A gearbox fault diagnosis method based on deep transfer learning

Abstract

The present application belongs to the technical field of fault diagnosis based on transfer learning, and more particularly relates to a gearbox fault diagnosis method based on deep transfer learning. The method comprises collecting original vibration signals of a gearbox under different operating conditions by using an accelerometer; dividing the collected data into source domain data and target domain data and performing preprocessing; constructing a pre-training model and inputting the source domain training set and the target domain training set for pre-training, and saving the best pre-training model with the highest accuracy; reading the best pre-training model, constructing a domain discriminator, an auxiliary discrimination module and a multi-label classifier after a feature extractor to obtain a fault diagnosis model and performing adversarial training; inputting the source domain test set and the target domain test set into the fault diagnosis model, and obtaining a fault diagnosis result. The present application solves the problems of the prior art, such as the limitation that the fault types of the gearbox between two fields must be the same, the limitation that the data set must be large and well labeled, and the gradient disappearance and explosion existing in the adversarial training process.

Description

Technical Field

[0001] This invention belongs to the technical field of fault diagnosis based on transfer learning, and more specifically, relates to a gearbox fault diagnosis method based on deep transfer learning. Background Technology

[0002] Fault diagnosis is a key technology in modern industrial systems, aiming to identify the causes of failures, prevent unplanned downtime, and reduce losses. Gearboxes, as commonly used components in these systems, are prone to failure and even breakdown due to harsh working environments, high mechanical loads, and long operating times. Therefore, bearings and gears within gearboxes are the primary research focus. Gearbox fault diagnosis involves monitoring the gearbox, identifying the time of failure, and classifying the fault. To identify and classify faults, multiple sensors are installed to collect data, such as vibration data or thermal imaging data. This data is then processed to determine if a fault has occurred and subsequently classified. Many deep learning-based diagnostic models have been studied for fault identification and operational mode prediction. However, existing deep learning methods typically adhere to two principles: 1) the dataset must be large and well-labeled; 2) the training and test datasets must follow the same distribution. These principles limit the application of existing deep learning methods.

[0003] Chinese patent document CN110907176B discloses a fault diagnosis method for deep adversarial transfer networks based on Wasserstein distance. The method includes measuring the distance between the feature distributions of two domains in the feature space using Wasserstein distance, adapting the feature distributions to reduce the differences between the two domains, learning domain-independent features to train an effective classifier, which is responsible for mapping domain-independent features to the category space to complete the classification task.

[0004] Compared with the above literature, the present invention removes the restriction that the fault types in the two domains within the gearbox must be the same, and adds a weight term to identify the fault types that newly appear in the target domain.

[0005] In summary, compared to deep learning, transfer learning relaxes a fundamental assumption: it allows for different probability distributions between the source and target domains. This means that a new but relevant task in the target domain can be effectively solved using knowledge learned from the source domain. However, most existing transfer learning algorithms assume that samples in the target domain necessarily belong to classes in the source domain, meaning there are no new fault classes in the target domain. In practical applications, due to the complexity and uncertainty of gearboxes, the fault types in the target domain may differ from those in the source domain under different operating conditions, potentially leading to new faults in the target domain. Unlike previous transfer learning assumptions, the new transfer learning approach assumes that only a few classes are shared between the source and target domains, with the remaining classes being completely different. This challenge severely limits the application of existing intelligent fault diagnosis methods based on transfer learning. In this situation, most existing transfer learning methods cannot be effectively utilized. The main problem is that while aligning target and source samples requires reducing domain differences, alignment is impossible due to the lack of new fault class samples during source domain training, and this is compounded by gradient vanishing and exploding problems inherent in adversarial training within transfer learning. Summary of the Invention

[0006] The present invention aims to overcome at least one of the defects of the prior art and provide a gearbox fault diagnosis method based on deep transfer learning.

[0007] The detailed technical solution of this invention is as follows:

[0008] In order to solve the above-mentioned technical problems, this invention provides a gearbox fault diagnosis method based on deep transfer learning, which solves the problems of gradient vanishing and exploding that exist in the training process, which are limited by the requirement that the dataset be large and well labeled, and that the training dataset and the test dataset follow the same distribution.

[0009] To achieve the above objectives, this invention provides a gearbox fault diagnosis method based on deep transfer learning, comprising the following steps:

[0010] S1. Use an accelerometer to collect the original vibration signals of the gearbox under different operating conditions;

[0011] S2. The collected raw vibration signal is divided into source domain data and target domain data and preprocessed. The source domain data is divided into source domain training set and source domain test set, and the target domain data is divided into target domain training set and target domain test set.

[0012] S3. Construct a pre-trained model and input the source domain training set and the target domain training set for pre-training. Set the number of iterations N. When the number of pre-training iterations is reached, switch to adversarial training and save the best pre-trained model with the highest accuracy.

[0013] The preprocessing specifically includes: extracting the time-domain waveform sequence from the original vibration signal to obtain the time-domain vibration signal; and then performing a fast Fourier transform on the time-domain vibration signal to convert the time-domain vibration signal into a frequency-domain vibration signal.

[0014] S4. Read the best pre-trained model, construct a domain discriminator, an auxiliary discriminator module and a multi-label classifier after the feature extractor to obtain a fault diagnosis model, set the Wasserstein distance of the domain discriminator, implement Lipschitz constraints on the domain discriminator and perform adversarial training, set the number of iterations K, and complete the training when the number of adversarial training reaches the number of iterations.

[0015] The domain discriminator includes one convolutional layer and two fully connected layers. After the convolution operation, the leakyReLU function is used for activation, and the leakyReLU activation function is used between the two fully connected layers.

[0016] The auxiliary discrimination module assigns weight terms to the samples distinguished by the domain discriminator, which are used to distinguish between known and unknown categories in the target domain training set;

[0017] S5. Input the source domain test set and the target domain test set into the fault diagnosis model and obtain the fault diagnosis results.

[0018] The pre-trained model includes: a feature extractor and a multi-label classifier;

[0019] The feature extractor specifically includes a residual network and an attention mechanism, with the attention mechanism added before the first residual block and after the last residual block of the residual network, respectively.

[0020] The multi-label classifier includes two fully connected layers. The first fully connected layer uses the Tanh activation function, and the second fully connected layer uses the Softmax function for classification. The domain discriminator is connected to the first fully connected layer of the multi-label classifier.

[0021] The pre-trained model initializes the feature extractor model parameters θ using the features obtained from the initial pre-training. f and multi-label classifier parameters θ c During the iteration process, the feature extractor model parameters θ f and multi-label classifier parameters θ c By minimizing the classification loss function L P Update, classification loss function L P Defined as the cross-entropy between the softmax prediction probability distribution and the one-hot encoding of the source domain data sampling labels:

[0022]

[0023] Where, x s These are the actual labels of the source domain training set, y s It is the predicted label of the source domain training set, n s It refers to the amount of data in the source domain. It is an indicator function; It is the k-th dimension value of the i-th source domain data in the predicted distribution, where K is the number of categories.

[0024] The specific steps of the S3 pre-training include:

[0025] S31. After the source domain training set and the target domain training set are processed by the first layer of convolution, the convolutional feature map is obtained by normalization and optimization using the ReLU activation function.

[0026] S32. Input the convolutional feature map into the max pooling layer, and then input it into the attention mechanism for optimization to extract the maximum feature value of the convolutional filter size β and the max pooling stride size Ipl.

[0027] S33. The largest feature value is input into the n-layer residual block and then into the attention mechanism, and finally into the average pooling layer to obtain the final feature value.

[0028] S34. The multi-label classifier performs fault classification on the received features through a fully connected layer and maps them to probability output through the softmax function.

[0029] The specific steps of S4 are as follows:

[0030] S41. Read the best pre-trained model, input the source domain training set and the target domain training set into the feature extractor to extract the source domain features and the target domain features, input them into the multi-label classifier and use the domain discriminator to distinguish whether the input data belongs to the source domain or the target domain.

[0031] S42. Set the Wasserstein distance of the domain discriminator and apply Lipschitz constraints on the domain discriminator to prune the weights of the domain discriminator to a compact space;

[0032] S43. The auxiliary discrimination module is used to assign a weight term w(xis) to the target domain samples distinguished by the domain discriminator, which is used to show the similarity with the source domain class level α.

[0033] S44. The multi-label classifier performs fault classification on the features processed by the domain discriminator through a fully connected layer, and maps them to probability output through the softmax function in the multi-label classifier.

[0034] S45. Perform iterative adversarial training, set the number of iterations K, and stop when the adversarial training reaches the number of iterations to obtain the anti-training transfer diagnostic model.

[0035] A gradient penalty term is added to the original GAN ​​objective function of adversarial training. The feature representation in the gradient penalty term consists of generated source domain features, target domain features, and points randomly selected along the straight line between the source domain features and the target domain features.

[0036] During adversarial training, the feature extractor G is optimized to increase the cross-entropy L. P (x s ,y s Minimize the binary crossover loss L. C (x t i) Domain adversarial loss L advj (x s ,x t Simultaneously, the multi-label classifier P attempts to minimize the cross-entropy L. P (x s ,y s ) and binary cross loss Simultaneously, the domain adversarial loss L advj (x s ,x t Maximize; w is the weight parameter in the auxiliary discrimination module, and the target loss function of the entire network is:

[0037]

[0038]

[0039] in, Indicates the actual label of the source domain. Indicates the source domain predicted label. Represents the target domain sample. This represents the binary crossover loss. L represents the classification loss. advj (x s ,x t ) represents the domain adversarial loss, LossC1 is the training objective of the multi-label classifier P, and LossC2 is the training objective of the feature extractor G.

[0040] S46. Based on the constructed target loss function, the fault diagnosis model is iteratively trained using the stochastic gradient descent algorithm to optimize the network parameters. After training, the optimal network parameters of the model are saved and the trained transfer diagnosis model is obtained.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] (1) The present invention provides a gearbox fault diagnosis method based on deep transfer learning, which uses the features in the known labeled source domain of the gearbox to diagnose new related diagnostic tasks through the deep transfer model, and distinguishes fault types that have not appeared in the source domain, without the need for any labeled samples.

[0043] (2) The present invention provides a gearbox fault diagnosis method based on deep transfer learning. The domain discriminator learns shared feature representation by minimizing the Wasserstein distance between the source domain and the target domain distribution, and adds a gradient penalty term to the original GAN ​​objective function of adversarial training to avoid gradient vanishing and exploding problems in the adversarial training process.

[0044] (3) The present invention provides a gearbox fault diagnosis method based on deep transfer learning, which assigns weight terms to the samples distinguished by the domain discriminator through an auxiliary discrimination module. The classifier displays similarity to the source domain at the class level; the multi-label classifier uses weight terms to establish a pseudo-decision boundary to distinguish between known fault classes and new fault classes in the target domain. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0046] Figure 2 This is a framework diagram of the gearbox fault diagnosis model in Embodiment 1 of the present invention.

[0047] Figure 3 This is a schematic diagram of the feature extractor structure in Embodiment 1 of the present invention. Detailed Implementation

[0048] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0049] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0051] Where there is no conflict, the embodiments and features described herein can be combined with each other.

[0052] Example 1

[0053] This embodiment provides a gearbox fault diagnosis method based on deep transfer learning, such as... Figure 1 As shown:

[0054] S1. Use an accelerometer to collect the original vibration signals of the gearbox under different operating conditions (different speeds and loads).

[0055] S2. Divide the collected data into source domain data and target domain data and perform preprocessing. The source domain data is divided into source domain training set and source domain test set, and the target domain data is divided into target domain training set and target domain test set.

[0056] The source domain data contains labeled known fault classes; the target domain data consists of unlabeled known fault classes and unlabeled new fault categories.

[0057] The original vibration signal acquired in step S1 is preprocessed. The preprocessing includes extracting the time-domain waveform sequence from the original vibration signal to obtain the time-domain vibration signal, and performing Fast Fourier Transform (FFT) processing on the time-domain vibration signal to convert the time-domain vibration signal into a frequency-domain vibration signal.

[0058] The preprocessed source domain data and target domain data are randomly shuffled. The source domain data is divided into a source domain training set and a source domain dataset in an 8:2 ratio, and the target domain data is divided into a target domain training set and a target domain dataset in an 8:2 ratio, with no overlap in data sampling.

[0059] S3. Construct a pre-trained model and input the source domain training set and target domain training set for pre-training. Set the number of iterations N. When the number of pre-training iterations reaches the required number, switch to adversarial training and save the best result with the highest accuracy of the pre-trained model.

[0060] The pre-trained model includes: a feature extractor and a multi-label classifier;

[0061] The feature extractor specifically includes a residual network and an attention mechanism. The attention mechanism is added to the first residual block and the last residual block of the residual network, respectively. Figure 3 As shown;

[0062] Specifically, the feature extractor model adds channel attention and spatial attention mechanisms to the one-dimensional ResNet-18; using the source domain training set X processed in step S2... s and the target domain training set X t Instances within the two mini-batch and For example, both instances are extracted by the feature extractor's extraction function r. f:x→Η and the network parameters θ of the feature extractor f Generate source domain features h s =r f (x s ) and target domain features h t =r f (x t );

[0063] The multi-label classifier includes two fully connected layers. The first fully connected layer uses the Tanh activation function, and the second fully connected layer uses the Softmax function for classification.

[0064] The pre-trained model initializes the feature extractor model parameters θ using the features obtained from the initial pre-training. f and multi-label classifier parameters θ c During the iteration process, the feature extractor model parameters θ f and multi-label classifier parameters θ c By minimizing the classification loss function L P Update, classification loss function L P Defined as the cross-entropy between the softmax prediction probability distribution and the one-hot encoding of the source domain data sampling labels:

[0065]

[0066] Where, x s These are the actual labels of the source domain training set, y s It is the predicted label of the source domain training set, n s It is the amount of data in the source domain. It is an indicator function. It is the k-th dimension value of the i-th source domain data in the predicted distribution, where K is the number of categories.

[0067] One-hot encoding encodes a tag using an N-bit state register to encode N states. Each state has its own independent register bit, and at any given time, only one bit is valid, which makes the distance calculation between features more reasonable.

[0068] The specific steps of the S3 pre-training include:

[0069] S31. After the source domain training set and the target domain training set are processed by the first layer of convolution, the convolutional feature map is obtained by normalization and optimization using the ReLU activation function.

[0070] S32. Input the convolutional feature map into the max pooling layer, and then into the attention mechanism for optimization, extracting the maximum eigenvalues ​​of the convolutional filter size β and the max pooling stride size Ipl.

[0071] S33. The largest feature value is input into the n-layer residual block and then into the attention mechanism, and finally into the average pooling layer to obtain the final feature value.

[0072] S34. The multi-label classifier performs fault classification on the received features through a fully connected layer and maps them to probability output through the softmax function.

[0073] S4. Read the optimal pre-trained model, construct a domain discriminator, an auxiliary discriminator module, and a multi-label classifier after the feature extractor to obtain a fault diagnosis model, and set the Wasserstein distance of the domain discriminator for adversarial training. Set the number of iterations K, and complete the training when the number of adversarial training iterations is reached. Figure 2 As shown;

[0074] The domain discriminator includes one convolutional layer and two fully connected layers. After the convolution operation, the leakyReLU function is used for activation, and the leakyReLU activation function is used between the two fully connected layers.

[0075] The domain discriminator is connected to the first fully connected layer of the multi-label classifier. The domain discriminator is used to distinguish whether the data belongs to the source domain or the target domain, and at the same time, it reduces the distribution difference between the source domain training set and the target domain training set.

[0076] The auxiliary discrimination module assigns weight terms to the samples distinguished by the domain discriminator, which are used to distinguish between known and unknown categories in the target domain training set.

[0077] The specific steps of S4 are as follows:

[0078] S41. Read the optimal pre-trained model, input the source domain training set and the target domain training set into the feature extractor to extract the source domain features h. s =r f (x s ) and target domain features h t =r f (x t The input is processed by a multi-label classifier, and a domain discriminator distinguishes whether the input data belongs to the source domain or the target domain.

[0079] S42. Set the Wasserstein distance of the domain discriminator and apply Lipschitz constraints on the domain discriminator to prune the weights of the domain discriminator to a compact space;

[0080] The source domain training set is transferred to the target domain training set to obtain the domain-invariant features of the fully connected layer FC, where FC is the first fully connected layer in the multi-label classifier. A domain discriminator is used to estimate the distribution difference of the fully connected layer FC.

[0081] During the migration process, the domain discriminator uses the maximum domain adversarial loss L advj With parameter θ d The relationship is used for optimization, and the domain-invariant feature is achieved by minimizing the domain adversarial loss L. advj Classification loss L C With parameter θ fc get;

[0082] Domain adversarial loss L advj The definition is as follows:

[0083]

[0084] Wherein, D(r) f (x s )) is a source domain feature r f (x s The set of Lipschitz constraints applied; D(r) f (x t )) is the target domain feature r f (x t The set of sets subject to Lipschitz constraints; n s n t represent the number of training sets in the source domain and the number of training sets in the target domain, respectively; λ represents the penalty coefficient. Representative feature representation The set to be subject to Lipschitz constraints.

[0085] The formula for Wasserstein distance is as follows:

[0086]

[0087] Where ρ(x,y) is the distance function; x and y are samples in set M; ∏(P,Q) is the set of all probability measures on M×M, with boundary values ​​P and Q. Let (μ k ) k∈N It is the Wasserstein space P p Let μ be a probability sample sequence in (M), and let P be a probability sample sequence. p Another sample of (M), when the Wasserstein distance is close to zero, i.e., W p (μ k When μ approaches 0, μ k In P p (M) weakly converges to μ;

[0088] The Wasserstein distance between the feature distributions of the source domain and the target domain is obtained by maximizing the domain discriminant loss L. wd express:

[0089]

[0090] Lipschitz constraints are implemented on the domain discriminator, clipping its weights to a compact space [-c, c]. A gradient penalty term is added to the original GAN ​​objective function. The feature representation in the gradient penalty term consists of generated source domain features, target domain features, and points randomly selected along the straight line between the source and target domain feature pairs. The loss function is shown below:

[0091]

[0092] in, The feature representation is derived from A random sample is uniformly sampled along a straight line between pairs of points sampled from the feature distributions of the source and target domains. It is a set of feature distributions of the source and target domains.

[0093] S43, The auxiliary discrimination module is used to assign weight terms to the target domain samples distinguished by the domain discriminator. Used to display similarity to the source domain class level α;

[0094] The source domain class level α is calculated by the following formula:

[0095]

[0096] Among them, y i Let n represent the predicted label of the i-th input sample. t Indicates the number of input samples

[0097] Each sample weight term Calculated by the following formula:

[0098]

[0099] Where p(x) i K represents the probability that each sample belongs to the source domain. s Indicates the number of tag categories, x i This indicates that the i-th sample is input.

[0100] S44. The multi-label classifier performs fault classification on the features processed by the domain discriminator through a fully connected layer, and maps them to probability output through the softmax function in the multi-label classifier.

[0101] The known and unknown categories in the target domain are distinguished based on the weights in the auxiliary domain discriminator; a pseudo-decision boundary is established between the known fault class and the new fault class by training a multi-label classifier P, thus identifying the target sample as a new sample; for each target sample... The probability of a new class of fault; the source domain class level α is used to determine whether a sample belongs to a known class or a new class. If the target sample If the weight term is less than the source domain class level α, then the sample is considered to be of a known class;

[0102] The binary cross-entropy loss function LC is defined as follows:

[0103]

[0104] S45. Perform iterative adversarial training, set the number of iterations K, and stop when the adversarial training reaches the number of iterations to obtain the anti-training transfer diagnostic model.

[0105] At the same time, adversarial training is used to reduce the distribution differences between the source and target datasets;

[0106] Adversarial training consists of two steps: first, training a domain discriminator to maximize the Wasserstein distance between the source and target domains; then, adjusting the feature mapper parameters θ. f To determine the parameters of the domain discriminator that minimize the Wasserstein distance, the domain-invariant feature learning strategy is expressed as:

[0107]

[0108] Among them, L wd Discriminating loss by domain representation; L gp This represents the gradient penalty term loss;

[0109] The iterative adversarial training process includes:

[0110] A gradient penalty term is added to the original GAN ​​objective function of adversarial training. The feature representation in the gradient penalty term consists of generated source domain features, target domain features, and points randomly selected along the straight line between the source domain features and the target domain features.

[0111] During adversarial training, the feature extractor G is optimized to make L P (x s ,y s Minimize, while making and L advj (x s ,x t Simultaneously, the multi-label classifier P attempts to minimize L. P (x s ,y s )and Simultaneously make L advj (x s ,x tMaximize; w is the weight parameter in the auxiliary discrimination module. The objective loss function of the entire network is:

[0112]

[0113]

[0114] in, Indicates the actual label of the source domain. Indicates the source domain predicted label. Represents the target domain sample. This represents the binary crossover loss. L represents the classification loss. advj (x s ,x t The domain adversarial loss is used, where LossC1 is the training objective of the multi-label classifier P and LossC2 is the training objective of the feature extractor G.

[0115] S46. Based on the constructed target loss function, use the stochastic gradient descent algorithm to perform iterative adversarial training on the network to optimize the network parameters. After training, save the optimal network parameters of the model and obtain the trained transfer diagnostic model.

[0116] S5. Input the source domain test set and the target domain test set into the fault diagnosis model and obtain the fault diagnosis results.

[0117] The source domain test set and the target domain test set are input into the transfer diagnostic model trained in step S4 for fault diagnosis, and the diagnostic results of different fault type samples under different working conditions are output.

[0118] The data test results are as follows:

[0119] To demonstrate the feasibility of the above method, the algorithm will be tested using a dataset from a certain university (SEU dataset).

[0120] The SEU dataset is a gearbox dataset provided by a university. It consists of two subsets: bearing and gear datasets, both collected from a transmission system dynamics simulator. Two operating conditions, with speed-load configurations set to 20Hz-0V and 30Hz-2V, are considered distinct tasks, denoted as Task 0 and Task 1, respectively.

[0121] As shown in Table 1, each subset contains five conditions: one healthy state and four fault states; four migration methods are set, as shown in Table 2.

[0122] Table 1 Experimental Gearbox Dataset

[0123]

[0124] Table 2 Migration Task List

[0125]

[0126] After multiple iterations of training, the accuracy rates for known faults and new class identification in Task T1 were 90.13% and 94.65%, respectively; in Task T2, they were 91.75% and 89.43%; in Task T3, they were 95.16% and 88.97%; and in Task T4, they were 85.65% and 87.66%. This demonstrates that the method can effectively learn and extract information features under different operating conditions, achieving cross-domain diagnosis.

[0127] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A gearbox fault diagnosis method based on deep transfer learning, characterized in that, include; S1. Use an accelerometer to collect the original vibration signals of the gearbox under different operating conditions; S2. Divide the collected data into source domain data and target domain data and perform preprocessing. The source domain data is divided into source domain training set and source domain test set, and the target domain data is divided into target domain training set and target domain test set. S3. Construct a pre-trained model and input the source domain training set and the target domain training set for pre-training. Set the number of iterations N. When the number of pre-training iterations is reached, switch to adversarial training and save the best pre-trained model with the highest accuracy. The pre-trained model includes: a feature extractor and a multi-label classifier; The specific steps of the S3 pre-training include: S31. After the source domain training set and the target domain training set are processed by the first layer of convolution, the convolutional feature map is obtained by normalization and optimization using the ReLU activation function. S32. Input the convolutional feature map into the max pooling layer, and then input it into the attention mechanism for optimization to extract the maximum feature value of the convolutional filter size β and the max pooling stride size Ipl. S33. The largest feature value is input into the n-layer residual block and then into the attention mechanism, and finally into the average pooling layer to obtain the final feature value. S34. The multi-label classifier performs fault classification on the received features through a fully connected layer and maps them to probability output through the softmax function. S4. Read the best pre-trained model, construct a domain discriminator, an auxiliary discriminator module and a multi-label classifier after the feature extractor to obtain a fault diagnosis model, set the Wasserstein distance of the domain discriminator, implement Lipschitz constraints on the domain discriminator and perform adversarial training, set the number of iterations K, and complete the training when the number of adversarial training reaches the number of iterations. The domain discriminator includes one convolutional layer and two fully connected layers. After the convolution operation, the leakyReLU function is used for activation, and the leakyReLU activation function is used between the two fully connected layers. The auxiliary discrimination module assigns weight terms to the samples distinguished by the domain discriminator, which are used to distinguish between known and unknown categories in the target domain training set; The specific steps of S4 are as follows: S41. Read the best pre-trained model, input the source domain training set and the target domain training set into the feature extractor to extract the source domain features and the target domain features, input them into the multi-label classifier and use the domain discriminator to distinguish whether the input data belongs to the source domain or the target domain. S42. Set the Wasserstein distance of the domain discriminator and apply Lipschitz constraints on the domain discriminator to prune the weights of the domain discriminator to a compact space; S43, The auxiliary discrimination module is used to assign weight terms to the target domain samples distinguished by the domain discriminator. Used to display the difference between the source domain class level and the source domain class level. Similarity; S44. The multi-label classifier performs fault classification on the features processed by the domain discriminator through a fully connected layer, and maps them to probability output through the softmax function in the multi-label classifier. S45. Perform iterative adversarial training, set the number of iterations K, and stop when the adversarial training reaches the number of iterations to obtain the anti-training transfer diagnostic model. S46. Based on the constructed target loss function, use the stochastic gradient descent algorithm to perform iterative adversarial training on the model to optimize the network parameters. After training, save the optimal network parameters of the model and obtain the trained transfer diagnostic model. The iterative adversarial training process includes: A gradient penalty term is added to the original GAN ​​objective function of adversarial training. The feature representation in the gradient penalty term consists of generated source domain features, target domain features, and points randomly selected along the straight line between the source domain features and the target domain features. During adversarial training, the feature extractor G is optimized to make... Minimize, while making and Maximize; meanwhile, the multi-label classifier P attempts to minimize and At the same time maximize; The weight parameters in the auxiliary discrimination module; the objective loss function of the entire network is: （2）； （3）； in, Indicates the actual tag of the source domain. Indicates the source domain predicted label. Represents the target domain sample. This represents the binary crossover loss. Represents classification loss. Domain confrontation losses; S5. Input the source domain test set and the target domain test set into the fault diagnosis model and obtain the fault diagnosis results.

2. The gearbox fault diagnosis method based on deep transfer learning according to claim 1, characterized in that, The preprocessing specifically includes: extracting the time-domain waveform sequence from the original vibration signal to obtain the time-domain vibration signal; and then performing a fast Fourier transform on the time-domain vibration signal to convert the time-domain vibration signal into a frequency-domain vibration signal.

3. The gearbox fault diagnosis method based on deep transfer learning according to claim 2, characterized in that, The feature extractor specifically includes a residual network and an attention mechanism, with the attention mechanism added before the first residual block and after the last residual block of the residual network, respectively. The multi-label classifier includes two fully connected layers. The first fully connected layer uses the Tanh activation function, and the second fully connected layer uses the Softmax function for classification. The domain discriminator is connected to the first fully connected layer of the multi-label classifier.

4. The gearbox fault diagnosis method based on deep transfer learning according to claim 1, characterized in that, The pre-trained model initializes the feature extractor model parameters using the features obtained from the initial pre-training. Multi-label classifier parameters ; During the iteration process, the feature extractor model parameters Multi-label classifier parameters By minimizing the classification loss function Update, classification loss function Defined as the cross-entropy between the softmax prediction probability distribution and the one-hot encoding of the source domain data sampling labels: （1） in, These are the actual labels of the source domain training set. It is the predicted label of the source domain training set. It is the amount of data in the source domain. It is an indicator function. It is the k-th dimension value of the i-th source domain data in the predicted distribution, where K is the number of categories.

Images