Any sample fault diagnosis method based on hierarchical feature learning

Abstract

The invention relates to an arbitrary sample fault diagnosis method based on hierarchical feature learning, and belongs to the field of industrial fault diagnosis. The method comprises the following steps: constructing a hierarchical auto-encoder, gradually extracting potential feature representations of data samples from a bottom layer to a high layer by using the hierarchical auto-encoder, and extracting potential attribute representations of the data samples at the same time; aligning the extracted potential feature characterization and the extracted potential attribute characterization; inputting the aligned potential feature representations and data samples into an attribute classifier, predicting semantic attributes of the data samples through the attribute classifier, and performing zero sample and small sample fault diagnosis based on the predicted semantic attributes; and calculating a fault category posterior probability of the data sample based on the aligned potential feature representation and the data sample, and performing generalized zero sample and generalized small sample fault diagnosis according to the fault category posterior probability. According to the invention, high adaptability to complex industrial scenes is realized, and actual industrial requirements are met.

Description

Technical Field

[0001] This invention belongs to the field of industrial fault diagnosis and relates to an arbitrary sample fault diagnosis method based on hierarchical feature learning. Background Technology

[0002] Industrial fault diagnosis is crucial for ensuring the safe and efficient operation of modern industrial systems. As industrial processes become increasingly complex, accurate fault diagnosis has become an indispensable means of reducing downtime, lowering maintenance costs, and preventing catastrophic failures. Traditional fault diagnosis methods based on machine learning or deep learning are typically considered supervised learning tasks, relying on large amounts of labeled data for training. However, in real-world industrial scenarios, obtaining sufficient fault data for all potential fault types is often costly and difficult, resulting in scarce or even unavailable fault samples. More importantly, traditional methods assume that all fault categories in the test set have already appeared during the training phase (i.e., "visible faults"), making them ineffective at diagnosing new fault categories not present during training (i.e., "unseen faults"). For example, when facing new faults in newly built factories, production lines, or equipment, these methods are severely limited in their practical application value due to the lack of fault data needed to support modeling.

[0003] Zero-shot learning (ZSL), as an advanced learning method, can use semantic attributes to predict the category of unseen samples, providing a new approach to solving the problem of missing data. Although this method has demonstrated excellent performance in machine vision, its application in industrial scenarios still faces many challenges. Furthermore, in real-world industrial scenarios, due to limitations in fault occurrence probability and sampling conditions, the number of samples for different fault categories often varies significantly. Specifically, during the training phase, there may be fault categories with no samples (zero-shot case) and fault categories with only a very small number of samples (small-sample case); while during the testing phase, fault categories may include both visible categories that have appeared in the training set and new unseen categories (generalized zero-shot case). This "arbitrary sample scenario" where small-sample, zero-sample, and generalized zero-sample coexist is more universal and has greater practical significance, bringing new challenges to industrial fault diagnosis.

[0004] For zero-shot fault diagnosis, patent application CN119862498A discloses a zero-shot fault diagnosis method for industrial processes based on similarity consistency relationships. This method constructs knowledge similarity between fault categories, extracts features from visible fault data, quantifies fault feature similarity, and establishes a mapping relationship from knowledge to data, thus achieving the diagnosis of unseen faults. Patent application CN113177584A proposes a composite fault diagnosis method based on zero-shot learning. It constructs semantic vectors of fault categories as auxiliary information and uses a single fault sample to effectively classify composite faults. Patent application CN114383845A discloses a bearing composite fault diagnosis method based on an embedded zero-shot learning model, achieving composite fault diagnosis by constructing an embedded model. However, under the generalized zero-shot setting, these methods often easily misclassify unseen faults as visible faults. For generalized zero-shot fault diagnosis, patent application CN118427672A discloses a generalized zero-shot bearing composite fault diagnosis method based on multi-label learning. By training the model using only single-fault samples, it reduces the dependence on composite fault samples and solves the feature learning problem caused by the scarcity of composite fault samples. Patent application CN113609569A discloses a discriminative generalized zero-shot learning fault diagnosis method, which decomposes the generalized zero-shot diagnosis task into two sub-tasks: supervised learning and zero-shot learning. Existing technologies typically process zero-shot faults and generalized zero-shot faults separately, requiring the establishment of separate diagnostic models. Therefore, constructing a unified diagnostic model that can adaptively handle insufficient or missing samples to achieve accurate fault diagnosis has become a critical problem that urgently needs to be solved.

[0005] In summary, existing technologies have the following main shortcomings in arbitrary sample fault diagnosis tasks: 1. Adapting to the complexity and uncertainty of industrial scenarios: In real industrial environments, fault samples may be completely missing, partially present, or only a small number available, making it difficult for traditional methods to respond flexibly.

[0006] 2. Closed-category assumption: Traditional data-driven fault diagnosis relies on the closed-category assumption, while fault types may dynamically increase in industrial scenarios.

[0007] 3. Ignoring feature hierarchy information: Hierarchical feature extraction can capture multi-scale fault information from raw data—low-level features reflect global anomalies, while high-level features focus on local details. Introducing hierarchical features can significantly improve diagnostic accuracy and robustness, but ignoring feature hierarchy and relying solely on single-layer features can easily lead to misjudgment of complex faults. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide an arbitrary sample industrial fault diagnosis method based on hierarchical feature learning. By constructing a hierarchical feature space with semantic attribute description, and using inter-class-intra-class constraints and cross-correlation matrices to achieve consistent mapping between features and semantics, this invention solves the feature transfer problem from visible fault categories to unseen fault categories. It can achieve unified fault diagnosis under different data conditions such as zero samples and small samples, thereby improving the adaptability and reliability of industrial equipment fault diagnosis.

[0009] To achieve the above objectives, the present invention provides the following technical solution: An arbitrary sample fault diagnosis method based on hierarchical feature learning, comprising: First, a hierarchical autoencoder is constructed, and the latent feature representation of the data sample is extracted step by step from the bottom to the top layer, while the latent attribute representation of the data sample is also extracted. Align the extracted latent feature representations and latent attribute representations; The aligned latent feature representation and data samples are input into the attribute classifier. The attribute classifier predicts the semantic attributes of the data samples and performs zero-sample and small-sample fault diagnosis based on the predicted semantic attributes. Based on the aligned latent feature representation and data samples, the posterior probability of the fault category of the data samples is calculated, and generalized zero-sample and generalized small-sample fault diagnosis is performed according to the posterior probability of the fault category.

[0010] Furthermore, the hierarchical autoencoder includes an encoder part and a decoder part; The encoder part includes multiple feature encoders and one attribute encoder. The multiple feature encoders extract latent feature representations of the input data samples layer by layer, and the attribute encoder extracts latent attribute representations based on the semantic attributes of the input data. The decoder part includes a feature decoder and multiple attribute decoders. The feature decoder reconstructs the input data sample based on the output of the first feature encoder. The multiple attribute decoders correspond one-to-one with each feature encoder and reconstruct the semantic attributes of the data sample based on the output of each feature encoder. At the same time, the feature decoder reconstructs the data sample based on the output of the attribute encoder, and the first attribute decoder reconstructs the semantic attributes of the data sample based on the output of the attribute encoder.

[0011] Furthermore, a dual alignment mechanism combining inter-class and intra-class constraints and cross-correlation matrices is adopted to align latent feature representations and latent attribute representations.

[0012] Specifically, the cross-entropy classification loss function with inter-class and intra-class constraints minimizes the latent representation distance with the same label and maximizes the latent representation distance with different labels.

[0013] In the formula, and They represent the first i Latent feature representation and latent attribute representation of each sample; Represents the L2 norm; p The number of visible fault categories. This represents the number of visible fault samples. For the first k The semantic attributes of visible faults. For attribute encoders; The Barlow Twins method is used to approximate the diagonal dominance matrix of the cross-correlation matrix of latent feature representations and latent attribute representations, thereby aligning the latent feature representations and latent attribute representations. The cross-correlation matrix is ​​calculated using the following formula:

[0014] In the formula, b For batch samples; and They represent the first k The first normalized latent feature representation i The element and the first k The first normalized latent attribute representation j Each element.

[0015] Furthermore, the semantic attributes of the predicted data samples are represented by the attribute classifier as follows:

[0016] In the formula, For data samples, For latent feature characterization, For the first i Individual attribute classifier, For data samples x The corresponding semantic attributes For the first i Each attribute value For the predicted semantic attributes, The number of semantic attributes is given; then the fault category is obtained by the nearest neighbor matching method that minimizes the Euclidean distance.

[0017] In the formula, q Number of unseen fault categories For the semantic attribute of the k-th class where no fault was found, This represents the L2 norm.

[0018] Furthermore, based on the aligned latent feature representation and data samples, the fault category posterior probability of the data samples is calculated. A gating mechanism is adopted to distinguish between visible faults and unseen faults according to the fault category posterior probability. Then, fault category judgment is performed in different label spaces to achieve generalized zero-sample and generalized small-sample fault diagnosis.

[0019] The posterior probability of the fault category for the data sample is calculated using the following formula:

[0020] In the formula, Label the fault category; Visible faults and unseen faults are distinguished based on the maximum posterior probability of the fault category:

[0021] In the formula, For visible faults, No fault was found.

[0022] If no faults are found in the data sample, the fault category label is obtained using the following formula:

[0023] If the data sample represents a visible fault, the fault category label is obtained using the following formula:

[0024] The beneficial effects of this invention are as follows: (1) Applicability to all scenarios: This invention constructs a unified diagnostic framework applicable to any sample scenario (including zero sample, few sample, generalized zero sample and generalized few sample), which breaks through the application limitations of traditional methods in sample-scarce scenarios and significantly improves the applicability to industrial sites.

[0025] (2) Improved diagnostic performance: This invention uses a hierarchical autoencoder to extract fault features at multiple scales, while capturing the global pattern (low-level features) and local detail features (high-level features) of the fault. By fusing hierarchical features, the quality of fault sample representation is improved, thus achieving effective diagnostic performance.

[0026] (3) Cross-modal alignment: This invention achieves accurate mapping between semantic space and feature space by combining inter-class-intra-class constraints and cross-correlation matrix dual alignment mechanism.

[0027] By constructing a hierarchical autoencoder, aligning semantic features, and employing fault diagnosis strategies, this invention achieves strong adaptability to complex industrial scenarios, meets actual industrial needs, and provides a new technical paradigm for fault diagnosis in the field of intelligent manufacturing.

[0028] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a structural block diagram of an arbitrary sample fault diagnosis method based on hierarchical feature learning provided in an embodiment of the present invention. Figure 2 This is a block diagram of a hierarchical autoencoder structure. Figure 3 This is a schematic diagram of the TE process. Detailed Implementation

[0030] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed 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 are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0031] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0032] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0033] This invention provides an arbitrary sample fault diagnosis method based on hierarchical feature learning. It uses semantic attribute description as a foundation and a hierarchical autoencoder to progressively extract hierarchical feature representations of faults from the bottom to the top. Semantic-feature consistency alignment is achieved through inter-class and intra-class constraints and cross-correlation matrices, ensuring knowledge transfer from visible fault categories to unseen fault categories.

[0034] The core idea of ​​Hierarchical Autoencoders (HAEs) is to progressively extract multi-level feature representations of data through a layered approach, achieving a gradual abstraction from low to high levels, with each layer capturing different levels of information in the data. Unlike bidirectional autoencoders, HAEs employ multiple encoder and decoder layers to progressively learn the abstract features of the data, thereby better capturing the hierarchical structural information of the data.

[0035] Furthermore, since the features extracted from the fault samples and the semantic attributes derived from the attribute descriptions belong to two different modalities, the alignment of their latent representations is crucial to improving fault diagnosis performance. Therefore, the semantic-feature alignment module is used to achieve the alignment of semantics and features in the latent space of HAE.

[0036] The fault diagnosis strategy is implemented by using transfer learning to achieve industrial fault diagnosis for zero-shot and few-shot tasks, while introducing an additional sample identification mechanism to distinguish test samples for generalized zero-shot and generalized few-shot tasks, thereby enabling industrial fault diagnosis in different search spaces.

[0037] Based on the above, an embodiment of the present invention provides an arbitrary sample fault diagnosis method as follows: Figure 1 As shown, the details are as follows: Part 1: Construction of a Hierarchical Autoencoder like Figure 2 As shown, the encoder section consists of multiple feature encoders and one attribute encoder. Composition. Each feature encoder The input data is progressively compressed into higher-level latent representations. The decoder section contains a feature decoder. and multiple attribute decoders, each attribute decoder The attributes are gradually reconstructed from the corresponding potential representations.

[0038] During the encoding process, for the input x First, through the first feature encoder Obtain latent feature representation , represented as Subsequently, each subsequent feature encoder It can learn feature representations in a hierarchical manner, gradually capturing higher-level abstract features from initial low-level features. The specific process is as follows:

[0039] in N This represents the number of feature encoders. Similarly, latent attribute representations... Accessible via Attribute Encoder Obtain, that is , For input samples x The corresponding semantic attributes.

[0040] In the decoding phase, the main goal is to generate output data that is highly similar to the input, thereby achieving data reconstruction. (Reconstructed sample) and its attributes This can be represented by the refactoring process, with the specific expressions being as follows: and .

[0041] Furthermore, the model achieves bidirectional reconstruction of cross-modal data, reconstructing attributes from features and features from attributes. Specifically, the cross-modal reconstructed samples... With cross-modal reconstruction properties It can be represented as:

[0042]

[0043] 2: Semantic Feature Alignment Since the features extracted from fault samples and the semantic attributes derived from attribute descriptions belong to two different modalities, aligning their potential representations is crucial for achieving better fault diagnosis.

[0044] The cross-entropy classification loss function with inter-class and intra-class constraints minimizes the latent representation distance with the same label and maximizes the latent representation distance with different labels. Its expression is:

[0045] in, and They represent the first i Latent feature representation and latent attribute representation of each sample. This represents the L2 norm. p The number of visible fault categories. n s This represents the number of visible fault samples. For the first k The semantic attributes of visible faults.

[0046] The Barlow Twins method, a self-supervised learning technique, calculates the covariance matrix of the output vectors of two networks and approximates it to the form of a diagonal dominance matrix. This ensures that learned features are as independent as possible while maintaining similarity; that is, different features are encoded separately, while similar / correlated features receive similar representations. This method not only improves model efficiency but also enables effective pre-training in scenarios with limited data. Therefore, the cross-correlation matrix... C It is used to align latent feature representations and latent attribute representations.

[0047]

[0048] in, b This is a batch sample. and They represent the first k The first normalized latent feature representation i The element and the first k The first normalized latent attribute representation j Each element. , , .

[0049] Three: Fault Diagnosis Implementation In zero- or small-sample diagnostics, based on the law of total probability and the conditional probability theorem, the posterior probability of the fault category of the test sample can be defined as:

[0050] in, For the semantic attribute of the k-th class where no fault was found, q This represents the number of fault categories not found.

[0051] Given attribute description matrix and samples x semantic attributes At that time, fault category label y All the necessary important information has been determined, that is, it meets the requirements. At the same time, due to With a definite and unique value, the posterior probability of the fault category of the test sample can be rewritten as:

[0052] In attribute prediction, attribute classifiers The training phase can learn from all visible fault samples and their corresponding attribute vectors. During the testing phase, the trained attribute classifier is transferred and applied to attribute prediction of the test samples.

[0053] In zero-sample and small-sample fault diagnosis tasks, given unknown samples x Its predicted attribute vector is , m The number of semantic attributes. Fault category labels are obtained using the nearest neighbor matching method that minimizes Euclidean distance. ,Right now

[0054] In generalized zero-sample and generalized small-sample fault diagnosis tasks, test samples contain both visible and unseen fault samples. When prediction results are in the same label space, unseen fault samples are easily misclassified as visible faults, indicating a domain shift problem. Employing a gating mechanism to effectively identify and distinguish test samples can mitigate the inherent domain shift and improve diagnostic accuracy. Specifically, the class identification of test samples is achieved by determining the maximum posterior probability of the sample, as defined below:

[0055] This identification process can be viewed as a binary classification problem, where the visible categories are... Unseen Category It exhibits symmetry. Using 0.5 as the symmetry threshold means that the model maintains the same confidence level in distinguishing between the two classes, ensuring that the decision rule is neither overly biased towards the visible class nor towards the unseen class. Furthermore, when the model has the maximum probability for all visible classes... When the threshold exceeds 0.5, the classifier can confidently classify the sample into a visible category; otherwise, it will determine that the sample does not belong to any visible category.

[0056] Subsequently, predictions are made within different label spaces. If the test sample is diagnosed as an unseen sample ( If the fault category label is not found in the Euclidean distance, then the fault category label can be obtained using the nearest neighbor matching method that minimizes the Euclidean distance. Otherwise, the fault category label can be obtained as follows:

[0057] The evaluation metric is the average accuracy of visible faults. Average accuracy without faults and harmonic mean H :

[0058] like Figure 3The TE process shown serves as a benchmark platform for process monitoring and fault diagnosis. It simulates a complex chemical process based on a real industrial circulating reactor system, including core equipment components such as reactors, condensers, compressors, stripping towers, and vapor-liquid separators. The process generates 52 monitoring variables, including 11 control variables, 19 component measurement variables, and 22 continuous process variables. Table 1 lists 15 typical fault categories, each containing 480 samples.

[0059] Table 1

[0060] According to Table 1, the semantic attribute description of the fault category is defined using 20 specified attributes based on one-hot coding technology.

[0061] Four diagnostic tasks were set up, as shown in Table 2. Each task randomly selected 12 faults as visible faults, and the rest were designated as unseen faults.

[0062] Table 2

[0063] Tables 3-6 present the experimental results for zero-sample fault diagnosis, small-sample fault diagnosis, generalized zero-sample fault diagnosis, and generalized small-sample fault diagnosis, respectively. The experimental results demonstrate that the proposed arbitrary-sample industrial fault diagnosis method based on hierarchical feature learning further enhances fault diagnosis capabilities and is applicable to diagnostic scenarios with different fault sample settings.

[0064] Table 3 method A B C D average DAP 54.2 62.6 40.1 55.5 53.1 IAP 55.5 60.7 45.0 36.3 53.1 SJE 74.6 33.1 34.0 63.9 49.4 ESZSL 57.2 33.1 39.5 39.7 51.4 DEVISE 50.6 74.2 32.9 51.5 42.4 SAE 45.7 74.3 33.5 65.6 52.3 FDAT 80.3 62.6 67.6 72.4 54.8 SCE 89.5 78.1 62.4 76.0 70.7 ZSIDM-OC 86.0 54.2 57.1 65.7 76.5 CPA 78.3 63.6 57.4 71.3 65.8 FGN 85.3 76.7 70.7 95.7 74.0 DSECMR-VAE 93.3 81.1 69.0 76.4 82.1 MetaEvolver 79.9 64.3 93.3 69.4 79.9 FAGAN 84.5 76.9 62.5 74.6 76.7 Free 81.0 75.6 71.5 78.8 74.6 SRWGAN 79.2 77.6 69.4 77.4 75.9 SSB-ZSL-1DCNN 84.3 76.9 62.1 59.7 70.8 AFT 66.8 41.4 40.7 36.8 46.4 GLA-ZSL 97.4 86.2 54.8 77.7 79.0 This invention 93.4 93.2 68.2 81.8 84.2 Table 4 method A B C D average 1-shot 94.4 87.5 70.2 76.9 82.3 3-shot 95.4 96.9 68.9 76.8 84.5 5-shot 95.8 95.5 68.8 77.3 84.3 Table 5 method A B C D average ESZSL 7.8 8.7 10.7 7.8 8.8 FDAT 0.0 0.0 0.7 0.0 0.2 SCE 0.0 0.0 1.6 0.0 0.4 ALE 6.0 7.5 10.6 8.9 8.5 SJE 0.0 7.5 4.6 10.9 5.8 CVAE 34.4 41.9 21.7 40.2 34.6 FAGAN 0.0 0.0 4.1 0.3 1.1 GatingAE 21.3 28.7 2.3 34.0 21.6 MetaEvolver 47.6 50.6 52.7 49.8 50.2 D-GZSL 34.0 47.5 28.1 42.2 38.0 DSECMR-VAE 58.3 42.8 47.2 39.9 42.1 e2eNN 74.7 41.6 45.2 32.9 49.8 This invention 65.3 56.9 48.0 58.9 54.9 Table 6 method A B C D average 1-shot 53.1 53.0 46.5 55.6 52.0 3-shot 56.7 62.3 46.9 55.4 55.3 5-shot 55.4 63.0 51.1 53.5 55.7 Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fault diagnosis method for arbitrary samples based on hierarchical feature learning, characterized in that, First, a hierarchical autoencoder is constructed, and the latent feature representation of the data sample is extracted step by step from the bottom to the top layer, while the latent attribute representation of the data sample is also extracted. Align the extracted latent feature representations and latent attribute representations; The aligned latent feature representations and data samples are input into an attribute classifier. The attribute classifier predicts the semantic attributes of the data samples, and zero-shot and small-shot fault diagnosis is performed based on the predicted semantic attributes. Based on the aligned latent feature representation and data samples, the posterior probability of the fault category of the data samples is calculated, and generalized zero-sample and generalized small-sample fault diagnosis is performed according to the posterior probability of the fault category.

2. The method according to claim 1, characterized in that, A hierarchical autoencoder consists of an encoder part and a decoder part; The encoder part includes multiple feature encoders and one attribute encoder. The multiple feature encoders extract latent feature representations of the input data samples layer by layer, and the attribute encoder extracts latent attribute representations based on the semantic attributes of the input data. The decoder part includes a feature decoder and multiple attribute decoders. The feature decoder reconstructs the input data sample based on the output of the first feature encoder. The multiple attribute decoders correspond one-to-one with each feature encoder and reconstruct the semantic attributes of the data sample based on the output of each feature encoder. At the same time, the feature decoder reconstructs the data sample based on the output of the attribute encoder, and the first attribute decoder reconstructs the semantic attributes of the data sample based on the output of the attribute encoder.

3. The method according to claim 2, characterized in that, A dual alignment mechanism combining inter-class and intra-class constraints and cross-correlation matrices is adopted to align latent feature representations and latent attribute representations.

4. The method according to claim 3, characterized in that, The classification loss function, which employs inter-class and intra-class constraints, minimizes the latent representation distance with the same label and maximizes the latent representation distance with different labels. In the formula, and They represent the first i Latent feature representation and latent attribute representation of each sample; Represents the L2 norm; p The number of visible fault categories. This represents the number of visible fault samples. For the first k The semantic attributes of visible faults. For attribute encoders; The Barlow Twins method is used to approximate the diagonal dominance matrix of the cross-correlation matrix of latent feature representations and latent attribute representations, thereby aligning the latent feature representations and latent attribute representations. The cross-correlation matrix is ​​calculated using the following formula: In the formula, b For batch samples; and They represent the first k The first normalized latent feature representation i The element and the first k The first normalized latent attribute representation j Each element.

5. The method according to claim 1, characterized in that, The semantic attributes of data samples predicted by the attribute classifier are represented as follows: In the formula, For data samples, For latent feature characterization, For the first i Individual attribute classifier, For data samples x The corresponding semantic attributes For the first i Each attribute value For the predicted semantic attributes, The number of semantic attributes is given; then the fault category is obtained by the nearest neighbor matching method that minimizes the Euclidean distance. : In the formula, q Number of unseen fault categories For the semantic attribute of the k-th class where no fault was found, This represents the L2 norm.

6. The method according to claim 1, characterized in that, Based on the aligned latent feature representation and the calculation of the fault category posterior probability of the data sample, a gating mechanism is adopted to distinguish between visible faults and unseen faults according to the fault category posterior probability. Then, the fault category is judged in different label spaces to realize the generalized zero sample and generalized small sample fault diagnosis.

7. The method according to claim 6, characterized in that, The posterior probability of the fault category for the data sample is calculated using the following formula: In the formula, For data samples, For latent feature characterization, For data samples x The corresponding semantic attributes Label the fault category; Visible faults and unseen faults are distinguished based on the maximum posterior probability of the fault category: In the formula, For visible faults, No fault was found; If no faults are found in the data sample, the fault category label is obtained using the following formula: If the data sample represents a visible fault, the fault category label is obtained using the following formula: In the formula, q Number of unseen fault categories For the semantic attribute of the k-th class where no fault was found, Describing the L2 norm, For the semantic attributes of the k-th type of visible fault, Data samples predicted by the attribute classifier x The semantic attributes of.

Images