A data-driven method for fault diagnosis of heavy truck electric drive axles

Abstract

The application discloses a heavy truck electric drive axle fault diagnosis method based on data driving, first collects the running state data of the heavy truck electric drive axle in various use scenarios, and labels the fault categories to construct an original data set; a generative adversarial network based on weight noise is used for sample generation, data expansion is realized, and the expanded data and the original data are combined to form a training data set; a self-encoder based on boundary smoothing is selected as a feature dimension reduction model, and the feature dimension reduction model is trained by using the training data set; a probabilistic neural network based on quantum probability guidance is used as a classifier model, and the classifier model is trained by using the features output by the feature dimension reduction model; finally, the original data of the heavy truck electric drive axle to be evaluated are collected, and are input into the trained feature dimension reduction model for feature processing; then, the processed features are input into the trained classifier model for classification, and a classification result is obtained. The application can realize high-precision and high-robustness of electric drive axle fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of heavy truck electric drive axle fault diagnosis, and in particular to a data-driven heavy truck electric drive axle fault diagnosis method. Background Art

[0002] As a crucial power transmission device for heavy-duty trucks, the stability and reliability of electric axles are directly linked to vehicle safety and service life. However, in diverse scenarios, such as high-speed driving, complex urban traffic, and inclement weather, heavy-duty truck electric axles are susceptible to failures due to long-term loads, environmental fluctuations, and component aging. These failures can include bearing damage, motor failure, sensor failure, brake system issues, and tire anomalies. These issues not only impact vehicle performance but can also lead to serious economic losses and safety risks.

[0003] The current methods for solving heavy truck electric drive axle faults have the following shortcomings:

[0004] 1. In the electric drive axle fault diagnosis task, the traditional generative adversarial network lacks effective diversity control in data generation, and the generated data is prone to pattern collapse. When the training samples are insufficient, the generated data has poor adaptability to complex working conditions.

[0005] 2. In the electric drive axle fault diagnosis task, the existing feature dimensionality reduction model is difficult to take into account both the compression of high-dimensional data and the processing of boundary anomalies, resulting in insufficient performance of the reduced features in sparse areas, affecting the robustness and accuracy of the diagnosis.

[0006] 3. In the electric drive axle fault diagnosis task, traditional probabilistic neural networks have limited performance when faced with uncertain information, and the classification accuracy of complex fault modes is insufficient. The classifier has low processing capabilities for small amounts of data or abnormal data, and cannot meet the fault diagnosis needs under complex working conditions. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the present invention proposes a data-driven method for diagnosing faults in heavy-duty truck electric drive axles. The specific technical solution is as follows:

[0008] A data-driven method for diagnosing faults in a heavy truck electric drive axle comprises the following steps:

[0009] Step 1: Collect operating status data of heavy-duty truck electric drive axles in various usage scenarios, annotate fault categories, and construct an original dataset;

[0010] Step 2: Use a weighted noise-based generative adversarial network to generate samples and achieve data expansion; the generative adversarial network includes a generator and a discriminator, and noise is injected into the weights of the generator. The generator generates data based on the current weights and the injected noise; the discriminator evaluates the difference between the generated data and the real data and feeds it back to the generator. The generator dynamically adjusts the intensity of the noise injected into the generator weights based on this feedback, thereby training the weighted noise-based generative adversarial network to obtain expanded samples output by the trained model; and the expanded samples are merged with the original data to form a training data set;

[0011] Step 3: Construct and train a feature dimensionality reduction model; the feature dimensionality reduction model is a boundary-smoothing autoencoder, including an encoder and a decoder. The encoder compresses the input data to a low-dimensional representation. To address outliers and boundary effects in the dimensionality reduction process, the encoder output is smoothed using a smoothing function; the decoder reconstructs the low-dimensional representation into a feature vector that is as similar as possible to the input data; the data is processed by forward propagation, and then the autoencoder parameters are updated using gradient descent through error backpropagation;

[0012] Step 4: Construct a classifier model and train it using the dimensionality-reduced data output from step 3. The classifier model is a probabilistic neural network guided by quantum probability. The weights and activation function parameters in the probabilistic neural network are dynamically adjusted using quantum state estimation methods. The classification process is optimized by simulating quantum behavior, so that the network can better adapt to changes in data distribution in each step of training. The data is processed by forward propagation, and then the weights and quantum parameters of the probabilistic neural network are updated by error backpropagation.

[0013] Step 5: Collect the original data of the heavy-duty truck electric drive axle to be evaluated and input it into the trained feature dimensionality reduction model for feature processing; then input the processed features into the trained classifier model for classification to obtain the classification results.

[0014] Further, the fault categories include normal operation, bearing damage, motor failure, sensor failure, brake system problem, and tire abnormality.

[0015] Furthermore, the operating status data of the heavy-duty truck electric drive axle include bearing temperature Ra, wheel hub vibration amplitude Da, wheel axle torque Sa, brake temperature Ta, tire pressure Ca, speed sensor data Pa, GPS positioning data Qa, electric drive axle current Ka, electric drive axle voltage Ma and control unit feedback signal Na.

[0016] Furthermore, the training process of the weighted noise-based generative adversarial network in step 2 is as follows:

[0017] S201: Initialize generator G c Weight and the discriminator D c Weight The initialization method is expressed as:

[0018]

[0019] Where randn() represents a random number generated from a standard normal distribution; ← is a parameter update operation; d c represents the dimension of the input noise, m c Represents the dimension of the middle layer of the generator, n c Indicates the dimension of the output data;

[0020] S202: In each iteration, the weight of the generator is dynamically adjusted based on the feedback from the discriminator:

[0021]

[0022]

[0023] Where z c is the input noise of the generator, is the loss function of the generative adversarial network, η c is the learning rate of the generative adversarial network, represents the gradient of the generator weight; G c () is a generator function; D c () is the discriminator function; is the weight update amount of the generator; is a real sample; represents the partial derivative of the generator output with respect to its weight, calculated by the backpropagation algorithm;

[0024] S203: The generator generates data based on the current weights and injected noise:

[0025]

[0026] Where, is the sample generated by the generator; σ c is the noise intensity parameter, which is used to adjust the amplitude of noise injection so that the intensity of noise injection can be dynamically adjusted according to the performance of the discriminator, thereby more finely controlling the quality and diversity of generated data; represents the error rate of the discriminator on the current generated data, α c Is the first parameter that controls the decay rate; β c It is the second parameter that controls the decay speed;

[0027] S204: The discriminator evaluates the difference between the generated data and the real data and feeds it back to the generator; the parameter update increment calculation method of the discriminator is expressed as:

[0028]

[0029] Where, represents the gradient of the discriminator weight;

[0030] S205: Repeat iterative steps S202-S204 until the preset stop iteration condition is met and the model training is completed.

[0031] Furthermore, the training process of the feature dimensionality reduction model is as follows:

[0032] S301: Initialize the parameters of the boundary smoothing autoencoder:

[0033]

[0034] Where, represents the weight of the encoder, Indicates the encoder bias; represents the weight of the decoder, Indicates the bias of the decoder; randn() represents the random function; d r,in is the dimension of the encoder input layer, d r,mid is the dimension of the middle layer of the encoder, d r,out is the dimension of the encoder output layer; zeros() is a function that generates an all-zero vector;

[0035] S302: During the training of the autoencoder, data is processed by forward propagation, and then the autoencoder parameters are updated by gradient descent through error backpropagation.

[0036] Among them, the loss function L of the autoencoder r The calculation method is expressed as:

[0037]

[0038] Where x r is the input data of the autoencoder, is the output reconstructed by the decoder, Sig() is the Sigmoid activation function; ∥∥ is the L2 norm; is the output of the smoothed autoencoder; is the weight coefficient of the i-th dimension self-encoder; n r is the vector dimension corresponding to the samples of the current batch input to the autoencoder;

[0039] The update amount of the autoencoder parameters is calculated as follows:

[0040]

[0041] wherein, is the weight parameter update amount of the encoder; is the bias parameter update amount of the encoder; is the weight parameter update amount of the decoder; is the bias parameter update amount of the decoder; is the learning rate of the self-encoder at the tthiteration; is the learning rate of the initial self-encoder; is the learning rate of the self-encoder at the tthiteration; γ r is the learning rate decay factor of the self-encoder, is the average diameter of the convex hull at the tthiteration, is the diameter of the initial convex hull;

[0042] S303: Calculate the updated self-encoder parameters by the following formula:

[0043]

[0044] wherein, denotes the parameter update operation.

[0045] S304: Repeat S302-S303 until the preset stopping iteration condition is met, and the feature dimension reduction model training is completed.

[0046] Further, the training process of the classifier model includes the following sub-steps:

[0047] S401: Initialize the weight and quantum state parameters of the probabilistic neural network:

[0048]

[0049] wherein, denotes the initial weight matrix, d u is the input dimension of the probabilistic neural network, h u is the hidden layer dimension of the probabilistic neural network, denotes the initial parameters of the quantum state;

[0050] S402: Dynamically adjust the weight and activation function parameters in the probabilistic neural network using the quantum state estimation method; optimize the classification process by simulating quantum behavior, so that the network can better adapt to the changes in data distribution at each step of training;

[0051] wherein, the update increment of the quantum state parameters is denoted as:

[0052]

[0053] wherein η u is the learning rate of the probabilistic neural network, Y u is the true label, is the model prediction, denotes the partial derivative with respect to the quantum state parameters, L u () is the loss function of the probabilistic neural network; denotes the gradient of the loss function of the probabilistic neural network with respect to the quantum state parameters;, is the partial derivative of the loss function with respect to the predicted output, denotes the gradient of the predicted value with respect to the quantum state parameters; φ u (X uj ) is the activation function after weight normalization, ψ u (X uj , θ u ) is the quantum state function, i * is the imaginary unit;

[0054] S403: In each iteration, the features input to the probabilistic neural network are forward propagated through the quantum probability guided neural network, and the neural network calculates the output according to the current parameter state, in this process, the network processes the input data through the quantum probability method, and adjusts the activation state by using the quantum state estimation;

[0055] wherein the model prediction label is calculated in the following manner:

[0056]

[0057] wherein X u is the feature input to the probabilistic neural network, b u is the bias term of the probabilistic neural network, and f u () is the quantum probability guided activation function;

[0058] The calculation manner of the quantum probability guided activation function is represented as:

[0059]

[0060] wherein a u =W u ·X u +b u is the weighted input, and α u is a dynamic adjustment coefficient, which is calculated as:

[0061]

[0062] wherein γ u is a first hyperparameter for adjusting the intensity of the activation; and β uis the second hyperparameter used to adjust the activation strength; C u Indicates the number of categories;

[0063] S404: Perform forward propagation of data according to S403 to obtain the output of the network, that is, the model prediction label, and calculate the loss function value in combination with the real label, and use the backpropagation algorithm to update the network weights and quantum state parameters to minimize the loss function; the loss function L u The cross entropy loss is used for calculation as follows:

[0064]

[0065] Where Y uc is the true label of the c-th category; is the probability of the model predicted label for the cth category;

[0066] Furthermore, the gradient of the loss function with respect to the weights When used to update weights, the calculation method of the updated increment is expressed as:

[0067]

[0068] In the formula, ΔW u is the weight update increment of the probabilistic neural network; η u is the learning rate of the probabilistic neural network;

[0069] S405: After each iteration, the weights and quantum state parameters of the probabilistic neural network are updated according to the results of back propagation, and after each parameter update, the quantum state parameters θ u Perform normalization to ensure the stability of quantum probability distribution;

[0070] Among them, the update method of the weights and quantum state parameters of the probabilistic neural network is expressed as:

[0071] W u ←W u +ΔW u

[0072] θ u ←θ u +Δθ u

[0073]

[0074] Where, ∥θ u ∥ is the L2 norm of the quantum state parameter, θ u,j is the jth element of the quantum state parameter.

[0075] S406: Repeat S402-S405 until a preset stop iteration condition is met, and the model training is completed.

[0076] The beneficial effects of the present application are as follows:

[0077] 1. The present application introduces noise into the generator weight of the traditional generative adversarial network, dynamically adjusts the noise intensity to improve the authenticity and diversity of the generated data, solves the problem of poor model generalization ability caused by insufficient training samples, and significantly improves the effectiveness of data augmentation.

[0078] 2. The present application dynamically processes boundary anomalies through a boundary smoothing strategy, improves the adaptability to feature boundaries and noise in the dimension reduction process, and enhances the robustness of the model and the processing capability of the model for sparse region data.

[0079] 3. The present application is based on quantum state estimation and neural network, and optimizes the weight and activation function through quantum probability, enhances the adaptability of the network to uncertain data, and effectively improves the recognition accuracy and stability of the classifier for complex fault patterns. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 The flowchart of the data-driven heavy truck electric drive axle fault diagnosis method of the present application. DETAILED DESCRIPTION

[0081] The purpose and effect of the present application will become more apparent from the following detailed description of the drawings and preferred embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0082] As shown in Figure 1 The data-driven heavy truck electric drive axle fault diagnosis method of the present application includes the following steps:

[0083] Step 1: Collect the running state data of the heavy truck electric drive axle in various use scenarios, and label the categories to build the original data set.

[0084] In one embodiment, the labeled categories include normal operation, bearing damage, motor failure, sensor failure, brake system problem, and tire anomaly. The running state data includes bearing temperature Ra, hub vibration amplitude Da, axle torque Sa, brake temperature Ta, tire pressure Ca, speed sensor data Pa, GPS positioning data Qa, electric drive axle current Ka, electric drive axle voltage Ma, and control unit feedback signal Na.

[0085] In practical applications, the number of data attributes is usually more than 10, and the number of data attributes may reach dozens or even hundreds.

[0086] Step 2: Use a weighted noise-based generative adversarial network to generate samples and achieve data expansion; the generative adversarial network includes a generator and a discriminator. Noise is injected into the weights of the generator, and the generator generates data based on the current weights and the injected noise; the discriminator evaluates the difference between the generated data and the real data and feeds it back to the generator. The generator dynamically adjusts the intensity of the noise injected into the generator's weights based on this feedback, thereby training the weighted noise-based generative adversarial network and obtaining expanded samples output by the trained model; and the expanded samples are merged with the original data to form a training data set.

[0087] It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. In traditional generative adversarial networks, the generator and the discriminator compete with each other to improve the authenticity of the generated data. The present invention injects noise into the weights of the generator, which can not only increase the diversity of the data generated by the model, but also help the network avoid overfitting during the learning process and improve the generalization ability of the model.

[0088] Step 2 specifically includes the following sub-steps:

[0089] S201: Initialize generator G c Weight and the discriminator D c Weight The initialization method is expressed as:

[0090]

[0091] Where randn() represents a random number generated from a standard normal distribution; ← is a parameter update operation; d c represents the dimension of the input noise, m c Represents the dimension of the middle layer of the generator, n c Indicates the dimension of the output data.

[0092] S202: In each iteration, the noise injected into the generator weights is dynamically adjusted according to the feedback from the discriminator. Specifically, the generator weights are dynamically adjusted according to the feedback from the discriminator. The weight calculation method is expressed as:

[0093]

[0094] Where z c is the input noise of the generator, is the loss function of the generative adversarial network, ηc is the learning rate of the generative adversarial network, represents the gradient of the generator weight; G c () is the generator function; D c () is the discriminator function; is the weight update amount of the generator; is the real sample; represents the partial derivative of the generator output with respect to its weight, which is calculated by the back propagation algorithm.

[0095] S203: The generator generates data based on the current weight and the injected noise; the generated data should cover the distribution of the real data in the feature space as much as possible, while increasing a certain diversity to simulate different fault conditions; The specific generation method is represented as:

[0096]

[0097] In the formula, is the sample generated by the generator; σ c is the noise intensity parameter, used to adjust the amplitude of noise injection, so that the intensity of noise injection can be dynamically adjusted according to the performance of the discriminator, so as to more finely control the quality and diversity of the generated data; represents the error rate of the discriminator on the current generated data, α c is the first parameter for controlling the decay rate; β c is the second parameter for controlling the decay rate.

[0098] S204: The discriminator evaluates the difference between the generated data and the real data, and feeds back to the generator, and the generator adjusts its parameters according to this feedback to optimize the authenticity and diversity of the generated data, and the parameter update increment of the discriminator is calculated as:

[0099]

[0100] In the formula, represents the gradient of the discriminator weight.

[0101] S205: Repeat steps S202-S204 until the preset stopping iteration condition is met, and the model training is completed.

[0102] In one embodiment, the preset stopping iteration condition is to reach a preset maximum iteration number, preferably, the preset maximum iteration number is set to 1000 times.

[0103] Step 3: Build and train a feature dimensionality reduction model. The feature dimensionality reduction model is an autoencoder based on boundary smoothing, which includes an encoder and a decoder. The encoder compresses the input data to a low-dimensional representation. To address outliers and boundary effects in the dimensionality reduction process, the encoder output is smoothed using a smoothing function. The decoder reconstructs the low-dimensional representation into a feature vector that is as similar as possible to the input data. The data is processed through forward propagation, and then the autoencoder parameters are updated using gradient descent through error backpropagation.

[0104] The training of the feature dimensionality reduction model specifically includes the following steps:

[0105] S301: Initialize the parameters of the boundary smoothing-based autoencoder;

[0106] In one embodiment, the initialization is expressed as:

[0107]

[0108]

[0109] Where, represents the weight of the encoder, Indicates the encoder bias; represents the weight of the decoder, Indicates the bias of the decoder; randn() represents the random function; d r,in is the dimension of the encoder input layer, d r,mid is the dimension of the middle layer of the encoder, d r,out is the dimension of the encoder output layer; zeros() is a function that generates an all-zero vector.

[0110] S302: During the training of the autoencoder, data is processed by forward propagation, and then the autoencoder parameters are updated using the gradient descent method by error backpropagation.

[0111] During the training process of the autoencoder, the loss function of the autoencoder is calculated as follows:

[0112]

[0113] Where, L r is the loss function of the autoencoder; x r is the input data of the autoencoder, is the output reconstructed by the decoder, Sig() is the Sigmoid activation function; ∥∥ is the L2 norm; is the output of the smoothed autoencoder; is the weight coefficient of the i-th dimension self-encoder; n ris the vector dimension corresponding to the sample input to the autoencoder in the current batch.

[0114] In one embodiment, the weight coefficients of the autoencoder are used to adjust the contribution of each dimension feature in the error; these weights are dynamically allocated according to the importance of the features; the weight coefficients of the autoencoder are calculated as follows:

[0115]

[0116] wherein, is the feature variance of the input data of the ith dimension autoencoder; is the feature of the input data of the ith dimension autoencoder; ∈ re is a small constant to avoid division by zero; Var() is the variance function.

[0117] In order to handle outliers and boundary effects in the dimension reduction process, an automatic boundary smoothing strategy is adopted to enhance the robustness of the model by dynamically adjusting the boundary value processing. Specifically, the output of the encoder is smoothed by a smoothing function, so the output of the smoothed autoencoder is represented as:

[0118]

[0119] wherein, τ r () is the smoothing function, δ r is the smoothing intensity parameter; θ r is an automatically adjusted threshold parameter used to control the degree of smoothing processing and improve the adaptability to the boundary of the input data.

[0120] In one embodiment, the smoothing intensity parameter is determined by an adaptive method based on the local density of the data, and the calculation method is represented as:

[0121]

[0122] wherein, calculates the kernel density estimation in its neighborhood , which can accurately process data in high-density and sparse areas, and improve the robustness and adaptability of the overall model; denotes the neighborhood; k r () is the kernel density estimation function; mean() is the average function; ∈ rea is a small preset constant. Preferably, ∈ rea is set to 0.001.

[0123] ​​The training of the autoencoder is completed by optimizing the loss function in each iteration, and the calculation of the parameter update amount of the parameter update of the autoencoder is represented as:

[0124]

[0125] In the formula, is the weight parameter update amount of the encoder; is the bias parameter update amount of the encoder; is the weight parameter update amount of the decoder; is the bias parameter update amount of the decoder; is the learning rate of the autoencoder in the tth iteration.

[0126] In the training process of the autoencoder, a convex hull convergence strategy is used to optimize the training process of the autoencoder, and the network is quickly converged by adjusting the learning step and the weight update strategy; the learning rate is dynamically adjusted according to the data distribution characteristics in the current iteration, and the formula is as follows:

[0127]

[0128] In the formula, is the learning rate of the initial autoencoder; is the learning rate of the autoencoder in the tth iteration; γ r is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the tth iteration, is the diameter of the initial convex hull. The learning rate decreases with the decrease of the diameter of the convex hull to prevent instability caused by too large step size in the later learning.

[0129] S303: Calculate the updated autoencoder parameters by the following formula:

[0130]

[0131]

[0132] In the formula, <- represents the parameter update operation.

[0133] S304: Repeat S302-S303 until the preset stopping iteration condition is met, and the model training is completed. In an embodiment, the preset stopping iteration condition is to reach the preset maximum number of iterations, and preferably, the preset maximum number of iterations is set to 1000 times.

[0134] Step four: input the reduced data into the classifier model for training of the classifier model; a probability neural network based on quantum probability guidance is used as the classifier model.

[0135] The classifier model of the application is based on the traditional probabilistic neural network, adopts the performance of quantum probability enhanced network in processing uncertain information, optimizes the weight and activation function through quantum state estimation technology, and thereby improves the classification accuracy and stability.

[0136] Specifically, the training process of the quantum probability guided probabilistic neural network algorithm is as follows:

[0137] S401: initialize the weight and quantum state parameters of the probabilistic neural network, and the initialization mode is represented as:

[0138]

[0139] In the formula, represents the initial weight matrix, d u is the input dimension of the probabilistic neural network, h u is the hidden layer dimension of the probabilistic neural network, represents the initial parameters of the quantum state.

[0140] S402: dynamically adjust the weight and activation function parameters in the probabilistic neural network by using the quantum state estimation method, optimize the classification process by simulating quantum behavior, and make the network better adapt to the change of data distribution in each step of training.

[0141] The update increment of the quantum state parameter is represented as:

[0142]

[0143] In the formula, η u is the learning rate of the probabilistic neural network, Y u is the real label, is the model predicted label, represents the partial derivative of the quantum state parameter, L u () is the loss function of the probabilistic neural network.

[0144] The gradient of the loss function of the probabilistic neural network with respect to the quantum state parameter is calculated in the following manner:

[0145]

[0146] ψ u (X uj ,θ u )=cos(θ u X uj )+i * sin(θ u X uj )

[0147] In the formula, is the partial derivative of the loss function with respect to the predicted output, denotes the gradient of the predicted value with respect to the quantum state parameter; φ u (X uj ) is a weight-normalized activation function, ψ u (X uj , θ u ) is a quantum state function, i * is the imaginary unit.

[0148] S403: In each iteration, the features input to the probabilistic neural network are forward propagated through the quantum probability guided neural network, and the neural network calculates the output according to the current parameter state, in this process, the network processes the input data through the quantum probability method, and adjusts the activation state by using the quantum state estimation.

[0149] Model predicted label is calculated as:

[0150]

[0151] In the formula, X u is the feature input to the probabilistic neural network, b u is the bias term of the probabilistic neural network, and f u () is a quantum probability guided activation function.

[0152] In one embodiment, the quantum probability guided activation function is calculated as:

[0153]

[0154] In the formula, a u =W u ·X u +b u is the weighted input, α u is a dynamic adjustment coefficient, which is calculated as:

[0155]

[0156] In the formula, γ u is a first hyperparameter for adjusting the activation intensity; β u is a second hyperparameter for adjusting the activation intensity; C u denotes the number of categories.

[0157] S404, according to the output of the probabilistic neural network and the real label, the loss function value is calculated, and the weights and quantum state parameters of the network are updated by using the back propagation algorithm to minimize the loss function.

[0158] The calculation method of the loss function of the probabilistic neural network is represented as:

[0159]

[0160] where Y uc and are the true label Y u and the model predicted label

[0161] the gradient of the loss function with respect to the weights The way to calculate the updated increment for updating the weights is represented as:

[0162]

[0163] where ΔW u is the weight update increment of the probabilistic neural network; η u is the learning rate of the probabilistic neural network. Preferably, η u is set to 0.01.

[0164] S405: After each iteration, the weights and quantum parameters of the network are updated according to the results of backpropagation, and the way to update is represented as:

[0165] W u ← W u + ΔW u

[0166] θ u ← θ u + Δθ u

[0167] After each parameter update, the quantum state parameter θ u is normalized to ensure the stability of the quantum probability distribution, which is represented as:

[0168]

[0169] where || θ u || is the L2 norm of the quantum state parameter, and the way to calculate is represented as:

[0170]

[0171] where θ u,j is the jth element of the quantum state parameter.

[0172] S406: Repeat the iteration until the preset stopping iteration condition is met, and the model training is completed. In an embodiment, the preset stopping iteration condition is to reach the preset maximum number of iterations, and preferably, the preset maximum number of iterations is set to 1000.

[0173] Step 5: Collect the original data of the heavy-duty truck electric drive axle to be evaluated and input it into the trained feature dimensionality reduction model for feature processing; then input the processed features into the trained classifier model for classification to obtain the classification results.

[0174] Those skilled in the art will understand that the foregoing descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art will still be able to modify the technical solutions described in the foregoing examples or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the invention shall be included within the scope of protection of the invention.

Claims

1. A data-driven method for diagnosing faults in heavy truck electric drive axles, characterized in that: The steps include: Step 1: Collect operating status data of heavy-duty truck electric drive axles in various usage scenarios, annotate fault categories, and construct an original dataset; Step 2: Using a weighted noise-based generative adversarial network to generate samples and achieve data augmentation; the generative adversarial network includes a generator and a discriminator, and noise is injected into the weights of the generator. The generator generates data based on the current weights and the injected noise; The discriminator evaluates the difference between the generated data and the real data and provides feedback to the generator. The generator dynamically adjusts the intensity of the noise injected into the generator weights based on this feedback, thereby training the weight-noise-based generative adversarial network and obtaining expanded samples of the trained model output; The expanded samples and original data are merged to form a training data set; Step 3: Build and train the feature dimensionality reduction model; The feature dimensionality reduction model is an autoencoder based on boundary smoothing, which includes an encoder and a decoder. The encoder compresses the input data into a low-dimensional representation. To address outliers and boundary effects in the dimensionality reduction process, the encoder output is smoothed by a smoothing function. The decoder reconstructs the low-dimensional representation into a feature vector that is as similar as possible to the input data; the data is processed by forward propagation, and then the autoencoder parameters are updated using gradient descent by backpropagation of the error; Step 4: Construct a classifier model and train it using the dimensionality-reduced data output from step 3. The classifier model is a probabilistic neural network guided by quantum probability. The weights and activation function parameters in the probabilistic neural network are dynamically adjusted using quantum state estimation methods. The classification process is optimized by simulating quantum behavior, so that the network can better adapt to changes in data distribution in each step of training. The data is processed by forward propagation, and then the weights and quantum parameters of the probabilistic neural network are updated by error backpropagation. Step 5: Collect the original data of the heavy-duty truck electric drive axle to be evaluated and input it into the trained feature dimensionality reduction model for feature processing; then input the processed features into the trained classifier model for classification to obtain the classification results.

2. The data-driven heavy truck electric drive axle fault diagnosis method according to claim 1 is characterized in that: Fault categories include normal operation, bearing damage, motor failure, sensor failure, brake system problems, and tire abnormalities.

3. The data-driven heavy truck electric drive axle fault diagnosis method according to claim 1 is characterized in that: The operating status data of the heavy-duty truck electric drive axle include bearing temperature Ra, wheel hub vibration amplitude Da, wheel axle torque Sa, brake temperature Ta, tire pressure Ca, speed sensor data Pa, GPS positioning data Qa, electric drive axle current Ka, electric drive axle voltage Ma and control unit feedback signal Na.

4. The data-driven heavy truck electric drive axle fault diagnosis method according to claim 1 is characterized in that: The training process of the weighted noise-based generative adversarial network in step 2 is as follows: S201: Initialize generator G c Weight and the discriminator D c Weight The initialization method is expressed as: Where randn() represents a random number generated from a standard normal distribution; ← is a parameter update operation; d c represents the dimension of the input noise, m c Represents the dimension of the middle layer of the generator, n c Indicates the dimension of the output data; S202: In each iteration, the weight of the generator is dynamically adjusted based on the feedback from the discriminator: Where z c is the input noise of the generator, is the loss function of the generative adversarial network, η c is the learning rate of the generative adversarial network, represents the gradient of the generator weight; G c () is a generator function; D c () is the discriminator function; is the weight update amount of the generator; is a real sample; represents the partial derivative of the generator output with respect to its weight, calculated by the backpropagation algorithm; S203: The generator generates data based on the current weights and injected noise: Where, is the sample generated by the generator; σ c is the noise intensity parameter, which is used to adjust the amplitude of noise injection so that the intensity of noise injection can be dynamically adjusted according to the performance of the discriminator, thereby more finely controlling the quality and diversity of generated data; represents the error rate of the discriminator on the current generated data, α c It is the first parameter that controls the decay rate; β c It is the second parameter that controls the decay speed; S204: The discriminator evaluates the difference between the generated data and the real data and feeds it back to the generator; the parameter update increment calculation method of the discriminator is expressed as: Where, represents the gradient of the discriminator weight; S205: Repeat iterative steps S202-S204 until the preset stop iteration condition is met and the model training is completed.

5. The data-driven heavy truck electric drive axle fault diagnosis method according to claim 4 is characterized in that: The training process of the feature dimensionality reduction model is as follows: S301: Initialize the parameters of the boundary smoothing autoencoder: Where, represents the weight of the encoder, Indicates the encoder bias; represents the weight of the decoder, Indicates the bias of the decoder; randn() represents the random function; d r,in is the dimension of the encoder input layer, d r,mid is the dimension of the middle layer of the encoder, d r,out is the dimension of the encoder output layer; zeros() is a function that generates an all-zero vector; S302: During the training of the autoencoder, data is processed by forward propagation, and then the autoencoder parameters are updated by gradient descent through error backpropagation. Among them, the loss function L of the autoencoder r The calculation method is expressed as: Where x r is the input data of the autoencoder, is the output reconstructed by the decoder, Sig() is the Sigmoid activation function; ∥∥ is the L2 norm; is the output of the smoothed autoencoder; is the weight coefficient of the i-th dimension self-encoder; n r is the vector dimension corresponding to the samples of the current batch input to the autoencoder; The update amount of the autoencoder parameters is calculated as follows: Where, is the update amount of the encoder's weight parameters; is the encoder bias parameter update amount; is the update amount of the decoder weight parameter; is the update amount of the decoder's bias parameter; is the learning rate of the autoencoder at the tth iteration; is the learning rate of the initial autoencoder; is the learning rate of the autoencoder at the tth iteration; γ r is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the t-th iteration, is the diameter of the initial convex hull; S303: Calculate the updated autoencoder parameters using the following formula: Where ← represents the parameter update operation; S304: Repeat S302 to S303 until the preset stop iteration condition is met and the feature dimensionality reduction model training is completed.

6. The data-driven heavy truck electric drive axle fault diagnosis method according to claim 4 is characterized in that: The training process of the classifier model includes the following sub-steps: S401: Initialize the weights and quantum state parameters of the probabilistic neural network: Where, represents the initial weight matrix, d u is the input dimension of the probabilistic neural network, h u is the hidden layer dimension of the probabilistic neural network, represents the initial parameters of the quantum state; S402: Dynamically adjust the weights and activation function parameters in the probabilistic neural network using quantum state estimation methods; optimize the classification process by simulating quantum behavior, so that the network can better adapt to changes in data distribution at each step of training; Among them, the update increment of the quantum state parameter is expressed as: Where η u is the learning rate of the probabilistic neural network, Y u is the true label, is the model prediction, represents the partial derivative of the quantum state parameter, L u () is the loss function of the probabilistic neural network; Represents the gradient of the loss function of the probabilistic neural network with respect to the quantum state parameters; is the partial derivative of the loss function with respect to the predicted output, represents the gradient of the predicted value with respect to the quantum state parameters; φ u (X uj ) is the activation function after weight normalization, ψ u (X uj ,θ u ) is the quantum state function, i * is an imaginary unit; S403: In each iteration, the features input to the probabilistic neural network are forward propagated through the quantum probability-guided neural network. The neural network calculates the output based on the current parameter state. During this process, the network processes the input data through quantum probabilistic methods and adjusts the activation state using quantum state estimation. Among them, the model predicts the label The calculation method is expressed as: Where, X u is the feature input to the probabilistic neural network, b u is the bias term of the probabilistic neural network, f u () is the quantum probability guided activation function; The calculation method of the quantum probability guided activation function is expressed as: Where a u =W u ·X u +b u is the weighted input, α u is the dynamic adjustment coefficient, which is calculated as follows: Where, γ u is the first hyperparameter used to adjust the activation strength; β u is the second hyperparameter used to adjust the activation strength; C u Indicates the number of categories; S404: Perform forward propagation of data according to S403 to obtain the output of the network, that is, the model prediction label, and calculate the loss function value in combination with the real label, and use the backpropagation algorithm to update the network weights and quantum state parameters to minimize the loss function; the loss function L u The cross entropy loss is calculated as follows: Where Y uc is the true label of the c-th category; is the probability of the model predicted label for the cth category; The gradient of the loss function with respect to the weights When used to update weights, the calculation method of the updated increment is expressed as: In the formula, ΔW u is the weight update increment of the probabilistic neural network; η u is the learning rate of the probabilistic neural network; S405: After each iteration, the weights and quantum state parameters of the probabilistic neural network are updated according to the results of back propagation, and after each parameter update, the quantum state parameters θ u Perform normalization to ensure the stability of quantum probability distribution; Among them, the update method of the weights and quantum state parameters of the probabilistic neural network is expressed as: IN u ←In u +ΔW u i u ←θ u +Δθ u Where, ∥θ u ∥ is the L2 norm of the quantum state parameter, θ u,j is the jth element of the quantum state parameter; S406: Repeat S402 to S405 until the preset stop iteration condition is met and the model training is completed.

Images