Rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation

Abstract

The application discloses a rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic field self-adaption, which comprises the following steps: (1) collecting vibration signals of rolling bearings under different equipment, different working conditions and different health states, and constructing a migration diagnosis task from a source field to a target field; (2) calculating a spectral coherence function of the vibration signals, and generating spectral coherence graph data sets of the source field and the target field; (3) constructing a deep multi-representation residual convolutional neural network model; (4) inputting labeled data of the source field and unlabeled data of the target field into the network model at the same time, and performing gradient back propagation training by using a source field cross-entropy loss and a multi-representation dynamic distribution difference loss function to update model parameters; and (5) after the model training is completed, the model is used for a cross-working condition and cross-equipment migration diagnosis task. The application can achieve a higher classification accuracy in the cross-working condition and cross-equipment migration diagnosis task of the rolling bearing.

Description

Technical Field

[0001] This invention belongs to the field of rolling bearing fault diagnosis, and in particular relates to a rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation. Background Technology

[0002] As the most commonly used but easily damaged key component in rotating machinery, the accurate, timely and automatic diagnosis of the health status of rolling bearings is of great engineering significance.

[0003] Deep learning-based intelligent diagnostic methods for rolling bearings have achieved high diagnostic accuracy in some scenarios, but they typically require two prerequisites: (1) sufficient and labeled training data; and (2) test data and training data following the same distribution. However, these conditions are difficult to meet in actual industrial settings, limiting their further application. Domain-adaptive diagnostic methods address the issues of unlabeled actual test data (target domain) and labeled data from similar but different domains (source domain), where the distributions of the two domains differ. By employing probability distribution alignment, diagnostic knowledge from the source domain is transferred and applied to the classification and diagnosis of the target domain. Therefore, deep domain-adaptive diagnostic methods, by combining the advantages of deep learning and domain adaptation, further improve the accuracy of intelligent diagnostic methods for rolling bearings.

[0004] For example, Chinese patent document CN117451360A discloses a method for diagnosing small-sample rolling bearing faults under varying working conditions based on a self-attention mechanism. This method improves the stability and robustness of traditional deep learning methods and incorporates a self-attention mechanism into the original domain adaptive method.

[0005] Chinese patent document CN116028876A discloses a rolling bearing fault diagnosis method based on transfer learning. First, a residual network is used to extract deep features of source domain data and target domain data. Second, the domain adaptation method in transfer learning is applied to calculate the local maximum mean difference between labeled source domain data and unlabeled target domain data. This difference and the classification error of the source domain samples are used as loss functions, and the backpropagation algorithm is used to train the network.

[0006] However, existing diagnostic methods based on deep domain adaptation have low diagnostic accuracy in some scenarios, such as low signal-to-noise ratio and complex working conditions, and have the following limitations: (1) directly using time domain signals as input to the deep network model, lacking signal preprocessing before input; (2) the feature extraction module in the deep network model has a single structure, and the features that can be extracted are not comprehensive enough; (3) in probability distribution alignment, the relative importance of marginal distribution and conditional distribution is ignored and dynamic adjustment is lacking. Summary of the Invention

[0007] This invention provides a rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation. By combining a deep multi-representation residual convolutional neural network based on cyclic spectral coherence graphs with a mechanism for dynamically aligning edges and conditional distributions, it can improve the classification accuracy of rolling bearing migration diagnosis even when the target domain data lacks labels.

[0008] A rolling bearing migration diagnostic method based on cyclic stationary analysis and multi-representation dynamic domain adaptation includes the following steps:

[0009] (1) Collect vibration signals of rolling bearings under different equipment, different working conditions and different health conditions, and construct a migration diagnosis task from the source domain to the target domain.

[0010] (2) Calculate the spectral coherence function of the vibration signal and generate a dataset of spectral coherence maps of the source and target domains;

[0011] (3) Construct a deep multi-representation residual convolutional neural network model and its built-in multi-representation dynamic domain adaptive method;

[0012] (4) Input the labeled data from the source domain and the unlabeled data from the target domain into the network model constructed in step (3) at the same time, and use the source domain cross-entropy loss and the multi-representation dynamic distribution difference loss function to perform gradient backpropagation training to update the model parameters.

[0013] (5) After multiple iterations of training, the loss function tends to stabilize and the model training is completed. The target domain data is then input into the model for testing, which is used for cross-working conditions and cross-equipment migration diagnosis tasks.

[0014] In step (2), the spectral coherence function of the vibration signal is calculated. The specific process is as follows:

[0015] (2-1) Calculate the time-varying autocorrelation function R of the vibration signal x(t). x (t,τ):

[0016]

[0017] In the formula, t is time; τ is time delay; * is the conjugate symbol; E[·] is the statistical averaging operator;

[0018] (2-2) Calculate the cyclic autocorrelation function R of the vibration signal x(t). x (α,τ):

[0019]

[0020] In the formula, T is the period; α is the cycle frequency; e is the natural constant; and j is the imaginary number.

[0021] (2-3) Calculate the spectral correlation function S of the vibration signal x(t). x (α,f):

[0022]

[0023] In the formula, f is the spectral frequency, i.e., the carrier frequency;

[0024] (2-4) Calculate the spectral coherence function γ of the vibration signal x(t). x (α,f):

[0025]

[0026] In step (3), the constructed deep multi-representation residual convolutional neural network model includes three parts: feature extraction module, classification module, and domain adaptation module;

[0027] The feature extraction module consists of a deep residual convolution module and a multi-representation module. The deep residual convolution module uses the backbone of the ResNet-50 network and includes five stages. Stage 1 includes a two-dimensional convolutional layer, a batch normalization (BN) layer, a ReLU activation function, and a max pooling layer. Stages 2-5 each consist of a residual block I and several residual blocks II. The multi-representation module uses multiple substructures to extract feature maps from the low-pixel image output by the backbone of the ResNet-50 network and uses global average pooling to generate multi-representation sub-vectors. In the multi-representation module, each substructure consists of two two-dimensional convolutional layers and a global average pooling layer, and the convolutional kernel sizes of different substructures are different.

[0028] The classification module includes a concatenation operation that merges multiple representation subvectors into a single vector, a fully connected layer, and a softmax layer; during the network training phase, it calculates the cross-entropy loss function of the source domain, and during the testing phase, it outputs diagnostic results.

[0029] The domain adaptation module calculates the average value of the probability distribution difference between the source domain and the target domain of each sub-vector as the loss function of the domain adaptation part.

[0030] In step (4), gradient backpropagation training is performed using the source domain cross-entropy loss and the multi-representation dynamic distribution difference loss function to update the model parameters, specifically including:

[0031] (4-1) Concatenate the multi-representation subvectors generated by the multi-representation module into a single vector V in sequence:

[0032] V = [M1(R(X)), M i (R(X)),...,M n (R(X))]

[0033] In the formula, X represents the spectral coherence map of the source or target domain of the input depth residual convolutional neural network model; R(·) represents the depth residual convolution module, which transforms high-pixel images into low-pixel images; M i (·) represents the i-th substructure in the multi-representation module, which transforms the low-pixel image into the i-th multi-representation subvector; n represents the number of substructures in the multi-representation module.

[0034] (4-2) Input vector V into the fully connected layer and softmax layer in the classification module to obtain the output probability distribution P:

[0035] P=softmax(linear(V))=[p1,p c ,...,p C ]

[0036] In the formula, linear(·) represents a fully connected layer; softmax(·) represents a softmax layer; p c This represents the softmax value of the c-th health state of the rolling bearing; C represents the number of health state categories of the rolling bearing.

[0037] (4-3) Calculate the source domain data cross-entropy loss function L c :

[0038]

[0039] In the formula, n s y represents the number of data in the source domain. i p is the label for the i-th source domain data; c This represents the softmax value of the c-th health state of the rolling bearing; I(·) is an indicator function, when y i When y = c, I = 1, when y = c i When ≠c, I = 0.

[0040] (4-4) For a given subvector generated by the multi-representation module, the marginal probability distribution difference (MKMMD) between the source and target domains is calculated using the multi-kernel maximum mean discrepancy (MKMMD) metric. m :

[0041]

[0042] In the formula, MMD mt The marginal distribution MMD corresponding to the t-th kernel function is represented by q; q represents the number of kernel functions; n s and n t These represent the number of data points in the source and target domains, respectively. For the i-th source domain data; Let be the j-th target domain data; φ(·) is the mapping function corresponding to the t-th kernel function; For the regenerated nucleus Hilbert space.

[0043] (4-5) For a given subvector generated by the multi-representation module, use the MKMMD index to calculate the difference in conditional probability distributions between the source and target domains (MKMMD). c :

[0044]

[0045] In the formula, MMD ct Let represent the conditional distribution MMD corresponding to the t-th kernel function; q represents the number of kernel functions; and Let n represent the data sets in the source and target domains that belong to the c-th health state of the rolling bearing, respectively. c and m c They represent and The amount of data in the data.

[0046] (4-6) For a given subvector generated by the multi-representation module, calculate the dynamic factor μ used to measure the relative importance of marginal and conditional distribution MMD:

[0047] d m =2(1-2ε(h))

[0048] d c =2(1-2ε(h)) (c) )

[0049]

[0050] In the formula, d m and d c Marginal and conditional probability distributions, respectively Distance; ε(h) is the classification error of the linear classifier h when distinguishing between the source and target domain data; ε(h) (c) Let be the classification error of the linear classifier h when it distinguishes between the source and target domain data of the rolling bearing in the c-th health state; C represents the number of categories of the rolling bearing health state.

[0051] (4-7) For a given subvector generated by the multi-representation module, calculate the dynamic distribution difference loss function MKMMD. μ :

[0052] MKMMD μ =μMKMMD m +(1-μ)MKMMD c

[0053] (4-8) Steps (4-4) to (4-7) only yielded the dynamic distribution difference loss function MKMMD corresponding to a certain substructure in the multi-representation module. μi The loss function L represents the dynamic distribution difference. d It can be obtained by summing the dynamic distribution difference loss functions of all substructures:

[0054]

[0055] (4-9) Calculate the total loss function L during network training, including the cross-entropy loss function L of the source domain data. c And multiple representations of the dynamic distribution difference loss function L d The specific formula is as follows:

[0056] L = L c +λL d

[0057] In the formula, λ is a trade-off parameter used to control the loss function L of the multi-representation dynamic distribution difference. d The relative importance of.

[0058] (4-10) Perform gradient backpropagation on the network based on the total loss function, and update the network parameters using the mini-batch stochastic gradient descent algorithm.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention proposes a preprocessing method for rolling bearing vibration signals based on cyclic stationary analysis. It reveals the distribution of modulation frequency and carrier frequency of the fault signal through a bispectral diagram, reduces the components of interference signals, and facilitates further feature extraction in subsequent deep network models.

[0061] 2. This invention proposes a deep multi-representation residual convolutional neural network model. Through the multi-representation module, features can be extracted from multiple perspectives, making the feature representation more comprehensive.

[0062] 3. The dynamic edge alignment and conditional distribution mechanism proposed in this invention can dynamically adjust the relative importance between edge distribution alignment and conditional distribution alignment, further reducing the distribution difference between the source domain and the target domain and improving the accuracy of migration diagnosis classification. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating a rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to the present invention.

[0064] Figure 2This is a schematic diagram of the deep multi-representation residual convolutional neural network model and the multi-representation dynamic domain adaptation method contained in this embodiment of the invention;

[0065] Figure 3 This is a schematic diagram of the structure of residual blocks I and II in an embodiment of the present invention;

[0066] Figure 4 The time-domain diagrams and spectral coherence diagrams are shown for four different health states of rolling bearings in two different devices in this embodiment of the invention.

[0067] Figure 5 This is a schematic diagram of the rotor bearing test bench in an example of the present invention;

[0068] Figure 6 This is a schematic diagram of the centrifugal pump set test bench in an example of the present invention;

[0069] Figure 7 This is a confusion matrix diagram of the three comparison methods in the examples of this invention and the method of this invention. Detailed Implementation

[0070] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0071] like Figure 1 As shown, a rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation includes the following steps:

[0072] S01: Collect vibration signals of rolling bearings from different equipment, under different working conditions, and in different health states to construct a migration diagnostic task from the source domain to the target domain.

[0073] S02, calculate the spectral coherence function of the vibration signal and generate a dataset of spectral coherence maps of the source and target domains;

[0074] (2-1) Calculate the time-varying autocorrelation function R of the vibration signal x(t). x (t,τ):

[0075]

[0076] In the formula, t is time; τ is time delay; * is the conjugate symbol; E[·] is the statistical averaging operator;

[0077] (2-2) Calculate the cyclic autocorrelation function R of the vibration signal x(t). x (α,τ):

[0078]

[0079] In the formula, T is the period; α is the cycle frequency; e is the natural constant; and j is the imaginary number.

[0080] (2-3) Calculate the spectral correlation function S of the vibration signal x(t). x (α,f):

[0081]

[0082] In the formula, f is the spectral frequency (carrier frequency);

[0083] (2-4) Calculate the spectral coherence function γ of the vibration signal x(t). x (α,f):

[0084]

[0085] S03, Construct a deep multi-representation residual convolutional neural network model and its built-in multi-representation dynamic domain adaptive method;

[0086] In S03, such as Figure 2 As shown, the constructed deep multi-representation residual convolutional neural network model consists of three parts: a feature extraction module, a classification module, and a domain adaptation module.

[0087] The feature extraction module consists of a deep residual convolution module and a multi-representation module.

[0088] The deep residual convolution module adopts the backbone of the common ResNet-50 network and consists of 5 stages. Stage 1 includes a 2D convolutional layer, a Batch Normalization (BN) layer, a ReLU activation function, and a max pooling layer. Stages 2 through 5 each consist of a residual block I and several residual blocks II. The structures of residual blocks I and II are as follows: Figure 3 As shown.

[0089] The multi-representation module employs multiple substructures to extract feature maps from the low-resolution image output by the backbone of the ResNet-50 network, and generates multi-representation sub-vectors using global average pooling. Each substructure in the multi-representation module consists of two 2D convolutional layers and one global average pooling layer, with different kernel sizes for different substructures.

[0090] The classification module includes a concatenation operation that merges multiple representation subvectors into a single vector, a fully connected layer, and a softmax layer. During network training, the cross-entropy loss function of the source domain is calculated; during testing, diagnostic results are output.

[0091] The domain adaptation module calculates the average value of the probability distribution difference between the source domain and the target domain of each sub-vector as the loss function of the domain adaptation part.

[0092] S04: Input both labeled data from the source domain and unlabeled data from the target domain into the network model constructed in S03. Use the source domain cross-entropy loss and the multi-representation dynamic distribution difference loss function to perform gradient backpropagation training to update the model parameters.

[0093] (4-1) Concatenate the multi-representation subvectors generated by the multi-representation module into a single vector V in sequence:

[0094] V = [M1(R(X)), M i (R(X)),...,M n (R(X))]

[0095] In the formula, X represents the spectral coherence map of the source or target domain of the input depth residual convolutional neural network model; R(·) represents the depth residual convolution module, which transforms high-pixel images into low-pixel images; M i (·) represents the i-th substructure in the multi-representation module, which transforms the low-pixel image into the i-th multi-representation subvector; n represents the number of substructures in the multi-representation module.

[0096] (4-2) Input vector V into the fully connected layer and softmax layer in the classification module to obtain the output probability distribution P:

[0097] P=softmax(linear(V))=[p1,p c ,...,p C ]

[0098] In the formula, linear(W) represents a fully connected layer; softmax(W) represents a softmax layer; p c This represents the softmax value of the c-th health state of the rolling bearing; C represents the number of health state categories of the rolling bearing.

[0099] (4-3) Calculate the source domain data cross-entropy loss function L c :

[0100]

[0101] In the formula, n s y represents the number of data in the source domain. i p is the label for the i-th source domain data; c This represents the softmax value of the c-th health state of the rolling bearing; I(·) is an indicator function, when y i When y = c, I = 1, when y = c i When ≠c, I = 0.

[0102] (4-4) For a given subvector generated by the multi-representation module, the marginal probability distribution difference (MKMMD) between the source and target domains is calculated using the multi-kernel maximum mean discrepancy (MKMMD) metric. m :

[0103]

[0104] In the formula, MMD mt The marginal distribution MMD corresponding to the t-th kernel function is represented by q; q represents the number of kernel functions; n s and n t These represent the number of data points in the source and target domains, respectively. For the i-th source domain data; Let be the j-th target domain data; φ(·) is the mapping function corresponding to the t-th kernel function; For the regenerated nucleus Hilbert space.

[0105] (4-5) For a given subvector generated by the multi-representation module, use the MKMMD index to calculate the difference in conditional probability distributions between the source and target domains (MKMMD). c :

[0106]

[0107] In the formula, MMD ct Let represent the conditional distribution MMD corresponding to the t-th kernel function; q represents the number of kernel functions; and Let n represent the data sets in the source and target domains that belong to the c-th health state of the rolling bearing, respectively. c and m c They represent and The amount of data in the data.

[0108] (4-6) For a given subvector generated by the multi-representation module, calculate the dynamic factor μ used to measure the relative importance of marginal and conditional distribution MMD:

[0109] d m =2(1-2ε(h))

[0110] d c =2(1-2ε(h)) (c) )

[0111]

[0112] In the formula, d m and d c Marginal and conditional probability distributions, respectively Distance; ε(h) is the classification error of the linear classifier h when distinguishing between the source and target domain data; ε(h) (c) Let be the classification error of the linear classifier h when it distinguishes between the source and target domain data of the rolling bearing in the c-th health state; C represents the number of categories of the rolling bearing health state.

[0113] (4-7) For a given subvector generated by the multi-representation module, calculate the dynamic distribution difference loss function MKMMD. μ :

[0114] MKMMD μ =μMKMMD m +(1-μ)MKMMD c

[0115] (4-8) Steps (4-4) to (4-7) only yielded the dynamic distribution difference loss function MKMMD corresponding to a certain substructure in the multi-representation module. μi The loss function L represents the dynamic distribution difference. d It can be obtained by summing the dynamic distribution difference loss functions of all substructures:

[0116]

[0117] (4-9) Calculate the total loss function L during network training, including the cross-entropy loss function L of the source domain data. c The multiple representations of the dynamic distribution difference loss function L d The specific formula is as follows:

[0118] L = L c +λL d

[0119] In the formula, λ is a trade-off parameter used to control the loss function L of the multi-representation dynamic distribution difference. d The relative importance of.

[0120] (4-10) Perform gradient backpropagation on the network based on the total loss function, and update the network parameters using the mini-batch stochastic gradient descent algorithm.

[0121] S05, after multiple iterations of training, the loss function tends to stabilize, the model training is completed, and the target domain data is input into the model for testing, used for transfer diagnostic tasks across working conditions and equipment.

[0122] To verify the effectiveness of this invention, the cross-operating condition and cross-equipment migration diagnostic task of rolling bearings was analyzed, as follows:

[0123] Vibration data of rolling bearings were collected on two separate test benches: a rotor bearing test bench and a centrifugal pump test bench. Schematic diagrams of the two test benches are shown below. Figure 5 and Figure 6 As shown in the figure. Rolling bearings in different health states were replaced in the test bench to achieve data acquisition for different health states. There were four different health states: healthy state (NC), inner ring fault (IF), outer ring fault (OF), and roller fault (BF). Faults were manually injected: wire cutting was used to cut the bearing raceway. Data acquisition for the rotor bearing test bench was conducted under three different operating conditions (speeds): 2100 rpm, 2250 rpm, and 2400 rpm; the operating conditions for the centrifugal pump set test bench were 2100 rpm and 2400 rpm. The rolling bearing model used was SKF 6205-2RSH. The sampling frequency for vibration data acquisition was 25.6 kHz, and the sampling time for each sample was 0.5 s. A total of 400 samples were collected for each fault under different operating conditions for different equipment. The data acquisition information for the two equipment is shown in Table 1.

[0124] Table 1

[0125]

[0126] There are eight types of cross-operating-condition migration tasks: A→B, A→C, B→A, B→C, C→A, C→B, E→F, and F→E; and eight types of cross-device migration tasks: A→E, A→F, B→E, B→F, E→B, E→C, F→B, and F→C. Taking A→B as an example: A represents the source domain dataset, and B represents the target domain dataset to be diagnosed.

[0127] The vibration time-domain signal was converted into a spectral coherence map using a cyclic stationary analysis method, and a spectral coherence map dataset was constructed accordingly. The time-domain maps and spectral coherence maps corresponding to four different health states of rolling bearings from two different devices are shown below. Figure 4 As shown.

[0128] To fully demonstrate the superiority of this invention, five existing classical methods (T-ResNet, T-DTCA, T-DAN, T-DSAN, T-DAAN) and one variant of this invention (T-MDAN) are compared with the method of this invention (C-MDAN) in comparative experiments. Note: "T" and "C" represent that the network input is a time-domain graph and the spectral coherence graph used in this invention, respectively; "T-ResNet" represents a network input of a time-domain graph and uses a residual network structure but does not use a domain adaptation method, i.e., it does not contain a transfer learning strategy; "T-DTCA" represents a method that automatically extracts time-domain graph features using a residual network and then performs traditional TCA; "T-DAN", "T-DSAN", and "T-DAAN" are three classical deep domain adaptation methods with a time-domain graph as their network input; "T-MDAN" represents a method whose structure is completely identical to the method of this invention except that the network input is a time-domain graph.

[0129] The classification accuracy rates for cross-operating condition and cross-equipment migration diagnostic tasks are shown in Tables 2 and 3, respectively. The proposed method (C-MDAN) achieved the highest classification accuracy in all cross-operating condition migration tasks, with an average accuracy of 98.39%. The proposed method (C-MDAN) also achieved the highest classification accuracy in the vast majority of cross-equipment migration tasks, with an average accuracy of 91.32%. This demonstrates the effectiveness of the proposed method's use of transfer learning and cyclic stationarity analysis as a signal preprocessing method, and also proves the superiority of the deep multi-representation residual convolutional neural network model in the proposed method compared to existing deep domain adaptive methods.

[0130] Table 2 Classification accuracy of cross-condition migration diagnostic tasks

[0131]

[0132]

[0133] Table 3 Classification accuracy of cross-device migration diagnostic tasks

[0134] Task T-ResNet T-DTCA T-DAN T-DSAN T-DAAN T-MDAN C-MDAN A→E 63.06 65.88 73.98 91.34 85.36 77.09 97.11 A→F 25.00 6.00 27.90 26.40 35.25 27.13 83.21 B→E 50.56 63.81 81.26 97.08 54.14 92.40 83.16 B→F 25.00 1.31 25.34 35.60 11.35 27.38 79.06 E→B 27.19 27.00 60.39 52.75 45.43 62.35 96.56 E→C 25.00 25.00 23.79 16.91 16.83 18.52 95.39 F→B 25.00 17.19 14.36 45.14 24.99 40.19 98.75 F→C 24.81 37.25 37.13 28.79 24.79 35.40 97.29 average value 33.20 30.43 43.02 49.25 37.27 47.56 91.32

[0135] To further demonstrate the superiority of the deep multi-representation residual convolutional neural network model and its inherent multi-representation dynamic domain adaptation method proposed in this invention, the input in the comparison methods also uses the spectral coherence diagram of this invention. That is, the comparison methods (C-ResNet, C-DTCA, C-DAN, C-DSAN, C-DAAN) are compared with the method of this invention (C-MDAN). The classification accuracy of the cross-device migration diagnostic task is shown in Table 4. The method of this invention (C-MDAN) achieved the highest classification accuracy in most cross-device migration tasks, with an average accuracy of 91.32%. Taking task E→C as an example... Figure 7 The confusion matrix of the classification results of three comparative methods (C-DAN, C-DSAN, and C-DAAN) and the method of this invention (C-MDAN) is presented. Overall, the method of this invention achieves the highest classification accuracy for the four health states of rolling bearings, reaching 95.39%. This demonstrates the superiority of the deep multi-representation residual convolutional neural network model of this invention and its inherent multi-representation dynamic domain adaptive method. Furthermore, comparing the results in Tables 3 and 4 shows that the classification accuracy of the comparative methods is significantly improved after the input is replaced with a spectral coherence map, proving the effectiveness of using cyclostationary analysis as a signal preprocessing method in this invention.

[0136] Table 4. Classification accuracy of cross-device migration diagnostic tasks (inputs are all spectral coherence maps)

[0137] Task C-ResNet C-DTCA C-DAN C-DSAN C-DAAN C-MDAN A→E 71.00 89.00 96.43 96.41 96.03 97.11 A→F 79.31 76.31 81.26 81.36 81.66 83.21 B→E 64.44 68.06 82.76 84.78 78.24 83.16 B→F 60.38 71.50 76.76 79.04 74.93 79.06 E→B 51.31 56.12 95.45 94.40 92.85 96.56 E→C 64.31 60.88 93.43 91.33 93.81 95.39 F→B 63.75 75.31 98.67 97.99 98.69 98.75 F→C 64.63 82.94 97.20 96.55 98.66 97.29 average value 64.89 72.51 90.25 90.23 89.36 91.32

[0138] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing the migration of rolling bearings based on cyclic stationary analysis and multi-representation dynamic domain adaptation, characterized in that, Includes the following steps: (1) Collect vibration signals of rolling bearings under different equipment, different working conditions and different health conditions, and construct a migration diagnosis task from the source domain to the target domain; (2) Calculate the spectral coherence function of the vibration signal and generate a dataset of spectral coherence maps of the source and target domains; (3) Construct a deep multi-representation residual convolutional neural network model and its built-in multi-representation dynamic domain adaptation method; the constructed deep multi-representation residual convolutional neural network model includes three parts: feature extraction module, classification module and domain adaptation module; The feature extraction module consists of a deep residual convolution module and a multi-representation module. The deep residual convolution module uses the backbone of the ResNet-50 network and includes five stages. Stage 1 includes a two-dimensional convolutional layer, a batch normalization (BN) layer, a ReLU activation function, and a max pooling layer. Stages 2-5 each consist of a residual block I and several residual blocks II. The multi-representation module uses multiple substructures to extract feature maps from the low-pixel image output by the backbone of the ResNet-50 network and uses global average pooling to generate multi-representation sub-vectors. In the multi-representation module, each substructure consists of two two-dimensional convolutional layers and a global average pooling layer, and the convolutional kernel sizes of different substructures are different. The classification module includes a concatenation operation that merges multiple representation subvectors into a single vector, a fully connected layer, and a softmax layer; during the network training phase, it calculates the cross-entropy loss function of the source domain, and during the testing phase, it outputs diagnostic results. The domain adaptation module uses the average value of the probability distribution difference between the source domain and the target domain of each sub-vector as the loss function of the domain adaptation part. (4) Input the labeled data from the source domain and the unlabeled data from the target domain into the network model constructed in step (3) simultaneously, and use the source domain cross-entropy loss and the multi-representation dynamic distribution difference loss function to perform gradient backpropagation training to update the model parameters; specifically including: (4-1) Concatenate the multi-representation sub-vectors generated by the multi-representation module into a single vector in sequence. ; (4-2) Transform the vector The fully connected layer and softmax layer in the input classification module are used to obtain the output probability distribution. ; (4-3) Calculate the source domain data cross-entropy loss function ; (4-4) For a subvector generated by the multi-representation module, the difference in marginal probability distributions between the source and target domains is calculated using the multi-kernel maximum mean difference (MKMMD) index. : In the formula, Indicates the corresponding number Marginal distribution MMD of kernel functions; Indicates the number of kernel functions; and These represent the number of data points in the source and target domains, respectively. For the first Data from one source domain; For the first Data for each target domain; In order to be with the first The mapping function corresponding to each kernel function; For the regenerating nucleus Hilbert space; (4-5) For a given subvector generated by the multi-representation module, use the MKMMD index to calculate the difference in conditional probability distributions between the source and target domains. : In the formula, Indicates the corresponding number The conditional distribution MMD of each kernel function; Indicates the number of kernel functions; and These represent the first and second elements belonging to the rolling bearing in the source and target domains, respectively. A dataset of health statuses; and They represent and The amount of data in the data; (4-6) For a given subvector generated by the multi-representation module, calculate the dynamic factor used to measure the relative importance of marginal and conditional distributions MMD. : In the formula, and Marginal and conditional probability distributions, respectively distance; For linear classifiers Classification error when distinguishing between source and target domain data; For linear classifiers For rolling bearings The classification error when distinguishing source and target domain data under a healthy state; The number of categories representing the health status of rolling bearings; (4-7) For a given subvector generated by the multi-representation module, calculate the dynamic distribution difference loss function. : (4-8) Steps (4-4) to (4-7) only yielded the dynamic distribution difference loss function corresponding to a certain substructure in the multi-representation module. The loss function represents the dynamic distribution difference. The result is obtained by summing the dynamic distribution difference loss functions of all substructures: (4-9) Calculate the total loss function during network training. This includes the cross-entropy loss function of the source domain data. The sum represents the dynamic distribution difference loss function. The specific formula is as follows: In the formula, To balance the parameters used to control the loss function of dynamic distribution differences across multiple representations. The relative importance; (4-10) Based on the total loss function Gradient backpropagation is performed on the network, and the network parameters are updated using the mini-batch stochastic gradient descent algorithm; (5) After multiple iterations of training, the loss function tends to stabilize and the model training is completed. The target domain data is then input into the model for testing and used for cross-working conditions and cross-equipment transfer diagnostic tasks.

2. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In step (2), the spectral coherence function of the vibration signal is calculated. The specific process is as follows: (2-1) Calculate the vibration signal Time-varying autocorrelation function : In the formula, For time; For time delay; The conjugate symbol; For statistical averaging operators; (2-2) Calculate the vibration signal Cyclic autocorrelation function : In the formula, For periodicity; It is the cycle frequency; It is a natural constant; It is the symbol for imaginary numbers; (2-3) Calculate the vibration signal spectral correlation function : In the formula, This refers to the spectral frequency, i.e., the carrier frequency. (2-4) Calculate the vibration signal spectral coherence function : 。 3. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In step (4-1), the multi-representation sub-vectors generated by the multi-representation module are concatenated into a single vector in sequence. The formula is: In the formula, The input depth represents the source or target domain spectral coherence graph of the residual convolutional neural network model; This indicates a depth residual convolution module that transforms high-resolution images into low-resolution images. This indicates the first in the multiple representation module. The substructure transforms the low-pixel image into the first substructure. Multiple representation subvectors; This indicates the number of substructures in a module.

4. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In step (4-2), the output probability distribution is obtained. The formula is: In the formula, Indicates a fully connected layer; Indicates a softmax layer; Indicates the rolling bearing number The softmax value of a health status; This indicates the number of categories representing the health status of rolling bearings.

5. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In step (4-3), the cross-entropy loss function of the source domain data is calculated. The formula is: In the formula, The number of data in the source domain; For the first Tagging of source domain data; Indicates the rolling bearing number The softmax value of a health status; It is an indicator function, when hour, ,when hour, .

6. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In step (4-4), the difference in marginal probability distributions between the source and target domains is calculated using the multi-kernel maximum mean difference (MKMMD) index. The formula is: In the formula, Indicates the corresponding number Marginal distribution MMD of kernel functions; Indicates the number of kernel functions; and These represent the number of data points in the source and target domains, respectively. For the first Data from one source domain; For the first Data for each target domain; In order to be with the first The mapping function corresponding to each kernel function; For the regenerated nucleus Hilbert space.

7. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In steps (4-5), the MKMMD index is used to calculate the difference in conditional probability distributions between the source and target domains. The formula is: In the formula, Indicates the corresponding number The conditional distribution MMD of each kernel function; Indicates the number of kernel functions; and These represent the first and second elements belonging to the rolling bearing in the source and target domains, respectively. A dataset of health statuses; and They represent and The amount of data in the data.

8. The rolling bearing migration diagnosis method based on cyclic stationary analysis and multi-representation dynamic domain adaptation according to claim 1, characterized in that, In steps (4-6), a dynamic factor is calculated to measure the relative importance of marginal and conditional distributions of MMD. The formula is as follows: In the formula, and Marginal and conditional probability distributions, respectively distance; For linear classifiers Classification error when distinguishing between source and target domain data; For linear classifiers For rolling bearings The classification error when distinguishing source and target domain data under a healthy state; This indicates the number of categories representing the health status of rolling bearings.

Images