A method and system for evaluating intelligent operation and maintenance of EMUs based on reliability engineering

Abstract

The present invention discloses a method and system for evaluating the intelligent operation and maintenance of multiple-unit trains based on credibility engineering. The method relates to the field of intelligent operation and maintenance technology and includes the following steps: acquiring multimodal data, preprocessing it, and generating multimodal data features; loading and dynamically adjusting a physical law constraint library, prioritizing the constraint rules in the physical law constraint library using an isolation forest algorithm, verifying the multimodal data features by priority, and screening out credible features; training and evaluating the physical credible feature dataset to generate an evaluation report on the operation status of the multiple-unit train; and utilizing the evaluation report on the operation status of the multiple-unit train to perform real-time monitoring of the multimodal data features and optimize operation and maintenance decisions. The present invention utilizes the isolation forest algorithm to perform anomaly detection on the multimodal data features, thereby achieving efficient identification of abnormal patterns in the operation of the multiple-unit train.

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation and maintenance technology, and in particular to a method and system for evaluating intelligent operation and maintenance of electric train sets based on credibility engineering. Background Art

[0002] With the rapid development of high-speed railway systems, the operational safety and reliability of EMUs have become critical factors in ensuring transportation efficiency and passenger safety. Traditional EMU operation and maintenance methods rely primarily on periodic inspections and threshold alarm mechanisms based on sensor data, making it difficult to fully perceive and dynamically assess the state of complex systems. In recent years, with the integration of technologies such as big data, artificial intelligence, and the Industrial Internet of Things, intelligent operation and maintenance systems have gradually become a research hotspot. Among them, multimodal data fusion technology is widely used in equipment condition monitoring, integrating information from multiple sensors to improve the accuracy and response speed of fault identification. At the same time, the combination of physical modeling and data-driven methods is also gaining increasing attention, aiming to enhance the interpretability and generalization capabilities of models.

[0003] However, there is still a significant problem in existing intelligent operation and maintenance technologies: due to the lack of effective constraints and screening of multimodal features at the level of physical laws, noise or abnormal features are easily introduced during the model training process, thereby affecting the credibility of the evaluation results; especially when faced with complex working conditions, traditional methods constrain and screen multimodal features by setting fixed thresholds and statistical-based anomaly detection techniques, such as using statistical quantities such as standard deviation and mean to identify outliers; however, these methods show limitations under complex working conditions and find it difficult to dynamically adapt to changing environments. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent operation and maintenance evaluation method for EMUs based on credibility engineering to solve the problem of lack of effective constraints and screening of multimodal features at the level of physical laws.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an intelligent operation and maintenance evaluation method for EMUs based on credibility engineering, which includes acquiring multimodal data, and preprocessing it to generate multimodal data features; the multimodal data features include vibration energy features, average temperature values, harmonic amplitudes, and voiceprint energy; loading and dynamically adjusting a physical law constraint library, using an isolation forest algorithm to prioritize the constraint rules in the physical law constraint library, verifying the multimodal data features according to the priority, screening out credible features, and calculating the credibility score based on the credible features to generate a physical credible feature data set; training and evaluating the physical credible feature data set to generate an evaluation report on the EMU operating status; using the evaluation report on the EMU operating status to monitor the multimodal data features in real time and optimize the operation and maintenance decisions to generate a real-time operation and maintenance decision report.

[0008] As an optimal solution for the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in the present invention, the physical law constraint library refers to the initial constraint rule set stored in the local file of the edge computing unit, which is used to verify whether the multimodal data characteristics conform to the EMU operation characteristics.

[0009] As a preferred solution of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in the present invention, wherein: the priority of the constraint rules in the physical law constraint library is sorted by using the isolation forest algorithm, specifically as follows:

[0010] Taking multimodal data features as multimodal samples, the abnormal patterns in the multimodal samples are detected through a tree structure, and an abnormal score for each multimodal sample is generated;

[0011] Based on the distribution characteristics of anomaly scores, a detection threshold is set to filter and mark multimodal samples with anomaly scores higher than the detection threshold as anomaly samples;

[0012] Compare the abnormal samples one by one with the constraint rule range in the adjusted physical law constraint library, and analyze the contribution of each constraint rule;

[0013] Assign priorities based on contribution and generate a prioritized physical law constraint library.

[0014] As a preferred solution of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in the present invention, the training and evaluation of the physical credibility feature data set are as follows:

[0015] The physical trustworthy feature dataset is defined as the operating state sample;

[0016] Divide the running status samples into training set and test set;

[0017] Performing weighted processing on the multimodal data features in the running state samples to generate weighted multimodal data features;

[0018] Construct multiple decision trees based on weighted multimodal data features;

[0019] Randomly extract a subset of running status samples and a subset of running status features from the training set;

[0020] The extracted running status sample subset and running status feature subset are combined into the training subset of each decision tree;

[0021] Through recursive segmentation, the running status samples in the training subset are classified to generate the classification rules of each decision tree;

[0022] Each decision tree makes predictions based on the classification rules to generate a single decision tree prediction result;

[0023] The prediction results of all decision trees are integrated through the majority voting method to generate the classification results of the training phase.

[0024] As a preferred solution of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in the present invention, the recursive segmentation refers to starting from the root node of the training subset, dividing the operating status samples into two sub-nodes, selecting multimodal data features and segmentation points for the sub-nodes, and completing the recursive segmentation through continuous repetition.

[0025] As a preferred solution of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in the present invention, the operation and maintenance decision optimization refers to using the evaluation report of the EMU operating status, defining the environment and intelligent agent, setting the reward function, and perceiving the operating status and selecting operation and maintenance actions through the interaction between the intelligent agent and the environment.

[0026] As a preferred solution of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in the present invention, the intelligent agent interacts with the environment to perceive the operating status and selects the operation and maintenance action, as follows:

[0027] Define the intelligent agent as the decision-making entity that selects operation and maintenance actions;

[0028] Define the environment as multimodal data features;

[0029] The agent perceives the current operating status of the EMU based on the multimodal data features of the current time window and selects an entity operation and maintenance action to perform;

[0030] The agent perceives the new operating status of the EMU based on the multimodal data features of the next time window, obtains rewards based on the reward function, adjusts the operation and maintenance strategy based on the reward value, and selects new entity operation and maintenance actions to perform;

[0031] The operation and maintenance strategy is an initial strategy set based on the evaluation report of the operation status.

[0032] In a second aspect, the present invention provides an EMU intelligent operation and maintenance evaluation system based on credibility engineering, comprising:

[0033] An acquisition module is used to acquire multimodal data, preprocess it, and generate multimodal data features; the multimodal data features include vibration energy features, average temperature values, harmonic amplitudes, and voiceprint energy;

[0034] The verification module is used to load and dynamically adjust the physical law constraint library, use the isolation forest algorithm to prioritize the constraint rules in the physical law constraint library, verify the multimodal data features according to the priority, filter out the credible features, and calculate the credibility score based on the credible features to generate a physical credible feature dataset;

[0035] The evaluation module is used to train and evaluate the physical trustworthy feature dataset and generate an evaluation report on the EMU operation status;

[0036] The optimization module is used to utilize the evaluation report of the EMU operation status to conduct real-time monitoring of multimodal data features and optimize operation and maintenance decisions, and generate real-time operation and maintenance decision reports.

[0037] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for intelligent operation and maintenance evaluation of EMUs based on credibility engineering as described in the first aspect of the present invention is implemented.

[0038] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the EMU intelligent operation and maintenance evaluation method based on credibility engineering as described in the first aspect of the present invention.

[0039] The beneficial effects of the present invention are as follows: By utilizing the isolation forest algorithm to perform anomaly detection on multimodal data features, the present invention achieves efficient identification of abnormal patterns in EMU operation. The isolation forest algorithm evaluates the anomaly scores of data points by constructing multiple tree structure models, accurately locating a small number of abnormal samples without the need for labels, making it particularly suitable for intelligent operation and maintenance under complex working conditions. It effectively distinguishes between normal fluctuations and potential fault characteristics, significantly reducing false alarm rates and missed detections. By assigning an anomaly score to each multimodal data feature and filtering out untrustworthy features, the accuracy of subsequent physical law constraint verification and credibility score calculation is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 The figure is a flow chart of the intelligent operation and maintenance evaluation method of EMU based on reliability engineering.

[0042] Figure 2 Schematic diagram of the EMU intelligent operation and maintenance evaluation system based on reliability engineering.

[0043] Figure 3 Flowchart for prioritizing constraint rules for Isolation Forest.

[0044] Figure 4 Training graph for the random forest model. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0048] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a method for evaluating the intelligent operation and maintenance of an EMU based on reliability engineering, comprising the following steps:

[0049] S1. Acquire multimodal data, preprocess it, and generate multimodal data features.

[0050] Furthermore, multimodal data is collected through a variety of sensors, including vibration acceleration, shaft temperature data, three-phase current data and sound pressure data;

[0051] Vibration sensors are installed on the axleboxes of the bogie wheelsets of the EMU to collect vibration acceleration (unit: square meters per second squared) to reflect the wheel-rail contact dynamics;

[0052] Install temperature sensors on the traction motor bearing housings of EMUs to collect shaft temperature data (unit: degrees Celsius) and monitor the motor thermal status;

[0053] Install the current sensor at the output end of the brake assembly motor of the EMU to collect three-phase current data (unit: ampere) and detect the electrical performance of the motor;

[0054] An acoustic sensor is installed on the surface of the gearbox housing of the EMU to collect sound pressure data (unit: Pascal) and capture the gear meshing soundprint;

[0055] The collected multimodal data is transmitted to the local file of the edge computing unit through the 5G communication network;

[0056] After receiving the multimodal data in the local file of the edge computing unit, high-frequency noise is removed through sliding window mean filtering. The sliding window size is set according to the vibration frequency of the EMU, for example, 5 milliseconds, retaining the signal details of vibration acceleration, axle temperature data, three-phase current data, and sound pressure data;

[0057] Missing multimodal data are supplemented by linear interpolation; sliding windows and linear interpolation are both common methods for data preprocessing;

[0058] The collected vibration acceleration, axle temperature data, three-phase current data, and sound pressure data are grouped according to the EMU operating conditions. Operating conditions refer to the operating status of the EMU, including acceleration, cruising, and braking. Based on the acceleration, cruising, and braking status of the EMU, the vibration acceleration, axle temperature data, three-phase current data, and sound pressure data are grouped into groups corresponding to the operating conditions. The multimodal data is divided using a fixed time length to generate multimodal data grouped by time window. The time length is set according to the operating condition switching frequency, for example, 3 minutes. At the same time, the collection timestamps of the vibration acceleration, axle temperature data, three-phase current data, and sound pressure data are synchronized to ensure time consistency.

[0059] Fast Fourier transform is used to extract vibration energy characteristics, harmonic amplitudes, and voiceprint energy from multimodal data, as follows:

[0060] The vibration acceleration is grouped into fixed time periods (e.g., 1 second) to generate a time domain signal. The time domain signal is converted into a frequency domain signal using a fast Fourier transform. The vibration frequency range associated with the EMU wheel-rail contact (e.g., the frequency range of 50-250 Hz) in the frequency domain signal is selected. The signal strength of the frequency domain signal within the vibration frequency range is summarized to obtain the vibration energy signature.

[0061] The three-phase current data is grouped into fixed time periods to generate time domain signals. Fast Fourier transform is applied to convert the time domain signals into frequency domain signals. The current frequency range related to the EMU motor characteristics is selected from the frequency domain signal (for example, the frequency range is 150 Hz). The current signal amplitude within the current frequency range is extracted to obtain the harmonic amplitude (for example, the third harmonic amplitude. Because the EMU motor current fundamental frequency is 50 Hz, the harmonic amplitude is exactly within the frequency range of 150 Hz, which can reflect problems such as winding asymmetry and magnetic field distortion).

[0062] The sound pressure data is grouped into fixed time periods to generate time domain signals. Fast Fourier transform is applied to convert the time domain signals into frequency domain signals. The sound pressure frequency range related to the EMU gear meshing in the frequency domain signal is selected (for example, the sound pressure frequency range is 800-1000 Hz). The root mean square value of the sound pressure data signal amplitude within the sound pressure frequency range is calculated to obtain the sound pressure root mean square value, which is the voiceprint energy. The root mean square value is expressed as:

[0063] ;

[0064] in, is the root mean square value, is the sound pressure data amplitude (unit: Pascal), which indicates the signal strength within the sound pressure frequency range. is the number of sound pressure data signal amplitudes within the sound pressure frequency range;

[0065] It should be noted that Fast Fourier Transform is a commonly used technology for data preprocessing and is used for feature extraction. It converts time domain signals into frequency domain signals to extract frequency features such as vibration energy characteristics, harmonic amplitudes, and voiceprint energy.

[0066] Extract the average temperature value from the axle temperature data of the EMU. Specifically, the axle temperature data is grouped by a fixed time period (for example, a time period of 1 minute), and the sum of all axle temperature data within the fixed time period is calculated and divided by the number of axle temperature data within the fixed time period to obtain the average temperature value.

[0067] The vibration energy characteristics, average temperature values, harmonic amplitudes, and root mean square values ​​of sound pressure are summarized as a multimodal feature vector (8 dimensions), which is then converted into a standardized multimodal feature matrix using the maximum-minimum normalization method and stored as a CSV file.

[0068] S2. Load and dynamically adjust the physical law constraint library, use the isolation forest algorithm to prioritize the constraint rules in the physical law constraint library, verify the multimodal data features according to the priority, filter out the credible features, and calculate the credibility score based on the credible features to generate a physical credible feature dataset.

[0069] Furthermore, the physical law constraint library is called from the local file of the edge computing unit as the initial constraint rules for verifying the multimodal data features. The initial constraint rules include the following:

[0070] Based on the relationship between wheel-rail contact force and vibration intensity in EMU wheel-rail mechanics, a vibration energy constraint rule is defined, stipulating that the vibration energy is within the wheel-rail contact force range (e.g., 10-500 square meters per second to the fourth power) to ensure that the vibration energy meets the wheel-rail contact characteristics during EMU operation;

[0071] Based on the heat balance characteristics of the motor during operation in the EMU heat conduction law, define the average temperature value constraint rule, stipulating that the average temperature value is within the motor operating temperature range (for example, 30-80 degrees Celsius) to ensure that the average temperature value conforms to the EMU motor thermal state;

[0072] Based on the relationship between current anomalies and harmonic amplitudes in EMU motor electromagnetics, a harmonic amplitude constraint rule is defined, stipulating that the harmonic amplitude is within the abnormal motor current range (e.g., 0-1 ampere) to ensure that the harmonic amplitude meets the electrical performance requirements of the EMU motor.

[0073] Based on the intensity characteristics of the gear meshing soundprint in the EMU acoustic law, a soundprint energy constraint rule is defined, stipulating that the soundprint energy is within the gear meshing soundprint range (for example, 0.1-5 Pascals) to ensure that the soundprint energy conforms to the EMU gear meshing characteristics;

[0074] Based on the loaded physical law constraint library, the constraint rule range in the physical law constraint library is dynamically adjusted according to the real-time operating conditions of the EMU (acceleration, cruising, braking) and the trends of the EMU's vibration, temperature, current, and sound pressure historical data during operation, including the following:

[0075] Based on the characteristics of EMU vibration and temperature rise under acceleration conditions, adjust the vibration energy range (e.g., 10-600 square meters per second to the fourth power), the average temperature range (e.g., 30-85 degrees Celsius), the harmonic amplitude range (e.g., 0-1.2 amps), and the soundprint energy range (e.g., 0.1-5.5 Pascals);

[0076] Based on the characteristics of stable operation of the EMU under cruising conditions, the vibration energy range (e.g., 10-550 square meters per second to the fourth power), average temperature range (e.g., 30-82 degrees Celsius), harmonic amplitude range (e.g., 0-1.1 amperes), and soundprint energy range (e.g., 0.1-5.2 Pascals) are adjusted;

[0077] Based on the characteristics of the EMU current and sound pressure changes under braking conditions, the vibration energy range (e.g., 10-520 square meters per second to the fourth power), average temperature range (e.g., 30-83 degrees Celsius), harmonic amplitude range (e.g., 0-1.3 amps), and soundprint energy range (e.g., 0.1-5.3 Pascals) are adjusted.

[0078] Based on the adjusted vibration energy, average temperature value, harmonic amplitude, and voiceprint energy constraint rule range, the physical law constraint library is updated to generate a dynamically adjusted physical law constraint library;

[0079] Extracting multimodal data features from the standardized multimodal feature matrix;

[0080] Multimodal data features are used as multimodal samples. The isolation forest algorithm is used to detect abnormal patterns in multimodal samples. The number of trees (e.g., 100 trees) and the abnormality ratio (e.g., 0.1) are set to generate an abnormality score for each multimodal sample (ranging from 0 to 1, e.g., 0.9 indicates a high probability of abnormality). A detection threshold (e.g., 0.8) is set based on the distribution characteristics of the abnormality score. Multimodal samples with abnormality scores above the detection threshold are screened out and marked as abnormal samples (e.g., for a multimodal sample with a vibration energy of 800 m2 / s to the fourth power, 100 abnormal samples are screened out).

[0081] By comparing the multimodal data features of abnormal samples with the range of each constraint rule in the adjusted physical law constraint library one by one, the contribution of each constraint rule to the abnormal sample is analyzed. The violations of each constraint rule in all abnormal samples are counted (for example, among 100 abnormal samples, the wheel-rail mechanics rule is violated 80 times, with an abnormal frequency of 0.8, and the heat conduction rule is violated 50 times, with an abnormal frequency of 0.5). The violation frequency is calculated (for example, the wheel-rail mechanics rule has a violation frequency of 0.8, and the heat conduction rule has a violation frequency of 0.5). The violation frequencies are then compared to identify the constraint rule with the greatest contribution (for example, the wheel-rail mechanics rule).

[0082] The constraints are prioritized according to their contribution (for example, the total priority weight is 1, the constraint rule with the highest contribution has a priority of 0.4, and the remaining priority of other constraints is 0.6). A prioritized physical law constraint library (JSON format) is generated and stored in the local file of the edge computing unit.

[0083] It should be noted that the isolation forest algorithm is a conventional technology in the field of anomaly detection. In the intelligent operation and maintenance evaluation of EMUs, it combines anomaly detection with rule optimization to improve the verification efficiency and accuracy of the intelligent operation and maintenance evaluation of EMUs.

[0084] Using the prioritized physical law constraint library, match and verify the multimodal data features in the standardized multimodal feature matrix one by one according to the priority from high to low, including the following:

[0085] Extract vibration energy from the standardized multimodal feature matrix, match the prioritized wheel-rail mechanical constraints according to the current EMU operating conditions, and check whether the vibration energy is within the adjusted vibration energy range. If the vibration energy is within the adjusted vibration energy range, mark it as a physically credible feature. If the vibration energy is not within the adjusted vibration energy range, set it to zero and trigger an alarm (for example, the alarm content is "vibration energy exceeds the limit").

[0086] Extract the average temperature value from the standardized multimodal feature matrix, match the prioritized heat conduction constraint rules according to the current EMU operating conditions, and check whether the average temperature value is within the adjusted average temperature range. If the average temperature value is within the adjusted average temperature range, mark it as a physically credible feature. If the average temperature value is not within the adjusted average temperature range, set it to zero and trigger an alarm (for example, the alarm content is "average temperature exceeds the limit").

[0087] Extract the harmonic amplitude from the standardized multimodal feature matrix, match the electromagnetic constraint rules after priority sorting based on the current EMU operating conditions, and check whether the harmonic amplitude is within the adjusted harmonic amplitude range. If the harmonic amplitude is within the adjusted harmonic amplitude range, mark it as a physically credible feature. If the harmonic amplitude is not within the adjusted harmonic amplitude range, set it to zero and trigger an alarm (for example, the alarm content is "harmonic amplitude exceeds the limit").

[0088] Extract voiceprint energy from the standardized multimodal feature matrix, match the prioritized acoustic constraint rules according to the current EMU operating conditions, and check whether the voiceprint energy is within the adjusted voiceprint energy range. If the voiceprint energy is within the adjusted voiceprint energy range, mark it as a physically trusted feature. If the voiceprint energy is not within the adjusted voiceprint energy range, set it to zero and trigger an alarm (for example, the alarm content is "voiceprint energy exceeds the limit").

[0089] From the standardized multimodal feature matrix after matching verification, filter and retain features marked as physically trustworthy (in the multimodal data features) and set features not marked as physically trustworthy (in the multimodal data features) to zero, generating a verified standardized multimodal feature matrix. Zeroed features represent untrustworthy features. Retaining zero values ​​maintains the structure of the standardized multimodal feature matrix and conveys untrustworthy information to the subsequent intelligent assessment and operation model steps.

[0090] Based on each multimodal data feature in the standardized multimodal feature matrix after verification, the credibility score of each multimodal data feature is calculated using SHAP (added explanation value), and the expression is:

[0091] ;

[0092] in, is the credibility score, which is used to measure the credibility of the multimodal data features in the standardized multimodal feature matrix. is the eigenvalue of the verified multimodal data, for example, the vibration energy value is the fourth power of 400 square meters per second, is the center value of the constraint range, calculated by the upper and lower limits of the constraint range. is the width of the constraint range, which is obtained by subtracting the lower limit from the upper limit of the constraint range;

[0093] It should be noted that this formula is based on the deviation measurement between multimodal data features and the scope of constraint rules, combined with the idea of ​​normalization, to calculate the credibility score of multimodal data features. Among them, SHAP is an interpretation method based on game theory, which is used to calculate the contribution value of multimodal data features to evaluate the credibility of multimodal data features.

[0094] Generate a verification report (JSON format, including the vibration energy verification results, average temperature verification results, harmonic amplitude verification results, voiceprint energy verification results, credibility score, and abnormality reasons) based on the verification results and credibility score of each multimodal data feature in the standardized multimodal feature matrix. For example, "Vibration energy passed, credibility 0.84"

[0095] The verified standardized multimodal feature matrix and the credibility score of each multimodal feature are stored together as a physical credibility feature dataset (CSV format) and stored in a local file of the edge computing unit.

[0096] S3. Train and evaluate the physical trusted feature dataset to generate an evaluation report on the EMU operating status.

[0097] Furthermore, a physical trust feature dataset is loaded from a local file of the edge computing unit, which contains the verified normalized multimodal feature matrix and the trustworthiness score of the multimodal data features;

[0098] The physical trustworthy feature dataset is defined as the operating state samples (including multimodal data features and trustworthiness scores);

[0099] Divide the running status samples into training and test sets according to the time window (for example, in a ratio of 8:2);

[0100] Use the random forest model as the basic framework of the intelligent evaluation and operation model, and set the basic parameters of the intelligent evaluation and operation model (for example, the number of decision trees is 100);

[0101] Use the training set as input for the intelligent evaluation and maintenance model;

[0102] Use the credibility score to weight multimodal data features (for example, a vibration energy feature with a credibility score of 0.84 is given a high weight, while a voiceprint energy feature with a credibility score of 0.65 is given a low weight). This generates weighted multimodal data features to increase the focus of the intelligent assessment and operation model on high-credibility feature values.

[0103] Based on the weighted multimodal data features, the intelligent evaluation and operation model is used to construct multiple decision trees (for example, 100 trees) for training, as follows:

[0104] For each decision tree (e.g., the first tree), a subset of operating status samples is randomly extracted from the training set using the Bootstrap sampling method (e.g., 70% of the operating status samples, with repeated sampling allowed). A subset of operating status features is also randomly selected (e.g., 50% of multimodal data features, such as vibration energy and average temperature). The extracted subset of operating status samples and the subset of operating status features are combined to form the training subset of the current decision tree to ensure diversity in each decision tree. The Bootstrap sampling method is a statistical resampling technique proposed by Efron in 1979 and is widely used in machine learning and data analysis, particularly in ensemble learning models such as random forests.

[0105] Starting from the root node of the training subset, which contains all extracted operating status samples, select a multimodal data feature (such as the vibration energy eigenvalue) and a split point (such as whether the vibration energy is greater than the fourth power of 300 square meters per second) to divide the operating status samples in the root node into two subnodes (for example, the "greater than 300" subnode contains operating status samples with vibration energy eigenvalues ​​greater than the fourth power of 300 square meters per second, and the "less than or equal to 300" subnode contains operating status samples with vibration energy eigenvalues ​​less than or equal to the fourth power of 300 square meters per second). For the "greater than 300" subnode, continue to select another multimodal data feature (such as the average temperature eigenvalue) and split point (such as whether the average temperature is greater than 50 degrees Celsius) to divide the operating status samples of the "greater than 300" subnode into smaller subnodes (for example, the "greater than 50 degrees Celsius" subnode and the "less than or equal to 50 degrees Celsius" subnode).

[0106] Repeat the segmentation process until the stopping condition is met (for example, the maximum depth is reached, such as 5 layers, or the purity of the child node running status samples is high enough, such as 90% of the samples in the child node belong to the "normal" category);

[0107] Finally, through recursive partitioning of the random forest model, classification rules from root nodes to leaf nodes are generated (for example, if the vibration energy eigenvalue is greater than the fourth power of 300 square meters per second and the average temperature eigenvalue is greater than 50 degrees Celsius, then the prediction is "abnormal");

[0108] After the decision tree training is completed, each operating status sample in the training subset is predicted according to the generated classification rules (for example, if the vibration energy is greater than 300 and the average temperature is greater than 50, it is predicted to be abnormal), and the independent prediction result of the current decision tree is obtained (for example, the first tree predicts normal, and the second tree predicts abnormal);

[0109] Repeat the decision tree training process, repeat the above sampling, segmentation, and prediction steps for each decision tree (for example, 100 trees) to generate independent prediction results for each decision tree;

[0110] The prediction results of all decision trees are integrated through majority voting to generate the classification results of the training phase (for example, if 70 out of 100 trees predict "normal", 20 predict "abnormal", and 10 predict "potential fault", the classification result of the training phase is "normal");

[0111] Based on the classification results of the training phase, complete the training of the intelligent evaluation and operation model;

[0112] Use the operating status samples in the test set as input to the intelligent evaluation and maintenance model;

[0113] The trained intelligent evaluation and operation model is used to predict the test set. Each decision tree independently predicts the EMU operating status of the operating status sample (such as "normal", "abnormal" and "potential failure"), and the classification result (i.e., the predicted category) in the test phase is generated through majority voting.

[0114] Based on the classification results in the test phase, the evaluation index (for example, the accuracy is 0.92) is calculated to evaluate the performance of the intelligent evaluation and operation model. The evaluation index expression is:

[0115] ;

[0116] in, It is an indicator for evaluating the performance of the intelligent evaluation and operation maintenance model. It represents the accuracy and measures the classification accuracy of the intelligent evaluation and operation maintenance model on the test set. The range is 0 to 1. The number of samples whose predicted category is consistent with the actual category and whose category is normal, The number of running status samples whose predicted categories are consistent with the actual categories and whose categories are abnormal or potential faults, The number of running status samples whose predicted category is normal but whose actual category is abnormal or potential failure, The number of running status samples that were predicted to be abnormal or potential fault but were actually normal; when the actual category is the collected multimodal data, the running status label (e.g., "normal" and "abnormal") provided by the EMU operation monitoring;

[0117] If the classification results of the test phase indicate that the EMU operating state is abnormal (for example, a potential fault caused by abnormal vibration energy characteristic values ​​of the operating state sample), a warning message is generated (for example, the warning content is "the EMU operating state is abnormal, it is recommended to check the vibration energy");

[0118] If the performance of the intelligent evaluation and operation maintenance model does not meet expectations (for example, the accuracy is lower than 0.9), adjust the basic parameters of the intelligent evaluation and operation maintenance model (for example, increase the number of random forest trees), return to the training step to retrain the intelligent evaluation and operation maintenance model, and re-evaluate until the performance requirements are met;

[0119] Based on the classification results of the intelligent evaluation and operation and maintenance model on the test set, an evaluation report on the EMU's operating status is generated (including the classification results of the test phase and operation and maintenance recommendations (such as "recommendation to inspect the EMU's wheel-rail contact components")), providing support for EMU intelligent operation and maintenance;

[0120] The evaluation report of the EMU operating status is stored in the local file of the edge computing unit to provide support for subsequent operation and maintenance decisions and real-time monitoring.

[0121] S4. Utilize the evaluation report of the EMU operation status to conduct real-time monitoring of multimodal data features and optimize operation and maintenance decisions, and generate a real-time operation and maintenance decision report.

[0122] Furthermore, the evaluation report of the EMU operation status is loaded from the local file of the edge computing unit;

[0123] The reinforcement learning algorithm uses the EMU operation status evaluation report to dynamically adjust the EMU operation and maintenance status, including the following:

[0124] Define the environment of the reinforcement learning algorithm as multimodal data features;

[0125] Define a reinforcement learning agent that selects a decision-making entity for operation and maintenance actions based on the current state of the EMU (including "continue monitoring," "check vibration-related components," "check temperature-related components," and "emergency stop").

[0126] Set the reward function as the initialization basic parameter of the reinforcement learning algorithm to evaluate the effectiveness of the agent's operation and maintenance actions;

[0127] Based on the evaluation report of the operating status, an initial operation and maintenance strategy is set. For example, if the evaluation report recommends "If the vibration energy exceeds the limit, check the vibration-related components", the initial strategy of the intelligent agent will prioritize the "check vibration-related components" action as the starting point for optimization;

[0128] The reinforcement learning agent interacts with the environment to monitor multimodal data features in real time and optimize the operation and maintenance strategy. Specifically, for the multimodal data features in the current time window, the agent perceives the current operating state (for example, the vibration energy characteristic value is 480 square meters per second to the fourth power, and the average temperature characteristic value is 72 degrees Celsius); the agent selects an entity operation and maintenance action based on the current strategy (for example, the optimal operation and maintenance action is "check vibration-related components") and executes the entity operation and maintenance action; then, for the multimodal data features in the next time window, the agent perceives the new operating state (for example, the operating state changes to "normal") and obtains rewards based on the reward function (for example, the reward for the operating state changing from "abnormal" to "normal" is +10, the reward for no improvement or deterioration is -5, the reward for maintaining "normal" is +2, and the reward for other cases is 0), and optimizes the operation and maintenance strategy through the reward value; and the interaction process is repeated continuously (for example, 1,000 times), using the agent to select the entity operation and maintenance action that can maximize the long-term reward;

[0129] Generate a real-time operation and maintenance decision report based on the operation and maintenance actions selected by the reinforcement learning agent and the changes in the EMU's operating status. The real-time operation and maintenance decision report records the operation and maintenance actions selected by the agent, the changes in the EMU's operating status, and the reward value, and is stored in JSON format;

[0130] If the EMU's operating status is abnormal or has a potential fault, the real-time operation and maintenance decision report will include a detailed analysis of the cause of the abnormality and operation and maintenance recommendations (for example, "it is recommended to inspect the EMU's wheel-rail contact components"), providing guidance for real-time EMU operation and maintenance.

[0131] The real-time operation and maintenance decision report is stored in the local file of the edge computing unit to complete the intelligent operation and maintenance of the EMU.

[0132] This embodiment also provides an EMU intelligent operation and maintenance evaluation system based on reliability engineering, including:

[0133] The acquisition module is used to acquire multimodal data, preprocess it, and generate multimodal data features; multimodal data features include vibration energy features, average temperature value, harmonic amplitude, and voiceprint energy;

[0134] The verification module is used to load and dynamically adjust the physical law constraint library, use the isolation forest algorithm to prioritize the constraint rules in the physical law constraint library, verify the multimodal data features according to the priority, filter out the credible features, and calculate the credibility score based on the credible features to generate a physical credible feature dataset;

[0135] The evaluation module is used to train and evaluate the physical trustworthy feature dataset and generate an evaluation report on the EMU operation status;

[0136] The optimization module is used to utilize the evaluation report of the EMU operation status to conduct real-time monitoring of multimodal data features and optimize operation and maintenance decisions, and generate real-time operation and maintenance decision reports.

[0137] This embodiment also provides a computer device, which is suitable for the case of an EMU intelligent operation and maintenance evaluation method based on credibility engineering, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the EMU intelligent operation and maintenance evaluation method based on credibility engineering proposed in the above embodiment.

[0138] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0139] This embodiment further provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the method for implementing intelligent operation and maintenance assessment of an EMU based on reliability engineering as proposed in the above embodiment is implemented. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0140] In summary, the present invention utilizes the isolation forest algorithm to perform anomaly detection on multimodal data features, achieving efficient identification of abnormal patterns in EMU operation. The isolation forest algorithm constructs multiple tree-structured models to evaluate the anomaly scores of data points, accurately locating a small number of abnormal samples without the need for labels. It is particularly suitable for intelligent operation and maintenance under complex operating conditions. It effectively distinguishes between normal fluctuations and potential fault characteristics, significantly reducing false alarm rates and missed detections. By assigning an anomaly score to each multimodal data feature and filtering out untrustworthy features, the accuracy of subsequent physical law constraint verification and credibility score calculations is improved.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for evaluating intelligent operation and maintenance of EMUs based on reliability engineering, characterized by: include, Acquire multimodal data, preprocess it, and generate multimodal data features; the multimodal data features include vibration energy features, average temperature values, harmonic amplitudes, and voiceprint energy; Load and dynamically adjust the physical law constraint library, use the isolation forest algorithm to prioritize the constraint rules in the physical law constraint library, verify the multimodal data features according to the priority, filter out the credible features, and calculate the credibility score based on the credible features to generate a physical credible feature dataset; Train and evaluate the physical trustworthy feature dataset to generate an assessment report on the EMU's operating status; Using the EMU operating status assessment report, multimodal data features are monitored in real time, operation and maintenance decision optimization is performed, and a real-time operation and maintenance decision report is generated; The physical law constraint library is called as the initial constraint rules for verifying the multimodal data features. The initial constraint rules include the following: Based on the relationship between wheel-rail contact force and vibration intensity in EMU wheel-rail mechanics, a vibration energy constraint rule is defined, which stipulates that the vibration energy is within the range of wheel-rail contact force. According to the heat balance characteristics of the motor during operation in the EMU heat conduction law, the average temperature value constraint rule is defined to stipulate that the average temperature value is within the motor operating temperature range; Based on the relationship between current anomaly and harmonic amplitude in EMU motor electromagnetics, a harmonic amplitude constraint rule is defined to stipulate that the harmonic amplitude is within the abnormal range of motor current; According to the intensity characteristics of gear meshing soundprint in the EMU acoustic law, the soundprint energy constraint rule is defined, which stipulates that the soundprint energy is within the gear meshing soundprint range; Based on the loaded physical law constraint library, dynamically adjust the constraint rule range in the physical law constraint library; Based on the adjusted vibration energy, average temperature value, harmonic amplitude, and voiceprint energy constraint rule range, the physical law constraint library is updated to generate a dynamically adjusted physical law constraint library; The isolation forest algorithm is used to prioritize the constraint rules in the physical law constraint library, as follows: Taking multimodal data features as multimodal samples, the abnormal patterns in the multimodal samples are detected through a tree structure, and an abnormal score for each multimodal sample is generated; Based on the distribution characteristics of anomaly scores, a detection threshold is set to filter and mark multimodal samples with anomaly scores higher than the detection threshold as anomaly samples; Compare the abnormal samples one by one with the constraint rule range in the adjusted physical law constraint library, and analyze the contribution of each constraint rule; Assign priorities based on contribution and generate a prioritized physical law constraint library.

2. The method for intelligent operation and maintenance evaluation of EMUs based on reliability engineering according to claim 1, characterized in that: The physical law constraint library refers to the initial constraint rule set stored in the local file of the edge computing unit, which is used to verify whether the multimodal data characteristics conform to the EMU operation characteristics.

3. The method for intelligent operation and maintenance evaluation of EMUs based on reliability engineering according to claim 1, characterized in that: The training and evaluation of the physical trustworthy feature dataset are as follows: The physical trustworthy feature dataset is defined as the operating state sample; Divide the running status samples into training set and test set; Performing weighted processing on the multimodal data features in the running state samples to generate weighted multimodal data features; Construct multiple decision trees based on weighted multimodal data features; Randomly extract a subset of running status samples and a subset of running status features from the training set; The extracted running status sample subset and running status feature subset are combined into the training subset of each decision tree; Through recursive segmentation, the running status samples in the training subset are classified to generate the classification rules of each decision tree; Each decision tree makes predictions based on the classification rules to generate a single decision tree prediction result; The prediction results of all decision trees are integrated through the majority voting method to generate the classification results of the training phase.

4. The method for intelligent operation and maintenance assessment of EMUs based on reliability engineering according to claim 3 is characterized in that: The recursive segmentation refers to starting from the root node of the training subset, dividing the running state samples into two sub-nodes, selecting multimodal data features and segmentation points for the sub-nodes, and completing the recursive segmentation through continuous repetition.

5. The method for intelligent operation and maintenance evaluation of EMUs based on reliability engineering according to claim 1, characterized in that: The operation and maintenance decision optimization refers to using the evaluation report of the EMU operation status, defining the environment and intelligent agent, setting the reward function, and perceiving the operation status and selecting operation and maintenance actions through the interaction between the intelligent agent and the environment.

6. The method for intelligent operation and maintenance evaluation of EMUs based on reliability engineering according to claim 5 is characterized in that: The agent interacts with the environment to perceive the operating status and selects operation and maintenance actions, as follows: Define the intelligent agent as the decision-making entity that selects operation and maintenance actions; Define the environment as multimodal data features; The agent perceives the current operating status of the EMU based on the multimodal data features of the current time window and selects an entity operation and maintenance action to perform; The agent perceives the new operating status of the EMU based on the multimodal data features of the next time window, obtains rewards based on the reward function, adjusts the operation and maintenance strategy based on the reward value, and selects new entity operation and maintenance actions to perform; The operation and maintenance strategy is an initial strategy set based on the evaluation report of the operation status.

7. A train intelligent operation and maintenance evaluation system based on credibility engineering, based on the train intelligent operation and maintenance evaluation method based on credibility engineering according to any one of claims 1 to 6, characterized in that: include, An acquisition module is used to acquire multimodal data, preprocess it, and generate multimodal data features; the multimodal data features include vibration energy features, average temperature values, harmonic amplitudes, and voiceprint energy; The verification module is used to load and dynamically adjust the physical law constraint library, use the isolation forest algorithm to prioritize the constraint rules in the physical law constraint library, verify the multimodal data features according to the priority, filter out the credible features, and calculate the credibility score based on the credible features to generate a physical credible feature dataset; The evaluation module is used to train and evaluate the physical trustworthy feature dataset and generate an evaluation report on the EMU operation status; The optimization module is used to utilize the evaluation report of the EMU operation status to conduct real-time monitoring of multimodal data features and optimize operation and maintenance decisions, and generate real-time operation and maintenance decision reports.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the EMU intelligent operation and maintenance evaluation method based on credibility engineering described in any one of claims 1 to 6 are implemented.

Images