Time-varying working condition mechanical fault migration diagnosis method based on passive progressive pseudo label

Abstract

This invention discloses a time-varying mechanical fault transfer diagnosis method based on passive progressive pseudo-labels. It involves collecting labeled source domain sample sets and unlabeled target domain sample sets of the mechanical equipment to be diagnosed. A mechanical fault diagnosis model is constructed according to actual needs. First, the source domain sample set is used to train the mechanical fault diagnosis model to obtain a source domain model. Then, based on a progressive pseudo-label selection strategy, the target domain sample set is used to transfer train the source model to obtain a target domain model. The loss function for transfer training is obtained by weighted summation of information entropy loss and distribution alignment loss. After training, the trained mechanical fault diagnosis model can be used to diagnose faults in the mechanical equipment to be diagnosed. This invention employs a progressive pseudo-label selection strategy, gradually utilizing easily learnable samples to guide difficult-to-learn samples to self-correct, thereby enhancing the model's adaptability in complex dynamic data structures and improving the accuracy of transfer diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis technology, and more specifically, relates to a method for diagnosing the migration of mechanical faults under time-varying operating conditions based on passive progressive pseudo-labels. Background Technology

[0002] When mechanical equipment malfunctions, it can lead to anything from reduced equipment performance to significant economic losses and casualties. Fault diagnosis is therefore a crucial measure to ensure the safe and stable operation of equipment. With the development of neural networks, their powerful feature learning and fitting capabilities are being widely applied to mechanical equipment fault diagnosis, providing a new approach to accurately identify faults.

[0003] However, in practical applications, to achieve optimal service performance, mechanical equipment often actively or passively adjusts its operating posture, leading to changes in operating conditions such as speed and load, causing distributional deviations between neural network training and testing data. Furthermore, monitoring data is trapped in "data silos" due to inter-enterprise competition and data privacy protection needs. To address this, a diagnostic method based on passive domain adaptation has been proposed as an effective solution. However, existing passive domain adaptation diagnostic methods do not consider the uncertainty and large distribution characteristics of target prediction under time-varying conditions, resulting in the method's effectiveness only applicable to scenarios with stable data distribution. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a time-varying mechanical fault transfer diagnosis method based on passive progressive pseudo-labels. This method gradually utilizes easy-to-learn samples to guide difficult-to-learn samples to perform self-correction, thereby enhancing the model's adaptive ability in complex dynamic data structures and improving the accuracy of transfer diagnosis.

[0005] To achieve the above-mentioned objectives, the present invention provides a time-varying mechanical fault migration diagnosis method based on passive progressive pseudo-tags, comprising the following steps:

[0006] S1: Collect several mechanical data samples and their corresponding fault category labels according to actual needs, and record the number of fault categories as follows: Thus, the source domain sample set is obtained;

[0007] S2: Collect several mechanical data samples for the mechanical equipment to be diagnosed to obtain the target domain sample set;

[0008] S3: Construct a mechanical fault diagnosis model according to actual needs, including a data feature extraction module and a classifier. The data feature extraction module is used to extract features from mechanical data samples, and the classifier is used to predict the corresponding fault category based on the extracted features.

[0009] S4: The source domain model is obtained by training the mechanical fault diagnosis model using the source domain sample set;

[0010] S5: Based on a progressive pseudo-label selection strategy, the source model is transferred and trained using the target domain sample set to obtain the target domain model. The loss function in each training batch is calculated as follows:

[0011] Let the target domain sample set of the current training batch be . Each target domain sample Input the current mechanical fault diagnosis model to obtain its predicted category; set the predicted category as... The target domain samples constitute the sample set , Then, the following formula is used to filter each category. Reliable pseudo-label sample set :

[0012] ,

[0013] in, Indicates the number of individuals in the set. Represents the source domain sample set The true category is The number of samples, Represents the target domain sample. This indicates that the classifier in the mechanical fault diagnosis model is for each category. The generated log odds, , , The parameter is The mapping function from the sample space to the label space in the target domain model; This indicates sorting based on confidence level from highest to lowest. , This indicates a preset proportional threshold. , This indicates the preset exponential threshold.

[0014] The information entropy loss is calculated using the following formula. :

[0015] ,

[0016] in, This represents the set of reliable pseudo-label samples for the current batch. Represents logarithmic probability The corresponding probability is calculated using the following formula:

[0017] ;

[0018] The distribution alignment loss is calculated using the following formula. :

[0019] ,

[0020] in, Represents the source domain sample set The actual labels in the middle are categories The source domain samples are composed of the source feature matrix consisting of features extracted by the data feature extraction module in the mechanical fault diagnosis model. Represents a reliable set of pseudo-label samples. The target domain samples are composed of target feature matrices formed by features extracted by the data feature extraction module in the mechanical fault diagnosis model. Represents the source feature matrix and target feature matrix The distribution differences;

[0021] The final loss is then calculated using the following formula. :

[0022] ,

[0023] in, , Represents information entropy loss Alignment loss with distribution The weights;

[0024] S6: When it is necessary to perform fault diagnosis on the mechanical equipment to be diagnosed, collect mechanical data samples at the current moment, input them into the mechanical fault diagnosis model trained in step S5, and obtain the corresponding fault category.

[0025] This invention relates to a time-varying mechanical fault transfer diagnosis method based on passive progressive pseudo-labels. It involves collecting labeled source domain sample sets and unlabeled target domain sample sets of the mechanical equipment to be diagnosed. A mechanical fault diagnosis model is constructed according to actual needs. First, the source domain sample set is used to train the mechanical fault diagnosis model to obtain a source domain model. Then, based on a progressive pseudo-label selection strategy, the target domain sample set is used to transfer train the source model to obtain a target domain model. The loss function for transfer training is obtained by weighted summation of information entropy loss and distribution alignment loss. After training, the trained mechanical fault diagnosis model can be used to diagnose faults in the mechanical equipment to be diagnosed.

[0026] The present invention has the following beneficial effects:

[0027] 1) This invention designs a novel progressive pseudo-label selection strategy, which gradually utilizes easy-to-learn samples to guide difficult-to-learn samples to self-correct, thereby ensuring the effectiveness of model adaptation and improving the accuracy of transfer diagnosis.

[0028] 2) This invention aligns source domain features and target domain features to explicitly reduce cross-domain distribution differences during model adaptation. When calculating distribution alignment loss, it proposes a scale-robust subspace calibration method to further reduce cross-domain distribution differences. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a specific implementation method of the time-varying mechanical fault migration diagnosis method based on passive progressive pseudo-labels of the present invention.

[0030] Figure 2 This is a schematic diagram illustrating the principle of CAS-Softmax loss in this embodiment;

[0031] Figure 3 This is a schematic diagram of the progressive pseudo-label selection strategy in this invention;

[0032] Figure 4 This is a schematic diagram of transfer training in this embodiment;

[0033] Figure 5 This is a schematic diagram of subspace calibration measurement in this embodiment;

[0034] Figure 6 This is a structural diagram of the self-test simulation test bench in this embodiment. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0036] Example

[0037] Figure 1 This is a flowchart illustrating a specific implementation of the time-varying mechanical fault migration diagnosis method based on passive progressive pseudo-tags of the present invention. Figure 1 As shown, the specific steps of the time-varying mechanical fault migration diagnosis method based on passive progressive pseudo-tags of the present invention include:

[0038] S101: Obtain the source domain sample set:

[0039] Based on actual needs, collect a number of mechanical data samples and their corresponding fault category labels, denoted as the number of fault categories. Thus, the source domain sample set is obtained.

[0040] In this embodiment, the data sample is the vibration signal collected by the accelerometer. It can collect a segment of the original vibration signal of a mechanical equipment under different fault categories, and then use the sliding window technique to obtain mechanical fault data samples. The vibration signal characteristic data at each sampling time can be set according to actual needs.

[0041] S102: Collect target domain sample set:

[0042] Several mechanical data samples are collected for the mechanical equipment to be diagnosed, resulting in a target domain sample set.

[0043] S103: Constructing a mechanical fault diagnosis model:

[0044] A mechanical fault diagnosis model is constructed according to actual needs, including a data feature extraction module and a classifier. The data feature extraction module is used to extract features from mechanical data samples, and the classifier is used to predict the corresponding fault category based on the extracted features.

[0045] In practical applications, the specific structure of each model in the mechanical fault diagnosis model can be set according to actual needs. In this embodiment, the data feature extraction module is implemented based on a convolutional neural network, and the fault prediction module is implemented based on a fully connected network.

[0046] S104: Training the source domain model:

[0047] The source domain model is obtained by training the mechanical fault diagnosis model using the source domain sample set.

[0048] Given labeled source domain data, under traditional supervised learning, the source model can easily be solved using classic cross-entropy loss. Learned the decision boundary of overfitting:

[0049] (1)

[0050] in, and These represent the source domain data sample and its data sample space, respectively. and These represent the source domain label and its label space, respectively. Indicates the number of categories. This indicates that the classifier in the mechanical fault diagnosis model is for each category. The generated logit. , , The parameter is The mapping function from the sample space to the label space in the source domain model. Indicates based on logarithmic odds Generated categories The probability is usually obtained using the Softmax activation function, and the calculation formula is as follows:

[0051] (2)

[0052] However, under time-varying conditions, the target domain prediction results obtained through the aforementioned source model exhibit a wider distribution range, leading to significantly higher prediction uncertainty. Therefore, only a small portion of easily learnable target samples can be correctly classified, while a large number of difficult-to-learn samples are pushed into the wrong category. Taking a binary classification task as an example, for a difficult-to-learn target sample from category 2 (meaning misclassification, i.e., prediction vector [a, b], a>b), its prediction output is... Due to overfitting of the source model and distribution shift, the actual output may be distorted to [0.9, 0.1], leading to even more inaccurate pseudo-labels. Once these erroneous pseudo-labels dominate the self-supervised process, the model will lose its ability to recalibrate predictions, ultimately causing the entire model adaptation process to collapse.

[0053] Therefore, this embodiment proposes a CAS-Softmax loss during the pre-training of the source model to adaptively adjust the decision boundary and mitigate the adverse effects of overfitting. The calculation formula is as follows:

[0054] (3)

[0055] (4)

[0056] (5)

[0057] in, and These represent the source domain data sample and its data sample space, respectively. and These represent the source domain label and its label space, respectively. This indicates that the classifier in the mechanical fault diagnosis model is for each category. The generated log odds, , , The parameter is The mapping function from the sample space to the label space in the source domain model. This represents the classifier's classification for the true label in a mechanical fault diagnosis model. The generated log odds, , These represent the categories corresponding to the actual labels. and categories The probability, Indicates the category corresponding to the actual label. The relevant decision margin.

[0058] According to equation (4), determining the decision interval is the core issue. Furthermore, due to differences in data distribution among different categories, the decision interval for each category varies. They should be different, that is . Figure 2 This is a schematic diagram illustrating the principle of CAS-Softmax loss in this embodiment. For example... Figure 2 As shown, by controlling It can adjust the degree of overfitting at the decision boundary.

[0059] Since feature distributions are typically skewed, the log-odds of each category's output can be approximated by a log-normal distribution, i.e. Its corresponding probability density function is:

[0060] (6)

[0061] in, Indicates category The natural logarithm mean of log odds Indicates category The natural logarithmic standard deviation of log odds.

[0062] Combining the log-normal distribution model and actual observations of each category probability value Distribution parameters of each category The following formula can be used for estimation:

[0063] (7)

[0064] Subsequently, categories and categories Overlap factor between the natural log-normal distribution of log odds It can be represented as:

[0065] (8)

[0066] Ultimately, decision interval The definition is as follows:

[0067] (9)

[0068] When label knowledge is learned accurately That is, the category corresponding to the real label. Corresponding The decision interval is greater than any other category. This is used to avoid overfitting caused by overly strict decision boundaries; while when That is, the category corresponding to the real label. Corresponding This is not the maximum value, indicating significant overlap in distributions between categories. The source model should focus more on learning discriminative features. In this case, let... .

[0069] S104: Transfer training yields the target domain model:

[0070] Next, the source model needs to be transferred and trained using the target domain sample set to obtain the target domain model.

[0071] Figure 3 This is a schematic diagram of the progressive pseudo-label selection strategy in this invention. For example... Figure 3 As shown, under time-varying conditions, dynamic data structures cause significant changes in data distribution. Therefore, in the initial stage of model adaptation, only a small number of easily learnable target samples can produce correct prediction probability distributions, while most difficult-to-learn samples are prone to incorrect predictions. However, in existing passive domain adaptation paradigms, most pseudo-label mechanisms uniformly incorporate the source model's predictions for all target samples into self-supervised learning throughout the entire model adaptation process. This indiscriminate pseudo-label strategy inevitably allows noisy, difficult-to-learn samples to interfere with easily learnable samples, thereby weakening the model's self-correction ability and ultimately leading to negative transfer. To address this, this invention proposes a progressive pseudo-label selection strategy to dynamically identify high-confidence pseudo-labels and gradually optimize the training process. An explicit distribution alignment mechanism gradually increases the number of reliable pseudo-labels during training, thereby ensuring the effectiveness of model adaptation. Specifically, this strategy faces two key issues: how to select reliable pseudo-labels and how to achieve effective distribution alignment. Figure 4 This is a schematic diagram of transfer training in this embodiment. For example... Figure 4 As shown, two losses need to be set during transfer training: one is the information entropy loss based on reliable pseudo-labels. Alignment loss with distribution .

[0072] 1) Select reliable pseudo-labels:

[0073] Generally, a higher probability value corresponding to a label index position indicates a higher prediction confidence. Intuitively, setting a uniform threshold to filter easily learnable samples is a feasible method. However, due to the differences in data distribution characteristics between different categories, a separate threshold needs to be set for each category, which inevitably increases the complexity of hyperparameter optimization. Based on the above considerations, this invention directly selects the samples with the highest confidence ranking in each category as reliable pseudo-label samples, thereby filtering out samples from each category. Reliable pseudo-label sample set The specific method is as follows:

[0074] Let the target domain sample set of the current training batch be . Each target domain sample Input the current mechanical fault diagnosis model to obtain its predicted category. Assign the predicted category as... The target domain samples constitute the sample set , Then, the following formula is used to filter each category. Reliable pseudo-label sample set :

[0075] (10)

[0076] in, Indicates the number of individuals in the set. Represents the source domain sample set The true category is The number of samples, Represents the target domain sample. This indicates that the classifier in the mechanical fault diagnosis model is for each category. The generated log odds, , , The parameter is The mapping function from the sample space to the label space in the target domain model. This indicates that the values ​​are sorted from highest to lowest based on confidence level. , This indicates a preset proportional threshold. In this embodiment , . This represents a preset exponential threshold, in this embodiment... .

[0077] According to the above formula, when the sample set The number of samples meets the requirements. ), while each sample in the category The logarithmic odds meet the requirements ( If the sample set is selected, then the sample set is taken. The top of the medium confidence level The samples constitute a reliable pseudo-label sample set When the sample size does not meet the requirements, or each sample is in a certain category If the log-odds ratio does not meet the requirements, a reliable pseudo-label sample set will be used. Set it directly to an empty set. (Based on the quantity condition) To ensure the effectiveness of subsequent distribution alignment, the log-odds condition is used. Improve the reliability of target prediction results.

[0078] Subsequently, information entropy loss is introduced. For the selected reliable pseudo-label sample set Self-supervised learning is employed to reduce the influence of uncertain samples while enhancing the discriminative power of target domain features and the stability of the model's adaptive process. Information entropy loss is used. The calculation formula is as follows:

[0079] (11)

[0080] in, This represents the set of reliable pseudo-label samples for the current batch. Represents logarithmic probability The corresponding probability is calculated using the following formula:

[0081] (12)

[0082] 2) Distribution alignment:

[0083] After selecting the target pseudo-label, it is necessary to explicitly reduce the feature distribution difference between it and the source domain features, thereby gradually generating more high-confidence target prediction results and ensuring the effectiveness of the method of this invention. Therefore, this invention also introduces distribution alignment loss. :

[0084] (13)

[0085] in, Represents the source domain sample set The actual labels in the middle are categories The source domain samples are composed of the source feature matrix consisting of features extracted by the data feature extraction module in the mechanical fault diagnosis model. Represents a reliable set of pseudo-label samples. The target domain samples are composed of target feature matrices formed by features extracted by the data feature extraction module in the mechanical fault diagnosis model. Represents the source feature matrix and target feature matrix The distribution differences.

[0086] In practical applications, to protect the data privacy of the source domain, the source feature matrix... The following method can be used to generate the categories: features are extracted from the source domain samples by the data feature extraction module in the pre-trained source model, and then the categories are obtained according to the categories corresponding to the true labels. The source domain feature set is used to construct the source feature matrix from all source domain features. Then the source feature matrix Simply save the data; there is no need to save the source domain samples or regenerate the model using the current target domain model, which can effectively improve efficiency.

[0087] To accurately calculate the distribution difference measure, this embodiment proposes a distribution difference measure method based on scale-robust subspace calibration. Figure 5 This is a schematic diagram of the subspace calibration metric in this embodiment. For example... Figure 5 As shown, for the source feature matrix and target feature matrix First, the Singular Value Decomposition (SVD) algorithm is used for decomposition:

[0088] (14)

[0089] (15)

[0090] Among them, superscript Indicates transpose. , , Let represent the left singular vector matrix, singular value matrix, and right singular vector matrix of the source domain characteristic matrix, respectively. , , Let represent the left singular vector matrix, singular value matrix, and right singular vector matrix of the feature matrix of the target domain, respectively. It is important to note that... and It can be viewed as an orthogonal basis of the characteristic space, and the singular value matrix. It is a diagonal matrix, that is , Indicates the feature dimension.

[0091] Assuming the scale of the characteristic matrix is ​​characterized by its Frobenius norm, the following derivation can be obtained:

[0092] (16)

[0093] in, It is an identity matrix.

[0094] As can be seen from equation (13), the characteristic scale depends only on the singular value matrix. Therefore, to mitigate scale fluctuations caused by noise, this embodiment introduces a subspace calibration metric based on orthogonal bases to characterize the inherent cross-domain distribution differences. Intuitively, using cosine similarity... To quantify the differences between cross-domain basis vectors. (Left singular vector matrix) For example, this metric can be expressed as:

[0095] (17)

[0096] Similarly, we can obtain the right singular vector matrix. Corresponding cosine similarity To maintain consistency with the general property that "a larger numerical value indicates a greater difference" in variability measures, we will... and The cosine terms in the equation are converted to their corresponding sine forms. Therefore, the final subspace calibration metric can be defined as:

[0097] (18)

[0098] in, Represents the left singular vector matrix , No. The main angle of the dimensional feature, Describing a right singular vector matrix , No. The main angle of the dimensional feature.

[0099] It can be seen that when and When both are equal to 1, the distribution difference between the source domain and the target domain is minimized.

[0100] Then, the information entropy loss is applied. Alignment loss with distribution By performing weighted fusion, the final loss is obtained. :

[0101] (19)

[0102] in, , Represents information entropy loss Alignment loss with distribution The weight.

[0103] In addition, the parameters of the data feature extraction module can be frozen, the feature data feature extraction module can be provided to the target domain, and the parameter updates of the classifier can be made available to the target domain, thereby protecting the privacy and security of the source model while meeting the model transfer requirements.

[0104] S106: Target Domain Fault Diagnosis

[0105] When it is necessary to perform fault diagnosis on the mechanical equipment to be diagnosed, collect mechanical data samples at the current moment, input them into the mechanical fault diagnosis model trained in step S5, and obtain the corresponding fault category.

[0106] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the invention. In this embodiment, a self-testing simulation test bench is designed to collect experimental datasets. Figure 6 This is a structural diagram of the self-test simulation test bench in this example. (See diagram for example.) Figure 6 As shown, the self-test simulation test bench in this embodiment consists of a drive motor, a single-stage planetary gearbox, a magnetic powder brake, and a controller. The simulated faulty component is the planetary gear in the gearbox. This embodiment collects vibration monitoring signals of the planetary gear in six health states: normal state (NC), surface spalling (SP), root crack (RC), tooth wear (TW), micropitting (MP), and tooth fracture (TF). The sampling frequency of the accelerometer is set to 16384Hz. Three time-varying speed states are simulated: 0~600 rpm (V1), 0~1200 rpm (V2), and 0~2400 rpm (V3), with change rates of 100 rpm / s, 200 rpm / s, and 400 rpm / s, respectively. Finally, a constant speed state of 600 rpm (C) is also simulated. The sample dimension is set to 3072 to ensure that each sample contains at least one fault cycle. Sliding sampling is used to divide the sample size for each health state into 1000 samples. Based on the monitoring signals collected above, six time-varying migration diagnostic tasks can be constructed, namely V1→C, V2→C, V3→C, C→V1, C→V2 and C→V3.

[0107] To demonstrate the effectiveness and superiority of the time-varying mechanical fault migration diagnosis method based on passive progressive subspace calibration proposed in this invention, representative existing domain generalization diagnosis methods (WDCNN, Whitening-net, and I-Softmax), domain adaptation diagnosis methods (DANN, DDC, and MMSD), and passive domain adaptation diagnosis methods (SHOT, KTDA, and SFPSC) were selected as comparative methods. Table 1 compares the experimental results of the present invention and the comparative methods on six migration tasks in this embodiment.

[0108]

[0109] Table 1

[0110] As shown in Table 1, the method of the present invention has higher migration diagnosis accuracy and stronger generalization ability. In particular, the average diagnostic accuracy of the present invention on six cross-bearing migration diagnosis tasks reached over 80%, and its overall diagnostic accuracy was improved by approximately 25% compared to other migration diagnosis methods.

[0111] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A time-varying working condition mechanical fault migration diagnosis method based on passive progressive pseudo-labeling, characterized in that, Includes the following steps: S1: According to the actual needs, a plurality of mechanical data samples and their corresponding fault category labels are collected, and the number of fault categories is denoted as , thereby obtaining the source domain sample set; S2: Collect several mechanical data samples for the mechanical equipment to be diagnosed to obtain the target domain sample set; S3: Construct a mechanical fault diagnosis model according to actual needs, including a data feature extraction module and a classifier. The data feature extraction module is used to extract features from mechanical data samples, and the classifier is used to predict the corresponding fault category based on the extracted features. S4: The source domain model is obtained by training the mechanical fault diagnosis model using the source domain sample set; S5: Based on a progressive pseudo-label selection strategy, the source model is transferred and trained using the target domain sample set to obtain the target domain model. The loss function in each training batch is calculated as follows: Let the target domain sample set of the current training batch be . Each target domain sample Input the current mechanical fault diagnosis model to obtain its predicted category; set the predicted category as... The target domain samples constitute the sample set , Then, the following formula is used to filter each category. Reliable pseudo-label sample set : ， in, Indicates the number of individuals in the set. Represents the source domain sample set The true category is The number of samples, Represents the target domain sample. This indicates that the classifier in the mechanical fault diagnosis model is for each category. The generated log odds, , , The parameter is The mapping function from the sample space to the label space in the target domain model; This indicates sorting based on confidence level from highest to lowest. , This indicates a preset proportional threshold. , This indicates the preset exponential threshold. The information entropy loss is calculated using the following formula : ， wherein, represents the reliable pseudo-label sample set of the current batch, represents the logit the corresponding probability, the calculation formula is as follows: ； The distribution alignment loss is calculated using the following equation : ， in, Represents the source domain sample set The actual labels in the middle are categories The source domain samples are composed of the source feature matrix consisting of features extracted by the data feature extraction module in the mechanical fault diagnosis model. Represents a reliable set of pseudo-label samples. The target domain samples are composed of target feature matrices formed by features extracted by the data feature extraction module in the mechanical fault diagnosis model. Represents the source feature matrix and target feature matrix The distribution differences; The final loss is then calculated using the following equation : ， wherein, , represents a weight of information entropy loss and distribution alignment loss . S6: When it is necessary to perform fault diagnosis on the mechanical equipment to be diagnosed, collect mechanical data samples at the current moment, input them into the mechanical fault diagnosis model trained in step S5, and obtain the corresponding fault category.

2. The time-varying operating condition mechanical fault migration diagnostic method of claim 1, wherein, The formula for calculating the loss function of the source model training in step S4 is as follows: ， ， ， in, and These represent the source domain data sample and its data sample space, respectively. and These represent the source domain label and its label space, respectively. This indicates that the classifier in the mechanical fault diagnosis model is for each category. The generated log odds, , , The parameter is The mapping function from the sample space to the label space in the source domain model. This represents the classifier's classification for the true label in a mechanical fault diagnosis model. The generated log odds, , These represent the categories corresponding to the actual labels. and categories The probability, Indicates the category corresponding to the actual label. The relevant decision margin.

3. The time-varying operating condition mechanical fault migration diagnostic method of claim 1, wherein, The decision margin The calculation method is as follows: Get each category The natural logarithm of the log odds and standard deviation , The category is calculated using the following formula. and categories Overlap factor between the natural log-normal distribution of log odds : ， The decision interval is then calculated using the following equation : 。 4. The time-varying operating condition mechanical fault migration diagnostic method of claim 1, wherein, The distribution difference The calculation method is as follows: source feature matrix and target feature matrix are decomposed using singular value decomposition algorithm respectively: ， ， Among them, superscript Indicates transpose. , , Let represent the left singular vector matrix, singular value matrix, and right singular vector matrix of the source domain characteristic matrix, respectively. , , Let represent the left singular vector matrix, singular value matrix, and right singular vector matrix of the target domain feature matrix, respectively; The distribution difference is then calculated using the following equation : ， wherein denotes a left-singular vector matrix , the principal angle of the denotes a right-singular vector matrix , the principal angle of the 5. The time-varying operating condition mechanical fault migration diagnostic method of claim 1, wherein, The source feature matrix The generation method is as follows: features are extracted from the source domain samples in advance by the data feature extraction module in the pre-trained source model, and then the categories are obtained according to the categories corresponding to the true labels. The source domain feature set is used to construct the source feature matrix from all source domain features. .

6. The method for diagnosing the migration of mechanical faults under time-varying operating conditions according to claim 1, characterized in that, The parameters of the data feature extraction module are frozen during the transfer training, the feature data feature extraction module is provided to the target domain through a calling interface, and the parameter updates of the classifier are made available to the target domain.

Images