Multi-source semantic alignment adaptive meta transfer learning method for industrial fault diagnosis

Abstract

The invention discloses an industrial fault diagnosis-oriented multi-source semantic alignment adaptive meta transfer learning method. The method comprises the following steps of: 1, constructing a task of a meta training stage by adopting a leave-one method; step 2, pre-training; step 3, meta-training; step 4, fine adjustment of elements; and 5, calculating a test sample for testing. The key problems of fault sample scarcity and multi-source domain knowledge migration in the field of industrial fault diagnosis are effectively solved; more refined meta-knowledge extraction is realized, and cross-domain element loss is remarkably reduced; according to the method, prototype instance calibration and a micro-regression reprojection method are adopted, the influence of distribution difference among multiple domains on diagnosis performance is effectively reduced, in the fine adjustment stage, a reprojection prototype is replaced with a learnable matrix, a self-adaptive dynamically-adjusted prototype is obtained, optimization of meta-learner parameters is achieved through a semantic alignment bidirectional embedding module, and the accuracy and accuracy of diagnosis are improved. And the diagnosis precision and generalization ability of the model are further improved.

Description

Technical Field

[0001] This invention relates to the field of big data processing, and in particular to a multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis. Background Technology

[0002] In complex industrial systems, the combined effects of multiple factors, such as accumulated mechanical fatigue, sensor performance degradation, and dynamic changes in operating conditions, can lead to equipment failure, disrupting the continuity and stability of the production process and causing serious economic losses and safety risks. Therefore, researching and developing fault diagnosis methods with high accuracy and robustness is of great significance for ensuring the reliable operation of industrial systems, optimizing maintenance strategies, and improving overall economic efficiency.

[0003] The existing mainstream industrial fault diagnosis methods can be roughly divided into two categories: knowledge-driven methods and data-driven methods.

[0004] Knowledge-driven fault diagnosis methods formalize knowledge such as domain expert experience, system models, physical laws, and rule bases, and combine this with reasoning mechanisms to infer potential fault types from system input / output or observation data. However, with the increasing complexity and nonlinearity of industrial systems, acquiring expert knowledge has become increasingly difficult. Furthermore, these methods require high completeness and accuracy of knowledge and struggle to handle highly nonlinear or uncertain problems, limiting their application in fault diagnosis. In contrast, data-driven methods achieve higher diagnostic accuracy by mining the implicit patterns in collected data to accurately identify equipment status. Deep learning, as an important branch of data-driven methods, with network structures such as CNN, AE, and LSTM, is widely used in fault diagnosis due to its powerful feature extraction and representation learning capabilities. However, these methods typically assume that source and target data have the same distribution so that the learned model can effectively process test data.

[0005] To address the aforementioned issues, many scholars have introduced transfer learning into the field of fault diagnosis. Transfer learning-based methods utilize techniques such as feature matching, model and parameter transfer to help models identify differences in data distribution across different operating conditions. Cao et al. addressed the problem of large cross-domain data distribution differences in bearing fault diagnosis based on relative entropy and metric learning. Yin et al. proposed an end-to-end fault diagnosis model to solve the knowledge transfer between different operating condition domains. Besides these classic transfer learning-based methods, domain adaptation (DA) methods have been widely used in existing research in the field of fault diagnosis. Existing research mainly revolves around two core ideas. The first idea is to embed domain difference measurements into deep networks to reduce distribution inconsistencies. Li et al. combined maximum mean difference (MMD) and domain adversarial training to propose an intelligent fault diagnosis technique based on an improved domain adaptation method, effectively solving the problem of domain mismatch in monitoring data during industrial processes. Yu et al. addressed the dependence of traditional domain adaptation methods on labeled data in complex industrial scenarios by designing an adaptive feature transfer strategy. Another approach is to use domain adversarial neural networks inspired by generative adversarial networks to learn domain-uninformative features. Han et al. proposed a dual-domain adversarial network to effectively address the problem of contaminated characteristics of rolling bearing vibration signals in high-noise industrial environments. Lv et al. proposed a fault identification algorithm based on domain adversarial neural networks to solve the problem of cross-operating condition adaptability. Domain adaptation (DA) methods effectively reduce the impact of data distribution differences, but when the number of collected fault samples is small, DA methods struggle to accurately estimate distribution difference metrics, leading to feature alignment failure. Furthermore, the model also carries the risk of overfitting.

[0006] Despite significant progress in fault diagnosis methods, most existing industrial fault diagnosis methods are based on the assumptions of a single source domain and a single target domain. In real-world industrial scenarios, the risk of system failure due to the collection of equipment health status data often makes it difficult to obtain fault samples. This phenomenon leads to a decrease in the diagnostic accuracy of the constructed fault diagnosis model. Recent literature has attempted to address this issue through multi-source domain approaches. Zhang et al. designed a tensor-based nonlinear classifier using multi-source domains to improve the model's fault diagnosis performance. Ren et al. designed a domain generalization framework based on multi-source domains and MAML to solve the generalization problem of fault diagnosis models under low-resource heterogeneous data. While these existing studies have achieved some success, they have not fully considered the impact of distributional differences between multi-source domains on model performance and generalization ability. Therefore, the problem of multi-source knowledge transfer in multi-source domain fault diagnosis scenarios remains a challenging one.

[0007] Definitions: N-way K-shot is a core concept in few-shot learning, used to describe the combination of the number of classes (N) and the number of samples (K) for each class in a classification task.

[0008] One-hot encoding is a technique for converting categorical variables into binary vectors. It uses an N-bit state register to independently encode N states, with each state having only one valid bit (1) and the rest being 0. It is widely used in machine learning to process unordered categorical variables. Summary of the Invention

[0009] Based on this, the purpose of this application is to provide a multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis.

[0010] A multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis includes the following steps: Step 1: Construct the tasks for the meta-training phase using the leave-one-out method; Step 2: Pre-train the shared meta-learner using the semantically aligned bidirectional embedding module to obtain the pre-training parameters of the meta-learner. ; Step 3, Meta-training: Using pre-trained parameters The pre-trained meta-learner is obtained by initializing the meta-learner parameters during the training phase. Using pre-trained meta-learners Feature extraction is performed on the support set samples and query set samples in each pseudo-task. After prototype instance calibration and prototype reprojection, the reprojected prototype W*P is obtained. Finally, the classification result of the query sample is obtained according to Eq(6). During meta-training, the loss of each training round is calculated. , , This represents the first cross-entropy loss of the i-th pseudo-task. The training parameters are obtained after convergence or reaching the preset number of training iterations. and the corresponding optimized meta-learner W represents the projection matrix obtained by reprojecting the prototype, P represents the calibration prototype obtained by calibrating the prototype instance; Eq(6) is the probability that the query sample in the query set belongs to the nth category; Step 4, Meta-fine-tuning: Using a learnable matrix Instead of the reprojection prototype W*P, where... Let Z represent the learnable matrix, P be the initial value of the calibrated prototype obtained from the prototype instance calibration, T denote the matrix inversion, and Q denote the query sample features. Represents the regularization parameter. Let W represent the identity matrix, W represent the projection matrix obtained by reprojecting the prototype, and P represent the calibration prototype obtained by calibrating the prototype instance. Then use the optimized metalearner Feature extraction is performed using the second cross-entropy loss. Dynamically adjust R(Z) to Further optimization was performed, and R(Z) was used as the final prototype for target domain testing. express() Step 5: Calculate the test samples using Eq(8) with the optimized meta-learner parameters and the learnable matrix R(Z). The classification result is given by Eq(8), which represents the probability that the test sample belongs to the nth category, xt represents the test sample, and yt represents the label of the corresponding sample.

[0011] Further improvements are made, and the specific steps of step one are as follows: Given k known domains, each known domain takes turns acting as the target domain, and the remaining k-1 known domains serve as the source domains, forming a k-group training task, denoted as Task. k The known domain that acts as the source domain is designated as the pseudo-source domain, denoted as . The known domain that plays the role of the target domain is taken as the pseudo-target domain, denoted as... After determining the roles of the pseudo-source domain and pseudo-target domain in each task, each task is further divided into sub-tasks: In the i-th task, two known domains are randomly selected as the source domain and target domain for cross-domain training, with the source domain denoted as S. S The target domain is denoted as S. T The cross-domain training task constitutes a pseudo-task, which together form... Group pseudo-tasks, pseudo-tasks are denoted as The pseudo-target domain samples were randomly divided into two parts. ; This is used to fine-tune the training samples, specifically to fine-tune the parameters of the meta-learner and the meta-learner itself. The test samples are used for fine-tuning and final testing of the meta-learner; in each pseudo-task, data randomly drawn from the source domain is used as the support set samples, denoted as... Following the N-way K-shot approach, data randomly sampled from the target domain is used as the query set sample, denoted as... ; Indicates a pseudo-support set sample. express The corresponding pseudo-support set sample labels, This represents a sample of a pseudo-query set. express The corresponding pseudo-query set sample labels.

[0012] Further improvements are made, and the specific steps of step two are as follows: from Randomly sample from the data to form m sets of pre-training data. , This represents the i-th sample in the m-th domain. Let represent the j-th sample in the m-th domain; in the semantic alignment bidirectional embedding module, obtain the label reconstruction loss and alignment constraint loss of each group of pre-trained samples, and then calculate the average of the label reconstruction loss and alignment constraint loss as the loss function for each iteration to optimize the meta-learner parameters. The pre-trained parameters of the meta-learner are obtained when the loss function converges or reaches the preset number of iterations. and pre-trained meta-learners .

[0013] In a further improvement, the label reconstruction loss in step two is as follows: ; in, and These represent one-hot tags and refactored tags, respectively. , These represent the parameters of the tag encoder and tag decoder, respectively; S represents... Quantity; Represents the reconstruction loss of the labels. express and Cross-entropy loss between; The alignment constraint loss is as follows: ; in, , These represent label embedding and feature embedding, respectively. This represents the parameters of the meta-learner and the MLP network. Representation domain D S The number of training samples in the data. This represents the alignment constraint loss.

[0014] In a further improvement, in step two, the semantic alignment bidirectional embedding module is used for label reconstruction and feature embedding; The steps for tag reconstruction are as follows: One-hot encoding is performed on the original labels from different source domains to obtain one-hot labels, which are then fed into the label encoder. The label encoder maps the one-hot labels to the semantic space to obtain the label embedding in the semantic space. The label decoder reconstructs the label embedding in the semantic space into the original labels and constrains the label reconstruction process with the label reconstruction loss. The steps of feature embedding are as follows: A meta-learner is used to extract features from the original input data from multiple sources. The features are then mapped to the semantic space through an MLP network to generate feature embeddings. Alignment constraint loss is used to ensure the consistency between feature embeddings and label embeddings.

[0015] In a further improvement, the prototype instance calibration method in step three is as follows: The category prototype is calculated by reweighting each sample in each category based on its similarity to the remaining samples in that category: , , in, This represents the adaptive weight calculated based on the similarity score between the i-th sample in the n-th category and the remaining samples in that category; This represents the i-th sample in the n-th category of the support set. This represents the j-th sample in the n-th class of the support set. This represents the feature of the j-th sample in the n-th class of the support set. This represents the feature of the i-th sample in the n-th category of the support set. This indicates that for input X, the meta-learner... The obtained features; This represents the prototype of the nth category. This represents the similarity of the Kth sample in the support set. This represents the feature of the Kth sample in the nth category of the support set, where N represents the number of categories in the support set.

[0016] In a further improvement, the method for prototype reprojection in step three is as follows: The reprojection matrix is ​​constructed using a differentiable regression approach. : ; in, , Q represents the query sample feature. Denotes the regularization parameter, and W represents the projection matrix. This represents the prototype of the i-th category. Let R represent the optimal projection matrix. dm*dm dm represents the dimension of W, and dm represents the feature dimension; Based on the closed-form solution obtained from the original expression Eq(4) of ridge regression, the prototype is reprojected as: ; Where n represents the number of categories supporting the samples in the set; Indicates the projection prototype. dm represents the projection prototype dimension, and dm represents the feature dimension. After obtaining the projection prototype, for a randomly given query sample The probability that it belongs to the nth category. Represented as: ; This represents the calibration prototype for the nth category. Exponential function, Euclidean distance calculation function The pre-trained meta-learner is trained using the first cross-entropy loss. Parameters in: ; in, This represents the loss of the pseudo-task, which is also the first cross-entropy loss of the n categories; S represents S S T Expectations between them.

[0017] In a further improvement, in step four, R(Z) is used as the final prototype for target domain testing. ; ; in, This represents the second cross-entropy loss. Indicates the query sample X f The probability of belonging to the nth category. For adaptive adjustment of the prototype for the nth category, This represents the query sample during the meta-fine-tuning phase.

[0018] In summary, this application includes the following beneficial technical effects: (1) An adaptive meta-transfer learning framework (MSA-AMTL) based on multi-source domain semantic alignment is proposed, which effectively solves the key problems of scarce fault samples and multi-source domain knowledge transfer in the field of industrial fault diagnosis.

[0019] (2) A novel task partitioning strategy was designed. By further partitioning pseudo-tasks within the meta-task, more refined meta-knowledge extraction was achieved, significantly reducing cross-domain meta-loss.

[0020] (3) A Prototype Feature Mapping Network (PFRN) was designed, and a prototype instance calibration and differentiable regression reprojection method were adopted to effectively reduce the impact of distribution differences between multiple domains on diagnostic performance. In the fine-tuning stage, a learnable matrix was used to replace the reprojected prototype to obtain an adaptively dynamically adjusted prototype.

[0021] (4) A semantically aligned bidirectional embedding module (SABE) was designed. By semantically embedding the latent attributes of the labels, the parameters of the meta-learner were optimized, which further improved the diagnostic accuracy and generalization ability of the model. Attached Figure Description

[0022] Figure 1 The structure diagram for the semantically aligned bidirectional embedding module, which pre-trains the shared meta-learner; Figure 2 A diagram illustrating the division of tasks.

[0023] Figure 3 A schematic diagram of the process of reprojecting prototype features. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. An adaptive meta-transfer learning method for multi-source semantic alignment in industrial fault diagnosis consists of three learning stages and two basic modules, such as... Figure 1 As shown, the first part proposes a novel task partitioning strategy based on meta-learning for learning high-level meta-knowledge between different source domains. The second part proposes a SemanticAligned Bidirectional Embedding module based on the semantic embedding idea of ​​the first part, to obtain a more generalized meta-learner. The third part proposes a Prototypical Feature Reprojection Network (PFRN), which calibrates prototypes of different categories by calculating the similarity of different samples within the same category, and reprojects the prototypes according to the relationship between the calibrated prototypes and query features. Based on the modules mentioned above, the fourth part introduces the overall framework and steps of this invention.

[0025] Part 1: Task Split Strategy In industrial fault diagnosis setups, the collected samples are typically time-series data, represented as... L represents the length of the time series, and d is the number of sensors collecting data. The input space is represented as X, the label space as Y, and the neighborhood D is defined as the joint distribution P of X×Y. X×Y .domain Let DT be the sample data collected from domain S. Assume the source domain consists of a set of domains S = {D1, ..., Dk}, and these different source and target domains DT can be represented by a joint distribution. The inconsistency is used to define it.

[0026] To learn advanced meta-knowledge across different source domains and extract advanced domain-invariant representations with low meta-loss, this paper designs a novel task partitioning strategy based on the characteristics of task partitioning in meta-learning. It employs a leave-one-out method to construct tasks for the meta-training phase. In other words, for k known domains, each domain takes turns acting as the target domain, and the remaining k-1 domains serve as source domains, forming k sets of meta-training tasks, denoted as Task. k Among them, the domain that plays the role of the source domain is called the pseudo-source domain, denoted as . The domain that acts as the target domain is designated as the pseudo-target domain, denoted as... After determining the roles of the source and target domains in each task, each task is further divided into subtasks. Specifically, in the i-th task, two domains are randomly selected as the source and target domains for cross-domain training, with the source domain denoted as S. S The target domain is denoted as S. T The cross-domain training task constructed in this way serves as a pseudo-meta-training task, comprising... Group pseudo-tasks, pseudo-tasks are denoted as The pseudo-target domain samples were randomly divided into two parts. , Used for fine-tuning the model. Used for final testing. In each pseudo-task, data randomly sampled from the source domain is used as the support set, denoted as . Following the N-way K-shot approach, data randomly sampled from the target domain is used as the query set, denoted as... Taking four domains as an example, the task division diagram is as follows: Figure 2 As shown. At this point, according to the description above, there are a total of 4 groups of meta-training tasks, and each group of meta-tasks is further divided into 3 groups of pseudo-tasks. Task 4 For example, S1, S2, and S3 are... S4 as .right Task 4 Further divide the pseudo-tasks into 3 groups, namely , and In the fine-tuning and testing phases of each pseudo-task, As the source domain for these two stages, and These are used as the target domains for the two stages, respectively. This approach allows the meta-learner to effectively extract shared intrinsic features from different source domains, enhancing its ability to extract domain-invariant representations while minimizing meta-loss and improving model performance in the target domain.

[0027] Part 2: Semantic Aligned Bidirectional Embedding

[0028] To maximize the use of data from different source domains, achieve cross-domain training, and address the problem of insufficient data resources, this paper designs a semantically aligned bidirectional embedding module to pre-train a shared meta-learner, such as... Figure 1 As shown. The semantically aligned bidirectional embedding module is inspired by the idea of ​​semantic embedding. Through feature embedding alignment, label embedding into the semantic space, and reconstructing label representations, it fully utilizes the latent attributes of labels to optimize the parameters of the meta-learner. Specifically, firstly, the original labels from different source domains are one-hot encoded and then fed into the label encoder. The label encoder maps the one-hot labels into the semantic space. Then, the label decoder reconstructs the original labels from the label embeddings in the semantic space. To avoid the loss of label information during the encoding-decoding process, a reconstruction loss is used to constrain this process. The reconstruction loss is expressed as:

[0029] in, and These represent one-hot tags and refactored tags, respectively. , These represent the parameters of the tag encoder and tag decoder, respectively; S represents... The number of [items / objects] is determined. Simultaneously, a meta-learner is used to extract features from the original input data, and an MLP network is used to map these features to the semantic space, generating feature embeddings. To ensure consistency between the feature embeddings and the label embeddings, an alignment constraint loss is used to align them. The alignment constraint loss is expressed as:

[0030] in, , These represent label embedding and feature embedding, respectively. This represents the parameters of the tag encoder. This represents the parameters of the meta-learner and the MLP network. Representation domain D S The number of training samples in the dataset.

[0031] In traditional supervised learning, the original labels are fixed and serve as supervisory signals to guide model training. This optimization process is unidirectional. In the semantically aligned bidirectional embedding module proposed in this paper, we achieve bidirectional optimization. On one hand, one-hot labels are mapped to the semantic space by a label encoder, forming learnable label embeddings, which dynamically guide the meta-learner to generate corresponding feature embeddings through an MLP. On the other hand, the label embeddings in the semantic space are reconstructed into one-hot labels by a label decoder, and the reconstruction loss is used to optimize the parameters of the label encoder and decoder. Feature embeddings influence the optimization of label embeddings, while label embeddings, in turn, adjust the feature distribution, thus forming bidirectional optimization.

[0032] Part 3: Prototypical Feature Reprojection Network 3.1 Prototype Instance Calibration In traditional prototype networks, given a one-dimensional time series sample x, its features can be represented as: , Let represent the meta-learner, B represent the batch size, and dm represent the feature dimension. The prototype of the corresponding category is obtained by calculating the mean of the features of each category's samples. The prototype of the nth category is represented as... The query sample is classified by calculating the distance between it and the class prototype. During the meta-learning scenario training process, the support set samples follow an N-way K-shot setup. Due to the scarcity of samples, this leads to a significant problem: during testing with samples from the target domain, the network consistently uses these limited samples to build fixed prototypes. If outliers exist among the samples used to compute the prototypes, this can cause severe inductive bias during gradient descent.

[0033] To address this, this paper designs a prototype fine-tuning strategy, which reweights each sample based on its similarity to the remaining samples in each category, and calculates the category prototype: , , in, This represents the adaptive weight calculated based on the similarity score between the i-th sample in the n-th category and the remaining samples in that category. This simple method allows for easy recalibration of instances to obtain the correct prototype.

[0034] 3.3.2 Prototypical Feature Reprojection After obtaining the prototypes for each category, the traditional approach is to calculate the distance between the query sample and the support prototypes to classify the query sample. However, in our scenario, the support set and query set samples come from different domains, resulting in domain drift. Therefore, our goal is to reproject the prototypes onto the query sample space. Based on this idea, we construct the reprojection matrix using differentiable regression. To learn this kind of embedding, such as Figure 3 As shown.

[0035] Previous work has shown that reprogramming these prototypes using differentiable closed-loop solutions demonstrates their effectiveness:

[0036] in, , , indicates that the query sample features. Let represent the regularization parameter. The specific reprojection is learned by preserving the spatial dimension R of the feature map. Based on the closed-form solution of Eq(4), the prototype is reprojected as:

[0037] Where n represents the number of categories supporting the samples in the set.

[0038] After obtaining the projection prototype, for a randomly given query sample The probability that it belongs to the nth category can be expressed as:

[0039] Where Ed(·) represents the Euclidean distance calculation function. Cross-entropy loss is used... Training meta-learners Parameters in:

[0040] in, This represents the loss of the pseudo-task, which is also the cross-entropy loss of the n categories.

[0041] Part Four: Adaptive MSA-AMTL Framework like Figure 1 As shown, in MSA-AMTL, we summarize the learning process of multi-source domain cross-domain few-shot learning (MSCDFSL) into four progressive processes: pre-training, meta-training, meta-fine-tuning, and meta-testing.

[0042] 1) Pre-training: In the pre-training phase, from Randomly sample from the data to form m sets of pre-training data. , Let represent the i-th sample in the m-th domain. The goal of pre-training is to learn a more generalized meta-learner. In the semantic alignment bidirectional embedding module, the label reconstruction loss and alignment constraint loss for each pre-trained sample group are calculated at each iteration. The average of the label reconstruction loss and alignment constraint loss for each pre-trained sample group is used as the loss for each iteration to optimize the meta-learner. Finally, the generalized network parameters are obtained. .

[0043] 2) Meta-training: After the pre-training phase, the pre-training parameters are used. As the initial parameters of the meta-learner during the training phase. In each meta-task, the meta-learner is used to extract features from the support set samples and query set samples in each pseudo-task, and a reprojected prototype is obtained after prototype instance calibration and prototype reprojection. Finally, the classification result of the query sample is calculated according to Eq(6). The loss of each training round is calculated according to the cross-entropy loss of each pseudo-task. ,right Optimization results .

[0044] 3) Meta-fine-tuning: During the meta-training phase, the projected prototype W*P obtained through differentiable regression remains fixed. Due to matrix inverse operations, it is merely a near-optimal embedding. Here, a learnable matrix is ​​used. Instead of calculating W*P, use... Feature extraction is performed through R(Z) is dynamically adjusted. Simultaneously, Further optimization was performed. R(Z) was then used as the final prototype for testing in the target domain.

[0045]

[0046] 4) Meta-testing: In the final testing phase, the optimized meta-learner parameters and the projected prototype R(Z) are used to calculate the test samples via Eq(8). The classification results.

[0047] Part 5: Verification Method In comparative experiments, the MSA-AMTL of this invention is compared with several state-of-the-art and fundamental methods. First, MSA-AMTL is compared with transfer learning-based methods, such as DANN, CIATL, and DMsrTTLN. Second, MSA-AMTL is compared with CDFSL (cross-domain few-shot learning)-based methods, such as Proto, LOGO, ATA, and MetaOptNet-SVM. Finally, MSA-AMTL is compared with state-of-the-art MSCDFSL (multisource cross-domain few-shot learning) methods, such as DAC-Net, SMRL, Meta-GENE, and MRTN. It should be noted that, for comparison methods using two-dimensional image data as input, to ensure comparability with the one-dimensional time-series data presented in this paper, the 2D-CNN feature embedding networks in the relevant comparison methods are uniformly adjusted to 1D-CNN.

[0048] To enhance the scientific rigor and credibility of the comparative experiments, this paper adjusted the parameters of each method based on relevant literature to obtain optimal performance. For fair comparison, in the case of multi-source domain transfer, for each task, all methods employing the single-source domain to single-target domain diagnostic model were trained separately for each source domain. The performance of each single-source transfer model was evaluated individually, and the optimal value among the three results was taken as the representative performance of the method. The classification accuracy for the i-th task is represented by the average accuracy. We conducted experimental validation at the Tennessee Eastman Process (TEP), Casey Western Reserve University Rolling Bearing Dataset (CWRU), and Aluminum Electrolysis Process (AEP):

[0049]

[0050]

[0051] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and shown herein.

Claims

1. A multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis, characterized in that: Comprising the following steps: Step one, using leave-one-out method to construct the task of meta-training stage; Step two, pre-training the meta-learner by the semantic alignment bidirectional embedding module to obtain the pre-training parameters of the meta-learner ; Step three, meta-training: using pre-training parameters initializing the meta-learner parameters in the training phase to obtain a pre-trained meta-learner ; using the pre-trained meta-learner Feature extraction is performed on the support set samples and query set samples in each pseudo task, and the re-projection prototype W*P is obtained through prototype instance calibration and prototype re-projection. Finally, the classification result of the query sample is obtained according to Eq(6); during meta-training, the loss of each training round is calculated , , represents the first cross-entropy loss of the i-th pseudo task, converge or reach the preset number of training times to obtain the optimized training parameters and the corresponding optimized meta-learner ; W represents the projection matrix obtained by prototype re-projection, and P represents the calibration prototype obtained by prototype instance calibration; Eq(6) is the probability that the query sample in the query set sample belongs to the n-th class Step four, meta-fine tuning: using learnable matrix Instead of re-projecting prototypes W*P, where, represents a learnable matrix, the calibration prototypes P obtained by prototype instance calibration as the initial value of Z, T represents matrix inversion, Q represents query sample features, represents a regularization parameter, represents an identity matrix, W represents a projection matrix obtained by prototype re-projection, and P represents a calibration prototype obtained by prototype instance calibration; Then use the optimized meta-learner Feature extraction is performed, and a second cross-entropy loss R(Z) is dynamically adjusted, and Optimization is performed again, and R(Z) is used as the final prototype for target domain testing. Indicates Step five, using the optimized meta-learner parameters and the learnable matrix R(Z) to calculate the classification results of the test samples by Eq (8) Eq (8) represents the probability that the test sample belongs to the nth class, xt represents the test sample, and yt represents the label of the corresponding sample.

2. The multi-source semantic alignment self-adaptive meta-transfer learning method for industrial fault diagnosis according to claim 1, wherein: The specific steps of step one are as follows: For k known domains, let each known domain take turns to play the role of target domain, and the remaining k-1 known domains as source domains, to form k groups of meta-training tasks, denoted as Task k ; the known domain playing the role of source domain as a pseudo-source domain, denoted as , the known domain playing the role of target domain as a pseudo-target domain, denoted as ; after determining the roles of the pseudo-source domain and the pseudo-target domain in each task, further divide the sub-tasks in each task: in the i-th task, randomly select 2 known domains as source domain and target domain for cross-domain training, the source domain is denoted as S S , the target domain is denoted as S T , the cross-domain training task formed is a pseudo-task, and a total of pseudo-tasks are formed, denoted as ; The pseudo-target domain samples are randomly divided into two parts, ; For fine-tuning training samples, used for fine-tuning the meta-learner parameters and the meta-learner, For fine-tuning test samples, used for final testing of the meta-learner; in each pseudo task, the data randomly extracted from the source domain is used as the support set sample, denoted as ; Following N-way K-shot, the data randomly sampled from the target domain as the query set samples, denoted as ; denotes the pseudo-support set samples, denotes the corresponding pseudo-support set sample labels, denotes the pseudo-query set samples, denotes the corresponding pseudo-query set sample labels.

3. The multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis according to claim 2, characterized in that: The specific steps of step two are as follows: from m groups of pre-training data are composed of random sampling from , represents the i-th sample in the m-th domain, represents the j-th sample in the m-th domain; the label reconstruction loss and the alignment constraint loss of each group of pre-training samples are obtained in the semantic alignment bidirectional embedding module, then the label reconstruction loss and the alignment constraint loss are averaged as the loss function optimization meta-learner parameters in each iteration, and the pre-training parameters of the meta-learner are obtained when the loss function converges or reaches the preset number of iterations and the pre-trained meta-learner .

4. The multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis according to claim 3, characterized in that: In step two, the label reconstruction loss is as follows: ； wherein, and denote one hot label and reconstruction label, respectively; , denote parameters of label encoder and label decoder, respectively; S denotes the number of denotes reconstruction loss of label, denotes and cross-entropy loss between The alignment constraint loss is as follows: ； where, , denote the label embedding and the feature embedding, respectively, denote the parameters of the meta-learner and the MLP network, denote the number of training samples in domain D S denote the alignment constraint loss.​ 5. The multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis of claim 1, wherein: In step two, the semantic alignment bidirectional embedding module is used for label reconstruction and feature embedding; The steps of label reconstruction are as follows: The original labels from different source domains are one-hot encoded to obtain one-hot labels, which are then fed into the label encoder to map the one-hot labels to the semantic space to obtain label embeddings in the semantic space, and the label embeddings in the semantic space are reconstructed into original labels through the label decoder, and the label reconstruction process is constrained through the label reconstruction loss; The steps of feature embedding are as follows: The meta-learner is used to extract features from the original input data of multiple source domains, the features are mapped to the semantic space through the MLP network to generate feature embeddings, and the consistency of the feature embeddings and the label embeddings is ensured through the alignment constraint loss.

6. The multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis of claim 1, wherein: In step three, the prototype instance calibration method is as follows: According to the similarity of each sample in each category to the rest of the samples in the category, each sample is reweighted, and the category prototype is calculated: ， ， wherein, represents the adaptive weight of the similarity score calculation of the i-th sample in the n-th class with the rest of the samples in the class; represents the i-th sample in the n-th class of the support set, represents the j-th sample in the n-th class of the support set, represents the features of the j-th sample in the n-th class of the support set, represents the features of the i-th sample in the n-th class of the support set, represents the features obtained by the meta-learner for the input X, the features obtained by the meta-learner for the input X. represents a prototype of the n-th class, represents a similarity of the K-th sample of the support set, represents a feature of the K-th sample of the n-th class of the support set, N represents a number of classes of the support set.

7. The multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis according to claim 1, characterized in that: In step three, the method of prototype re-projection is as follows: Constructing a re-projection matrix by means of a differentiable regression : ; wherein, , , Q denotes query sample features, denotes a regularization parameter, W denotes a projection matrix, denotes a prototype of the i-th class, denotes an optimal projection matrix, R dm*dm denotes a dimension of W, dm denotes a feature dimension; According to the closed solution obtained from the original expression Eq(4) of ridge regression, the prototype is re-projected as: ; wherein n represents the number of classes of the support set samples; denotes the projection prototype, denotes the projection prototype dimension, and dm denotes the feature dimension; After obtaining the projection prototypes, for a randomly given query sample the probability that it belongs to the nth class is represented as: ; represents a calibration prototype of the n-th class, an exponential function, represents a Euclidean distance computation function Training parameters in a pre-trained meta-learner by a first cross-entropy loss : ; wherein, represents the loss of the pseudo task, which is also the first cross-entropy loss of n classes; represents S S , S T between expectations.

8. The multi-source semantic alignment adaptive meta-transfer learning method for industrial fault diagnosis of claim 1, wherein: In step four, R(Z) is used as the final prototype for target domain testing; ; ; wherein, denotes a second cross-entropy loss, denotes a query sample X f a probability that the query sample belongs to the nth class, is an adaptive adjustment prototype for the nth class, denotes a meta-fine-tuning stage query sample.

Images