State monitoring and fault early warning method for magnetic levitation vehicle system

Abstract

The invention provides a state monitoring and fault early warning method for a maglev vehicle system, relates to the technical field of urban rail transit data processing, and solves the problem of how to effectively combine an empirical threshold mode and a machine learning technology and give full play to a combination effect. The method comprises the steps of collecting operation data of a magnetic levitation vehicle system, performing corresponding data preprocessing operation, performing analysis and calculation in a mode of constructing a mathematical model, and evaluating the health state of the magnetic levitation vehicle system; constructing and training a fault prediction model, and inputting the evaluation data into the trained fault prediction model to obtain potential fault data of the magnetic levitation vehicle system; and generating early warning information based on the potential fault data, retrieving an expert experience library by using the early warning information, generating an intelligent decision result, and feeding back the intelligent decision result to the operation and maintenance personnel. According to the method, machine learning and expert experience are effectively combined, so that the real-time performance and the accuracy of fault early warning are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rail transit data processing and is applied to a maglev vehicle system. Specifically, it relates to a method for state monitoring and fault warning of a maglev vehicle system. Background Art

[0002] As a typical representative of new rail transit technologies, the maglev vehicle system realizes contactless suspension and guidance through electromagnetic force, and has the advantages of low noise and high ride comfort while eliminating mechanical friction losses. With the continuous growth of the demand for efficient and intelligent transportation, this technology has gradually entered the commercial operation stage from the experimental verification stage, and its application scope covers intercity high-speed trunk lines, urban rail transit, and special scenario transportation systems. Against the background of the continuous improvement of the system operation reliability requirements, the intelligent state monitoring and fault warning technology has become the core support technology for ensuring operation safety and optimizing maintenance strategies.

[0003] The existing state monitoring technology mainly relies on a multi-source heterogeneous sensor network to construct a perception system, and implements real-time acquisition of key parameters such as suspension gap, electromagnetic field strength, track deformation, and temperature rise characteristics. At the level of fault warning methods, there are two typical solutions in the current technical system: the rule-based threshold determination method and the machine learning-based data-driven method. The former realizes rapid response by setting expert experience thresholds. For example, when the suspension control current fluctuation exceeds ±15% of the rated value, a primary warning is triggered, and when the temperature sensor reading breaks through the preset value, a forced cooling program is started. Although this method has the advantage of simple implementation, it is difficult to effectively cope with the system's non-linear characteristics and multi-parameter coupling effects, especially in dealing with hidden faults and compound faults, there are warning blind spots.

[0004] The intelligent warning method based on machine learning deeply mines dynamic characteristics such as the vibration spectrum of the suspension system and the change gradient of the guiding force by constructing time series network models such as LSTM and GRU. Existing research shows that by using technologies such as gated recurrent unit networks, it is possible to achieve a certain degree of early warning ability while maintaining good accuracy. However, this type of method also faces three major technical bottlenecks: First, a large amount of effective working condition data needs to be collected to ensure the generalization ability of the model, and the data acquisition cost and annotation difficulty restrict its practical application; second, the explosive growth of the feature dimension under the coupling action of multiple physical fields leads to a significant increase in the calculation delay, making it difficult to meet the real-time warning requirements; third, the existing models have a low recognition accuracy for small probability events such as sudden mechanical shocks and transient electromagnetic interferences.

[0005] At the system integration application level, the existing technologies generally have the problem of insufficient coordination between monitoring and early warning and intelligent control. Although intelligent transportation system technologies have achieved advanced functions such as autonomous driving and dynamic formation, their data interaction with the equipment health management system still remains in the stage of post-event analysis, which may cause the control strategy adjustment to lag behind the evolution of fault characteristics and miss the best intervention window; the competition for computing resources during multi-node alarms is also likely to result in the loss of critical early warning information.

[0006] It can be seen that the traditional threshold method is difficult to meet the accurate early warning requirements under complex working conditions, and the data-driven method is limited by the bottlenecks of real-time performance and reliability; these technical defects will directly affect the availability and maintenance economy of the maglev system. Therefore, it is necessary to build a new architecture of intelligent monitoring and early warning system for the maglev vehicle system to improve the business effects of corresponding scenarios. Summary of the Invention

[0007] Based on the current situation in the background technology, the purpose of the present invention is to solve the problem of how to effectively combine the empirical threshold method and machine learning technology and give full play to the combined effect during the monitoring and early warning process of the existing maglev vehicle system. Therefore, a method for state monitoring and fault early warning of a maglev vehicle system is proposed. When applying machine learning technology, the present invention not only relies on data driving but also combines rules and thresholds related to expert experience, so that the fault early warning is improved in terms of both real-time performance and accuracy, and the preventive maintenance early warning effect before the possible occurrence of faults is successfully achieved.

[0008] The present invention adopts the following technical solutions to achieve the purpose:

[0009] A method for state monitoring and fault early warning of a maglev vehicle system, the method comprising the following steps:

[0010] S1. Collect the operation data of the maglev vehicle system and perform corresponding data preprocessing operations;

[0011] S2. Based on the operation data, analyze and calculate by constructing a mathematical model to evaluate the health state of the maglev vehicle system and obtain evaluation data;

[0012] S3. Construct and train a fault prediction model, input the evaluation data into the trained fault prediction model to obtain the potential fault data of the maglev vehicle system;

[0013] S4. Generate early warning information based on the potential fault data; use the early warning information to retrieve the expert experience database, generate an intelligent decision result and feedback it to the operation and maintenance personnel.

[0014] Specifically, in step S1, the train control system TCMS is used to collect the operation data generated by the maglev vehicle system during operation at a preset frequency. The categories of the operation data include speed, acceleration, suspension gap, motor current, and motor voltage.

[0015] Preferably, data preprocessing operations are performed on the operation data, including data cleaning, data denoising, and data normalization. Among them, data cleaning is implemented by median filtering, data denoising is implemented by Kalman filtering, and data normalization is implemented by the min-max normalization method.

[0016] Furthermore, in step S2, corresponding health state evaluations are performed for the linear motor and suspension performance of the maglev vehicle system. The categories of the mathematical models constructed during the evaluation process include linear regression models, support vector machine models, and neural network models. After the mathematical models are established and corresponding training is completed, the input is the operation data, and the output is the evaluation data.

[0017] Specifically, the establishment and training of the mathematical model are as follows: historical data of the maglev vehicle system is collected in advance and divided into a training set and a test set. Data features including speed, acceleration, suspension gap, motor current, and motor voltage are extracted from the historical data. Linear regression models, support vector machine models, and neural network models are established respectively, and the three types of models established are trained using the training set. During the training process, the model parameters are adjusted correspondingly to enable each model to predict the health state of the system as accurately as possible. Finally, the optimal evaluation data is selected and determined from the outputs of the three types of models, and the performance of the three types of models is tested using the test set.

[0018] Furthermore, in step S3, the categories of the fault prediction models constructed include linear regression models, support vector machine models, and neural network models. After the fault prediction models are established and corresponding training is completed, the input is the evaluation data, and the output is the potential fault data. The establishment and training methods of the fault prediction models are the same as those of the mathematical models.

[0019] Specifically, in step S4, the generated warning information includes: the fault type, fault location, and fault occurrence time after the potential fault corresponding to the potential fault data is transformed into an actual fault.

[0020] Preferably, in step S4, the expert experience library stores fault diagnosis rule information, fault repair strategy information, and fault maintenance experience information. After retrieving and matching the corresponding information pre-stored in the expert experience library according to the warning information, the intelligent decision result is generated.

[0021] Preferably, the method further includes: S5. When the maglev vehicle system is actually operating, the remote control center continuously and real-time obtains the operation data, the evaluation data, the potential fault data, the warning information, and the intelligent decision-making result based on the Internet of Things technology, so as to quickly take higher-level countermeasures to reduce the impact of possible faults; meanwhile, classify and store all the obtained data, and correspondingly establish a fault warning maintenance record; use the fault warning maintenance record to update the expert experience database to provide more experienced data support for future fault analysis and prevention.

[0022] Specifically, the remote control center uses the Hadoop distributed storage system to classify and store all the obtained data, and at the same time uses the Spark big data analysis platform technology to correspondingly establish a fault warning maintenance record after completing the data analysis.

[0023] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows:

[0024] The present invention collects the operation data of the maglev vehicle system through the TCMS, and uses the linear motor and suspension data analysis technology to evaluate the health status of the system, realizing the real-time monitoring of the maglev vehicle system. At the same time, by analyzing the operation data through machine learning algorithms, possible faults are predicted, so as to give a warning before the actual occurrence of the faults, improving the reliability and safety of the system.

[0025] When the present invention analyzes the operation data through machine learning algorithms, it not only relies on data-driven, but also combines expert experience, and makes the fault warning more accurate by means of preset rules and thresholds on the basis of the analysis. Compared with the method that only relies on machine learning, this feature of the present invention can still maintain an effective fault warning effect in the case of insufficient data volume or limited computing resources.

[0026] Through the method of the present invention, potential faults of the maglev vehicle system are predicted and warnings are issued in a timely manner, so as to achieve the effect of preventive maintenance, thereby avoiding the actual occurrence of faults and further reducing the maintenance cost. Therefore, on the basis of improving the reliability and safety of the system, the present invention improves the operation efficiency and is beneficial to improving the travel experience of passengers. Brief Description of the Drawings

[0027] Figure 1 It is a schematic diagram briefly describing the overall process of the method of the present invention. Detailed Embodiment

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention generally described and illustrated in the figures herein can be arranged and designed in a variety of different configurations.

[0029] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0030] Embodiment

[0031] A method for state monitoring and fault warning of a maglev vehicle system, Figure 1 The overall process of the method is briefly described, and the steps of the method can be summarized as follows for synchronous reference:

[0032] S1. Collect the operation data of the maglev vehicle system and perform corresponding data preprocessing operations;

[0033] S2. Based on the operation data, perform analysis and calculation by constructing a mathematical model to evaluate the health status of the maglev vehicle system and obtain evaluation data;

[0034] S3. Construct and train a fault prediction model, and input the evaluation data into the trained fault prediction model to obtain potential fault data of the maglev vehicle system;

[0035] S4. Generate warning information based on the potential fault data; use the warning information to retrieve the expert experience database, generate an intelligent decision result, and feedback it to the operation and maintenance personnel.

[0036] This embodiment will explain and introduce the detailed content of each of the above steps.

[0037] In step S1, through the train control system TCMS, the operation data generated by the maglev vehicle system during operation is collected at a preset frequency. The categories of the operation data include speed, acceleration, suspension gap, motor current, and motor voltage. Subsequently, data preprocessing operations are performed on the operation data, including data cleaning, data denoising, and data normalization; among them, data cleaning is implemented by the median filtering method, data denoising is implemented by the Kalman filtering method, and data normalization is implemented by the min-max normalization method.

[0038] During the operation of the maglev vehicle system in this embodiment, the train control system (TCMS) obtains multi-dimensional operation parameters in real time through an integrated sensor network. For example, the data acquisition module in this system can synchronously acquire the original signals generated by the suspension control unit, the power drive unit, and the on-vehicle monitoring device with a period of 50 ms. Specifically, it can cover the pulse frequency signal of the speed sensor, the vibration waveform of the triaxial accelerometer, the millimeter-level displacement of the suspension gap probe, the instantaneous current of the traction motor, and the voltage pulsation parameters. For the characteristics of different physical quantities collected, in addition to using the standard sampling strategy with a fixed frequency, the system can also achieve the optimal data acquisition effect through an adaptive sampling strategy: for high-frequency change parameters such as the dynamic fluctuation of the suspension gap, a 1000 Hz high-speed acquisition mode is enabled; for other basic parameters such as the slowly changing temperature parameter, it is switched to a 1 Hz low-frequency monitoring mode to achieve the optimal balance between data acquisition efficiency and storage resource occupancy.

[0039] After the collected original operation data is transmitted to the edge computing node through the bus, a multi-level preprocessing process can be executed in sequence. In the data cleaning stage, a sliding window dynamic median filtering algorithm is adopted, and a 3-second time window is set to suppress the sudden jump noise of the suspension gap. When the data points in the window deviate from the median value by more than ±2σ, they are automatically replaced with the weighted average of adjacent data points. For the high-frequency oscillation of the current signal caused by electromagnetic interference, the system constructs an improved Kalman filter with a time-varying noise covariance. Its state equation is established based on the equivalent circuit model of the linear traction motor, and the observation equation fuses the cross-validation data of the current sensor and the voltage sensor to eliminate the measurement deviation caused by the electromagnetic coupling effect through recursive calculation. After the denoising process, each parameter is uniformly subjected to dynamic range normalization: for the motor current signal, the normalization reference is set according to the rated working current of 100 A, and the actual measurement value is mapped to the [0,1] interval by piecewise linear transformation; the suspension gap data is nonlinearly compressed according to the track geometric tolerance range to ensure the distribution consistency of parameters with different dimensions.

[0040] The operation data after preprocessing can be transmitted to the system processing unit through a dual-redundancy communication channel and perform an integrity check before temporary storage. The system dedicated to state monitoring and fault warning can build an abnormal data traceability mechanism. When it detects the absence of data for 5 consecutive sampling periods, it automatically triggers the data replenishment program of adjacent nodes and reconstructs the complete time series through the timestamp alignment algorithm. To cope with the sudden storage failure in the on-vehicle environment, the intermediate results of the preprocessing are written into a circular buffer with an error-checking function in real time to ensure the data recoverability under extreme conditions. The operation data after the complete preprocessing process is finally stored in the corresponding database in a time-space joint indexing manner as the input source for the subsequent step S2.

[0041] In step S2, a corresponding health status assessment is carried out for the linear motor and suspension performance of the maglev vehicle system; the types of mathematical models constructed during the assessment process include linear regression models, support vector machine models, and neural network models. After the mathematical models are established and corresponding training is completed, their input is operation data and the output is assessment data.

[0042] Generally speaking, the establishment and training of the mathematical model are specifically as follows: historical data of the maglev vehicle system is collected in advance and divided into a training set and a test set; data features including speed, acceleration, suspension gap, motor current, and motor voltage are extracted from the historical data; linear regression models, support vector machine models, and neural network models are respectively established, and the three established models are respectively trained using the training set. During the training process, the model parameters are adjusted correspondingly, and the optimal assessment data is selected and determined from the outputs of the three models. Subsequently, the performance of the three models is tested using the test set.

[0043] In the health status assessment link of the maglev vehicle system in this embodiment, a dynamic diagnosis system is constructed by adopting an evaluation strategy of multi-model fusion for the performance degradation characteristics of the linear motor and the suspension system. During the model training process, the system calls historical operation data, covering three types of data sets: normal working conditions, known fault modes, and boundary conditions, and can specifically divide the training set, validation set, and test set according to a ratio of 7:2:1. In the feature engineering stage, a time-domain and frequency-domain joint analysis method is adopted to deeply analyze the original operation data: for the linear motor current and voltage signals, electrical characteristics such as three-phase unbalance degree, harmonic distortion rate, and ripple coefficient are extracted; for the suspension system, the vibration energy distribution in the 0-500Hz frequency band is obtained through fast Fourier transform, and the root mean square value and peak-to-peak ratio of the gap fluctuation amount are calculated.

[0044] During the model construction process, the linear regression model focuses on analyzing the linear mapping relationship between motor efficiency and operation parameters, selects 12 core features such as current harmonic content and voltage fluctuation coefficient through the stepwise regression method, and introduces an L2 regularization term to control the overfitting risk.

[0045] The support vector machine model uses a Gaussian radial basis kernel function to handle nonlinear classification problems, sets the relaxation variable tolerance to 0.1, optimizes the combined parameters of the penalty factor C and the kernel width γ through the grid search method, and constructs a decision boundary for the stability assessment of the suspension system.

[0046] The neural network model is designed with a dual-channel input structure, including an electrical feature analysis channel with 5 convolutional modules and a mechanical vibration analysis channel with 3 LSTM units. The attention mechanism is introduced into the fusion layer to dynamically adjust the feature weights, and the final output layer uses a Sigmoid activation function to generate a health index in the 0-1 interval.

[0047] In this embodiment, the cross-validation strategy is adopted to optimize hyperparameters in the model training stage. Among them, the linear regression model fits the parameter matrix by the least squares method, the support vector machine model uses the sequential minimal optimization algorithm to accelerate the solution process, and the neural network model adopts the Adam optimizer combined with the dynamic learning rate decay strategy. During the training process, the F1 score and mean squared error metrics on the validation set are monitored in real time. When the performance has not improved for 10 consecutive epochs, the early stopping mechanism is automatically triggered. In the model selection link, a multi-dimensional evaluation system is established, comprehensively considering three indicators: prediction accuracy, calculation efficiency, and memory occupancy. The weights of each indicator are determined by the entropy weight method. Finally, the neural network model can be selected as the core engine for online evaluation, and the linear regression model is retained for rapid trend prediction.

[0048] In this embodiment, an adversarial sample verification mechanism is introduced in the model testing stage, and 5% of noisy perturbation data is added to the test set to test the robustness of the model. The system synchronously establishes a model degradation monitoring module. When the deviation between the prediction results and the linear regression model exceeds the threshold for 100 consecutive times, the model retraining process is automatically triggered. The evaluation data output adopts the dynamic confidence annotation method, and abnormal probability annotations are added to the data points where the health index fluctuates more than ±3σ, providing a traceable decision basis for subsequent fault prediction.

[0049] In step S3, the types of the constructed fault prediction models include the linear regression model, the support vector machine model, and the neural network model. After the fault prediction model is established and corresponding training is completed, its input is the evaluation data, and the output is the potential fault data. The overall fault prediction model and training method constructed in this step are similar to the mathematical model in step S2, only with adaptive adjustments and differences in specific sub-business functions.

[0050] In step S4, the generated warning information includes: the fault type, fault location, and fault occurrence time after the potential fault corresponding to the potential fault data is transformed into an actual fault. The expert experience library stores fault diagnosis rule information, fault repair strategy information, and fault maintenance experience information; according to the warning information, the corresponding information pre-stored in the expert experience library is retrieved and matched to generate an intelligent decision result.

[0051] In the above-mentioned fault prediction stage of this embodiment, the system constructs a multi-modal fusion prediction system, and fault prediction models are established for typical fault modes such as linear motor winding deterioration and suspension electromagnet demagnetization. The input end of the fault prediction model receives the multi-dimensional feature vector corresponding to the evaluation data, such as derivative indicators such as motor efficiency decay rate and suspension energy consumption growth coefficient.

[0052] The fault prediction model architecture adopts a cascaded design. Among them, the linear regression model focuses on capturing the long-term trend of parameter changes. By constructing an autoregressive equation with time series lags, it predicts the performance degradation curve within the next 3 maintenance cycles. The support vector machine model deals with multi-fault coupling scenarios. It adopts a one-versus-all classification strategy to establish the discrimination boundaries of up to 12 typical fault modes, and the kernel function parameters are dynamically adjusted according to the feature space distribution. The neural network model is designed as a deep time series prediction network, which includes a combined structure of a bidirectional GRU layer and a self-attention mechanism. It extracts the dynamic evolution features of the evaluation data through a sliding time window and outputs a three-dimensional probability distribution map of the fault occurrence probability and the remaining effective operating time.

[0053] In the training of the fault prediction model, this embodiment also adopts a transfer learning strategy. First, the basic network is pre-trained on the historical fault data set, and then the newly collected evaluation data is fine-tuned through an online learning module. A dynamic weight allocation mechanism is introduced during the training process. A weighted coefficient of 0.7 is assigned to the operation data within the past three months, which can ensure that the model continuously adapts to the feature drift caused by system aging. For the prediction requirements of different fault modes, the system establishes a differentiated performance evaluation system: the mean square error index is used to optimize progressive faults, and the F2 score is emphasized for sudden faults to improve the recall rate. At the model output stage, confidence calibration is implemented. When the differences in the potential fault data output by the three models exceed the threshold, a decision fusion algorithm based on the D-S evidence theory is started to generate the final prediction result by integrating the credibility weights of each model.

[0054] Based on this prediction result reflecting the potential fault data, the warning information generation module of the system maps it to the physical topology structure of the vehicle system. The root cause location of the fault can be determined through fault propagation tree analysis. The built-in spatio-temporal association engine of the system combines the real-time position information of the train and the track feature database, calculates the geographical location probability distribution of the fault occurrence and the specific position of the fault point in the vehicle, and deduces the fault evolution time axis based on the vehicle dynamics model. The warning levels issued externally can be divided into a four-level system: the first-level warning corresponds to a planned maintenance event within 72 hours, the second-level warning indicates a potential risk that requires intervention within 24 hours, the third-level warning requires an immediate downgrade of the operation mode, and the fourth-level warning triggers the emergency braking and passenger evacuation procedures.

[0055] In this embodiment, the expert experience database adopts a hybrid architecture that combines a knowledge graph and case-based reasoning, and can contain the structured features and solutions of a large number of historical maintenance cases. The nodes of the knowledge graph cover entity types such as equipment components, failure modes, and maintenance actions, and the edge relationships define the causal associations of failure propagation paths and disposal strategies. When a warning message is received, the system first performs three-dimensional matching of the failure mode: matching the equipment topology location in the spatial dimension, analyzing the failure development speed in the time dimension, and comparing the sensor data patterns in the feature dimension. The fuzzy inference algorithm is used in the matching process to calculate the similarity between the current warning and historical cases. When the similarity exceeds 85%, the optimal disposal plan is automatically pushed. Otherwise, a decision tree based on reinforcement learning is started to generate a new strategy.

[0056] In order to generate intelligent decision results based on the retrieval and matching method of the expert experience database, a multi-objective optimization model can be introduced, and evaluation functions are established in three dimensions: maintenance cost, operation impact, and safety risk. The system is connected to the shift plan of the operation scheduling system and the maintenance resource database in real time, and the comprehensive benefits of different disposal plans are predicted through Monte Carlo simulation. The output of the decision result includes a timing plan diagram of the disposal steps, a list of required spare parts, and the expected recovery time, and a three-dimensional positioning view of the faulty component is displayed to the on-site personnel through an augmented reality terminal. The system synchronously establishes a decision effect feedback mechanism, analyzes the differences between the actual maintenance results and the predicted results, and automatically updates the case weights and association rules in the experience database to form a continuously optimized intelligent decision-making closed loop.

[0057] Finally, the above method of this embodiment further includes: S5. When the maglev vehicle system is actually operating, the remote control center continuously and real-time obtains operation data, evaluation data, potential fault data, warning information, and intelligent decision results based on the Internet of Things technology, classifies and stores all the obtained data, and correspondingly establishes a fault warning maintenance record; the expert experience database is updated using the fault warning maintenance record. Specifically, the remote control center uses the Hadoop distributed storage system to classify and store all the obtained data, and at the same time uses the Spark big data analysis platform technology to correspondingly establish a fault warning maintenance record after completing the data analysis.

[0058] In this embodiment, at the level of closed-loop management of system operation and maintenance, the remote control center constructs a full-life-cycle data management system based on a cloud-edge collaborative architecture. Through the communication base station group deployed along the line, the system can real-time converge the multi-source data streams uploaded by the vehicle-mounted edge computing nodes with a delay of 100ms level. After corresponding to various types of data in the method, that is, including original sensor readings, health assessment vectors, fault prediction probability distributions, and disposal decision plans, etc. The data access layer can use the Kafka message queue to realize the buffering and diversion of high-concurrency data streams, and set priority channels to ensure the instant transmission of warning information above level three.

[0059] In this embodiment, the storage architecture adopts the Hadoop ecosystem. The HDFS distributed file system divides the storage strategy according to data types: time-series operation data adopts Parquet columnar storage to optimize the compression ratio, unstructured maintenance records are stored in the HBase wide-table database, and knowledge graph data is associated and stored through the JanusGraph graph database.

[0060] In this embodiment, the data processing engine builds a hybrid computing paradigm based on Spark. The Structured Streaming module is enabled for window aggregation analysis of real-time data streams, and batch processing tasks are executed through the optimized Spark SQL. For the generation of fault warning maintenance records, the system designs a multi-dimensional feature extraction pipeline: first, the time stamps of abnormal events are marked through the time series pattern recognition algorithm, then the fault location analysis is carried out in combination with the device topology tree, and finally the effectiveness index of the disposal measures is extracted by associating with the maintenance work order system. Each maintenance record can contain up to 17 feature dimensions, covering key elements such as the fault evolution stage mark, disposal response time limit, and maintenance resource consumption. The storage format adopts the Avro serialization scheme to ensure cross-platform compatibility.

[0061] The expert experience library update mechanism adopts a combination of incremental learning and case-based reasoning. For example, the system starts the knowledge distillation job every day at midnight, matches the disposal solutions in the newly added maintenance records with historical cases, and automatically triggers the creation process of new knowledge nodes when the cosine similarity is lower than 0.75. The knowledge graph update module uses graph neural networks for relationship reasoning, generates the embedding vectors of device fault modes through the Node2Vec algorithm, and optimizes the relationship weights between entities using the TransE model. For new fault cases involving multi-system coupling, the system starts a semi-supervised learning process and constructs knowledge edges with confidence annotations in combination with the annotation feedback of operation and maintenance personnel.

[0062] In summary, when the present invention uses machine learning-related algorithms to perform state monitoring and analysis on the data of the maglev vehicle system, it not only relies on data-driven, but also combines expert experience, while ensuring the accuracy and real-time effect of fault warning.

Claims

1. A method for state monitoring and fault warning of a maglev vehicle system, characterized in that: The method comprises the following steps: S1, collecting the operating data of the maglev vehicle system and performing corresponding data preprocessing operations; S2. Based on the operation data, analyzing and calculating by constructing a mathematical model, evaluating the health status of the maglev vehicle system, and obtaining evaluation data; S3, constructing and training a fault prediction model, inputting the evaluation data into the trained fault prediction model, and obtaining potential fault data of the maglev vehicle system; S4. Generate warning information based on the potential fault data; The warning information is used to search the expert experience database, generate intelligent decision results and feed them back to the operation and maintenance personnel.

2. The condition monitoring and fault early warning method according to claim 1 is characterized in that: In step S1, the operation data generated by the maglev vehicle system during operation is collected at a preset frequency through the train control system TCMS, and the categories of the operation data include speed, acceleration, suspension gap, motor current and motor voltage.

3. The condition monitoring and fault early warning method according to claim 2 is characterized in that: The operating data is subjected to data preprocessing operations, including data cleaning, data denoising and data normalization; wherein data cleaning is implemented by a median filter, data denoising is implemented by a Kalman filter, and data normalization is implemented by a min-max normalization method.

4. The condition monitoring and fault early warning method according to claim 1 is characterized in that: In step S2, a corresponding health status assessment is performed on the linear motor and suspension performance of the magnetic levitation vehicle system; the categories of the mathematical models constructed during the assessment process include linear regression models, support vector machine models and neural network models. After the mathematical model is established and trained accordingly, its input is the operating data and its output is the evaluation data.

5. The condition monitoring and fault early warning method according to claim 4 is characterized in that: The establishment and training of the mathematical model specifically includes: pre-collecting historical data of the maglev vehicle system, dividing the historical data into a training set and a test set; extracting data features including speed, acceleration, suspension gap, motor current and motor voltage from the historical data; A linear regression model, a support vector machine model and a neural network model are respectively established, and the three types of models are trained respectively using the training set. The model parameters are adjusted accordingly during the training process, and the optimal evaluation data is selected and determined from the outputs of the three types of models. Subsequently, the performance of the three types of models is tested using the test set.

6. The condition monitoring and fault early warning method according to claim 1 is characterized in that: In step S3, the categories of the fault prediction model constructed include linear regression model, support vector machine model and neural network model. After the fault prediction model is established and trained accordingly, its input is the evaluation data and its output is the potential fault data.

7. The condition monitoring and fault early warning method according to claim 1 is characterized in that: In step S4, the generated warning information includes: the fault type, fault location and fault occurrence time of the potential fault corresponding to the potential fault data after it is transformed into an actual fault.

8. The condition monitoring and fault early warning method according to claim 1, characterized in that: In step S4, the expert experience database stores fault diagnosis rule information, fault repair strategy information and fault maintenance experience information; based on the warning information, the corresponding information pre-stored in the expert experience database is retrieved and matched to generate the intelligent decision result.

9. The condition monitoring and fault early warning method according to claim 1, characterized in that: The method also includes: S5, when the maglev vehicle system is actually operating, the remote control center continuously obtains the operating data, the evaluation data, the potential fault data, the warning information and the intelligent decision-making results in real time based on the Internet of Things technology, and classifies and stores all the acquired data, and establishes corresponding fault warning maintenance records; uses the fault warning maintenance records to update the expert experience library.

10. The condition monitoring and fault early warning method according to claim 9, characterized in that: The remote control center uses the Hadoop distributed storage system to classify and store all acquired data, and uses the Spark big data analysis platform technology to establish corresponding fault warning maintenance records after completing data analysis.

Images