A motor fault diagnosis method based on deep autoencoder and feature optimization

Abstract

The present invention proposes a transferable motor fault diagnosis method based on deep autoencoder and domain-invariant feature extraction. First, the dual-tree complex wavelet packet transform is used to process the motor vibration signal and extract hybrid domain features. Secondly, a domain-invariant feature selection method based on importance score and inter-domain difference measurement is proposed to select features with high fault resolution ability and domain-invariant characteristics. Then, the deep autoencoder (source model) is pre-trained using the selected domain-invariant features to enhance the fault characterization ability of the deep features. The parameters of the source model are transferred to a target model with the same structure, and the normal state feature data of the target domain is used to fine-tune the target model. Finally, the fine-tuned target model is used to classify the target domain faults. The method proposed by the present invention can significantly improve the accuracy of cross-domain fault diagnosis, and has stronger usability, stability and advantages in actual industrial scenarios.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and in particular to a motor fault diagnosis method based on deep autoencoder and feature optimization. Background Art

[0002] With the rapid development of modern industry, rotating machinery such as motors are developing towards integration and complexity. At the same time, the working conditions are also more complex and harsh, so the equipment failure rate is high. The failure of a certain component usually leads to the failure of other components, and even causes huge human and economic losses. Therefore, it is of great practical significance to study the intelligent fault diagnosis method of motors for actual industrial scenarios.

[0003] In recent years, research in the field of intelligent fault diagnosis generally includes frameworks based on traditional machine learning (TML), deep learning (DL), and transfer learning (TL). The framework based on TML includes three steps: signal processing and feature extraction, feature selection and dimensionality reduction, and fault classification. At present, the analysis signals of motor intelligent fault diagnosis mostly come from the vibration signals of the motor, and the time-frequency analysis method is generally used for signal processing and feature extraction. This method can effectively help extract fault features, but the resulting high-dimensional feature set often contains interference and redundant features. Therefore, feature selection and dimensionality reduction are key steps before fault mode classification. The main limitation of the TML-based framework is that when the diagnostic model performs fault diagnosis on different working conditions and machines, it relies heavily on expert knowledge. Therefore, the DL-based framework introduces a deep learning algorithm that can automatically mine hidden features, which has been widely studied in the field of intelligent fault diagnosis. However, most DL-based frameworks have limitations: the premise that the training set and the test set are distributed in the same way makes the model generalization ability insufficient; a large amount of labeled data is required. When the labeled data is insufficient, overfitting will occur during diagnosis, resulting in a decrease in diagnostic accuracy. In view of the above limitations, domain adaptation based on the TL framework has become a research hotspot as a recently developed new technology. Its goal is to use labeled source domains and unlabeled target domains for cross-domain learning to achieve ideal fault classification in the target domain. At present, deep transfer models have achieved many successful applications in cross-domain fault diagnosis of rotating machinery, but how to improve the fault characterization ability of deep features and the stability of diagnostic models under different working conditions is still a challenging task. To address this problem, this paper proposes a transferable motor fault diagnosis method (TFDD) based on deep auto-encoder (DAE) and domain-invariant feature selection under variable working conditions. It aims to use a small number of training samples to train the source model and perform transfer learning of the target model to achieve intelligent fault diagnosis of motors under variable working conditions and improve the generalization ability of the model. Summary of the invention

[0004] In order to improve the adaptability and stability of motor fault diagnosis methods in actual industrial scenarios and to cope with complex and changeable mechanical working conditions, the present invention provides a motor fault diagnosis method based on deep autoencoder and feature optimization, which satisfies the high diagnostic accuracy of the motor fault diagnosis model under the condition of less training data while also having superior cross-domain diagnostic performance.

[0005] The present invention is implemented by the following technical solution: a motor fault diagnosis method based on deep autoencoder and feature optimization, specifically as follows:

[0006] Step 1, using dual-tree complex wavelet packet transform to perform signal processing and statistical feature extraction on the original vibration signals under different working conditions, and construct an initial statistical feature set;

[0007] Step 2: evaluate the fault resolution ability and domain invariance of the features in the initial statistical feature set constructed in step 1, select domain invariant features that can characterize the motor state, and construct a domain invariant feature subset of the sample.

[0008] Step 3: Use the domain-invariant feature subset of the source domain obtained in step 2 to train the source DAE model, obtain model parameters, and transfer the parameters to the target DAE model. Use the normal state domain-invariant feature subset of the target domain obtained in step 2 to fine-tune the target model, and complete the construction of the deep transfer autoencoder model DTAE.

[0009] Step 4: test the deep transfer autoencoder model DTAE constructed in step 3 and output the fault diagnosis result.

[0010] Preferably, the specific steps of step 1 are as follows: the original vibration signal samples from different working conditions are divided into vibration data set I and vibration data set II, wherein vibration data set I is the source domain, and the samples of its data set are labeled; vibration data set II is the target domain, and the samples of its data set are unlabeled, and the diagnostic model is trained using known label samples from the source domain, and the trained model is used to perform state classification on the target domain.

[0011] Preferably, the specific process in step 2 is: utilizing the labeled fault feature data from the initial statistical feature set of the source domain, and the normal state feature data of the source domain and the target domain, and adopting a domain-invariant feature selection method based on importance score and domain invariance to evaluate the features of the initial statistical feature set, obtaining a domain-invariant feature selection index RIM, and using the sorted RIM sequence to select a domain-invariant feature subset of the initial statistical feature set from the target domain.

[0012] Preferably, for the labeled fault feature data in the source domain obtained in step 1, the importance score IS of each feature is calculated by a feature importance evaluation method based on random forest to evaluate the fault resolution ability of the feature; the inter-domain difference of a single feature in the source domain and the target domain is analyzed, and the maximum mean difference MMD is used to estimate the distribution difference MMDS of the same feature between different domains; for the IS and MMDS of each feature obtained, the ratio of IS to MMDS is calculated, a domain-invariant feature selection index RIM is constructed, and the RIM sequence is sorted in descending order; for the unlabeled feature data from the original feature set of the target domain, the sorted RIM sequence is directly used for domain-invariant feature selection.

[0013] Preferably, the specific steps of the feature importance evaluation method based on random forest are as follows:

[0014] For the N motor vibration signal samples in the vibration data set Ι (source domain), after the signal processing and statistical feature extraction process in step 1, each vibration signal sample can obtain Q kinds of statistical features, which can constitute the initial statistical feature set F = [f 1 ,f 2 ,…,f N ] T , where f i ={f i 1 ,f i 2 ,…f i Q}, i∈[1,N] represents the set of Q statistical features of the i-th vibration signal sample. The RF-based feature selection method is used to calculate the importance score (IS) of each feature in the original feature set. The main idea of ​​RF is to build a decision tree (DT). Here, it is assumed that an RF has n DT It is composed of decision trees.

[0015] Using F j and its label Y to construct the set F j is the jth statistical feature set of the vibration data set Ι, y i Represents the motor state label, C and N represent the number of categories and samples respectively. Calculate the decision tree DT constructed by S n The Gini index GI of each node of , GI(t) represents DT n The GI value at node t is used to quantify the impurity of node t and is defined as follows:

[0016]

[0017] Where, the probability of occurrence of the k-th sample of node t. Feature F j The importance score IS can be defined by the following formula:

[0018]

[0019] Where n DT It can be expressed as the number of DTs in RF, t∈DT n Represents a collection of split nodes.

[0020] Finally, in the original high-dimensional feature set, the importance score IS(j) of each feature can be calculated, and the IS sequence of Q features IS={IS(1),IS(2),…,IS(Q)} can be obtained. The larger the IS value, the greater the fault resolution ability of the feature. According to the sorted IS sequence, features with high IS values ​​can be selected to construct feature subsets, thereby eliminating many features with small IS values ​​and improving classification performance.

[0021] For the normal state feature data of vibration data set Ι (source domain) and vibration data set Ⅱ (target domain), the distribution difference of the same feature in different domains is evaluated by MMD, and the value of MMD is used as a quantitative indicator of feature domain invariance. The normal state feature data from the source domain and the target domain are defined as SF and TF, respectively, both containing N samples, and each sample contains Q statistical features after signal processing and statistical feature extraction in step 1. The expressions of SF and TF are as follows:

[0022]

[0023] Among them, the j-th column elements of SF and TF represent N samples ((j∈[1,2,…,Q])) of the j-th feature from the source domain and the target domain, respectively, and their expressions are as follows:

[0024]

[0025] SF j and TF j The empirical distance estimate between the distributions is defined as follows:

[0026]

[0027] In the formula, ||·|| H represents the RKHS norm, and φ represents the kernel-induced feature map. We can further calculate the MMD sequence of Q statistical features MMD = {MMD(1), MMD(2), …, MMD(Q)}. The smaller the value of MMD, the better the domain invariance of the feature.

[0028] Based on the IS and MMD values ​​of each feature obtained in the first two steps, a new feature selection index RIM is constructed to select domain-invariant features from RFS. The RIM of the jth feature is defined as follows:

[0029]

[0030] For Q kinds of statistical features, a RIM sequence RIM = {RIM (1), RIM (2), ..., RIM (Q)} can be obtained. The RIM sequence is sorted, and features with higher RIM are selected, and the cross-domain fault diagnosis model is trained with the RIM sequence sorted in descending order.

[0031] Preferably, the specific steps of constructing the deep transfer autoencoder model DTAE are as follows:

[0032] Step 3.1 Build the source DAE model. DAE is an unsupervised deep neural network, which is composed of several basic autoencoders stacked together. Each AE consists of an encoder and a decoder, including an input layer x = {x 1 ,x 2 ,...,x m}、Hidden layer h={h 1 ,h 2 ,...,h n}、Output layer z={z 1 ,z 2 ,...,z m};

[0033] The encoder passes through the activation function Map the input data x to the hidden layer h. The mapping process is as follows:

[0034]

[0035] Where W and b are the weight matrix and bias vector of the encoder respectively;

[0036] The decoder uses the activation function Map the data in the hidden layer h to the output data z. The mapping process is as follows:

[0037]

[0038] In the formula, W′ and b′ are the weight matrix and bias vector of the decoder respectively, the hidden layer represents the new features of the input data, and the output data is the reconstruction of the input data;

[0039] The parameters of the AE model are learned by minimizing the reconstruction error between the input layer and the output layer. The reconstruction error is defined as follows:

[0040]

[0041] Where N is the number of samples, θ = {W, b, W′, b′}, the hidden layer features of an AE are used as the input of the next AE model, and a DAE is constructed by stacking multiple AEs;

[0042] Step 3.2 constructs a target DAE model with the same structure as the source model: 1 input layer, 3 hidden layers, and 1 softmax layer. At the same time, the parameter weight matrix W learned during the source model training is (S) and the bias vector b (S) Passed to the target model, that is, W (S) =W (T) and b(S) =b (T) ;

[0043] Step 3.3 fine-tunes the target DAE model using the normal state feature data from the target domain, and uses the fine-tuned target model to test the remaining subset of unlabeled domain-invariant features from the target domain.

[0044] Preferably, in step 4, the unlabeled domain-invariant feature subset of the test set is used as input to train the classification performance of the constructed target DAE model, and finally the cross-domain diagnosis result is output, that is, the label information of the target domain is obtained.

[0045] Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0046] In order to solve the problem of redundant features in the original feature set after signal processing and statistical feature extraction of the original vibration signal samples, the present invention proposes a domain-invariant feature selection method DISD based on fault resolution and domain invariance. The method evaluates the fault discrimination ability of each feature and estimates the distribution differences of the same feature in different domains, and selects features with both high fault resolution and domain invariance, so as to improve the running speed and accuracy of subsequent diagnostic algorithms.

[0047] To address the problem of insufficient stability of diagnostic models based on traditional transfer learning methods under different working conditions, a deep transfer autoencoder (DTAE) was constructed, and the domain-invariant features selected by DISD were used for model training. This not only improves the fault characterization ability of deep features, but also improves the stability of the diagnostic model in cross-domain fault diagnosis.

[0048] While ensuring high accuracy of fault diagnosis, the availability and generalization ability of the fault diagnosis model in cross-domain fault diagnosis are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the structural block diagram of the deep migration autoencoder model;

[0050] Figure 2 It is a structural diagram of the system framework for motor cross-domain fault diagnosis of the present invention;

[0051] Figure 3 Here are the sample raw vibration signal graphs for 5 motor states;

[0052] Figure 4 It is the wavelet packet node reconstruction signal diagram obtained by decomposing the normal motor vibration signal;

[0053] Figure 5These are the 352 statistical feature graphs extracted from the NM and BF vibration signals at 1730 rpm and 1750 rpm (the feature data have been normalized);

[0054] Figure 6 It is the MMD graph of 352 statistical features in the source domain;

[0055] Figure 7 It is the IS graph of 352 statistical features in the source domain;

[0056] Figure 8 It is the RIM graph of 352 statistical features in the source domain;

[0057] Fig. 9 This is a schematic diagram of the cross-domain fault diagnosis results of the TFDD framework;

[0058] Fig.10 This is a diagram of cross-domain fault diagnosis results of the comparison model based on transfer learning;

[0059] Fig.11 It is a comparison chart of the fault diagnosis results of the comparison model. DETAILED DESCRIPTION

[0060] like Figure 1 and 2 As shown in FIG. 1 , a motor fault diagnosis method based on deep autoencoder and feature optimization includes four processes, as follows:

[0061] Process 1. Signal processing

[0062] The motor vibration signals collected at different speeds are respectively defined as the source domain and target domain required by the present invention, wherein the source domain is a labeled sample (i.e., the motor state is known), and the target domain uses an unlabeled sample. DTCWPT is used to perform signal processing on each sample, decompose it into different grouping nodes, and obtain a feature data set characterizing the motor operation state by calculating 11 statistical parameters such as extreme difference, mean, standard deviation, kurtosis, energy, energy entropy, skewness, crest factor, pulse factor, shape factor and latitude factor. Four-layer DTCWPT decomposition is performed on each vibration signal sample in the source domain and the target domain, 16 terminal wavelet packet nodes and corresponding wavelet packet coefficients are obtained, and 16 terminal nodes are reconstructed by a single wavelet packet, 16 single reconstructed signals can be obtained, and 11 statistical parameters of the reconstructed signal are calculated to generate 176 time domain statistical features. The Hilbert envelope spectrum HES of the 16 reconstructed signals is then obtained, and 176 frequency domain statistical features are generated by calculating the 11 statistical parameters of HES. By combining the time domain and frequency domain features of a vibration sample, 352 mixed domain statistical parameters of each vibration signal sample can be obtained to form the initial statistical feature set.

[0063] Process 2. Feature selection

[0064] Using the labeled fault feature data from the initial statistical feature set of the source domain and the normal state feature data of the source domain and the target domain, the DISD method is used to evaluate the features of the initial statistical feature set to obtain the domain invariant feature selection index RIM. The sorted RIM sequence is used to select the domain invariant feature subset of the initial statistical feature set from the target domain.

[0065] In order to reduce redundant features in the high-dimensional feature set (HRFS), domain-invariant features (features that can identify faults but are not affected by the operating state) are selected. The present invention quantifies the ability of a single feature to distinguish motor faults under different operating states by calculation, selects the original feature set with strong resolution, and discards the weak resolution. When selecting domain-invariant features, the fault resolution and domain invariance of the features are quantitatively analyzed as the basis for feature selection of the original feature set obtained in process 1.

[0066] For the labeled fault feature data in the source domain OSFS obtained in step 1, the random forest-based

[0067] The feature importance evaluation method of Random Forest (RF) is used to calculate the importance score (IS) of each feature to evaluate the fault resolution ability of the feature; the inter-domain differences of single features in the source domain and the target domain OSFS are analyzed, and the maximum mean discrepancy (MMD) is used to estimate the distribution difference MMDS of the same feature between different domains. The size of MMDS represents the domain invariance of the feature. For the IS and MMDS of each feature obtained, the ratio of IS to MMDS is calculated to construct the domain invariant feature selection index RIM. The RIM sequence is sorted in descending order. The higher the RIM value, the higher the fault resolution ability and domain invariance of the feature, which is more conducive to cross-domain fault diagnosis. For the unlabeled feature data from the original feature set of the target domain, the sorted RIM sequence is directly used for domain invariant feature selection.

[0068] The specific steps of the DISD method are as follows:

[0069] Step 2.1 For the N motor vibration signal samples in the vibration data set Ι (source domain), after the signal processing and statistical feature extraction process in step 1, each vibration signal sample can obtain Q kinds of statistical features, which can constitute the initial statistical feature set F = [f 1 ,f 2 ,…,f N ] T , where f i ={f i 1 ,f i 2,…f i Q}, i∈[1,N] represents the set of Q statistical features of the i-th vibration signal sample. The RF-based feature selection method is used to calculate the importance score (IS) of each feature in the original feature set. The main idea of ​​RF is to build a decision tree (DT). Here, it is assumed that an RF has n DT It is composed of decision trees.

[0070] Using F j and its label Y to construct the set F j is the jth statistical feature set of the vibration data set Ι, y i Represents the motor state label, C and N represent the number of categories and samples respectively. Calculate the decision tree DT constructed by S n The Gini index GI of each node of , GI(t) represents DT n The GI value at node t is used to quantify the impurity of node t and is defined as follows:

[0071]

[0072] Where, the probability of occurrence of the k-th sample of node t. Feature F j The importance score IS can be defined by the following formula:

[0073]

[0074] Where n DT It can be expressed as the number of DTs in RF, t∈DT n Represents a collection of split nodes.

[0075] Finally, in the original high-dimensional feature set, the importance score IS(j) of each feature can be calculated, and the IS sequence of Q features IS={IS(1),IS(2),…,IS(Q)} can be obtained. The larger the IS value, the greater the fault resolution ability of the feature. According to the sorted IS sequence, features with high IS values ​​can be selected to construct feature subsets, thereby eliminating many features with small IS values ​​and improving classification performance.

[0076] Step 2.2 For the normal state feature data of vibration data set I (source domain) and vibration data set II (target domain), the distribution difference of the same feature in different domains is evaluated by MMD, and the value of MMD is used as a quantitative indicator of feature domain invariance. The normal state feature data from the source domain and the target domain are defined as SF and TF, respectively, both containing N samples, and each sample contains Q statistical features after signal processing and statistical feature extraction in step 1. The expressions of SF and TF are as follows:

[0077]

[0078] Among them, the j-th column elements of SF and TF represent N samples ((j∈[1,2,…,Q])) of the j-th feature from the source domain and the target domain, respectively, and their expressions are as follows:

[0079]

[0080] SF j and TF j The empirical distance estimate between the distributions is defined as follows:

[0081]

[0082] In the formula, ||·|| H represents the RKHS norm, and φ represents the kernel-induced feature map. We can further calculate the MMD sequence of Q statistical features MMD = {MMD(1), MMD(2), …, MMD(Q)}. The smaller the value of MMD, the better the domain invariance of the feature.

[0083] Step 2.3: Based on the IS and MMD values ​​of each feature obtained in the first two steps, a new feature selection index RIM is constructed to select domain-invariant features from RFS. The RIM of the jth feature is defined as follows:

[0084]

[0085] For Q kinds of statistical features, a RIM sequence RIM = {RIM (1), RIM (2), ..., RIM (Q)} can be obtained. The RIM sequence is sorted, and features with higher RIM are selected, and the cross-domain fault diagnosis model is trained with the RIM sequence sorted in descending order.

[0086] Process 3. Deep Transfer Autoencoder Model Construction

[0087] The domain-invariant feature subset from the source domain selected in step 2 is used as the training set, the unlabeled domain-invariant feature subset from the target domain is used as the test set, and the normal state feature data from the target domain is used to fine-tune the model. The deep transfer autoencoder model (DTAE) is constructed, including the steps of constructing the source DAE model, constructing the target DAE model, parameter transfer, and fine-tuning the target model.

[0088] The specific steps to build DTAE are as follows:

[0089] Step 3.1 Build the source DAE model. DAE is an unsupervised deep neural network, which is composed of several basic autoencoders (AE) stacked together. Each AE consists of two parts: encoder and decoder, including the input layer x = {x 1 ,x 2 ,...,xm}、Hidden layer h={h 1 ,h 2 ,...,h n}、Output layer z={z 1 ,z 2 ,...,z m}.

[0090] The encoder passes through the activation function Map the input data x to the hidden layer h. The mapping process is as follows:

[0091]

[0092] Where W and b are the weight matrix and bias vector of the encoder respectively.

[0093] The decoder uses the activation function Map the data in the hidden layer h to the output data z. The mapping process is as follows:

[0094]

[0095] Where W′ and b′ are the weight matrix and bias vector of the decoder respectively. The hidden layer represents the new features of the input data, and the output data is the reconstruction of the input data.

[0096] The parameters of the AE model are learned by minimizing the reconstruction error between the input layer and the output layer. The reconstruction error is defined as follows:

[0097]

[0098] Where N is the number of samples, θ = {W, b, W′, b′}. The hidden layer features of an AE are used as the input of the next AE model, and a DAE can be constructed by stacking multiple AEs. DAE can mine the deep features of the input feature data, thereby improving the accuracy of fault classification.

[0099] The present invention is configured to include 1 input layer, 3 hidden layers, and 1 softmax layer. The softmax layer is used to classify and represent deep features. The source DAE model is trained using the training set, and the model parameters are obtained, including the weight matrix W (S) With the bias vector b (S) .

[0100] Step 3.2 constructs a target DAE model with the same structure as the source model, i.e., 1 input layer, 3 hidden layers, and 1 softmax layer. At the same time, the parameter weight matrix W learned during the source model training is (S) and the bias vector b (S) Passed to the target model, that is, W (S) =W(T) and b (S) =b (T) .

[0101] Step 3.3 fine-tunes the target DAE model using the normal state feature data from the target domain, and uses the fine-tuned target model to test the remaining subset of unlabeled domain-invariant features from the target domain.

[0102] Process 4. Pattern Recognition

[0103] The domain-invariant features of the target domain are input into the trained DTAE model, and the softmax layer of the DTAE model can realize fault classification of the test data.

[0104] 1 Experimental data introduction

[0105] The test bench is the SQI-MFS test bench, which supports motor speeds of 1200-1800 rpm. The vibration signal is sampled by acquisition cards and acceleration sensors installed at the motor drive end and fan end, with a sampling frequency of 16kHz. The experimental equipment includes a motor fault data acquisition part and multiple detachable faulty motors.

[0106] Here, the cross-domain fault diagnosis is realized by collecting vibration signals of two motors with speeds of 1730rpm and 1750rpm. The main parameters of the motors are shown in Table 1. The experimental bench has a normal state (NM) motor and four faulty motors, corresponding to five motor states: Figure 3 . The faulty motors are set with broken rotor bar fault (BF), winding fault (WF), rotor bending fault (RF) and single-phase voltage unbalance fault (SF). For each motor state, 30 and 60 vibration data samples are randomly selected as training samples and test samples, respectively, and each sample contains 5000 continuous sampling points. The detailed introduction of the motor dataset is shown in Table 2. The motor state category labels are set to 1-5, which are used to match the sample labels of the source domain with the state recognition results. Based on the vibration data at 1730rmp and 1750rmp speeds, we set two cross-domain fault diagnosis tasks (Task 1 and Task 2), as shown in Table 3. The vibration samples at 1730rmp and 1750rmp speeds are selected as the source domains of Task 1 and Task 2 respectively; the vibration samples at 1750rmp and 1730rmp speeds are selected as the target domains of Task 1 and Task 2 respectively. The source domain and target domain of the two tasks contain 150 and 300 vibration samples respectively.

[0107] Table 1: Description of main parameters of motor

[0108]

[0109] Table 2: Introduction to the motor vibration dataset

[0110]

[0111] Table 3: Introduction of cross-domain fault diagnosis tasks for five motor states

[0112]

[0113] 2 Experimental steps

[0114] 2.1 Motor fault diagnosis model based on DISD and DAE

[0115] The sample raw data in the motor fault experiment includes sampled vibration signals of five motor states at two speeds of 1730 rpm and 1750 rpm. The figure clearly shows that there are obvious distribution differences in vibration signals at different speeds. The original vibration signal is processed by DTCWPT, and statistical features are generated by calculating the statistical parameters of the single reconstructed signal of the wavelet packet node.

[0116] First, four layers of DTCWPT processing are performed on each original vibration signal sample to generate 16 terminal wavelet packet nodes (TWPN). For the corresponding reconstructed signal (RS), the 11 statistical parameters described in the invention process 1 are calculated to generate a total of 176 time domain statistical features; at the same time, these 11 statistical parameters are calculated for the 16 Hilbert envelope spectra (HES) of the reconstructed signal to generate a total of 176 frequency domain statistical features. Finally, for each vibration sample, a total of 352 statistical features are generated to form the original feature set (RFS). The vibration samples collected by the NM motor are subjected to wavelet packet decomposition, and the corresponding RS of the obtained TWPN is as follows: Figure 4 . Figure 5 These are 352 normalized statistical features extracted from the vibration signals collected from the NM and BF motors at two speeds of 1730 rpm and 1750 rpm.

[0117] The domain invariant feature selection method DSID proposed in this invention is used to evaluate the fault resolution capability and domain invariance of 352 statistical features of RFS. For the 352 statistical features from the source domain, the feature selection index RIM is calculated. The calculated IS, MMD and RIM are respectively as follows: Figure 6-8 As shown in the figure, the calculated RIM sequence containing 352 statistical features is sorted in descending order, and the sorted RIM sequence is used to select domain-invariant features to construct a domain-invariant feature subset; for the target domain, the RIM sequence is directly used to select domain-invariant features.

[0118] The DTAE model is trained using a subset of domain-invariant features. For the source domain, domain-invariant features are selected to train the source DAE. After the training is completed, the learned parameters W (S) and b(S) Migrate to the target DTAE model with the same structure as the source model. Use the normal feature data from the target domain to fine-tune the target DTAE model. Finally, input the test data (unlabeled feature data from the target domain) into the DTAE model, and use the softmax layer of the DTAE model to classify the test data. Some parameters used in the DTAE model training of the present invention are as follows: the number of hidden layers is 4, the sizes are 400, 100, 50 and 50 respectively, and the number of iterations is set to 200. .

[0119] The experimental results of the proposed TFDD motor diagnosis model are shown in Fig. 9 and are given in Table 4, where dfn is the number of domain-invariant features selected by the DISD method. As can be seen from Table 4, when the number of domain-invariant features is set to 352, that is, all 352 statistical features of the source domain are used to train the DTAE model, the diagnostic accuracy of Task 1 and Task 2 is only 69.00% and 66.67%, respectively. When the proposed domain-invariant feature selection method DISD is executed before the DTAE model, that is, the number of domain-invariant features is limited, the diagnostic accuracy can be significantly improved. When the number of domain-invariant features is set to 101, Task 1 achieves a maximum diagnostic accuracy of 81.68%; when the number of domain-invariant features is set to 140, Task 2 achieves a maximum diagnostic accuracy of 82.67%. Fig. 9 The diagnostic accuracy of Task 1 and Task 2 is shown when the number of domain invariant features is set to 40 to 352. It can be concluded that when the appropriate number of domain invariant features is set, the proposed DISD-based TFDD diagnostic model can improve the accuracy of cross-domain fault diagnosis.

[0120] Table 4: Cross-domain fault diagnosis results of TFDD diagnosis model

[0121]

[0122] 2.2 Experimental analysis of motor fault diagnosis model based on DISD and DAE

[0123] In order to verify the advantages of the motor fault diagnosis model based on DISD and DAE proposed in this invention in cross-domain fault diagnosis, this experiment builds corresponding comparative models based on the commonly used methods of cross-domain fault diagnosis for experimental analysis. Table 5 shows a series of fault diagnosis comparative models. These comparative models are roughly divided into two categories: (1) Models that introduce transfer learning, including RFS-SVM, RFS-KNN, RFS-DAE, RFS-DBN and RFS-CNN models, perform DTCWPT signal processing and feature extraction on the original vibration signal samples in the source domain, and use the original feature set (RFS) containing 352 statistical features as the input of SVM (support vector machine), KNN (K nearest neighbor classification), DAE (deep autoencoder), DBN (deep belief network), CNN (convolutional neural network) to train the fault diagnosis model, and use the trained diagnosis model to perform pattern classification on the original feature set (RFS) in the target domain. (2) Models that introduce transfer learning, including RFS-TCA, RFS-JDA, RFS-DSID-TCA, RFS-DSID-JDA and TFDD models. Among them, RFS-TCA and RFS-JDA refer to inputting the RFS obtained in the source domain and the RFS obtained in the target domain into TCA and JDA at the same time. Transfer component analysis (TCA) refers to mapping the source domain and the target domain to the high-dimensional reproducing kernel Hilbert space at the same time to minimize the distance between the source domain and the target domain; the target domain of joint distribution adaptation (JCA) adopts a pseudo-label strategy, and iterates repeatedly to reduce the marginal distribution and conditional distribution of the source domain and the target domain. Finally, the source domain features after transfer learning are input into the SVM classifier to train the fault diagnosis model, and the trained diagnosis model is used to perform state recognition and classification on the target domain features after transfer learning. RFS-DSID-TCA and RFS-DSID-JDA refer to first using the proposed DSID method to select domain invariant features for RFS, and then performing subsequent transfer learning and fault diagnosis.

[0124] Table 5: Comparison model based on common methods of cross-domain fault diagnosis

[0125]

[0126] The experimental results of the above comparison models are shown in Table 6 and Figure 10-11As shown in Table 6, the diagnostic accuracy of other comparison models is significantly lower than that of the method TFDD proposed in this paper. The models RFS-TCA, RFS-JDA, RFS-DSID-TCA, and RFS-DSID-JDA based on transfer learning have better diagnostic performance in cross-domain fault diagnosis compared with the models without transfer learning. For the transfer learning model using the DISD method, its diagnostic performance is significantly higher than that of the transfer learning model without the DISD method. Taking Task 1 as an example, under the premise of selecting the most appropriate number of domain-invariant features, the diagnostic accuracy of the RFS-DSID-TCA model is 71.67%, which is 5.67% higher than that of the RFS-TCA model; the diagnostic accuracy of the RFS-DSID-JDA model is 77.00%, which is 4.67% higher than that of the RFS-JDA model. For Task 2, the transfer learning model using the DISD method has a more obvious effect. Under the premise of selecting the most appropriate number of domain-invariant features, the diagnostic accuracy of the RFS-DSID-TCA model is 77%, which is 7.33% higher than that of the RFS-TCA model; the diagnostic accuracy of the RFS-DSID-JDA model is 78.00%, which is 8.00% higher than that of the RFS-JDA model. The model using the TFDD proposed by the present invention has a higher diagnostic accuracy than RFS-DSID-TCA and RFS-DSID-JDA.

[0127] Based on the above experimental analysis results, we can conclude that:

[0128] (1) For Task 1 and Task 2, the proposed TFDD model can effectively classify the five working states of the motor, and the maximum diagnostic accuracy can reach more than 80%. The proposed domain-invariant feature selection method DSID can effectively select features with high fault discrimination and domain invariance, and select the appropriate number of domain-invariant features to improve the cross-domain diagnostic performance accuracy.

[0129] (2) The comparison results show that the diagnostic performance of TFDD is significantly better than the comparison models shown in Table 6. In addition, the diagnostic model combined with the transfer learning strategy is helpful to improve the cross-domain diagnostic accuracy; the DTAE model proposed in the present invention has a better cross-domain fault diagnosis effect than the general transfer method.

[0130] The SQI-MFS test bench was used to conduct multiple comparative tests, verifying that the proposed DISD and DTAE can improve the recognition accuracy of the motor cross-domain fault diagnosis model. The above content is a further detailed description of the present invention in combination with the specific implementation methods of actual cases. It cannot be determined that the specific implementation of the present invention is limited to these descriptions. Without departing from the concept of the present invention, technicians in the relevant fields who make some simple deductions or substitutions should be regarded as belonging to the protection scope of the present invention.

Claims

1. A motor fault diagnosis method based on deep autoencoder and feature optimization, characterized in that: The details are as follows: Step 1, using dual-tree complex wavelet packet transform to perform signal processing and statistical feature extraction on the original vibration signals under different working conditions, and construct an initial statistical feature set; Step 2: evaluate the fault resolution ability and domain invariance of the features in the initial statistical feature set constructed in step 1, select domain invariant features that can characterize the motor state, and construct a domain invariant feature subset of the sample. Step 3: Use the domain-invariant feature subset of the source domain obtained in step 2 to train the source DAE model, obtain model parameters, and transfer the parameters to the target DAE model. Use the normal state domain-invariant feature subset of the target domain obtained in step 2 to fine-tune the target model, and complete the construction of the deep transfer autoencoder model DTAE. Step 4, testing the deep transfer autoencoder model DTAE constructed in step 3, and outputting the fault diagnosis result; The specific steps of step 1 are as follows: the original vibration signal samples from different working conditions are divided into vibration data set I and vibration data set II, where vibration data set I is the source domain and its data set samples are labeled; vibration data set II is the target domain and its data set samples are unlabeled, the diagnostic model is trained using samples with known labels from the source domain, and the trained model is used to classify the state of the target domain; The specific process in step 2 is as follows: using the labeled fault feature data from the initial statistical feature set of the source domain and the normal state feature data of the source domain and the target domain, the features of the initial statistical feature set are evaluated by using the domain-invariant feature selection method based on importance score and domain invariance, and the domain-invariant feature selection index RIM is obtained. The sorted RIM sequence is used to select the domain-invariant feature subset of the initial statistical feature set from the target domain.

2. The motor fault diagnosis method based on deep autoencoder and feature optimization according to claim 1, characterized in that: For the labeled fault feature data in the source domain obtained in step 1, the importance score IS of each feature is calculated through the feature importance evaluation method based on random forest to evaluate the fault resolution ability of the feature; the inter-domain difference of a single feature in the source domain and the target domain is analyzed, and the maximum mean difference MMD is used to estimate the distribution difference MMDS of the same feature between different domains. For the IS and MMDS of each feature obtained, the ratio of IS to MMDS is calculated, and the domain-invariant feature selection index RIM is constructed, and the RIM sequence is sorted in descending order; for the unlabeled feature data from the original feature set of the target domain, the sorted RIM sequence is directly used for domain-invariant feature selection.

3. The motor fault diagnosis method based on deep autoencoder and feature optimization according to claim 2 is characterized in that: The specific steps of the feature importance evaluation method based on random forest are as follows: For N motor vibration signal samples in the source domain, after the signal processing and statistical feature extraction process in step 1, each vibration signal sample is obtained to contain Q statistical features, thereby forming an initial statistical feature set F = [f1, f2, ..., f N ] T ,in Represents the set of Q statistical features of the i-th vibration signal sample; the RF-based feature selection method is used to calculate the importance score IS for each feature in the original feature set. The main idea of ​​RF is to construct a decision tree DT. Assume that an RF has n DT The decision tree is composed of 10 decision trees; Using F j and its label Y to construct the set F j is the jth statistical feature set of the vibration data set Ι, y i Represents the motor state label, C and N represent the number of categories and samples respectively; calculate the decision tree DT constructed by S n The Gini index GI of each node of , GI(t) represents DT n The GI value at node t is used to quantify the impurity of node t and is defined as follows: Where, the probability of occurrence of the k-th sample of node t; feature F j The importance score IS can be defined by the following formula: Where n DT It can be expressed as the number of DTs in RF, t∈DT n represents a set of split nodes; Finally, in the original high-dimensional feature set, the importance score IS(j) of each feature is calculated to obtain the IS sequence IS={IS(1), IS(2),…, IS(Q)} of Q features. According to the sorted IS sequence, features with high IS values ​​are selected to construct a feature subset.

4. The motor fault diagnosis method based on deep autoencoder and feature optimization according to claim 2 is characterized in that: For the normal state feature data of the source domain and the target domain, MMD is used to evaluate the distribution difference of the same feature in different domains. The value of MMD is used as a quantitative indicator of the invariance of the feature domain. The normal state feature data from the source domain and the target domain are defined as SF and TF, respectively, both containing N samples, and each sample contains Q statistical features after signal processing and statistical feature extraction in step 1. The expressions of SF and TF are as follows: Among them, the j-th column elements of SF and TF represent N samples ((j∈[1,2,…,Q])) of the j-th feature from the source domain and the target domain, respectively, and their expressions are as follows: SF j and TF j The empirical distance estimate between the distributions is defined as follows: In the formula, ||·|| H represents the RKHS norm, φ represents the kernel-induced feature map, and the MMD sequence of Q statistical features is calculated MMD={MMD(1),MMD(2),…,MMD(Q)}.

5. The motor fault diagnosis method based on deep autoencoder and feature optimization according to claim 2 is characterized in that: The IS and MMD values ​​of each feature are used to construct a new feature selection index RIM, which is used to select domain-invariant features from RFS. The RIM of the jth feature is defined as follows: For q kinds of statistical features, a RIM sequence RIM = {RIM(1), RIM(2), ..., RIM(q)} can be obtained. The RIM sequence is sorted, and features with higher RIM are selected. The cross-domain fault diagnosis model is trained with the RIM sequence sorted in descending order.

6. The motor fault diagnosis method based on deep autoencoder and feature optimization according to claim 1, characterized in that: The specific steps to build the deep transfer autoencoder model DTAE are as follows: Step 3.1 Build the source DAE model. DAE is an unsupervised deep neural network, which is composed of several basic autoencoders stacked together. Each AE consists of an encoder and a decoder, including an input layer x = {x1, x2, ..., x m }、Hidden layer h={h1,h2,...,h n }、Output layer z={z1,z2,...,z m }; The encoder passes through the activation function Map the input data x to the hidden layer h. The mapping process is as follows: Where W and b are the weight matrix and bias vector of the encoder respectively; The decoder uses the activation function Map the data in the hidden layer h to the output data z. The mapping process is as follows: In the formula, W′ and b′ are the weight matrix and bias vector of the decoder respectively, the hidden layer represents the new features of the input data, and the output data is the reconstruction of the input data; The parameters of the AE model are learned by minimizing the reconstruction error between the input layer and the output layer. The reconstruction error is defined as follows: Where N is the number of samples, θ = {W, b, W′, b′}, the hidden layer features of an AE are used as the input of the next AE model, and a DAE is constructed by stacking multiple AEs; Step 3.2 constructs a target DAE model with the same structure as the source model: 1 input layer, 3 hidden layers, and 1 softmax layer. At the same time, the parameter weight matrix W learned during the source model training is (S) and the bias vector b (S) Passed to the target model, that is, W (S) =W (T) and b (S) =b (T) ; Step 3.3 fine-tunes the target DAE model using the normal state feature data from the target domain, and uses the fine-tuned target model to test the remaining subset of unlabeled domain-invariant features from the target domain.

7. The motor fault diagnosis method based on deep autoencoder and feature optimization according to claim 1 is characterized in that: In step 4, the unlabeled domain-invariant feature subset of the test set is used as input to train the classification performance of the constructed target DAE model, and finally the cross-domain diagnosis result is output, that is, the label information of the target domain is obtained.

Images