Variable working condition mechanical fault diagnosis method based on multi-source domain class separation adversarial network

Abstract

The present application relates to mechanical parameter pattern recognition method, more particularly to a kind of variable working condition mechanical fault diagnosis method based on multi-source domain class separation counter network, for solving the domain counter fault diagnosis technology proposed for the fault diagnosis problem of variable working condition at present Although the alignment source domain and target same working condition fault feature distribution, it is mainly from the global perspective to carry out the alignment of global distribution, cannot guarantee the alignment of global domain alignment while intra-domain class distribution, cause In the final fault diagnosis classification, the phenomenon of intra-domain fault class confusion occurs, and then lead to the deficiency of the reduction of the diagnosis precision of model.The variable working condition mechanical fault diagnosis method based on multi-source domain class separation counter network constructs a kind of multi-source domain class separation counter network, for realizing the fault diagnosis of each multi-source domain migration fault diagnosis task, the network can realize the fine-grained alignment of each domain in cross-domain sub-class, effectively learn the fault diagnosis knowledge of multiple domains.

Description

TECHNICAL FIELD

[0001] The present application relates to a mechanical parameter pattern recognition method, in particular to a variable working condition mechanical fault diagnosis method based on a multi-source domain class separation adversarial network. BACKGROUND

[0002] Bearing is the most widely used component in major rotating machinery, which directly affects the health status of rotating machinery. With the rapid innovation of modern industry, bearings as the key core components of modern industrial equipment, higher reliability and safety requirements are put forward for their maintenance management. In industrial applications, fault diagnosis is the key to keep the equipment running in good condition. In recent years, with the development of artificial intelligence technology, intelligent fault diagnosis methods based on deep learning technology have received extensive attention. Past research usually assumes that there is only one source domain. Due to the particularity of single-source training samples, overfitting risk may occur in single-source domain transfer fault diagnosis, which weakens the generalization ability of the model under the working condition of the target domain. However, there are multiple related labeled source domain data with different distributions in actual industrial scenarios, which can obtain better, more comprehensive and more general diagnostic knowledge from multiple source domains, thereby improving the diagnostic performance. Therefore, how to reduce the transfer between all source distributions and target distributions to align multiple domains is still a challenge.

[0003] For the problem of fault diagnosis under variable working conditions, scholars have proposed a domain adversarial technology transfer fault diagnosis model to solve the problem of poor generalization ability caused by the difference in fault feature distribution under variable working conditions. The main idea is to construct a domain adversarial network composed of a feature extractor, a classifier and a domain discriminator, to train the feature extractor and the classifier using sample data of source and target domains, to train the feature extractor and the domain discriminator using an adversarial training strategy, and to train the softmax classifier using labeled source domain sample data, to learn domain-invariant fault features and obtain a better fault diagnosis model, and to improve the generalization performance of the model.

[0004] Han et al. in their published paper "A novel adversarial learning framework in deep convolutional neural network for intelligent diagnosis of mechanical faults," (Knowl.-Based Syst. 165 (2019) 474-487) proposed a method based on deep convolutional neural network and domain adversarial for transfer learning for rotating machinery fault diagnosis under variable working conditions. The steps of the method are: first, collect the vibration data of the planetary gearbox under different working conditions; second, use the data under one working condition as the source domain and the data under other working conditions as the target domain, train a basic CNN to initialize the parameters, pass the obtained parameters to the feature descriptor and the discriminative classifier, and train the DAN model using the adversarial strategy to obtain the domain-invariant features; finally, use the labeled source domain data to train the Softmax classifier to achieve good fault diagnosis ability in the target domain.

[0005] The disadvantage of the above diagnosis method is that only the global domain alignment between different domains is considered, without considering the fine-grained alignment of sub-classes in the global domain, which cannot guarantee the separation between sub-classes while aligning the global domain, and cannot guarantee the extraction of domain-invariant fault features that separate different classes in the global domain. Due to the insufficient domain alignment and domain adaptation ability, the phenomenon of confusion of fault classes in the classification result still occurs, which further leads to insufficient fault classification accuracy of the model and poor generalization ability in different transfer learning tasks under complex working conditions. SUMMARY

[0006] The purpose of the present application is to solve the problem of the existing domain adversarial fault diagnosis technology for variable working condition fault diagnosis. Although the domain adversarial fault diagnosis technology can align the fault feature distribution of the source domain and the target same working condition, it mainly aligns the global distribution from the global perspective, cannot guarantee the alignment of the domain class distribution while aligning the global domain, causes the phenomenon of confusion of fault classes in the final fault diagnosis classification, and further leads to the problem of low diagnosis accuracy of the model. The present application provides a variable working condition mechanical fault diagnosis method based on multi-source domain class separation adversarial network.

[0007] In order to solve the above-mentioned problems existing in the prior art, the present application provides the following technical solutions:

[0008] The variable working condition mechanical fault diagnosis method based on multi-source domain class separation adversarial network has the speciality that it comprises the following steps:

[0009] Step 1, setting a multi-source domain transfer fault diagnosis task;

[0010] Step 1.1, collect the vibration frequency domain signals of the bearing under M working conditions as samples, and the fault types of all samples under each working condition are divided into N types; M≥3, N≥12;

[0011] Step 1.2, set M multi-source domain migration fault diagnosis tasks;

[0012] Each multi-source domain migration fault diagnosis task takes the samples of one of the M working conditions as the target domain sample set D T , and the samples of the remaining M-1 working conditions as the source domain sample set D k is the source domain sample set number, k=1, ……M-1; all samples of the target domain sample set D T are unmarked fault types, and all samples of the source domain sample set D are marked with fault types, and the fault types are divided into N types; the target domain sample set D T of different multi-source domain migration fault diagnosis tasks is different; each multi-source domain migration fault diagnosis task takes all samples of the source domain sample set D , part of the target domain sample set D T as the training set, and the remaining samples of the target domain sample set D T as the test set;

[0013] Step 2, for each multi-source domain migration fault diagnosis task set in step 1, steps 2.1 to 2.6 are executed in turn to obtain the fault diagnosis result of each multi-source domain migration fault diagnosis task;

[0014] Step 2.1, construct a feature extractor G f , and input the training set into the feature extractor G f to extract deep features F;

[0015] Step 2.2, construct a fault classifier, and input the deep features F extracted in step 2.1 into the fault classifier to obtain fault prediction probability P and predicted fault type

[0016] Step 2.3, construct a multi-source domain discriminator G d , input the deep features F extracted in step 2.1 after gradient flipping into the multi-source domain discriminator G d to obtain the domain prediction label d i ;

[0017] Step 2.4, construct a class separation adversarial discriminator According to the fault prediction probability P and the predicted fault type obtained in step 2.2, the deep features F extracted in step 2.1 are weighted and gradient flipped, and input into the class separation adversarial discriminator corresponding to the predicted fault class Obtain the class separation domain prediction label cd i ; It is a c-th type of adversarial discriminant used for inter-class separation, where c = 1, 2, ..., N;

[0018] Step 2.5: Train the multi-source domain class separation adversarial network;

[0019] Set the initial learning rate η0 and the number of iterations, and input the training set into the feature extractor module G. f The deep features F are obtained and then sequentially input into the fault classifier and the multi-source domain discriminator G. d and class separation adversarial discriminant Obtain the fault prediction probability P and the predicted fault type from the deep feature F. Combined with the domain prediction label d obtained in step 2.3 i Step 2.4 Obtain the class separation domain prediction label cd i Calculate the multi-source domain classification loss L respectively. y Multi-source domain discrimination loss L D Class separation discriminant loss L cs Finally, the total loss function L is calculated;

[0020] The gradient optimization algorithm and adaptive step-size learning rate η are used to minimize the loss, and the feature extractor G is updated iteratively. f Fault classifier, multi-source domain discriminator G d and class separation adversarial discriminant The parameters are adjusted until the maximum number of iterations is reached, thus completing the training of the multi-source domain class separation adversarial network.

[0021] Step 2.6: Perform bearing fault diagnosis on the target domain operating conditions;

[0022] Input the test set into the multi-source domain class separation adversarial network trained in step 2.5, and output the fault category prediction result for each sample; select the maximum value from the fault category prediction results for each sample as the fault diagnosis category for that sample, and output the predicted fault type.

[0023] Further, in step 2.1, the construction of the feature extractor G f Specifically, a 16-layer convolutional neural network consisting of four identical sub-modules stacked and connected together is constructed as the feature extractor G. f Each sub-module consists of a convolutional layer, a batch normalization layer, a non-linear activation function layer, and a pooling layer.

[0024] The parameter settings for each submodule are as follows:

[0025] The input dimension is 1024, the size of the first layer of convolution layer is 32*1, the size of the BN layer is 3, and the size of the pooling layer is 2; the size of the second layer of convolution layer is 32*1, the size of the BN layer is 5, and the size of the pooling layer is 3; the size of the third layer of convolution layer is 32*1, the size of the BN layer is 8, and the size of the pooling layer is 4; the size of the fourth layer of convolution layer is 16*1, the size of the BN layer is 10, and the size of the pooling layer is 4; ReLU is used as a nonlinear activation function in the output convolution layer of each layer.

[0026] Further, the step 2.2 is specifically:

[0027] Step 2.2.1, a two-layer fully connected network and a Softmax activation function layer are constructed as a fault classifier, and the nodes of the fully connected network are set to: 40-20-N.

[0028] Step 2.2.2, the deep features F extracted in step 2.1 are input into the fault classifier to obtain fault prediction probability P and predicted fault type

[0029] The calculation formula of the fault prediction probability P is as follows:

[0030]

[0031] Wherein, x i represents a sample, y i represents a labeled fault type, and θ=[θ1, θ2, … θ C ] T represents the weight parameters of the fully connected layer, represents the weight parameters of the fully connected layer connected to the jth neuron of the last layer, ∑ represents the accumulation operation, C represents that the fault classifier is divided into C classes, and C=N.

[0032] Further, in step 2.3, the multi-source domain discriminator G d is constructed. d The structure parameters of the multi-source domain discriminator G d are set to: 40-20-M.

[0033] Further, in step 2.4, the class separation adversarial discriminator is constructed. The structure parameters of the class separation adversarial discriminator are set to: 40-20-M.

[0034] Further, the step 2.5 is specifically:

[0035] Step 2.5.1, set the initial learning rate η0 is 10 -3 , and the number of iterations is 200;

[0036] Step 2.5.2, input the training set into the feature extractor module G f to obtain the deep feature F, and then input the deep feature F into the fault classifier, the multi-source domain discriminator G d and the class separation discriminator in turn, respectively to obtain the fault prediction probability P and the predicted fault type of the deep feature F

[0037] Step 2.5.3, according to the prediction probability P of the feature obtained in step 2.5.2, the domain prediction label d i obtained in step 2.3, and the class separation domain prediction label cd i obtained in step 2.4, respectively calculate the multi-source domain classification loss L y , the multi-source domain discrimination loss L D and the class separation discrimination loss L cs ;

[0038] The calculation formula of the multi-source domain classification loss L y is as follows:

[0039]

[0040] Wherein, k represents the kth source domain, is the number of samples of the kth source domain, K = M-1; θ f represents the network weight parameter of the feature extractor G f ;

[0041] The calculation formula of the multi-source domain discrimination loss L D is as follows:

[0042]

[0043] Wherein, n t represents the number of samples of the target domain belonging to the training set, L d represents the multi-source domain discriminator G d loss, G f (x i ) represents the deep feature of the sample (x f ) extracted by the feature extractor G i , m represents the total number of samples of the training set, represents the gradient inversion layer of the deep feature F; θ d represents the network weight parameter of the multi-source domain discriminator G d , θd T denotes the transpose of θ d .

[0044] The calculation formula of the class separation discriminant loss L cs is as follows:

[0045]

[0046] wherein, is the cross-entropy loss of the c-th class separation adversarial discriminator, denotes the probability that the sample belongs to the c-th class, denotes the probability that the sample does not belong to the c-th class; denotes the network weight parameter of the c-th class separation adversarial discriminator;

[0047] Step 2.5.4, according to the multi-source domain classification loss L y , the multi-source domain discriminant loss L D and the class separation discriminant loss L cs obtained in step 2.5.3, the total loss function L is calculated in the following manner:

[0048]

[0049] wherein θ f is the parameter of the feature extractor G f , θ c is the parameter of the fault classifier, and α is the weighting parameter between the classification loss of the fault classifier and the adversarial loss of the multi-source domain adversarial discriminator and the class separation adversarial discriminator, α = 1, θ d is the parameter of the multi-source domain discriminator G d , is the parameter of the class separation adversarial discriminator ;

[0050] Step 2.5.5, the parameters of the feature extractor G f , the fault classifier, the multi-source domain discriminator G d and the class separation adversarial discriminator are updated iteratively in turn to minimize the loss by using the Adam optimization algorithm and the adaptive step learning rate η until the maximum number of iterations, i.e., the training of the multi-source domain class separation adversarial network is completed;

[0051] The calculation formula of the adaptive step learning rate η is as follows:

[0052]

[0053] wherein epoch is the maximum number of iterations of the model, and z represents the z-th iteration;

[0054] update the parameters of the feature extractor module G f , the fault classifier, the multi-source domain discriminator module G d , and the class separation adversarial discriminator module :

[0055]

[0056]

[0057]

[0058]

[0059] wherein θ d is the parameter of the multi-source domain discriminator G d , denotes the partial derivative.

[0060] Compared with the prior art, the present application has the following beneficial effects:

[0061] (1) The variable working condition mechanical fault diagnosis method based on a multi-source domain class separation adversarial network of the present application constructs M domains by collecting data of bearings under M working conditions as samples, and sets M multi-source domain transfer fault diagnosis tasks according to the differences between source domains and target domains; further, the present application constructs a multi-source domain class separation adversarial network for realizing fault diagnosis of each multi-source domain transfer fault diagnosis task, which can realize fine-grained alignment of sub-classes in each domain after cross-domain and effectively learn fault diagnosis knowledge of multiple domains.

[0062] (2) The multi-source domain class separation adversarial network constructed by the present application considers alignment within the same fault type and separation across domains between different fault types, and can extract fault features with intra-class compactness and inter-class separation across domains through multi-domain adversarial learning and class separation adversarial learning strategies.

[0063] (3) The present application obtains fault diagnosis results for each multi-source domain transfer fault diagnosis task through the multi-source domain class separation adversarial network, which shows strong robustness and superiority under different tasks. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is the flowchart of the variable working condition mechanical fault diagnosis method based on a multi-source domain class separation adversarial network of the present application;

[0065] Figure 2 is the vibration time domain signal waveform graph of the bearing with normal fault type in step 1.1 of the embodiment of the present application;

[0066] Figure 3The vibration time domain signal waveform diagram of the bearing whose fault type is ball fault and whose fault diameter is 0.007 inch in step 1.1 of the embodiment of the present application;

[0067] Figure 4 The vibration time domain signal waveform diagram of the bearing whose fault type is ball fault and whose fault diameter is 0.014 inch in step 1.1 of the embodiment of the present application;

[0068] Figure 5 The vibration time domain signal waveform diagram of the bearing whose fault type is ball fault and whose fault diameter is 0.021 inch in step 1.1 of the embodiment of the present application;

[0069] Figure 6 The vibration time domain signal waveform diagram of the bearing whose fault type is inner ring fault and whose fault diameter is 0.007 inch in step 1.1 of the embodiment of the present application;

[0070] Figure 7 The vibration time domain signal waveform diagram of the bearing whose fault type is inner ring fault and whose fault diameter is 0.014 inch in step 1.1 of the embodiment of the present application;

[0071] Figure 8 The vibration time domain signal waveform diagram of the bearing whose fault type is inner ring fault and whose fault diameter is 0.021 inch in step 1.1 of the embodiment of the present application;

[0072] Figure 9 The vibration time domain signal waveform diagram of the bearing whose fault type is outer ring fault, whose fault diameter is 0.007 inch and whose fault direction is vertical 3 o'clock direction in step 1.1 of the embodiment of the present application;

[0073] Figure 10 The vibration time domain signal waveform diagram of the bearing whose fault type is outer ring fault, whose fault diameter is 0.007 inch and whose fault direction is horizontal 6 o'clock direction in step 1.1 of the embodiment of the present application;

[0074] Figure 11 The vibration time domain signal waveform diagram of the bearing whose fault type is outer ring fault, whose fault diameter is 0.014 inch and whose fault direction is horizontal 6 o'clock direction in step 1.1 of the embodiment of the present application;

[0075] Figure 12 The vibration time domain signal waveform diagram of the bearing whose fault type is outer ring fault, whose fault diameter is 0.021 inch and whose fault direction is vertical 3 o'clock direction in step 1.1 of the embodiment of the present application;

[0076] Figure 13 The vibration time domain signal waveform diagram of the bearing whose fault type is outer ring fault, whose fault diameter is 0.021 inch and whose fault direction is horizontal 6 o'clock direction in step 1.1 of the embodiment of the present application;

[0077] Figure 14 A comparison chart of the fault diagnosis results of the embodiment of the present application and six prior arts. DETAILED DESCRIPTION

[0078] The present application will be further described below in conjunction with the accompanying drawings and exemplary embodiments.

[0079] The hardware platform of the simulation experiment of the present embodiment is: the central processing unit is Intel(R) Core(TM) i5-7500 CPU, the main frequency is 3.40 GHZ, and the memory is 16 G.

[0080] The software platform of the simulation experiment of the present embodiment is: WINDOWS 7 operating system and Python 3.7.

[0081] The present embodiment adopts the bearing data set of Case Western Reserve University (CWRU) laboratory. The experimental device of the CWRU laboratory mainly consists of an induction motor, a test bearing, a loading motor, and a test bearing support motor shaft. The motor bearing is implanted through electric spark machining.

[0082] Referring to Figure 1 , the variable working condition mechanical fault diagnosis method based on multi-source domain class separation adversarial network includes the following steps:

[0083] Step 1, constructing a source domain sample set and a target domain sample set D T ;

[0084] Step 1.1, collecting vibration time domain signals of bearings under M working conditions and converting them into vibration frequency domain signals as samples. The fault types of all samples under each working condition are divided into N types; M = 4; N = 12;

[0085] The four working conditions are as follows: 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm, denoted as HP0, HP1, HP2, and HP3.

[0086] The 12 fault types are shown in Table 1:

[0087] Table 1

[0088]

[0089]

[0090] Figures 2 to 13The vibration time domain signal waveforms corresponding to 12 kinds of bearing faults are further described, wherein the vertical coordinate represents the amplitude of the vibration signal, and the horizontal coordinate represents time; Figure 2 It is indicated that the fault type of the bearing is normal. Figure 3 It is indicated that the fault type of the bearing is a ball fault, and the fault diameter is 0.007 inches. Figure 4 It is indicated that the fault type of the bearing is a ball fault, and the fault diameter is 0.014 inches. Figure 5 It is indicated that the fault type of the bearing is a ball fault, and the fault diameter is 0.021 inches. Figure 6 It is indicated that the fault type of the bearing is an inner ring fault, and the fault diameter is 0.007 inches. Figure 7 It is indicated that the fault type of the bearing is an inner ring fault, and the fault diameter is 0.014 inches. Figure 8 It is indicated that the fault type of the bearing is an inner ring fault, and the fault diameter is 0.021 inches. Figure 9 It is indicated that the fault type of the bearing is an outer ring fault, the fault diameter is 0.007 inches, and the fault orientation is vertical 3 o'clock direction. Figure 10 It is indicated that the fault type of the bearing is an outer ring fault, the fault diameter is 0.007 inches, and the fault orientation is horizontal 6 o'clock direction. Figure 11 It is indicated that the fault type of the bearing is an outer ring fault, the fault diameter is 0.014 inches, and the fault orientation is horizontal 6 o'clock direction. Figure 12 It is indicated that the fault type of the bearing is an outer ring fault, the fault diameter is 0.021 inches, and the fault orientation is vertical 3 o'clock direction. Figure 13 It is indicated that the fault type of the bearing is an outer ring fault, the fault diameter is 0.021 inches, and the fault orientation is horizontal 6 o'clock direction.

[0091] Step 1.2, referring to Table 2, M kinds of multi-source domain migration fault diagnosis tasks are set.

[0092] Each multi-source domain migration fault diagnosis task takes the samples of one of the M kinds of working conditions as the target domain sample set D T , and the samples of the remaining M-1 kinds of working conditions as the source domain sample set k is the source domain sample set sequence number, k = 1, …, M-1; all samples of the target domain sample set D T are unmarked for fault types, and all samples of the source domain sample set are marked for fault types, and the fault types are divided into N kinds; the target domain sample sets D T of different multi-source domain migration fault diagnosis tasks are different.

[0093] Each multi-source domain migration fault diagnosis task takes all samples of the source domain sample set and the target domain sample set D Tpart of the samples of the source domain as the training set, and the remaining samples of the target domain as the test set. T f ;

[0094] Table 2

[0095] Task Source domain Target domain T1 Working condition 0, working condition 1, working condition 2 Working condition 3 T2 Working condition 1, working condition 2, working condition 3 Working condition 0 T3 Working condition 0, working condition 2, working condition 3 Working condition 1 T4 Working condition 0, working condition 1, working condition 3 Working condition 2

[0096] Step 2, for each multi-source domain migration fault diagnosis task, steps 2.1 to 2.6 are sequentially executed to obtain the fault diagnosis result of each multi-source domain migration fault diagnosis task.

[0097] Step 2.1, constructing a feature extractor G f ;

[0098] Step 2.1.1, constructing a 16-layer convolutional neural network stacked by four sub-modules with the same structure as the feature extractor G f , and each sub-module is composed of a convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a pooling layer.

[0099] The parameter settings of each sub-module are as follows:

[0100] The input dimension is 1024, the size of the first layer of convolutional layer is 32x1, the size of the BN layer is 3, and the size of the pooling layer is 2; the size of the second layer of convolutional layer is 32x1, the size of the BN layer is 5, and the size of the pooling layer is 3; the size of the third layer of convolutional layer is 32x1, the size of the BN layer is 8, and the size of the pooling layer is 4; the size of the fourth layer of convolutional layer is 16x1, the size of the BN layer is 10, and the size of the pooling layer is 4; ReLU is used as the nonlinear activation function in the output convolutional layer of each layer.

[0101] Step 2.1.2, inputting the training set into the feature extractor G f to extract the deep features F;

[0102] Step 2.2, constructing a fault classifier;

[0103] Step 2.2.1, constructing a fault classifier by stacking a 2-layer fully connected network and a Softmax activation function layer, and setting the nodes of the fully connected network as: 40-20-N.

[0104] Step 2.2.2, inputting the deep features F extracted in step 2.1 into the fault classifier to obtain the fault prediction probability P and the predicted fault type

[0105] The calculation formula of the fault prediction probability P is as follows:

[0106]

[0107] where xi y represents a sample, i θ represents a labeled fault type, θ = [θ1, θ2, … θN] represents a fault type, C ] T W represents the weight parameters of the fully connected layer, W represents the weight parameters of the fully connected layer connected to the jth neuron of the last layer, ∑ represents the accumulation operation, C represents that the fault classifier is divided into C classes, C = N;

[0108] Step 2.3, constructing a multi-source domain discriminator G d ;

[0109] Step 2.3.1, constructing a multi-source domain discriminator G d with 2 layers of fully connected network and Softmax activation function layer, the structure parameters of the multi-source domain discriminator G d are set as: 40-20-M;

[0110] Step 2.3.2, after gradient flipping of the deep features F extracted in step 2.1, input the multi-source domain discriminator G d to obtain the domain prediction label d i ;

[0111] Step 2.4, constructing class separation adversarial discriminators

[0112] Step 2.4.1, constructing C class separation adversarial discriminators with 2 layers of fully connected network and Softmax activation function layer, the structure parameters of the class separation adversarial discriminators are set as: 40-20-M; is the cth class separation adversarial discriminator for class separation, c = 1, 2 … N;

[0113] Step 2.4.2, according to the fault prediction probability P and the predicted fault type obtained in step 2.2, the deep features F extracted in step 2.1 are weighted and gradient flipped, and input into the class separation adversarial discriminator corresponding to the predicted fault class to obtain the class separation domain prediction label cd i ;

[0114] Step 2.5, training the multi-source domain class separation adversarial network;

[0115] Step 2.5.1, setting the initial learning rate η0 is 10 -3 , and the iteration number is 200;

[0116] Step 2.5.2, inputting the training set into the feature extractor module G fObtain the deep feature F, and then input the fault classifier and the multi-source domain discriminator G in sequence respectively d and the class separation adversarial discriminator Obtain the fault prediction probability P and the predicted fault type of the deep feature F

[0117] Step 2.5.3, according to the prediction probability P of the feature obtained in step 2.5.2, the domain prediction label d obtained in step 2.3.2 i , the class separation domain prediction label cd obtained in step 2.4 i , respectively calculate the multi-source domain classification loss L y , the multi-source domain discrimination loss L D and the class separation discrimination loss L cs ;

[0118] The calculation formula of the multi-source domain classification loss L y is as follows:

[0119]

[0120] Wherein, k represents the kth source domain, is the sample number of the kth source domain, K=M-1; θ f represents the network weight parameter of the feature extractor G f ;

[0121] The calculation formula of the multi-source domain discrimination loss L D is as follows:

[0122]

[0123] Wherein, n t represents the sample number of the target domain belonging to the training set, L d represents the multi-source domain discriminator G d loss, G f (x i ) represents the deep feature of the sample (x f ) extracted by the feature extractor G i , and m represents the total sample number of the training set, represents the gradient inversion layer of the deep feature F; θ d represents the network weight parameter of the multi-source domain discriminator G d , θ d T represents the transpose of θ d ;

[0124] The calculation formula of the class separation discrimination loss L cs is as follows:

[0125]

[0126] wherein, is the cross-entropy loss of the c-th class class-wise disentangled adversarial discriminator, denotes the probability that the sample belongs to the c-th class, denotes the probability that the sample does not belong to the c-th class; denotes the network weight parameters of the c-th class class-wise disentangled adversarial discriminator;

[0127] Step 2.5.4. Calculate the total loss function L according to the multi-source domain classification loss L y , the multi-source domain discrimination loss L D and the class-wise disentangled discrimination loss L cs , in the following way:

[0128]

[0129] wherein θ f are the parameters of the feature extractor G f , θ c are the parameters of the fault classifier, and α is the weighting parameter between the classification loss of the fault classifier and the adversarial loss of the multi-source domain adversarial discriminator and the class-wise disentangled adversarial discriminator, α = 1, θ d are the parameters of the multi-source domain discriminator G d , are the parameters of the class-wise disentangled adversarial discriminator ;

[0130] Step 2.5.5. Update the parameters of the feature extractor G f , the fault classifier, the multi-source domain discriminator G d and the class-wise disentangled adversarial discriminator in turn by using the Adam optimization algorithm and the adaptive step learning rate η until the maximum number of iterations is reached, i.e. the training of the multi-source domain class-wise disentangled adversarial network is completed;

[0131] The calculation formula of the adaptive step learning rate η is as follows:

[0132]

[0133] wherein epoch is the maximum number of iterations of the model, and z represents the z-th iteration;

[0134] Update the parameters of the feature extractor module G f , the fault classifier, the multi-source domain discriminator module G d and the class-wise disentangled adversarial discriminator module in the following way:

[0135]

[0136]

[0137]

[0138]

[0139] wherein, θ d is a parameter of the multi-source domain discriminator G d , denotes partial derivative;

[0140] Step 2.6, bearing fault diagnosis is performed on the target domain working condition;

[0141] Step 2.6.1, the test set is input into the multi-source domain class separation adversarial network trained in step 2.5, and a fault class prediction result of each sample is output;

[0142] Step 2.6.2, a maximum value is selected from the fault class prediction result of each sample as a fault diagnosis class of the sample, and a predicted fault type is output

[0143] To verify the performance of the present application, based on the settings in step 1 of the embodiments of the present application, the present application is compared with the fault diagnosis results of 6 prior arts, as shown in Table 3 and Figure 14 The 6 prior arts are respectively a transfer learning method based on TCA, a transfer learning method based on JDA, a transfer learning method based on BDA, a transfer learning method based on DDC, a transfer learning method based on D-CORAL and a transfer learning method based on DAN.

[0144] Table 3

[0145] Task T1 T2 T3 T4 Average The present invention 0.9933 0.9917 0.9967 0.9983 0.9949 TCA 0.7433 0.7283 0.7450 0.6800 0.7200 JDA 0.7467 0.7550 0.7383 0.7183 0.7396 BDA 0.7233 0.7717 0.8333 0.8333 0.7904 DDC 0.9792 0.8858 0.9692 0.9725 0.9517 D-CORAL 0.9758 0.9133 0.9 0.9808 0.9635 DAN 0.9767 0.9508 0.9808 0.9750 0.9708

[0146] Among them, the transfer learning method based on JDA refers to the transfer learning method proposed by M. Long et al. in "Transfer feature learning with joint distribution adaptation, in: 2013 IEEE Int. Conf. Comput. Vis., IEEE, Sydney, Australia, 2013: pp. 2200-2207"; the transfer learning method based on BDA refers to the transfer learning method proposed by J. Wang et al. in "Balanced distribution adaptation for transfer learning, in: 2017 IEEE Int. Conf. Data Min. ICDM, IEEE, New Orleans, LA, 2017: pp. 1129-1134"; the transfer learning method based on DDC refers to the transfer learning method proposed by E. Tzeng et al. in "Deep coral: correlation alignment for deep domain adaptation, in: Eur. Conf. Comput. Vis., Springer, 2016: pp. 443-450"; the transfer learning method based on D-CORAL refers to the transfer learning method proposed by B. Sun et al. in "A new deep transfer learning method for bearing fault diagnosis under different working conditions, IEEE Sens. J. 20 (2020) 8394-8402"; and the transfer learning method based on DAN refers to the transfer learning method proposed by T. Han et al. in "A novel adversarial learning framework in deep convolutional neural network for intelligent diagnosis of mechanical faults, Knowl.-Based Syst. 165 (2019) 474-487".

[0147] As can be seen from Table 3, the fault diagnosis accuracy of the embodiment of the present application averages 0.9949, each task is higher than 0.99, and is significantly higher than other methods, which are 0.7200, 0.7396, 0.7904, 0.9517, 0.9635 and 0.9754 respectively. Compared with the traditional transfer learning methods TCA, JDA and BDA, the fault diagnosis accuracy of the embodiment of the present application is increased by 0.2749, 0.2553 and 0.2049 respectively, indicating that the embodiment of the present application has good diagnostic performance. In addition, the average fault diagnosis accuracy of other transfer learning methods based on DDC and D-CORAL is lower than that of the embodiment of the present application by 0.0414 and 0.0314 respectively, which further indicates the effectiveness of the embodiment of the present application. Similarly, the average fault diagnosis accuracy of DAN is lower than that of the embodiment of the present application by 0.0241, which shows that the embodiment of the present application has better class separation capability. Therefore, through the above comparative analysis, the effectiveness of the embodiment of the present application is verified.

[0148] The fault diagnosis accuracy is equal to the number of samples in the test set whose fault type is predicted correctly divided by the total number of samples in the test set.

[0149] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit it. For ordinary skilled in the art, the specific technical solutions recorded in the above embodiments can be modified, or some technical features can be replaced by equivalent, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions protected by the present application.

Claims

1. A variable working condition mechanical fault diagnosis method based on multi-source domain class separation adversarial network, characterized in that, The steps comprise the following: Step 1, setting a multi-source domain migration fault diagnosis task; Step 1.1, collecting vibration frequency domain signals of bearings under M working conditions as samples, and the fault types of all samples under each working condition are divided into N types; M≥3, N≥12; Step 1.2, setting M multi-source domain migration fault diagnosis tasks; Each multi-source domain migration fault diagnosis task takes samples of one of M working conditions as a target domain sample set D T , and samples of the remaining M-1 working conditions as a source domain sample set k is the ordinal number of the source domain sample set, k = 1, …, M-1; all samples of the target domain sample set D T are unmarked for fault types, and all samples of the source domain sample set are marked for fault types, and the fault types are divided into N types; the target domain sample set D T is different for different multi-source domain migration fault diagnosis tasks; each multi-source domain migration fault diagnosis task takes all samples of the source domain sample set and part of the samples of the target domain sample set D T as a training set, and takes the remaining samples of the target domain sample set D T as a test set; Step 2, for each multi-source domain migration fault diagnosis task set in step 1, steps 2.1 to 2.6 are executed in turn to obtain the fault diagnosis result of each multi-source domain migration fault diagnosis task; Step 2.1, constructing a 16-layer convolutional neural network stacked by four sub-modules with the same structure as the feature extractor G f , wherein the structure of each sub-module is composed of a convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a pooling layer, and inputting the training set into the feature extractor G f to extract the deep features F; Step 2.2, construct a 2-layer fully connected network with Softmax activation function layer as the fault classifier, and input the deep features F extracted in step 2.1 into the fault classifier to obtain the fault prediction probability P and the predicted fault type Step 2.

3. Constructing a two-layer fully connected network with a Softmax activation function layer as the multi-source domain discriminator G d After gradient flipping of the deep features F extracted in step 2.1, input the multi-source domain discriminator G d Obtain the domain prediction label d i ; Step 2.

4. Constructing C class-separation adversarial discriminators with 2 layers of fully connected networks and Softmax activation function layer as class separation obtaining the failure prediction probability P and the predicted failure type according to step 2.2 performing weighted processing on the deep features F extracted in step 2.1 and gradient flipping, and inputting into the class-separation adversarial discriminator corresponding to the predicted failure class obtaining the class-separation domain prediction label cd i ; is the c-th class-separation adversarial discriminator for class separation, c = 1, 2, …, N. Step 2.5, training a multi-source domain class separation adversarial network; Set initial learning rate η0, iteration number, input training set into feature extractor module G f Obtain deep feature F, and then input fault classifier and multi-source domain discriminator G d And class separation adversarial discriminator Obtain fault prediction probability P and predicted fault type of deep feature F Combine the domain prediction label d obtained in step 2.3 i , the class separation domain prediction label cd obtained in step 2.4 i , respectively calculate multi-source domain classification loss L y , multi-source domain discrimination loss L D And class separation discrimination loss L cs Finally, calculate the total loss function L; The gradient optimization algorithm and the adaptive step size learning rate η are used to minimize the loss, and the feature extractor G is iteratively updated in turn f , the fault classifier, the multi-source domain discriminator G d and the class separation adversarial discriminator parameters until the maximum number of iterations is reached, that is, the training of the multi-source domain class separation adversarial network is completed; Step 2.6, bearing fault diagnosis under target domain working condition; The test set is input into the multi-source domain class separation adversarial network trained in step 2.5, and the fault class prediction result of each sample is output; the maximum value is selected from the fault class prediction result of each sample as the fault diagnosis class of the sample, and the predicted fault type is output 2. The method of claim 1, wherein the method is based on a multi-source domain class separation adversarial network. In step 2.1, the parameter settings of each sub-module are as follows: The input dimension is 1024, the size of the first convolution layer is 32×1, the size of the BN layer is 3, and the size of the pooling layer is 2; the size of the second convolution layer is 32×1, the size of the BN layer is 5, and the size of the pooling layer is 3; the size of the third convolution layer is 32×1, the size of the BN layer is 8, and the size of the pooling layer is 4; the size of the fourth convolution layer is 16×1, the size of the BN layer is 10, and the size of the pooling layer is 4; ReLU is used as a nonlinear activation function in the output convolution layer of each layer.

3. The multi-source domain class separation adversarial network-based mechanical fault diagnosis method under varying working conditions according to claim 2, characterized in that, The step 2.2 is specifically as follows: Step 2.2.1, constructing a fault classifier composed of a 2-layer fully connected network and a Softmax activation function layer, and the node settings of the fully connected network are: 40-20-N; Step 2.2.2, input the extracted deep features F in step 2.1 into the fault classifier to obtain a fault prediction probability P and a predicted fault type The calculation formula of the fault prediction probability P is as follows: where x i represents the sample, y i represents the labeled fault type, θ = [θ1, θ2, … θ C ] T represents the weight parameters of the fully connected layer, represents the weight parameters of the fully connected layer connected to the jth neuron of the last layer, ∑ represents the accumulation operation, C represents that the fault classifier is divided into C classes, C = N.

4. The method of claim 3, wherein the method is based on a multi-source domain class separation adversarial network. In step 2.3, the multi-source domain discriminator G d The structure parameter is set as: 40-20-M.

5. The method of claim 4, wherein the multi-source domain class separation adversarial network-based mechanical fault diagnosis method is characterized in that: In step 2.4, the class-separated adversarial discriminator The structure parameter is set as: 40-20-M.

6. The multi-source domain class separation adversarial network-based mechanical fault diagnosis method under varying working conditions according to claim 5, characterized in that, The step 2.5 is specifically as follows: Step 2.5.1, set initial learning rate η0is 10 -3 number of iterations is 200; Step 2.5.2, input the training set into the feature extractor module G f Obtain the deep feature F, and then input it into the fault classifier, the multi-source domain discriminator G d and the class separation adversarial discriminator in turn Obtain the fault prediction probability P and the predicted fault type of the deep feature F Step 2.5.3, the predicted probability P of the feature obtained according to step 2.5.2, the domain prediction label d obtained in step 2.3 i , the class separation domain prediction label cd obtained in step 2.4 i , respectively, calculate the multi-source domain classification loss L y , the multi-source domain discrimination loss L D and the class separation discrimination loss L cs ; The multi-source domain classification loss L y The calculation formula is as follows: wherein k represents the kth source domain, is the number of samples of the kth source domain, K = M - 1; θ f represents the network weight parameters of the feature extractor G f . The multi-source domain discriminative loss L D The calculation formula is as follows: wherein n t represents the number of samples in the target domain belonging to the training set, L d represents the multi-source domain discriminator G d loss, G f (x i ) represents the deep feature of the sample (x i ) extracted by the feature extractor G f , and m represents the total number of samples in the training set, represents the gradient inversion layer of the deep feature F; θ d represents the network weight parameters of the multi-source domain discriminator G d , and θ d T represents the transpose of θ d . The class separation discriminative loss L cs The calculation formula is as follows: wherein, is a cross-entropy loss of the c-th class-specific counterfactual discriminator, denotes a probability that the sample belongs to the c-th class, denotes a probability that the sample does not belong to the c-th class; denotes network weight parameters of the c-th class-specific counterfactual discriminator; Step 2.5.

4. Obtain the multi-source domain classification loss L according to the multi-source domain classification loss L obtained in step 2.5.

3. y , the multi-source domain discrimination loss L and the class separation discrimination loss L cs , the total loss function L is calculated as follows: where θ f is the parameter of the feature extractor G f , θ c is the parameter of the fault classifier, and α is the trade-off parameter between the classification loss of the fault classifier and the adversarial loss of the multi-source domain adversarial discriminator and the class separation adversarial discriminator, α = 1, θ d is the parameter of the multi-source domain discriminator G d , is the parameter of the class separation adversarial discriminator ; Step 2.5.5, iteratively update the feature extractor G with Adam optimization algorithm and adaptive step size learning rate η to minimize the loss f , the fault classifier, the multi-source domain discriminator G d and the class separation adversarial discriminator parameters until the maximum number of iterations, that is, the training of the multi-source domain class separation adversarial network is completed; The calculation formula of the adaptive step learning rate η is as follows: Wherein, epoch is the maximum iteration number of the model, and z represents the zth iteration. The parameters of the feature extractor module G are updated according to the following formula f , the failure classifier, the multi-source domain discriminator module G d and the class separation adversarial discriminator module are updated according to the following formula: where θ d is a parameter of the multi-source domain discriminator G d , denotes partial derivation.

Images