Fault prediction method based on railway vehicle test data

Abstract

The invention relates to the technical field of rail vehicle test data processing, and discloses a rail vehicle test data-based fault prediction method, which comprises the following steps of: in a rail vehicle test process, acquiring dynamic monitoring data, static configuration data and environmental parameter data in real time as multi-source input data, performing standardization processing on the multi-source input data, and performing fault prediction on the multi-source input data; the method comprises the following steps: respectively extracting a state feature vector, a configuration feature set and an environment feature sequence, carrying out cross-dimension correlation analysis through a time sequence correlation model to generate a joint feature group, identifying an abnormal state, a potential fault point and a trend evolution path based on the joint feature group, forming a fault early warning information set, and finally carrying out hierarchical organization according to a time sequence and a serious grade. And generating a fault prediction report which covers the whole test period and is dynamically updated along with the test process. According to the method, the comprehensiveness and timeliness of fault prediction are improved through multi-source data fusion and time sequence correlation analysis, and effective support is provided for safe operation and research and development of railway vehicles.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rail vehicle test data processing, in particular to a fault prediction method based on rail vehicle test data. BACKGROUND

[0002] In the development and testing process of rail vehicles, accurate fault prediction is crucial for ensuring vehicle operation safety and improving development efficiency. Traditional rail vehicle fault prediction methods have many shortcomings in practical application. Most methods rely only on a single type of monitoring data, such as focusing only on dynamic monitoring data while ignoring static configuration data and environmental parameter data, resulting in an incomplete assessment of the vehicle's state and difficulty in accurately capturing potential fault risks under the combined action of multiple factors.

[0003] Existing data processing methods lack systematization and standardization, the time synchronization processing of dynamic monitoring data is not accurate enough, the version checking and missing field supplement of static configuration data are not timely, and the interference value filtering in environmental parameter data is not thorough, all of which greatly reduce the usability and reliability of the data. In addition, in the feature extraction and correlation analysis stage, traditional methods do not fully consider the time sequence correlation between different dimensional data, and cannot effectively mine the internal relationship between dynamic state, configuration parameter and environmental condition, resulting in a lag in identifying abnormal states and difficulty in accurately predicting potential fault points and trend evolution paths.

[0004] Existing fault prediction reports often cannot be dynamically updated according to the test progress, and cannot reflect the state changes of the vehicle in different test stages in real time, limiting the timeliness and practicality of the prediction results. With the continuous development of rail vehicle technology, higher requirements are put forward for the accuracy and real-time performance of test data processing and fault prediction, and there is an urgent need for a fault prediction method that can comprehensively utilize multi-source data, perform standardized processing and cross-dimensional correlation analysis, and dynamically update the prediction content. SUMMARY

[0005] The purpose of the present application is to provide a fault prediction method based on rail vehicle test data to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides a fault prediction method based on rail vehicle test data, which comprises: During the test process of the rail vehicle, real-time acquisition of dynamic monitoring data, static configuration data and environmental parameter data in the test stage as multi-source input data; Standardized processing of the acquired multi-source input data, wherein the dynamic monitoring data is labeled with a time stamp and calibrated for sampling frequency, the static configuration data is checked for version information and supplemented with missing fields, and the environmental parameter data is filtered for interference values and recorded for measurement location; The normalized dynamic monitoring data, static configuration data and environmental parameter data are respectively subjected to parameter extraction to form a state feature vector, a configuration feature set and an environmental feature sequence; The state feature vector, the configuration feature set and the environmental feature sequence are subjected to cross-dimension correlation analysis through a time sequence correlation model to generate a joint feature group containing time correlation; Based on the joint feature group, abnormal states, potential fault points and trend evolution paths in the test process are identified to form a fault early warning information set; According to the time sequence and severity level of the fault early warning information set, hierarchical organization is performed to generate a fault prediction report covering the whole test cycle, and the prediction content is dynamically updated with the test progress.

[0007] Preferably, the obtained multi-source input data is subjected to normalization processing, and the specific steps are as follows: The dynamic monitoring data is aligned to a unified time axis by a time synchronization module through different sensor sampling points to generate sequence data with continuous time stamps; The static configuration data is compared with historical archive records by a version checking module to correct error fields and supplement unrecorded component model information; The environmental parameter data is filtered by a threshold filtering module to eliminate abnormal values outside the normal working range of the device, and the remaining data is labeled with the specific installation position of the temperature and humidity sensor or the barometer.

[0008] Preferably, the normalized dynamic monitoring data, static configuration data and environmental parameter data are respectively subjected to parameter extraction, and the specific steps are as follows: The dynamic monitoring data is subjected to sliding window algorithm to intercept continuous data segments of a fixed time length, and the mean and variance of each data segment are calculated as the state feature vector; The static configuration data is filtered by a field extraction module to select key parameters related to faults, including component life, rated load and maintenance cycle, to form a configuration feature set; The environmental parameter data is subjected to trend analysis module to calculate the change rate per unit time, and the temperature and humidity fluctuation value and the air pressure change amount are extracted as the environmental feature sequence.

[0009] Preferably, the state feature vector, the configuration feature set and the environmental feature sequence are subjected to cross-dimension correlation analysis through a time sequence correlation model, and the specific steps are as follows: A time anchor mechanism is established to align the state feature vector, the configuration feature set and the environmental feature sequence to the test stage division nodes according to the time stamp; The correlation strength between different dimension features is calculated by a weight distribution module, and the feature information of the dynamic state, the corresponding configuration parameter and the environmental condition in the same test stage is highlighted; The associated features are dimensionally fused to generate a joint feature group containing time node, state dimension and influence dimension information.

[0010] Preferably, the joint feature group is used to identify abnormal states, potential fault points and trend evolution paths during the test process, and the specific steps are as follows: The state classification module is used to perform pattern clustering on the joint feature group to determine the normal state interval and abnormal state boundary of each stage of the test; In the normal state interval of each stage, the threshold detection module is used to identify feature values exceeding twice the standard deviation of the historical mean as potential fault points; The trend deduction module is used to analyze the change direction of continuous abnormal features to extract the evolution path that may cause faults.

[0011] Preferably, the fault warning information set is hierarchically organized according to the time sequence and severity level, and the specific steps are as follows: The abnormal states are arranged in chronological order according to the test process, and the corresponding test stage and time node are listed under each abnormal state; In each abnormal state, the potential fault points are sorted according to their occurrence frequency and deviation degree, and the fault points that repeatedly occur or deviate from the normal interval are prioritized; The trend evolution path is listed separately, and the key components that may be affected and the expected development time are marked to form a hierarchical information list with clear levels.

[0012] Preferably, a fault prediction report covering the entire test cycle is generated and dynamically updated, and the specific steps are as follows: The hierarchical information list is converted into a structured text, and the content is organized using the "time-stage-exception-trend" expression logic; When the test enters a new test stage, repeat the steps of multi-source data acquisition, normalization processing, parameter extraction, correlation analysis and fault identification to obtain new warning information; Insert the new warning information into the corresponding stage or trend path position, adjust the structure and content of the original report, and keep the prediction synchronized with the test process.

[0013] Preferably, when real-time acquisition of dynamic monitoring data, static configuration data and environmental parameter data is performed, the specific acquisition steps are as follows: Dynamic monitoring data is collected by sensor arrays installed on bogies and traction systems, and a filtering algorithm is used to suppress electromagnetic interference noise; Static configuration data is real-time retrieved through the vehicle management system interface, including component number, factory date and technical specification parameters, forming a multi-type multi-source input data set; The environmental parameter data are collected by an environmental monitoring station arranged at the test site, and the temperature, humidity and atmospheric pressure data of the test area are recorded synchronously.

[0014] Preferably, when the weight distribution module calculates the correlation strength between different dimensional features, the specific calculation steps are as follows: Taking the test stage node as the reference, the state feature vector, the configuration feature set and the environmental feature sequence in the same stage are combined to form a feature cluster; The matching degree of the dynamic state and the configuration parameter is calculated for each cluster feature; The correlation between the environmental condition and the state change is calculated; The correlation strength is distributed according to the matching degree and the correlation result, and the higher the strength is, the closer the information correlation between the dimensions is.

[0015] Preferably, when the dynamic monitoring data is aligned to the unified time axis by the time synchronization module, the specific alignment steps are as follows: The dynamic monitoring data is divided into independent data streams according to the sensor type, and the time stamp verification is performed on each data stream to identify the sampling start and end deviation position; The data stream with time deviation is subjected to interpolation processing, and the monitoring value of the missing time point is supplemented by the linear interpolation algorithm; The time axis of the interpolated data stream is rearranged, and the time identification of all data streams is unified based on the main sensor time stamp as the reference, serving as the sequence data with continuous time stamp.

[0016] Compared with the prior art, the beneficial effects of the present application are: By acquiring the dynamic monitoring data, the static configuration data and the environmental parameter data as multi-source input data in real time, all kinds of key information in the vehicle test process can be covered comprehensively, and the limitation of a single data source is avoided, thereby laying a solid data foundation for accurate fault prediction. The multi-source input data are subjected to standardized processing, including labeling time stamp and calibrating sampling frequency of the dynamic monitoring data, checking version information and supplementing missing fields of the static configuration data, filtering interference values and recording measurement positions of the environmental parameter data, so that the quality and reliability of the data are effectively improved, and the subsequent analysis is ensured to be based on accurate and complete data. The standardized data are subjected to parameter extraction respectively to form the state feature vector, the configuration feature set and the environmental feature sequence, so that the effective conversion from the original data to the key features is realized, and the important information related to the fault is highlighted.

[0017] The cross-dimension correlation analysis of the features through the time sequence correlation model generates a joint feature group containing time correlation, which can deeply mine the internal relationship between different dimension data, accurately grasp the change rule and mutual influence of the data in the time dimension, and more accurately identify the abnormal state, potential fault point and trend evolution path in the test process. The fault warning information set formed based on the joint feature group can comprehensively and timely reflect the potential fault risk of the vehicle, and provide a strong basis for fault prevention and processing.

[0018] According to the time sequence and severity level of the fault warning information, a fault prediction report covering the whole test cycle is generated, and the prediction content is dynamically updated with the test process, so that the timeliness and practicability of the prediction result are ensured, accurate fault prediction information can be provided for the test personnel in real time, and the test personnel can adjust the test scheme in time, and the test efficiency and vehicle research and development quality are improved. The method significantly improves the accuracy, real-time and comprehensiveness of the fault prediction of the railway vehicle through multi-source data fusion, standardized processing, time sequence correlation analysis and dynamic updating mechanism, and provides strong technical support for the safe operation and efficient research and development of the railway vehicle. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A working principle diagram of the fault prediction method based on test data of a railway vehicle is described. Figure 2 A flowchart of standardized processing of multi-source data is described. Figure 3 A flowchart of multi-source data parameter extraction is described. Figure 4 A flowchart of dynamic updating of the prediction report is described. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0021] Please refer to Figures 1-4 The present application provides a fault prediction method based on test data of a railway vehicle, which comprises.

[0022] During the test of the rail vehicle, dynamic monitoring data is collected by a sensor array installed on the bogie and traction system, and a filtering algorithm is used to suppress electromagnetic interference noise during collection; static configuration data is real-time accessed through a vehicle management system interface, which contains component number, factory date, technical specification parameters, etc.; environmental parameter data is collected by an environmental monitoring station arranged in the test site, and temperature, humidity and atmospheric pressure data in the test area are recorded synchronously, which are used as multi-source input data.

[0023] The obtained multi-source input data is normalized. The dynamic monitoring data is aligned to a unified time axis by a time synchronization module, specifically: the dynamic monitoring data is divided into independent data streams according to sensor type, time stamp verification is performed on each data stream, the start and end deviation positions of sampling are identified, interpolation processing is performed on the data stream with time deviation, the monitoring values of missing time points are supplemented by linear interpolation algorithm, the time axis of the interpolated data stream is rearranged, and the time identifiers of all data streams are unified based on the main sensor time stamp to generate sequence data with continuous time stamp; the static configuration data is compared with historical archive records by a version verification module to correct error fields and supplement unrecorded component model information; the environmental parameter data is filtered by a threshold filtering module to remove abnormal values outside the normal working range of the device, and the remaining data is labeled with the specific installation position of the temperature and humidity sensor or barometer.

[0024] The normalized dynamic monitoring data, static configuration data and environmental parameter data are respectively subjected to parameter extraction. The dynamic monitoring data is subjected to sliding window algorithm to intercept continuous data segments of fixed length, and the mean and variance of each data segment are calculated as state feature vectors; the static configuration data is subjected to field extraction module to screen key parameters related to faults, including component life, rated load and maintenance period, to form a configuration feature set; the environmental parameter data is subjected to trend analysis module to calculate the change rate per unit time, and the temperature and humidity fluctuation value and air pressure change amount are extracted as environmental feature sequence.

[0025] The state feature vector, configuration feature set and environmental feature sequence are subjected to cross-dimensional correlation analysis by a time series correlation model. A time anchor mechanism is established, and the state feature vector, configuration feature set and environmental feature sequence are aligned to test phase division nodes according to time stamp; the correlation strength between different dimensional features is calculated by a weight distribution module, specifically: based on the test phase node, the state feature vector, configuration feature set and environmental feature sequence in the same phase are combined into a feature cluster, the matching degree of dynamic state and configuration parameter is calculated for each cluster feature, the correlation between environmental condition and state change is calculated, and the correlation strength is distributed according to the matching degree and correlation result, and the higher the strength, the closer the information correlation between dimensions; the correlated features are dimensionally fused to generate a joint feature group containing time node, state dimension and influence dimension information.

[0026] The abnormal state, potential fault point and trend evolution path in the test process are identified based on the joint feature group. The state classification module is used for pattern clustering of the joint feature group to determine the normal state interval and abnormal state boundary of each stage of the test. In the normal state interval of each stage, the threshold detection module is used to identify the feature value exceeding twice the standard deviation of the historical mean as a potential fault point. The trend deduction module is used to analyze the change direction of the continuous abnormal features to extract the evolution path that may cause the fault and form a fault warning information set.

[0027] The fault warning information set is hierarchically organized according to the time sequence and severity level. The abnormal states are arranged in the time sequence of the test process, and the corresponding test stage and time node are listed under each abnormal state. In each abnormal state, the potential fault points are sorted according to the occurrence frequency and deviation degree, and the fault points that repeatedly occur or deviate from the normal interval are prioritized. The trend evolution path is listed separately, and the key components that may be affected and the predicted development time are marked to form a hierarchical information list with clear levels.

[0028] A fault prediction report covering the entire test cycle is generated, and the prediction content is dynamically updated as the test progresses. The hierarchical information list is converted into a structured text, and the content is organized using the expression logic of "time-stage- abnormality-trend". When the test enters a new test stage, the multi-source data acquisition, normalization processing, parameter extraction, correlation analysis and fault identification steps are repeated to obtain new warning information. The new warning information is inserted into the corresponding stage or trend path position, and the structure and content of the original report are adjusted to maintain the synchronization of the prediction and the test process.

[0029] Embodiment 1: During the normalization processing of the obtained multi-source input data, dynamic monitoring data is an important part of the normalization. The dynamic monitoring data is collected by a sensor array installed on the bogie, traction system and other parts. These sensors include but are not limited to pressure sensors, speed sensors, vibration sensors, etc. Different sensors may have different sampling frequencies and time bases, which may cause deviations in the time of the collected dynamic monitoring data, making it impossible to directly perform unified analysis.

[0030] In order to align the data collected by these different sensors to a unified time axis, a time synchronization module is needed. The dynamic monitoring data is divided by sensor type, such as the bogie-related sensor data is classified into one category, and the traction system-related sensor data is classified into another category, forming independent data streams. The purpose of this is to facilitate the processing of data from different types of sensors separately, as different types of sensors may have different characteristics and time deviation conditions.

[0031] Timestamp verification is performed on each segmented data stream. Timestamp verification mainly checks whether the sampling start time and end time of data in each data stream are accurate, and identifies the positions where there is deviation in sampling start and end. For example, a certain sensor may have a clock error or other reasons, causing its sampling start time to be later than the actual specified time by a certain number of milliseconds, or there may be uneven time intervals during sampling, which need to be discovered through timestamp verification.

[0032] After identifying the data streams with time deviation, interpolation processing needs to be performed on these data streams. Linear interpolation algorithm is used here to supplement the monitoring values of missing time points. The basic principle of linear interpolation algorithm is to estimate the value of the missing time point in the middle according to the monitoring values of the adjacent two time points through linear calculation. For example, assuming that there is a missing time point t between time points t1 and t2, the monitoring value at t1 is v1, and the monitoring value at t2 is v2, then the estimated value v at t can be calculated by Through this way, the data streams with time deviation become continuous in time sequence, reducing the impact of data loss on subsequent analysis.

[0033] After completing the interpolation processing, the time axis of the interpolated data streams needs to be rearranged. This step is to unify the time identifiers of all data streams based on the timestamp of the main sensor. The main sensor is usually selected from those sensors with high precision, good stability and dominant role in the whole system. For example, in the test of rail vehicles, the speed sensor installed in the key position with high sampling precision can be selected as the main sensor. Based on the timestamp of the main sensor, the data streams of other sensors are adjusted and arranged in time sequence, so that the time identifiers of all data streams are consistent with the timestamp of the main sensor, thereby generating sequence data with continuous timestamp. After such processing, the data collected by different sensors has consistency in time, providing a unified time basis for subsequent parameter extraction, correlation analysis and other steps.

[0034] For the normalization processing of static configuration data, version verification module is needed to compare historical archive records. Static configuration data contains important information such as component number, manufacturing date, technical specification parameters, etc., and the accuracy of these information is crucial for fault prediction. During comparison, if it is found that there are error fields in the current static configuration data, such as incorrect component number, technical specification parameters inconsistent with actual situation, etc., it needs to be corrected in time; at the same time, if it is found that there are unrecorded component model information, such as newly replaced components or previously omitted component models, it needs to be supplemented. Through such processing, the accuracy and completeness of static configuration data can be ensured, so that it can truly reflect the configuration situation of rail vehicles.

[0035] The normalization processing of the environmental parameter data is to eliminate abnormal values beyond the normal working range of the device through a threshold filtering module. The environmental parameter data includes temperature, humidity, atmospheric pressure, etc. in the test area, and the collection of these data is completed by the environmental monitoring station arranged in the test site. In the actual collection process, abnormal values may be collected due to various reasons, such as sensor failure, external interference, etc. If these abnormal values are not eliminated, it will mislead the subsequent analysis. Therefore, by setting the threshold range of the normal working of the device, the data beyond the range is considered as abnormal value and is eliminated. After eliminating the abnormal values, the remaining data needs to be labeled with the specific installation position of the temperature and humidity sensor or the barometer. The purpose of labeling the installation position is to enable more accurate analysis in combination with the specific position information when analyzing the influence of the environmental parameters on the device in the subsequent analysis, because the environmental parameters at different positions may be different.

[0036] Through the above normalization processing of the dynamic monitoring data, the static configuration data and the environmental parameter data, the multi-source input data is optimized in terms of time, accuracy and integrity, etc. to provide high-quality data support for each step of the subsequent fault prediction method, and to ensure the reliability and accuracy of the entire fault prediction process. In this process, each step is closely connected, and improper handling of any link may affect the final fault prediction result. Therefore, in actual application, the multi-source input data needs to be normalized according to the above steps to ensure the effectiveness of the fault prediction method.

[0037] Example 2: In the parameter extraction of the normalized dynamic monitoring data, static configuration data and environmental parameter data, the parameter extraction of the dynamic monitoring data is one of the key links. After the normalization processing, the dynamic monitoring data has become sequence data with continuous time stamps, which reflects various dynamic running states of the rail vehicle in the test process. In order to extract the key features representing the device state from these massive data, a sliding window algorithm is needed to intercept a continuous data segment with a fixed time length.

[0038] The core of the sliding window algorithm is to set a suitable window length, which can be determined according to the specific test requirements and data characteristics. For example, in the monitoring of some key systems of the rail vehicle, the window length can be set to 10 seconds, 30 seconds or 1 minute, etc. With the passage of time, this window slides forward on the dynamic monitoring data sequence according to a certain time step, and each sliding intercepts a continuous data segment with a fixed time length. The advantage of this is that the continuous data stream can be divided into multiple small segments with time sequence, which facilitates independent analysis of the data in each small segment.

[0039] For each segment of the intercepted continuous data segment, the mean and variance thereof need to be calculated. The mean can reflect the overall level of the monitoring values in the data segment, for example, in a certain 10-second window, the current data mean of the traction system can reflect the average size of the current in the time period; the variance can reflect the dispersion degree of the data, that is, the fluctuation of the data, the greater the variance, the more intense the fluctuation of the data in the time period, and vice versa. By calculating the mean and variance of each segment of data, the state feature vector is formed. These state feature vectors are a high generalization and abstraction of the original dynamic monitoring data, which can succinctly describe the running state of the equipment in different time periods, and provide an important feature basis for subsequent correlation analysis and fault identification.

[0040] For parameter extraction of static configuration data, the field extraction module is needed to filter out the key parameters related to faults from a large amount of configuration data. Static configuration data contains component numbers, factory dates, technical specification parameters and a large amount of information, but not all information has equal importance for fault prediction. Therefore, it is necessary to determine which parameters are key parameters related to faults, such as component life, rated load and maintenance period.

[0041] Component life information can help determine whether the component is close to or exceeds its designed service life, thereby predicting possible faults; the rated load parameter is compared with the actual running load of the equipment, and if the actual load exceeds the rated load, it may cause the equipment to run overload, increasing the probability of failure; the maintenance period information can prompt whether the component needs to be maintained, and failure may be caused by the performance degradation of the component due to the lack of timely maintenance. After the field extraction module filters out these key parameters, they are combined together to form a configuration feature set. This set concentrates important configuration information that affects the equipment failure, so that the influence of configuration factors on the equipment state can be more targeted in subsequent analysis.

[0042] The parameter extraction of environmental parameter data is to calculate the change rate per unit time through the trend analysis module, and to extract the temperature and humidity fluctuation value and the air pressure change as the environmental feature sequence. Environmental parameter data includes temperature, humidity, atmospheric pressure, etc., and the changes of these parameters may affect the running state of the railway vehicle. For example, a sharp rise in temperature may cause the equipment to overheat, affecting its performance; changes in humidity may cause some components to be damp, causing failure; changes in air pressure may affect the pneumatic system.

[0043] To capture the trend of environmental parameter changes, the trend analysis module needs to calculate the change amount of environmental parameters within each time unit, i.e., the change rate. The time unit here can be minutes, hours, etc., which is determined according to the time scale of the test and the speed of change of the environmental parameters. By calculating the change rate, we can understand the magnitude and direction of the change of the environmental parameters within a unit of time. For example, the temperature increased by 5°C in 1 hour, the humidity decreased by 10% in 30 minutes, etc. In addition to the change rate, the temperature and humidity fluctuation value and the pressure change amount are also important features. The temperature and humidity fluctuation value can reflect the degree of instability of the environment temperature and humidity, the greater the fluctuation, the more unstable the environment, and the more significant the impact on the equipment may be; the pressure change amount can reflect the change of atmospheric pressure.

[0044] Arranging these calculated change rates, temperature and humidity fluctuation values, and pressure change amounts in chronological order forms the environmental feature sequence. This sequence can clearly show the trend of environmental parameter changes over time, providing data support for analyzing the relationship between environmental factors and equipment failures. In practical applications, through the analysis of the environmental feature sequence, we can determine whether environmental changes are one of the reasons for equipment abnormalities, or predict the possible failures of the equipment under specific environmental changes.

[0045] The parameter extraction process is an important process that transforms the original multi-source input data into feature vectors, feature sets, and feature sequences with representation capabilities. By conducting targeted parameter extraction on dynamic monitoring data, static configuration data, and environmental parameter data, subsequent cross-dimensional correlation analysis and fault identification can be based on more concise and representative data, improving the efficiency and accuracy of fault prediction. In this process, the processing method and parameter selection of each step need to be determined according to the specific test scenario and equipment characteristics to ensure that the extracted features accurately reflect the state of the equipment and the influencing factors.

[0046] Example 3: In the cross-dimension association analysis of the state feature vector, the configuration feature set and the environment feature sequence through the time sequence association model, a time anchor mechanism needs to be established first. Since the rail vehicle test is usually divided into multiple stages, such as the initial debugging stage, the load increasing stage, the continuous running stage and the like, the test conditions and the equipment running states in each stage are different, and therefore the feature data in different dimensions need to be aligned to the test stage division nodes according to the time stamps. Specifically, the starting and ending time points of each test stage are determined first, for example, the initial debugging stage is from 0 hour to 2 hours after the test starts, the load increasing stage is from 2 hours to 6 hours, and the like. Then, for each data point in the state feature vector, the configuration feature set and the environment feature sequence, it is judged according to the time stamp of the data point which test stage the data point belongs to, and the data point is associated with the corresponding test stage node, so that the different dimension feature data in the same test stage are under the same reference in time, and the uniformity in time is provided for the subsequent association analysis.

[0047] Next, the association strength between the feature data in different dimensions is calculated through the weight distribution module. Taking each test stage node as the reference, the state feature vector, the configuration feature set and the environment feature sequence in the same stage are combined into a feature cluster. For example, at a certain time node in the load increasing stage, the state feature vector at this time (such as the average value of the traction system current, the variance of the bogie vibration and the like), the configuration feature set (such as the rated load of the traction motor, the maintenance period of the gear box and the like) and the environment feature sequence (such as the temperature change rate in this stage, the humidity fluctuation value and the like) are combined into a feature cluster.

[0048] The matching degree of the dynamic state and the configuration parameter is calculated for each cluster of features. The dynamic state is represented by the state feature vector, the configuration parameter comes from the configuration feature set, and the calculation of the matching degree needs to consider whether the dynamic state conforms to the specified range of the configuration parameter. For example, if the rated current of the traction motor in the configuration feature set is 100A±5A, and the average value of the current of the motor in the state feature vector is 108A, which exceeds the rated range, then the matching degree of the dynamic state and the configuration parameter is low; if the average value of the current is 103A, which is within the rated range, then the matching degree is high. In this way, the degree of conformity between the dynamic state and the configuration parameter is quantified.

[0049] Simultaneously, the correlation between environmental conditions and state changes is calculated. Environmental conditions are represented by a sequence of environmental characteristics, while state changes are reflected through changes in the state characteristic vector. Calculating the correlation requires analyzing whether changes in environmental conditions are related to changes in the state characteristic vector. For example, when the ambient temperature rises sharply within a certain time period, observe whether the monitored value of the equipment temperature in the state characteristic vector also rises accordingly. If there is a clear synchronous upward trend, it indicates a high correlation between environmental conditions and state changes; if the ambient temperature rises but the monitored value of the equipment temperature does not change significantly, the correlation is low. The degree of correlation between environmental conditions and state changes is determined through statistical analysis and other methods.

[0050] Association strength is assigned based on matching degree and relevance results. A higher matching degree indicates a better fit between the dynamic state and configuration parameters, and a stronger association between them. A higher relevance indicates a greater influence of environmental conditions on state changes, and a stronger association between environmental conditions and state features. Association strength can be represented numerically; a larger value indicates a stronger association between dimensions. For example, dimensions with both high matching degree and high relevance are assigned higher association strength values, while those with low matching degree and low relevance are assigned lower association strength values.

[0051] After assigning correlation strength, the correlated features are dimensionally fused. Dimensional fusion integrates correlated features from the state feature vector, configuration feature set, and environmental feature sequence to generate a joint feature set containing time nodes, state dimensions, and influence dimensions. The time node corresponds to a specific point in the test phase, the state dimension includes the feature values ​​of each feature in the state feature vector, and the influence dimension includes configuration parameters and environmental features associated with the state dimension. For example, at a certain time node, the joint feature set might contain state dimension information such as the mean traction system current and bogie vibration variance at that moment, as well as influence dimension information such as the corresponding traction motor rated load and the rate of change of ambient temperature, while also indicating which test phase the time node belongs to.

[0052] This cross-dimensional correlation analysis organically combines previously independent dynamic monitoring data, static configuration data, and environmental parameter data, fully exploring the correlations between features of different dimensions. The resulting joint feature sets not only contain feature information from each dimension but also reflect their temporal correlations and influence relationships. These joint feature sets can more comprehensively and deeply describe the state of the rail vehicle during testing, providing a rich information foundation for subsequent identification of abnormal states, potential fault points, and trend evolution paths during testing, thus contributing to more accurate fault prediction. In practice, the accuracy of the time anchoring mechanism, the reasonable composition of feature clusters, the calculation methods for matching degree and correlation, and the dimensional fusion methods all need to be optimized based on specific test data and equipment characteristics to ensure the effectiveness and reliability of the correlation analysis.

[0053] Example 4: In the process of identifying abnormal states, potential failure points and trend evolution paths based on joint feature groups, the state classification module is needed to perform pattern clustering on the joint feature groups. Taking the test of the traction system of a rail vehicle as an example, the joint feature groups contain time node, traction motor current mean value, voltage fluctuation variance and other state dimension information in different test stages, as well as motor rated load, environmental temperature change rate and other influence dimension information. The state classification module will input these data with multi-dimensional features into clustering algorithms such as K-means clustering or DBSCAN clustering, and by calculating the similarity between data points, data with similar features will be classified into the same cluster.

[0054] In the clustering process, the algorithm will automatically divide different clusters according to the distribution characteristics of the data, and each cluster corresponds to a typical operating state in a certain stage of the test. For example, in the initial debugging stage of the traction system, the current mean value in the joint feature group is usually low and the fluctuation is small, and the environmental temperature change rate is also relatively gentle. These data will be clustered into a class, forming the normal state interval of this stage. In the load increasing stage, the current mean value will gradually increase with the increase of the load, and the fluctuation amplitude may also increase. The corresponding joint feature group data will be clustered into another class. In this way, the normal state interval and abnormal state boundary of each stage of the test are determined, and the feature range of the normal operating state and the starting point of the abnormal state are clarified.

[0055] In the normal state interval of each stage, the threshold detection module is used to identify potential failure points. Still taking the test of the traction system as an example, it is assumed that in the continuous running stage, the historical data shows that the historical mean value of the traction motor current mean value is 200A, and the standard deviation is 10A. Therefore, the two times standard deviation is 20A, and the normal state interval can be set to 180A to 220A. When the threshold detection module detects that the current mean value of a certain time node is 230A, which exceeds the two times standard deviation of the historical mean value, i.e. exceeds the upper limit 220A of the normal state interval, this feature value will be identified as a potential failure point. This is because the current value deviates from the normal range by a large margin, which may indicate potential problems such as motor overload and winding failure.

[0056] In identifying potential failure points, not only the deviation of a single feature value should be considered, but also other related features should be considered. For example, if the current mean value exceeds the threshold value at the same time, the feature value of the motor temperature also exceeds the normal range, and the environmental temperature change rate does not increase significantly, which indicates that there may be a fault. The threshold detection module will comprehensively analyze multiple feature values in the joint feature group to avoid misjudgment due to accidental fluctuations of a single feature, and improve the accuracy of identifying potential failure points.

[0057] The trend inference module analyzes the change direction of the continuous abnormal features, and extracts the evolution path that may cause the fault. Assuming that in the traction system test, the average current of continuous multiple time nodes is 225 A, 230 A, 235 A in turn from a certain time point, showing a continuous upward trend, and each time exceeds the upper limit of the normal state interval. The trend inference module will analyze this set of continuous abnormal features, judge that the change direction is continuously increasing, and combine the configuration features and environmental features such as the rated load of the motor, the heat dissipation condition, etc. to infer the possible fault evolution path.

[0058] For example, if the rated load of the motor is 250 A, the current is continuously rising and approaching the rated load, the temperature of the motor is also gradually rising, and the environmental temperature changes little, the trend inference module may extract the evolution path of “continuous overload of current → temperature rise of motor → accelerated insulation aging → possible short circuit fault of winding”. When extracting the evolution path, the time point of each key node, the change of feature value, and the key components that may be affected, such as the winding of the traction motor in the above example, are marked.

[0059] In the test of the bogie of the railway vehicle, the joint feature group may include the mean and variance of the bogie vibration acceleration, the state dimension information such as the wheelset wear, and the influence dimension information such as the maintenance period of the bogie and the flatness of the test track. The state classification module will determine the normal vibration state interval of the bogie in different test stages according to the data clustering. When the threshold detection module finds that the mean of the vibration acceleration is continuously higher than the normal interval in a certain time period, and the wheelset wear also exceeds the normal range, it will be identified as a potential fault point. The trend inference module analyzes the continuous abnormal vibration features and the wear change, and may extract the evolution path of “intensified wheelset wear → increased bogie vibration → fatigue of suspension system components → possible suspension device failure”.

[0060] Through the pattern clustering, potential fault point identification and trend evolution path extraction of the joint feature group, the possible fault signs in the test process can be comprehensively and systematically identified, and a fault warning information set is formed. These warning information not only includes the current abnormal state and potential fault point, but also predicts the possible development trend and path of the fault, providing detailed fault prediction information for the test personnel, so as to take timely measures to avoid the occurrence or expansion of the fault. In actual application, the clustering algorithm selection of the state classification module, the threshold setting of the threshold detection, and the analysis model of the trend inference module all need to be adjusted and optimized according to the specific railway vehicle system and test data, to ensure the accuracy and effectiveness of fault identification and prediction.

[0061] Example 5: In the hierarchical organization according to the time sequence and severity level of the fault early warning information set and the generation of the fault prediction report, taking the rail vehicle traction system test as an example, it is assumed that the test process includes four stages of initial debugging, load increment, continuous running and fault simulation. First, arrange the abnormal states according to the time sequence of the test process, and list the corresponding test stage and time node under each abnormal state. For example, in the 3rd hour and 15 minutes of the load increment stage, the average current of the traction motor is abnormally high, so in the hierarchical organization, the abnormal state will be first classified into the load increment stage, and the specific time node of its occurrence will be clearly marked.

[0062] In each abnormal state, the potential fault points need to be sorted according to the frequency of occurrence and the degree of deviation. Continuing with the traction system test as an example, it is assumed that two potential fault points occur in the continuous running stage, one is that the average current value exceeds the normal range many times and the deviation is large, and the other is that the voltage fluctuation occasionally exceeds the normal range and the deviation is small. At this time, the fault point of the current average is arranged first, because its frequency of occurrence is higher and the degree of deviation from the normal range is larger, which is more likely to cause serious equipment failure. In the arrangement, detailed information such as the specific characteristic value of each potential fault point, the deviation amplitude from the normal range, and the number of occurrences is listed.

[0063] The trend evolution path needs to be listed separately, and the key components that may be affected and the predicted development time are marked. For example, in the traction system test, it is found that the average current value has been rising continuously since the 4th hour, increasing by 5-8A every 30 minutes or so, and the motor temperature is also rising at a rate of 2-3℃ per hour. Through analysis, it is determined that the trend may lead to the evolution path of "continuous current overload → motor temperature rise → insulation layer accelerated aging → possible winding short circuit failure in the next 2-3 hours", and the key components that may be affected are marked as traction motor winding, and it is predicted that the fault may occur in 2-3 hours.

[0064] After completing the organization of the hierarchical information list, it needs to be converted into structured text, using the expression logic of "time-stage-exception-trend" to organize the content. Taking an abnormal situation in the traction system test as an example, the structured text may be expressed as: "At the 3rd hour and 15 minutes of the test, in the load increment stage, the average current of the traction motor is detected as 230A, exceeding the upper limit of the normal range (180A-220A) of this stage. This potential fault point has occurred 2 times in the past 30 minutes, deviating from the normal average by 20A. The current trend shows that the current value is continuously rising at a rate of about 3A every 10 minutes, which may lead to a traction motor winding overheating fault, and is expected to affect the normal operation of the motor in the next 1.5-2 hours." When the test enters a new test phase, it is necessary to repeatedly perform multi-source data acquisition, normalization processing, parameter extraction, correlation analysis and fault identification, etc. to obtain new early warning information. For example, after the test enters the fault simulation phase from the continuous operation phase, it is necessary to re-collect the dynamic monitoring data, static configuration data and environmental parameter data of this phase, normalize these data, extract the state feature vector, configuration feature set and environmental feature sequence, and then perform cross-dimensional correlation analysis through the time sequence correlation model to generate new joint feature groups, and further identify the abnormal state, potential fault point and trend evolution path of this phase to obtain new early warning information.

[0065] After obtaining the new early warning information, it is necessary to insert it into the corresponding phase or trend path position, and adjust the structure and content of the original report to maintain the synchronization of prediction and test progress. Assuming that a new abnormal state is found in the fault simulation phase through analysis, which is associated with the prediction result of a trend evolution path in the continuous operation phase, then the new abnormal state needs to be inserted into the corresponding position of the fault simulation phase, and the predicted development time and impact range of the related trend evolution path in the continuous operation phase are updated, so that the entire fault prediction report can accurately reflect the latest progress of the test and the real-time state of the equipment.

[0066] Taking the test of the bogie of a rail vehicle as an example, it is found in the initial debugging phase that the mean value of the vibration acceleration of the bogie is 0.8g, which exceeds the normal range (0.3g-0.6g), and this potential fault point occurs 3 times within 1 hour, deviating from the normal mean value of 0.2g. The trend evolution path shows that the vibration acceleration is rising at a rate of 0.1g every 15 minutes, which may cause fatigue failure of the suspension system components of the bogie, and is expected to affect the suspension device within 45 minutes to 1 hour in the future. After the test enters the load increment phase, new data is collected and it is found through analysis that the mean value of the vibration acceleration of the bogie has reached 1.0g, and the temperature of the suspension system has also risen, which belongs to new early warning information. At this time, it is necessary to insert this new abnormal state of the load increment phase into the report, and adjust the predicted development time of the trend evolution path in the initial debugging phase, updating the original 45 minutes to 1 hour to 30 minutes to 45 minutes, and supplementing the impact information of the temperature rise of the suspension system.

[0067] Through this hierarchical organization and dynamic updating, the generated fault prediction report can cover the entire test cycle and continuously update the content as the test progresses, providing timely and accurate fault warning information for test personnel. Each abnormal state, potential fault point and trend evolution path in the report has a clear time identifier, phase attribution and severity level, which facilitates test personnel to quickly locate the problem, assess the risk and take appropriate measures to ensure the safe conduct of the rail vehicle test and the reliable operation of the equipment.

[0068] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify a subject or action, without necessarily requiring or implying any such actual relationship or order between such subjects or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0069] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and variations of the embodiments can be undertaken without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A method for failure prediction based on test data of a rail vehicle, characterized in that, The method comprises the following steps: During the test process of the rail vehicle, dynamic monitoring data, static configuration data and environmental parameter data of the test stage are acquired in real time as multi-source input data; The acquired multi-source input data is normalized, wherein the dynamic monitoring data is labeled with a time stamp and calibrated in sampling frequency, the static configuration data is checked for version information and supplemented with missing fields, and the environmental parameter data is filtered for interference values and recorded for measurement location; The normalized dynamic monitoring data, static configuration data and environmental parameter data are respectively parameterized to form a state feature vector, a configuration feature set and an environmental feature sequence; The state feature vector, the configuration feature set and the environmental feature sequence are cross-dimensionally analyzed by a time sequence correlation model to generate a joint feature group containing time correlation; Based on the joint feature group, abnormal states, potential fault points and trend evolution paths in the test process are identified to form a fault warning information set; According to the time sequence and severity level of the fault warning information set, the fault warning information set is hierarchically organized to generate a fault prediction report covering the whole test cycle, and the prediction content is dynamically updated with the test process.

2. The method for failure prediction based on test data of a rail vehicle according to claim 1, characterized in that, The normalized multi-source input data is normalized, and the specific steps are as follows: The dynamic monitoring data is aligned to a unified time axis by a time synchronization module, and sequence data with continuous time stamps is generated; The static configuration data is compared with historical archive records by a version verification module to correct error fields and supplement missing component model information; The environmental parameter data is filtered by a threshold filtering module to remove abnormal values outside the normal working range of the device, and the remaining data is labeled with the specific installation location of the temperature and humidity sensor or the barometer.

3. The method of claim 1, wherein the method further comprises: The normalized dynamic monitoring data, static configuration data and environmental parameter data are respectively parameterized, and the specific steps are as follows: The dynamic monitoring data is cut off by a sliding window algorithm to obtain continuous data segments of a fixed time length, and the mean and variance of each data segment are calculated as a state feature vector; The static configuration data is filtered by a field extraction module to obtain key parameters related to faults, including component life, rated load and maintenance cycle, to form a configuration feature set; The environmental parameter data is calculated by a trend analysis module to obtain the change rate per unit time, and the temperature and humidity fluctuation value and the air pressure change value are extracted as environmental feature sequences.

4. The method for failure prediction based on test data of a rail vehicle according to claim 3, characterized in that, The state feature vector, the configuration feature set and the environmental feature sequence are cross-dimensionally analyzed by a time sequence correlation model, and the specific steps are as follows: A time anchor mechanism is established to align the state feature vector, the configuration feature set and the environmental feature sequence to the test stage division nodes according to the time stamp; The correlation strength between different dimension features is calculated by a weight distribution module, and the feature information of the dynamic state, the corresponding configuration parameter and the environmental condition in the same test stage is highlighted; The correlated features are dimensionally fused to generate a joint feature group containing time node, state dimension and influence dimension information.

5. The method for failure prediction based on test data of a rail vehicle according to claim 4, characterized in that, Based on the joint feature group, abnormal states, potential fault points and trend evolution paths in the test process are identified, and the specific steps are as follows: The joint feature group is pattern clustered by a state classification module to determine the normal state interval and the abnormal state boundary of each stage of the test; In the normal state interval of each stage, the threshold detection module identifies the feature values exceeding twice the standard deviation of the historical mean as potential fault points; The trend extrapolation module analyzes the change direction of continuous abnormal features to extract the evolution path that may trigger a fault.

6. The method for failure prediction based on test data of a rail vehicle according to claim 5, characterized in that, The fault warning information set is hierarchically organized according to the time sequence and severity level, and the specific steps are as follows: Arrange the abnormal states in chronological order according to the test process, and list the corresponding test stages and time nodes under each abnormal state; In each abnormal state, sort the potential fault points according to their occurrence frequency and deviation degree, and prioritize the fault points that repeatedly occur or deviate significantly from the normal interval; The trend evolution path is listed separately, with the possible impact on key components and the predicted development time, forming a hierarchical information list with clear levels.

7. The method for failure prediction based on test data of a rail vehicle according to claim 6, characterized in that, Generate a fault prediction report covering the entire test cycle and update it dynamically, with the following specific steps: Convert the hierarchical information list to structured text, using the "time-stage-exception-trend" expression logic to organize the content; When the test enters a new test stage, repeat the steps of multi-source data acquisition, normalization processing, parameter extraction, correlation analysis, and fault identification to obtain new warning information; Insert the new warning information into the corresponding stage or trend path position, adjust the structure and content of the original report, and maintain the synchronization of the prediction and test process.

8. The method for failure prediction based on test data of a rail vehicle according to claim 1, characterized in that, When acquiring dynamic monitoring data, static configuration data, and environmental parameter data in real time, the specific acquisition steps are as follows: Dynamic monitoring data is collected by sensor arrays installed on the bogie and traction system, and a filtering algorithm is used to suppress electromagnetic interference noise; Static configuration data is retrieved in real time through the vehicle management system interface, including component number, factory date, and technical specification parameters, forming a multi-type multi-source input data set; Environmental parameter data is collected by environmental monitoring stations arranged in the test site, recording temperature, humidity, and atmospheric pressure data in the test area simultaneously.

9. The method for failure prediction based on test data of a rail vehicle according to claim 4, characterized in that, When calculating the correlation strength between different dimensional features through the weight allocation module, the specific calculation steps are as follows: Take the test stage node as the reference, and group the state feature vector, configuration feature set, and environmental feature sequence within the same stage into a feature cluster; Calculate the matching degree of dynamic state and configuration parameters for each cluster of features; Calculate the correlation between environmental conditions and state changes; According to the matching degree and correlation results, assign the correlation strength, with higher strength indicating closer information correlation between dimensions.

10. The method of claim 2, wherein, When aligning different sensor sampling points to a unified time axis through the time synchronization module, the specific alignment steps are as follows: Divide the dynamic monitoring data into independent data streams according to sensor type, and perform timestamp verification on each data stream to identify the sampling start and end deviation positions; Interpolate the data streams with time deviation, and supplement the monitoring values at missing time points through linear interpolation algorithm; After interpolation, rearrange the time axis of the data stream, and unify the time identifiers of all data streams based on the main sensor timestamp as the reference, as a sequence data with continuous timestamps.

Images