Domain adaptation cross-working-condition fault diagnosis method and device based on inter-domain decision difference minimization

Abstract

The invention discloses a domain adaptation cross-working-condition fault diagnosis method and device based on inter-domain decision difference minimization, and relates to the technical field of variable working condition fault diagnosis, and the method comprises the steps: 1, obtaining source domain data and target domain data; 2, constructing a diagnosis model comprising a feature generator, a domain discriminator and a classifier; 3, generating domain private features and domain invariant features; 4, training a domain discriminator and a diagnosis model, and enabling the classifier to establish a preliminary decision boundary; 5, fixing model parameters of the feature generator and the domain discriminator, training a classifier, and determining the upper definite bound of decision differences; 6, fixing model parameters of the classifier and the domain discriminator, training a feature generator, and reducing the feature distribution difference between the source domain data and the target domain data to obtain a mature diagnosis model; and 7, outputting a diagnosis result by using the mature diagnosis model. The technical problem that the accuracy of fault diagnosis of the rotating machine is insufficient in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of variable operating condition fault diagnosis, and in particular to a domain-adaptive cross-operating condition fault diagnosis method and device based on minimizing inter-domain decision differences. Background Art

[0002] Rotating machinery, the most widely used type of mechanical equipment in current production and daily life, plays a critical role in providing and transmitting power within large, complex systems. Bearings and gears, core components of rotating machinery, have a significant impact on their operating efficiency and the reliability of the entire system.

[0003] Currently, conventional fault diagnosis models are commonly used to diagnose rotating machinery faults. However, these models are typically trained based on data from a single operating condition. Since the operating conditions of the equipment often change with the external environment, the distribution of the collected data changes, causing the performance of conventional fault diagnosis models to degrade when performing cross-operating condition diagnosis. Relabeling the unlabeled data for the new operating conditions and building a new model is costly. To address this variable operating condition scenario, domain adaptation methods must still be used for modeling to save costs. Accurately measuring the distribution differences between the source and target operating conditions is key to efficient migration.

[0004] Publication number CN114492533A discloses a method for constructing and applying a variable-condition bearing fault diagnosis model. This technology constructs a model comprising a feature extraction module, a generation module, a domain classifier, and a label classifier. The combination of these three components effectively generates additional training data. While ensuring that the domain classifier can obtain discriminative features within the domain, the domain classifier's sensitivity to different data is reduced by maximizing the domain discrimination loss, making it difficult for the domain classifier to process and distinguish these data. This, in turn, enables the feature extraction module to extract more domain-independent temporal features, achieving end-to-end joint training. After continuous model training, the label classifier is refined as much as possible, and the domain classifier is generalized as much as possible, thereby improving the model's generalization for unseen bearing conditions. However, analysis has found that in complex and variable working environments, the data distribution between the source and target conditions is complex, and the proportions of marginal distributions and conditional distributions in the distribution composition are fuzzy, making the distribution boundaries difficult to clearly define and quantify. In this case, when the model is adapted to the domain, the measurement accuracy of the data distribution differences between working conditions will be reduced, resulting in deviations in the adaptation of the distribution of similar data between domains, making the final migration performance of the model unsatisfactory, thereby affecting the accuracy of fault diagnosis.

[0005] Therefore, for scenarios with complex data distribution, an efficient cross-operating condition fault diagnosis technology is needed to ensure the operational reliability of rotating machinery. Summary of the Invention

[0006] In order to overcome the above-mentioned technical problems existing in the prior art, the present invention proposes a domain-adaptive cross-operating condition fault diagnosis method and device based on minimizing inter-domain decision differences. The present invention trains the model in a staged manner by first establishing a model and a preliminary decision boundary, then fixing the model parameters, finding the supremum of the decision difference, fixing the decision boundary, adjusting the feature distribution, and minimizing the feature distribution difference, and improves the accuracy of the decision difference calculation by introducing the nuclear norm, thereby solving the technical problem of insufficient accuracy of the prior art in rotating machinery fault diagnosis.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences, which comprises the following steps: S1. Obtain labeled vibration signal data of rotating machinery under different fault states under source working conditions and unlabeled vibration signal data under different fault states under target working conditions to be migrated, and define these two types of data as source domain data and target domain data respectively; S2. Construct a diagnostic model comprising a feature generator, a domain discriminator, and a classifier, wherein the feature generator comprises a domain-private feature generator and a domain-invariant feature generator, the domain discriminator is connected to the domain-private feature generator, and the classifier is connected to the domain-private feature generator and the domain-invariant feature generator, respectively; S3. Input the source domain data and the target domain data into the domain-private feature generator to generate domain-private features; input the source domain data and the target domain data into the domain-invariant feature generator to generate domain-invariant features; S4. Use domain-private features to train the domain discriminator, and use domain-private features and domain-invariant features to supervise the training of the diagnostic model, so that the classifier of the diagnostic model establishes a preliminary decision boundary; S5. Fix the model parameters of the feature generator and domain discriminator, introduce the decision difference loss function value and the decision difference nuclear norm value, and train the classifier using the decision difference maximization objective function to update the classifier model parameters and determine the supremum of the decision difference; S6. Fix the model parameters of the classifier and domain discriminator, introduce the decision difference loss function value and the decision difference nuclear norm value, and train the feature generator using the decision difference minimization objective function to update the model parameters of the feature generator and narrow the feature distribution difference between the source domain data and the target domain data. After training, a mature diagnostic model is obtained. S7. Input the test data into the mature diagnosis model and output the diagnosis results.

[0008] In S2, the domain-private feature generator and the domain-invariant feature generator have the same structure, both consisting of two convolutional layers, each of which is connected to the activation function ReLU and the average pooling layer respectively; in the first convolutional layer, the convolution kernel size is 15, the number of input channels is 1, the number of output channels is 16, the sliding step is 1, the convolution kernel length of the pooling layer is 15, and the sliding step is 2; in the second convolutional layer, the convolution kernel size is 15, the number of input channels is 16, the number of output channels is 16, and the sliding step is 1; the convolution kernel length and sliding step of the average pooling layer are 15 and 2, respectively.

[0009] In S2, the domain discriminator consists of three fully connected layers, where the second fully connected layer is sequentially connected to a batch normalization layer, a ReLU activation function layer, and a Dropout layer; the first fully connected layer has 4000 input nodes and 256 output nodes; the second fully connected layer has 256 input nodes and 256 output nodes; the third fully connected layer has 256 input nodes and 2 output nodes.

[0010] In S2, the classifier consists of three fully connected layers, where the second fully connected layer is connected to the batch normalization layer, the ReLU activation function layer and the Dropout layer in sequence; the number of input nodes of the first fully connected layer is 8000, and the number of output nodes is 256; the number of input nodes of the second fully connected layer is 256, and the number of output nodes is 256; the number of input nodes of the third fully connected layer is 256, and the number of output nodes is the number of categories.

[0011] In S3, the domain-invariant feature generator and domain-private feature generator are set to be and , domain-invariant features and domain-private features are and ,but:

[0012] .

[0013] In S4, the domain-private features are input into the domain discriminator, and the predicted domain label is output. The first cross-entropy loss value of the true label and the predicted domain label is then calculated using the first objective function, and the gradient backpropagation of the first cross-entropy loss value is performed to train the domain discriminator.

[0014] In S4, the formula for calculating the first cross entropy loss value using the first objective function is: (1) in, (2) In formula (1) and (2), is the first cross entropy loss value, For the Domain-private features of samples, 、 are the number of samples in the source domain and the target domain respectively; is the predicted domain label, which is a probability vector of length 2. If the data comes from the source domain, the predicted domain label is 0; if the data comes from the target domain, the predicted domain label is 1. They are represented by one-hot encoding vectors as follows:

[0015] .

[0016] In S4, the process of supervised training of the diagnostic model is as follows: domain-private features and domain-invariant features are concatenated and fused, and then input into the classifier to output the predicted label; then the second cross-entropy loss value of the true label and the predicted label is calculated using the second objective function, and the gradient of the second cross-entropy loss value is back-propagated to train the feature generator and classifier, so that the diagnostic model can establish a preliminary decision boundary.

[0017] In S4, the formula for calculating the second cross entropy loss value using the second objective function is: (3) in, (4) In formula (3) and (4), is the second cross entropy loss value, is the true label of the source domain data, which is of length The one-hot encoded vector of The predicted label output by the diagnostic model classifier for the source domain data is of length The probability vector of 、 The source domain The domain-invariant features and domain-private features of samples, is the number of source domain samples.

[0018] In S5, the objective function of maximizing decision difference is: (5) In formula (5), is the objective function value of maximizing the decision difference, is the loss function value of decision difference, is the norm value of the decision difference.

[0019] In S6, the objective function for minimizing decision differences is: (6) In formula (6), is the objective function value of maximizing the decision difference, is the loss function value of decision difference, is the norm value of the decision difference.

[0020] The method for obtaining the loss function value of the decision difference is: S61. Given source domain data and target domain data , the number of categories is c, and the distribution of the number of samples in each category is the same across domains; let Represents a classifier On the source domain, the prediction belongs to the category The number of samples, let Represents a classifier On the target domain, the prediction belongs to the category The number of samples, then the expression of the loss function value defining the decision difference is: (7) In formula (7), is the loss function value of decision difference; For the classifier The symmetric hypothesis space of is the upper bound.

[0021] S62. The expression of the loss function value of the decision difference is transformed into the following formula (8): (8) in, is a probability vector, that is , and satisfies the following formula (9): (9); S63. According to formula (9), the relationship between the number of predicted categories and the probability of the predicted results is obtained: (10;) S64. Substitute equation (10) into equation (8) to obtain the final expression of the loss function value of the decision difference: (11); S65. Use formula (11) to calculate the loss function value of decision difference.

[0022] The calculation formula of the nuclear norm value of the decision difference is: (12) In formula (12), is the single batch data prediction probability matrix The largest singular value.

[0023] In a second aspect, the present invention provides a domain-adaptive cross-operating-condition fault diagnosis device based on minimizing inter-domain decision differences, comprising: The data acquisition module is used to obtain the labeled vibration signal data of the rotating machinery in different fault states under the source working condition and the unlabeled vibration signal data of different fault states under the target working condition to be migrated, and define these two types of data as source domain data and target domain data respectively; a model construction module, for constructing a diagnostic model comprising a feature generator, a domain discriminator, and a classifier, wherein the feature generator comprises a domain-private feature generator and a domain-invariant feature generator, the domain discriminator is connected to the domain-private feature generator, and the classifier is connected to the domain-private feature generator and the domain-invariant feature generator, respectively; A feature generation module, configured to input source domain data and target domain data into a domain-private feature generator and generate domain-private features; and to input source domain data and target domain data into a domain-invariant feature generator and generate domain-invariant features; A decision boundary establishment module is used to train a domain discriminator using domain-private features, and to supervise the training of the diagnosis model using domain-private features and domain-invariant features, so that the classifier of the diagnosis model establishes a preliminary decision boundary; The classifier training module is used to fix the model parameters of the feature generator and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and train the classifier using the decision difference maximization objective function to update the classifier model parameters and determine the supremum of the decision difference; The feature generator training module is used to fix the model parameters of the classifier and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and use the decision difference minimization objective function to train the feature generator to update the model parameters of the feature generator and narrow the feature distribution difference between the source domain data and the target domain data; after the training is completed, a mature diagnostic model is obtained.

[0024] The model application module is used to input the test data into the mature diagnosis model and output the diagnosis results.

[0025] Compared with the prior art, the beneficial effects of the present invention are: In the method provided by the present invention, the advantages of S2 and S3 are that the domain-private feature generator can be used to extract as much difference information between domains as possible, thereby laying the foundation for the subsequent calculation of decision differences. The advantage of S4 is that it is conducive to ensuring that the classifier can form accurate classification capabilities in the source domain. The advantage of S5 is that by fixing the distribution of data in the feature space, it is conducive to more accurately locating the supremum of decision differences. The advantage of S6 is that by fixing the position of the decision boundary, the distribution position of data in the feature space can be more accurately adjusted, thereby promoting feature alignment between similar data in two fields and narrowing the decision differences.

[0026] In summary, the present invention adopts the above-mentioned specific method, which helps avoid the mutual influence between the two steps of maximizing the decision difference and minimizing the decision difference by first establishing a model and establishing a preliminary decision boundary, then fixing the model parameters, finding the supremum of the decision difference, fixing the decision boundary, adjusting the feature distribution, and minimizing the feature distribution difference. The introduction of the nuclear norm helps to improve the accuracy of the decision difference calculation, so that the model can achieve more accurate intelligent fault diagnosis of bearings in rotating machinery, providing a solid foundation for fault diagnosis of rotating machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of the method of the present invention; Figure 2 This is a principle block diagram of the diagnostic model of the present invention; Figure 3 It is a structural diagram of the diagnostic model; Figure 4 Schematic diagram of the connection structure of the device of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention but are not used to limit the present invention.

[0029] Example 1 like Figure 1 As shown, the present invention provides a domain-adaptive cross-operating condition fault diagnosis method based on minimizing inter-domain decision differences, which includes the following steps: S1. During the operation of the rotating machinery, obtain labeled vibration signal data of the rotating machinery under different fault states from the bearing acceleration sensor under the source working condition. Under the target working condition to be migrated, obtain unlabeled vibration signal data of the rotating machinery under different fault states from the bearing acceleration sensor. Define these two types of data as source domain data and target domain data, respectively.

[0030] S2. Figure 2As shown in the figure, a diagnostic model including a feature generator, a domain discriminator and a classifier is constructed. The feature generator includes a domain-private feature generator and a domain-invariant feature generator. The domain discriminator is connected to the domain-private feature generator, and the classifier is connected to the domain-private feature generator and the domain-invariant feature generator respectively.

[0031] Specifically, such as Figure 3 As shown in (a), the domain-private feature generator and the domain-invariant feature generator have the same structure, both consisting of two convolutional layers, each of which is connected to the activation function ReLU and the average pooling layer respectively; in the first convolutional layer, the convolution kernel size is 15, the number of input channels is 1, the number of output channels is 16, the sliding step is 1, the convolution kernel length of the pooling layer is 15, and the sliding step is 2; in the second convolutional layer, the convolution kernel size is 15, the number of input channels is 16, the number of output channels is 16, and the sliding step is 1; the convolution kernel length and sliding step of the average pooling layer are 15 and 2, respectively.

[0032] like Figure 3 As shown in (b), the domain discriminator consists of three fully connected layers, where the second fully connected layer is sequentially connected to the batch normalization layer, the ReLU activation function layer and the Dropout layer; the number of input nodes of the first fully connected layer is 4000, and the number of output nodes is 256; the number of input nodes of the second fully connected layer is 256, and the number of output nodes is 256; the number of input nodes of the third fully connected layer is 256, and the number of output nodes is 2.

[0033] like Figure 3 As shown in (c), the classifier consists of three fully connected layers, and its basic structure is basically the same as that of the domain discriminator. The second fully connected layer is sequentially connected to the batch normalization layer, the ReLU activation function layer, and the Dropout layer; the number of input nodes of the first fully connected layer is 8000, and the number of output nodes is 256; the number of input nodes of the second fully connected layer is 256, and the number of output nodes is 256; the number of input nodes of the third fully connected layer is 256, and the number of output nodes is the number of categories.

[0034] S3. Input the source domain data and the target domain data into the domain-private feature generator to generate domain-private features; input the source domain data and the target domain data into the domain-invariant feature generator to generate domain-invariant features.

[0035] Specifically, the domain-invariant feature generator and domain-private feature generator are set to be and , domain-invariant features and domain-private features are and ,but:

[0036] .

[0037] S4. Use domain-private features to train the domain discriminator, and use domain-private features and domain-invariant features to supervise the training of the diagnosis model, so that the classifier of the diagnosis model can establish a preliminary decision boundary.

[0038] Specifically, the method of using domain-private features to train a domain discriminator is as follows: the domain-private features are input into the domain discriminator, and the predicted domain label is output. Then, a first objective function is used to calculate a first cross-entropy loss value between the true label and the predicted domain label, and gradient backpropagation is performed on the first cross-entropy loss value to train the domain discriminator.

[0039] Furthermore, the formula for calculating the first cross entropy loss value using the first objective function is: (1) in, (2) In formula (1) and (2), is the first cross entropy loss value, For the Domain-private features of samples, 、 are the number of samples in the source domain and the target domain respectively; is the predicted domain label, which is a probability vector of length 2. If the data comes from the source domain, the predicted domain label is 0; if the data comes from the target domain, the predicted domain label is 1. They are represented by one-hot encoding vectors as follows:

[0040] .

[0041] Specifically, the process of supervised training of the diagnostic model is as follows: domain-private features and domain-invariant features are concatenated and fused, and then input into the classifier to output the predicted label; then the second cross-entropy loss value of the true label and the predicted label is calculated using the second objective function, and the second cross-entropy loss value is gradient backpropagated to train the feature generator and classifier, so that the diagnostic model can establish a preliminary decision boundary.

[0042] Furthermore, the formula for calculating the second cross entropy loss value using the second objective function is: (3) in, (4) In formula (3) and (4), is the second cross entropy loss value, is the true label of the source domain data, which is of length The one-hot encoded vector of The predicted label output by the diagnostic model classifier for the source domain data is of length The probability vector of 、 The source domain The domain-invariant features and domain-private features of samples, is the number of source domain samples.

[0043] It should be noted that the decision boundary in this step refers to the boundary where the diagnostic model separates different categories in the feature space. It is the dividing line used by the diagnostic model to divide samples of different categories.

[0044] S5. Fix the model parameters of the feature generator and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and train the classifier using the decision difference maximization objective function to update the model parameters of the classifier and determine the supremum of the decision difference.

[0045] Specifically, the objective function of maximizing the decision difference is: (5) In formula (5), is the objective function value of maximizing the decision difference, is the loss function value of decision difference, is the norm value of the decision difference.

[0046] It should be noted that this step refers to the difference in prediction results of the diagnostic model classifier for similar data in two fields (source field and target field).

[0047] S6. Fix the model parameters of the classifier and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and use the decision difference minimization objective function to train the feature generator to update the model parameters of the feature generator and narrow the feature distribution difference between the source domain data and the target domain data; after training, a mature diagnostic model is obtained.

[0048] Specifically, the objective function for minimizing decision differences is: (6) In formula (6), is the objective function value of maximizing the decision difference, is the loss function value of decision difference, is the norm value of the decision difference.

[0049] Specifically, the method for obtaining the loss function value of decision difference is: S61. Given source domain data and target domain data , the number of categories is c, and the distribution of the number of samples in each category is the same across domains; let Represents a classifier On the source domain, the prediction belongs to the category The number of samples, let Represents a classifier On the target domain, the prediction belongs to the category The number of samples, then the expression of the loss function value defining the decision difference is: (7) In formula (7), is the loss function value of decision difference; For the classifier The symmetric hypothesis space of is the upper bound.

[0050] S62. The expression of the loss function value of the decision difference is transformed into the following formula (8): (8) in, is a probability vector, that is , and satisfies the following formula (9): (9); S63. According to formula (9), the relationship between the number of predicted categories and the probability of the predicted results is obtained: (10;) S64. Substitute equation (10) into equation (8) to obtain the final expression of the loss function value of the decision difference: (11); S65. Use formula (11) to calculate the loss function value of decision difference.

[0051] Specifically, the calculation formula for the nuclear norm value of the decision difference is: (12) In formula (12), is the single batch data prediction probability matrix The largest singular value.

[0052] S7. Input the test data into the mature diagnosis model and output the diagnosis results.

[0053] In addition, the applicant named the fault types based on the Case Western Reserve University (CWRU) bearing data set in Table 1 below, and conducted multiple tests under cross-operating condition combinations on the same CWRU data set. The accuracy of bearing fault prediction was compared using the scheme of the present invention, the Base algorithm, the MMD method, and the DANN neural network. As shown in Table 2 below, in the case of multiple tests, the average accuracy obtained using the scheme was the highest, at 99.96%. This shows the advantages of the present method in cross-operating condition fault diagnosis applications.

[0054]

[0055]

[0056] Example 2 Based on the same inventive concept, the present invention also provides a domain-adaptive cross-operating-condition fault diagnosis device based on minimizing inter-domain decision differences, such as Figure 4 As shown, it includes: The data acquisition module is used to obtain the labeled vibration signal data of the rotating machinery in different fault states under the source working condition and the unlabeled vibration signal data of different fault states under the target working condition to be migrated, and define these two types of data as source domain data and target domain data respectively; a model construction module, for constructing a diagnostic model comprising a feature generator, a domain discriminator, and a classifier, wherein the feature generator comprises a domain-private feature generator and a domain-invariant feature generator, the domain discriminator is connected to the domain-private feature generator, and the classifier is connected to the domain-private feature generator and the domain-invariant feature generator, respectively; A feature generation module, configured to input source domain data and target domain data into a domain-private feature generator and generate domain-private features; and to input source domain data and target domain data into a domain-invariant feature generator and generate domain-invariant features; A decision boundary establishment module is used to train a domain discriminator using domain-private features, and to supervise the training of the diagnosis model using domain-private features and domain-invariant features, so that the classifier of the diagnosis model establishes a preliminary decision boundary; The classifier training module is used to fix the model parameters of the feature generator and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and train the classifier using the decision difference maximization objective function to update the classifier model parameters and determine the supremum of the decision difference; The feature generator training module is used to fix the model parameters of the classifier and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and use the decision difference minimization objective function to train the feature generator to update the model parameters of the feature generator and narrow the feature distribution difference between the source domain data and the target domain data; after the training is completed, a mature diagnostic model is obtained.

[0057] The model application module is used to input the test data into the mature diagnosis model and output the diagnosis results.

[0058] In detail, the functions of the above functional modules correspond to the contents of the corresponding steps in Example 1, and the parts not fully described will be omitted.

[0059] The above description is only a specific embodiment of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all disclosed features, or all steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A domain-adaptive cross-operating condition fault diagnosis method based on minimizing inter-domain decision differences, characterized by The following steps are involved: S1. Obtain labeled vibration signal data of rotating machinery under different fault states under source working conditions and unlabeled vibration signal data under different fault states under target working conditions to be migrated, and define these two types of data as source domain data and target domain data respectively; S2. Construct a diagnostic model comprising a feature generator, a domain discriminator, and a classifier, wherein the feature generator comprises a domain-private feature generator and a domain-invariant feature generator, the domain discriminator is connected to the domain-private feature generator, and the classifier is connected to the domain-private feature generator and the domain-invariant feature generator, respectively; S3. Input the source domain data and the target domain data into the domain-private feature generator to generate domain-private features; Input the source domain data and the target domain data into the domain-invariant feature generator to generate domain-invariant features; S4. Use domain-private features to train the domain discriminator, and use domain-private features and domain-invariant features to supervise the training of the diagnostic model, so that the classifier of the diagnostic model establishes a preliminary decision boundary; S5. Fix the model parameters of the feature generator and domain discriminator, introduce the decision difference loss function value and the decision difference nuclear norm value, and train the classifier using the decision difference maximization objective function to update the classifier model parameters and determine the supremum of the decision difference; S6. Fix the model parameters of the classifier and domain discriminator, introduce the decision difference loss function value and the decision difference nuclear norm value, and train the feature generator using the decision difference minimization objective function to update the model parameters of the feature generator and narrow the feature distribution difference between the source domain data and the target domain data; After the training is completed, a mature diagnostic model is obtained; S7. Input the test data into the mature diagnosis model and output the diagnosis results.

2. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 1 is characterized by: In S2, the domain-private feature generator and the domain-invariant feature generator have the same structure, both consisting of two convolutional layers, each of which is connected to the activation function ReLU and the average pooling layer respectively; in the first convolutional layer, the convolution kernel size is 15, the number of input channels is 1, the number of output channels is 16, the sliding step is 1, the convolution kernel length of the pooling layer is 15, and the sliding step is 2; in the second convolutional layer, the convolution kernel size is 15, the number of input channels is 16, the number of output channels is 16, and the sliding step is 1; the convolution kernel length and sliding step of the average pooling layer are 15 and 2, respectively.

3. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 1 is characterized by: In S2, the domain discriminator consists of three fully connected layers, where the second fully connected layer is sequentially connected to a batch normalization layer, a ReLU activation function layer, and a Dropout layer; the first fully connected layer has 4000 input nodes and 256 output nodes; the second fully connected layer has 256 input nodes and 256 output nodes; the third fully connected layer has 256 input nodes and 2 output nodes.

4. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 1 is characterized by: In S2, the classifier consists of three fully connected layers, where the second fully connected layer is connected to the batch normalization layer, the ReLU activation function layer and the Dropout layer in sequence; the number of input nodes of the first fully connected layer is 8000, and the number of output nodes is 256; the number of input nodes of the second fully connected layer is 256, and the number of output nodes is 256; the number of input nodes of the third fully connected layer is 256, and the number of output nodes is the number of categories.

5. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 1 is characterized by: In S3, the domain-invariant feature generator and domain-private feature generator are set to be and , domain-invariant features and domain-private features are and ,but: 。 6. A domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to any one of claims 1 to 5, characterized in that: In S4, the domain-private features are input into the domain discriminator, and the predicted domain label is output. The first cross-entropy loss value of the true label and the predicted domain label is then calculated using the first objective function, and the gradient backpropagation of the first cross-entropy loss value is performed to train the domain discriminator.

7. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 6 is characterized by: In S4, the formula for calculating the first cross entropy loss value using the first objective function is: （1） in, （2） In formula (1) and (2), is the first cross entropy loss value, For the Domain-private features of samples, 、 are the number of samples in the source domain and the target domain respectively; is the predicted domain label, which is a probability vector of length 2. If the data comes from the source domain, the predicted domain label is 0; if the data comes from the target domain, the predicted domain label is 1. They are represented by one-hot encoding vectors as follows: 。 8. A domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to any one of claims 1 to 5, characterized in that: In S4, the process of supervised training of the diagnostic model is as follows: domain-private features and domain-invariant features are concatenated and fused, and then input into the classifier to output the predicted label; then the second cross-entropy loss value of the true label and the predicted label is calculated using the second objective function, and the gradient of the second cross-entropy loss value is back-propagated to train the feature generator and classifier, so that the diagnostic model can establish a preliminary decision boundary.

9. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 8, characterized in that: In S4, the formula for calculating the second cross entropy loss value using the second objective function is: （3） in, （4） In formula (3) and (4), is the second cross entropy loss value, is the true label of the source domain data, which is of length The one-hot encoded vector of The predicted label output by the diagnostic model classifier for the source domain data is of length The probability vector of 、 The source domain The domain-invariant features and domain-private features of samples, is the number of source domain samples.

10. A domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to any one of claims 1-5, 7 or 9, characterized in that: In S5, the objective function of maximizing decision difference is: （5） In formula (5), is the objective function value of maximizing the decision difference, is the loss function value of decision difference, is the norm value of the decision difference.

11. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 10, characterized in that: In S6, the objective function for minimizing decision differences is: （6） In formula (6), is the objective function value of maximizing the decision difference, is the loss function value of decision difference, is the norm value of the decision difference.

12. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 11, characterized in that: The method for obtaining the loss function value of the decision difference is: S61. Given source domain data and target domain data , the number of categories is c, and the distribution of the number of samples in each category is the same across domains; let Represents a classifier On the source domain, the prediction belongs to the category The number of samples, let Represents a classifier On the target domain, the prediction belongs to the category The number of samples, then the expression of the loss function value defining the decision difference is: （7） In formula (7), is the loss function value of decision difference; For the classifier The symmetric hypothesis space of is the upper bound. S62. The expression of the loss function value of the decision difference is transformed into the following formula (8): （8） in, is a probability vector, that is , and satisfies the following formula (9): （9）； S63. According to formula (9), the relationship between the number of predicted categories and the probability of the predicted results is obtained: （10；） S64. Substitute equation (10) into equation (8) to obtain the final expression of the loss function value of the decision difference: （11）； S65. Use formula (11) to calculate the loss function value of decision difference.

13. The domain-adaptive cross-operating-condition fault diagnosis method based on minimizing inter-domain decision differences according to claim 12, characterized in that: The calculation formula of the nuclear norm value of the decision difference is: （12） In formula (12), is the single batch data prediction probability matrix The largest singular value.

14. A domain-adaptive cross-operating fault diagnosis device based on minimizing inter-domain decision differences, characterized by include: The data acquisition module is used to obtain the labeled vibration signal data of the rotating machinery in different fault states under the source working condition and the unlabeled vibration signal data of different fault states under the target working condition to be migrated, and define these two types of data as source domain data and target domain data respectively; a model construction module, for constructing a diagnostic model comprising a feature generator, a domain discriminator, and a classifier, wherein the feature generator comprises a domain-private feature generator and a domain-invariant feature generator, the domain discriminator is connected to the domain-private feature generator, and the classifier is connected to the domain-private feature generator and the domain-invariant feature generator, respectively; A feature generation module, configured to input source domain data and target domain data into a domain-private feature generator and generate domain-private features; and to input source domain data and target domain data into a domain-invariant feature generator and generate domain-invariant features; A decision boundary establishment module is used to train a domain discriminator using domain-private features, and to supervise the training of the diagnosis model using domain-private features and domain-invariant features, so that the classifier of the diagnosis model establishes a preliminary decision boundary; The classifier training module is used to fix the model parameters of the feature generator and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and train the classifier using the decision difference maximization objective function to update the classifier model parameters and determine the supremum of the decision difference; The feature generator training module is used to fix the model parameters of the classifier and domain discriminator, introduce the loss function value of the decision difference and the nuclear norm value of the decision difference, and train the feature generator using the decision difference minimization objective function to update the model parameters of the feature generator and narrow the feature distribution difference between the source domain data and the target domain data. After training, a mature diagnostic model is obtained; The model application module is used to input the test data into the mature diagnosis model and output the diagnosis results.

Images