A device status diagnosis method based on wavelet packet frequency graph and ResNet migration network

Abstract

The present invention discloses an equipment state diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network, which relates to the field of fault diagnosis. The wavelet packet transform is used to perform time-frequency analysis on nonlinear and non-stationary vibration signals, and a wavelet packet time-frequency feature diagram of the vibration signal is constructed. Aiming at the characteristics of small size and feature dispersion of WPT-TFF, a multi-group parallel ResNet network structure is designed to extract the deep features of WPT-TFF. The MK-MMD is used to evaluate the distribution difference between the deep features of source domain and target domain data. The deep feature extraction network is optimized by combining with the classification loss of the source domain sample set, so that the deep features have better cross-domain invariance and fault state distinction ability. By comparing with other multi-group fault diagnosis models, the WMGRNMM diagnosis model proposed in the present invention has good variable working condition fault diagnosis and noise resistance in rotating machinery and equipment, and has strong adaptability and generalization ability in actual industrial scenarios.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and in particular to a device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network. Background Art

[0002] Rotating machinery is widely used in important engineering fields of national economy and people's livelihood, such as aerospace, rail transportation, petrochemical, automobile manufacturing and wind power generation, and plays a vital role in national economic production. Research on intelligent fault diagnosis methods of rotating machinery has important theoretical significance and engineering value for ensuring the safe and efficient operation of equipment and avoiding huge economic losses and major production safety accidents. The vibration signals generated by rotating machinery contain rich state information and are often used for fault diagnosis. Traditional machine learning methods such as BP neural network, support vector machine, fuzzy C-means clustering and other methods have achieved a series of research results in rotating machinery diagnosis. However, traditional machine learning requires effective signal analysis methods and prior knowledge as guidance, which cannot meet the development needs of big data in the industrial field.

[0003] Compared with traditional machine learning, deep learning uses end-to-end learning ideas to adaptively extract deep semantic features from vibration signals, reducing the involvement of empirical knowledge and being more suitable for processing complex vibration signals of rotating machinery and equipment. However, as the network depth increases, the deep CNN network is prone to gradient diffusion or explosion problems during the training process. In order to solve this problem, the present invention uses ResNet to solve the model degradation problem caused by the CNN network being too deep. Since the time-frequency vibration signals of rotating machinery and equipment contain rich damage mechanisms, and wavelet packet decomposition has a powerful nonlinear and non-stationary vibration signal analysis capability, wavelet packet analysis is used to pre-process the vibration signal, hoping to extract the time-frequency information in the vibration signal that can reflect the fault state. The WPT vibration signal time-frequency feature map construction method can effectively utilize the time-frequency information of vibration signals of different scales and resolutions, and can effectively improve the accuracy of model fault diagnosis when combined with deep learning.

[0004] In variable operating condition application scenarios, the data in the source domain and the target domain do not satisfy the assumption of independent and identical distribution, which makes it difficult for the model trained with the source domain data to obtain good results on the target domain data. Therefore, it is necessary to introduce a transfer learning method to solve this problem. However, in real application scenarios, even a small amount of target domain data labels may not be available; if only the target domain data can be obtained but not their labels, then the domain distribution adaptation transfer learning method is a good choice. The domain distribution adaptation method solves the variable operating condition migration problem by finding a distribution space similar to the source domain and the target domain. The present invention utilizes the multi-core maximum mean difference (MK-MMD) method to obtain the optimal kernel through multi-core linear combination to solve the problem of improper kernel parameter selection caused by a single kernel, analyze the distribution difference of the deep features of the source domain and the target domain, and optimize the deep feature extraction network to solve the problem of decreased model diagnostic ability caused by the distribution difference between the deep features of the source domain and the target domain. On this basis, the present invention proposes a device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network, which combines WPT time-frequency signal analysis with MK-MMD, greatly improving the fault diagnosis rate and generalization ability of the model under variable working conditions. Summary of the invention

[0005] The technical problem to be solved by the present invention is to meet the demand for fault diagnosis of rotating machinery equipment in actual industrial production, ensure the safe and reliable operation of rotating machinery, and reduce the risk of safety accidents caused by its failure. The present invention proposes an equipment status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network to meet the high-precision fault classification requirements of the rotating machinery equipment fault diagnosis model.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] A device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network specifically includes the following steps:

[0008] Step 1, using wavelet packet transform to perform time-frequency analysis on the nonlinear and non-stationary vibration signal, and constructing a wavelet packet decomposition time-frequency feature diagram of the vibration signal;

[0009] Step 2, establish a deep feature extraction network based on ResNet, and extract the deep features of the wavelet packet decomposition time-frequency feature map of the vibration signal;

[0010] Step 3: Use MK-MMD to evaluate the distribution difference between the deep features of the source domain and target domain data, and combine it with the classification loss of the source domain sample set to optimize the deep feature extraction network so that the deep features have better cross-domain invariance and fault state differentiation capabilities; where MK-MMD is the multi-core maximum mean error.

[0011] As a further preferred solution of the device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network of the present invention, the step 1 is specifically as follows:

[0012] Step 1.1, performing four-layer, five-layer, and six-layer wavelet packet decomposition on the collected vibration signal of the rotating mechanical equipment, and then obtaining its wavelet packet node coefficients;

[0013] Step 1.2, using the coefficients of a single wavelet packet leaf node to construct a coefficient matrix, concatenating the coefficient matrices of all leaf nodes in a Z-shaped order to establish a wavelet packet decomposition time-frequency feature map of the vibration signal.

[0014] As a further preferred solution of the device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network of the present invention, the step 2 is specifically as follows:

[0015] Step 2.1, decompose the vibration signal using 4-layer, 5-layer, and 6-layer wavelet packet decomposition, and construct the corresponding time-frequency feature map WPT-TFF;

[0016] Step 2.2, establish a deep feature extraction network based on ResNet, and extract the deep features of the wavelet packet decomposition time-frequency feature map of the vibration signal.

[0017] Among them, the deep feature extraction network based on ResNet is as follows:

[0018] (1) The WPT-TFF time-frequency feature diagram of the vibration signal is small in size, and the state characteristics are scattered in the time-frequency diagram:

[0019] For the small-sized time-frequency feature map WPT-TFF, 3*3 convolution kernels are used in the ResNet network for feature extraction, the convolution step is set to 1, and the pooling operation of the middle layer is cancelled. At the same time, the depth of the ResNet network is limited, and a 7-layer residual network deep feature extraction module ResNet Block is designed to reduce the impact of small image size and feature dispersion on the generalization ability of the model.

[0020] (2) The time-frequency feature maps obtained by wavelet packet decomposition at different levels have their own structural characteristics:

[0021] Three 7-layer ResNet Blocks are designed to form the ResNet network structure MGRN, which extract the feature information of 4-layer, 5-layer and 6-layer WPT-TFF respectively. Three 7-layer ResNet Blocks with the same structure and independent parameters are more effective in extracting the personalized structural features of different types of time-frequency feature maps WPT-TFF. The deep features extracted by the three 7-layer ResNet Blocks are spliced ​​into combined deep features and sent to the fully connected layer, and finally enter the Softmax classifier for classification.

[0022] As a further preferred solution of the device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network of the present invention, the step 3 is specifically as follows:

[0023] Step 3.1: Send the labeled source domain data and the unlabeled target domain data to the parallel network MGRN in step 2 respectively to obtain the deep features of the source domain and target domain datasets, denoted as F s and F t , calculate F s and F t The MK-MMD loss between As shown in formula (1),

[0024]

[0025] Where P is the marginal probability distribution of the source domain, Q is the marginal probability distribution of the target domain, E represents the mathematical expectation, φ represents the mapping of the regenerated Hilbert space, and H k represents the kernel k used to reproduce the Hilbert space;

[0026] At the same time, F s After the fully connected layer and the Softmax classifier, the predicted category information is obtained, and the cross entropy loss is calculated by combining the label information, which is recorded as As shown in formula (2),

[0027]

[0028] Where: M is the number of categories; is a symbolic function, and is 0 or 1; p ic is the predicted probability that sample i belongs to category c;

[0029] Calculate the overall loss of the transfer classification model As shown in formula (3):

[0030]

[0031] In the formula, λ MK-MMD The trade-off parameter of total loss; F sis the source domain data; F t is the target domain data;

[0032] Step 3.2, optimize the deep network and fully connected network parameters using back propagation and gradient descent methods;

[0033] Step 3.3: In the test phase, the deep features of the target domain sample time-frequency graph are extracted using the trained MGRN network, and the deep features are sent to the trained FC1 to obtain the category label of the sample and complete the fault classification.

[0034] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0035] The present invention discloses an equipment state diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network, which has good variable working condition migration diagnosis ability, including state recognition accuracy, scene adaptability and noise resistance, and can ensure the safe and reliable operation of rotating machinery, reduce the risk of safety accidents caused by its failure, and meet the high-precision fault classification requirements of the rotating machinery equipment fault diagnosis model; in view of the problem that the existing signal analysis method of rotating machinery equipment easily ignores the time-frequency characteristics of vibration signals, the present invention selects 4, 5, 6 wavelet packet analysis to pre-process the vibration signal, reconstructs it into a grayscale wavelet packet time-frequency feature diagram, and extracts the time-frequency information that can reflect the fault state of the rotating machinery equipment; in view of the fact that the traditional deep learning model cannot fully extract the effective features of different time-frequency diagrams, three RNB-7 networks with the same structure and independent parameters are constructed; in view of the problem that the model diagnosis ability is insufficient in variable working conditions, MK-MMD is used to evaluate the distribution difference between the deep features of source domain and target domain data, and the source domain classification loss is combined to complete the optimization of the deep feature extraction network, thereby improving the variable working condition model fault diagnosis accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of constructing a time-frequency diagram of wavelet packet decomposition according to the present invention;

[0037] Figure 2 It is the characteristic graph of the wavelet packet WPT-TFF with different layers of the present invention;

[0038] Figure 3 It is the ResNet Block flow chart of the present invention;

[0039] Figure 4 It is a structural diagram of the parallel network MGRN of the present invention;

[0040] Figure 5 This is a WMGRNMM fault condition migration diagnosis structure diagram of the present invention;

[0041] Figure 6 This is a comparison experiment diagram of various network model fault classifications;

[0042] Among them, (a) is the various migration modes of the data set 0HP, (b) is the various migration modes of the data set 1HP, (c) is the various migration modes of the data set 2HP, and (d) is the various migration modes of the data set 3HP;

[0043] Figure 7(a) shows the change of recognition accuracy of each model state with the number of iterations in the 0->1 migration mode;

[0044] Figure 7(b) shows the change of recognition accuracy of each model state with the number of iterations in the 0->2 migration mode;

[0045] Figure 7(c) shows the change of recognition accuracy of each model state with the number of iterations in the 0->3 migration mode;

[0046] Figure 7(d) shows the change of recognition accuracy of each model state with the number of iterations in the 1->0 migration mode;

[0047] Figure 7(e) shows the change of recognition accuracy of each model state with the number of iterations in the 1->2 migration mode;

[0048] Figure 7(f) shows the change of recognition accuracy of each model state with the number of iterations in the 1->3 migration mode;

[0049] Figure 7(g) shows the change of recognition accuracy of each model state with the number of iterations in the 2->0 migration mode;

[0050] Figure 7(h) shows the variation of recognition accuracy of each model state with the number of iterations in the 2->1 migration mode;

[0051] Figure 7(i) shows the variation of recognition accuracy of each model state with the number of iterations in the 2->3 migration mode;

[0052] Figure 7(j) shows the variation of recognition accuracy of each model state with the number of iterations in the 3->0 migration mode;

[0053] Figure 7(k) shows the variation of recognition accuracy of each model state with the number of iterations in the 3->1 migration mode;

[0054] Figure 7(l) shows the variation of recognition accuracy of each model state with the number of iterations in the 3->2 migration mode;

[0055] FIG8( a ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WMGRN model under the 3-1 migration mode of the present invention;

[0056] FIG8( b ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WMGRNMM model under the 3-1 migration mode of the present invention;

[0057] FIG8( c ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WRNMM model under the 3-1 migration mode of the present invention;

[0058] FIG8( d ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WRN18MM model under the 3-1 migration mode of the present invention;

[0059] FIG8( e ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WRN34MM model under the 3-1 migration mode of the present invention;

[0060] Fig. 9 It is a flow diagram of a device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0062] The present invention will be further supplemented with the accompanying drawings and specific embodiments below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Those skilled in the art should understand that without departing from the above-mentioned technical ideas of the present invention, various replacements and changes to the present invention based on the basic technical knowledge and means in the field should be included in the protection scope of the present invention.

[0063] The present invention implements a device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network with the following technical scheme: Fig. 9 As shown, the details are as follows:

[0064] Step 1. Wavelet packet decomposition time-frequency diagram construction:

[0065] The present invention performs four-layer, five-layer, and six-layer wavelet packet decomposition on the collected vibration signal of the rotating mechanical equipment, and then obtains its wavelet packet node coefficients. Since different wavelet packet node coefficients represent different physical meanings, we propose to use the coefficients of a single wavelet packet leaf node to construct a coefficient matrix. In order to facilitate the subsequent two-dimensional convolution kernel to extract the deep features of the wavelet coefficients, we splice the coefficient matrices of all leaf nodes in a Z-shaped order to establish the wavelet packet decomposition time-frequency feature map of the vibration signal (abbreviated as WPT-TFF).

[0066] Here, we take the vibration signal sample of rotating machinery equipment with 4096 sampling points as an example to introduce the process of constructing the time-frequency feature map by wavelet packet decomposition. Figure 1 shown.

[0067] 1) After the vibration signal sample is decomposed into 4 layers of wavelet packets, 16 wavelet packet leaf nodes can be obtained, and each wavelet packet leaf node has 256 wavelet packet coefficients.

[0068] 2) Construct each leaf node wavelet packet coefficient into a 16*16 coefficient matrix.

[0069] 3) The coefficient matrices of the 16 wavelet packet leaf nodes are zigzag-joined according to the numbers to obtain the WPT-TFF of the vibration signal samples of the rotating mechanical equipment.

[0070] Step 2. Deep feature extraction network based on ResNet:

[0071] In order to better analyze the time-frequency characteristics of the vibration signal of rotating machinery equipment at different scales and resolutions, 4-layer, 5-layer, and 6-layer wavelet packet decomposition are used to decompose the vibration signal, and the corresponding time-frequency feature graphs WPT-TFF(4), WPT-TFF(5), and WPT-TFF(6) are constructed. The vibration signal WPT-TFF under different fault conditions is shown in Figure 2. Figure 1 As shown. A deep feature extraction network based on ResNet is established to extract the deep features of WPT-TFF(4), WPT-TFF(5), and WPT-TFF(6). Figure 2 According to the analysis, the ResNet deep feature extraction network faces the following two problems:

[0072] (1) The WPT-TFF size of the vibration signal is small, and the state characteristics are scattered in the time-frequency diagram.

[0073] (2) The time-frequency feature maps obtained by wavelet packet decomposition of different layers have their own structural characteristics.

[0074] For problem (1), first, for the small-sized WPT-TFF, 3*3 convolution kernels are used in the ResNet network for feature extraction, the convolution step size is set to 1, and the pooling operation of the intermediate layer is cancelled; then, unlike the conventional ResNet that improves the feature extraction capability by increasing the network depth, the depth of the ResNet network is limited here, and a 7-layer residual network deep feature extraction module ResNet Block (here referred to as RNB-7) is designed, as shown in Figure 3 As shown, this method can reduce the impact of small image size and feature dispersion on the generalization ability of the model.

[0075] For problem (2), three RNB-7s are designed to form a parallel network MGRN to extract the feature information of 4-layer, 5-layer, and 6-layer WPT-TFF respectively, as follows: Figure 4As shown in the figure, three RNB-7s with the same structure and independent parameters can more effectively extract personalized structural features of different types of WPT-TFF. The deep features extracted by the three RNB-7s are concatenated into combined deep features and sent to the fully connected layer, and finally enter the Softmax classifier for classification.

[0076] Step 3. Design of migration diagnosis model based on MK-MMD

[0077] By training the fault diagnosis model in step 2 with labeled data from the source domain of rotating machinery, we can obtain deep features that express the fault status of the source domain data. However, when there is a distribution difference between the deep features of the source domain and the target domain data, the recognition accuracy of the model in the target domain will drop significantly. In order to improve the similarity of the probability distribution of the deep features of the source domain data and the target domain data, MK-MMD is used here to perform domain adaptation measurement on the deep features of the source domain data and the target domain data. The labeled source domain data and the unlabeled target domain data are respectively sent to the MGRN in step 2 to obtain the deep features of the source domain and the target domain data sets, denoted as F s and F t , calculate F s and F t The MK-MMD loss between As shown in formula (1),

[0078]

[0079] Where P is the marginal probability distribution of the source domain, Q is the marginal probability distribution of the target domain, E represents the mathematical expectation, φ represents the mapping of the regenerated Hilbert space, and H k represents the kernel k used to reproduce the Hilbert space.

[0080] At the same time, F s After the fully connected layer and the Softmax classifier, the predicted category information is obtained. Combined with the label information, the cross entropy loss can be calculated, which is recorded as As shown in formula (2),

[0081]

[0082] Where: M is the number of categories; is the sign function (0 or 1); p ic The predicted probability that sample i belongs to class c.

[0083] Calculate the overall loss of the transfer classification model As shown in formula (3):

[0084]

[0085] In the formula, λ MK-MMDThe trade-off parameter of total loss; F s is the source domain data; F t It is the target domain data.

[0086] Here, the source domain and the target domain share the parameters of the deep feature extraction network. The goal of the entire network is to minimize the total loss. The back propagation and gradient descent methods are used to optimize the parameters of the deep network and the fully connected network. In the test phase, the deep features of the target domain sample time-frequency graph are extracted using the trained MGRN network, and the deep features are sent to the trained FC1 to obtain the category label of the sample and complete the fault classification. The deep features extracted by the MK-MMD-based migration diagnosis model can express the fault status of the source domain data. At the same time, the difference in the deep features of the source domain and target domain data is reduced, and the generalization ability of the model is improved.

[0087] The specific embodiments are as follows:

[0088] Step 1: Collect bearing vibration signals of various states under four different powers of 0HP, 1HP, 2HP, and 3HP, and divide them into source domain data and target domain data. The source domain data contains both bearing sample data and corresponding sample labels, and the target domain data only contains bearing sample data, but does not contain labels corresponding to bearing samples. The bearing vibration signals are divided into 12 fault types according to the degree of damage to the inner ring, outer ring, and rolling element, and the labels are 0 to 11.

[0089] Step 2, perform wavelet packet decomposition on the sample signals of the training set and the test set to obtain the wavelet packet coefficients of each sample signal after four-layer, five-layer, and six-layer wavelet packet decomposition. The obtained wavelet packet node coefficients represent different physical meanings of the time-frequency characteristics of the vibration signal at different scales. The present invention uses the coefficients of a single wavelet packet leaf node to construct a coefficient matrix; the vibration signal samples of 4096 sampling points are used to construct a time-frequency feature map for a four-layer wavelet packet. First, the vibration signal sample is decomposed through four layers of wavelet packets to obtain 16 wavelet packet leaf nodes, and each wavelet packet leaf node has 256 wavelet packet coefficients; secondly, the wavelet packet coefficients of each leaf node are constructed into a 16*16 coefficient matrix; finally, the coefficient matrices of the 16 wavelet packet leaf nodes are zigzag spliced ​​according to the numbers to obtain the WPT-TFF of the vibration signal sample. The construction of the time-frequency feature map using five-layer and six-layer wavelet packets is the same as that of the four-layer wavelet packet.

[0090] Step 3: In order to better analyze the time-frequency characteristics of the source domain and target domain bearing vibration signals at different scales and resolutions, the problem of the dispersion of state features in the time-frequency diagram caused by the small size of WPT-TFF is solved. The corresponding WPT-TFF (4), WPT-TFF (5), and WPT-TFF (6) time-frequency feature maps are sent to the 7-layer RNB-7 network to reduce the impact of small image size and feature dispersion on the generalization ability of the model.

[0091] Step 4: In order to better extract the combined deep features of the four-layer, five-layer, and six-layer wavelet packet time-frequency graphs of the source and target domains, three RNB-7s are designed to form a parallel network MGRN to extract the feature information of 4-layer, 5-layer, and 6-layer WPT-TFFs respectively. The three RNB-7s with the same structure and independent parameters can more effectively extract the personalized structural features of different types of WPT-TFFs. The deep features extracted by the three RNB-7s are spliced ​​into combined deep features. The deep features of the source domain samples are sent to the fully connected layer FC1, and the cross entropy loss is calculated by combining the softmax classifier and the source domain label information.

[0092] Step 5: To improve the similarity of the probability distribution of the deep features of the source domain data and the target domain data, MK-MMD is used here to perform domain adaptation measurement on the deep features of the source domain data and the target domain data. The MK-MMD distance of the deep features of the source domain and the target domain samples is calculated to obtain the MK-MMD loss of the deep feature space of the source domain and the target domain.

[0093] Step 6: Use the source domain sample classification loss Back propagation optimizes the fully connected layer FC1 and calculates loss and back-propagate to optimize MGRN; iterate steps 3)-5) until The number of loss iterations reaches the target requirement, the model parameters that minimize the loss are achieved, and the trained MGRN and FC1 are obtained; the deep features of the target domain sample time-frequency graph are extracted using the trained MGRN network, and the deep features are sent to the trained FC1 to obtain the category label of the sample. Here, the source domain and the target domain share the deep feature extraction network parameters. The goal of the entire network is to minimize the total loss, and the back propagation and gradient descent methods are used to optimize the deep network and fully connected network parameters. The deep features extracted by the MK-MMD-based migration diagnosis model can express the fault status of the source domain data. At the same time, the difference in the deep features of the source domain and target domain data is reduced, which improves the generalization ability of the model.

[0094] 1 Experimental data introduction

[0095] The bearing fault diagnosis experimental platform data set is used. The experimental data set includes 12kHz drive end, 12kHz fan end, and 48kHz drive end vibration data at four different powers of 0HP, 1HP, 2HP, and 3HP. The speeds corresponding to different horsepower are shown in Table 1. Table 1 shows the speed and benchmark data corresponding to the data set. This experiment analyzed the drive end vibration data with a sampling frequency of 12kHz, including 12 different bearing conditions: (0) Normal (1) Inner ring damage, diameter is 0.007 inches (2) Outer ring damage, diameter is 0.007 inches (3) Rolling element damage, diameter is 0.007 inches (4) Inner ring damage, diameter is 0.014 inches (5) Outer ring damage, diameter is 0.014 inches (6) Rolling element damage, diameter is 0.014 inches (7) Inner ring damage, diameter is 0.021 inches (8) Outer ring damage, diameter is 0.021 inches (9) Rolling element damage, diameter is 0.021 inches (10) Inner ring damage, diameter is 0.028 inches (11) Rolling element damage, diameter is 0.028 inches.

[0096] Table 1

[0097]

[0098]

[0099] The vibration signal sequence of 4096 sampling points is taken as a sample, and 30 vibration signal samples of the bearing in each state under a single power are obtained. The vibration data under four different powers are used to establish four different working condition data sets, namely 0HP data set, 1HP data set, 2HP data set, and 3HP data set. Each data set includes 360 vibration signal samples, as shown in Table 2. Here, the data sets under the four working conditions are subjected to mutual migration diagnosis tests, and there are 12 migration modes (0HP->1HP, 0-2, 0-3, 1-0, 1-2, 1-3, 2-0, 2-1, 2-3, 3-0, 3-1, 3-2), where 0-1 means that the training set (source domain) is the 0HP data set and the test set (target domain) is the 1HP data set.

[0100] Table 2

[0101]

[0102] 2 Experimental steps

[0103] In order to verify the migration diagnosis effect of the bearing fault diagnosis method based on WPT time-frequency analysis and feature migration between different working conditions proposed in this invention, we set up four groups of experiments to verify the effectiveness of the proposed method. In the first group of experiments, we compared the WMGRNMM model with the WMGRN model. The WMGRN model has the same network structure as the WMGRNMM model, but does not use MK-MMD for domain adaptation optimization. In the second group of experiments, we compared the WMGRNMM model with the WRNMM model. Compared with the WMGRNMM model, the WRNMM model changed the deep feature extraction network and used a single RNB-7 network instead of MGRN, and input WP-TFF(4), WP-TFF(5), and WP-TFF(6) as three channels into a single RNB-7. In the third group of experiments, we compared the WMGRNMM model with the WRN18MM and WRN34MM models. The WRN18MM and WRN34MM models increased the depth of Resnet and used the classic ResNet18 and ResNe34 networks instead of the RNB-7 network designed in this paper respectively. In the fourth group of experiments, in order to demonstrate the anti-noise ability of the WMGRNMM model, four types of noise with signal-to-noise ratios of 3dB, 5dB, 8dB, and 10dB were added to the original vibration signal to test the impact of the noise signal on the vibration model. The present invention analyzes the accuracy of bearing fault diagnosis in four groups through experimental comparison, which is described in detail as follows.

[0104] In the first group of experiments, the experimental results of the WMGRNMM model and the WMGRN model are shown in Table 3. From the table, we can see that the classification accuracy of the WMGRNMM model is much higher than that of the WMGRN model in each working condition migration process. The average classification accuracy of the WMGRNMM model is 9.17% higher than that of the WMGRN model. Moreover, the WMGRN model is unstable in various working condition migration diagnosis modes, with the highest reaching 97% and the lowest only 81%, and the accuracy difference is too large. This shows that the deep features of the source domain and target domain data extracted by WPT and MGRN have certain distribution differences under different working conditions, resulting in the diagnosis model trained by the source domain data unable to achieve good recognition accuracy in the target domain. After adding the MK-MMD domain migration method, the diagnosis results of the WMGRNMM model in various modes have been greatly improved, and the results are stable at more than 97%, with an average accuracy of 98.52%. It can be seen that the MK-MMD loss function can better measure the distance between the source domain and the target domain, reduce the distribution difference between the two domains, maintain model stability, and better optimize the model classification accuracy.

[0105] Table 3

[0106]

[0107] The second group of experiments, WMGRNMM model and WRNMM model, the experimental results of WRNMM model are shown in Table 4. From the table, we can see that the classification accuracy of WMGRNMM model is much higher than that of WMGRN model in the process of migration of each working condition. Moreover, the average classification accuracy of WMGRNMM model is 98.52%, while the average classification accuracy of WRNMM model is only 95.09%, which is 3.43% lower than the average classification accuracy of WMGRNMM model. It can be seen that the single RNB-7 network model cannot take into account the structural characteristics of the time-frequency feature map obtained by wavelet packet decomposition of different layers, resulting in the WRNMM model being unable to fully extract the different features contained in WPT-TFF, thereby reducing the accuracy of the migration model fault. Therefore, we can conclude that for the time-frequency features of different resolutions in WP-TFF (4), WP-TFF (5), and WP-TFF (6), the effect of using a parallel network with three independent parameters is better than that of using a single network to extract deep features.

[0108] Table 4

[0109]

[0110] The third group of experiments, the experimental results of the WMGRNMM model and the WRN18MM and WRN34MM models are shown in Table 5. As the network depth increases, the recognition accuracy of the two models does not increase significantly. Instead, it decreases in many migration modes. The classification accuracy of the WMGRNMM model is better than that of the WRN18MM and WRN34MM models. The average classification accuracy of the WRN18MM model is 93.51%, which is better than the average classification accuracy of the WRN34MM model of 92.41%. Because the WPT-TFF size of the bearing vibration signal is small, and the state features are scattered in the time-frequency diagram, as the network depth increases, the scattered feature map of WP-TFF in the network model may cause a large amount of feature information redundancy, thereby affecting the classification accuracy of the model.

[0111] Table 5

[0112]

[0113]

[0114] In order to better compare the accuracy of each model with the WMGRNMM model in migration conditions, the Figure 6 .

[0115] Among them, Figure (a) shows various migration modes of the data set 0HP, Figure (b) shows various migration modes of the data set 1HP, Figure (c) shows various migration modes of the data set 2HP, and Figure (d) shows various migration modes of the data set 3HP.

[0116] It can be concluded from the figure that because the WMGRN model does not use the MK-MMD loss function to calculate the weight factor between the source domain and the target domain to reduce the distribution difference between the two domains, the bearing fault migration classification accuracy of the WMGRN model is the lowest. It can be seen that MK-MMD has greatly improved the accuracy of the migration fault diagnosis model.

[0117] Figure 7(a)-Figure 7(l) shows the change of recognition accuracy of each model state under various migration modes with the number of iterations.

[0118] Figure 7(a) shows the variation of the recognition accuracy of each model state in the 0->1 migration mode with the number of iterations; Figure 7(b) shows the variation of the recognition accuracy of each model state in the 0->2 migration mode with the number of iterations; Figure 7(c) shows the variation of the recognition accuracy of each model state in the 0->3 migration mode with the number of iterations; Figure 7(d) shows the variation of the recognition accuracy of each model state in the 1->0 migration mode with the number of iterations; Figure 7(e) shows the variation of the recognition accuracy of each model state in the 1->2 migration mode with the number of iterations; Figure 7(f) shows the variation of the recognition accuracy of each model state in the 1->3 migration mode with the number of iterations; Figure 7(g) shows the variation of the recognition accuracy of each model state in the 2->0 migration mode with the number of iterations; Figure 7(h) shows the variation of the recognition accuracy of each model state in the 2->1 migration mode with the number of iterations; Figure 7(i) shows the variation of the recognition accuracy of each model state in the 2->3 migration mode with the number of iterations; Figure 7(j) shows the variation of the recognition accuracy of each model state in the 3->0 migration mode with the number of iterations; Figure 7(k) shows the variation of the recognition accuracy of each model state in the 3->1 migration mode with the number of iterations; Figure 7(l) shows the variation of the recognition accuracy of each model state in the 3->2 migration mode with the number of iterations;

[0119] It can be found that all models can converge after 4200 iterations. Generally, the recognition accuracy tends to be stable after 2000 iterations and reaches the best level. Overall, the proposed WMGRNMM model performs better than other models in various migration modes, followed by WRNMM. The results show that the WMGRNMM model has good stability. The results of the WRN18MM and WRN34MM models are better than the WRNMM model in some modes, but they are prone to overfitting problems caused by the deepening of the network depth, and the overall effect is worse than the WRNMM model.

[0120] In order to compare the adaptive capabilities of deep features extracted by different models, we use the t-SNE method to analyze the deep features extracted by each model, and use the confusion matrix of sample labels and predicted labels to analyze the characteristics of state recognition results.

[0121] Figure 8(a)-Figure 8(e) It is the t-SNE feature analysis diagram of each diagnostic model under the 3-1 migration mode and the confusion matrix of the diagnostic results.

[0122] Among them, Figure 8 (a) is a t-SNE feature analysis diagram and a confusion matrix of diagnostic results of the WMGRN model under the 3-1 migration mode of the present invention;

[0123] FIG8( b ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WMGRNMM model under the 3-1 migration mode of the present invention;

[0124] FIG8( c ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WRNMM model under the 3-1 migration mode of the present invention;

[0125] FIG8( d ) is a t-SNE feature analysis diagram and a confusion matrix of diagnosis results of the WRN18MM model under the 3-1 migration mode of the present invention;

[0126] Figure 8(e) is the t-SNE feature analysis diagram and diagnosis result confusion matrix of the WRN34MM model under the 3-1 migration mode of the present invention.

[0127] The t-SNE graph shows that compared with WMGRN, after the introduction of MK-MMD loss, the deep features extracted by the WMGRNMM model have better category separability and improve the distance between signal samples of different categories. The confusion matrix shows that the types and number of misclassifications are significantly reduced. Compared with the WRNMM, WRN18MM and WRN34MM models, the deep features of the WMGRNMM model also have better category separability.

[0128] In the fourth group of experiments, the WMGRNMM model was used to diagnose and analyze the data with various signal-to-noise ratios, and the comparison results with no noise are shown in Table 6. It can be seen from Table 6 that for 8dB and 10dB noise data, the accuracy of the WMGRNMM model did not drop significantly. For 5dB noise data, the diagnostic accuracy of the model dropped slightly, but it was still able to maintain an average accuracy of more than 97%. For 3dB noise data, the accuracy of the WMGRNMM model was affected to a certain extent, with an average decrease of 3%. The results of the noise comparison experiment show that when the signal-to-noise ratio is greater than 5db, the accuracy of the WMGRNMM model is less affected by noise, and the WMGRNMM model has good noise resistance.

[0129] Table 6

[0130]

[0131]

[0132] Based on the above experimental results, the following conclusions can be drawn:

[0133] (1) The proposed vibration signal WPT-TFF construction method can effectively extract the time-frequency features reflecting the fault state from the nonlinear and non-stationary vibration signal, with strong anti-noise ability, which is conducive to the subsequent deep feature extraction.

[0134] (2) A ResNet-based deep feature extraction network RNB7 was designed, which can solve the problems of small size and feature dispersion of WPT-TFF. The MGRN network structure can more effectively extract deep features from multi-scale and multi-resolution WPT-TFFs with different layers.

[0135] (3) By introducing MK-MMD to optimize the deep feature extraction network, the diagnostic accuracy of the diagnosis model under variable operating conditions can be greatly improved. The experimental results on the fault test bench show that under the condition of limited training data sample data, WMGRNMM can well solve the problem of operating condition migration diagnosis and has strong noise resistance.

[0136] The present invention proposes a device status diagnosis method based on wavelet packet time-frequency graph and ResNet migration network. The fault identification accuracy and generalization ability of the proposed WMGRNMM method are verified by comparing multiple experiments on the Case Western Reserve University test bench.

[0137] The above content is a further detailed description of the present invention in combination with 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 may make some simple deductions or substitutions, which should be deemed to fall within the scope of protection of the present invention.

Claims

1. A device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network, characterized by: The specific steps include: Step 1, using wavelet packet transform to perform time-frequency analysis on the nonlinear and non-stationary vibration signal, and constructing a wavelet packet decomposition time-frequency feature diagram of the vibration signal; The step 1 is specifically as follows: Step 1.1, performing four-layer, five-layer, and six-layer wavelet packet decomposition on the collected vibration signal of the rotating mechanical equipment, and then obtaining its wavelet packet node coefficients; Step 1.2, using the coefficients of a single wavelet packet leaf node to construct a coefficient matrix, concatenating the coefficient matrices of all leaf nodes in a Z-shaped sequence to establish a wavelet packet decomposition time-frequency feature map of the vibration signal; Step 2, establish a deep feature extraction network based on ResNet, and extract the deep features of the wavelet packet decomposition time-frequency feature map of the vibration signal; Three 7-layer ResNet Blocks are designed to form the ResNet network structure MGRN, which extracts the feature information of four-layer, five-layer, and six-layer WPT-TFF respectively; Step 3: Use MK-MMD to evaluate the distribution difference between the deep features of the source domain and target domain data, and combine it with the classification loss of the source domain sample set to optimize the deep feature extraction network so that the deep features have better cross-domain invariance and fault state differentiation capabilities; MK-MMD is the multi-core maximum mean error; The step 3 is as follows: Step 3.1: Send the labeled source domain data and the unlabeled target domain data to the parallel network MGRN in step 2 respectively to obtain the deep features of the source domain and target domain datasets, denoted as F s and F t , calculate F s and F t The MK-MMD loss between MK-MMD , as shown in formula (1), Where P is the marginal probability distribution of the source domain, Q is the marginal probability distribution of the target domain, Ε represents the mathematical expectation, Φ represents the mapping of the regenerated Hilbert space, and H k represents the kernel k used to reproduce the Hilbert space; At the same time, F s After the fully connected layer and the Softmax classifier, the predicted category information is obtained, and the cross entropy loss is calculated by combining the label information, which is recorded as L c , as shown in formula (2), Where: M is the number of categories; is a symbolic function, and is 0 or 1; p ic is the predicted probability that sample i belongs to category c; Calculate the overall loss L of the migration classification model, as shown in formula (3): L=L c +λ MK-MMD L MK-MMD (F s ,F t ) (3) In the formula, λ MK-MMD is the trade-off parameter of the total loss; F s is the source domain data; F t is the target domain data; Step 3.2, optimize the deep network and fully connected network parameters using back propagation and gradient descent methods; Step 3.3: In the test phase, the deep features of the target domain sample time-frequency graph are extracted using the trained MGRN network, and the deep features are sent to the trained fully connected layer FC1 to obtain the category label of the sample and complete the fault classification.

2. According to claim 1, a device status diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network is characterized in that: The step 2 is specifically as follows: Step 2.1, decompose the vibration signal using 4-layer, 5-layer, and 6-layer wavelet packet decomposition, and construct the corresponding time-frequency feature map WPT-TFF; Step 2.2, establish a deep feature extraction network based on ResNet, and extract the deep features of the wavelet packet decomposition time-frequency feature map of the vibration signal; Among them, the deep feature extraction network based on ResNet is as follows: (1) The WPT-TFF time-frequency feature diagram of the vibration signal is small in size, and the state characteristics are scattered in the time-frequency diagram: For the small-sized time-frequency feature map WPT-TFF, 3*3 convolution kernels are used in the ResNet network for feature extraction, the convolution step is set to 1, and the pooling operation of the middle layer is cancelled. At the same time, the depth of the ResNet network is limited, and a 7-layer residual network deep feature extraction module ResNet Block is designed to reduce the impact of small image size and feature dispersion on the generalization ability of the model. (2) The time-frequency feature maps obtained by wavelet packet decomposition at different levels have their own structural characteristics: Three 7-layer ResNet Blocks with the same structure and independent parameters can more effectively extract personalized structural features of different types of time-frequency feature maps WPT-TFF. The deep features extracted by the three 7-layer ResNet Blocks are concatenated into combined deep features and sent to the fully connected layer, and finally enter the Softmax classifier for classification.

Images