Intelligent diagnosis method and system for railway locomotive wheel tread damage based on multi-source information fusion

Abstract

This application belongs to the field of railway locomotive wheel damage detection technology, and relates to an intelligent diagnostic method and system for railway locomotive wheel tread damage through multi-source information fusion. It comprehensively extracts time-domain, frequency-domain, and time-frequency-domain features from collected multi-source data to form a multi-dimensional feature vector. Simultaneously, it constructs a working condition weight function based on working condition features, and corrects the spectral energy by combining the influence of different working conditions on the damage signal spectrum. Then, through a working condition adaptive attention submodule, it calculates the attention weights of corresponding features of each sensor based on the working condition features, and generates preliminary fusion features adapted to the working conditions after weighted summation. Next, through a sensor interaction modeling submodule, it analyzes the correlation between the features of each sensor to generate corrected features. The preliminary fusion features and corrected features are then weighted and fused according to a preset ratio to obtain the final fusion features. This ensures that the fusion features can accurately reflect the damage state under different working conditions, effectively reducing the risk of misjudgment caused by working condition fluctuations.

Description

Technical Field

[0001] This application belongs to the field of railway locomotive wheel damage detection technology, and more specifically, relates to an intelligent diagnostic method and system for railway locomotive wheel tread damage by multi-source information fusion. Background Technology

[0002] As the core transportation tool of rail transit, the operational safety of railway locomotives is directly related to the efficiency of rail network transportation and the safety of personnel and property. As the direct contact component between the locomotive and the track, the wheel tread is prone to damage such as peeling and cracking due to friction and impact during long-term operation. If it is not diagnosed in a timely and accurate manner, it will aggravate wheel and rail wear and even cause serious safety accidents such as derailment, while significantly increasing maintenance costs and downtime losses. Therefore, achieving high-precision, real-time intelligent diagnosis of wheel tread damage is a key technical problem that urgently needs to be solved in the field of railway transportation.

[0003] Currently, wheel damage diagnosis technology has gradually evolved from traditional manual inspection and offline testing to online intelligent diagnosis. Traditional manual inspection relies on the experience of maintenance personnel, which is inefficient, subjective, and difficult to adapt to the inspection needs of high-density locomotive operation. Offline testing requires the locomotive to be stopped and tested with specialized equipment. Although it has high accuracy, it affects transportation scheduling and cannot achieve real-time damage warning. In order to overcome the above defects, sensor-based online diagnostic methods have emerged in the existing technology. These methods identify damage by collecting single-dimensional data. However, such methods can only capture local features of damage, are easily affected by fluctuations in locomotive operating conditions, have poor diagnostic robustness, and are difficult to cope with the damage identification needs in complex environments. Summary of the Invention

[0004] This invention provides a method and system for intelligent diagnosis of railway locomotive wheel tread damage based on multi-source information fusion, aiming to solve the technical problem that current locomotive wheel tread damage diagnosis can only capture local features of the damage and is easily affected by fluctuations in locomotive operating conditions.

[0005] On the one hand, this invention provides an intelligent diagnostic method for railway locomotive wheel tread damage based on multi-source information fusion, comprising the following steps: Multi-source data acquisition: Multiple sensors are deployed near the wheels of railway locomotives to simultaneously collect time-series signals related to wheel surface damage, and locomotive operating condition parameters are obtained in real time from the locomotive control system. A unified time source is used to ensure that the timestamps of the sensor-collected signals and operating condition parameters are consistent. Data preprocessing: Outlier removal and noise filtering are performed on the collected raw time series signals and operating parameters respectively. After outlier removal, gaps are filled by the mean of adjacent normal data. After noise filtering, irrelevant interference signals are eliminated. Then, different types of data are standardized to a unified dimension to obtain the preprocessed time series signals and preprocessed operating parameters. Feature extraction: Extract time-domain features, frequency-domain features, and time-frequency-domain features from the preprocessed time series signal, and extract features from the preprocessed operating parameters. Depending on the type of the operating parameters, discretization processing or direct use as operating features is selected. Based on the extracted operating features, the frequency-domain features are adaptively adjusted to adapt to the current locomotive operating conditions, forming a feature vector set grouped by signal type and operating features. Multi-source information fusion: The constructed attention mechanism integrates the grouped feature vector set. The constructed attention mechanism includes a working condition adaptive attention submodule and a sensor interaction modeling submodule. First, the working condition adaptive attention submodule calculates the attention weight of the corresponding features of each sensor in combination with the working condition features. After weighted summation, a preliminary fusion feature adapted to the working condition is generated. Then, the sensor interaction modeling submodule analyzes the correlation between the features of each sensor and generates a corrected feature. The preliminary fusion feature and the corrected feature are weighted and fused according to a preset ratio to obtain the final fusion feature. Damage diagnosis: The final fused features are input into a pre-trained deep learning classifier. The classifier analyzes the features layer by layer and outputs the type and severity of wheel tread damage. The reliability of the result is judged by the confidence score output by the classifier, and a diagnostic result containing damage information and reliability assessment is generated.

[0006] This invention lays the foundation for operational condition correlation analysis by simultaneously collecting multi-source data and operational condition data. Based on the collected multi-source data, it comprehensively extracts time-domain, frequency-domain, and time-frequency-domain features to form multi-dimensional feature vectors. Simultaneously, it constructs an operational condition weighting function based on the extracted operational condition features, and corrects the spectral energy by combining the influence of different operational conditions on the spectrum of the damage signal, making the features adaptable to the current operating conditions and avoiding the limitations of single-dimensional features. Then, through the operational condition adaptive attention submodule, it calculates the attention weights of corresponding features of each sensor based on the operational condition features, and generates preliminary fusion features adapted to the operational conditions after weighted summation. Next, through the sensor interaction modeling submodule, it analyzes the correlation between the features of each sensor to generate corrected features. The preliminary fusion features and corrected features are then weighted and fused according to a preset ratio to obtain the final fusion features. This makes the fusion features more discriminative, ensuring that the fusion features can accurately reflect the damage state under different operating conditions and effectively reducing the risk of misjudgment caused by operational condition fluctuations.

[0007] Preferably, the time-domain features include mean, variance, root mean square, peak factor, and kurtosis; the frequency-domain features include spectral peak value, spectral energy, frequency center, and frequency variance; the time-frequency domain features are extracted through wavelet transform, specifically including wavelet energy spectrum, wavelet entropy, and wavelet peak coefficient.

[0008] Preferably, the adaptive adjustment in feature extraction is achieved through a constructed working condition weight function, wherein the working condition weight function is based on the extracted working condition features and combined with the influence law of different working conditions on the spectrum of wheel tread damage signal to determine the weight coefficient of the spectrum energy of each frequency band, and uses the weight coefficient to make targeted corrections to the spectrum energy in the frequency domain features to match the spectrum energy distribution characteristics under the current working condition.

[0009] Preferably, the working condition adaptive attention submodule maps the extracted working condition features into a weight matrix corresponding to each sensor feature through a two-layer fully connected network, calculates the attention weight of each sensor feature using the weight matrix, and obtains the preliminary fused features after weighting each sensor feature.

[0010] Preferably, the sensor interaction modeling submodule treats the sensors as independent graph nodes, calculates the initial edge weights between nodes based on the Pearson correlation coefficient to characterize the correlation strength of each sensor feature, and then learns and optimizes the attention coefficients between nodes through a graph attention network to perform weighted aggregation of each node feature, eliminates the interaction interference between sensors, and obtains the corrected features.

[0011] On the other hand, the present invention provides a multi-source information fusion-based intelligent diagnostic system for railway locomotive wheel tread damage, used to implement the multi-source information fusion-based intelligent diagnostic method for railway locomotive wheel tread damage described in the present invention, wherein the system includes: Data acquisition module: includes multiple types of deployed sensors that synchronously collect time-series signals and operating parameters related to the wheel tread; Data preprocessing module: used to perform outlier removal, noise filtering and standardization on the collected raw data, and output preprocessed data; Feature extraction module: used to extract time-domain features, frequency-domain features and time-frequency-domain features from preprocessed time series signals. It performs feature extraction on preprocessed operating parameters and selects discretization processing or direct use as operating feature based on parameter type. Based on the extracted operating features, it adaptively adjusts the frequency-domain features to adapt to the current operating conditions and generates a grouped feature vector set. Multi-source information fusion module: Based on the constructed attention mechanism, it integrates grouped feature vectors and outputs the final fused features; Damage diagnosis module: This module processes the final fused features using a deep learning classifier and outputs the damage diagnosis results and confidence levels.

[0012] Preferably, the multi-source information fusion module includes a working condition adaptive attention submodule and a sensor interaction modeling submodule. The working condition adaptive attention submodule calculates the sensor feature attention weights by combining working condition features and generates preliminary fusion features by weighted summation. The sensor interaction modeling submodule analyzes the correlation between sensor features to eliminate interaction interference and generate corrected features. The initial fusion features and the modified features are weighted and fused according to a preset ratio to output the final fusion features.

[0013] Preferably, the working condition adaptive attention submodule is configured with a two-layer fully connected network to map the working condition features into a weight matrix corresponding to each sensor feature, calculate the attention weight through the weight matrix, and sum the sensor features by weight.

[0014] Preferably, the sensor interaction modeling submodule is configured with a graph attention network, which treats each sensor as an independent graph node, calculates the initial edge weights between nodes based on the Pearson correlation coefficient to characterize the association strength, optimizes the attention coefficients through network learning and weighted aggregates the node features to eliminate interaction interference between sensors.

[0015] Preferably, the deep learning classifier in the damage diagnosis module includes a multilayer perceptron classifier and a convolutional neural network classifier.

[0016] The beneficial effects of the invention include: This invention lays the foundation for operational condition correlation analysis by simultaneously collecting multi-source data and operational condition data. Based on the collected multi-source data, it comprehensively extracts time-domain, frequency-domain, and time-frequency-domain features to form multi-dimensional feature vectors. Simultaneously, it constructs an operational condition weighting function based on the extracted operational condition features, and corrects the spectral energy by combining the influence of different operational conditions on the spectrum of the damage signal, making the features adaptable to the current operating conditions and avoiding the limitations of single-dimensional features. Then, through the operational condition adaptive attention submodule, it calculates the attention weights of corresponding features of each sensor based on the operational condition features, and generates preliminary fusion features adapted to the operational conditions after weighted summation. Next, through the sensor interaction modeling submodule, it analyzes the correlation between the features of each sensor to generate corrected features. The preliminary fusion features and corrected features are then weighted and fused according to a preset ratio to obtain the final fusion features. This makes the fusion features more discriminative, ensuring that the fusion features can accurately reflect the damage state under different operating conditions and effectively reducing the risk of misjudgment caused by operational condition fluctuations. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The above is a flowchart illustrating the overall steps of an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating the multi-source information fusion process provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0021] Example 1 See Figure 1 As shown, this embodiment provides an intelligent diagnostic method for railway locomotive wheel tread damage based on multi-source information fusion, including the following steps: Multi-source data acquisition: One vibration sensor is deployed at the wheel axle box of each locomotive car to capture mechanical impact vibrations caused by tread spalling and cracks; one acoustic sensor is deployed 12 to 15 cm away from the wheel flange on the side of the wheel to collect high-frequency noise generated by damage; one temperature sensor is attached to the surface of the wheel spokes to monitor the correlation between frictional heat accumulation and damage. In this embodiment, all sensors are encapsulated with vibration-proof and waterproof technology (protection level of IP67) to adapt to the high temperature, high humidity and strong vibration environment of railway operation.

[0022] Operating parameters are read in real time via the RS485 interface of the locomotive control system, including operating speed, load weight, and ambient temperature. The acquisition frequency is synchronized with the sensor data to ensure that each frame of sensor data corresponds to a unique operating parameter. Furthermore, a high-precision clock module inside the locomotive is used as a unified time source to add timestamps to all sensor signals and operating parameters, ensuring data correlation.

[0023] In this embodiment, multiple types of sensors are deployed and collected synchronously, covering the mechanical, acoustic, and thermal characteristics of wheel tread damage. Combined with operating parameters, this provides context for subsequent adaptive diagnosis, solving the problem of insufficient data dimensions from a single sensor.

[0024] Data preprocessing: The interquartile range (IQR) is used to process the sensor time series signals. The upper quartile (Q3), lower quartile (Q1), and interquartile range (IQR) of each sensor data are calculated. Data that exceed the range of [Q1-1.5IQR, Q3+1.5IQR] are identified as outliers. Outliers are filled with the mean of three adjacent normal data to avoid data sequence breakpoints affecting subsequent signal analysis.

[0025] For operating parameters, threshold screening is set, for example: operating speed 0-350km / h, load weight 0-100t, ambient temperature -40℃-60℃. Parameters exceeding the threshold are marked as invalid data, and the sensor data of the corresponding time block is removed simultaneously.

[0026] In this embodiment, a differentiated filtering strategy is adopted for the noise characteristics of different sensors. Vibration sensor data contains mechanical resonance noise, so a sliding window averaging method with a window size of 10 sampling points is used to smooth the data, updating it every 5ms. This filters out high-frequency random noise while retaining the abrupt changes in the impact of damage. For acoustic sensor data, which is affected by ambient background noise, a 3-level wavelet decomposition is performed using the db4 wavelet basis function. The high-frequency noise coefficients are set to zero before reconstructing the signal, retaining the characteristic frequency components corresponding to the damage. For temperature sensor data, which changes slowly, a first-order sliding average method (window size of 5 sampling points) is used in this embodiment to smooth the data and eliminate interference from minute fluctuations.

[0027] Then, the Z-score standardization method is used to convert all sensor signals and operating parameters into a standard form with a mean of 0 and a variance of 1. The mean (μ) and standard deviation (σ) are calculated by grouping according to sensor type and parameter type. For example, the statistics of vibration signals are calculated separately to avoid cross-interference between different types of data.

[0028] Feature extraction: In this embodiment, feature extraction includes time-domain features, frequency-domain features, and time-frequency-domain features; The time-domain features include the mean, variance, root mean square, peak factor, and kurtosis calculated from vibration, acoustic, and temperature signals. The peak factor and kurtosis are sensitive to impact signals caused by tread abrasion and cracks, and can capture early minor damage. The mean and variance of the temperature signal are used to reflect the degree of frictional heat accumulation, indirectly characterizing the severity of damage.

[0029] The frequency domain features are obtained by performing a fast Fourier transform on the vibration and acoustic signals to obtain the power spectrum, and then extracting the spectral peak, spectral energy, frequency center, and frequency variance. The spectral peak corresponds to the resonance frequency caused by the damage, and the frequency center reflects the shift of the spectral centroid, which can distinguish the frequency features of different types of damage.

[0030] Then, wavelet packet decomposition (db4 wavelet basis, three-level decomposition) is used to process the vibration and acoustic signals, extracting wavelet energy spectrum, wavelet entropy, and wavelet peak coefficients. Based on this, time-frequency domain features are obtained. Among them, wavelet entropy reflects the signal complexity. The entropy value increases significantly when damage occurs, which is suitable for damage identification of non-stationary signals.

[0031] Regarding the operating conditions, in this embodiment, the operating speed is discretized into three levels: low speed (v < 30 km / h), medium speed (30 ≤ v < 80 km / h), and high speed (v ≥ 80 km / h). The load weight is also discretized into three levels: light load (< 30 t), medium load (30-60 t), and heavy load (> 60 t). The ambient temperature is directly used as a continuous feature to form the operating condition feature vector.

[0032] Furthermore, in this embodiment, to improve the robustness of frequency domain features to changes in operating conditions, a weighting function for operating conditions is used. The amplitude at each frequency point in the power spectrum is adaptively weighted and then integrated to obtain the corrected spectral energy, which compensates for the frequency distribution shift of the damage characteristics under different operating conditions, as follows: Definition and trend of working condition weighting function: The frequency point weighting coefficients corresponding to the two-dimensional operating conditions (speed v, load l) are implemented based on a two-dimensional lookup table trained on historical data, and are adapted to different operating conditions according to frequency bands. In this embodiment, the value selection rules are as follows: High load (>60t), low speed (<30km / h): Damage impact energy is concentrated at low frequencies, therefore the 0-500Hz frequency band. Frequency bands above 500Hz ; Low load (<30t), high speed (≥80km / h): Damage characteristics shift to mid-to-high frequencies, therefore the 500-2000Hz frequency band. Other frequency bands ; Medium load (30-60t), medium speed (30≤v<80km / h): Stable operating conditions, full frequency band. .

[0033] and Linear interpolation is used to cover the continuous operating range, ensuring a smooth transition of weights.

[0034] The spectral energy is corrected based on the operating condition weighting function, as follows: ; In the formula: This represents the spectral energy after adaptive adjustment under operating conditions; This represents the power spectral amplitude at frequency f; , To pre-determine the analysis frequency band, in this embodiment, the frequency range of 0-5000Hz is not included, covering the typical characteristic frequencies of wheel damage.

[0035] In this embodiment, the weighting function The characteristic contribution of each frequency point under different operating conditions is calibrated. When the low-frequency weight is increased, the low-frequency band is strengthened. Contribution to total energy; when the mid-to-high frequency weighting is increased, the corresponding frequency band is amplified. The impact of this makes Maintaining a consistent response to the same damage under different operating conditions solves the problem of feature distribution shift caused by changes in operating conditions and improves feature robustness.

[0036] Finally, the extracted features are grouped according to signal type to form vibration feature vector, acoustic feature vector, temperature feature vector set, and working condition feature vector, and stored using NumPy arrays, with each row corresponding to a feature set of a 10-second data block.

[0037] Multi-source information fusion: In this implementation, a dual-attention mechanism module is used to integrate multi-source features to address the issues of operating condition changes and sensor interaction interference. See [link to relevant documentation]. Figure 2 As shown, the details are as follows: The working condition adaptive attention submodule maps the working condition feature vector into a weight matrix through two fully connected layers. The activation function used is the ReLU activation function to ensure that the weights are non-negative and have non-linear fitting ability. The number of neurons in the first fully connected layer is twice the dimension of the working condition feature, and the second layer is consistent with the dimension of the sensor feature.

[0038] Then, based on the weight matrix, the attention weights of each sensor feature are calculated using the softmax function. The specific calculation formula is as follows: ; In the formula: This represents the attention weight for the i-th type of sensor; Represents the weight matrix The submatrix of the features of the i-th type of sensor; This represents the feature vector of the i-th type of sensor; The types of sensors are identified as vibration sensors, acoustic sensors, and temperature sensors. Represents the weight matrix The submatrix of the k-th type of sensor features; in this embodiment, the vibration feature weights under high load conditions. Increased to 0.4-0.5, acoustic feature weights under high-speed operating conditions. Increased to 0.4-0.5; Temperature characteristic weights under extreme temperature conditions. Increase to 0.3-0.4; The feature vectors of each sensor are weighted and summed according to their corresponding attention weights to obtain the preliminary fused features, as shown in the following expression: ; In the formula: Indicates preliminary fusion characteristics; Represents the vibration eigenvector; Represents the acoustic eigenvector; This represents the temperature eigenvector.

[0039] The sensor interaction modeling submodule treats vibration, acoustic, and temperature sensors as three nodes in a graph structure, with each node's feature being a vibration feature vector. Acoustic feature vectors and temperature eigenvectors Based on the Pearson correlation coefficient The initial edge weights between nodes are calculated to characterize the correlation strength of sensor features. The specific calculation formula is as follows: ; In the formula: Indicates feature dimension; This represents the k-th component of the feature of the i-th node; This represents the mean of the features of the i-th node; A stronger positive correlation indicates a more pronounced interaction, while a negative correlation indicates complementary features.

[0040] Then, by optimizing the variable weights in a single attention layer of a Graph Attention Network (GAT), the attention coefficients are learned. The specific expression is as follows: ; In the formula: Represents the attention vector; Represents the shared feature transformation matrix; This represents a vector concatenation operation; Represents the set of adjacent nodes of node i; The activation function retains negative weight interaction information; This represents the transpose of the attention vector; This represents the node feature vector corresponding to the i-th sensor, which corresponds to the vibration, acoustic, and temperature sensors, respectively. This represents the feature vector corresponding to the neighboring node j of node i; Represents the set of adjacent nodes of node i. The feature vector corresponding to any node k in the vector.

[0041] The learned attention coefficients are used to weight and aggregate the features of each node to obtain the interaction correction features of each node. : ; In the formula: This represents the sigmoid activation function, which normalizes the weighted aggregated feature values ​​to the [0,1] interval.

[0042] Finally, the initial fusion features will be... The mean of the interactive correction features is weighted and fused according to a preset ratio, and the fusion weight is... The contribution of balancing operating condition adaptation and interactive correction is determined through training with historical data, and the specific expression is as follows: ; In the formula: Indicates the final fusion characteristics; This indicates the interactive correction characteristics of the vibration sensor; This represents the interactive correction characteristics of acoustic sensors; This indicates the interactive correction characteristics of the temperature sensor.

[0043] In this embodiment, the dual-module design addresses the issues of working condition adaptation and sensor interaction interference, thereby improving the discriminative power of the final fused features.

[0044] Damage diagnosis: The fused features are analyzed by a deep learning classifier to output the diagnostic results. In this embodiment, the deep learning classifier includes a multilayer perceptron and a convolutional neural network classifier. The multilayer perceptron is used for real-time diagnosis on the vehicle side. It includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The first hidden layer has 64 neurons, the second hidden layer has 32 neurons, and the activation function is ReLU. The output layer has 4 neurons, corresponding to 4 states: no damage, slight peeling, medium fracture, and severe crack. The output probability distribution is output using the Softmax activation function.

[0045] The convolutional neural network classifier is used for ground center depth diagnosis. It reconstructs the final fused features into an 8×N two-dimensional feature map, where N is the dimension of the final fused features / 8. It sets two convolutional layers with a convolution sum of 3×3 and a stride of 1. It also sets two max pooling layers with a pooling kernel size of 2×2 to extract local correlation features. Subsequently, it connects one fully connected layer (32 neurons) to the output layer to achieve accurate identification of early minor damage.

[0046] The specific type of deep learning classifier mentioned above is not a limitation of this application. The focus of this application is on the construction process of the final fused features. Therefore, the type of deep learning classifier mentioned above is only an exemplary technical solution. As for the training of the model, given the above technical solution, how to train the model is a conventional technical means in this field. Therefore, no further details will be made in this embodiment.

[0047] Example 2 This embodiment provides an intelligent diagnostic system for railway locomotive wheel tread damage based on multi-source information fusion, used to implement the above-described method embodiment. The system includes a data acquisition module, a data preprocessing module, a feature extraction module, a multi-source information fusion module, and a damage diagnosis module.

[0048] The data acquisition module is used to simultaneously acquire multi-source signals and operating parameters, specifically including: Sensor unit: includes vibration sensor, acoustic sensor, and temperature sensor. Each wheel corresponds to a set of sensor groups, which are deployed in the same location as in the method embodiment. The sensor outputs analog signals, which are converted into 16-bit digital signals by a built-in AD converter. Operating Parameter Acquisition Unit: Adopts RS485 communication module, interfaces with locomotive TCMS system, and reads operating speed, load weight and ambient temperature in real time to ensure synchronization with sensor data; Time synchronization unit: Based on the clock module built into the locomotive system, timestamps are added to sensor signals and operating parameters to avoid data timing misalignment; By deploying and synchronizing multiple sensors, the system is provided with comprehensive and consistent raw data, laying the foundation for subsequent data processing.

[0049] The data preprocessing module is used for data cleaning and standardization, and includes an outlier handling unit, a noise filtering unit, and a standardization unit, which are integrated into the vehicle-mounted embedded processor. Outlier handling unit: Built-in quartile algorithm and operating parameter threshold filtering rules, automatically identify and remove outlier data, and use mean imputation strategy to repair data breakpoints to ensure data integrity; Noise filtering unit: Preset sliding window averaging method, wavelet denoising method, and first-order moving average method, automatically switching according to sensor type to eliminate different noise interferences in a targeted manner; Standardization Unit: Implements the Z-score standardization algorithm, calculates statistics by grouping data types, and outputs standardized data with unified dimensions, providing consistent input for the feature extraction module.

[0050] The feature extraction module is used to extract multi-dimensional features and adaptively adjust them, including a time-domain feature extraction unit, a frequency-domain feature extraction unit, a time-frequency-domain feature extraction unit, a working condition feature processing unit, and a working condition adaptive adjustment unit; The time-domain, frequency-domain, and time-frequency-domain feature extraction units have built-in algorithms such as mean, variance, FFT, and wavelet packet decomposition (see the method implementation for details) to automatically extract corresponding features from the preprocessed data and output a highly representative feature set. The working condition feature processing unit discretizes or directly extracts the working condition parameters to generate working condition feature vectors, providing a basis for adaptive adjustment. The adaptive adjustment unit adjusts the spectral energy dynamically according to the current operating condition characteristics based on a pre-trained operating condition weight function lookup table, and outputs optimized features that are suitable for the operating condition. In this embodiment, damage information is captured comprehensively through multi-dimensional feature extraction, and the working condition adaptive adjustment unit counteracts working condition interference, improves feature robustness, and provides high-quality feature input for multi-source fusion.

[0051] The multi-source information fusion module includes a working condition adaptive attention submodule, a sensor interaction modeling submodule, and a fusion submodule; The working condition adaptive attention submodule is configured with a two-layer fully connected network to realize the mapping of working condition features to weight matrix, calculate the attention weight of sensor features, output preliminary fused features, and dynamically adapt to changes in working conditions. The sensor interaction modeling submodule configures a graph attention network, constructs a sensor node graph, calculates initial edge weights based on the Pearson correlation coefficient, learns and optimizes attention coefficients, outputs interaction correction features, and eliminates sensor coupling interference. Fusion submodule: The initial fusion features and the average of the interactive correction features are weighted according to a preset ratio, and the final fusion features are output.

[0052] In this embodiment, the two sub-modules work together to solve the problems of working condition adaptation and sensor interaction, respectively. Finally, they merge and balance the contributions of the two to generate highly discriminative features, providing core support for accurate diagnosis.

[0053] The damage diagnosis module includes a model storage unit, an inference unit, and an online fine-tuning unit; The model storage unit is used to store the optimal parameters of pre-trained MLP and CNN classifiers, and supports model calls and updates; The inference unit is used to load the classifier model, perform feature verification, forward inference, confidence determination, and output diagnostic results. The online fine-tuning unit is used to collect incremental data, perform small-batch fine-tuning, update model weights, and adapt to dynamic changes in operating conditions.

[0054] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent diagnosis of railway locomotive wheel tread damage based on multi-source information fusion, characterized in that, Includes the following steps: Multi-source data acquisition: Multiple sensors are deployed near the wheels of railway locomotives to simultaneously collect time-series signals related to wheel surface damage, and locomotive operating condition parameters are obtained in real time from the locomotive control system. A unified time source is used to ensure that the timestamps of the sensor-collected signals and operating condition parameters are consistent. Data preprocessing: Outlier removal and noise filtering are performed on the collected raw time series signals and operating parameters respectively. After outlier removal, gaps are filled by the mean of adjacent normal data. After noise filtering, irrelevant interference signals are eliminated. Then, different types of data are standardized to a unified dimension to obtain the preprocessed time series signals and preprocessed operating parameters. Feature extraction: Extract time-domain features, frequency-domain features, and time-frequency-domain features from the preprocessed time series signal, and extract features from the preprocessed operating parameters. Depending on the type of the operating parameters, discretization processing or direct use as operating features is selected. Based on the extracted operating features, the frequency-domain features are adaptively adjusted to adapt to the current locomotive operating conditions, forming a feature vector set grouped by signal type and operating features. Multi-source information fusion: The constructed attention mechanism integrates the grouped feature vector set. The constructed attention mechanism includes a working condition adaptive attention submodule and a sensor interaction modeling submodule. First, the working condition adaptive attention submodule calculates the attention weight of the corresponding features of each sensor in combination with the working condition features. After weighted summation, a preliminary fusion feature adapted to the working condition is generated. Then, the sensor interaction modeling submodule analyzes the correlation between the features of each sensor and generates a corrected feature. The preliminary fusion feature and the corrected feature are weighted and fused according to a preset ratio to obtain the final fusion feature. Damage diagnosis: The final fused features are input into a pre-trained deep learning classifier. The classifier analyzes the features layer by layer and outputs the type and severity of wheel tread damage. The reliability of the result is judged by the confidence score output by the classifier, and a diagnostic result containing damage information and reliability assessment is generated.

2. The intelligent diagnostic method for railway locomotive wheel tread damage based on multi-source information fusion according to claim 1, characterized in that, The time-domain features include mean, variance, root mean square, peak factor, and kurtosis; the frequency-domain features include spectral peak value, spectral energy, frequency center, and frequency variance; the time-frequency domain features are extracted through wavelet transform, specifically including wavelet energy spectrum, wavelet entropy, and wavelet peak coefficient.

3. The intelligent diagnostic method for railway locomotive wheel tread damage based on multi-source information fusion according to claim 1, characterized in that, The adaptive adjustment in feature extraction is achieved through a constructed working condition weight function. The working condition weight function is based on the extracted working condition features and combines the influence of different working conditions on the spectrum of wheel tread damage signals to determine the weight coefficient of the spectrum energy of each frequency band. The weight coefficient is used to make targeted corrections to the spectrum energy in the frequency domain features to match the spectrum energy distribution characteristics under the current working condition.

4. The intelligent diagnostic method for railway locomotive wheel tread damage based on multi-source information fusion according to claim 1, characterized in that, The working condition adaptive attention submodule maps the extracted working condition features into weight matrices corresponding to the features of each sensor through a two-layer fully connected network. The attention weights of each sensor feature are calculated using the weight matrices, and the preliminary fused features are obtained after weighting each sensor feature.

5. The intelligent diagnostic method for railway locomotive wheel tread damage based on multi-source information fusion according to claim 1, characterized in that, The sensor interaction modeling submodule treats sensors as independent graph nodes, calculates the initial edge weights between nodes based on the Pearson correlation coefficient to characterize the correlation strength of each sensor feature, and then learns and optimizes the attention coefficients between nodes through a graph attention network to perform weighted aggregation of the features of each node, eliminates the interaction interference between sensors, and obtains the corrected features.

6. A multi-source information fusion-based intelligent diagnostic system for railway locomotive wheel tread damage, characterized in that, The intelligent diagnostic method for multi-source information fusion of railway locomotive wheel tread damage as described in any one of claims 1 to 5, wherein the system comprises: Data acquisition module: includes multiple types of deployed sensors that synchronously collect time-series signals and operating parameters related to the wheel tread; Data preprocessing module: used to perform outlier removal, noise filtering and standardization on the collected raw data, and output preprocessed data; Feature extraction module: used to extract time-domain features, frequency-domain features and time-frequency-domain features from preprocessed time series signals. It performs feature extraction on preprocessed operating parameters and selects discretization processing or direct use as operating feature based on parameter type. Based on the extracted operating features, it adaptively adjusts the frequency-domain features to adapt to the current operating conditions and generates a grouped feature vector set. Multi-source information fusion module: Based on the constructed attention mechanism, it integrates grouped feature vectors and outputs the final fused features; Damage diagnosis module: This module processes the final fused features using a deep learning classifier and outputs the damage diagnosis results and confidence levels.

7. The intelligent diagnostic system for railway locomotive wheel tread damage based on multi-source information fusion according to claim 6, characterized in that, The multi-source information fusion module includes a working condition adaptive attention submodule and a sensor interaction modeling submodule. The working condition adaptive attention submodule calculates the sensor feature attention weights by combining working condition features and generates preliminary fusion features by weighted summation. The sensor interaction modeling submodule analyzes the correlation between sensor features to eliminate interaction interference and generate corrected features. The initial fusion features and the modified features are weighted and fused according to a preset ratio to output the final fusion features.

8. The intelligent diagnostic system for railway locomotive wheel tread damage based on multi-source information fusion according to claim 7, characterized in that, The working condition adaptive attention submodule is configured with a two-layer fully connected network to map working condition features into weight matrices corresponding to the features of each sensor. Attention weights are calculated through the weight matrices and the sensor features are weighted and summed.

9. The intelligent diagnostic system for railway locomotive wheel tread damage based on multi-source information fusion according to claim 7, characterized in that, The sensor interaction modeling submodule configures a graph attention network to treat each sensor as an independent graph node, calculates the initial edge weights between nodes based on the Pearson correlation coefficient to characterize the association strength, optimizes the attention coefficients through network learning and weighted aggregation of node features to eliminate interaction interference between sensors.

10. The intelligent diagnostic system for railway locomotive wheel tread damage based on multi-source information fusion according to claim 6, characterized in that, The deep learning classifiers in the damage diagnosis module include multilayer perceptron classifiers and convolutional neural network classifiers.

Images