A multi-condition mechanical fault diagnosis method based on spectral norm GAN

Abstract

The present invention discloses a multi-operating-condition mechanical fault diagnosis method based on spectral norm GAN, which includes: data acquisition and signal preprocessing based on short-time Fourier transform, taking the original time-domain vibration signal of the machine as the input sample, collecting experimental data of variable working conditions as samples of different fields, and marking field labels, wherein the data with artificially marked labels are used as source domain data; the labels that are not artificially marked are target domain data; the source domain data and the target domain data are converted into feature time-frequency graphs by short-time Fourier transform; the source domain feature time-frequency graphs are convoluted to extract fault features, and an initialized convolutional neural network model is constructed to obtain a pre-trained model. The method obtains cross-domain invariant features through adversarial learning between feature extractors and field discriminators, realizes migration between different fields, and realizes intelligent diagnosis of variable working-condition mechanical faults.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical fault diagnosis and artificial intelligence technology, and specifically is a multi-operating condition mechanical fault diagnosis method based on spectral norm GAN. Background Art

[0002] In the context of intelligent manufacturing, along with the development of information technology, the deep integration of informatization and industrialization has become an important entry point for promoting the strategy of manufacturing power, and intelligent fault diagnosis of mechanical equipment has become one of the key technologies of modern industry. Intelligent fault diagnosis integrates the technologies of mechanical, computer, artificial intelligence and other disciplines to identify the operation mode of mechanical equipment, which can reduce the economic losses caused by equipment failure, improve economic benefits, and more importantly, protect the life safety of equipment users. Therefore, it is particularly important to implement intelligent fault diagnosis for mechanical equipment.

[0003] The essence of equipment fault diagnosis is a pattern recognition problem, including a series of feature extraction and fault identification. In recent years, researchers in the field of fault diagnosis have proposed various fault diagnosis models based on deep learning. These models can achieve quite high accuracy under ideal laboratory conditions, but are difficult to apply in practical applications. There are mainly three reasons: 1) The target domain data (actual diagnosis data) lacks labels or even has no labels. When the actual equipment is running, the signals of key components are difficult to collect, and the fault signals are far less than the normal signals. 2) Even if a large amount of fault data is collected, manual calibration of labels is time-consuming and laborious and requires expert knowledge. 3) The operating conditions of mechanical equipment are complex. During the actual operation of the equipment, the load, speed and other working conditions of the equipment are not constant, which leads to a certain degree of deviation between the data distribution of model training and the data distribution of the final test. This difference in data distribution leads to the low accuracy and lack of robustness of the fault diagnosis model under complex working conditions.

[0004] In 2014, Goodfellow proposed the Generative Adversarial Networks (GAN), which enables the neural network to learn the feature representation of data and the distance measurement between data from different sources in an unsupervised manner by alternately training the generator network and the discriminator network. The input of the generator network is a latent space vector that follows a certain distribution. It generates data with the same distribution as the training samples through the generator. The input of the discriminator network is the generated data and the original data. Its function is to distinguish whether the given input is real data or generated data through the discriminator. The generator and the discriminator are alternately trained to update the parameters and repeat this process until the training is completed. The sign of the completion of training is that the two networks reach Nash equilibrium, that is, the discriminator with discriminative ability cannot distinguish whether the input is real data or fake data generated by the generator. The ability of GAN to learn feature representation and distance measurement in an unsupervised manner helps to realize the knowledge transfer of the mechanical fault diagnosis model under variable working conditions. Even if the diagnostic target domain lacks sample labels, the accuracy of the mechanical fault diagnosis model can still be improved. Summary of the invention

[0005] To achieve the above object, the present invention provides the following technical solution: a multi-condition mechanical fault diagnosis method based on spectral norm GAN, comprising the following steps:

[0006] S1, data collection and signal preprocessing based on short-time Fourier transform. The original time-domain vibration signal of the mechanical equipment is used as the input sample. The experimental data under variable working conditions are collected as samples in different fault diagnosis fields, and the fault diagnosis field labels are marked. The data with manual labels are used as source domain data, and the data with no manual labels are used as target domain data. The source domain data and the target domain data are subjected to short-time Fourier transform to convert the time-domain vibration signal into a characteristic time-frequency diagram.

[0007] S2, model pre-training, performing convolution operation on the characteristic time-frequency graph of the source domain, extracting fault features, constructing an initialized convolutional neural network model, forming a high-precision fault diagnosis model for the source domain, using the high-precision fault diagnosis model of the source domain as a pre-training model, and migrating the parameters in the convolutional neural network model to the fault diagnosis model of the target domain, wherein the convolutional neural network model includes a feature extractor and a fault identification classifier;

[0008] S3, based on the feature representation learning of spectral norm GAN, the generator in GAN is used to improve the feature extractor of the convolutional neural network model, and a feature extractor for the target domain is constructed, so that the convolutional neural network model can extract effective fault features; the parameters of the feature extractor are initialized, and the weights and biases of the feature extractor are obtained through training and optimization; the parameters of the feature extractor are regularized using the spectral norm, so that the feature extractor can meet the 1-Lipshcitz condition and improve the robustness of the fault diagnosis model;

[0009] S4, domain discriminator based on spectral norm GAN, uses the discriminator in GAN to build a multi-classification domain discriminator for different fault diagnosis fields. The domain discriminator can learn the differences between different fault diagnosis fields, identify the fault features in S3, and determine the field to which the fault features belong; under the supervision of the domain discriminator, the feature extractor can learn fault features that are invariant across domains; the domain discriminator uses the spectral norm to regularize the network weights, so that the domain discriminator satisfies the 1-Lipshcitz condition, thereby improving the robustness of the fault diagnosis model;

[0010] S5, adversarial knowledge transfer and diagnosis, including convolutional neural network model optimization, the objective function of optimizing the convolutional neural network model in S2 is:

[0011]

[0012] Where L CE is the cross entropy loss, y s are the predicted labels and manual labels for the source domain data, respectively. They are the predicted labels and pseudo labels for the target domain data, and the pseudo labels can be used with the predicted labels To calculate:

[0013]

[0014] The objective function for optimizing the feature extractor in S3 is:

[0015]

[0016] in d s are the domain prediction value and domain label of the source domain data respectively. Domain prediction value for target domain data;

[0017] The objective function for optimizing the domain discriminator in S4 is:

[0018]

[0019] in is the domain label of the target domain data;

[0020] Alternately train the following joint optimization objective:

[0021]

[0022]

[0023] To achieve adversarial knowledge transfer of the diagnostic model and reduce the differences in fault features between domains; when the training tends to be stable, the feature extractor can learn fault features that are invariant across domains; during the adversarial training process, the fault recognition classifier, feature extractor, and domain fault recognition classifier are optimized, so that the fault recognition classifier can efficiently diagnose fault data from different working conditions based on fault features that are invariant across domains.

[0024] Further, preferably, S2 further comprises:

[0025] S2.1, convolutional neural network model, including feature extractor and fault identification classifier, after multi-layer convolutional neural network model optimization training, a pre-trained model for source domain data is obtained, in which the convolutional layers are stacked to form an efficient feature extractor;

[0026] S2.2, pre-training model, used to pre-train the convolutional neural network model parameters with source domain data before performing adversarial knowledge transfer;

[0027] S2.3, optimization of the convolutional neural network model, optimizing the parameters of the neural network through the back propagation algorithm until the optimization of the convolutional neural network model is completed.

[0028] Further, preferably, the S3 further comprises:

[0029] S3.1, initialize the parameters of the feature extractor, where the initialization parameters of the feature extractor and the fault identification classifier are pre-trained initialization parameters obtained from the source domain high-precision fault diagnosis model;

[0030] S3.2, feature representation learning of the target domain feature extractor, transfer the learned parameters to the feature extractor to initialize the model parameters, and through training optimization, enable the feature extractor to learn and extract fault features that are invariant across domains;

[0031] S3.3, spectral norm regularization of feature extractor and domain discriminator, introduces spectral norm regularization so that the mapping function of the feature extractor satisfies the spectral norm constraint of Lipschitz constant 1, and at the same time makes the Lipschitz constant of each layer of the domain fault identification classifier less than 1.

[0032] Further, preferably, the S4 further comprises:

[0033] S4.1, a domain discriminator based on a GAN model, wherein the GAN-based domain discriminator is used to identify the working condition information of the fault data and constrain the training of the feature extractor;

[0034] S4.2, domain discriminator under spectral norm regularization constraint, spectral norm regularization is introduced into the domain discriminator.

[0035] Further, preferably, the S5 further comprises:

[0036] S5.1, optimization of domain discriminator, extracts the same fault features in different domains, and the feature extractor is constrained by the domain discriminator;

[0037] S5.2, Optimization of feature extractor and fault identification classifier,The feature extractor optimizes parameters by adversarial knowledge transfer.,The optimal feature extractor can confuse the domain discriminator, that is, the extracted features make the domain discriminator unable to accurately identify the domain to which the fault features belong,,to extract fault features that are invariant across domains.

[0038] Further, preferably, the multi-operating-condition mechanical fault diagnosis method further includes:

[0039] 1. Construct a fault diagnosis data set using the characteristic time-frequency graphs of different fields in S1;

[0040] 2. Mechanical fault diagnosis model based on spectral norm GAN;

[0041] 3. The multi-condition mechanical fault diagnosis method based on spectral norm GAN obtains the fault diagnosis results under variable conditions.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) Traditional mechanical fault diagnosis methods are difficult to effectively diagnose mechanical equipment under variable working conditions. Therefore, the present invention adopts the method of adversarial knowledge transfer to improve the diagnostic accuracy and robustness of the fault diagnosis model under variable working conditions.

[0044] (2) The multi-condition mechanical fault diagnosis method can realize knowledge transfer but does not have the learning ability of feature differences. The present invention adopts a self-learning domain discriminator to replace a specific distance metric, and utilizes the ability of a feature extractor to learn feature representation, effectively capturing the essential characteristics of source domain and target domain fault signals in different domains;

[0045] (3) Different from the traditional knowledge transfer scenario where there is only one source domain and one target domain, the method proposed in the present invention can realize multi-condition fault diagnosis in multiple target domains, which not only greatly improves the robustness of the fault diagnosis model, but also avoids the deficiency of the traditional method that multiple transfers are required for different target domains;

[0046] (4) The present invention introduces spectral norm regularization into the feature extractor and domain discriminator to improve the robustness of the fault diagnosis model and stabilize adversarial training, thereby effectively improving the effectiveness of feature extraction and the accuracy of fault diagnosis;

[0047] (5) This method introduces GAN and knowledge transfer into fault diagnosis. It does not require labeling of target domain data or identification of specific operating conditions of equipment. However, it can still ensure that the multi-target domain fault classification accuracy reaches more than 98%, which is particularly suitable for actual industrial application scenarios and meets actual industrial needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the process of the present invention;

[0049] Figure 2 It is a structural schematic diagram of the feature extractor in the present invention;

[0050] Figure 3 It is a structural schematic diagram of the front view of the fault identification classifier in the present invention;

[0051] Figure 4 Schematic diagram of the structure of the domain discriminator in the present invention;

[0052] Figure 5 It is a structural schematic diagram of the experimental platform in the present invention;

[0053] Figure 6 Schematic diagram of multi-target domain transfer learning in the present invention.

[0054] In the figure: 1. Drive motor; 2. Loading device; 3. Test bearing. DETAILED DESCRIPTION

[0055] See also Figures 1 to 6 In an embodiment of the present invention, a multi-operating-condition mechanical fault diagnosis method based on spectral norm GAN includes the following steps:

[0056] S1, data collection and signal preprocessing based on short-time Fourier transform (SFT), taking the original vibration signal of the mechanical equipment as the input sample, collecting experimental data under variable working conditions (such as variable load) as samples in different fault diagnosis fields, and marking different fault diagnosis field labels; the data with manual labeling is used as the source domain data (the source domain data is the data under known working conditions, and its labels can be manually marked), and the data without manual labeling is the target domain data (the target domain data is the equipment operation data collected in real time, and it is impossible and there is no way to perform real-time labeling); the source domain data and the target domain data are short-time Fourier transform to convert the time domain vibration signal into a characteristic time-frequency diagram, and construct a variable working condition fault diagnosis data set;

[0057] Signal preprocessing based on short-time Fourier transform is as follows:

[0058] Acquire data from the source domain through sensors represents the signal data in the source domain, The label representing the source domain, The labels representing the source domain working conditions are obtained in the same way as the dataset of the target domain. They represent the signal data of the target domain and the labels of the target domain working conditions respectively. Different from the traditional transfer learning diagnosis model, the target domain is not limited to one. The transfer learning of multiple target domains is as follows: Figure 6 As shown in the figure, it can use the fault recognition classifier learned in the source domain to diagnose fault data in different fields. The data in the source domain and the target domain are short-time Fourier transformed and L2 regularized according to the following formula:

[0059]

[0060] Where T window is the window function, the default is the Hamming window, and the time-frequency diagram is constructed as the input of the variable working condition fault diagnosis model;

[0061] S2, model pre-training, performing convolution operation on the source domain feature time-frequency graph, extracting fault features, constructing an initialized convolutional neural network model, realizing a high-precision fault diagnosis model for the source domain, using the high-precision fault diagnosis model of the source domain as a pre-training model, and migrating the parameters of the convolutional neural network model to the fault diagnosis model of the target domain, wherein the convolutional neural network model includes a feature extractor and a fault identification classifier;

[0062] S3, feature representation learning based on spectral norm GAN, using the generator in GAN to improve the feature extractor of the convolutional neural network model, constructing a feature extractor for the target domain, so that the convolutional neural network model can extract effective fault features; initializing the parameters of the feature extractor, and obtaining the weights and biases of the feature extractor through training and optimization; using the spectral norm to regularize the parameters of the feature extractor so that the feature extractor meets the 1-Lipshcitz condition, improving the robustness of the fault diagnosis model;

[0063] S4, domain discriminator based on spectral norm GAN, uses the discriminator in GAN to build a multi-classification domain discriminator for different domains. The domain discriminator can learn the differences between domains, identify the fault features of S3, and determine the domain to which the fault features belong; under the supervision of the domain discriminator, the feature extractor can learn fault features that are invariant across domains; the domain discriminator uses the spectral norm to regularize the network weights, so that the domain discriminator satisfies the 1-Lipshcitz condition and improves the robustness of the fault diagnosis model;

[0064] S5, adversarial knowledge transfer and diagnosis, convolutional neural network model optimization, the objective function of optimizing the convolutional neural network model in S2 is:

[0065]

[0066] Where L CE is the cross entropy loss, y s are the predicted labels and manual labels for the source domain data, respectively. They are the predicted labels and pseudo labels for the target domain data, and the pseudo labels can be used with the predicted labels To calculate:

[0067]

[0068] The objective function for optimizing the feature extractor in S3 is:

[0069]

[0070] in d s are the domain prediction value and domain label of the source domain data respectively. Domain prediction value for target domain data;

[0071] The objective function for optimizing the domain discriminator in S4 is:

[0072]

[0073] in is the domain label of the target domain data;

[0074] Alternately train the following joint optimization objective:

[0075]

[0076]

[0077] To achieve adversarial knowledge transfer of the diagnostic model and reduce the differences in fault features between domains; when the training tends to be stable, the feature extractor can learn fault features that are constant across domains; during the adversarial training process, the fault recognition classifier, feature extractor, and domain fault recognition classifier are optimized, so that the fault recognition classifier can efficiently diagnose fault data from different working conditions based on fault features that are constant across domains;

[0078] Based on the target domain knowledge transfer and diagnosis of transfer learning theory, the back propagation algorithm is used to optimize the models in S3 and S4, and alternate training is performed to achieve adversarial knowledge transfer of diagnostic knowledge and reduce the differences in fault features between different domains. When the training tends to Nash equilibrium, the feature extractor can learn fault features that are invariant across domains. In order to ensure that the fault recognition classifier also has a high recognition accuracy for the updated fault features, the fault recognition classifier obtained by S2 is optimized during the adversarial training process, so that the fault recognition classifier can efficiently diagnose fault data from different working conditions.

[0079] See also Figure 2 and Figure 3 In this embodiment, S2 further includes:

[0080] S2.1, convolutional neural network model, including feature extractor and fault recognition classifier, after multi-layer convolutional neural network model optimization training, a pre-trained model for source domain data is obtained, in which convolution layers are stacked to form an efficient feature extractor; starting from the input layer, fault features are extracted layer by layer through convolution kernels, and finally input into the fault recognition classifier to realize fault pattern recognition. The feature vector of the lth layer can be expressed as:

[0081]

[0082] Where φ is the activation function, * represents the convolution operation, the convolution kernel performs a convolution operation on the feature vector of the previous layer, adds the bias b, and then activates it through the activation function as the input of the next convolution layer.

[0083] Batch Normalization (BN) is used between convolutional layers to reduce the distribution offset between layers:

[0084]

[0085] Where E[x (k) ] is the mean value of the data in this layer, Var[x (k) ] is the variance of the data in this layer.

[0086] The convolutional neural network of the source domain fault data is then supervised trained, and its objective function is the cross entropy function:

[0087]

[0088] Where y is the label value of the source domain data, and is the predicted output of the convolutional neural network, which is obtained by the following softmax function:

[0089]

[0090] O represents the output vector of the previous convolutional layer, and b is the additional bias. After the softmax function, the neural network output is converted into a predicted probability value.

[0091] Finally, after multi-layer convolutional neural network optimization training, a convolutional network diagnosis pre-training model for source domain data can be obtained, in which the convolutional layers are stacked to form an efficient feature extractor;

[0092] S2.2, pre-training model, used to pre-train the convolutional neural network model parameters with source domain data before performing adversarial knowledge transfer;

[0093] S2.3, optimization of the convolutional neural network model, optimize the parameters of the neural network through the back propagation algorithm until the convolutional neural network model optimization is completed. The convolutional neural network can obtain the predicted value through forward propagation, and the error between the true value and the predicted value can be obtained by calculating the loss function. The prediction error can be reduced through gradient back propagation and the model can be optimized. The gradient descent back propagation formula is as follows:

[0094]

[0095] θ is the parameter to be updated, η is the learning rate, which measures the step size of learning. is the gradient of the loss function with respect to the updated variable.

[0096] See also Figure 3 As a preferred embodiment, the S3 further comprises:

[0097] S3.1, initialize the parameters of the feature extractor, where the initialization parameters of the feature extractor and the fault identification classifier are pre-trained initialization parameters obtained from the source domain high-precision fault diagnosis model;

[0098] Combines parameter-based and feature-based transfer learning for fault diagnosis. Transfer learning based on parameters in S3,

[0099]

[0100]

[0101] The parameters of the feature extractor and fault identification classifier are initialized by the optimal parameters obtained through supervised pre-training;

[0102] S3.2, feature representation learning of the target domain feature extractor, transfer the learned parameters to the feature extractor to initialize the model parameters, and through training optimization, enable the feature extractor to learn and extract fault features that are invariant across domains;

[0103] according to Figure 1 The parameters learned in S3 are transferred to the feature extractor as the initialization of the model parameters. FE(·) represents the feature map of the feature extractor, which maps the time-frequency diagram of the signal into the features of the high-dimensional space.

[0104] f=FE(x)

[0105] The time-frequency graphs of the source domain and the target domain are simultaneously input into the feature extractor to obtain the high-dimensional features of the source domain and the target domain.

[0106] GAN achieves learning of the original data distribution under the constraints of a learnable discriminator, and the optimization of its generator can be expressed as:

[0107]

[0108] Therefore, based on GAN theory, the feature extractor of the target domain is improved to achieve:

[0109]

[0110] Where Div represents a distance metric, which is applied in the form of a discriminator in the present invention. They are the feature distribution learned by the target domain feature extractor and the fixed feature distribution in the source domain. Through training optimization, FE can learn and extract features that are independent of the working condition information;

[0111] S3.3, spectral norm regularization of feature extractor and domain discriminator, introduces spectral norm regularization so that the mapping function of the feature extractor satisfies the spectral norm constraint of Lipschitz constant 1, and at the same time makes the Lipschitz constant of each layer of the domain fault identification classifier less than 1.

[0112] The spectral norm regularization is introduced to make the mapping function of the feature extractor satisfy the constraint that the Lipschitz constant is 1. The spectral norm is defined as:

[0113]

[0114] That is, the largest singular value of matrix A.

[0115] Lipschitz constraint, that is, a function f(x) satisfies

[0116]

[0117] Where M is a constant, and the gradient of the function that satisfies this constraint is limited, so the function will not change too quickly, which is conducive to model training.

[0118] By the inequality:

[0119]

[0120] It can be seen that in order to make the feature extractor satisfy the 1-Lipschitz constraint, the Lipschitz constant of each layer of the discriminator D is less than 1, that is,

[0121]

[0122] in Represents the mapping relationship from the input layer to the output layer, W l is the weight matrix ||a 1 || Lip is the Lipschitz constant of the activation function.

[0123] The spectral norm regularization formula introduced is:

[0124]

[0125] in is the weight matrix constrained by regularization, W is the original weight matrix, and it is noted that after the weight matrix of the feature extractor is regularized by the spectral norm, the Lipschitz constant of each layer has an upper bound of 1, so the entire model satisfies the 1-Lipschitz constraint.

[0126] In this embodiment, S4 further includes:

[0127] S4.1, a domain discriminator based on a GAN model, the domain discriminator based on GAN is used to identify the working condition information of fault data and constrain the training of the feature extractor. The discriminator of the GAN is a two-fault recognition classifier, which evaluates the difference between the distribution of real data and generated data. The domain discriminator based on the GAN model is a multi-classification domain discriminator, which is used to identify the working condition information of fault data and learn a suitable distance metric to achieve feature-based transfer learning. Combined with the GAN theory, for a certain field, the loss function of the improved domain discriminator is a multi-classification cross entropy function:

[0128]

[0129] in It is the probability estimate of the domain of the feature made by the discriminator, which is calculated by the domain discriminator and the softmax function

[0130]

[0131] Where O is the output of the previous layer of the discriminator. After learning, the domain discriminator can identify the working conditions to which multiple different fault features belong, thereby providing constraints for the learning of the feature extractor;

[0132] S4.2, domain discriminator under spectral norm regularization constraint, spectral norm regularization is introduced in the domain discriminator. The domain discriminator under spectral norm regularization constraint, the GAN-based transfer learning method, introduces spectral norm regularization in the domain discriminator, therefore, the spectral regularization formula derived from step 3 regularizes the weights of the domain discriminator:

[0133]

[0134] in is the discriminator weight matrix constrained by regularization, W d is the original discriminator weight matrix. The Lipschitz constant of each layer of D has an upper bound of 1, so the entire discriminator D satisfies the 1-Lipschitz constraint. The feature extractor also satisfies the 1-Lipschitz constraint, thereby achieving stable adversarial knowledge transfer learning.

[0135] In this embodiment, S5 further includes:

[0136] S5.1 Optimization of domain discriminator

[0137] In order to extract the same features in different domains, the feature extractor must be constrained by the domain discriminator, so the domain classification loss function for multiple target domains is:

[0138]

[0139] in d s is the predicted value and label value of the domain discriminator for the domain to which the feature belongs, The domain discriminator predicts the values ​​and label values ​​of the domains to which multiple target domains belong, and jointly updates the domain discriminator by back-propagating the gradient, so that it can effectively identify the domains described by the features.

[0140] S5.2 Optimization of feature extractor and fault identification classifier

[0141] Optimization of feature extractor and fault identification classifier. The feature extractor optimizes parameters by adversarial knowledge transfer. The optimal feature extractor can confuse the domain discriminator, that is, the extracted features make the domain discriminator unable to accurately identify the domain to which the fault features belong, so as to extract fault features that are invariant across domains. In order to extract domain-invariant features of multiple target domains, the loss function of the feature extractor for multiple target domains is:

[0142]

[0143] in d s is the predicted value and label value of the domain discriminator for the domain to which the feature belongs, d s It represents the domain prediction value of the target domain and the domain label of the source domain. Therefore, after optimizing the objective function, the difference in features between domains is reduced, so that the extracted features can make the domain discriminator unable to distinguish which domain it comes from.

[0144] In order to ensure that the fault recognition classifier still has a high diagnostic accuracy for the updated features, the fault recognition classifier is updated while alternately optimizing the feature extractor and the fault recognition classifier.

[0145]

[0146] in y s are the predicted values ​​and corresponding labels of the source domain. The predicted value and pseudo label of the target domain fault identification classifier. Because the label of the target domain data set is unknown, in order to improve the diagnostic accuracy of the target domain, the present invention introduces pseudo labels. The dimension of the original predicted value is the number of classifications, so the neuron with the largest prediction probability is 1, and the rest of the neurons are 0, that is:

[0147]

[0148] Therefore, after adversarial training and training of the fault identification classifier, the overall objective function of the present invention can be summarized as:

[0149]

[0150]

[0151] Among them, the first part is the optimization of the domain discriminator, and the weight coefficient is λ d The second part is the optimization of feature extraction and fault identification classifier, with a weight coefficient of λ c and λ f , alternately training the domain discriminator, feature extractor and fault identification classifier makes it possible to extract unified feature representations under different working conditions and obtain higher diagnostic accuracy under different working conditions.

[0152] In this embodiment, a multi-operating condition mechanical fault diagnosis method based on spectral norm GAN includes:

[0153] 1. Dataset construction: construct a fault diagnosis dataset using the characteristic time-frequency graphs of different fields in S1;

[0154] 2. Establish a mechanical fault diagnosis model based on spectral norm regularization and GAN transfer learning. The convolutional neural network architecture of the feature extractor and fault identification classifier is as follows: Figure 2 and Figure 3 As shown in the figure, the input is a time-frequency graph after short-time Fourier transform and L2 regularization (convolution kernel weight spectrum regularization). After layer-by-layer feature extraction, the output of the feature extractor is a 4*4@256 feature map, which is used as the input of the fault recognition classifier and the domain discriminator. The fault recognition classifier identifies / globally pools the extracted feature map to determine the type of fault. The convolutional neural network architecture of the domain discriminator is shown in the figure. Figure 4 As shown, identify the field to which the feature map belongs;

[0155] 3. Based on spectral norm regularization, GAN and transfer learning, the fault diagnosis method of variable operating conditions machinery is used to obtain the fault diagnosis results under variable operating conditions.

[0156] The fault diagnosis results of this embodiment are shown in Table 1:

[0157] Table 1

[0158]

[0159] By analyzing the experimental results of the multi-condition mechanical fault diagnosis method based on spectral norm GAN on the bearing data set, the following conclusions can be drawn:

[0160] 1. Through feature learning based on spectral norm regularized generative adversarial networks, features that are independent of domain information (features with domain invariance) can be learned. Through learning, the domain discriminator can learn the ability to identify different working condition data;

[0161] 2. Through adversarial knowledge transfer learning, the distance between the source domain working conditions and the target domain working conditions can be reduced, and feature-based transfer learning can be achieved;

[0162] 3. Through classification diagnosis based on domain-invariant features, the overall generalization ability of the model can be improved. From the test results, it can be seen that this method has achieved good accuracy in various variable working condition tasks. In particular, the present invention can diagnose faults in multiple target domains, and by utilizing the knowledge of the source domain and multiple target domains, the accuracy of the model can be greatly improved.

[0163] As a preferred embodiment, the present invention also includes an experimental platform for data acquisition, which includes a drive motor 1, a loading device 2, a test bearing 3 and a NIPXle-1082 data acquisition system. The drive motor 1 is connected to the test bearing 3 and rotates the test bearing 3. The loading device 2 is electrically connected to the test bearing 3 and is used to adjust the load of the bearing. The NIPXle-1082 data acquisition system collects vibration signals of the test bearing 3 at a sampling frequency of 10 kHz, fully considering the sampling frequency and fault frequency, and each sample consists of 2048 data points.

[0164] The fault diagnosis data set is constructed by the collected fault signals as shown in Table 2:

[0165] Table 2 Fault diagnosis dataset

[0166]

[0167] The test bearings are normal bearings, outer ring fault bearings, inner ring fault bearings, and rolling element fault bearings. The fault sizes of the faulty bearings are 0.2mm, 0.4mm, and 0.6mm, respectively. Therefore, there are 10 types of health states. 400 training samples and 200 test samples are collected for each type of health state. The load of 0KN, 1KN, 2KN, and 3KN is applied by the loading device to simulate the scene of variable working conditions.

[0168] Specifically, the variable-condition mechanical fault diagnosis method based on spectral norm GAN is mainly divided into five steps:

[0169] 1. Data acquisition and signal preprocessing based on short-time Fourier change;

[0170] 2. Model pre-training based on source domain data and convolutional neural network;

[0171] 3. Feature representation learning based on spectral norm GAN;

[0172] 4. Domain discriminator based on spectral norm GAN;

[0173] 5. Counteract knowledge transfer and diagnosis.

[0174] What is described above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-condition mechanical fault diagnosis method based on spectral norm GAN, It is characterized in that The following steps are involved: S1, data collection and signal preprocessing based on short-time Fourier transform, taking the original time-domain vibration signal of the mechanical equipment as the input sample, collecting experimental data of variable working conditions as samples of different fields, and marking the field labels, among which the data with manual labeling is used as the source domain data; the labels that are not manually labeled are the target domain data; the source domain data and the target domain data are short-time Fourier transform to convert the time-domain vibration signal into a characteristic time-frequency diagram; S2, model pre-training, performing convolution operation on the source domain feature time-frequency graph, extracting fault features, constructing an initialized convolutional neural network model, realizing a high-precision fault diagnosis model for the source domain, using the high-precision fault diagnosis model of the source domain as a pre-training model, and migrating the parameters of the convolutional neural network model to the fault diagnosis model of the target domain, wherein the convolutional neural network model includes a feature extractor and a fault identification classifier; S3, feature representation learning based on spectral norm GAN, uses the generator in GAN to improve the feature extractor of the convolutional neural network model, and constructs a feature extractor for the target domain, so that the convolutional neural network model can extract effective fault features; Initialize the parameters of the feature extractor, and obtain the weights and biases of the feature extractor through training and optimization; S4, a domain discriminator based on spectral norm GAN, uses the discriminator in GAN to build a multi-classification domain discriminator for different domains. The domain discriminator can learn the differences between domains and identify the fault features of S3 to determine the domain to which the fault features belong; Under the supervision of the domain discriminator, the feature extractor is enabled to learn fault features that are invariant across domains; S5, adversarial knowledge transfer and diagnosis, including convolutional neural network model optimization, the objective function of optimizing the convolutional neural network model in S2 is: Where L CE is the cross entropy loss, y s are the predicted labels and manual labels for the source domain data, respectively. They are the predicted labels and pseudo labels for the target domain data, and the pseudo labels can be used with the predicted labels To calculate: The objective function for optimizing the feature extractor in S3 is: in d s are the domain prediction value and domain label of the source domain data respectively. Domain prediction value for target domain data; The objective function for optimizing the domain discriminator in S4 is: wherein is the domain label of the target domain data; Alternately train the following joint optimization objective: This is to achieve adversarial knowledge transfer of diagnostic models and reduce the differences in fault characteristics between domains.

2. According to claim 1, a multi-condition mechanical fault diagnosis method based on spectral norm GAN, Features: The S2 includes: S2.1, convolutional neural network model, including feature extractor and fault identification classifier, after multi-layer convolutional neural network model optimization training, a pre-trained model for source domain data is obtained, in which the convolutional layers are stacked to form an efficient feature extractor; S2.2, pre-training model, used to pre-train the convolutional neural network model parameters with source domain data before performing adversarial knowledge transfer; S2.3, optimization of the convolutional neural network model, optimizing the parameters of the neural network through the back propagation algorithm until the optimization of the convolutional neural network model is completed.

3. According to claim 1, a multi-condition mechanical fault diagnosis method based on spectral norm GAN, Features: The S3 includes: S3.1, initialize the parameters of the feature extractor, where the initialization parameters of the feature extractor and the fault identification classifier are pre-trained initialization parameters obtained from the source domain high-precision fault diagnosis model; S3.2, feature representation learning of the target domain feature extractor, transfer the learned parameters to the feature extractor to initialize the model parameters, and through training optimization, enable the feature extractor to learn and extract fault features that are invariant across domains; S3.3, spectral norm regularization of feature extractor and domain discriminator, introduces spectral norm regularization so that the mapping function of the feature extractor satisfies the spectral norm constraint of Lipschitz constant 1, and at the same time makes the Lipschitz constant of each layer of the domain fault identification classifier less than 1.

4. According to claim 1, a multi-condition mechanical fault diagnosis method based on spectral norm GAN, Features: The S4 includes: S4.1, a domain discriminator based on a GAN model, wherein the GAN-based domain discriminator is used to identify the working condition information of the fault data and constrain the training of the feature extractor; S4.2, domain discriminator under spectral norm regularization constraint, spectral norm regularization is introduced into the domain discriminator.

5. According to claim 1, a multi-condition mechanical fault diagnosis method based on spectral norm GAN, Features: The S5 includes: S5.1, optimization of domain discriminator, extracts the same fault features in different domains, and the feature extractor is constrained by the domain discriminator; S5.2,Optimization of feature extractor and fault identification classifier,The feature extractor optimizes parameters by adversarial knowledge transfer.,The optimal feature extractor can confuse the domain discriminator,,that is, the extracted features make the domain discriminator unable to accurately identify the domain,to which the fault features belong, to extract fault features that are invariant across domains.

6. According to claim 1, a multi-condition mechanical fault diagnosis method based on spectral norm GAN, It is characterized in that The multi-operating-condition mechanical fault diagnosis method further includes: (1) Constructing a fault diagnosis data set using the characteristic time-frequency graphs of different fields in S1; (2) Mechanical fault diagnosis model based on spectral norm GAN; (3) The multi-condition mechanical fault diagnosis method based on spectral norm GAN obtains the fault diagnosis results under variable conditions.

Images