Cross-multi-source-domain industrial fault diagnosis method based on generalized zero sample learning

Abstract

The invention relates to a cross-multi-source-domain industrial fault diagnosis method based on generalized zero sample learning, and belongs to the field of industrial fault diagnosis. According to the method, fault diagnosis of a source domain and a target domain is realized through joint learning of multi-classifier and domain alignment consistency, and a latent hyper-sphere space with orthogonal constraint is constructed for connecting a feature space and a semantic attribute space, so that discriminative information is extracted and diagnosis of visible and non-visible faults is realized. According to the method, visible and non-visible faults can be efficiently and accurately diagnosed under the condition that no sample missing exists, and the generalization performance and practicability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the field of industrial fault diagnosis and relates to a cross-multi-source domain industrial fault diagnosis method based on generalized zero-sample learning. Background Art

[0002] Industrial fault diagnosis is one of the important technologies to ensure the stable, efficient and safe operation of industrial systems. With the increasing complexity of industrial equipment and systems, traditional model- and knowledge-based diagnostic methods are gradually being replaced by data-driven technologies. In recent years, machine learning methods, especially deep learning, transfer learning and zero-shot learning (ZSL) techniques, have been widely used in industrial fault diagnosis tasks. However, since factories cannot allow equipment to run into a fault state to obtain fault samples for training diagnostic models, and it is very difficult to obtain sufficient labeled data, traditional supervised learning methods have great limitations in practical applications.

[0003] As an advanced learning method, ZSL provides a new solution to the problem of missing data by leveraging semantic attributes to predict unseen samples. Although the ZSL method has performed well in the field of computer vision, its application in industrial scenarios still faces many challenges. The industrial field usually involves a variety of working conditions and environmental noise, and existing ZSL methods are insufficient in cross-domain generalization. Especially in multi-source domain fault diagnosis tasks and generalized zero-shot settings, how to achieve effective knowledge transfer between different domains and ensure the generalization ability of the model has become the focus of current technological development.

[0004] Most of the existing industrial fault diagnosis methods adopt the supervised learning mode and train the model through a large amount of historical data. For the diagnosis problem of few samples or unseen faults, transfer learning and zero-sample learning techniques are gradually introduced. ZSL belongs to the category of machine learning, which predicts zero-sample faults that have not been encountered in the training stage by using prior knowledge of visible faults (such as semantic attributes). Patent CN112763214B discloses a rolling bearing fault diagnosis method based on multi-label zero-sample learning, which pioneered fault diagnosis under zero-sample conditions, migrated visible class fault attributes to unseen classes, and effectively diagnosed unseen class faults. Patent CN114383845B discloses a bearing composite fault diagnosis method based on an embedded zero-sample learning model, which uses an embedded zero-sample learning model for composite fault diagnosis. However, under the setting of generalized zero samples, these methods often easily misclassify unseen faults as visible faults. In order to solve the problem of diagnosing both visible and unseen faults, patent CN113609569B discloses a discriminative generalized zero-shot learning fault diagnosis method, which decomposes the generalized zero-shot diagnosis task into supervised learning and zero-shot learning tasks by discriminating fault samples. However, the existing methods have not considered the distribution difference between the source domain and the target domain of actual working conditions, which limits the generalization ability in multiple source domains and different working conditions.

[0005] In summary, the existing technologies have the following shortcomings in multi-source domain fault diagnosis tasks:

[0006] (1) Poor cross-domain adaptability: Existing methods assume that the source domain and the target domain have similar distributions, which makes it difficult to adapt to distribution differences in actual industrial applications.

[0007] (2) Dependence on a single source domain: Most fault diagnosis methods focus on a single source domain and fail to effectively utilize multi-source domain data for the migration and fusion of diagnostic knowledge, resulting in insufficient generalization ability of the model under different working conditions.

[0008] (3) Poor generalization in general scenarios: Due to the lack of unseen fault samples, existing zero-shot fault diagnosis methods perform poorly in more practical general scenarios, especially in the classification task of unseen faults. Summary of the invention

[0009] In view of this, the purpose of the present invention is to provide a cross-multi-source domain industrial fault diagnosis method based on generalized zero-shot learning, to solve the problem of insufficient generalization ability of existing industrial fault diagnosis technology under multi-source domain conditions, especially the limitation of being unable to effectively distinguish between visible faults and unseen faults in generalized zero-shot fault diagnosis tasks. Existing methods usually rely on the same distribution assumption between the source domain and the target domain, and are limited to single-source domain scenarios, making it difficult to cope with distribution differences under different working conditions. The present invention proposes an unsupervised framework, which overcomes the limitation of the existing technology in insufficient diagnosis of unseen faults under multi-source domain conditions through multi-source domain knowledge transfer and latent space mapping, thereby achieving accurate diagnosis of visible and unseen faults in the absence of unseen samples, and improving the generalization performance and practicality of the model.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A cross-multi-source domain industrial fault diagnosis method based on generalized zero-shot learning is proposed. In order to achieve efficient and accurate classification of seen and unseen faults in industrial fault diagnosis tasks, especially when there are differences in data distribution in multiple source domains, fault diagnosis of source and target domains is achieved through joint learning of multiple classifiers and domain alignment consistency. A latent hypersphere space with orthogonal constraints is constructed to connect the feature space and the semantic attribute space, thereby extracting discriminative information and realizing the diagnosis of seen and unseen faults.

[0012] The method specifically comprises the following steps:

[0013] S1: Joint learning of multiple classifiers and cross-domain alignment consistency;

[0014] An independent classifier is constructed for each source domain, and the distribution consistency between the source domain and the target domain is achieved by aligning the source domain features with the unlabeled sample features of the target domain. The domain alignment strategy adopts the feature maximum mean difference method to ensure that the feature distribution of the source domain and the target domain, and the inter-domain distribution between multiple domains are unified in the same feature space.

[0015] S2: Fault prototype and semantic embedding learning;

[0016] A latent hypersphere space with orthogonal constraints is constructed for mapping learning between feature space and semantic attribute space. For visible faults in different source domains, learning is performed by obtaining corresponding sample labels during the training phase. Through domain alignment technology, all samples belonging to a specific category are averaged to obtain the prototype representation of the fault in that category. In the context of generalized zero-shot fault diagnosis, a strategy based on semantic similarity is adopted to determine the semantic attribute description of visible faults similar to unseen faults by calculating the Euclidean distance, thereby achieving effective reasoning for unseen faults. By learning the mapping containing the information of unseen faults, the connection between feature space and latent space and between attribute space and latent space is achieved respectively.

[0017] S3: Comprehensive fault diagnosis;

[0018] Based on visible faults, a model that can be classified across multiple source domains is learned, and a classifier with strong generalization ability is trained for the unlabeled target domain through domain alignment; this method diagnoses test samples containing visible faults and unseen faults in feature space, latent space and fault attribute space respectively, and at the same time, by introducing visible information in the multi-source domain learning process, a comprehensive diagnosis strategy is adopted to obtain the diagnosis results.

[0019] Furthermore, in step S1, the multi-classifier learning specifically includes: for the k-th source domain, which includes There are c fault samples and c fault categories, and the classifier C k Continuously optimize by calculating the classification loss J; the formula of the loss function is as follows:

[0020]

[0021] Among them, y i ∈S is the i-th sample x i The label of S is the visible fault category; I S (y ij ) is y i The jth value in the one-hot encoding vector of y, which can be regarded as an indicator function; if y i =j, then I S (y ij )=1; otherwise I S (y ij )=0;w j is the connection feature f(x i ,ω) and fault classifier C k The weight of the mth output neuron in , the superscript T indicates transposition; then, the multi-class learning loss L cls It can be expressed as:

[0022]

[0023] Where K is the number of classifiers, is the first classifier C k The loss function of .

[0024] Further, in step S2, the fault prototype and semantic embedding learning specifically includes: the fault prototype FP of the i-th visible class i s It can be obtained by:

[0025]

[0026] Among them, m i is the number of samples of the i-th visible class fault in all source domains, X s is the visible class sample space;

[0027] For the jth unseen fault, the first l (1<l≤c) similar visible attribute descriptions r(i) can be expressed as:

[0028]

[0029] Among them, Dis is the Euclidean distance function, is the jth unseen class semantic attribute vector, A s is the visible class attribute matrix, rank(t,Dis) represents the ascending order of t in Dis;

[0030] The jth unseen fault prototype FP j u It can be expressed as:

[0031]

[0032] Among them, FP s (r(i)) is the fault prototype with similar visible attributes to the i-th one.

[0033] Further, in step S3, the comprehensive diagnosis strategy specifically includes: assuming that the weighted distance set is in, is the weighted (normalized) distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th fault, is the weighted distance between the feature of the i-th sample in the latent space and the fault prototype of the j-th fault, is the weighted distance between the feature of the i-th sample in the attribute space and the fault prototype of the j-th fault, n te is the number of faulty test samples of the visible class, q is the number of unseen classes; the comprehensive label set It can be represented by the corresponding label value in the minimum distance space; at the same time, it is necessary to ensure that the label value is consistent with the classification result C[f(x i)] remain consistent; sample x i Tags It can be expressed as follows:

[0034]

[0035] in, θ ti,j represents the tth element of vector θ when i and j are determined; is the predicted label of the i-th sample in the feature space, is the predicted label of the i-th sample in the latent space, is the predicted label of the i-th sample in the attribute space; the predicted label It is expressed as follows:

[0036]

[0037] in, and is the normalized weighted distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th fault, is the normalized weighted distance between the feature of the i-th sample in the latent space and the fault prototype of the j-th fault, is the normalized weighted distance between the feature of the i-th sample in the attribute space and the fault prototype of the j-th fault; is the number of test samples with unseen class faults; represents FP uT , H u and A uT , * represents FP uT , H u and A uT I, B and D in T B, FP uT , H u and A uT They represent the prototype of the unseen fault class, the latent space of the unseen fault class and the attribute of the unseen fault class respectively, I represents the identity matrix, B represents the mapping between the feature space and the latent space, and D represents the mapping between the attribute space and the latent space; f k Represents the characteristics of the kth test sample.

[0038] Further, in step S3, the predicted label of the test sample The expression is:

[0039]

[0040] Among them, f i is the feature of the i-th test sample, n te is the number of test samples; Represents the matrix FP T The j-th column vector of (i.e., the FP of the j-th type of fault); is the weighted (normalized) distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th type of fault; Y is the set of fault types.

[0041] Furthermore, in step S3, in the latent space, the predicted label can be obtained by the learned mapping B:

[0042]

[0043] in, is the predicted label in the latent space, f i is the feature of the i-th test sample, n te is the number of test samples; H j is the latent space of the jth fault category; Y is the set of fault types.

[0044] Furthermore, in step S3, in the fault attribute space, by mapping B and D, the following prediction labels can be obtained:

[0045]

[0046] in, is the predicted label in the fault attribute space, f i is the feature of the i-th test sample, n te is the number of test samples; is the attribute matrix of the i-th fault class; Y is the fault class label; Y is the set of fault types.

[0047] The beneficial effects of the present invention are:

[0048] (1) Performance improvement: The unsupervised generalized zero-shot fault diagnosis framework proposed in this paper effectively improves the fault diagnosis performance in multi-source domains. Without relying on the same distribution assumption, it can achieve accurate diagnosis of visible and unseen faults under different working conditions (multi-source domains). By combining cross-domain alignment and semantic attribute space mapping, the performance and accuracy of fault diagnosis are significantly improved, especially when faced with large differences in data distribution, it can still maintain efficient fault diagnosis capabilities.

[0049] (2) Wide applicability: The method of the present invention has good generalization ability on different devices and under various working conditions, which effectively solves the problem that traditional methods fail under different working conditions. The existing technology usually assumes that the distribution of training and test data is the same, which limits the application scenarios of the model. The present invention overcomes the problem of distribution differences between different domains by aligning and transferring knowledge between multiple source domains, and is suitable for more diversified industrial scenarios.

[0050] (3) Cost advantage brought by unsupervised learning: Existing technologies mostly rely on a large amount of labeled data, but in industrial scenarios, the cost of obtaining labeled data is extremely high, especially when no fault is found. The present invention adopts unsupervised learning, which reduces the reliance on manual labeling, and reduces the cost of data labeling while ensuring diagnostic performance. This label-free diagnostic capability makes the method more economical and efficient in actual industrial applications.

[0051] (4) Improved accuracy and reliability of fault diagnosis: By constructing a latent hypersphere space with orthogonal constraints, the present invention can extract more discriminative representations between feature space and semantic attribute space. The discriminativeness of this representation ensures accurate identification of different fault categories and effectively avoids misclassifying unseen faults as visible faults, greatly improving the accuracy and reliability of diagnosis.

[0052] (5) Expanded scope of application: The present invention is not only applicable to a single device or working condition, but can also be applied in complex industrial scenarios in multiple source domains to meet the needs of different devices, working conditions and even different factories. Through cross-source domain learning, the present invention can achieve wider adaptability, greatly expanding its scope of application.

[0053] (6) Specific effects of achieving the purpose of the invention: Through unsupervised learning and comprehensive diagnosis strategies across multiple source domains, the present invention achieves effective knowledge transfer and successfully applies diagnostic knowledge in multiple source domains to the target domain. Even when fault data in the target domain is scarce or even unlabeled, it can still complete efficient fault diagnosis tasks, meet actual industrial needs, and demonstrate significant economic value and application prospects.

[0054] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0056] Figure 1 It is a schematic diagram of the generalized zero-sample fault diagnosis framework of the present invention;

[0057] Figure 2 It is a schematic diagram of a closed-loop continuous stirred reactor;

[0058] Figure 3 is the fault semantic attribute matrix of the continuous stirred reactor. DETAILED DESCRIPTION

[0059] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0060] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0061] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0062] See also Figure 1 to Figure 3 The present invention provides a method for cross-multi-source domain industrial fault diagnosis based on generalized zero-shot learning. The specific process is as follows:

[0063] Step 1: Multi-classifier learning;

[0064] For the kth source domain, it contains There are c fault samples and c fault categories, and the classifier C k Continuously optimize by calculating the classification loss J. The formula of the loss function is as follows:

[0065]

[0066] Among them, y i ∈S is the i-th sample x i I S (y ij ) is y iThe jth value in the one-hot encoding vector of y, which can be regarded as an indicator function. i =j, then I S (y ij )=1; otherwise I S (y ij )=0. j is the connection feature f(x i ,ω) and fault classifier C k The weight of the mth output neuron in . Subsequently, the multi-class learning loss can be expressed as:

[0067]

[0068] Step 2: Fault prototype and semantic embedding learning;

[0069] Fault prototype FP of the i-th visible class i s It can be obtained by:

[0070]

[0071] Among them, m i is the number of samples of the i-th visible class fault in all source domains.

[0072] For the jth unseen fault, the first l (1<l≤c) similar visible attribute descriptions can be expressed as:

[0073]

[0074] Where Dis is the Euclidean distance function, and rank(t,Dis) represents the ascending order of t in Dis. The jth unseen fault prototype can be expressed as:

[0075]

[0076] Step 3: Comprehensive fault diagnosis;

[0077] The predicted label of the test sample can be expressed as:

[0078]

[0079] Among them, f i is the feature of the i-th test sample, n te is the number of test samples. Represents the matrix FP T The j-th column vector of (i.e., the FP of the j-th type of fault). It is the weighted (normalized) distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th type of fault.

[0080] In the latent space, the predicted label can be obtained by the learned mapping B:

[0081]

[0082] Similarly, in the fault attribute space, by mapping B and D, we can get the following predicted labels:

[0083]

[0084] Collection-based By introducing the visible information in the multi-source domain learning process, a diagnostic strategy is proposed. Let the weighted distance set be Comprehensive label set It can be represented by the corresponding label value in the minimum distance space. At the same time, it is necessary to ensure that the label value is consistent with the classification result C[f(x i )] remain consistent. Sample x i Tags It can be expressed as follows:

[0085]

[0086] in, θ ti,j represents the tth element of vector θ when i and j are determined. Prediction label It is expressed as follows:

[0087]

[0088] in, and is the number of test samples without failure. · indicates FP uT , H u and A uT , * represents FP uT , H u and A uT I, B and D T B.

[0089] Verification experiment:

[0090] The experimental validation was based on a continuous stirred reactor system, which generated process data under different operating conditions by operating T during 20 hours of data acquisition. A total of 4 working conditions (C1: T = 430.88K, C2: T = 440.88K, C3: T = 400.88K and C4: T = 460.88K) were included, forming 4 different domains. In each case, 7 process variables with Gaussian noise were collected. In addition, the input C i 、T i and TCi is dynamically changing, randomly changing its nominal value every 60 minutes. The sampling interval is set to 1 minute, and the fault is introduced after 200 minutes. Table 1 lists the 10 initial process failure modes.

[0091] Table 1 Early failure categories of continuous stirred reactor system

[0092]

[0093] Table 2 Two different diagnostic tasks for a continuous stirred reactor process

[0094]

[0095]

[0096] Table 3 Comparison results in the continuous stirred reactor data set (%)

[0097]

[0098] The training samples only include 6 types of visible faults from 3 labeled source domains and 1 unlabeled target domain, with 350 samples for each fault. The total number of training samples is 8400 (350×6×4). The test samples include all faults in the target domain, with 350 instances for each type of visible fault and 525 instances for each type of unseen fault. The total number of test samples is 4200 (350×6+525×4).

[0099] The evaluation index is the average accuracy of the seen and unseen faults Acc s and Acc u , harmonic mean H1, and overall accuracy H2:

[0100]

[0101] Combining the experimental results in Table 2 and Table 3, it can be seen that the unsupervised generalized zero-shot fault diagnosis method proposed in the present invention further improves the ability of generalized zero-shot fault diagnosis. In general, this method can achieve satisfactory fault diagnosis performance, which is highly consistent with the goal of the generalized zero-shot learning task.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for cross-multi-source domain industrial fault diagnosis based on generalized zero-shot learning, characterized in that: The method specifically comprises the following steps: S1: Joint learning of multiple classifiers and cross-domain alignment consistency; An independent classifier is constructed for each source domain, and the distribution consistency between the source domain and the target domain is achieved by aligning the source domain features with the unlabeled sample features of the target domain. The domain alignment strategy adopts the feature maximum mean difference method to ensure that the feature distribution of the source domain and the target domain, and the inter-domain distribution between multiple domains are unified in the same feature space. S2: Fault prototype and semantic embedding learning; A latent space with orthogonal constraints is constructed for mapping learning between feature space and semantic attribute space. For visible faults in different source domains, learning is performed by obtaining corresponding sample labels during the training phase. All samples belonging to a specific category are averaged through domain alignment technology to obtain the prototype representation of the fault in that category. In the context of generalized zero-shot fault diagnosis, a strategy based on semantic similarity is adopted to determine the semantic attribute description of visible faults similar to unseen faults by calculating the Euclidean distance. The connection between feature space and latent space and between attribute space and latent space is achieved by learning the mapping containing the information of unseen faults. S3: Comprehensive fault diagnosis; Based on visible faults, a model that can be classified across multiple source domains is learned, and a classifier with strong generalization ability is trained for the unlabeled target domain through domain alignment; in the feature space, latent space and fault attribute space, the test samples containing visible faults and unseen faults are diagnosed respectively, and the diagnostic results are obtained by adopting a comprehensive diagnosis strategy by introducing the visible information in the multi-source domain learning process.

2. The method for cross-multi-source domain industrial fault diagnosis according to claim 1, characterized in that: In step S1, the multi-classifier learning specifically includes: for the kth source domain, which contains n k s There are c fault samples and c fault categories, and the classifier C k Continuously optimize by calculating the classification loss J; the formula of the loss function is as follows: Among them, y i ∈S is the i-th sample x i The label of S is the visible fault category, S = {y1,y2,···,y c };I S (y ij ) is y i The jth value in the one-hot encoding vector of ; if y i =j, then I S (y ij )=1; otherwise I S (y ij )=0;w j is the connection feature f(x i ,ω) and the fault classifier C k The weight of the mth output neuron in; subsequently, the multi-class learning loss L cls It is expressed as: Where K is the number of source domains, is the first classifier C k The loss function of .

3. The method for cross-multi-source domain industrial fault diagnosis according to claim 2 is characterized in that: In step S2, the fault prototype and semantic embedding learning specifically includes: the fault prototype FP of the i-th visible class i s Obtained through: Among them, m i is the number of samples of the i-th visible class fault in all source domains, X s is the visible class sample space; For the jth unseen fault, the first l similar visible attribute descriptions r(i) are expressed as: Among them, Dis is the Euclidean distance function, is the jth unseen class semantic attribute vector, A s is the visible class attribute matrix, rank(t,Dis) represents the ascending order of t in Dis; The jth unseen fault prototype It is expressed as: Among them, FP s (r(i)) is the fault prototype with similar visible attributes to the i-th one.

4. The method for cross-multi-source domain industrial fault diagnosis according to claim 3 is characterized in that: In step S3, the comprehensive diagnosis strategy specifically includes: assuming that the weighted distance set is in, is the weighted distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th fault, is the weighted distance between the feature of the i-th sample in the latent space and the fault prototype of the j-th fault, is the weighted distance between the feature of the i-th sample in the attribute space and the fault prototype of the j-th fault, n te is the number of test samples, q is the number of unseen classes; the comprehensive label set It is represented by the corresponding label value in the minimum distance space; at the same time, it is necessary to ensure that the label value is consistent with the classification result C[f(x i )] remain consistent; sample x i Tags It is expressed as follows: in, θ ti,j represents the tth element of vector θ when i and j are determined; is the predicted label of the i-th sample in the feature space, is the predicted label of the ith sample in the latent space, is the predicted label of the i-th sample in the attribute space; the predicted label It is expressed as follows: in, and is the normalized weighted distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th fault, is the normalized weighted distance between the feature of the i-th sample in the latent space and the fault prototype of the j-th fault, is the normalized weighted distance between the feature of the i-th sample in the attribute space and the fault prototype of the j-th fault; is the number of test samples with unseen class faults; represents FP uT , H u and A uT , * represents FP uT , H u and A uT I, B and D in T B, FP uT , H u and A uT They represent the prototype of the unseen fault class, the latent space of the unseen fault class and the attribute of the unseen fault class respectively, I represents the identity matrix, B represents the mapping between the feature space and the latent space, and D represents the mapping between the attribute space and the latent space; f k Represents the characteristics of the kth test sample.

5. The method for cross-multi-source domain industrial fault diagnosis according to claim 4 is characterized in that: In step S3, the predicted label of the test sample The expression is: Among them, f i is the feature of the i-th test sample, n te is the number of test samples; Represents the matrix FP T The j-th column vector of is the FP of the j-th type of fault; is the weighted (normalized) distance between the feature of the i-th sample in the feature space and the fault prototype of the j-th type of fault; Y is the fault class label.

6. The method for cross-multi-source domain industrial fault diagnosis according to claim 4, characterized in that: In step S3, in the latent space, the predicted label is obtained by the learned mapping B: in, is the predicted label in the latent space, f i is the feature of the i-th test sample, n te is the number of test samples; H j is the latent space of the jth fault category; Y is the fault class label.

7. The method for cross-multi-source domain industrial fault diagnosis according to claim 4, characterized in that: In step S3, in the fault attribute space, by mapping B and D, the following prediction labels are obtained: in, is the predicted label in the fault attribute space, f i is the feature of the i-th test sample, n te is the number of test samples; is the attribute matrix of the i-th fault class; Y is the fault class label.

Images