Rail transit equipment fault prediction method and system

Abstract

The invention provides a rail transit equipment fault prediction method and system, and the system comprises a data perception and fusion layer which is used for collecting and fusing multi-source heterogeneous data; the digital twinning and intelligent prediction layer is used for realizing accurate prediction and early fault early warning of the health state of the equipment by constructing a lightweight digital twinning model of the equipment and combining a space-time diagram attention network and a self-adaptive learning algorithm; the interpretable analysis and decision support layer is used for analyzing the root cause of the prediction result and providing visual analysis; and the maintenance linkage and execution layer is used for automatically generating a maintenance strategy and scheduling resources according to a prediction result, relates to the technical field of rail transit operation and maintenance, and overcomes the defects of single data, static model, weak early warning capability, poor interpretability and disjunction between prediction and maintenance in the prior art. And the full-life-cycle, self-adaptive and intelligent management of the fault prediction of the rail transit equipment is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit operation and maintenance technology, and in particular to a rail transit equipment fault prediction method and system. BACKGROUND

[0002] As the backbone of urban public transportation, the safety and reliability of rail transit is of great importance. Fault prediction and health management technology is the key to ensuring operational safety and reducing maintenance costs. By monitoring the state of key equipment and predicting faults, the transition from planned repair to state repair can be achieved, major safety accidents can be avoided, and operational efficiency can be significantly improved.

[0003] However, existing rail transit equipment fault prediction techniques still have many defects, mainly in the following aspects:

[0004] Insufficient data utilization: Existing methods rely heavily on single vibration signals or current signals, and fail to effectively integrate multi-source heterogeneous information such as operating conditions, environmental data, and historical maintenance records, resulting in insufficient data value mining and poor model generalization ability.

[0005] Weak model adaptability: Equipment performance degrades over time, and traditional data-driven models are difficult to adapt to this time-varying degradation process, resulting in a decline in prediction accuracy over time and an inability to effectively track the state over a long period of time.

[0006] Insufficient sensitivity of early fault detection: For weak early fault features, existing models lack sufficient detection sensitivity, are prone to false positives or false negatives, and cannot provide effective early warnings at the early stages of failure.

[0007] Poor interpretability of decision support: Most deep learning-based prediction models are like a black box, with their prediction results lacking reasonable explanations, making it difficult for maintenance personnel to understand the basis for the model's decisions and locate the root cause of the fault, thereby reducing the trustworthiness and effectiveness of the decision.

[0008] Disconnection between prediction and maintenance: Prediction results are often presented in the form of simple alarms or life values, and are not effectively integrated with the back-end maintenance management system, making it difficult to directly translate prediction information into efficient maintenance actions and creating an information silo.

[0009] In summary, developing a fault prediction method and system that can fully utilize multi-source data, has adaptive learning capabilities, provides interpretable decision support, and can be integrated with maintenance execution in a closed loop, has become a technical problem that needs to be solved in the field. SUMMARY

[0010] The present application aims to overcome the deficiencies of the prior art, and provides a rail transit equipment fault prediction method and system based on digital twinning and adaptive learning, so as to solve the problems of insufficient data utilization, weak model adaptive ability, poor interpretability, and disconnection between prediction and maintenance in the prior art.

[0011] The technical solution adopted by the embodiments of the present application to solve the technical problems is:

[0012] The rail transit equipment fault prediction method and system comprises a data perception and fusion layer, a digital twinning and intelligent prediction layer, an interpretable analysis and decision support layer, and a maintenance linkage and execution layer.

[0013] The data perception and fusion layer is used to collect real-time sensor data, running state data, environmental data, and historical maintenance data of the rail transit equipment, and to clean, align, and fuse the data to form a unified feature data set.

[0014] The digital twinning and intelligent prediction layer is connected with the data perception and fusion layer and comprises a digital twinning modeling module and an adaptive prediction module.

[0015] The digital twinning modeling module is used to establish a lightweight digital twinning model based on a physical information neural network for the rail transit equipment to output a baseline state prediction.

[0016] The adaptive prediction module is used to utilize the unified feature data set and construct a health state prediction model based on a spatio-temporal graph attention network to obtain a preliminary health state and a remaining useful life prediction.

[0017] The interpretable analysis and decision support layer is connected with the digital twinning and intelligent prediction layer and is used to perform interpretability analysis on the output of the health state prediction model, locate the fault root cause and propagation path, and generate a health state evaluation report.

[0018] The maintenance linkage and execution layer is connected with the interpretable analysis and decision support layer and is used to automatically generate a maintenance work order and a resource scheduling scheme according to the health state evaluation report and the predicted remaining useful life in combination with resource inventory information.

[0019] Preferably, the lightweight digital twinning model based on the physical information neural network established by the digital twinning modeling module is formed by embedding device physical equations as constraint conditions into neural network training, and is used to simulate the baseline running state of the device under ideal working conditions.

[0020] Preferably, in the adaptive prediction module, the health state prediction model adopts a spatio-temporal graph attention network.

[0021] The graph structure takes device components as nodes and physical connections or signal coupling relationships as edges, and the network captures the spatiotemporal evolution characteristics of the device state through a time attention mechanism and a graph attention mechanism.

[0022] Preferably, the adaptive prediction module dynamically adjusts the parameters of the health state prediction model using the output residual error of the digital twin model, specifically:

[0023] The residual error is used as part of the loss function or as a supplement to the model input features to update the weights of the health state prediction model online.

[0024] Preferably, the explainable analysis and decision support layer uses explainable artificial intelligence technology based on Shapley value and attention weights to perform attribution analysis on the prediction results of the health state prediction model and visually display the failure features and their contribution degrees.

[0025] Preferably, the adaptive prediction module also uses the output residual error of the digital twin model as feedback to dynamically adjust the parameters of the health state prediction model.

[0026] The method comprises the following steps:

[0027] S1: Collect and fuse multi-source heterogeneous data through the data perception and fusion layer to form a unified feature dataset;

[0028] S2: Run the equipment digital twin model through the digital twin modeling module of the digital twin and intelligent prediction layer to obtain a baseline state prediction, wherein the equipment digital twin model is a lightweight model based on a physical information neural network, which is trained by embedding the device physical equation as a constraint condition into the neural network;

[0029] S3: Input the unified feature dataset into the health state prediction model through the adaptive prediction module of the digital twin and intelligent prediction layer to obtain preliminary health state and remaining useful life prediction, and use the residual error between the baseline state prediction and real-time data to adaptively correct the health state prediction model, wherein the health state prediction model is constructed based on a spatiotemporal graph attention network and uses residual error to adaptively correct the model, specifically including:

[0030] The residual error is included in the backpropagation process of model training or is spliced with the original data as input features to realize online updating of model parameters;

[0031] S4: Perform root cause analysis on the prediction results through the explainable analysis and decision support layer to generate a health state evaluation report with explainability, wherein the root cause analysis uses explainable artificial intelligence technology to locate the key components and features that cause health state degradation by calculating the gradient or attention weight of the model output on the input features.

[0032] S5: According to the health state evaluation report and the remaining useful life, the maintenance linkage and the execution layer automatically trigger the maintenance process, generate and issue maintenance work orders and resource scheduling instructions.

[0033] The advantages of the embodiments of the present application are:

[0034] 1. Through the data perception and fusion layer, sensor data, working condition data, environmental data and historical maintenance data are integrated, overcoming the limitations of a single data source, and providing a comprehensive data foundation for accurate prediction.

[0035] 2. The lightweight digital twin model based on PINN is innovatively introduced, which contains both the physical laws of device operation and the learning ability of neural networks, providing a reliable benchmark state reference.

[0036] 3. By taking the output residual error of the digital twin model as feedback, the health state prediction model is dynamically corrected, enabling the model to adapt to long-term degradation of device performance and ensuring long-term stability of prediction accuracy.

[0037] 4. The health state prediction model can effectively capture weak anomalies and propagation paths of device state in the time and space dimensions, and combined with XAI technology, can clearly locate the root cause of failure, improving the transparency and credibility of decision-making.

[0038] 5. The system directly converts the prediction results into executable maintenance work orders and resource scheduling schemes, realizing seamless connection from prediction to execution, forming a complete operation and maintenance closed loop, and significantly improving operation and maintenance efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The architecture diagram of the rail transit equipment fault prediction system provided by the embodiments of the present application is shown in the figure;

[0040] Figure 2 The flowchart of the rail transit equipment fault prediction method provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. In addition, for the convenience of description, the "up", "down", "left", "right" and the like in the drawings are consistent with the "up", "down", "left", "right" and the like in the drawings, and "first", "second" and the like in the following text are for description and differentiation, and have no other special meanings.

[0042] The embodiment of the application provides a rail transit equipment fault prediction method and system, and solves the problems in the prior art.

[0043] The technical scheme in the embodiment of the application is used for solving the above problems, and the general idea is as follows:

[0044] Embodiment one

[0045] The embodiment provides a rail transit equipment fault prediction method and system, as shown in the figure, including a data perception and fusion layer, a digital twin and intelligent prediction layer, an interpretable analysis and decision support layer, and a maintenance linkage and execution layer. Figure 1 The data perception and fusion layer is used for collecting real-time sensor data, running state data, environment data and historical maintenance data of the rail transit equipment, and performing cleaning, alignment and fusion on the data to form a unified feature data set.

[0046] The data perception and fusion layer is used for collecting real-time sensor data, running state data, environment data and historical maintenance data of the rail transit equipment, and performing cleaning, alignment and fusion on the data to form a unified feature data set.

[0047] Further, the data perception and fusion layer is deployed on an edge computing gateway or a central server, and multi-source data is collected through various sensors such as vibration, temperature, acoustic sensors, train networks and management information systems. The layer uses data cleaning technology to process abnormal values, and uses a timestamp-based data alignment method to fuse data of different frequencies and sources onto a unified timeline, and finally forms a unified feature data set through feature engineering.

[0048] The digital twin and intelligent prediction layer is connected with the data perception and fusion layer, and includes a digital twin modeling module and an adaptive prediction module.

[0049] The digital twin modeling module is used for establishing a lightweight digital twin model based on a physical information neural network for the rail transit equipment to output a baseline state prediction.

[0050] The adaptive prediction module is used for utilizing the unified feature data set and constructing a health state prediction model based on a spatio-temporal graph attention network to obtain a preliminary health state and a remaining useful life prediction. The adaptive prediction module also uses the output residual error of the digital twin model as feedback to dynamically adjust the parameters of the health state prediction model.

[0051] In the adaptive prediction module, the health state prediction model adopts a spatio-temporal graph attention network.

[0052] Among them, the graph structure takes the device components as nodes and the physical connection or signal coupling relationship as edges, and the network captures the spatio-temporal evolution characteristics of the device state through time attention mechanism and graph attention mechanism.

[0053] The adaptive prediction module dynamically adjusts the parameters of the health state prediction model using the output residual error of the digital twin model, specifically:

[0054] The residual error is used as part of the loss function or as a supplement to the model input features to update the weights of the health state prediction model online.

[0055] Further, the digital twin modeling module establishes a PINN model for key equipment such as bogie bearings. During training, the loss function not only includes data fitting terms, but also adds physical constraint terms composed of device dynamics equations, so that it can simulate the baseline vibration or temperature response of the device under normal and ideal operating conditions. The adaptive prediction module constructs a health state prediction model that treats each measurement point of the device as a graph node and the physical connection between components as an edge to build a graph structure. It captures the dependence of the historical state through a time attention mechanism and captures the spatial correlation between different measurement points through a graph attention mechanism. The input of this module is a unified feature dataset, and the output is a health index and RUL. The residual error between real-time monitoring data and the baseline prediction value output by the PINN digital twin model is used as an additional supervision signal. This residual error reflects abnormalities or degradation that cannot be explained by the physical model. By feeding this residual error back to the health state prediction model and fine-tuning its network parameters through online learning, the model can continuously track the actual degradation trajectory of the device and achieve adaptive prediction.

[0056] The explainable analysis and decision support layer is connected to the digital twin and intelligent prediction layer, which is used for explainable analysis of the output of the health state prediction model, locating the root cause and propagation path of the fault, and generating a health state evaluation report.

[0057] Further, the explainable analysis and decision support layer uses attention weight visualization methods to analyze which time step and sensor node in the health state prediction model contributes most to the final fault prediction. The system can generate a heat map highlighting the key vibration measurement points that cause the health state of the bearing to decline and the time period when they appear abnormal. At the same time, combined with knowledge graph technology, it can infer the potential propagation path of the fault and generate a health state evaluation report containing quantitative indicators, root cause positioning, and maintenance recommendations.

[0058] The maintenance linkage and execution layer is connected to the explainable analysis and decision support layer, which is used to automatically generate maintenance work orders and resource scheduling plans based on the health state evaluation report and the predicted remaining useful life, combined with resource inventory information.

[0059] Further, the maintenance linkage and execution layer integrates with the existing asset management system, work order system and inventory management system of the enterprise through an API interface. When the received health status evaluation report indicates that the equipment health is below a threshold or the RUL is less than a preset value, the layer automatically creates a maintenance work order in the EAM system, and the work order content includes the faulty component and the recommended measures. At the same time, the system queries the inventory database, and if the spare parts are sufficient, the inventory is automatically locked; if not, a procurement warning is triggered. Finally, the maintenance instructions and resource scheduling scheme are issued to the corresponding maintenance team.

[0060] The lightweight digital twin model based on the physical information neural network established by the digital twin modeling module is trained by embedding the physical equation of the equipment as a constraint condition into the neural network, and is used to simulate the baseline operating state of the equipment under ideal working conditions.

[0061] The explainable analysis and decision support layer uses the explainable artificial intelligence technology based on Shapley value and explanation or attention weight to perform attribution analysis on the prediction results of the health status prediction model, and displays the failure features and their contribution degrees in a visual manner.

[0062] Embodiment Two

[0063] The embodiment provides a rail transit equipment fault prediction method and system. As shown in Figure 2 The embodiment provides a rail transit equipment fault prediction method and system. As shown in

[0064] S1: Collect and fuse multi-source heterogeneous data through the data perception and fusion layer to form a unified feature data set;

[0065] Further, real-time collection of vibration, speed, load, environmental temperature and humidity and historical maintenance records is performed, data cleaning, alignment and extraction of effective features are performed, and a unified feature data set is formed.

[0066] S2: Run the equipment digital twin model through the digital twin modeling module of the digital twin and intelligent prediction layer to obtain baseline state prediction, wherein the equipment digital twin model is a lightweight model based on a physical information neural network, which is trained by embedding the physical equation of the equipment as a constraint condition into the neural network;

[0067] Further, the PINN-based bearing digital twin model is run, the current working condition is input, and the ideal vibration response under the working condition is output as the baseline state prediction.

[0068] S3: input the unified feature dataset into the health state prediction model through the adaptive prediction module of the intelligent prediction layer by digital twinning, obtain the preliminary health state and remaining useful life prediction, and use the residual error of the benchmark state prediction and real-time data to adaptively correct the health state prediction model, wherein the health state prediction model is constructed based on a spatiotemporal graph attention network, and the model is adaptively corrected by using the residual error, specifically including:

[0069] The residual error is included in the back propagation process of model training, or is spliced as an input feature with original data to realize online updating of model parameters;

[0070] Further, the unified feature dataset obtained in S1 is input into the health state prediction model to obtain the preliminary HI and RUL, and at the same time, the residual error of real-time vibration data and the benchmark prediction in S2 is calculated, and the residual error is used to fine-tune the last one or several layers of parameters of the health state prediction model through an online gradient descent algorithm, so as to realize adaptive correction of the model.

[0071] S4: perform root cause analysis on the prediction results through the explainable analysis and decision support layer to generate a health state evaluation report with explainability, wherein the root cause analysis adopts explainable artificial intelligence technology to locate the key components and features causing health state degradation by calculating the gradient or attention weight of the input features of the model output;

[0072] Further, the prediction results of the corrected health state prediction model are analyzed for explainability, and it is located that the radial vibration sensor of a certain bearing appears abnormal at a specific frequency band, and it is judged as an early peeling fault, and a health state evaluation report containing this conclusion is generated.

[0073] S5: automatically trigger the maintenance process according to the health state evaluation report and the remaining useful life through the maintenance linkage and execution layer, generate and issue a maintenance work order and resource scheduling instruction.

[0074] Further, the system automatically generates a work order for planned bearing replacement in the work order system according to the report and the RUL prediction value, and queries the inventory system to reserve the corresponding type of bearing spare parts for the work order, and the maintenance personnel receive the work order and perform maintenance.

[0075] Finally, it should be noted that: obviously, the above embodiments are only examples for clearly illustrating the present application, and are not limitations on the embodiments. For ordinary skilled persons in the art, other different forms of changes or variations can be made on the basis of the above description. Here, it is not necessary and impossible to exhaust all the embodiments. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A rail transit equipment failure prediction system, characterized by, The data perception and fusion layer is used for collecting real-time sensor data, running state data, environment data and historical maintenance data of the rail transit equipment, and cleaning, aligning and fusing the data to form a unified feature data set. The digital twin and intelligent prediction layer is connected with the data perception and fusion layer and includes a digital twin modeling module and an adaptive prediction module. The digital twin modeling module is used for establishing a lightweight digital twin model based on a physical information neural network for the rail transit equipment to output a baseline state prediction. The adaptive prediction module is used for utilizing the unified feature data set and constructing a health state prediction model based on a spatio-temporal graph attention network to obtain a preliminary health state and a remaining useful life prediction. The explainable analysis and decision support layer is connected with the digital twin and intelligent prediction layer and is used for performing explainable analysis on the output of the health state prediction model, locating fault root causes and propagation paths, and generating a health state evaluation report. The maintenance linkage and execution layer is connected with the explainable analysis and decision support layer and is used for automatically generating a maintenance work order and a resource scheduling scheme according to the health state evaluation report and the predicted remaining useful life in combination with resource inventory information. The lightweight digital twin model based on the physical information neural network established by the digital twin modeling module is formed by embedding device physical equations as constraint conditions into neural network training and is used for simulating the baseline running state of the device under ideal working conditions.

2. The rail transit vehicle failure prediction system of claim 1, wherein, In the adaptive prediction module, the health state prediction model adopts a spatio-temporal graph attention network.

3. The rail transit vehicle failure prediction system of claim 1, wherein, The graph structure takes device components as nodes and physical connections or signal coupling relationships as edges, and the network jointly captures the spatio-temporal evolution characteristics of the device state through time attention mechanisms and graph attention mechanisms. The adaptive prediction module dynamically adjusts the parameters of the health state prediction model by using the output residual error of the digital twin model, specifically as follows:

4. The rail transportation equipment failure prediction system of claim 1, wherein, The residual error is used as part of a loss function or as a supplement to model input features to update the weights of the health state prediction model online. The explainable analysis and decision support layer adopts explainable artificial intelligence technology based on Shapley additive explanations or attention weights to perform attribution analysis on the prediction results of the health state prediction model and visually displays fault features and their contribution degrees.

5. The rail transportation equipment failure prediction system of claim 1, wherein, The adaptive prediction module also uses the output residual error of the digital twin model as feedback to dynamically adjust the parameters of the health state prediction model.

6. The rail transit vehicle failure prediction system of claim 1, wherein, The method includes the following steps:

7. The method for predicting the failure of the rail transit equipment, applied to the rail transit equipment failure prediction system according to any one of claims 1-6, characterized in that, S1: Collect and fuse multi-source heterogeneous data through the data perception and fusion layer to form a unified feature data set; S2: Run the device digital twin model through the digital twin modeling module of the digital twin and intelligent prediction layer to obtain a baseline state prediction; ​ S3: input the unified feature dataset into a health state prediction model through a self-adaptive prediction module of the intelligent prediction layer of the digital twin, obtain a preliminary health state and remaining useful life prediction, and perform self-adaptive correction on the health state prediction model by using a residual error of a benchmark state prediction and real-time data; S4: perform root cause analysis on the prediction result through the explainable analysis and decision support layer to generate a health state evaluation report with explainability; S5: automatically trigger a maintenance process, generate and issue a maintenance work order and resource scheduling instruction according to the health state evaluation report and the remaining useful life through the maintenance linkage and execution layer.

8. The rail transit equipment failure prediction method according to claim 7, characterized by, The device digital twin model in the step S2 is a lightweight model based on a physical information neural network, which is formed by embedding a device physical equation as a constraint condition into neural network training.

9. The rail transit equipment failure prediction method according to claim 7, characterized by, The health state prediction model in the step S3 is constructed based on a spatio-temporal graph attention network, and the model is adaptively corrected by using a residual error, specifically including: The residual error is included in a back propagation process of model training, or is spliced with original data as input features to realize online updating of model parameters.

10. The rail transit equipment failure prediction method according to claim 7, characterized by, The root cause analysis in the step S4 adopts an explainable artificial intelligence technology, and a gradient or attention weight of model output to input features is calculated to locate key components and features causing health state degradation.

Images