A rolling bearing fault diagnosis method based on pseudo-label transfer two-stage domain adaptation

Abstract

This invention discloses a two-stage domain-adaptive rolling bearing fault diagnosis method based on pseudo-label transfer. The invention aims to address the problem of significant data distribution differences between monitoring data of rolling bearings with known health states and those of rolling bearings with unknown health states during rolling bearing migration fault diagnosis, which hinders accurate diagnosis of the bearings to be diagnosed. By using available intermediate data to bridge the source and target data, a single domain-adaptive fault diagnosis process is transformed into a two-stage process. The domain adaptation method gradually reduces the data distribution difference, enabling the shared health state knowledge of the rolling bearings to be gradually transferred to the target data, thereby improving the fault diagnosis accuracy of the rolling bearings to be diagnosed.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and specifically to a rolling bearing fault diagnosis method based on pseudo-label transfer two-stage domain adaptive method. Background Technology

[0002] Rolling bearings operate under complex and variable speed, load, and temperature conditions for extended periods. Failures can severely impact the normal operation of machinery and even lead to injuries or fatalities. Therefore, fault diagnosis of rolling bearings is a pressing issue. The applicable conditions for intelligent fault diagnosis models based on constant operating conditions are: ① sufficient labeled information in the training data; ② the training and test data follow the same distribution. However, physical signal data collected under different operating conditions exhibit significant distributional differences, making fault diagnosis models based on constant operating conditions difficult to diagnose and identify fault data under other conditions. Furthermore, data labeling is costly and time-consuming; in practical engineering, performing fully labeled supervised training for each operating condition is inefficient and inconvenient. Therefore, utilizing existing rolling bearing monitoring data with known health status (source data) to assist in diagnosing rolling bearing monitoring data with unknown health status (target data) has become a practical requirement in current engineering projects.

[0003] In recent years, transfer learning techniques have provided solutions for the fault diagnosis of rolling bearings under different operating conditions. Transfer fault diagnosis can apply known supervised diagnostic knowledge to fault diagnosis and identification in related fields. Feature domain adaptation is one of the commonly used methods in transfer diagnosis. This type of method reduces the distribution difference between source data features and target data features, enabling the diagnostic model learned from the source data to be applied to the target data. However, most existing transfer diagnosis frameworks ignore the diagnostic difficulties caused by excessively large distribution differences between source and target data. When operating conditions change significantly, especially when there are large changes between different rolling bearings, direct transfer diagnosis may result in negative transfer, leading to low diagnostic accuracy for the target rolling bearing data. Summary of the Invention

[0004] The purpose of this invention is to address the problem that the large differences in the distribution of source domain data and target domain data during current transfer diagnosis between different mechanical equipment makes direct transfer diagnosis difficult and results in low diagnostic accuracy. This invention provides a cross-mechanical equipment fault diagnosis method based on a deep transfer learning framework. By utilizing a suitable intermediate domain to form a bridge between the source and target domains, the shared health status knowledge of the mechanical equipment is gradually transferred, thereby improving the diagnostic accuracy of the target domain.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] A rolling bearing fault diagnosis method based on pseudo-label transfer two-stage domain adaptive method, characterized by the following steps:

[0007] Step 1: Collect existing rolling bearing fault data as source data D S The signal data of the rolling bearing to be diagnosed is collected as the target data D. T ;

[0008] Step 2: First, use data similarity measurement methods to calculate the source data D. S With target data D T The similarity of data distribution between them, and then based on the calculation results and selection conditions, in the target data D T Selecting intermediate data D I ;

[0009] Step 3: Construct a rolling bearing fault diagnosis model based on a pseudo-label-transferring two-stage domain adaptive network;

[0010] Step 4: Utilize the source data D described in Steps 1 and 2. S Target data D T and intermediate data D I The model described in step 3 is used to diagnose the faults in the rolling bearing to be diagnosed.

[0011] Preferably, step 1 specifically includes:

[0012] Source data D S This refers to available rolling bearing data with known health status, including fault data under N operating conditions, i.e. in n S Indicates the number of sample items. Let i represent the i-th sample. C represents the corresponding fault label, and C represents the number of health status categories; target data D T This refers to rolling bearing signal data whose health status is unknown and awaits diagnosis, including fault data under M operating conditions. in Let j represent the j-th sample, and n represent the n-th sample. T Indicates the number of samples;

[0013] Preferably, step 2 specifically includes:

[0014] Step A: Use S(D) i D j )∈[0,1] represents data D i With data D j The similarity between them, when S(D) iD j When it is close to 1, it indicates that the data D i With data D j They are very similar, when S(D) i D j When the value is close to 0, it indicates that the data D is... i With data D j The data distributions differ significantly between them. A data similarity metric is constructed: D i and D j They are labeled as 0 and 1 respectively, and then a linear support vector machine classifier h is trained to distinguish D. i and D j The samples in the dataset are used to calculate the loss err(h) for h:

[0015]

[0016] In the formula, I[a] is an indicator function; I[a] is 1 when a is true, and 0 otherwise; x i and x j D represents i and D j Given m and m' samples obtained by sampling from each of the two samples, then S(D) i D j The calculation is as follows:

[0017]

[0018] Step B: Calculate the distribution similarity between source data containing N operating conditions and target data containing M operating conditions using the data similarity measurement method described in Step A. Where n∈{1,...,N} and m∈{1,...,N};

[0019] Step C: Based on the calculation results described in Step B, and in conjunction with the following conditions 1-3, select the intermediate data to be used.

[0020] Condition 1: To select intermediate data from the target data;

[0021] Condition 2: and The similarity between the source data and the intermediate data, and the similarity between the intermediate data and the target data are both greater than the similarity between the source data and the target data.

[0022] Condition 3: Make For intermediate data that satisfy condition 2 When multiple options exist, selection can enable... and The target data with the smallest absolute value of the difference is used as the final intermediate data;

[0023] Preferably, step 3 specifically includes:

[0024] Step A: The rolling bearing fault diagnosis model based on a pseudo-label-transfer two-stage domain adaptive network includes a two-stage network model, namely, the source data D in stage I. S To intermediate data D I The domain-adaptive network model, and the intermediate data D for stage II. I To target data D T A domain-adaptive network model is used, with Stage I and Stage II connected by a pseudo-label constraint method based on threshold decision. First, a deep diagnostic network model for Stage I is constructed and trained. The Stage I network mainly includes a feature extraction module, a classification module, and a domain adaptation module. The feature extraction module consists of convolutional layers, pooling layers, and a global average pooling layer. The mapping function of the feature extraction module is f. g1 (x, θ) g1 The parameter to be trained is θ. g1 The classification module consists of a fully connected layer, a Dropout layer, and a Softmax layer. The mapping function of the classification module is f. c1 (x, θ) c1 The parameter to be trained is θ. c1 The domain adaptation module employs a domain adaptation method based on the maximum mean difference to reduce the feature distribution differences between different data points. The mapping function is d. 2 The training process in Phase I is as follows: The source data sample set with known health status is... and intermediate data sample sets with unknown health status The inputs are fed into the network of stage I to optimize the model parameters for this part. Through the feature extraction module, corresponding source data features are generated from the source data and intermediate data, respectively. and intermediate data features Then, source data features The predicted labels of the source data samples will be obtained through the classification module in Stage I, and the error between the predicted labels and the true labels will be measured by the cross-entropy cost function shown below:

[0025]

[0026] In the formula, v represents the v-th dimension of the classification probability distribution finally output by the classification module. I{·} is an indicator function, which is defined when the following conditions are met. Its value is 1 if it is true, and 0 otherwise. Additionally, source data features... Features of intermediate data The distributional discrepancies are reduced through the domain adaptation module in Phase I. The loss function for this part is expressed as:

[0027]

[0028] Preferably, the domain adaptation method based on the maximum mean difference used in the domain adaptation module specifically includes:

[0029] Use H k Let RKHS represent the reproducing kernel Hilbert space (RKHS) associated with the feature kernel k. Given two distributions P and Q, then RKHSH exists. k There is a nonlinear mapping function φ(·)∈H k Map the data to RKHSH k The maximum mean difference used in the domain adaptation module can be theoretically calculated using the following formula:

[0030]

[0031] In the formula, sup(·) is the supremum of the input set. and Let P and Q represent the expected values ​​of distributions P and Q, respectively. This means that φ(·) is a series of functions within the unit sphere in RKHS. Two sets of samples are obtained independently and identically distributed from distributions P and Q, respectively. and The empirical calculation of the maximum mean difference is as follows:

[0032]

[0033] In the formula, the kernel mapping k(x) p x q )=<φ(x p ), φ(x q In multi-kernel maximum mean difference analysis, the optimal kernel function composed of multiple feature kernels can better approximate the distribution of the feature space. Multiple feature kernels... It can be defined as a convex combination of m cores:

[0034]

[0035] In the formula, the constraint conditions The derived multi-core k is guaranteed to be unique, where u is the number of cores, and β... u The constraint coefficients for the core.

[0036] Intermediate sample features after domain adaptation The classification module will obtain the predicted probability of the rolling bearing health status category for this sample. That is, the output of the last layer of the classification module:

[0037]

[0038] In the formula, For intermediate sample features, θ c1 These are the training parameters for the classification module. Given a one-dimensional probability distribution vector, the label corresponding to its maximum value represents the sample. The predicted health status category is determined. To avoid overfitting, L2 regularized loss is used. If the parameters are trained using a penalty model, then the objective function for stage I is expressed as:

[0039]

[0040] In the formula, α and β are penalty coefficients.

[0041] Step B: Construct a pseudo-label constraint method based on threshold decision-making, by analyzing intermediate data samples. In network models, a threshold is set for the prediction probability. The method is used to select intermediate data samples with high confidence. These selected intermediate data samples The generated pseudo-labels will be used for training in Phase II. It can be represented as:

[0042]

[0043]

[0044] In the formula, τ∈(0,1) is the threshold used to generate pseudo-labels. I{·} is the indicator function, which is satisfied when... hour And the sample The corresponding pseudo-tags will be generated. Otherwise, the sample will not generate a pseudo-label and will not be used in Phase II. When the number of intermediate data samples with pseudo-labels reaches ρ, a proportion of the total number of intermediate data samples used for training, the intermediate data samples with pseudo-labels are used for training in Phase II. Therefore, the set of intermediate data samples with high confidence is represented as follows:

[0045]

[0046] In the formula, ρ∈(0,1) is the scaling parameter used to determine whether to end training in phase I; N I This represents the total number of intermediate data samples used in the training.

[0047] Step C: Construct the deep diagnostic network structure for Stage II. The network structure for Stage II is similar to that for Stage I, also including a feature extraction module, a classification module, and a domain adaptation module. The corresponding mapping function includes f. g2 (x, θ) g2 ) and f c2 (x, θ) c2 The parameters to be trained include θ. g2 and θ c2 The training process in Phase II is as follows: The intermediate data sample set containing pseudo-label information is... and target data sample set with unknown health status The input is fed into the network in Stage II to optimize the parameters of this part of the model. Similar to the training process in Stage I, the objective function of Stage II is expressed as:

[0048]

[0049] In the formula, α and β are penalty coefficients.

[0050] Specifically, step 4 includes:

[0051] The network model parameters for stages I and II are optimized using the Adam backpropagation algorithm. The total number of training iterations for stages I and II is K. After reaching the training iteration count K, unlabeled target data samples are input into the completed network model. The Softmax layer of the classification module outputs the predicted probability distribution vector of the target domain samples, and the label corresponding to the highest probability represents the predicted health status category of the tested sample.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] In this application, compared with the prior art, the present invention introduces the concept of intermediate data, thereby transforming the traditional single migration diagnosis process from source data to target data into a two-stage migration diagnosis process: source data to intermediate data (Stage I) and intermediate data to target data (Stage II). The present invention proposes a data similarity measurement method to calculate the similarity between source data containing multiple operating conditions and target data containing multiple operating conditions. The present invention proposes three conditions that should be met when selecting usable intermediate data, used to select intermediate data from target data containing multiple operating conditions. The present invention uses a method of generating pseudo-labels for intermediate data samples to transfer fault labeling information to assist in diagnosing target data. The present invention proposes a pseudo-label constraint method based on threshold decision to filter intermediate data samples with high confidence, thereby constructing a pseudo-label transfer-based two-stage domain adaptive network and reducing the cumulative error of transfer-based migration. When the distribution differences between source data and target data are significant, the present invention can divide a source data to target data diagnosis process with significantly different data distributions into two diagnosis processes with similar data distributions: source data to intermediate data and intermediate data to target data. This invention proposes a two-stage domain adaptation method based on convolutional neural networks, which fully leverages the strong nonlinear feature mapping capabilities of deep neural networks to effectively improve the extraction of nonlinear features from the physical signals of rolling bearings. This invention avoids the negative transfer problem caused by direct transfer diagnosis from source data to target data, thereby accurately identifying the health status of rolling bearings and effectively improving the diagnostic accuracy of target data. It provides a valuable method for fault diagnosis of rolling bearings under actual industrial conditions. Attached image description:

[0054] Figure 1 This is a schematic diagram illustrating the principle of the two-stage field adaptation of the present invention;

[0055] Figure 2 This is a schematic diagram of the process of the present invention;

[0056] Figure 3 This is a schematic diagram of the pseudo-label-transferring two-stage domain adaptation network model of the present invention;

[0057] Figure 4 This is the result of the distribution similarity calculation between the source data and the target data in this invention;

[0058] Figure 5 This provides detailed information about the source and target bearing datasets for this invention.

[0059] Figure 6 This refers to the data distribution similarity calculation results between different datasets in this invention;

[0060] Figure 7These are the experimental test and diagnostic results of different methods of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0062] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0063] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0064] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0065] In the description of this invention, it should be noted that the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. These terms are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0066] like Figure 1 As shown, the schematic diagram illustrating the principle of this invention includes two parts, a and b. Figure 1 Part a illustrates an applicable scenario where the monitoring data of the rolling bearing to be diagnosed contains intermediate data, and its data distribution lies between the source data and the target data. After adding the intermediate data, a migration diagnostic process with a large distribution difference can be transformed into two migration diagnostic processes with smaller distribution differences, thus demonstrating that adding intermediate data can convert a diagnostic process with high migration difficulty into two diagnostic processes with lower migration difficulty. Figure 1Part b illustrates the principle of transitive two-stage domain adaptation, which transforms a direct transfer diagnostic process with a large domain offset into two transfer diagnostic stages (Stage I and Stage II) with smaller domain offsets. In this way, shared health status knowledge from the source data is progressively transferred to the target data, thereby achieving high-accuracy diagnosis of the target data sample.

[0067] like Figure 2 The diagram illustrates the invented pseudo-tag-transfer two-stage domain adaptation fault diagnosis process, which specifically includes the following steps:

[0068] Step 1: Collect existing rolling bearing fault data as source data D S The signal data of the rolling bearing to be diagnosed is collected as the target data D. T Source data D S This refers to available rolling bearing data with known health status, including fault data under N operating conditions, i.e. in n S Indicates the number of sample items. Let i represent the i-th sample. express The fault label, C represents the number of health status categories; the target data D T This refers to rolling bearing signal data whose health status is unknown and awaits diagnosis, including fault data under M operating conditions. in Let j represent the j-th sample, and n represent the n-th sample. T Indicates the number of samples;

[0069] Step 2: First, use data similarity measurement methods to calculate the source data D. S With target data D T The similarity of data distribution between them, and then based on the calculation results and selection conditions, in the target data D T Selecting intermediate data D I First, using S(D) i D j )∈[0,1) represents the data D i With data D j The similarity between them, when S(D) i D j When it is close to 1, it indicates that the data D i With data D j They are very similar, when S(D) i D j When the value is close to 0, it indicates that the data D is... i With data D jThe data distributions differ significantly between them. A data similarity metric is constructed: D i and D j They are labeled as 0 and 1 respectively, and then a linear support vector machine classifier h is trained to distinguish D. i and D j We use samples from h to estimate the loss err(h):

[0070]

[0071] In the formula, I[a] is an indicator function; I[a] is 1 when a is true, and 0 otherwise; x i and x j D represents i and D j Given m and m' samples obtained by sampling from each of the two samples, then S(D) i D j The calculation is as follows:

[0072]

[0073] Then, the distribution similarity between source data containing N operating conditions and target data containing M operating conditions is calculated using the aforementioned data similarity measurement method. Where n∈{1,...,N} and m∈{1,...,M}, the result is as follows Figure 4 As shown;

[0074] Finally, based on the calculation results described in step B, and in conjunction with the following conditions 1-3, select the available intermediate data.

[0075] Condition 1: That is, selecting intermediate data from the target data;

[0076] Condition 2: and That is, the similarity between the source data and the intermediate data, and the similarity between the intermediate data and the target data are both less than the similarity between the source data and the target data;

[0077] Condition 3: Make That is, when the intermediate data meets condition 2 When multiple options exist, selection can enable... and The target data with the smallest absolute value of the difference is used as the final intermediate data;

[0078] Step 3: Construct a rolling bearing fault diagnosis model based on a pseudo-label-transferring two-stage domain adaptive network. The rolling bearing fault diagnosis model based on the pseudo-label-transferring two-stage domain adaptive network includes a two-stage network model, namely, the source data D in stage I. S To intermediate data D I The domain-adaptive network model, and the intermediate data D for stage II. I To target data D T The domain-adaptive network model uses a threshold-based decision-making pseudo-label constraint method to establish a connection between stage I and stage II. Figure 3 As shown, the model structure of the constructed pseudo-label transitive two-stage domain adaptation network is illustrated.

[0079] First, a deep diagnostic network model for Stage I is constructed and trained. The Stage I network mainly includes a feature extraction module, a classification module, and a domain adaptation module. The feature extraction module consists of convolutional layers, pooling layers, and a global average pooling layer. The mapping function of the feature extraction module is f. g1 (x, θ) g1 The parameter to be trained is θ. g1 The classification module consists of a fully connected layer, a Dropout layer, and a Softmax layer. The mapping function of the classification module is f. c1 (x, θ) c1 The parameter to be trained is θ. c1 The domain adaptation module employs a domain adaptation method based on the maximum mean difference to reduce the feature distribution differences between different data points. The mapping function is d. 2 The training process in Phase I is as follows: The source data sample set with known health status is... and intermediate data sample sets with unknown health status The inputs are fed into the network of stage I to optimize the model parameters for this part. Through the feature extraction module, corresponding source data features are generated from the source data and intermediate data, respectively. and intermediate data features Then, source data features The predicted labels of the source data samples will be obtained through the classification module in Stage I, and the error between the predicted labels and the true labels will be measured by the cross-entropy cost function shown below:

[0080]

[0081] In the formula, v represents the v-th dimension of the classification probability distribution finally output by the classification module. I{·} is an indicator function, which is defined when the following conditions are met. Its value is 1 if it is true, and 0 otherwise. Additionally, source data features... Features of intermediate data The distributional discrepancies are reduced through the domain adaptation module in Phase I. The loss function for this part is expressed as:

[0082]

[0083] Specifically, the domain adaptation method based on the maximum mean difference used in the domain adaptation module includes:

[0084] Use H k Let RKHS represent the reproducing kernel Hilbert space (RKHS) associated with the feature kernel k. Given two distributions P and Q, then RKHSH exists. k There is a nonlinear mapping function φ(·)∈H k Map the data to RKHSH k The maximum mean difference used in the domain adaptation module can be theoretically calculated using the following formula:

[0085]

[0086] In the formula, sup(·) is the supremum of the input set. and Let P and Q represent the expected values ​​of distributions P and Q, respectively. This means that φ(·) is a series of functions within the unit sphere in RKHS. Two sets of samples are obtained independently and identically distributed from distributions P and Q, respectively. and The empirical calculation of the maximum mean difference is as follows:

[0087]

[0088] In the formula, the kernel mapping k(x) p x q )=<φ(x p ), φ(x q In multi-kernel maximum mean difference analysis, the optimal kernel function composed of multiple feature kernels can better approximate the distribution of the feature space. Multiple feature kernels... It can be defined as a convex combination of m cores:

[0089]

[0090] In the formula, the constraint conditions The derived multi-core k is guaranteed to be unique, where u is the number of cores, and β... u The constraint coefficients for the core.

[0091] Intermediate sample features after domain adaptation The classification module will obtain the predicted probability of the rolling bearing health status category for this sample. That is, the output of the last layer of the classification module:

[0092]

[0093] In the formula, For intermediate sample features, θ c1 These are the training parameters for the classification module. Given a one-dimensional probability distribution vector, the label corresponding to its maximum value represents the sample. The predicted health status category is determined. To avoid overfitting, L2 regularized loss is used. If the parameters are trained using a penalty model, then the objective function for stage I is expressed as:

[0094]

[0095] In the formula, α and β are penalty coefficients. The network model parameters for stage I are optimized using the Adam backpropagation algorithm.

[0096] Then, a pseudo-label constraint method based on threshold decision-making is constructed, which filters out intermediate data samples with high confidence by setting a threshold for the predicted probability of samples in the network model. These selected intermediate data samples The generated pseudo-labels will be used for training in Phase II. It can be represented as:

[0097]

[0098]

[0099] In the formula, τ∈(0,1) is the threshold used to generate pseudo-labels. I{·} is the indicator function, which is satisfied when... hour And the sample The corresponding pseudo-tags will be generated. Otherwise, the sample will not generate a pseudo-label and will not be used in Phase II. When the number of intermediate data samples with pseudo-labels reaches ρ, a proportion of the total number of intermediate data samples used for training, the intermediate data samples with pseudo-labels are used for training in Phase II. Therefore, the set of intermediate data samples with high confidence is represented as follows:

[0100]

[0101] In the formula, ρ∈(0,1) is the scaling parameter used to determine whether to end training in phase I; N I This represents the total number of intermediate data samples used in the training.

[0102] Finally, the deep diagnostic network structure for Stage II is constructed. The network structure for Stage II is similar to that of Stage I, also including a feature extraction module, a classification module, and a domain adaptation module. The corresponding mapping function includes f. g2 (x, θ) g2 ) and f c2 (x, θ) c2 The parameters to be trained include θ. g2 and θ c2 The training process in Phase II is as follows: The intermediate data sample set containing pseudo-label information is... and target data sample set with unknown health status The input is fed into the network in Stage II to optimize the parameters of this part of the model. Similar to the training process in Stage I, the objective function of Stage II is expressed as:

[0103]

[0104] In the formula, α and β are penalty coefficients. The network model parameters for stage II are optimized using the Adam backpropagation algorithm.

[0105] Step 4: Utilize the source data D described in Steps 1 and 2 S Target data D T and intermediate data D I The model described in step 3 is used to diagnose faults in the rolling bearing to be diagnosed. The total number of training iterations for stages I and II is K. After reaching the training iteration count K, unlabeled target data samples are input into the completed network model. The Softmax layer of the classification module outputs the predicted probability distribution vector of the target domain samples, and the label corresponding to the highest probability represents the predicted health status category of the tested sample.

[0106] Example: Taking the diagnosis of rolling bearing migration faults between two different test benches under different working conditions as an example, the feasibility of the present invention is verified.

[0107] Dataset P comes from the Bearing Data Center at Case Western Reserve University, using data on failed bearings on the drive side. The bearings are 6205-2RS deep groove ball bearings, and the bearing failures are single-point failures artificially induced by electrical discharge machining (EDM). The damaged diameter of the failed bearings used is 0.178 mm. The bearing health status includes normal condition, outer ring failure, and inner ring failure. Data was used under three operating conditions: motor speeds of 1797, 1772, and 1750 rpm. The sampling frequency was 48 kHz. A sliding window was used to extract samples, obtaining 2000 samples for each health status, with a training data:test data ratio of 7:3. Each sample contains 2048 sampling points.

[0108] Dataset Q comes from the bearing dataset of the University of Paderborn, Germany. The test bearings on this test bench are model 6203 deep groove ball bearings. Real damage data obtained from accelerated life testing are used. The bearing health status includes three states: normal, outer ring failure, and inner ring failure. Different operating conditions can be achieved by changing the rotational speed N, loading torque M, and loading radial force F. This embodiment uses bearing data collected under three operating conditions. The sampling frequency of the collected vibration data is 64kHz. A sliding window is used to extract samples, obtaining 2000 samples for each health status, with a training data:test data ratio of 7:3. Each sample contains 2048 sampling points.

[0109] like Figure 5 As shown, this displays details of the dataset used. Based on Figure 5 First, the distribution similarity between the target rolling bearing's various operating condition datasets and the source rolling bearing's various operating condition datasets is analyzed. The results are as follows: Figure 6 As shown.

[0110] according to Figure 6 Three transfer diagnostic tasks with relatively low data distribution similarity were established: T1:A1→C, T2:A2→C, and T3:A3→C to verify the feasibility of the invention. The aim is to utilize artificially generated knowledge of damaged bearing diagnosis to assist in the diagnosis of real damaged bearing data. Figure 6 Based on intermediate data selection conditions 1 and 2, in transfer tasks T1, T2, and T3, intermediate data is selected from B1 and B2, transforming the single transfer diagnostic process into a two-stage process. Taking transfer task T1: A1→C as an example, if B1 is selected as intermediate data, S(A1,B1) = 0.5467 and S(B1,C) = 0.6029; if B2 is selected, S(A1,B2) = 0.4267 and S(B2,C) = 0.701. Clearly, in transfer task T1, B1 better satisfies intermediate data selection condition 3, therefore, transfer task T1 selects B1 as intermediate data. Similarly, in transfer tasks T2 and T3, B1 is selected as intermediate data to establish a pseudo-label-transferring two-stage domain adaptation network. To verify the feasibility of this invention, two comparative methods were used: ① the WDCNN method without domain adaptation, and ② the CNN-MMD method with domain adaptation. The basic networks of both comparative methods are the same as the network structure of Stage I (or Stage II) in this invention, and the relevant parameter settings are also the same. However, the difference lies in that the WDCNN method is a common intelligent fault diagnosis method that does not use domain adaptation; while the CNN-MMD method differs from this invention in that it does not employ a transitive two-stage domain adaptation method, but is a direct transfer-based common domain adaptive diagnosis method. The hyperparameter in this invention is set as: ε = 10. -3 , τ=0.9, ρ=0.9, α=0.05, β=10-6 K = 15000, γ 2 ={10 -6 10 -5 10 -4 10 -3 ,0.01,0.1,1,5,10,15,20,25,30,35,100,10 3 10 4 10 5 10 6 The experiment was repeated 10 times, and the average value was taken as the final test accuracy of the method of this invention and the comparative method to ensure the reliability of the results. The experimental results are as follows: Figure 7 As shown.

[0111] The comparative method directly transfers diagnostics, while the method of this invention uses intermediate data B1 for transitive two-stage domain adaptive diagnostics. Figure 7 It can be seen that the average diagnostic accuracy of CNN-MMD on the three transfer diagnostic tasks is 64.61%, which is 26.54% higher than that of the WDCNN method. This indicates that the domain adaptation method can effectively improve the transfer diagnostic performance of the intelligent diagnostic model under different working conditions and between different rolling bearings, but there is still a large false diagnosis rate. The average accuracy standard deviation of the CNN-MMD method reached 5.82%, indicating that the ordinary transfer diagnostic framework is unstable when there is a large difference in data distribution between the source data and the target data, and cannot solve this problem well. The technical solution disclosed in this invention achieves an average diagnostic accuracy of 96.30% on the three transfer diagnostic tasks, which is 31.69% higher than that of the direct transfer CNN-MMD method, and the model's test performance is more stable. The reason for this is that the direct transfer domain adaptation method ignores the large distribution difference between the source data and the target data. When this problem exists, forcibly performing direct transfer will result in unsatisfactory diagnostic results. This invention transforms the single diagnostic process with a large transfer difficulty into two diagnostic processes with a smaller transfer difficulty. By breaking it down into two and transferring it step by step, a higher diagnostic accuracy for the target rolling bearing is achieved, proving the feasibility of this invention.

[0112] By comparing the present invention with traditional intelligent fault diagnosis methods (WDCNN) and direct transfer domain adaptive transfer diagnosis methods (CNN-MMD), it is shown that the present invention can effectively solve the problem of transfer fault diagnosis between different rolling bearings under different working conditions, especially the problem of unsatisfactory direct transfer accuracy of target data caused by significant differences in the distribution of source data and target data.

[0113] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on pseudo-label transfer two-stage domain adaptive method, characterized in that, Includes the following steps: Step 1: Collect existing rolling bearing fault data as source data. The signal data of the rolling bearing to be diagnosed is collected as the target data. ; Step 2: Calculate the source data using data similarity measurement methods. With target data The similarity of data distribution between them, and then based on the calculation results and selection criteria, in the target data Select intermediate data ; Step 3: Construct a rolling bearing fault diagnosis model based on a pseudo-label-transferring two-stage domain adaptive network; Step 4: Utilize the source data Target data and intermediate data The rolling bearing fault diagnosis model based on the pseudo-label transitive two-stage domain adaptive network is trained, and the fault diagnosis of the rolling bearing to be diagnosed is performed after training. In step 3, the rolling bearing fault diagnosis model based on a pseudo-label-transfer two-stage domain adaptive network includes a two-stage network model: a stage I network model and a stage II network model. The stage I network model is the source data. To intermediate data The domain-adaptive network model, wherein the stage II network model is intermediate data. To target data The domain-adaptive network model, wherein the connection between the stage I network model and the stage II network model is established by a pseudo-label constraint method based on threshold decision, The Stage I network model includes a feature extraction module, a classification module, and a domain adaptation module. The feature extraction module consists of convolutional layers, pooling layers, and a global average pooling layer. The mapping function of the feature extraction module is: The parameters to be trained are The classification module consists of a fully connected layer, a Dropout layer, and a Softmax layer. The mapping function of the classification module is... The parameters to be trained are The domain adaptation module employs a domain adaptation method based on the maximum mean difference to reduce the feature distribution differences between different data. The mapping function is... ; The training process in Phase I is as follows: The source data sample set with known health status... and intermediate data sample sets with unknown health status The inputs are fed into the Stage I network model to optimize the model parameters; through the feature extraction module, the source data samples and intermediate data samples respectively generate corresponding source data features. and intermediate data features The source data features The predicted labels of the source data samples are obtained through the classification module in the Stage I network model. The error between the predicted labels and the true labels of the source data samples is measured by the cross-entropy cost function, and the error is: ， The v The classification probability distribution output by the classification module is the first... v dimension, For an indicator function, when satisfying Its value is 1 when it is active, and 0 otherwise. The source data features With the intermediate data features The distribution discrepancy is reduced by the domain adaptation module in the Stage I network model, and the loss function is: ， The intermediate sample features after the domain adaptation module The classification module obtains the predicted probability of the rolling bearing health status category for each sample. Output for: ， The Features of intermediate samples The training parameters for the classification module, The vector is a one-dimensional probability distribution. The label corresponding to the maximum value is the sample. Predicted health status categories; To avoid model overfitting, L2 regularization loss is used. The parameters are trained using a penalty model, therefore the objective function of the Phase I network model is expressed as: ， The α , β This is the penalty coefficient; The method for constructing pseudo-label constraints based on threshold decision-making involves processing intermediate data samples. Predicting probabilities in network models Using a threshold method to filter out intermediate data samples with high confidence. The selected intermediate data samples The pseudo-labels generated were used for training the Phase II network model. for: ， The The threshold used to generate pseudo-labels; For indicator functions, when satisfying Sometimes, And the sample The corresponding pseudo-tags will be generated. Otherwise, the sample will not generate a pseudo-label and will not be used in the Stage II network model; when the number of intermediate data samples with pseudo-labels reaches the proportion of the total number of intermediate data samples participating in training. ρ Subsequently, intermediate data samples with pseudo-labels were used to train the Stage II network model; the set of high-confidence intermediate data samples is represented as follows: ， The , ; The proportional parameter used to determine whether to end Phase I network model training; This represents the total number of intermediate data samples used in training. The network structure of the Stage II network model also includes a feature extraction module, a classification module, and a domain adaptation module, and the corresponding mapping functions include... and The parameters to be trained include and ; The training process in Phase II is as follows: The intermediate data sample set containing pseudo-label information... and target data sample set with unknown health status The parameters are input into the Stage II network model to optimize the model. The training process for the Stage II network model is the same as that for the Stage I network model. The objective function of the Stage II network model is expressed as: ， The α , β This is the penalty coefficient.

2. The rolling bearing fault diagnosis method based on pseudo-tag transfer two-stage domain adaptive method as described in claim 1, characterized in that, Step 1 specifically includes: Source data For available rolling bearing data with known health status, the known health status rolling bearing data includes rolling bearing data in... N Fault data under various operating conditions, the ,in , The number of sample data entries in the source data. For the first i Sample, for Fault labels, C Indicates the number of health status categories; Target data This refers to rolling bearing signal data whose health status is unknown and needs to be diagnosed. The rolling bearing signal data with unknown health status and requiring diagnosis includes data related to the rolling bearing's performance in... M Fault data under various operating conditions, the ,in , Indicates the first j Sample, Indicates the number of target data samples.

3. The rolling bearing fault diagnosis method based on pseudo-tag transfer two-stage domain adaptive method as described in claim 2, characterized in that, The data similarity measurement method in step 2 is as follows: For data With data The similarity between them, when The closer the data is to 1, the more important it is to maintain the correct value. With data The more similar, when The closer the data is to 0, the more... With data The more significant the difference in data distribution between them, Marked as 0, Marked as 1, train a linear support vector machine classifier. h， The classifier h Used for discrimination and The samples in the data are used to calculate the results. h loss err ( h ): ； The I[ a ] is an indicator function, when a When true, I[ a If the value is 1, then the value is 0; otherwise, the value is 0. express Sampling m Sample, express Sampling m' One sample, then The calculation is as follows: ， Using the aforementioned data similarity measurement method, for data containing N The source data and contents of each working condition data M The distribution similarity between target data of each working condition is calculated. = ,in and , According to the above The calculation results that satisfy conditions 1-3 are the intermediate data. : Condition 1: ; Condition 2: and ; Condition 3: , making That is, when the intermediate data meets condition 2 When multiple options exist, selection can enable... and The target data with the smallest absolute value of the difference is used as the final intermediate data.

4. The rolling bearing fault diagnosis method based on pseudo-tag transfer two-stage domain adaptive method as described in claim 3, characterized in that, Step 4 specifically includes: The parameters of the Stage I and Stage II network models are optimized using the Adam backpropagation algorithm. The total number of training iterations for the Stage I and Stage II network models is [number missing]. K When the number of training iterations is reached K Subsequently, target data samples with unknown health status are input into the trained network model. The Softmax layer of the classification module outputs the predicted probability distribution vector of the target data samples, and the label corresponding to the highest probability is represented as the predicted health status category of the tested sample.

5. The rolling bearing fault diagnosis method based on pseudo-tag transfer two-stage domain adaptive method as described in claim 1, characterized in that, The domain adaptation method based on the maximum mean difference used by the domain adaptation module in step 3 specifically includes: use Representation and Characteristic Kernel k The associated regenerating kernel Hilbert space, given two distributions P , Q Then RKHS exists There is a nonlinear mapping function Map data to RKHS The domain adaptation module then uses the maximum mean difference for calculation: ， The Let be the supremum of the input set. Represents distribution P Expectations Represents distribution Q Expectations express It is a function within the unit sphere in RKHS, respectively, from the distribution P , Q Two groups of samples were obtained independently and identically distributed. and The empirical calculation of the maximum mean difference is as follows: ， The kernel mapping In the multi-kernel maximum mean difference model, the optimal kernel function composed of multiple feature kernels can better approximate the distribution of the feature space. Defined as m Convex combination of kernels: ， The constraints Guaranteed multi-core derivation k It is unique, among which u The number of cores, The constraint coefficient for the core.

Images