Heavy truck electric drive bridge fault diagnosis method based on generative adversarial network

Abstract

The invention discloses a heavy truck electric drive bridge fault diagnosis method based on a generative adversarial network, and the method achieves the precise analysis and diagnosis of the operation state of a heavy truck electric drive bridge through data collection and marking, data expansion, feature extraction model training, classifier model training and fault diagnosis. In the data expansion stage, a generative adversarial network based on adaptive stochastic disturbance is adopted, an entropy weight dynamic back propagation mechanism is introduced, the authenticity of generated data is improved by dynamically adjusting a learning rate, and distribution of the generated data is optimized by using a Gaussian mixture model and a feature dependency matrix; the feature extraction model is based on a full-connection neural network optimized by a bionic algorithm, a training rule is adjusted by adopting a self-correction mechanism, and the training efficiency and the model generalization ability are improved; the classifier adopts a high-order neural network of a reverse backtracking sparse strategy, the network structure and the feature weight are dynamically adjusted, and the sensitivity to abnormal data and the accuracy of fault classification are effectively enhanced. The method is high in fault diagnosis efficiency, high in precision and high in stability.

Description

Technical Field

[0001] The present invention relates to the field of heavy truck electric drive axle fault diagnosis, and in particular to a heavy truck electric drive axle fault diagnosis method based on a generative adversarial network. Background Art

[0002] With the rapid development of electric drive technology for new energy vehicles and heavy trucks, the importance of the electric drive axle system has become increasingly prominent, and its operating status directly affects the performance, energy efficiency and service life of the vehicle. However, the complex mechanical structure and electrical control system of the electric drive axle make it susceptible to a variety of factors, such as vibration, temperature fluctuations, current anomalies, etc., which can lead to performance degradation or failure. This complexity requires accurate and efficient fault diagnosis of the electric drive axle during operation to ensure the safety and reliability of the vehicle.

[0003] The existing heavy truck electric drive axle fault diagnosis method has the following shortcomings:

[0004] 1. In the intelligent diagnosis task of electric drive axle faults, the existing fault diagnosis data expansion methods are mostly based on simple data transformation or interpolation, and fail to specifically enhance the characteristics of heavy-duty truck electric drive axle fault data, resulting in a gap in statistical characteristics between the generated data and the real data, affecting the diagnostic effect of the expanded data.

[0005] 2. In the intelligent diagnosis task of electric drive axle fault, the correlation between multidimensional features was not fully considered during the data expansion process. The generated data samples deviated from the actual working conditions in terms of the correlation between features, which limited the reliability of the expanded data in the diagnosis task.

[0006] 3. In the intelligent diagnosis task of electric drive axle faults, traditional feature extraction methods are easily affected by problems such as gradient vanishing and local optimality when processing complex multi-dimensional data, which makes it difficult for the model to accurately extract fault features.

[0007] 4. In the intelligent diagnosis task of electric drive axle faults, most fault classification models use a fixed neural network structure, which makes it difficult to dynamically adjust the network structure and sparsity according to different data characteristics. The computational efficiency is low and cannot adapt to real-time diagnosis needs.

[0008] 5. In the intelligent diagnosis task of electric drive axle faults, the existing methods lack a dynamic evaluation mechanism for the diagnostic relevance of input features, resulting in poor robustness and adaptability of the model when processing new samples, making it difficult to accurately identify new types of faults. Summary of the invention

[0009] In view of the shortcomings of the prior art, the present invention proposes a heavy truck electric drive axle fault diagnosis method based on a generative adversarial network. The specific technical solution is as follows:

[0010] A heavy truck electric drive axle fault diagnosis method based on a generative adversarial network comprises the following steps:

[0011] Step 1: Collect the operating data of the heavy-duty truck electric drive axle under different working conditions and mark the fault category for each data;

[0012] Step 2: Construct a generative adversarial network based on adaptive random perturbation, and input the data obtained in step 1 into the generative adversarial network based on adaptive random perturbation to generate samples and achieve data expansion; merge the expanded data with the original data in step 1 to obtain a training data set;

[0013] The adaptive random perturbation-based generative adversarial network introduces an entropy-weighted dynamic back-propagation mechanism on the basis of the generative adversarial network, that is, the learning rate in the back-propagation process is adjusted according to the information entropy of the generated data, so as to optimize the learning efficiency of the generator and the quality of the generated data;

[0014] Step 3: construct a feature extraction model, and train the feature extraction model using the data of the training data set; the feature extraction model is a 6-layer fully connected neural network, and a bionic algorithm is used to optimize the training process of the feature extraction model during the training process;

[0015] Step 4: Construct a classifier model and use the feature-extracted data to train the classifier model; the classifier model is a high-order neural network based on reverse backtracking sparseness;

[0016] Step 5: Collect new data of the heavy-duty truck electric drive axle to be evaluated, input it into the trained feature extraction model for feature processing, and input the processed features into the trained classifier model for classification to obtain the classification results.

[0017] Furthermore, the training process of the adaptive random perturbation-based generative adversarial network is as follows:

[0018] S201: Initialize generator G c and the discriminator D c The generator receives a random noise vector as input and outputs data with the same structure as the real data; the discriminator determines the authenticity of the input data:

[0019]

[0020] in, represents the data generated by the generator, z c represents the input random noise vector, Represents the parameters of the generator; G c () is the generator function; y c represents the output of the discriminator, x cRepresents the input data, represents the parameters of the discriminator;

[0021] S202: In each iteration, the entropy weighted dynamic back propagation algorithm is used to calculate the entropy of the current generated data, and the learning rate η of the generative adversarial network is adjusted according to the entropy value c :

[0022]

[0023] In the formula, H c () represents the information entropy of generated data; Nce is the characteristic dimension of generated data; is the i-th feature of the generated data; p() is the probability distribution function, represents the probability distribution of the i-th eigenvalue of the generated data, estimated by the probability density function of the generated data;

[0024] Model the distribution of generated data and use Gaussian mixture model fitting to estimate the probability distribution of each eigenvalue:

[0025]

[0026] Mce is the number of Gaussian mixture components; is the weight of the jth mixture component, G s () is Gaussian distribution, and are the mean and variance of the jth mixture component respectively;

[0027] S203: The generator generates new data samples, which are evaluated by the discriminator. The output of the discriminator is used to calculate the loss, which is fed back to the generator and the discriminator so that they can update their own network parameters through the gradient descent method.

[0028] Among them, the iterative training process of the generator and the discriminator is expressed as:

[0029]

[0030] Where ← is the parameter update operation; and are the parameters of the generator and the discriminator respectively; represents the gradient with respect to the generator parameters; represents the gradient of the discriminator parameters; D c () is the discriminator function;

[0031] The loss function is specifically:

[0032]

[0033] in, represents the total loss function of the generative adversarial network, is the adversarial loss function of the generative adversarial network, is the adaptive weight loss function, α c is the weight factor of the first loss function; β c is the weight factor of the second loss function; represents expectation; ~ represents compliance with a specific distribution; p data is the distribution of real data; p z is the distribution of noise; is the weight of the data feature, which is used to adjust the contribution of different features in the loss; is the variance of the feature, which is used to measure the variability of the feature in the data set; ∈ cks is a small positive number to prevent the denominator from being zero; a c,i,j Represents the dependency between the i-th feature and the j-th feature, cov() represents the covariance, and are the standard deviations of the i-th feature and the j-th feature respectively; is the i-th feature of the generated data; is the jth feature of the generated data; cew It is the dynamic adjustment item adjustment coefficient, which is used to control the impact of the adjustment item;

[0034] S204: Repeat steps S202 to S203 until a preset stop iteration condition is met, and the training of the generative adversarial network based on adaptive random perturbation is completed.

[0035] Furthermore, the training process of the feature extraction model includes:

[0036] S301: According to the bionic algorithm initialization method, in the initialization stage, an initial population is generated, and each individual represents a configuration of network weights; let the population size be N p , the weights and biases for the i-th individual are initialized as:

[0037]

[0038] In the formula, Represents the weight matrix of the i-th individual in the initial state; represents the bias matrix of the i-th individual in the initial state; σ 2 represents the initialization variance; Indicates that the mean is 0 and the variance is σ 2 Normal distribution of is a normal distribution;

[0039] S302: For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function;

[0040] For the i-th individual, the loss on the training dataset is calculated using its weight and bias, expressed as:

[0041]

[0042] Where, L pi is the loss of the neural network corresponding to the i-th individual, and is defined as the fitness of the individual; mps represents the number of samples input in the current batch; l() represents the composite loss function, and fsig() represents the neural network model function; Represents the characteristics of the jth sample; represents the label of the jth sample; MSE() is the mean square error function, λ ps is the regularization parameter; Reg(W pi ) is the regularization term; W pi,k is the kth weight of the neural network corresponding to the i-th individual;

[0043] S303: According to the fitness of the individuals and based on the selection strategy, the best performing individuals are selected from the current population to be retained to form an excellent population as candidate solutions for the next generation;

[0044] The selection strategy is defined as:

[0045]

[0046] Where P select (i) represents the probability of the i-th individual being selected; γ pse is the parameter that controls the selection pressure; L pk is the loss of the neural network corresponding to the kth individual;

[0047] S304: Generate new individuals through crossover and mutation operations; the crossover operation is specifically to randomly select two individuals from the excellent population to exchange genes, which is expressed as:

[0048] W′ pi =α pcs W p1 +(1-α pcs )W p2

[0049] b′ pi =α pcs b p1 +(1-α pcs )b p2

[0050] In the formula, α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first individual selected, b p1 is the bias of the neural network corresponding to the first individual selected, W p2 is the weight of the neural network corresponding to the second individual selected, b p2 is the bias of the neural network corresponding to the second individual selected, W′ pi is the weight of the neural network corresponding to the individual after the crossover operation, b′ pi is the bias of the neural network corresponding to the individual after the crossover operation;

[0051] The mutation operation is to perform a small random perturbation on the individual weights in the excellent population, which can be expressed as:

[0052]

[0053] In the formula, τ 2 represents the variance of variation, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation;

[0054] S305: Repeat S302 to S304 until the preset stop iteration condition is met and the model training is completed.

[0055] Furthermore, in S303, the probabilities of individuals being selected are sorted from high to low, and the top 30% of individuals with high probabilities are selected as excellent populations.

[0056] Furthermore, the training process of the high-order neural network based on reverse backtracking sparse is as follows:

[0057] S401: Initialize the weights and biases of the high-order neural network. The structure of the high-order neural network is designed to be a 4-layer feedforward structure. The parameter initialization process of the high-order neural network is expressed as:

[0058]

[0059] In the formula, It means the mean is 0 and the variance is The normal distribution of and They represent the weight and bias of the kth layer of the high-order neural network respectively;

[0060] S402: After each forward propagation, the weight is adjusted according to the error between the output of each layer of the high-order neural network and the actual output. The error feedback affects the weight update in the form of high-order derivatives. The calculation method of the weight update amount is expressed as:

[0061]

[0062] In the formula, is the weight update amount of the kth layer of the high-order neural network; η ue is the learning rate of the high-order neural network, α u is the coefficient that adjusts the influence of higher-order derivatives, is the loss function of the high-order neural network; is the true category of the sample; u It is the classification category of high-order neural network;

[0063] S403: During the back propagation process, the connection sparsity of the high-order neural network is calculated according to the predefined rules and the performance indicators of the current network. If the sparsity index of the k-th layer of the high-order neural network is not less than the sparsity threshold θ u , then keep the connection of the kth layer of the high-order neural network, otherwise disconnect;

[0064] Sparsity index of the kth layer of high-order neural network It is expressed as:

[0065]

[0066] In the formula, β u is a binary vector that determines whether to maintain or disconnect the connection; θ u is the sparsity threshold, mce is the number of neurons in this layer, and ⊙ represents element-wise multiplication; is the weight of the jth neuron in the kth layer of the high-order neural network;

[0067] The sparsity threshold θ u It is expressed as:

[0068]

[0069] In the formula, λ u is the adjustment coefficient; std() is the function for calculating the standard deviation to ensure that the model maintains an appropriate balance between complexity and overfitting during training; is the mean value of the layer weights of the high-order neural network;

[0070] S404: Calculate the feature influence of each feature, calculate the weight update variable of the high-order neural network based on the feature influence, and further calculate the increment of the weight update variable:

[0071]

[0072] in, is the feature influence; It is the input feature of the kth layer of the high-order neural network; Update variables for the weights of high-order neural networks; γ u is the adjustment coefficient during the learning process; Increment the variable for weight update; is the loss function of the high-order neural network;

[0073] S405: Repeat forward propagation and back propagation. In each iteration, the weights and parameters of the high-order neural network are updated as follows:

[0074]

[0075] Where ← is the parameter update operation; η us is the learning rate of the high-order neural network; is the loss function of the high-order neural network;

[0076] S406: Repeat iterations S402 to S405 until a preset condition for stopping iteration is met, which means that the high-order neural network training based on reverse backtracking sparseness is completed.

[0077] Furthermore, the fault categories include normal operation, abnormal temperature, abnormal vibration, abnormal torque transmission, and unstable current.

[0078] Furthermore, the operating data of the heavy-duty truck electric drive axle under different working conditions includes the rotational speed of the electric drive axle, the torque value input to the electric drive axle, the temperature data of the key components of the electric drive axle, the vibration characteristics of the electric drive axle under different working conditions, the noise generated when the electric drive axle is running, the current change of the electric drive axle, the diagnostic code when a fault occurs, the location information of the electric drive axle in the vehicle, maintenance records and operating modes.

[0079] The beneficial effects of the present invention are as follows:

[0080] 1. The present invention adopts an improved generation mechanism of a generative adversarial network and a generative adversarial network based on adaptive random perturbation to expand data. The generated data is closer to the real data distribution, which expands the diversity of training samples, solves the problem of small sample training, and effectively enhances the accuracy of heavy-duty truck electric drive axle fault diagnosis.

[0081] 2. The present invention adopts a bionic algorithm to optimize the training process of the feature extraction model, improves the convergence speed and training stability of the model, dynamically adjusts the learning rate and feature dependency, and adapts to complex fault diagnosis tasks.

[0082] 3. The present invention adopts a dynamic sparse strategy to optimize the classifier model structure, which improves the prediction accuracy of the model while reducing the computational cost. The dynamic evaluation mechanism of feature importance significantly improves the sensitivity to abnormal data and the diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 The present invention is a flow chart of the intelligent diagnosis method for heavy truck electric drive axle fault based on generative adversarial network. DETAILED DESCRIPTION

[0084] The present invention will be described in detail below based on the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become more clear. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0085] like Figure 1 As shown, the heavy truck electric drive axle fault diagnosis method based on the generative adversarial network of the present invention comprises the following steps:

[0086] Step 1: Collect the operating data of the heavy-duty truck electric drive axle under different working conditions, and mark the fault category for each data.

[0087] The data acquisition of the present invention is derived from the operation monitoring system of the heavy truck electric drive axle. The real-time sensor network is used to capture the operation data of the electric drive axle under different working conditions. The acquisition method includes but is not limited to vibration sensing, temperature sensing and torque monitoring, and is stored in a database in a structured manner. In one embodiment, the attributes of the data include:

[0088] Speed ​​Ra, records the speed of the electric drive axle and reflects the power output;

[0089] Torque Ta, measures the torque value input to the electric drive axle;

[0090] Temperature Pa, records the temperature data of key components of the electric drive axle;

[0091] Vibration Va, capturing the vibration characteristics of the electric drive axle under different working conditions;

[0092] Sound Sa, analyzes the noise generated when the electric drive axle is running;

[0093] Current Da, monitors the current change of the electric drive bridge;

[0094] Fault code Fa, records the diagnostic code when the fault occurs;

[0095] Position Ga, indicating the position information of the electric drive axle in the vehicle;

[0096] Maintenance records Ma, including records of historical maintenance and parts replacement;

[0097] Operation mode Ka reflects the operating status of the electric drive axle.

[0098] It should be noted that this embodiment is only used to illustrate a data format and type of the present invention. In actual applications, the attributes of the data are usually more than 10 attributes, and the number of attributes of the data may reach dozens or even hundreds. The fault categories marked in this embodiment include normal operation, abnormal temperature, abnormal vibration, abnormal torque transmission, and unstable current.

[0099] Step 2: Construct a generative adversarial network based on adaptive random perturbation, and input the data obtained in step 1 into the generative adversarial network based on adaptive random perturbation for sample generation to achieve data expansion; merge the expanded data with the original data in step 1 to obtain a training data set.

[0100] The generative adversarial network consists of two parts: a generator and a discriminator. The generator is responsible for generating new data that is as close to the real data as possible, and the discriminator's task is to distinguish the generated data from the real data. The generative adversarial network based on adaptive random perturbation introduces an entropy-weighted dynamic back-propagation mechanism on the basis of the traditional generative adversarial network. This mechanism adjusts the learning rate in the back-propagation process according to the information entropy of the generated data, optimizing the learning efficiency of the generator and the quality of the generated data. The adaptive random perturbation is based on the adaptive weight adjustment of the data characteristics, so that the model can pay more attention to the key data features in the fault diagnosis of the heavy-duty truck electric drive axle during the training process, thereby improving the pertinence and effect of data expansion.

[0101] Specifically, the training process of the generative adversarial network based on adaptive random perturbation is as follows:

[0102] S201: Initialize generator G c and the discriminator D c The generator receives a random noise vector as input and outputs data with the same structure as the real data. The task of the discriminator is to determine the authenticity of the input data (whether real or generated).

[0103] Generator G c The principle can be simplified as follows:

[0104]

[0105] In the formula, represents the data generated by the generator, z c represents the input random noise vector, Represents the parameters of the generator; G c () is a generator function.

[0106] The principle of the discriminator can be simplified as follows:

[0107]

[0108] In the formula, y c represents the output of the discriminator (true or false judgment), x c Represents the input data (which can be real data or generated data), Represents the parameters of the discriminator.

[0109] S202: In each iteration, the learning rates of the generator and the discriminator are dynamically adjusted according to the performance of the discriminator in the last few batches.

[0110] Specifically, the entropy weighted dynamic back propagation algorithm is used to calculate the entropy of the currently generated data, and the learning rate is adjusted according to the entropy value to improve the learning efficiency and stability. The calculation method is expressed as:

[0111]

[0112] Where η c represents the learning rate of the generative adversarial network, H c () represents the information entropy of generated data; Nce is the characteristic dimension of generated data; is the i-th feature of the generated data; p() is the probability distribution function, represents the probability distribution of the i-th eigenvalue of the generated data, estimated by the probability density function of the generated data.

[0113] Model the distribution of generated data and use the Gaussian mixture model to fit and estimate the probability distribution of each eigenvalue. The calculation method is expressed as:

[0114]

[0115] Where, Mce is the number of Gaussian mixture components; is the weight of the jth mixture component, G s () is Gaussian distribution, and are the mean and variance of the jth mixture component, respectively.

[0116] S203: The generator generates new data samples, which are evaluated by the discriminator. The output of the discriminator is used to calculate the loss, which is fed back to the generator and the discriminator so that they can update their own network parameters through the gradient descent method. The iterative training process of the generator and the discriminator is expressed as:

[0117]

[0118] Where ← is the parameter update operation; and are the parameters of the generator and the discriminator respectively; represents the gradient with respect to the generator parameters; represents the gradient of the discriminator parameters; D c () is the discriminator function.

[0119] In traditional generative adversarial networks, the generator and discriminator are optimized with fixed strategies, ignoring the interdependence between different features. The present invention utilizes a dynamic feature dependency analysis and adjustment mechanism to evaluate and adjust the dependencies between features in the generated data in real time, so as to more accurately simulate the statistical characteristics of real data.

[0120] The calculation process of the loss function is as follows:

[0121] (1) Define the feature dependency matrix of generated data as:

[0122] A c ={a c,i,j}

[0123]

[0124] In the formula, A c To generate the feature dependency matrix of the data; a c,i,j Represents the dependency between the i-th feature and the j-th feature, cov() represents the covariance, and are the standard deviations of the i-th feature and the j-th feature respectively; is the i-th feature of the generated data; is the jth feature of the generated data.

[0125] (2) Based on the dependency matrix, define the dynamic adjustment term ΔL c Used to adjust the loss function to better guide the generator to learn the interdependence between features. The calculation method is expressed as:

[0126]

[0127] In the formula, λ cew It is the dynamic adjustment item adjustment coefficient, which is used to control the impact of the adjustment item.

[0128] (3) Calculate the loss function of the generative adversarial network. The loss function consists of two parts. One part is the traditional generative adversarial network loss, which is used to evaluate the similarity between the generated data and the real data. The other part is the adaptive weight loss, which automatically adjusts the weight according to the diagnostic relevance of different data attributes. The calculation method is expressed as:

[0129]

[0130] In the formula, represents the total loss function of the generative adversarial network, is the adversarial loss function of the generative adversarial network, is the adaptive weight loss function, α c is the weight factor of the first loss function; β c is the weight factor of the second loss function; represents expectation; ~ represents compliance with a specific distribution; p data is the distribution of real data; p z is the distribution of noise; is the weight of the data feature, which is used to adjust the contribution of different features in the loss. The calculation method is expressed as:

[0131]

[0132] in, is the variance of the feature, which is used to measure the variability of the feature in the data set; ∈ cks is a small positive number that prevents the denominator from being zero. Preferably, ∈ cks Set to 0.001.

[0133] S204: Repeat steps S202 to S203 until the preset stop iteration condition is met and the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0134] After the training of the generative adversarial network based on adaptive random perturbation is completed, the trained data expansion model is used to increase the number of samples. In one embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0135] Step 3: Build a feature extraction model and use the expanded data to train the feature extraction model. The feature extraction model is a 6-layer fully connected neural network, and the bionic algorithm is used to optimize the training process of the feature extraction model.

[0136] In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, they may encounter problems such as gradient vanishing, gradient exploding, or falling into local optimal solutions, which will affect the stability of training and the performance of the model. In the traditional bionic population evolution algorithm, the evolution of all individuals is based on fixed rules. The present invention uses a self-correction mechanism to automatically adjust the evolution rules according to the characteristics of the current training data, and adjusts the probability of crossover and mutation according to the changing trend of the loss function, so that the algorithm can adapt to different data distributions more flexibly and improve the generalization ability and training efficiency of the model.

[0137] The training process of the feature extraction model includes:

[0138] S301: According to the bionic algorithm initialization method, in the initialization stage, an initial population is generated, each individual represents a configuration of network weights. Specifically, the population size is N p , initialize the weights and biases for the i-th individual, expressed as:

[0139]

[0140] Where W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual; Represents the weight matrix of the i-th individual in the initial state; represents the bias matrix of the i-th individual in the initial state; σ 2 represents the initialization variance; Indicates that the mean is 0 and the variance is σ 1 Normal distribution of is a normal distribution. Preferably, σ 2 Set to 0.01.

[0141] S302: For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function.

[0142] Specifically, for the i-th individual, its weight and bias are used to calculate the loss on the training data set, expressed as:

[0143]

[0144] Where, L pi is the loss of the neural network corresponding to the i-th individual, and is defined as the fitness of the individual; mps represents the number of samples input in the current batch; l() represents the composite loss function, and fsig() represents the neural network model function; Represents the characteristics of the jth sample; represents the label of the jth sample.

[0145] The composite loss function contains a regularization term, which can increase the generalization ability of the model. The calculation method is expressed as:

[0146]

[0147] Where MSE() is the mean square error function, λ ps is the regularization parameter; Reg(W pi ) is the regularization term, and its calculation method is expressed as:

[0148]

[0149] Where W pi,k is the kth weight of the neural network corresponding to the i-th individual.

[0150] S303: According to the fitness of the individual, based on the selection strategy, select the best performing individuals from the current population to be retained to form an excellent population as the candidate solution for the next generation. Specifically, define a selection strategy function, which is a function of the individual fitness, and calculate the probability of the individual being selected. The selection strategy function is defined as:

[0151]

[0152] Where P select (i) represents the probability of the i-th individual being selected; γ pse is the parameter that controls the selection pressure; L pk is the loss of the neural network corresponding to the kth individual. Preferably, γ pse Set to 2.

[0153] Furthermore, based on the probability of individuals being selected, the 30% of individuals with the highest probability values ​​are selected as the excellent population.

[0154] S304: Generate new individuals through crossover and mutation operations.

[0155] The crossover operation allows two excellent individuals to exchange some genes to generate new offspring; the mutation operation randomly changes some genes in an individual to increase the diversity of the population. Specifically, the crossover operation randomly selects two individuals from the excellent population for gene exchange, which is expressed as:

[0156] W′ pi =α pcs W p1 +(1-α pcs )W p2

[0157] b′ pi =α pcs b p1 +(1-α pcs )b p2

[0158] In the formula, α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first individual selected, b p1 is the bias of the neural network corresponding to the first individual selected, W p2 is the weight of the neural network corresponding to the second individual selected, b p2is the bias of the neural network corresponding to the second individual selected, W′ pi is the weight of the neural network corresponding to the individual after the crossover operation, b′ pi is the bias of the neural network corresponding to the individual after the crossover operation. Preferably, α pcs Set to 0.3.

[0159] The mutation operation performs a small random perturbation on the individual weights in the excellent population, which can be expressed as:

[0160]

[0161] In the formula, τ 2 represents the variance of variation, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation. Preferably, τ 2 Set to 0.04.

[0162] S305: Repeat S302 to S304 until the preset stop iteration condition is met and the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0163] Step 4: Construct a classifier model and use the feature-extracted data to train the classifier model; the classifier model is a high-order neural network based on reverse backtracking sparseness.

[0164] The reverse backtracking sparsity strategy dynamically adjusts the connection sparsity in the neural network during the training process of the high-order neural network, optimizes the structure and computational efficiency of the model, which not only increases the generalization ability of the model, but also reduces the risk of overfitting, making the model more stable on new samples.

[0165] Specifically, the training process of the high-order neural network based on reverse backtracking sparse is as follows:

[0166] S401: Initialize the weights and biases of the high-order neural network. The structure of the high-order neural network is designed to be a 4-layer feedforward structure. The parameter initialization process of the high-order neural network is expressed as:

[0167]

[0168] In the formula, It means the mean is 0 and the variance is The normal distribution of and denote the weight and bias of the kth layer of the high-order neural network, respectively. Preferably, Set to 0.01.

[0169] S402: After each forward propagation, the weight is adjusted according to the error between the output of each layer of the high-order neural network and the actual output. The error feedback affects the weight update in the form of high-order derivatives. The calculation method of the weight update amount is expressed as:

[0170]

[0171] In the formula, is the weight update amount of the kth layer of the high-order neural network; η ue is the learning rate of the high-order neural network, α u is the coefficient that adjusts the influence of higher-order derivatives, is the loss function of the high-order neural network; is the true category of the sample; u It is a classification category for high-order neural networks.

[0172] S403: During the back propagation process, dynamically adjust the connection sparsity in the network according to predefined rules and performance indicators of the current network, including adjusting the number of connections, so that the high-order neural network algorithm can self-regulate and adapt to the complexity of the data;

[0173] The calculation method for sparsity determination in the dynamic reverse backtracking sparse strategy is expressed as:

[0174]

[0175] In the formula, is the sparsity index of the kth layer of the high-order neural network; β u is a binary vector that determines whether to maintain or disconnect the connection; θ u is the sparsity threshold, mce is the number of neurons in this layer, and ⊙ represents element-wise multiplication; is the weight of the jth neuron in the kth layer of the high-order neural network.

[0176] If the above formula is true, the connection of the kth layer of the high-order neural network is maintained, otherwise the connection is disconnected to achieve dynamic adjustment of the connection sparsity in the neural network.

[0177] Sparsity threshold θ u Dynamic adjustment is made according to the performance of the network during training. The calculation method is expressed as:

[0178]

[0179] In the formula, λ u is the adjustment coefficient; std() is the function for calculating the standard deviation to ensure that the model maintains an appropriate balance between complexity and overfitting during training; is the mean value of the layer weights of the high-order neural network.

[0180] S404: Adopt an adaptive feature impact feedback mechanism to dynamically evaluate and adjust the impact of each input feature on the model prediction during the training process, so as to achieve more efficient and accurate feature weight adjustment. Specifically, the feature impact is defined as the importance score of each feature in the current model prediction, based on the partial derivative of the feature to the model output, that is, the absolute value of the gradient, expressed as:

[0181]

[0182] in, is the feature influence; It is the input feature of the kth layer of the high-order neural network.

[0183] The weight update variable of the high-order neural network is calculated based on the feature influence, expressed as:

[0184]

[0185] In the formula, Update variables for the weights of high-order neural networks; γ u is the adjustment coefficient in the learning process. Preferably, γ u Set to 0.2.

[0186] Furthermore, the increment of the weight update variable depends not only on the gradient of the loss function, but also on the real-time impact evaluation of each feature, so that the algorithm pays more attention to the features that have a great impact on the output during the training process, thereby improving the training efficiency and model accuracy. The calculation method is expressed as:

[0187]

[0188] In the formula, Increment the variable for weight update; is the loss function of the high-order neural network.

[0189] S405: Repeat forward propagation and back propagation. In each iteration, the weights and parameters of the high-order neural network are updated as follows:

[0190]

[0191] Where ← is the parameter update operation; η us is the learning rate of the high-order neural network; is the loss function of the high-order neural network.

[0192] S406: Repeat iterations S402 to S405 until a preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0193] Step 5: Collect new data of the heavy-duty truck electric drive axle to be evaluated, input it into the trained feature extraction model for feature processing, and input the processed features into the trained classifier model for classification to obtain the classification results.

[0194] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A heavy truck electric drive axle fault diagnosis method based on generative adversarial network, characterized in that: The steps include: Step 1: Collect the operating data of the heavy-duty truck electric drive axle under different working conditions and mark the fault category for each data; Step 2: Construct a generative adversarial network based on adaptive random perturbation, and input the data obtained in step 1 into the generative adversarial network based on adaptive random perturbation to generate samples and achieve data expansion; merge the expanded data with the original data in step 1 to obtain a training data set; The adaptive random perturbation-based generative adversarial network introduces an entropy-weighted dynamic back-propagation mechanism on the basis of the generative adversarial network, that is, the learning rate in the back-propagation process is adjusted according to the information entropy of the generated data, so as to optimize the learning efficiency of the generator and the quality of the generated data; Step 3: construct a feature extraction model, and train the feature extraction model using the data of the training data set; the feature extraction model is a 6-layer fully connected neural network, and a bionic algorithm is used to optimize the training process of the feature extraction model during the training process; Step 4: Construct a classifier model and use the feature-extracted data to train the classifier model; the classifier model is a high-order neural network based on reverse backtracking sparseness; Step 5: Collect new data of the heavy-duty truck electric drive axle to be evaluated, input it into the trained feature extraction model for feature processing, and input the processed features into the trained classifier model for classification to obtain the classification results.

2. The heavy truck electric drive axle fault diagnosis method based on generative adversarial network according to claim 1 is characterized in that: The training process of the generative adversarial network based on adaptive random perturbation is as follows: S201: Initialize generator G c and the discriminator D c The generator receives a random noise vector as input and outputs data with the same structure as the real data; the discriminator determines the authenticity of the input data: in, represents the data generated by the generator, z c represents the input random noise vector, Represents the parameters of the generator; G c () is the generator function; y c represents the output of the discriminator, x c Represents the input data, represents the parameters of the discriminator; S202: In each iteration, the entropy weighted dynamic back propagation algorithm is used to calculate the entropy of the current generated data, and the learning rate η of the generative adversarial network is adjusted according to the entropy value c : In the formula, H c () represents the information entropy of generated data; Nce is the characteristic dimension of generated data; is the i-th feature of the generated data; p() is the probability distribution function, represents the probability distribution of the i-th eigenvalue of the generated data, estimated by the probability density function of the generated data; Model the distribution of generated data and use Gaussian mixture model fitting to estimate the probability distribution of each eigenvalue: Mce is the number of Gaussian mixture components; is the weight of the jth mixture component, G s () is Gaussian distribution, and are the mean and variance of the jth mixture component respectively; S203: The generator generates new data samples, which are evaluated by the discriminator. The output of the discriminator is used to calculate the loss, which is fed back to the generator and the discriminator so that they can update their own network parameters through the gradient descent method. Among them, the iterative training process of the generator and the discriminator is expressed as: Where ← is the parameter update operation; and are the parameters of the generator and the discriminator respectively; represents the gradient with respect to the generator parameters; represents the gradient of the discriminator parameters; D c () is the discriminator function; The loss function is specifically: in, represents the total loss function of the generative adversarial network, is the adversarial loss function of the generative adversarial network, is the adaptive weight loss function, α c is the weight factor of the first loss function; β c is the weight factor of the second loss function; represents expectation; ~ represents compliance with a specific distribution; p data is the distribution of real data; p z is the distribution of noise; is the weight of the data feature, which is used to adjust the contribution of different features in the loss; is the variance of the feature, which is used to measure the variability of the feature in the data set; ∈ cks is a small positive number to prevent the denominator from being zero; a c,i,j Represents the dependency between the i-th feature and the j-th feature, cov() represents the covariance, and are the standard deviations of the i-th feature and the j-th feature respectively; is the i-th feature of the generated data; is the jth feature of the generated data; cew It is the dynamic adjustment item adjustment coefficient, which is used to control the impact of the adjustment item; S204: Repeat steps S202 to S203 until a preset stop iteration condition is met, and the training of the generative adversarial network based on adaptive random perturbation is completed.

3. The heavy truck electric drive axle fault diagnosis method based on generative adversarial network according to claim 2 is characterized in that: The training process of the feature extraction model includes: S301: According to the bionic algorithm initialization method, in the initialization stage, an initial population is generated, and each individual represents a configuration of network weights; let the population size be N p , the weights and biases for the i-th individual are initialized as: In the formula, Represents the weight matrix of the i-th individual in the initial state; represents the bias matrix of the i-th individual in the initial state; σ 2 represents the initialization variance; Indicates that the mean is 0 and the variance is σ 2 Normal distribution of is a normal distribution; S302: For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function; For the i-th individual, the loss on the training dataset is calculated using its weight and bias, expressed as: Where, L pi is the loss of the neural network corresponding to the i-th individual, and is defined as the fitness of the individual; mps represents the number of samples input in the current batch; represents the composite loss function, and fsig() represents the neural network model function; Represents the characteristics of the jth sample; represents the label of the jth sample; MSE() is the mean square error function, λ ps is the regularization parameter; Reg(W pi ) is the regularization term; W pi,k is the kth weight of the neural network corresponding to the i-th individual; S303: According to the fitness of the individuals and based on the selection strategy, the best performing individuals are selected from the current population to be retained to form an excellent population as candidate solutions for the next generation; The selection strategy is defined as: Where P select (i) represents the probability of the i-th individual being selected; γ pse is the parameter that controls the selection pressure; L pk is the loss of the neural network corresponding to the kth individual; S304: Generate new individuals through crossover and mutation operations; the crossover operation is specifically to randomly select two individuals from the excellent population to exchange genes, which is expressed as: W p ′ i =a pcs W p1 +(1-a pcs )W p2 b ′ pi =a pcs b p1 +(1-a pcs )b p2 In the formula, α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first individual selected, b p1 is the bias of the neural network corresponding to the first individual selected, W p2 is the weight of the neural network corresponding to the second individual selected, b p2 is the bias of the neural network corresponding to the second individual selected, W p ′ i is the weight of the neural network corresponding to the individual after the crossover operation, b ′ pi is the bias of the neural network corresponding to the individual after the crossover operation; The mutation operation is to perform a small random perturbation on the individual weights in the excellent population, which can be expressed as: In the formula, τ 2 represents the variance of variation, W p ″ i is the weight of the neural network corresponding to the individual after the mutation operation, b ″ pi is the bias of the neural network corresponding to the individual after the mutation operation; S305: Repeat S302 to S304 until the preset stop iteration condition is met and the model training is completed.

4. The heavy truck electric drive axle fault diagnosis method based on generative adversarial network according to claim 3 is characterized in that: In S303, the probabilities of individuals being selected are sorted from high to low, and the top 30% of individuals with high probabilities are selected as excellent populations.

5. The heavy truck electric drive axle fault diagnosis method based on generative adversarial network according to claim 4 is characterized in that: The training process of the high-order neural network based on reverse backtracking sparse is as follows: S401: Initialize the weights and biases of the high-order neural network. The structure of the high-order neural network is designed to be a 4-layer feedforward structure. The parameter initialization process of the high-order neural network is expressed as: In the formula, It means the mean is 0 and the variance is The normal distribution of and They represent the weight and bias of the kth layer of the high-order neural network respectively; S402: After each forward propagation, the weight is adjusted according to the error between the output of each layer of the high-order neural network and the actual output. The error feedback affects the weight update in the form of high-order derivatives. The calculation method of the weight update amount is expressed as: In the formula, is the weight update amount of the kth layer of the high-order neural network; η ue is the learning rate of the high-order neural network, α u is the coefficient that adjusts the influence of higher-order derivatives, is the loss function of the high-order neural network; is the true category of the sample; u It is the classification category of high-order neural network; S403: During the back propagation process, the connection sparsity of the high-order neural network is calculated according to the predefined rules and the performance indicators of the current network. If the sparsity index of the k-th layer of the high-order neural network is not less than the sparsity threshold θ u , then keep the connection of the kth layer of the high-order neural network, otherwise disconnect; Sparsity index of the kth layer of high-order neural network It is expressed as: In the formula, β u is a binary vector that determines whether to maintain or disconnect the connection; θ u is the sparsity threshold, mce is the number of neurons in this layer, and ⊙ represents element-wise multiplication; is the weight of the jth neuron in the kth layer of the high-order neural network; The sparsity threshold θ u It is expressed as: In the formula, λ u is the adjustment coefficient; std() is the function for calculating the standard deviation to ensure that the model maintains an appropriate balance between complexity and overfitting during training; is the mean value of the layer weights of the high-order neural network; S404: Calculate the feature influence of each feature, calculate the weight update variable of the high-order neural network based on the feature influence, and further calculate the increment of the weight update variable: in, is the feature influence; It is the input feature of the kth layer of the high-order neural network; Update variables for the weights of high-order neural networks; γ u is the adjustment coefficient during the learning process; Increment the variable for weight update; is the loss function of the high-order neural network; S405: Repeat forward propagation and back propagation. In each iteration, the weights and parameters of the high-order neural network are updated as follows: Where ← is the parameter update operation; η us is the learning rate of the high-order neural network; is the loss function of the high-order neural network; S406: Repeat iterations S402 to S405 until a preset condition for stopping iteration is met, which means that the high-order neural network training based on reverse backtracking sparseness is completed.

6. The heavy truck electric drive axle fault diagnosis method based on generative adversarial network according to claim 1 is characterized in that: The fault categories include normal operation, abnormal temperature, abnormal vibration, abnormal torque transmission, and unstable current.

7. The heavy truck electric drive axle fault diagnosis method based on generative adversarial network according to claim 1 is characterized in that: The operating data of the heavy-duty truck electric drive axle under different working conditions include the speed of the electric drive axle, the torque value input to the electric drive axle, the temperature data of the key components of the electric drive axle, the vibration characteristics of the electric drive axle under different working conditions, the noise generated when the electric drive axle is running, the current change of the electric drive axle, the diagnostic code when a fault occurs, the location information of the electric drive axle in the vehicle, maintenance records and operating modes.

Images