A Deep Learning-Based Early Warning Method for Rail Vehicle Wheelset Maintenance Integrating Meteorological Big Data and LKJ Records

Abstract

This invention relates to the field of intelligent operation and maintenance technology for rail transit, specifically to a deep learning-based early warning method for rail vehicle wheelsets that integrates meteorological big data and LKJ records. The method includes the following steps: S1: Real-time acquisition of multi-source meteorological data and rail vehicle LKJ record data; S2: Real-time stream processing of the multi-source data acquired in S1 to obtain fused feature data; S3: Inputting the fused feature data obtained in S2 into a pre-trained deep learning model to output the corresponding wheelset damage index; S4: Determining the corresponding early warning level based on the wheelset damage index output in S3 and a preset threshold; S5: Executing corresponding maintenance response measures according to the early warning level determined in S4. This invention, by integrating meteorological big data and LKJ records to construct a deep learning prediction model, achieves real-time identification and graded early warning of wheelset damage, thereby improving the safe operation and intelligent maintenance capabilities of rail vehicles in complex meteorological environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for rail transit, and in particular to a deep learning-based early warning method for rail vehicle wheelsets that integrates meteorological big data and LKJ records. Background Technology

[0002] During long-term operation, the wheelsets of rail vehicles are key structural components that directly bear the vehicle load and rail impact. Their service condition directly affects driving safety and operational economy. As railway transportation develops towards high speed and heavy load, wheelsets are in a complex stress environment for a long time. Their material properties and structural stability are easily affected by both environmental meteorological factors and operating conditions. In particular, in actual operation, vehicles often experience extreme meteorological conditions such as high temperature radiation, rainstorm salt spray, and low temperature wind load. These environmental stresses will couple with the operating load, causing crack propagation, corrosion deepening, or fatigue accumulation on the metal surface of the wheelsets, ultimately leading to serious safety hazards such as wheel flange spalling and wheel rim fracture. Although the current railway system has deployed temperature monitoring, vibration detection, and infrared diagnostics, they mostly rely on single physical signals and lack the ability to jointly model changes in the meteorological environment and operating data, making it difficult to achieve dynamic prediction of the health status of the wheelsets.

[0003] Existing technologies for wheelset maintenance and early warning still have significant shortcomings: First, the monitoring data sources are limited, failing to integrate meteorological big data, LKJ operation records, and image detection results into a multimodal manner, resulting in a lack of environmental relevance in damage assessment; second, the model analysis methods are outdated, with traditional threshold alarms or statistical regression models struggling to capture the nonlinear damage evolution patterns under the coupling of meteorological and operational conditions; third, there is a lack of prediction and graded response mechanisms, meaning that once damage occurs, the system often only provides passive alarms and cannot quantify the risk level in advance. Therefore, it is necessary to propose a deep learning-based early warning method for rail vehicle wheelsets that integrates meteorological big data and LKJ records to improve the safe operation and maintenance decision-making capabilities of rail vehicles in complex meteorological environments. Summary of the Invention

[0004] To achieve the above objectives, this invention provides a deep learning-based early warning method for the maintenance of rail vehicle wheelsets that integrates meteorological big data and LKJ records.

[0005] A deep learning-based early warning method for rail vehicle wheelset maintenance, integrating meteorological big data and LKJ records, includes the following steps:

[0006] S1: Real-time acquisition of multi-source meteorological data and LKJ record data from rail vehicles;

[0007] S2: Perform real-time stream processing on the multi-source data collected by S1, including data cleaning and feature extraction, to obtain fused feature data;

[0008] S3: Input the fused feature data obtained in S2 into the pre-trained deep learning model and output the corresponding wheel pair damage index;

[0009] S4: Based on the wheelset damage index output by S3 and the preset threshold, determine the corresponding warning level;

[0010] S5: Implement corresponding maintenance response measures based on the warning level determined by S4.

[0011] Optionally, S1 specifically includes:

[0012] S11: By connecting with the open interface of the National Meteorological Administration, it collects the temperature, precipitation intensity, wind speed, humidity and air salinity parameters of the target line section in real time, and updates them at a frequency of no less than once every 10 minutes;

[0013] S12: Obtain supplementary micro-meteorological data through ground-deployed IoT meteorological sensor nodes, including temperature and humidity sensors and PM2.5 particulate matter sensors, with a collection frequency of no less than once every 5 minutes;

[0014] S13: Call the LKJ data access module provided by the train dispatch center to receive LKJ record data of all running trains in the current section in real time, including train number, timestamp, current speed, acceleration, braking status and track section information, with a sampling interval of 10 seconds.

[0015] Optionally, S2 specifically includes:

[0016] S21: Align the multi-source meteorological data collected in S1 with the LKJ recorded data according to a unified timestamp and section number to form a time-consistent data stream input;

[0017] S22: The sliding window mechanism is used to calculate the local mean and standard deviation of each type of sensor data, and outlier data points with absolute values ​​exceeding the set threshold are removed based on the Z-score method;

[0018] S23: Set the physical boundary value range for the corresponding data type, and perform physical filtering on data that exceeds the boundary range;

[0019] S24: Use a time linear interpolation completion strategy for short-term missing data points, and mark long-term missing data points as invalid and submit an abnormal data identifier;

[0020] S25: Based on continuous data within each time window, extract meteorological change rate, operational fluctuation characteristics, and environmental extreme indicators to obtain fused feature data.

[0021] Optionally, S24 specifically includes:

[0022] S241: When the time interval between consecutive missing data is less than a set threshold, this missing interval is marked as an interpolable interval, and the start time of the interpolation interval is set to... The end time is ;

[0023] S242: Obtain the observation values ​​of the two valid data points before and after the interpolable interval, denoted as follows: and And calculate the total time difference between intervals. ;

[0024] S243: Let the missing time be... ,and The corresponding estimated value is calculated using a time linear interpolation model. The formula is: ,in, For time points The interpolated estimate; Indicates a time interval;

[0025] S244: All interpolated values The data is filled into the corresponding time series and marked as interpolated data.

[0026] Optionally, S25 specifically includes:

[0027] S251: Within each sliding time window, calculate the rate of change of the temperature difference between the beginning and end of the window, as a representative feature of the meteorological rate of change.

[0028] S252: Extract the speed standard deviation from the LKJ record data to characterize the degree of train operation fluctuation within the time window, and use it as an operation fluctuation feature;

[0029] S253: Based on the historical statistical distribution of multi-source meteorological data, the extreme index is calculated using the Z-score method;

[0030] S254: The temperature change rate, operational fluctuation characteristic value obtained from S251 to S253 are time-aligned and stitched together with multiple meteorological extreme indicators to obtain fused characteristic data.

[0031] Optionally, S3 specifically includes:

[0032] S31: Construct a historical sample dataset, which includes wheelset image data, fused feature data for the corresponding time period, and wheelset damage labels obtained by manual detection;

[0033] S32: Input the wheelset image data into a pre-trained ResNet network to extract image features and obtain image feature vectors that reflect the surface cracks, corrosion and wear characteristics of the wheelset;

[0034] S33: Input the fused feature data of the corresponding time period into the LSTM network to extract the temporal feature representation of the environment and operating status;

[0035] S34: The image feature vector is concatenated and fused with the temporal feature vector output by the LSTM, and then mapped to the regression output layer through a fully connected layer to predict the wheelset damage index.

[0036] Optionally, S32 specifically includes:

[0037] S321: Preprocess the wheelset image data, including removing fixed image borders and watermark information, using a denoising algorithm to suppress high-frequency noise, performing brightness and contrast normalization processing, and maintaining consistent pixel aspect ratio.

[0038] S322: Perform wheelset target region localization on the preprocessed image, identify the tread and wheel flange regions based on edge detection and geometric constraints, generate the corresponding region of interest, and crop to obtain the target sub-image containing wheelset surface information;

[0039] S323: Perform size and channel normalization on the target sub-image, scale the image to the same size as the input of the pre-trained ResNet network, and perform numerical normalization and fixed mean and variance standardization on the channel to form a tensor that meets the requirements of the model input.

[0040] S324: Input the normalized image tensor into the pre-trained ResNet network, perform forward inference, and extract the intermediate layer feature maps;

[0041] S325: Perform global average pooling and channel compression on the intermediate layer feature map to obtain a fixed-length image feature vector;

[0042] S326: Perform amplitude normalization and anomaly detection on the obtained image feature vector, remove invalid features caused by blanks or occlusions, and mark the valid features as image feature vectors reflecting the surface cracks, corrosion and wear characteristics of the wheelset.

[0043] Optionally, S33 specifically includes:

[0044] S331: Based on the set time window and step size, the fused feature data of the corresponding time period are framed in order according to the timestamp to form a feature sequence with a fixed order;

[0045] S332: Perform mean and variance normalization on each feature dimension of the sequence in the same manner as during model training to ensure consistent input scale;

[0046] S333: For sequences that are shorter than the window length, use mask alignment to maintain consistent sequence lengths within the batch and generate corresponding mask identifiers;

[0047] S334: Input the normalized feature sequence and mask into the LSTM network, perform forward computation step by step, and obtain the hidden state sequence at each time step;

[0048] S335: Perform terminal time extraction or temporal pooling on the hidden state sequence to obtain a fixed-length temporal feature vector;

[0049] S336: Perform amplitude normalization and validity verification on the time series feature vector, and output the time series feature representation of the environment and operating state.

[0050] Optionally, S34 specifically includes:

[0051] S341: Receive the image feature vector obtained from S32 and the temporal feature vector obtained from S33, and perform scale unification and dimension calibration on the two;

[0052] S342: The calibrated image feature vector and the temporal feature vector are concatenated on the feature axis to form a single joint feature vector;

[0053] S343: Perform the set regularization and normalization processing on the joint feature vector, and output the fused stable input;

[0054] S344: The fused stable input is sequentially fed into the fully connected layer group for nonlinear mapping and feature compression to obtain an intermediate representation for regression calculation;

[0055] S345: Input the intermediate representation into the regression output layer to generate the corresponding wheelset damage index. .

[0056] Optionally, S4 specifically includes:

[0057] S41: Receiver wheelset damage index It is then compared with a preset multi-level risk threshold set, which includes four level boundary values, namely... , , , ,satisfy ;

[0058] S42: Based on wheelset damage index The corresponding warning level is determined based on the numerical range in which the value falls. The judgment rules include:

[0059] when At that time, it is set to level 0;

[0060] when At that time, it is set as a Level I warning;

[0061] when At that time, it was set as a Level II warning;

[0062] when At that time, it was set as a Level III warning;

[0063] when At that time, it was set as a Level IV warning;

[0064] S43: Output the corresponding warning level in a structured format, along with the current wheelset number, damage index value, warning level label, and corresponding risk advice text.

[0065] The beneficial effects of this invention are:

[0066] This invention integrates meteorological big data with LKJ operation records to construct a deep learning prediction model based on image and temporal features, enabling dynamic modeling and intelligent quantification of wheelset damage. It can simultaneously capture the interactive features of environmental changes and operational stress, and accurately identify damage trends under complex meteorological conditions such as high temperature, humid heat, and low temperature.

[0067] This invention combines the wheelset damage index output by the deep model with a multi-threshold early warning mechanism. The system can generate graded early warning results in real time and trigger corresponding maintenance responses, realizing closed-loop maintenance from minute-level monitoring to hour-level decision-making. This improves the accuracy and timeliness of wheelset fault prediction and significantly enhances the operational safety and intelligent maintenance level of rail vehicles in complex environments. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a schematic diagram of the deep learning-based maintenance early warning method for rail vehicle wheelsets according to an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram illustrating the training process of a deep learning model according to an embodiment of the present invention. Detailed Implementation

[0071] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0072] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0073] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0074] like Figures 1-2 As shown, a deep learning-based early warning method for rail vehicle wheelset maintenance, which integrates meteorological big data and LKJ records, includes the following steps:

[0075] S1: Real-time acquisition of multi-source meteorological data and LKJ record data from rail vehicles;

[0076] S2: Perform real-time stream processing on the multi-source data collected by S1, including data cleaning and feature extraction, to obtain fused feature data;

[0077] S3: Input the fused feature data obtained in S2 into the pre-trained deep learning model and output the corresponding wheel pair damage index;

[0078] S4: Based on the wheelset damage index output by S3 and the preset threshold, determine the corresponding warning level;

[0079] S5: Implement corresponding maintenance response measures based on the warning level determined by S4.

[0080] S1 specifically includes:

[0081] S11: By connecting with the open interface of the National Meteorological Administration, it collects the temperature, precipitation intensity, wind speed, humidity and air salinity parameters of the target line section in real time, and updates them at a frequency of no less than once every 10 minutes;

[0082] S12: Obtain supplementary micro-meteorological data through ground-deployed IoT meteorological sensor nodes, including temperature and humidity sensors and PM2.5 particulate matter sensors, with a data collection frequency of no less than once every 5 minutes;

[0083] S13: Call the LKJ data access module provided by the train dispatch center to receive LKJ record data of all running trains in the current section in real time, including train number, timestamp, current speed, acceleration, braking status and track section information, with a sampling interval of 10 seconds; the above steps collect data in layers from the macro meteorological platform, IoT micro monitoring and train LKJ operation records, providing high-precision and low-latency basic data support for subsequent damage modeling.

[0084] S2 specifically includes:

[0085] S21: Align the multi-source meteorological data collected in S1 with the LKJ recorded data according to a unified timestamp and section number to form a time-consistent data stream input;

[0086] S22: The sliding window mechanism is used to calculate the local mean and standard deviation of each type of sensor data, and outlier data points with absolute values ​​exceeding the set threshold are removed based on the Z-score method;

[0087] The formula for calculating the Z-score is: ,in, For the first Standard deviation of each data point; For the first Raw readings from each sensor; This represents the average value of the corresponding sensor within the current sliding window; The standard deviation of the corresponding sensor within the current sliding window; when the following conditions are met. When this happens, the data point is considered an outlier and removed, with a threshold value of [value missing]. Set to 3;

[0088] S23: Set the physical boundary value range for the corresponding data type, and perform physical filtering on data that exceeds the boundary range. The temperature range is set to -50 to 60 degrees Celsius, the precipitation intensity is set to 0 to 300 mm per hour, and the air salinity is set to 0 to 100 micrograms per cubic meter.

[0089] S24: Use a time linear interpolation completion strategy for short-term missing data points, and mark long-term missing data points as invalid and submit an abnormal data identifier;

[0090] S25: Based on continuous data within each time window, extract meteorological change rate, operational fluctuation characteristics, and environmental extreme indicators to obtain fused feature data, which serves as the data input for subsequent model building. The above steps, by performing standardized stream processing on multi-source heterogeneous data, can effectively remove outliers and unusable segments while ensuring data continuity and accuracy, and extract high-quality fused features that reflect actual environmental changes and operational fluctuations, ensuring that the data input received by the subsequent deep learning model has consistency and representativeness.

[0091] S24 specifically includes:

[0092] S241: When the time interval between consecutive missing data is less than a set threshold, this missing interval is marked as an interpolable interval, and the start time of the interpolation interval is set to... The end time is ;

[0093] S242: Obtain the observation values ​​of the two valid data points before and after the interpolable interval, denoted as follows: and And calculate the total time difference between intervals. ;

[0094] S243: Let the missing time be... ,and The corresponding estimated value is calculated using a time linear interpolation model. The formula is: ,in, For time points The interpolated estimate; Indicates a time interval;

[0095] S244: All interpolated values The data is filled into the corresponding time series and marked as interpolated data to maintain the integrity of the data time series. The above steps dynamically fill in the short-term missing data points based on the time linear interpolation model, which can maintain the stability of the data trend while ensuring physical continuity. It effectively reduces the data interruption problem caused by instantaneous packet loss or signal interference from sensors, and ensures the integrity and smoothness of the fused feature data in the time dimension, providing stable data support for the continuous input of subsequent models.

[0096] S25 specifically includes:

[0097] S251: To reflect the temporal trend of meteorological parameters, the rate of change of temperature difference between the beginning and end of each sliding time window is calculated as a representative feature of the meteorological change rate; the formula for calculating the rate of temperature change is: ,in, The rate of temperature change; This is the temperature value at the end of the current sliding window; This is the temperature value at the start of the current sliding window; The duration of the sliding window;

[0098] S252: Extract the speed standard deviation from the LKJ record data to characterize the degree of train operation fluctuation within a time window, serving as an operational fluctuation feature; the formula for calculating the operational fluctuation standard deviation is:

[0099] ,in, The standard deviation of train speed; For the first One velocity observation value; This is the mean of all velocity observations within the window; This represents the number of velocity data points within the window.

[0100] S253: Based on the historical statistical distribution of multi-source meteorological data, the extreme index is calculated using the Z-score method to reflect the intensity of anomalies; the formula is: ,in, Meteorological parameters Extreme indicators; The values ​​of meteorological parameters at the current moment (such as precipitation intensity, salinity, etc.); This represents the historical average value of the corresponding type of meteorological parameter; The historical standard deviation of the corresponding type of meteorological parameter;

[0101] S254: The temperature change rate and operational fluctuation characteristic values ​​obtained from S251 to S253 are time-aligned and stitched together with multiple meteorological extreme indicators to obtain fused feature data, which serves as effective input for subsequent deep learning models. The above steps, by extracting structured features from meteorological trends, train operation fluctuations, and environmental anomaly intensity, and constructing unified fused features, not only enhance the model's ability to express complex multi-source scenarios, but also improve the modeling accuracy of the damage evolution process of rail wheelsets under the combined effects of environmental and operational stresses, thereby achieving more targeted wheelset damage prediction and early warning judgment.

[0102] S3 specifically includes:

[0103] S31: Construct a historical sample dataset. The samples include wheelset image data, fused feature data for the corresponding time period, and wheelset damage labels obtained by manual detection. The fused feature data includes meteorological change rate, operational fluctuation characteristics, and extreme indexes.

[0104] S32: Input the wheelset image data into a pre-trained ResNet network to extract image features and obtain image feature vectors that reflect the surface cracks, corrosion and wear characteristics of the wheelset;

[0105] S33: Input the fused feature data of the corresponding time period into the LSTM network to extract the temporal feature representation of the environment and operating status, reflecting the lag effect of environmental stress evolution on the wheelset status;

[0106] S34: The image feature vector is concatenated and fused with the temporal feature vector output by the LSTM, and then mapped to the regression output layer through a fully connected layer to predict the wheelset damage index. By using the fused feature data and wheelset image data as training inputs, the model can simultaneously capture the complex relationship between environmental dynamic factors and wheelset surface state, achieving high-precision modeling of the wheelset damage index. This enhances the model's ability to identify and predict the health status of wheelsets in complex scenarios, providing more reliable data support for subsequent early warning decisions.

[0107] S32 specifically includes:

[0108] S321: Preprocess the wheelset image data, including removing fixed image borders and watermark information, using a denoising algorithm to suppress high-frequency noise, performing brightness and contrast normalization processing, and maintaining consistent pixel aspect ratio to ensure the stability of subsequent feature extraction.

[0109] S322: Perform wheelset target region localization on the preprocessed image, identify the tread and wheel flange regions based on edge detection and geometric constraints, generate the corresponding region of interest, and crop to obtain the target sub-image containing wheelset surface information;

[0110] S323: Perform size and channel normalization on the target sub-image, scale the image to the same size as the input of the pre-trained ResNet network, and perform numerical normalization and fixed mean and variance standardization on the channel to form a tensor that meets the requirements of the model input.

[0111] S324: Input the normalized image tensor into the pre-trained ResNet network, perform forward inference, and extract the intermediate layer feature maps;

[0112] S325: Perform global average pooling and channel compression on the intermediate layer feature map to obtain a fixed-length image feature vector;

[0113] S326: Perform amplitude normalization and anomaly detection on the obtained image feature vectors, remove invalid features caused by blanks, severe blurring, or occlusion, and mark the valid features as image feature vectors reflecting the surface cracks, corrosion, and wear of the wheelset, and output them for subsequent steps; the above steps complete region localization, size and channel normalization at the input end, and generate fixed-length image feature vectors using the intermediate layer features of the pre-trained ResNet network as carriers. This can stably extract information related to surface cracks, corrosion, and wear of the wheelset while ensuring input consistency, reduce the impact of image quality fluctuations on feature expression, and provide a highly discriminative and reusable feature foundation for subsequent damage index prediction.

[0114] The principle of feature extraction for wheel pair images based on ResNet is as follows:

[0115] During the operation of the rail vehicle, high-resolution images of the wheelset surface are acquired through fixed monitoring points or onboard image acquisition devices. The images are first processed by grayscale conversion, size normalization, and histogram equalization to eliminate noise and illumination differences. The processed images are then input into a pre-trained ResNet-50 convolutional neural network. The network extracts multi-level surface features through residual structures. Its shallow convolutional layers capture detailed information about crack edges and corrosion spots, the middle layer feature maps respond to wheel flange spalling and wear patterns, and the high-level semantic layer forms an abstract expression of the overall defects of the wheelset. In the output stage, the model performs global average pooling on the last layer of convolutional feature maps to obtain a 1024-dimensional image feature vector, which is used to quantitatively characterize the crack, corrosion, and wear features of the wheelset surface.

[0116] S33 specifically includes:

[0117] S331: Based on the set time window and step size, the fused feature data of the corresponding time period are framed in order according to the timestamp to form a feature sequence with a fixed order;

[0118] S332: Perform mean and variance normalization on each feature dimension of the sequence in the same manner as during model training to ensure consistent input scale;

[0119] S333: For sequences that are shorter than the window length, use mask alignment to maintain consistent sequence lengths within the batch and generate corresponding mask identifiers;

[0120] S334: Input the normalized feature sequence and mask into the LSTM network, perform forward computation step by step, and obtain the hidden state sequence at each time step;

[0121] S335: Perform terminal time extraction or temporal pooling on the hidden state sequence to obtain a fixed-length temporal feature vector;

[0122] S336: Perform amplitude normalization and validity verification on the time series feature vector, and output the time series feature representation of the environment and operating state for subsequent steps to call; the above steps, through chained processing of normalization, mask alignment and time series pooling, can stably extract the time series feature representation reflecting the environment and operating dynamics, and enhance the ability of subsequent damage index prediction to characterize time-dependent and abrupt signals.

[0123] The operational principle of extracting temporal features of environment and operating status based on LSTM network is as follows:

[0124] First, the fused feature data is divided into sliding time windows of length 30 according to the timestamp order, and a sliding step size of 5 steps is set between the windows to achieve overlapping sampling of continuous time segments; the feature data in each time window is standardized to make its mean 0 and variance 1, thereby maintaining the consistency of feature dimensions.

[0125] Subsequently, the normalized time-series data is input into the LSTM network structure. The LSTM unit controls the information flow through the input gate, forget gate and output gate to capture the long-term dependency between weather and operational status. At each time step, the network receives the fused feature input at the current time and recursively updates it by combining it with the hidden state at the previous time step to generate a new hidden state sequence.

[0126] After obtaining the complete hidden state sequence, a terminal time-step feature extraction strategy is adopted to extract high-level temporal feature representations from the output of the last time step, or average pooling is performed to generate a unified temporal feature vector.

[0127] S34 specifically includes:

[0128] S341: Receive the image feature vector obtained from S32 and the temporal feature vector obtained from S33, and perform scale unification and dimension calibration on the two to ensure that the numerical range and vector dimension meet the fusion input requirements;

[0129] S342: The calibrated image feature vector and the temporal feature vector are concatenated on the feature axis to form a single joint feature vector; the concatenation expression is as follows: ,in, Joint eigenvectors; Image feature vector; This is a time-series feature vector;

[0130] S343: Performs regularization and normalization on the joint feature vector to suppress abnormal amplitudes and overfitting risks, and outputs a stable fused input.

[0131] S344: The fused stable input is sequentially fed into the fully connected layer group for nonlinear mapping and feature compression to obtain an intermediate representation for regression calculation;

[0132] S345: Input the intermediate representation into the regression output layer to generate the corresponding wheelset damage index. The above steps introduce mathematical constraints in the feature splicing, normalization, and fully connected mapping stages to ensure that the image and temporal information have a consistent expression scale and a stable gradient propagation path after fusion, thereby achieving efficient coupling expression of spatial and temporal features and significantly improving the accuracy of damage index prediction and the model's generalization ability.

[0133] S4 specifically includes:

[0134] S41: Receiver wheelset damage index It is then compared with a preset multi-level risk threshold set, which includes four level boundary values, namely... , , , ,satisfy ;

[0135] Before determining the multi-level risk threshold, the mean and standard deviation of the wheelset damage index sequence within the last N time windows are calculated to construct a sliding window statistic. This statistic is used as a dynamic threshold correction factor to perform real-time offset correction on the fixed threshold set in order to improve the adaptive capability of the threshold.

[0136] S42: Based on wheelset damage index The corresponding warning level is determined based on the numerical range in which the value falls. The judgment rules include:

[0137] when At that time, it is set to level 0 (safe state);

[0138] when At that time, it is set as a Level I warning;

[0139] when At that time, it was set as a Level II warning;

[0140] when At that time, it was set as a Level III warning;

[0141] when At that time, it is set as a Level IV warning (emergency shutdown);

[0142] S43: Output the corresponding warning level in a structured form, along with the current wheelset number, damage index value, warning level label, and corresponding risk suggestion text; the above steps, by constructing a warning level determination rule based on a multi-threshold grading mechanism, can map the degree of wheelset damage to different risk level ranges, achieve accurate identification and quantitative expression of different risk levels, and effectively improve the hierarchical management capability of warning response and the targeted nature of operation and maintenance.

[0143] Example 1: Predictive maintenance of wheelsets under continuous high temperature conditions

[0144] In this embodiment, the rail vehicle operates in the hot and dry summer environment of North China, where the outside temperature is consistently between 38°C and 44°C, the relative humidity is maintained at 20%–30%, and the maximum rail surface radiation temperature can reach 60°C. The system obtains parameters such as air temperature, humidity, radiation intensity, and surface temperature at a sampling frequency of 10 seconds / time through a meteorological big data platform. At the same time, the LKJ recording system collects the train running speed (80–120 km / h in the section), axle temperature (maximum 95°C), braking frequency (average 4 times / km), and wheel flange temperature rise change (rise rate 3°C / min) in real time.

[0145] In the S2 stage, the collected data is subjected to streaming cleaning and synchronous calibration; through sliding processing with a time window of 30 seconds, sudden noise is removed and linear interpolation is performed to complete the data; then the meteorological change rate and the standard deviation of operational fluctuations are calculated to obtain a multi-dimensional fusion feature vector, which includes 12 feature components such as temperature rise rate, speed fluctuation amplitude and environmental radiation intensity.

[0146] In stage S3, the ResNet50 network receives wheelset images (1920×1080 pixels resolution, 2fps) captured by the trackside monitoring camera unit. Through multi-scale convolution, it extracts crack density, oxide layer texture, and wear depth features. Simultaneously, the LSTM network inputs a fused feature data sequence (30 time steps, 5-second stride) to extract features reflecting thermal fatigue and operational inertial response under high-temperature fluctuations. After fusion in step S34, a wheelset damage index is generated. ;

[0147] In stage S4, the set of thresholds is compared. If the current damage index is determined to be in the Level II warning range, an advance maintenance recommendation instruction will be automatically triggered, suggesting that infrared detection and visual inspection be performed within the next 12 hours.

[0148] Example 1 demonstrates the ability to identify the thermal fatigue trend of wheelset materials in real time under high temperature and long-term heavy load operation conditions, detecting the risk of crack initiation caused by high temperature 6-8 hours in advance, significantly improving the accuracy of maintenance prediction and the level of thermal stress management. In the test set under high temperature scenarios, the model achieved an accuracy rate of 94.2%, a false alarm rate of 3.1%, and a false alarm rate of 2.7%, outputting a Level II warning 6 hours in advance, accurately capturing the crack initiation trend caused by high temperature radiation, proving that the model has good thermal environment adaptability.

[0149] Model Comparison Analysis:

[0150] To verify the effectiveness of the method of the present invention, a comparative model comprising the following components was constructed:

[0151] Fixed threshold model: based on single early warning of wheelset temperature rise and vibration threshold;

[0152] Support Vector Machine Model: Features are extracted from meteorological and LKJ data and then input into the SVM classifier;

[0153] Single image path model: Only uses ResNet to extract wheel pair image features;

[0154] The fusion model of this invention adopts a ResNet+LSTM multimodal structure.

[0155] Table 1 shows the comparison results on the test set:

[0156] Model Name Accuracy (%) Missed report rate (%) False alarm rate (%) Fixed threshold model 81.3 11.2 7.5 SVM model 85.7 9.8 4.5 Single image model 88.9 7.3 3.8 This invention model 94.6 3.4 2

[0157] Model ablation experiment:

[0158] To evaluate the impact of each submodule on the overall model performance, image paths and temporal paths were removed separately for ablation comparison:

[0159] Table 2 Ablation Comparison Results

[0160] Model Structure Accuracy (%) Missed report rate (%) False alarm rate (%) Remove image paths (LSTM only) 89.1 6.8 4.1 Remove timing paths (ResNet only) 90.3 5.6 3.4 Preserve all paths (merge) 94.6 3.4 2

[0161] Example 2: Dynamic early warning of wheelsets under sudden rainstorms and salt spray corrosion scenarios

[0162] In this embodiment, the rail vehicle operates on a line in the southeast coastal port area, encountering a salt spray erosion environment caused by continuous heavy rain and sea breeze during operation. The system receives real-time rainfall intensity (maximum 38 mm / h), wind speed (18–22 m / s), air salinity (1.8–2.5 g / m³), and humidity (95% RH) data from the meteorological center interface at a frequency of 15 seconds / time. At the same time, the LKJ recording system collects vehicle speed (60–100 km / h), traction current fluctuation (±12%), brake temperature rise (8℃ / min), and wheelset vibration acceleration (peak value 2.8 g).

[0163] In the S2 phase, short-time interpolation is performed on all meteorological parameters with an interpolation interval of 5 seconds. The rate of change of rainfall and the standard deviation of operational fluctuations are calculated to obtain a fusion feature dataset that reflects the effects of a highly corrosive environment and dynamic load, which contains 15 feature components.

[0164] In the S3 stage, the ResNet network extracts corrosion patch distribution and oxidation color difference features from image frames (resolution 1280×720 pixels, frame rate 5fps); the LSTM network extracts temporal features, such as salinity fluctuations, humidity periodic changes, and running vibration energy trends; after feature concatenation and fully connected mapping, the model outputs a damage index D=86.2.

[0165] In the S4 stage, if D > T3 = 80, a Level III warning is triggered, generating a structured output with a label indicating accelerated corrosion risk, including wheelset number (e.g., WD-07A), damage index (86.2), risk level (Level III), and recommended actions (immediate rust prevention and surface inspection). The warning signal is pushed to the maintenance control center via a dedicated LTE-R link, with an average response delay of no more than 3 seconds.

[0166] Example 2 realizes the dynamic identification of wheelset corrosion and wear process under high humidity and salt spray environment, and completes damage trend prediction within 2 minutes after the occurrence of rainstorm, improving the response speed and risk control capability of corrosion prevention and maintenance under coastal lines; under rainstorm and salt spray environment, the system damage prediction accuracy is 93.5%, the false alarm rate is 4.3%, and the false alarm rate is 2.2%.

[0167] Example 3: Wheelset fatigue early warning under strong wind impact and low temperature frost scenarios

[0168] In this embodiment, the rail vehicle operates on a winter line in Northeast China, where the ambient temperature ranges from -25°C to -15°C and the peak gust wind speed reaches 25 m / s. The system collects meteorological data at a frequency of 10 seconds per cycle, including temperature, wind speed, air pressure, and surface icing status information. The LKJ recording system simultaneously collects traction output, wheel speed, acceleration, and bearing temperature (fluctuating within the range of 45–65°C during low-temperature operation).

[0169] In the S2 phase, the system performs time series analysis on temperature and wind speed data based on a 60-second sliding window, calculates the wind speed change rate (average 3.2 m / s²) and wheel speed fluctuation standard deviation (2.5 km / h), and extracts fused feature data representing the impact load characteristics under low temperature conditions.

[0170] In stage S3, the ResNet network extracts features such as frost cracks, ice pit distribution, and metal brightness decay on the wheelset surface; the LSTM network extracts the temporal dependencies of low temperature fluctuations, wind load impact, and acceleration changes, and after fusion in stage S34, a damage index D=65.4 is generated.

[0171] The threshold determination results of stage S4 show that T2=60<D<T3=80. The system automatically outputs a Level II warning and marks the risk of low-temperature fatigue accumulation, suggesting that surface de-icing and crack acoustic emission detection tasks be performed. The system can automatically adjust the damage prediction weights after detecting continuous low temperature for three hours, thereby improving the model sensitivity in low-temperature scenarios.

[0172] Example 3 integrates meteorological and operational data to achieve joint analysis of wind load impact and low temperature fatigue, detects microcracks caused by the superposition of temperature difference and mechanical stress in advance, and ensures the operational safety of lines in cold regions and the foresight of maintenance plans; in low temperature and strong wind scenarios, the system damage prediction accuracy is 91.8%, the false alarm rate is 5.2%, and the false alarm rate is 3.0%.

[0173] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0174] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based early warning method for rail vehicle wheelset maintenance, integrating meteorological big data and LKJ records, characterized in that... Includes the following steps: S1: Real-time acquisition of multi-source meteorological data and LKJ record data from rail vehicles; S2: Perform real-time stream processing on the multi-source data collected by S1, including data cleaning and feature extraction, to obtain fused feature data; S3: Input the fused feature data obtained in S2 into the pre-trained deep learning model and output the corresponding wheel pair damage index; S4: Based on the wheelset damage index output by S3 and the preset threshold, determine the corresponding warning level; S5: Implement corresponding maintenance response measures based on the warning level determined by S4.

2. The method for deep learning-based maintenance early warning of rail vehicle wheelsets, which integrates meteorological big data and LKJ records according to claim 1, is characterized in that... S1 specifically includes: S11: By connecting with the open interface of the National Meteorological Administration, it collects the temperature, precipitation intensity, wind speed, humidity and air salinity parameters of the target line section in real time, and updates them at a frequency of no less than once every 10 minutes; S12: Obtain supplementary micro-meteorological data through ground-deployed IoT meteorological sensor nodes, including temperature and humidity sensors and PM2.5 particulate matter sensors, with a collection frequency of no less than once every 5 minutes; S13: Call the LKJ data access module provided by the train dispatch center to receive LKJ record data of all running trains in the current section in real time, including train number, timestamp, current speed, acceleration, braking status and track section information, with a sampling interval of 10 seconds.

3. The method for deep learning-based early warning of rail vehicle wheelset maintenance, which integrates meteorological big data and LKJ records according to claim 1, is characterized in that... S2 specifically includes: S21: Align the multi-source meteorological data collected in S1 with the LKJ recorded data according to a unified timestamp and section number to form a time-consistent data stream input; S22: The sliding window mechanism is used to calculate the local mean and standard deviation of each type of sensor data, and outlier data points with absolute values ​​exceeding the set threshold are removed based on the Z-score method; S23: Set the physical boundary value range for the corresponding data type, and perform physical filtering on data that exceeds the boundary range; S24: Use a time linear interpolation completion strategy for short-term missing data points, and mark long-term missing data points as invalid and submit an abnormal data identifier; S25: Based on continuous data within each time window, extract meteorological change rate, operational fluctuation characteristics, and environmental extreme indicators to obtain fused feature data.

4. The method for deep learning-based maintenance early warning of rail vehicle wheelsets, which integrates meteorological big data and LKJ records according to claim 3, is characterized in that... S24 specifically includes: S241: When the time interval between consecutive missing data is less than a set threshold, this missing interval is marked as an interpolable interval, and the start time of the interpolation interval is set to... The end time is ; S242: Obtain the observation values ​​of the two valid data points before and after the interpolable interval, denoted as follows: and And calculate the total time difference between intervals. ; S243: Let the missing time be... ,and The corresponding estimated value is calculated using a time linear interpolation model. The formula is: ,in, For time points The interpolated estimate; Indicates a time interval; S244: All interpolated values The data is filled into the corresponding time series and marked as interpolated data.

5. The method for deep learning-based maintenance early warning of rail vehicle wheelsets, which integrates meteorological big data and LKJ records according to claim 3, is characterized in that... Specifically, S25 includes: S251: Within each sliding time window, calculate the rate of change of the temperature difference between the beginning and end of the window, as a representative feature of the meteorological rate of change. S252: Extract the speed standard deviation from the LKJ record data to characterize the degree of train operation fluctuation within the time window, and use it as an operation fluctuation feature; S253: Based on the historical statistical distribution of multi-source meteorological data, the extreme index is calculated using the Z-score method; S254: The temperature change rate, operational fluctuation characteristic value obtained from S251 to S253 are time-aligned and stitched together with multiple meteorological extreme indicators to obtain fused characteristic data.

6. The method for deep learning-based early warning of rail vehicle wheelset maintenance, which integrates meteorological big data and LKJ records according to claim 1, is characterized in that... S3 specifically includes: S31: Construct a historical sample dataset, which includes wheelset image data, fused feature data for the corresponding time period, and wheelset damage labels obtained by manual detection; S32: Input the wheelset image data into a pre-trained ResNet network to extract image features and obtain image feature vectors that reflect the surface cracks, corrosion and wear characteristics of the wheelset; S33: Input the fused feature data of the corresponding time period into the LSTM network to extract the temporal feature representation of the environment and operating status; S34: The image feature vector is concatenated and fused with the temporal feature vector output by the LSTM, and then mapped to the regression output layer through a fully connected layer to predict the wheelset damage index.

7. The method for deep learning-based maintenance early warning of rail vehicle wheelsets, which integrates meteorological big data and LKJ records according to claim 6, is characterized in that... Specifically, S32 includes: S321: Preprocess the wheelset image data, including removing fixed image borders and watermark information, using a denoising algorithm to suppress high-frequency noise, performing brightness and contrast normalization processing, and maintaining consistent pixel aspect ratio. S322: Perform wheelset target region localization on the preprocessed image, identify the tread and wheel flange regions based on edge detection and geometric constraints, generate the corresponding region of interest, and crop to obtain the target sub-image containing wheelset surface information; S323: Perform size and channel normalization on the target sub-image, scale the image to the same size as the input of the pre-trained ResNet network, and perform numerical normalization and standardization with fixed mean and variance by channel to form a tensor that meets the requirements of the model input. S324: Input the normalized image tensor into the pre-trained ResNet network, perform forward inference, and extract the intermediate layer feature maps; S325: Perform global average pooling and channel compression on the intermediate layer feature map to obtain a fixed-length image feature vector; S326: Perform amplitude normalization and anomaly detection on the obtained image feature vector, remove invalid features caused by blanks or occlusions, and mark the valid features as image feature vectors reflecting the surface cracks, corrosion and wear characteristics of the wheelset.

8. The method for deep learning-based maintenance early warning of rail vehicle wheelsets, which integrates meteorological big data and LKJ records according to claim 7, is characterized in that... S33 specifically includes: S331: Based on the set time window and step size, the fused feature data of the corresponding time period are framed in order according to the timestamp to form a feature sequence with a fixed order; S332: Perform mean and variance normalization on each feature dimension of the sequence in the same way as during model training to ensure consistent input scale; S333: For sequences that are shorter than the window length, use mask alignment to maintain consistent sequence lengths within the batch and generate corresponding mask identifiers; S334: Input the normalized feature sequence and mask into the LSTM network, perform forward computation step by step, and obtain the hidden state sequence at each time step; S335: Perform terminal time extraction or temporal pooling on the hidden state sequence to obtain a fixed-length temporal feature vector; S336: Perform amplitude normalization and validity verification on the time series feature vector, and output the time series feature representation of the environment and operating state.

9. A method for early warning of rail vehicle wheelset maintenance based on deep learning, integrating meteorological big data and LKJ records, as described in claim 8, is characterized in that... S34 specifically includes: S341: Receive the image feature vector obtained from S32 and the temporal feature vector obtained from S33, and perform scale unification and dimension calibration on the two; S342: The calibrated image feature vector and the temporal feature vector are concatenated on the feature axis to form a single joint feature vector; S343: Perform the set regularization and normalization processing on the joint feature vector, and output the fused stable input; S344: The fused stable input is sequentially fed into the fully connected layer group for nonlinear mapping and feature compression to obtain an intermediate representation for regression calculation; S345: Input the intermediate representation into the regression output layer to generate the corresponding wheelset damage index. .

10. A method for deep learning-based early warning of rail vehicle wheelset maintenance, based on meteorological big data and LKJ records, as described in claim 9, is characterized in that... S4 specifically includes: S41: Receiver wheelset damage index It is then compared with a preset multi-level risk threshold set, which includes four level boundary values, namely... , , , ,satisfy ; S42: Based on wheelset damage index The corresponding warning level is determined based on the numerical range in which the value falls. The judgment rules include: when At that time, it is set to level 0; when At that time, it is set as a Level I warning; when At that time, it was set as a Level II warning; when At that time, it was set as a Level III warning; when At that time, it was set as a Level IV warning; S43: Output the corresponding warning level in a structured format, along with the current wheelset number, damage index value, warning level label, and corresponding risk advice text.

Images