Fault diagnosis method for open set domain generalization under continuous variable working condition
By using a joint loss function of a class-specific autoencoder and a cross-domain alignment module, the problem of traditional methods struggling to identify unknown faults under continuous operating conditions is solved. This enables accurate fault diagnosis and identification of unknown faults in complex environments, improving the robustness and adaptability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional fault diagnosis methods struggle to cope with continuously changing operating conditions and identify unknown faults in complex and dynamic industrial environments, leading to decreased model performance and misjudgments.
By employing a class-specific autoencoder and a cross-domain alignment module, a joint loss function is constructed through semantic feature reconstruction and mutual information minimization to achieve cross-domain fault diagnosis and unknown fault identification.
It achieves cross-domain generalization capability and adaptive identification of unknown faults under continuously changing operating conditions, improving the robustness and adaptability of the model and ensuring the accuracy and stability of fault diagnosis.
Smart Images
Figure CN121744022A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent fault diagnosis of equipment, and in particular to a fault diagnosis method for open set domain generalization under continuous variable working conditions. BACKGROUND
[0002] With the development of intelligentization and complexification of industrial systems, equipment operation state monitoring and fault diagnosis technology plays an increasingly key role in modern manufacturing, transportation, power and aviation fields.
[0003] Traditional fault diagnosis methods are based on the assumption that training data and test data come from the same distribution, and usually rely on a large number of labeled samples and stable working conditions. However, in actual industrial environments, the working conditions of equipment are complex and variable, and the data distribution collected under different loads and speeds often differs, resulting in a sharp decline in model performance and making it difficult to meet the diversified needs in actual production.
[0004] In recent years, domain generalization methods have improved the model's generalization ability in unseen target domains by learning domain-independent feature representations in multiple source domains. However, in real industrial systems, due to equipment aging, material wear and tear, and other situations, the target domain may contain new fault samples that are not seen in the source domain. Traditional domain generalization models often misclassify unknown faults as known categories, lacking the ability to identify unknown samples.
[0005] In addition, the working conditions of industrial equipment often change continuously rather than discretely, such as dynamic changes in speed, load, etc. over time, forming a continuous domain distribution. Existing domain generalization methods are based on the assumption of discrete multi-domain, ignoring the smooth change characteristics between continuously changing working conditions, resulting in the model's inability to effectively model the working condition migration rules and gradual change characteristics.
[0006] Therefore, there is an urgent need for an intelligent fault diagnosis model that can achieve cross-domain fault classification under continuously changing working conditions while having the ability to identify unknown fault types. This method can significantly improve the robustness and adaptability of the model in complex, dynamic and unknown environments, and has important significance for the reliable operation of industrial equipment. SUMMARY
[0007] Therefore, the present application provides a fault diagnosis method for open set domain generalization under continuously changing working conditions, which can achieve fault diagnosis across continuously changing working conditions while having the ability to identify unknown fault types.
[0008] To solve the above technical problems, the present application is implemented as follows.
[0009] A fault diagnosis method for open set domain generalization under continuously changing working conditions, comprising: Step 1: Obtain time-domain data under different operating conditions and different fault categories, and construct training samples with operating condition index and fault category index; Step 2: Convert the time-domain data to the time-frequency domain, and extract semantic features using the feature extraction module; Step 3: A class-specific semantic reconstruction module consisting of a class-specific autoencoder and a classifier is used to classify semantic features. Each fault category corresponds to a class-specific autoencoder, which is trained using training samples from different operating conditions under the fault category. This allows the class-specific autoencoder to learn the common semantic representation of the corresponding fault category in different operating conditions and output the reconstruction error of the common semantic representation of the semantic features. The classifier generates the fault classification result based on the reconstruction error of all class-specific autoencoders. A cross-domain alignment module is used to estimate the mutual information between the semantic features of the training samples and the working condition index; the smaller the mutual information, the lower the degree of working condition dependence. Step 4: Construct a joint loss including cross-domain alignment loss; the cross-domain alignment loss is based on minimizing mutual information to constrain the consistency of feature distribution between continuous domains; backpropagation training is performed using the joint loss to enable the class-specific semantic reconstruction module to learn the ability to extract condition-invariant features. Step 5: During fault diagnosis, the time-domain data is converted to the time-frequency domain, and the fault classification results are obtained using the trained feature extraction module and class-specific semantic reconstruction module.
[0010] Preferably, in steps 2 and 5, the conversion of time-domain data to time-frequency domain is performed using continuous wavelet transform.
[0011] Preferably, in step 2, the semantic features extracted by the feature extraction module are: the time-frequency image converted to the time-frequency domain is pooled to a fixed size, and then input into a ResNet18 network with the classification head removed to extract semantic features.
[0012] Preferably, in step 5, when extracting the fault classification result using the trained feature extraction module, the classifier outputs the predicted label. Get predicted labels The reconstruction error output by the class-specific autoencoder for the corresponding fault category is compared with a threshold for the corresponding fault category. If the reconstruction error is greater than the threshold, it is determined to be an unknown category and used as the fault classification result; if it is less than or equal to the threshold, the predicted label is confirmed. The corresponding fault category is used as the fault classification result; the threshold is determined using the reconstruction error statistics of the training samples.
[0013] Preferably, the threshold is obtained by weighting the mean and standard deviation of the reconstruction error of the training samples.
[0014] Preferably, in step 3, the mutual information is obtained using a kernel-based mutual information estimation algorithm.
[0015] Preferably, the joint loss further comprises a classification loss and a reconstruction error loss; the classification loss adopts a classification cross-entropy loss; and the reconstruction error loss adopts a mean value of reconstruction errors of all training samples in a current round of epochs, the reconstruction error being an L2 norm of an error between the common semantics and the semantic features.
[0016] Preferably, the cross-domain alignment loss is calculated in the following manner: The mutual information is represented as a divergence between a joint distribution and a product distribution of two probability distributions corresponding to the semantic features and the working condition index. The divergence is approximated by a function using a linear combination of multiple Gaussian kernels, and parameters of the linear combination are The divergence is approximated by a function using a linear combination of multiple Gaussian kernels, and parameters of the linear combination are A convex optimization problem of minimizing the mutual information is solved to obtain an optimal estimation value of the parameter The optimal estimation value of the parameter is substituted into the mutual information estimation model to obtain an optimal estimation of the mutual information as the cross-domain alignment loss.
[0017] Advantages: (1) Realize cross-domain generalization under continuous working conditions: Traditional domain generalization methods are mostly based on discrete multi-domain assumption, ignoring the continuity of working condition changes, and are difficult to adapt to the smooth changes of load, speed and other working condition parameters in actual equipment operation. The present application introduces a continuous domain index strategy to calculate the mutual information between the semantic features and the continuous domain index (the working condition index). In training, the mutual information is added to the training target to effectively promote the extraction of domain-invariant features and ensure that the learned representation is robust to changes in working conditions, realizing cross-domain alignment. Therefore, the present application can learn domain-invariant features under continuously changing working conditions, realize smooth transfer from limited source domains to unseen target domains, and effectively improve the cross-condition robustness and generalization performance of the model.
[0018] (2) Has the ability of adaptive identification of unknown faults: In view of the problem that traditional methods can only identify known classes and are difficult to deal with unknown faults in actual production, the present application introduces a semantic reconstruction mechanism (class-specific semantic reconstruction module) combining auto-encoder and prototype learning after feature extraction, and measures whether the sample belongs to a known class by reconstruction error and prototype clustering distance, thereby realizing open set identification without relying on additional unknown samples. This design significantly improves the safety and practicality of the model, and can effectively avoid misjudgment of unknown faults as normal working conditions or other known classes.
[0019] (3) Unlike existing domain alignment strategies based on adversarial or statistical matching, the present application minimizes the mutual information between the feature distribution and the continuous domain index The mutual information between the continuous domain features and the cross-domain index is calculated, and a global consistency constraint on the continuous domain features is realized. The method effectively reduces the correlation between the domain features and the working condition changes while maintaining the discriminability of the features, so that the model can maintain stable diagnostic performance when the working condition gradually changes.
[0020] In summary, the open set domain generalization method for fault diagnosis under continuously changing working conditions introduced in the application introduces continuous domain index modeling, cross-domain class semantic reconstruction and cross-domain alignment mechanism under continuously changing working conditions, realizes the organic integration of domain generalization and open set identification. The method can accurately classify known faults and effectively identify unknown faults in complex dynamic working conditions, and significantly improves the robustness, reliability and practical application value of the intelligent fault diagnosis system. BRIEF DESCRIPTION OF DRAWINGS
[0021] Figure 1 The principle diagram of the fault diagnosis method for open set domain generalization under continuously changing working conditions is shown.
[0022] Figure 2 The flowchart of the fault diagnosis method for open set domain generalization under continuously changing working conditions is shown. DETAILED DESCRIPTION
[0023] The application provides a fault diagnosis method for open set domain generalization under continuously changing working conditions, which is used for mechanical fault diagnosis under continuously changing working conditions. The method combines methods such as class-specific semantic reconstruction (CSSR) and cross-continuous domain invariant feature extraction (CCIFE) to extract and model the features of cross-domain open set data of the continuous working condition index, and further updates the fault diagnosis method through joint loss function back propagation, so that the fault diagnosis method has stronger robustness and adaptability.
[0024] The application will be described in detail below with reference to the accompanying drawings and examples.
[0025] In this embodiment, intelligent fault diagnosis of rotating machinery is taken as an example to specifically describe the method of the application. The principle diagram and real-time flowchart of the application are shown in Figure 1 and Figure 2 , wherein steps 1-6 are the training process, and step 7 is the inference process. Referring to Figure 1 and Figure 2 , the method comprises the following steps: Step 1: obtaining training samples: obtaining time domain data under different working condition domains and different fault categories , constructing training samples with working condition index and fault category index , denoted as .
[0026] In this embodiment, the vibration acceleration signal data sets of rotating machinery under different working conditions are obtained, the invalid data at the beginning and end is removed, the vibration signals running under continuous changing working conditions are intercepted from the middle of the signal, and the data sets of different working condition domains are formed, each domain containing healthy and various fault data.
[0027] The one-dimensional time series signal data is divided into sliding windows to form several fixed-length data segments, and each time domain data segment is a single sample. Based on the time domain data segments collected and preprocessed, fault classification and working condition classification are performed to mark the fault classification labels and working condition labels to form training samples.
[0028] In practice, public data sets can also be used to form training samples by fault classification and working condition classification. The training samples have fault classification labels and working condition labels .
[0029] Step 2: Signal conversion and feature extraction: convert the time domain data to the time-frequency domain, and extract semantic features by the feature extraction module.
[0030] This step includes the following sub-steps: Step 2.1: Convert the time domain data to the time-frequency domain using continuous wavelet transform.
[0031] In this step, the input one-dimensional signal is analyzed by multi-scale wavelet kernel using continuous wavelet transform layer. First, the signal is subjected to fast Fourier transform (FFT), multiplied by the wavelet frequency domain kernel generated by a specified number of scales, and then subjected to inverse FFT and taking the amplitude to obtain the time-frequency feature map under different scales. This ensures multi-resolution representation of different frequency band signals and provides rich time-frequency information for subsequent convolution feature extraction.
[0032] Step 2.2: Extract semantic features using the feature extraction module.
[0033] In this step, the time-frequency map is unified to a fixed size by pooling, and then input into the ResNet18 network without the classification head to extract deep convolutional features.
[0034] The following steps 3 and 4 are parallel steps and there is no existing order.
[0035] Step 3: Class-specific semantic reconstruction across continuous domains.
[0036] This step uses a class-specific semantic reconstruction module composed of a class-specific autoencoder and a classifier, combining the autoencoder with prototype learning to construct a cross-continuous domain average prototype manifold for each class instead of a traditional single-point prototype, achieving semantic feature classification.
[0037] Specifically, each fault category corresponds to an independent class-specific autoencoder. ,These Instead of reconstructing the original image, the semantic features extracted by the network are used for reconstruction. Cross-domain data training is performed using training samples from different operating conditions under each fault category. A class-specific autoencoder for each fault category learns the common semantic representations of the corresponding fault category across different operating conditions, forming a domain-robust prototypical manifold. The reconstruction process employs... The norm measures the distance between semantic features and the prototype manifold, and is called the reconstruction error. This distance directly reflects the class affiliation of a sample; the classifier generates fault classification results based on the reconstruction error of all class-specific autoencoders.
[0038] First, a class-specific autoencoder is constructed. A class-specific autoencoder consists of an encoder, a latent encoding layer, and a decoder; the encoder... Convolutional layers are generated layer by layer according to the hidden layer list, progressively mapping the input features to high-dimensional hidden representations; latent encoding Convolution is performed on the output of the last hidden layer to obtain the latent feature representation. decoder This type of autoencoder reconstructs latent features back to the original input dimension through inverse hidden convolutional layers, recovering feature information layer by layer. It learns the feature distribution of each fault category through a process of "input → encoding convolution → latent encoding → decoding convolution → output reconstruction," achieving category-specific feature reconstruction for calculating reconstruction error or measuring sample attribution in open-set recognition.
[0039] Next, based on the number of fault categories, a corresponding number of autoencoders are created to initialize the category autoencoders. In this invention, the number of class-specific autoencoders is the same as the number of known fault categories, making them fault category-specific autoencoders. During training, samples from all operating conditions under the corresponding fault category are input, enabling the class-specific autoencoders to learn the common semantic representations of the corresponding fault category in different operating condition domains, serving as a prototype manifold.
[0040] Use each fault category (Category Labels) Class-specific autoencoders (use (representation) instead of category-specific point sets It will latent features As input, and output the reconstructed semantic representation. Then, using The regularization method calculates the reconstruction error as follows:
[0041] The reconstruction error of all class-specific autoencoders is input into the classifier, which generates fault classification labels. For example, for , is the sample set, let learn the class prototype, represents the probability that the class of the sample is . After normalization, the final probability is:
[0042] where is a hyperparameter that controls the difficulty of probability assignment, is the total number of classes. Considering the ideal solution that maximizes the output probability of the true class label, should first learn the minimum distance mapping from features to manifolds to minimize the reconstruction error with the true label of the class. At the same time, the manifolds of should also learn to maintain distances between each other to maximize the reconstruction error of other than corresponding to the true label.
[0043]
[0044] Step 4: Mutual information extraction for cross-continuous domain invariant feature extraction.
[0045] To further eliminate the distribution difference between different continuous domains, this step proposes a cross-domain alignment module based on mutual information, which estimates the mutual information between the semantic features of the training samples and the working condition index. The smaller the mutual information, the lower the dependence on the working condition. By designing the loss used for training, the mutual information is minimized, the feature and working condition correlation is reduced, and the class-specific semantic reconstruction module is trained to achieve domain-invariant feature extraction.
[0046] (1) Mutual information representation First, kernel-based mutual information estimation is performed. The mutual information between the semantic features and the continuous working condition index is calculated, and the mutual information calculation formula is:
[0047] where represents the set of continuous working condition indexes, ; is the set of sample time domain data ; is the semantic feature of the data ; represents taking the expectation value, denote the joint probability distribution, denote the joint probability,
[0048] mutual information is always non-negative, and reaches the minimum value of zero only when and are mutually independent. This means that when mutual information is minimized, the extraction of domain-invariant features is effectively promoted, ensuring that the learned features are robust to changes in working conditions and achieving cross-domain alignment.
[0049] The variational formula for the divergence between two probability distributions and is given by:
[0050] where denotes the supremum, denotes any possible function, denotes the probability under the probability distribution , denotes the probability under the probability distribution , and the formula traverses all possible functions , calculates the value of the expression inside the parentheses, and finally takes the supremum of these values. This maximum value is the divergence.
[0051] Mutual information can be interpreted as the divergence between the joint distribution and the product distribution . Therefore, the mutual information between the two can be re-expressed using the above divergence formula:
[0052] (2) Mutual information estimation When calculating the expected value in the expression for mutual information, the probability is expressed in a linear form. Using random sampling instances for small batch training, the mutual information can be estimated as:
[0053] where denotes the mutual information estimate, denotes the a sample, denotes the continuous working condition index of the sample,
[0054] In order to estimate the function , the idea of kernel method is used, a plurality of Gaussian kernels are used for function approximation, and a linear parameter is used to represent:
[0055] wherein, is an optimizable parameter, denotes the total number of Gaussian kernels, denotes the Gaussian kernel function, denotes a Gaussian kernel function with a bandwidth of , and is defined as:
[0056] The optimal estimation value of the parameter is obtained by solving the following convex optimization problem:
[0057] wherein, is a regularization coefficient for mitigating overfitting.
[0058] Finally, the optimal estimation parameter value is substituted into the mutual information expression, and the optimal mutual information can be estimated as:
[0059] In the subsequent training process, by adding the cross-domain alignment loss in the loss, the estimated mutual information is minimized, and the cross-domain alignment is realized.
[0060] Therefore, the way to obtain the mutual information in this step is summarized as follows: the mutual information is represented as the divergence between the joint distribution and the product distribution of the two probability distributions corresponding to the semantic features and the working condition index; a kernel-based mutual information estimation model is constructed, a linear combination of a plurality of Gaussian kernels is used for function approximation of the divergence, and the parameters of the linear combination are ; the optimal estimation value of the parameter is obtained by solving the convex optimization problem of minimizing the mutual information, substituted into the mutual information estimation model, and the optimal estimation of the mutual information is obtained as the cross-domain alignment loss. Step 5: Construct a joint loss.
[0061]
[0062] To realize the unified optimization of classification discrimination, feature reconstruction and cross-domain alignment, the application designs a joint loss function, including: a classification loss: based on cross-entropy loss to constrain the recognition performance of known categories; a reconstruction error loss: to constrain the reconstruction accuracy of the autoencoder, used to measure the difference before and after the sample reconstruction; a cross-domain alignment loss: based on mutual information minimization to constrain the consistency of feature distribution between continuous domains. The three are realized by weighted summation for end-to-end joint training, so as to have unknown class recognition ability while accurately classifying known faults, and realize adaptive generalization to continuous changing working conditions. According to the calculated joint loss, the model parameters are updated by back propagation.
[0063] Classification loss The classification cross-entropy loss is adopted, and the calculation method is:
[0064] Among them, is the output of the classifier; is the cross-entropy loss, based on the output of the classifier and the fault category label, the cross-entropy loss is calculated; represents the expectation of the joint distribution of the input sample and the label obeys, that is, the expectation of the cross-entropy loss is .
[0065] Reconstruction error loss The calculation method is:
[0066] Among them, represents the number of samples in each round epoch , and the reconstruction error of the th sample obtained by the aforementioned L2 regularization method is . In each epoch , the reconstruction error corresponding to all samples is averaged, that is, the reconstruction error loss.
[0067] Cross-domain alignment loss The calculation method is to minimize the estimated mutual information by solving the aforementioned convex optimization problem, as the cross-domain alignment loss .
[0068] The above loss functions are weighted and added to form a joint loss function for back propagation in the training process. The representation of the joint loss function is:
[0069] Among them is the weighted coefficient of the three.
[0070] Step 6: Model training.
[0071] Backpropagation training is performed using a joint loss. The goal of training the network is to minimize the combination of cross-entropy loss, reconstruction error loss, and estimated mutual information.
[0072] During training, all domains are cross-divided, and each time one domain is selected as the target domain, while the rest are treated as multiple source domains. A specific class is designated as the known class, and the rest as the unknown class. The model is trained using data from the known classes of the source domains.
[0073] In the backbone network and specific categories During the joint optimization process, the learning process occurs simultaneously. Furthermore, the model hyperparameters are set as follows: the learning rate is... The weight decays to , It is 32. The value is 100. The experimental environment is as follows: Intel(R) Xeon(R) Platinum 8358 CPU@2.60GHz, LINUX operating system, and NVIDIA L40 graphics card for GPU acceleration.
[0074] After training, the model can be tested and diagnosed. Several open-set fault diagnosis tasks are designed to evaluate the diagnostic performance of the proposed model. Test samples containing all categories in the target domain are input into the classifier, and the classifier outputs the diagnostic results, realizing the diagnosis of known faults and the identification of unknown faults in the target domain.
[0075] The evaluation metrics for the model include accuracy for known fault categories, accuracy for identifying unknown fault categories, and overall accuracy. The definitions and calculation formulas for these metrics are as follows: The number of correctly predicted samples for a given class; Number of correctly predicted samples for the unknown class; The total number of samples in each class is known. Total number of samples in the unknown class; Overall accuracy; Known class accuracy; Accuracy for unknown classes; It combines known classification accuracy with unknown detection recall.
[0076] Step 7: Fault diagnosis reasoning.
[0077] During fault diagnosis, only the time-frequency conversion, feature extraction module, and class-specific semantic reconstruction module are used.
[0078] First, the time-domain data is converted to the time-frequency domain. Then, semantic features are extracted using a pre-trained feature extraction module and input into a class-specific semantic reconstruction module. The classifier in this module outputs the predicted label. The classification probability and reconstruction error show a negative correlation, which is achieved through the module. Normalization yields the probability distribution of the final category, thus providing the predicted label. However, for unknown categories, the predicted label... The indicated fault category is incorrect because a threshold is used to determine whether it is an unknown category.
[0079] So, in obtaining predicted labels Then, further calls and Extract the reconstruction error from the corresponding class-specific autoencoder. Reconstruction error Thresholds corresponding to the fault categories The comparison is performed; the threshold is obtained by weighting the mean and standard deviation of the reconstruction error of the training samples, specifically:
[0080] in, The mean of the error. The standard deviation of the error. These are the weighting coefficients.
[0081] If reconstruction error Greater than the threshold If the reconstruction error is not specified, it is classified as an unknown category and used as the fault classification result; if the reconstruction error is not specified, it is considered an unknown category. Less than or equal to the threshold Then directly predict the label The corresponding fault category is used as the fault classification result.
[0082] This concludes the process.
[0083] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A fault diagnosis method for open set generalization under continuous variable working conditions, characterized in that, The method comprises the following steps: Step 1: Obtain time domain data under different working condition domains and different fault categories, and construct training samples with working condition indexes and fault category indexes; Step 2: Convert the time domain data to a time-frequency domain, and extract semantic features by a feature extraction module; Step 3: Class-specific semantic reconstruction modules composed of a class-specific autoencoder and a classifier are used to classify the semantic features; each fault category corresponds to a class-specific autoencoder, and training samples under different working condition domains of the fault category are used for training, so that the class-specific autoencoder learns the common semantic representation of the corresponding fault category in different working condition domains, and outputs the reconstruction error of the common semantic to the semantic features; the classifier generates a fault classification result based on the reconstruction error of all class-specific autoencoders; A cross-domain alignment module is used to estimate the mutual information between the semantic features of the training samples and the working condition indexes; the smaller the mutual information, the lower the working condition dependence; Step 4: A joint loss including a cross-domain alignment loss is constructed; the cross-domain alignment loss minimizes the consistency of feature distribution between continuous domains based on mutual information; the joint loss is used for back propagation training, so that the class-specific semantic reconstruction module learns the ability of extracting working condition invariant features; Step 5: During fault diagnosis, the time domain data is converted to a time-frequency domain, and the trained feature extraction module and class-specific semantic reconstruction module are used to obtain a fault classification result.
2. The method of claim 1, wherein, In steps 2 and 5, the conversion of the time domain data to the time-frequency domain uses continuous wavelet transform.
3. The method of claim 1, wherein, In step 2, the semantic features extracted by the feature extraction module are: the time-frequency image converted to the time-frequency domain is unified to a fixed size by pooling, and then input into a ResNet18 network without a classification head to extract semantic features.
4. The method of claim 1, wherein, In step 5, when extracting the fault classification result by using the trained feature extraction module, the classifier outputs a predicted label ; obtaining the predicted label The reconstruction error output by the corresponding fault category specific class-specific autoencoder is compared with the threshold value of the corresponding fault category. If it is greater than the threshold value, it is determined as an unknown category, which is used as the fault classification result; If less than or equal to threshold, confirm prediction label The corresponding failure class as the failure classification result; The threshold value is determined by statistical data of the reconstruction error of the training samples.
5. The method of claim 4, wherein, The threshold value is obtained by weighting the mean and standard deviation of the reconstruction error of the training samples.
6. The method of claim 1, wherein, In step 3, the mutual information is obtained by using a kernel-based mutual information estimation algorithm.
7. The method of claim 1, wherein, The joint loss further includes a classification loss and a reconstruction error loss; the classification loss uses a classification cross-entropy loss; the reconstruction error loss uses the mean of the reconstruction error of all training samples in the current round of epochs, and the reconstruction error is the L2 norm of the error between the common semantic and the semantic features.
8. The method of claim 1, wherein, The calculation method of the cross-domain alignment loss is: The mutual information is represented as the difference between the joint distribution and the product distribution of two probability distributions corresponding to the semantic features and the working condition index Divergence; constructing a kernel-based mutual information estimation model, using a linear combination of multiple Gaussian kernels to functionally approximate the Divergence, the parameters of the linear combination are Solving a convex optimization problem of minimizing mutual information to obtain the optimal estimation value of the parameter Substituting the mutual information estimation model to obtain the optimal estimation of the mutual information as the cross-domain alignment loss.