal, and a graded alarm is immediately triggered. It is associated with the track system fault prediction model and quantitatively outputs the probability and risk level of fault occurrence in a specific time window in the future, so as to realize the intelligent monitoring and prediction of the track system.

Images