Intelligent rail vehicle predictive maintenance system and method based on multi-source sensor fusion

Abstract

The application provides a smart rail vehicle predictive maintenance system and method based on multi-source sensing fusion, relates to the smart rail traffic operation and maintenance technical field, and solves the problems of low fault early warning precision, poor strategy dynamics and low resource scheduling efficiency in the existing smart rail vehicle predictive maintenance technology. In the system, multi-source sensing data of smart rail vehicle exclusive components are synchronously acquired, noise suppression, abnormal correction and feature extraction processing are performed on the multi-source sensing data, weight distribution and feature fusion are performed based on an attention mechanism, and a fusion feature vector is obtained; after being modeled into a digital twin model, the operation state and degradation trend of the vehicle components are dynamically mapped, thereby the health state quantitative evaluation and residual service life prediction of the smart rail vehicle exclusive components are performed, a dynamic maintenance plan is obtained, the configuration of maintenance resources is guided, and maintenance operations are performed. The application reduces maintenance cost and operation risk, and fills the gap of smart rail exclusive predictive maintenance technology.

Description

Technical Field

[0001] This invention relates to the field of intelligent rail transit operation and maintenance technology, specifically to a predictive maintenance system and method for intelligent rail vehicles based on multi-source sensor fusion. Background Technology

[0002] With the large-scale deployment and routine operation of intelligent rail transit systems in many cities, the reliability of key vehicle components has become a core element in ensuring public transportation safety and improving service quality. Relying on virtual track guidance technology and rubber-wheel load-bearing structures, intelligent rail trains' core components, such as guide wheel assemblies and articulated devices, bear complex dynamic loads during operation, resulting in a failure evolution mechanism significantly different from traditional steel-wheel and steel-rail rail transit vehicles. This structural uniqueness makes it difficult for existing maintenance systems to be precisely adapted, creating an urgent need for differentiated and refined predictive maintenance technologies.

[0003] Current fault prediction technologies often overemphasize single sensor data sources. Some methods collect vibration signals for feature encoding and neural network modeling, which, while possessing some recognition capabilities in specific scenarios, fail to effectively integrate multi-dimensional heterogeneous information such as visual monitoring images, operating condition parameters, and environmental variables. Since intelligent rail transit component faults often exhibit multi-physics coupling characteristics, a single data dimension cannot comprehensively reflect early degradation signs, especially in complex urban road conditions where interference is prevalent. This results in insufficient robustness of prediction models and limitations in the timeliness and accuracy of early warnings.

[0004] Most current technical solutions still follow the framework logic for general rail transit equipment. Although such systems introduce intelligent algorithms to optimize maintenance plans, they lack modeling of the failure mechanisms of intelligent rail-specific components and fail to build targeted analysis models that incorporate operational characteristics such as virtual guidance mechanisms and rubber wheel wear characteristics. Maintenance recommendations are weakly correlated with the actual condition of the vehicle, making it difficult to intervene in components and resulting in problems of strategy generalization and insufficient applicability.

[0005] Existing systems mostly employ preset fixed-cycle maintenance tasks, failing to establish a correlation between vehicle lifecycle data and dynamic factors such as real-time operational intensity, road conditions, and climate. The lack of continuous data interaction and state mapping between virtual models and physical entities hinders adaptive adjustments in maintenance decisions, easily leading to premature maintenance and resource waste, or delayed maintenance creating safety hazards, thus restricting the overall improvement of operational efficiency and resource utilization.

[0006] In summary, existing technologies have systemic limitations in terms of the breadth of sensing dimensions, the depth of component adaptation, prediction accuracy, and the dynamism of maintenance, making it difficult to support the comprehensive requirements of intelligent rail transit operation for accurate fault identification, real-time response to maintenance needs, and scientific resource scheduling. The industry urgently needs to break through the constraints of traditional technological paths and explore a technological direction that can deeply integrate multi-source sensing information, fit the characteristics of intelligent rail transit operation, and achieve dynamic optimization of maintenance strategies, in order to drive the intelligent rail transit operation and maintenance system towards continuous evolution towards high reliability, high efficiency, and high adaptability. Summary of the Invention

[0007] The purpose of this invention is to address the problems in existing predictive maintenance technologies for intelligent rail transit (IRT) vehicles, such as low accuracy in fault prediction, lack of dynamic response capabilities in maintenance strategies, and inefficient resource scheduling, caused by the single dimension of sensing data and insufficient adaptation to fault mechanisms of specific components. Therefore, this invention proposes an IRT vehicle predictive maintenance system and method based on multi-source sensor fusion. This invention achieves health status assessment of key IRT vehicle components, early fault warning, dynamic optimization of maintenance plans, and efficient resource scheduling, thereby reducing maintenance costs and operational risks and filling the gap in IRT-specific predictive maintenance technology.

[0008] The present invention employs the following technical solutions to achieve its objective: A predictive maintenance system for intelligent rail transit vehicles based on multi-source sensor fusion, comprising the following functional modules: The multi-source data acquisition module is configured to synchronously acquire multi-source sensing data of the intelligent rail vehicle's dedicated components, including operating parameter data, multimodal perception data, and cross-system operation data. The operating parameter data includes physical state parameters of at least one type of dedicated component, such as the guide wheel assembly and the articulation device. The multimodal perception data includes visual images and environmental monitoring information. The cross-system operation data is provided by the train control and management system and the automatic train monitoring system. The data fusion processing module is configured to perform noise suppression, anomaly correction and feature extraction processing on the multi-source sensor data, and to perform weight allocation and feature fusion on heterogeneous data sources based on the attention mechanism, and output a fused feature vector representing the real-time state of the vehicle. The digital twin modeling module is configured to construct a digital twin model corresponding to the physical vehicle structure and operating logic, and dynamically map the operating status and degradation trend of vehicle components based on the fused feature vector. The fault prediction and health assessment module is configured to combine the digital twin modeling module and the fused feature vector to perform a quantitative assessment of the health status and predict the remaining service life of the intelligent rail vehicle's dedicated components, and generate a health assessment result that includes fault risk level and warning threshold. The maintenance decision optimization module is configured to generate a dynamic maintenance plan that matches the actual condition of the vehicle based on the health assessment results, preset component association rules, and maintenance constraints. The resource scheduling and execution module is configured to parse the dynamic maintenance plan and output maintenance task instructions, resource allocation schemes and execution timing information to the maintenance terminal device.

[0009] Preferably, the multi-source data acquisition module includes an on-board sensing unit, a multimodal sensing unit, and a cross-system data interface unit; The on-board sensing unit is integrated into at least one type of dedicated component in the intelligent rail vehicle's guide wheel assembly, articulation device, and bogie, and is equipped with at least one of vibration sensors, displacement sensors, angle sensors, tension sensors, temperature sensors, and pressure sensors. The multimodal sensing unit includes industrial cameras, environmental parameter sensors, and virtual wheel-rail contact monitoring sensors distributed at the front, rear, and sides of the vehicle chassis. The cross-system data interface unit uses RESTful API, MQTT or OPC UA communication protocols to establish data connections with the train control management system, the automatic train monitoring system and the energy consumption management system to obtain at least one of real-time vehicle speed, cumulative mileage, passenger capacity, energy consumption data and historical maintenance records.

[0010] Preferably, in the vehicle-mounted sensing unit, the sensors deployed on the guide wheel assembly include a triaxial vibration sensor and a displacement sensor, used to collect three-dimensional vibration acceleration signals and wheel rim wear displacement data, respectively; The sensors deployed on the articulation device include angle sensors and tension sensors, which are used to monitor the change in articulation angle and the tension value of the connecting structure, respectively. Sensors deployed on the bogie include temperature sensors and pressure sensors, used to acquire bearing temperature and suspension system pressure parameters, respectively.

[0011] Preferably, in the multimodal sensing unit, the industrial camera is equipped with a supplementary lighting device and transmits image data via Ethernet; the environmental parameter sensor integrates a temperature and humidity detection element and a dust concentration detection element; the virtual wheel-rail contact monitoring sensor is an ultrasonic flaw detection sensor, used to collect stress distribution signals and wheel surface defect feature data in the virtual wheel-rail contact area.

[0012] Furthermore, the data fusion processing module includes an edge preprocessing unit and a cloud fusion unit; The edge preprocessing unit is used to perform adaptive Kalman filtering to suppress high-frequency noise, and combines anomaly data identification and interpolation correction based on statistical criteria and isolated forest algorithm with continuous projection algorithm to perform feature screening and dimensionality reduction on the original parameters. The cloud fusion unit concatenates the dimensionality-reduced temporal sensing features, visual image features extracted by the convolutional neural network, and structured operational data features into vectors, and inputs them into a hybrid neural network model containing a scaled dot product attention layer, a convolutional layer, and a long short-term memory network layer to calculate the dynamic weights of each data source and output the fused feature vector.

[0013] Preferably, the digital twin model constructed by the digital twin modeling module includes a three-dimensional geometric structure model of the intelligent rail vehicle's exclusive components, a physical degradation behavior model, and an operational logic mapping relationship; The fault prediction and health assessment module, based on the real-time state mapping results of the digital twin model and the fused feature vector, calls a deep learning prediction model to perform time-series extrapolation of the remaining service life of the component, and outputs a quantitative health index and graded early warning signal.

[0014] Preferably, the maintenance decision optimization module has a built-in component association rule library and maintenance constraint condition library. The component association rule library stores fault propagation logic and collaborative maintenance rules between different dedicated components, and the maintenance constraint condition library stores information on the availability status of maintenance resources, operating time windows, and safety specification requirements. The output information generated by the resource scheduling and execution module includes maintenance task sequence, list of required tools and equipment, technical personnel configuration instructions and task execution time nodes, and is sent to maintenance terminal equipment through communication link.

[0015] This invention also provides a predictive maintenance method for intelligent rail transit vehicles based on the aforementioned system, comprising the following steps: S1. The physical state parameters of at least one type of dedicated component in the intelligent rail vehicle, such as the guide wheel assembly, articulation device, and bogie, are collected synchronously through a multi-source sensor network. The chassis visual image and environmental monitoring data are acquired synchronously, and the operation scheduling data and historical maintenance records are extracted from the train control management system and the automatic train monitoring system through a standardized communication interface. S2. The collected multi-source sensor data are sequentially subjected to adaptive filtering and noise reduction, abnormal data identification and correction based on statistical criteria and isolated forest algorithm, feature screening and dimensionality reduction processing; the processed time-series sensor features, visual image features and structured operation data features are vectorized and input into a neural network model containing attention mechanism for dynamic weight allocation and feature fusion to generate a fused feature vector representing the real-time state of the vehicle. S3. Call the digital twin model corresponding to the physical vehicle structure and operating logic, input the fused feature vector into the digital twin model, and complete the real-time mapping of the operating status of the vehicle-specific components and the deduction of degradation trends. S4. Based on the mapping result of the digital twin model and the fused feature vector, a time-series prediction model is used to perform a quantitative assessment of the health status and a projection of the remaining service life of the dedicated component, and outputs a health assessment result including a fault risk level identifier and a warning threshold. S5. Based on the health assessment results, and combined with the preset component association rule library and maintenance constraint library, generate a dynamic maintenance plan that includes maintenance items, execution priorities and suggested time windows; S6. Parse the dynamic maintenance plan, generate maintenance task sequence, required tool and equipment list, technician configuration instructions and task execution time nodes, and transmit the maintenance information to the maintenance terminal equipment through the communication link.

[0016] Preferably, in step S2, the adaptive filtering and noise reduction uses an adaptive Kalman filter algorithm to process multi-source sensor data; the abnormal data identification and correction combines the three-standard-deviation statistical criterion and the isolated forest algorithm. After configuring a preset number of isolated trees and a preset number of samples, the data points determined to be abnormal are corrected using a combination of linear interpolation and spline interpolation; the feature selection and dimensionality reduction process uses a continuous projection algorithm to select a subset of features with a cumulative contribution rate of not less than a preset percentage. The feature fusion is as follows: the dimensionality-reduced temporal sensing features, visual image features extracted by the convolutional neural network, and structured operational data features are concatenated into an input vector, which is then input into a hybrid neural network model containing a scaled dot product attention layer, a convolutional layer, and a long short-term memory network layer. The convolutional layer contains two convolutional operations with a kernel size of 3×3. The scaled dot product attention layer calculates the weight coefficients of each data source, and the long short-term memory network layer extracts temporal dependent features and outputs a fused feature vector.

[0017] Specifically, in step S5, the component association rule library stores wear association rules between the guide wheel assembly and the virtual wheel-rail contact monitoring sensor, and mechanical maintenance rules between the articulation device and the bogie; the maintenance constraint condition library stores maintenance personnel qualification and scheduling data, tool and equipment inventory and calibration status information, available time windows for outages of operating lines, and safe operating procedures. When generating a dynamic maintenance plan, the system searches the component association rule library based on the fault risk level identifier in the health assessment results to determine the set of components that need to be maintained in conjunction with the maintenance plan. Combining the real-time resource status and time-series constraints in the maintenance constraint library, the system prioritizes maintenance projects and allocates time windows to generate a dynamic maintenance plan that includes a list of specific maintenance projects, execution priority sequence, suggested start and end times, and associated component identifiers.

[0018] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention constructs a multi-dimensional data acquisition system covering vehicle-specific components, visual perception, and the operating environment. It integrates key physical quantities such as guide wheel vibration and wear, and mechanical parameters of articulated devices with images and operating condition information, effectively overcoming the perception blind spots and interference limitations of single data sources, and providing a comprehensive and reliable data foundation for condition identification. A customized monitoring strategy designed for the guidance of the intelligent rail virtual track and the load-bearing structure of the rubber wheels ensures that the data acquisition and analysis process aligns with the failure mechanisms of specific components such as guide wheel assemblies and articulated devices, significantly enhancing the targeting and discrimination capabilities of fault feature extraction.

[0019] This invention employs a collaborative mechanism of real-time edge noise filtering and cloud-based intelligent fusion during data processing, combined with attention weight allocation and deep feature extraction techniques, to achieve high-quality integration and enhancement of key information from multi-source heterogeneous data. Its dynamic mapping relationship between the digital twin model and physical entities supports the visualization of vehicle component operating status and continuous deduction of degradation trajectories, providing a high-confidence model basis for health assessment.

[0020] In this invention, the maintenance decision-making process integrates quantitative health indices, component association rule bases, and real-time resource constraints. The generated dynamic maintenance plan can adaptively adjust maintenance items, priorities, and timing according to the actual vehicle status, avoiding resource mismatch caused by fixed-cycle maintenance. The resource scheduling module decomposes the optimized maintenance instructions into task sequences, resource configurations, and time nodes, and sends them to the execution terminal in real time, strengthening the closed-loop management and collaborative efficiency of the maintenance process. Overall, this invention promotes a systematic upgrade of the intelligent rail vehicle operation and maintenance system from passive response to proactive prediction, and from experience-driven to data-intelligent, providing reliable technical support for operational safety and service quality. Attached Figure Description

[0021] The present invention is described in detail with reference to the following figures, which include six figures as follows: Figure 1 This is a schematic diagram of the composition structure of the predictive maintenance system for intelligent rail transit vehicles of the present invention; Figure 2 This is a schematic diagram of the sensor deployment relationship in the vehicle-mounted sensing unit of the present invention; Figure 3 This is a functional flowchart of the data fusion processing module in the system of the present invention; Figure 4 This is a schematic diagram of the architecture of the hybrid neural network model in the system of the present invention; Figure 5 This is a schematic diagram illustrating the module functions from fault prediction and health assessment to resource optimization execution in this invention. Figure 6 This is a schematic diagram illustrating the overall process of the predictive maintenance method for intelligent rail transit vehicles according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] Example 1 A predictive maintenance system for intelligent rail transit vehicles based on multi-source sensor fusion. Figure 1 The system's components are shown and can be viewed synchronously; the system includes the following functional modules: The multi-source data acquisition module is configured to synchronously acquire multi-source sensing data of the intelligent rail vehicle's dedicated components, including operating parameter data, multimodal perception data, and cross-system operation data. The operating parameter data includes the physical state parameters of at least one type of dedicated component, such as the guide wheel assembly and the articulation device. The multimodal perception data includes visual images and environmental monitoring information. The cross-system operation data is provided by the train control and management system and the automatic train monitoring system. The data fusion processing module is configured to perform noise suppression, anomaly correction and feature extraction on multi-source sensor data, and to perform weight allocation and feature fusion on heterogeneous data sources based on the attention mechanism, and output a fused feature vector representing the real-time state of the vehicle. The digital twin modeling module is configured to construct a digital twin model corresponding to the physical vehicle structure and operating logic, and dynamically map the operating status and degradation trend of vehicle components based on the fused feature vectors. The fault prediction and health assessment module is configured to combine the digital twin modeling module and the fusion feature vector to perform a quantitative assessment of the health status and predict the remaining service life of the intelligent rail vehicle's dedicated components, and generate health assessment results that include fault risk level and warning threshold. The maintenance decision optimization module is configured to generate a dynamic maintenance plan that matches the actual condition of the vehicle based on health assessment results, preset component association rules, and maintenance constraints. The resource scheduling and execution module is configured to parse dynamic maintenance plans and output maintenance task instructions, resource allocation schemes, and execution timing information to maintenance terminal devices.

[0025] In this embodiment, the multi-source data acquisition module includes an on-board sensing unit, a multimodal sensing unit, and a cross-system data interface unit.

[0026] The on-board sensing unit implements sensing configurations for three major categories of components specific to intelligent rail vehicles: guide wheel sets, articulation devices, and bogies. Specific devices involved may include: traction inverters, braking energy storage devices, wheel-side motors, air springs, anti-roll torsion bars, hydraulic shock absorbers, high-voltage distribution boxes, and on-board controllers, etc.

[0027] like Figure 2 As shown, the guide wheel assembly is equipped with a 3-axis MEMS vibration sensor and a laser displacement sensor, which are used to collect three-dimensional vibration acceleration signals and wheel rim wear displacement data on the X / Y / Z axes, respectively.

[0028] An angle sensor and a tension sensor are installed on the hinge device to monitor the change in hinge angle and the tension value of the connecting structure, respectively.

[0029] The bogie is equipped with a PT100 temperature sensor and a piezoelectric pressure sensor to obtain bearing temperature and suspension system pressure parameters, respectively.

[0030] In this embodiment, the specific parameters collected by the vehicle-mounted sensing unit include core parameters such as vibration, temperature, pressure, wear, displacement, rotation angle, current, and voltage. The sampling frequency is configured remotely through the cloud, and the data is output in JSON format so that subsequent edge nodes can perform local caching processing.

[0031] In this embodiment, the multimodal sensing unit includes industrial cameras, environmental parameter sensors, and virtual wheel-rail contact monitoring sensors distributed at the front, rear, and sides of the vehicle chassis. The industrial cameras are equipped with LED supplementary lights to ensure good acquisition performance in low-light environments. Their image data, transmitted via Ethernet parameters, will be used to identify key visual features such as looseness, leakage, and deformation of chassis components.

[0032] The environmental parameter sensor integrates temperature and humidity detection elements and dust concentration detection elements, and the preferred installation location is the roof vent.

[0033] The virtual wheel-rail contact monitoring sensor is installed next to the bogie wheelset. It is an ultrasonic flaw detection sensor used to collect stress distribution signals and wheel surface defect feature data in the virtual wheel-rail contact area.

[0034] In this embodiment, the cross-system data interface unit adopts a standardized interface design, supporting RESTful API, MQTT, or OPC UA communication protocols, thereby ensuring compatibility with the data formats of existing Train Control Management System (TCMS), Automatic Train Monitoring System (ATS), and Energy Management System (EMS). The data collected by this unit includes: Operational dispatch data includes: real-time vehicle speed, cumulative mileage, number of starts and stops, current passenger capacity, route number, and arrival time deviation. Energy consumption data: traction energy consumption, auxiliary system energy consumption, regenerative braking energy, and high-voltage battery SOC value, etc. Historical data: Fault records, maintenance records, and parts replacement records for the past year.

[0035] In this embodiment, the data fusion processing module includes an edge preprocessing unit and a cloud fusion unit. Its processing logic can be found in [reference needed]. Figure 3 The illustrations are used for simultaneous understanding.

[0036] The edge preprocessing unit performs adaptive Kalman filtering to suppress high-frequency noise, making it suitable for time-series data such as vibration and temperature. Further, it combines anomaly identification and interpolation correction based on the 3σ statistical criterion and the isolated forest algorithm, employing the SPA continuous projection algorithm for feature selection and dimensionality reduction of the original parameters. As an edge computing node, the edge preprocessing unit is installed in the electrical cabinet of the intelligent rail vehicle and communicates with the vehicle control system via a CAN bus to collect control data. It also connects to sensors and communication modules via Ethernet to achieve subsequent cloud communication transmission.

[0037] In this embodiment, the isolated forest algorithm uses 100 isolated trees and 256 samples to identify anomalous data that exceeds the normal distribution range. After labeling, it performs completion and correction through a combination of linear interpolation and spline interpolation. The continuous projection algorithm SPA performs feature filtering on multiple original parameters, retaining only core features with a contribution rate greater than 85%. After dimensionality reduction, the feature dimension is controlled to be less than or equal to 8, thereby reducing the computational complexity of subsequent operations.

[0038] The cloud-based fusion unit concatenates the dimensionality-reduced temporal sensor features, visual image features extracted by a convolutional neural network, and structured operational data features into vectors. This concatenation is then fed into a hybrid neural network model containing scaled dot-product attention layers, convolutional layers, and long short-term memory (LSTM) layers. The model calculates the dynamic weights of each data source and outputs a fused feature vector. The cloud-based fusion unit is deployed in the intelligent rail transit operation control center using a server cluster architecture. Simultaneously, this unit can also have its own dedicated computing server deployed at the station's ground-level locations to manage a specific number of intelligent rail transit vehicle data points.

[0039] In this embodiment, the overall architecture of the hybrid neural network model used here is a CNN-LSTM hybrid model employing an attention mechanism, which can be referred to as... Figure 4 The diagram is shown below, and its specific structure is described in detail: The input layer receives the dimensionality-reduced sensor features, visual image features, and operational data features, and concatenates the three into a 128-dimensional feature vector; among which, the visual image features are extracted by an additional CNN model, and preferably the first 5 layers of the VGG16 architecture are used. The attention layer employs a scaled dot product attention mechanism to calculate the weight coefficients of different data sources. The weight values ​​range from 0 to 1, thereby highlighting the contribution of key information. The CNN feature extraction layer consists of two convolutional layers and one max pooling layer; the convolutional kernels of the convolutional layers are 3×3 in size, with 64 and 128 kernels respectively, and the activation function is ReLU; the max pooling kernel is 2×2 in size; the CNN feature extraction layer extracts the local spatial features of its input. The LSTM temporal modeling layer consists of two LSTM layers; the number of hidden layer neurons in the two LSTM layers are 256 and 128 respectively, and the dropout rate is 0.2; the LSTM temporal modeling layer captures feature correlations in the time dimension; The output layer uses a 64-dimensional fully connected layer with a Sigmoid activation function to output the component's health status probability value, which ranges from 0 to 1.

[0040] In this embodiment, the hybrid neural network model is also put into use after training. The training process uses the Adam optimizer. After setting the learning rate and decay rate, the loss function is set to cross-entropy loss. After setting the batch size, the model is trained for a preset number of iterations until it converges. The data required for training is the same type as the data collected by the multi-source data acquisition module, consisting of various normal and abnormal data accumulated and recorded during the long-term operation of the intelligent rail vehicle.

[0041] The trained hybrid neural network model can detect abnormal data and adjust the detection process by setting additional preset thresholds. In this embodiment, based on three years of historical operation and maintenance data of the intelligent rail transit vehicle, a quantile statistical method is used to set a dynamic threshold range for each parameter of each component. This dynamic threshold range can be updated with the operating mileage of the intelligent rail transit vehicle, for example, adjusted once every 10,000 kilometers.

[0042] After the model outputs the component health status probability value, it is compared with the preset threshold to determine the warning threshold and the fault threshold respectively. When the warning threshold is exceeded, an early warning is triggered, and when the fault threshold is exceeded, an emergency alarm is triggered. The component health status probability value will be further applied in the intelligent rail vehicle's dedicated digital twin and in the maintenance decision-making process.

[0043] In this embodiment, the digital twin model constructed by the digital twin modeling module includes a three-dimensional geometric structure model, a physical degradation behavior model, and an operational logic mapping relationship for the intelligent rail vehicle's specific components. The modeling process employs a Unity3D + ANSYS Space Claim joint modeling approach to fully recreate the overall structure of the multi-section intelligent rail vehicle, and can appropriately refine the internal structural shape and positional relationships of the guide wheel assembly, articulation device, and bogie.

[0044] After modeling is completed, these virtual components are given real physical properties, including material parameters, mechanical properties, electrical parameters, etc. The deviation between them and the real physical components is controlled within an acceptable range of attribute errors, so that the current digital twin model can perform multiphysics simulation.

[0045] The main purpose of multiphysics simulation is to enable the digital twin model of the intelligent rail transit vehicle to predict future faults and assess health based on the current input data. The simulation tools used are ANSYS Workbench mechanical simulation and Simplorer electrical simulation to construct a multiphysics coupled simulation model, in which: The mechanical field simulation is used to model the wear of the guide wheel assembly, bogie vibration, and deformation of the articulation device, and the wear amount is calculated based on the Archard wear model. Thermodynamic field simulation of temperature changes in traction and braking systems, and calculation of temperature distribution based on Fourier heat conduction equation; The electrical field simulates the voltage and current changes of the traction inverter and wheel-side motor.

[0046] In this embodiment, the real-time simulation step size is set to 10ms, and the offline degradation simulation step size is set to 1s, thereby supporting simulation of different operating conditions.

[0047] The digital twin model is thus constructed, capable of receiving input from various intelligent rail vehicle data represented by fused feature vectors. Furthermore, real-time data synchronization between physical components and the virtual model is achieved through 5G edge nodes, ensuring that the virtual model's state is consistent with the physical entity of the intelligent rail vehicle. In practical applications, if a control relationship is established between the digital twin model and the intelligent rail vehicle control system, the monitoring configuration parameters of the physical component sensors can be adjusted through the model. Based on the digital twin model, the evolution process of various common faults can be simulated with existing data, thereby providing a virtual scenario for maintenance training of intelligent rail vehicles.

[0048] In this embodiment, the fault prediction and health assessment module, based on the real-time state mapping results of the digital twin model and fused feature vectors, calls a deep learning prediction model to perform time-series extrapolation of the remaining service life of components and outputs a quantitative health index and graded early warning signals. (See also...) Figure 5This illustrates the system's functional processes from health assessment to final maintenance resource scheduling and execution.

[0049] The quantitative health index is divided into three levels: Component level: The component health status probability value output by the hybrid neural network model is directly converted into a health index of 0-100 points. Since the probability value itself represents the component failure probability, the health index is obtained by multiplying 100 by (1-component health status probability value). It is divided into 5 levels: excellent (90-100 points), good (80-89 points), average (60-79 points), warning (40-59 points), and failure (0-39 points).

[0050] Subsystem level: The health index of all components in a subsystem that belongs to the same major category, such as guide wheel assembly, articulation device and bogie, as well as electrical part, braking part, etc., is obtained by weighted summation; the weight of each component in the subsystem is determined by the importance of the component.

[0051] Vehicle-level: The overall vehicle health index is obtained by summing the health indices of multiple subsystems with equal weights.

[0052] In this embodiment, the remaining service life of components is extrapolated over time using a degradation model based on the Wiener process. By inputting historical degradation data and real-time health indices, the remaining service life can be predicted, with the unit preset to the operating kilometers of the intelligent rail transit vehicle. For different levels of quantitative health indices, corresponding graded early warning signals can be generated according to the specific level classification method at the component level, and these signals can serve as influencing factors in the subsequent maintenance plan generation process.

[0053] In this embodiment, the maintenance decision optimization module has a built-in component association rule library and maintenance constraint condition library. The component association rule library stores the fault propagation logic and collaborative maintenance rules between different dedicated components, and the maintenance constraint condition library stores information on the availability status of maintenance resources, operating time windows, and safety specification requirements.

[0054] Furthermore, when the maintenance plan is generated, specific maintenance levels are defined: based on the fault level and component importance, maintenance levels are divided into three levels; fault levels are divided into general faults / serious faults / emergency faults, and component importance is divided into critical components / general components; the three levels of maintenance are as follows: Level 1 Emergency Maintenance: If a critical component experiences an emergency failure, maintenance should be arranged within 2 hours, using a fault repair mode. Secondary preventive maintenance: If a critical component experiences an early warning or a general component fails, maintenance should be arranged within 24 hours, adopting a preventive maintenance model; Level 3 routine maintenance: If the warning or health index of general components shows a significant downward trend, it should be included in the weekly / monthly maintenance plan and a regular maintenance mode should be adopted.

[0055] In this embodiment, the output information generated by the resource scheduling and execution module includes the maintenance task sequence, the list of required tools and equipment, the technical personnel configuration instructions, and the task execution time nodes, and is sent to the maintenance terminal equipment through the communication link.

[0056] Combining the two types of databases preset in the maintenance decision optimization module, after receiving and storing relevant data, the maintenance plan can be automatically generated when the maintenance plan is generated, based on the intelligent rail operation diagram, the remaining life of components, and the idle status of maintenance resources. The plan includes maintenance time, maintenance content, required tools, spare parts models, operation procedures, etc., and supports manual adjustment.

[0057] During resource scheduling, the optimization objectives are prioritized as the lowest maintenance cost, shortest response time, and highest resource utilization, and a corresponding weighted optimization function is constructed. Constraints such as maintenance personnel skill matching, spare parts inventory, maintenance workstation availability, maintenance time, and operational gaps are also incorporated. The preferred optimization algorithm is the improved Gray Wolf Optimization Algorithm (IGWO). In this embodiment, Logistic chaotic mapping is used to generate the initial population during chaotic mapping initialization to improve population diversity. A nonlinear convergence factor is used to balance global and local searches, and Gaussian mutation is introduced to prevent the algorithm from getting trapped in local optima.

[0058] As a preferred embodiment, a dedicated feedback optimization module can also be introduced to achieve iterative optimization of the entire system, continuously improving the final maintenance prediction accuracy and scheduling efficiency. The maintenance effect needs to be evaluated, and the evaluation indicators are the improvement rate of component health index before and after maintenance, the recurrence rate of faults within a preset time, and the maintenance time deviation rate.

[0059] Based on the evaluation results, an incremental learning online learning mechanism is adopted to update the parameters of the hybrid neural network model, including its weight parameters and threshold ranges, thereby continuously improving the model's prediction accuracy. Relevant data is also stored in the fault knowledge base for updates, and combined with the component association rule base and maintenance constraint condition base, all fault-related information can be queried in a structured manner. When the system identifies a new fault type, it automatically records the fault characteristics, processing results, and process, which can be manually reviewed before being entered into the knowledge database to enrich its sample data.

[0060] Example 2 Based on Example 1, this example provides a predictive maintenance method for intelligent rail transit vehicles based on the system of Example 1. Figure 6 The overall process of this method is briefly described below and can be viewed concurrently; the key steps of this method can be summarized as follows: S1. The physical state parameters of at least one type of dedicated component in the intelligent rail vehicle, such as the guide wheel assembly, articulation device, and bogie, are collected synchronously through a multi-source sensor network. The chassis visual image and environmental monitoring data are acquired synchronously, and the operation scheduling data and historical maintenance records are extracted from the train control management system and the automatic train monitoring system through a standardized communication interface. S2. The collected multi-source sensor data are sequentially subjected to adaptive filtering and noise reduction, abnormal data identification and correction based on statistical criteria and isolated forest algorithm, feature screening and dimensionality reduction processing; the processed time-series sensor features, visual image features and structured operation data features are vectorized and input into a neural network model containing attention mechanism for dynamic weight allocation and feature fusion to generate a fused feature vector representing the real-time state of the vehicle. S3. Call the digital twin model corresponding to the physical vehicle structure and operating logic, input the fused feature vector into the digital twin model, and complete the real-time mapping of the operating status of the vehicle-specific components and the deduction of degradation trends. S4. Based on the mapping results and fused feature vectors of the digital twin model, a time-series prediction model is used to perform a quantitative assessment of the health status and extra service life of the dedicated component, and output a health assessment result including fault risk level identifier and warning threshold. S5. Based on the health assessment results, and combined with the preset component association rule library and maintenance constraint library, generate a dynamic maintenance plan that includes maintenance items, execution priorities and suggested time windows; S6. Parse the dynamic maintenance plan, generate maintenance task sequence, required tool and equipment list, technician configuration instructions and task execution time nodes, and transmit the maintenance information to the maintenance terminal equipment through the communication link.

[0061] In step S2 of this embodiment, adaptive filtering and noise reduction uses an adaptive Kalman filter algorithm to process multi-source sensor data; abnormal data identification and correction are performed by combining the 3σ standard deviation statistical criterion and the isolated forest algorithm. After configuring 100 isolated trees and 256 sample sampling numbers, the data points determined to be abnormal are corrected by a combination of linear interpolation and spline interpolation; feature selection and dimensionality reduction use a continuous projection algorithm to select feature subsets with a cumulative contribution rate of not less than 85%.

[0062] Feature fusion involves concatenating the dimensionality-reduced temporal sensing features, visual image features extracted by a convolutional neural network, and structured operational data features into an input vector. This vector is then fed into a hybrid neural network model containing a scaled dot product attention layer, a convolutional layer, and a long short-term memory network layer. The convolutional layer includes two 3×3 convolutional kernels. The scaled dot product attention layer calculates the weight coefficients of each data source, and the long short-term memory network layer extracts temporal dependent features, outputting a fused feature vector.

[0063] In step S5 of this embodiment, the component association rule library stores the wear association rules between the guide wheel assembly and the virtual wheel-rail contact monitoring sensor, and the mechanical maintenance rules between the articulation device and the bogie; the maintenance constraint condition library stores the qualification and scheduling data of maintenance personnel, the inventory and calibration status information of tools and equipment, the available time window for the outage of the operating line and the safe operation procedures. When generating a dynamic maintenance plan, the system searches the component association rule library based on the fault risk level identifier in the health assessment results to determine the set of components that need to be maintained in conjunction with the maintenance plan. Combining the real-time resource status and time-series constraints in the maintenance constraint library, the system prioritizes maintenance projects and allocates time windows to generate a dynamic maintenance plan that includes a list of specific maintenance projects, execution priority sequence, suggested start and end times, and associated component identifiers.

[0064] This embodiment will divide the method into several main stages and describe the specific sub-steps involved in each stage.

[0065] I. Data Collection Phase S101. The sensor collects data at a set frequency. The industrial camera captures one frame of image every 1 second, and the ultrasonic flaw detection sensor collects data once every 0.5 seconds. S102. The vehicle-mounted edge gateway receives raw data, performs format conversion, unifies it into JSON format, and synchronizes it with the timestamp; the NTP protocol is used here. S103 transmits data to ground edge nodes via a 5G private network, while locally caching data from the last 24 hours to avoid data loss due to network interruption; S104. The cross-system interface periodically collects operational data from the TCMS, ATS, and EMS systems and aligns it with the sensor data by timestamp.

[0066] II. Data Processing Stage S201. Edge nodes preprocess the data: Adaptive Kalman filtering is used to remove noise, the isolated forest algorithm is used to identify outliers, and linear interpolation is used to complete missing data. S202. Use the continuous projection algorithm SPA to perform feature dimensionality reduction on the preprocessed data and extract the core features; S203: Images acquired by industrial cameras are preprocessed and then input into the VGG16 model to extract visual features. S204. The sensor features, visual features, and operational data features are spliced ​​together and transmitted to the cloud multi-source fusion engine, which contains a hybrid neural network model. S205. The hybrid neural network model processes the input features and outputs the component health status probability value and abnormal label.

[0067] III. Modeling and Prediction Stage S301: The digital twin platform receives health status data output from the cloud, updates the status parameters of the virtual model, and achieves virtual-real synchronization. S302. Based on a multiphysics simulation model, input the current operating conditions and simulate the future degradation trend of each component. S303. Calculate the remaining life of the component using the Wiener process degradation model, combined with simulation results and historical data. S304. When the component health index or remaining lifespan is lower than the corresponding preset threshold, the corresponding early warning or emergency alarm is triggered.

[0068] IV. Decision-making and Implementation Stage S401 The intelligent decision-making module generates corresponding maintenance plans based on the warning level, component importance, and operation plan; S402 The resource optimization scheduler uses an improved gray wolf optimization algorithm to allocate maintenance personnel, maintenance workstations, and spare parts, and generate a scheduling plan. S403: Push maintenance tasks to maintenance personnel via industrial tablet, including fault location, handling process, required tools, spare parts information, and synchronize to the intelligent rail operation and dispatch system, reserving maintenance time window; S404. After the maintenance personnel arrive at the site, they confirm their identity and task by scanning the NFC tag of the industrial flat panel device, perform maintenance operations according to the instructions, and upload the operation process and results in real time.

[0069] V. Optimization Phase S501. Compare the component health index and fault recurrence before and after maintenance to evaluate the maintenance effect; S502. Based on the evaluation results, the parameters of the fusion model are updated using an incremental learning mechanism, and the threshold range of the components is adjusted. S503. Input new fault types, fault characteristics, and handling solutions into the fault knowledge base to enrich the fault sample. S504. Regularly generate maintenance effectiveness analysis reports, including indicators such as failure rate, maintenance cost, and resource utilization rate, to provide decision support for operation management.

[0070] The following is an example of the predictive maintenance method for intelligent rail vehicles in this embodiment, applied in a specific scenario. Taking the intelligent rail line T2 train in a certain area as an example, the application effect after system deployment is as follows: Fault Warning Case 1: One day, the system detected a continuous rise in the bearing temperature of the No. 1 bogie of a train, from 45℃ to 68℃, and the vibration acceleration increased from 0.3g to 0.75g. The health index dropped to 58 points, triggering a warning and predicting a remaining lifespan of 5200 kilometers. The system automatically generated a maintenance plan, scheduling maintenance at the depot from 23:00 that night to 5:00 the next morning. After the maintenance personnel replaced the bearing, the health index recovered to 92 points, preventing a shutdown accident caused by bearing seizure.

[0071] Fault Warning Case 2: One day, the system detected a slight leak in the middle articulation device of the train through visual image recognition. Simultaneously, the tension sensor data showed an increased fluctuation range in connection tension, and the health index dropped to 62 points, triggering a warning. Based on the operational plan, the system scheduled localized maintenance at a station from 12:00 to 13:00 the following day. After tightening the seals, the fault was resolved. The maintenance took only 45 minutes and did not affect normal operations.

[0072] Test results show that after the method of this embodiment is run in the system, all indicators meet the design requirements, which can effectively improve the accuracy and efficiency of intelligent rail vehicle maintenance, reduce operating costs and safety risks, and has practical engineering application value.

Claims

1. A predictive maintenance system for intelligent rail transit vehicles based on multi-source sensor fusion, characterized in that, The system includes the following functional modules: The multi-source data acquisition module is configured to synchronously acquire multi-source sensing data of the intelligent rail vehicle's dedicated components, including operating parameter data, multimodal perception data, and cross-system operation data. The operating parameter data includes physical state parameters of at least one type of dedicated component, such as the guide wheel assembly and the articulation device. The multimodal perception data includes visual images and environmental monitoring information. The cross-system operation data is provided by the train control and management system and the automatic train monitoring system. The data fusion processing module is configured to perform noise suppression, anomaly correction and feature extraction processing on the multi-source sensor data, and to perform weight allocation and feature fusion on heterogeneous data sources based on the attention mechanism, and output a fused feature vector representing the real-time state of the vehicle. The digital twin modeling module is configured to construct a digital twin model corresponding to the physical vehicle structure and operating logic, and dynamically map the operating status and degradation trend of vehicle components based on the fused feature vector. The fault prediction and health assessment module is configured to combine the digital twin modeling module and the fused feature vector to perform a quantitative assessment of the health status and predict the remaining service life of the intelligent rail vehicle's dedicated components, and generate a health assessment result that includes fault risk level and warning threshold. The maintenance decision optimization module is configured to generate a dynamic maintenance plan that matches the actual condition of the vehicle based on the health assessment results, preset component association rules, and maintenance constraints. The resource scheduling and execution module is configured to parse the dynamic maintenance plan and output maintenance task instructions, resource allocation schemes and execution timing information to the maintenance terminal device.

2. The predictive maintenance system for intelligent rail transit vehicles according to claim 1, characterized in that: The multi-source data acquisition module includes an on-board sensing unit, a multimodal sensing unit, and a cross-system data interface unit. The on-board sensing unit is integrated into at least one type of dedicated component in the intelligent rail vehicle's guide wheel assembly, articulation device, and bogie, and is equipped with at least one of vibration sensors, displacement sensors, angle sensors, tension sensors, temperature sensors, and pressure sensors. The multimodal sensing unit includes industrial cameras, environmental parameter sensors, and virtual wheel-rail contact monitoring sensors distributed at the front, rear, and sides of the vehicle chassis. The cross-system data interface unit uses RESTful API, MQTT or OPC UA communication protocols to establish data connections with the train control management system, the automatic train monitoring system and the energy consumption management system to obtain at least one of real-time vehicle speed, cumulative mileage, passenger capacity, energy consumption data and historical maintenance records.

3. The predictive maintenance system for intelligent rail transit vehicles according to claim 2, characterized in that: In the vehicle-mounted sensing unit, the sensors deployed on the guide wheel assembly include a triaxial vibration sensor and a displacement sensor, which are used to collect three-dimensional vibration acceleration signals and wheel flange wear displacement data, respectively. The sensors deployed on the articulation device include angle sensors and tension sensors, which are used to monitor the change in articulation angle and the tension value of the connecting structure, respectively. Sensors deployed on the bogie include temperature sensors and pressure sensors, used to acquire bearing temperature and suspension system pressure parameters, respectively.

4. The predictive maintenance system for intelligent rail transit vehicles according to claim 2, characterized in that: In the multimodal sensing unit, the industrial camera is equipped with a supplementary lighting device and transmits image data via Ethernet; the environmental parameter sensor integrates temperature and humidity detection elements and dust concentration detection elements; the virtual wheel-rail contact monitoring sensor is an ultrasonic flaw detection sensor, used to collect stress distribution signals and wheel surface defect feature data in the virtual wheel-rail contact area.

5. The predictive maintenance system for intelligent rail transit vehicles according to claim 1, characterized in that: The data fusion processing module includes an edge preprocessing unit and a cloud fusion unit; The edge preprocessing unit is used to perform adaptive Kalman filtering to suppress high-frequency noise, and combines anomaly data identification and interpolation correction based on statistical criteria and isolated forest algorithm with continuous projection algorithm to perform feature screening and dimensionality reduction on the original parameters. The cloud fusion unit concatenates the dimensionality-reduced temporal sensing features, visual image features extracted by the convolutional neural network, and structured operational data features into vectors, and inputs them into a hybrid neural network model containing a scaled dot product attention layer, a convolutional layer, and a long short-term memory network layer to calculate the dynamic weights of each data source and output the fused feature vector.

6. The predictive maintenance system for intelligent rail transit vehicles according to claim 1, characterized in that: The digital twin model constructed by the digital twin modeling module includes a three-dimensional geometric structure model of the intelligent rail vehicle's exclusive components, a physical degradation behavior model, and an operational logic mapping relationship. The fault prediction and health assessment module, based on the real-time state mapping results of the digital twin model and the fused feature vector, calls a deep learning prediction model to perform time-series extrapolation of the remaining service life of the component, and outputs a quantitative health index and graded early warning signal.

7. The predictive maintenance system for intelligent rail transit vehicles according to claim 1, characterized in that: The maintenance decision optimization module has a built-in component association rule library and maintenance constraint condition library. The component association rule library stores the fault propagation logic and collaborative maintenance rules between different dedicated components, and the maintenance constraint condition library stores information on the availability status of maintenance resources, operating time windows, and safety specification requirements. The output information generated by the resource scheduling and execution module includes maintenance task sequence, list of required tools and equipment, technical personnel configuration instructions and task execution time nodes, and is sent to maintenance terminal equipment through communication link.

8. A predictive maintenance method for intelligent rail transit vehicles according to claim 1, characterized in that, Includes the following steps: S1. The physical state parameters of at least one type of dedicated component in the intelligent rail vehicle, such as the guide wheel assembly, articulation device, and bogie, are collected synchronously through a multi-source sensor network. The chassis visual images and environmental monitoring data are acquired simultaneously, and the operation scheduling data and historical maintenance records are extracted from the train control management system and the automatic train monitoring system through a standardized communication interface. S2. The collected multi-source sensor data are sequentially subjected to adaptive filtering and noise reduction, abnormal data identification and correction based on statistical criteria and isolated forest algorithm, feature screening and dimensionality reduction processing; the processed time-series sensor features, visual image features and structured operation data features are vectorized and input into a neural network model containing attention mechanism for dynamic weight allocation and feature fusion to generate a fused feature vector representing the real-time state of the vehicle. S3. Call the digital twin model corresponding to the physical vehicle structure and operating logic, input the fused feature vector into the digital twin model, and complete the real-time mapping of the operating status of the vehicle-specific components and the deduction of degradation trends. S4. Based on the mapping result of the digital twin model and the fused feature vector, a time-series prediction model is used to perform a quantitative assessment of the health status and a projection of the remaining service life of the dedicated component, and outputs a health assessment result including a fault risk level identifier and a warning threshold. S5. Based on the health assessment results, and combined with the preset component association rule library and maintenance constraint library, generate a dynamic maintenance plan that includes maintenance items, execution priorities and suggested time windows; S6. Parse the dynamic maintenance plan, generate maintenance task sequence, required tool and equipment list, technician configuration instructions and task execution time nodes, and transmit the maintenance information to the maintenance terminal equipment through the communication link.

9. The predictive maintenance method for intelligent rail transit vehicles according to claim 8, characterized in that: In step S2, the adaptive filtering noise reduction uses an adaptive Kalman filter algorithm to process multi-source sensor data; the abnormal data identification and correction is performed by combining the three-standard-deviation statistical criterion and the isolated forest algorithm. After configuring a preset number of isolated trees and a preset number of sampled data, the data points determined to be abnormal are corrected by a combination of linear interpolation and spline interpolation. The feature selection and dimensionality reduction process uses a continuous projection algorithm to select a subset of features with a cumulative contribution rate of not less than a preset percentage. The feature fusion is as follows: the dimensionality-reduced temporal sensing features, visual image features extracted by the convolutional neural network, and structured operational data features are concatenated into an input vector, which is then input into a hybrid neural network model containing a scaled dot product attention layer, a convolutional layer, and a long short-term memory network layer. The convolutional layer contains two convolutional operations with a kernel size of 3×3. The scaled dot product attention layer calculates the weight coefficients of each data source, and the long short-term memory network layer extracts temporal dependent features and outputs a fused feature vector.

10. The predictive maintenance method for intelligent rail transit vehicles according to claim 8, characterized in that: In step S5, the component association rule library stores wear association rules between the guide wheel assembly and the virtual wheel-rail contact monitoring sensor, and mechanical maintenance rules between the articulation device and the bogie; the maintenance constraint condition library stores maintenance personnel qualification and scheduling data, tool and equipment inventory and calibration status information, available time windows for outages of operating lines, and safe operating procedures. When generating a dynamic maintenance plan, the system searches the component association rule library based on the fault risk level identifier in the health assessment results to determine the set of components that need to be maintained in conjunction with the maintenance plan. Combining the real-time resource status and time-series constraints in the maintenance constraint library, the system prioritizes maintenance projects and allocates time windows to generate a dynamic maintenance plan that includes a list of specific maintenance projects, execution priority sequence, suggested start and end times, and associated component identifiers.

Images