Generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning

Abstract

The application relates to a generalized zero-shot industrial fault diagnosis method based on feature generation and contrast learning, comprising the following steps: inputting to-be-diagnosed data into a class-aware sparse autoencoder to extract test features; obtaining seen-class fault features; and judging the fault category to which the test features belong through a contrast learning gate model based on the test features and the seen-class fault features. The application extracts discriminative features aligned with fault semantic attributes based on a class-aware sparse autoencoder, and overcomes the semantic feature misalignment problem by using sparse constraints and class-aware supervision; meanwhile, an attribute-guided generative adversarial network is used, and a cyclic ordering loss is introduced to generate high-quality pseudo features based on unseen-class attributes, so that the lack of training samples is made up; finally, through a contrast learning gate mechanism optimized based on a SINCERE loss, accurate identification and shunt diagnosis of seen-class and unseen-class samples are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial process safety monitoring and fault diagnosis technology, and more specifically, to a technology for effectively identifying fault types with completely missing samples in industrial processes. Specifically, this invention mainly relates to a generalized zero-sample industrial fault diagnosis method based on feature generation and contrastive learning. Background Technology

[0002] In modern industrial systems, whether it's large-scale petrochemical processes, precision semiconductor manufacturing, or complex rotating machinery systems, long-term stable operation is a core prerequisite for ensuring production safety, improving production efficiency, and reducing operating costs. However, industrial equipment typically operates under complex and harsh conditions for extended periods, including high temperatures, high pressures, corrosion, and high-speed rotation. Affected by multiple uncertainties such as natural equipment aging, operator errors, and external environmental interference, these systems are highly susceptible to various anomalies or malfunctions. If these malfunctions are not detected, diagnosed, and located in a timely manner, they can lead to anything from declining product quality, equipment damage, and production interruptions to serious environmental pollution and even severe accidents resulting in personal injury or death. Therefore, developing efficient, intelligent, and reliable fault diagnosis technologies is of paramount strategic and economic value for ensuring the safety and stability of industrial processes.

[0003] Traditional fault diagnosis methods are primarily based on mechanistic models or expert systems. Mechanistic models require a deep and precise understanding of the physicochemical properties of industrial processes, but building accurate mathematical models becomes extremely difficult or even infeasible when facing increasingly complex modern industrial systems. Expert systems, on the other hand, rely on manually summarized rule bases, which are costly to maintain and struggle to handle unknown fault modes. With the rapid development of sensor technology, the Internet of Things (IoT), and big data technology, data-driven fault diagnosis methods have made significant progress in recent years. These methods typically rely on deep learning or statistical machine learning techniques to automatically extract and classify features from massive amounts of monitoring data.

[0004] Although industrial systems generate massive amounts of data every second, the vast majority of this data is from normal operation. Many failures are characterized by extremely low frequency of occurrence but extremely severe consequences, making it difficult to obtain sufficient labeled failure samples covering all operating conditions from actual operational history. For reasons of safety and economic efficiency, enterprises are unlikely to deliberately operate expensive critical equipment to a state of failure solely for the purpose of collecting failure samples. This results in the complete absence of certain types of failure samples in the training set. This lack of "availability" of failure data has become a core problem restricting the performance of data-driven fault diagnosis models.

[0005] To overcome the limitations of missing samples, zero-shot learning techniques have emerged. The core idea is to construct a fault attribute matrix using expert knowledge and leverage the knowledge of seen classes acquired during training to identify completely unseen classes. However, in real-world industrial monitoring, test data may come from both seen and unseen classes; this task is called generalized zero-shot learning. In generalized zero-shot learning tasks, existing models often exhibit a strong bias towards seen classes, misclassifying almost all unseen class samples as seen classes, leading to a significant drop in overall diagnostic performance. Furthermore, existing industrial zero-shot fault diagnosis techniques still suffer from severe domain shift and semantic-feature space misalignment, making it difficult to establish stable and robust consistency alignment relationships, thus preventing attribute information from effectively guiding feature learning.

[0006] In summary, existing fault diagnosis technologies still suffer from performance limitations when dealing with complex industrial scenarios where seen and unseen faults coexist. Therefore, how to alleviate the domain offset and semantic feature misalignment problems of fault diagnosis models in generalized zero-sample scenarios, and thus improve the diagnostic accuracy of both seen and unseen faults, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, this invention provides a generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning. It aims to solve the core technical problems in actual industrial production caused by the lack of unseen fault samples, such as poor model generalization ability, domain shift and semantic feature misalignment. It is especially applicable to safety monitoring and fault diagnosis of complex industrial processes such as chemical industry, semiconductor industry and machinery manufacturing.

[0008] To achieve the above objectives, the present invention adopts the following technical solution;

[0009] A generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning includes:

[0010] The data to be diagnosed is input into a category-aware sparse autoencoder to extract test features;

[0011] Obtain the observed fault features, and based on the test features and the observed fault features, determine the fault category to which the test features belong by comparing and learning a gating model;

[0012] The contrastive learning gating model includes a contrastive learning unit and a gating unit. The contrastive learning unit is used to map test features and known fault features to an optimization metric space of intra-class cohesion and inter-class repulsion. The gating unit is used to calculate the maximum cosine similarity between the two and determine the visibility of the sample fault, so as to determine the fault category to which the test feature belongs based on the visibility of the sample fault.

[0013] The contrastive learning unit is trained by minimizing the SINCERE loss function based on the seen class fault features and the unseen class pseudo features. The seen class fault features are obtained by the category-aware sparse autoencoder based on the seen class fault samples and the seen class fault semantic attribute matrix. The unseen class pseudo features are obtained by the generator based on the unseen class fault semantic attribute matrix.

[0014] Preferably, the category-aware sparse autoencoder includes an encoder, a decoder, and a category-aware branch.

[0015] Preferably, the category-aware sparse autoencoder is trained based on known fault samples and the known fault semantic attribute matrix, and jointly performs reconstruction error loss, sparsity constraint loss and category-aware supervision loss.

[0016] Preferably, the reconstruction error loss is the mean square error;

[0017] The sparse constraint loss introduces KL divergence based on the hidden layer of the encoder in the category-aware sparse autoencoder, and is defined as:

[0018]

[0019] In the formula, Represents the sparsity constraint loss. This indicates the number of neurons in the encoder's hidden layer. For the index of neurons, This represents the preset target sparse activation probability. Indicates the first The actual average activation probability of each neuron.

[0020] The category-aware supervised loss is implemented based on the category-aware branch in the category-aware sparse autoencoder.

[0021] Preferably, the gating unit is used to determine the visibility of a sample fault based on the maximum cosine similarity between the two, including:

[0022] If the maximum cosine similarity exceeds a preset threshold, the sample is determined to belong to the "seen class domain"; if it is lower than the preset threshold, the sample is determined to belong to the "unseen class domain".

[0023] Based on the judgment result, the test features are input into the corresponding classifier for fault category identification.

[0024] Preferably, training is performed by minimizing the SINCERE loss function, which is:

[0025]

[0026] In the formula, Indicates the number of samples within a batch. Indicates the sample index. For all samples in the batch A set of positive samples of the same type Indicates the first One sample, Indicates pairs of similar samples The contrast loss between them is defined as:

[0027]

[0028] In the formula, Represents the vector dot product. This is a temperature coefficient used to adjust the smoothness of the similarity distribution. For the set of all negative samples of different classes, This represents the nth sample.

[0029] Preferably, the generator, discriminator, and attribute classifier together form an attribute-guided generative adversarial network, which is trained based on the generator's adversarial loss, attribute classification loss, and / or cyclic ranking loss.

[0030] Preferably, the adversarial loss function of the generator is:

[0031]

[0032] The objective function for optimizing the discriminator is as follows:

[0033]

[0034] In the formula, (D, G) represents the adversarial loss of the generator. (D, G) represents the objective function of the discriminator. This represents the discriminator's score for the input features. This represents the generated features output by the generator. Represents a random noise vector. Represents the standard normal distribution. This represents the semantic attribute matrix of unseen fault types. This represents the distribution of real fault feature data extracted by a category-aware sparse autoencoder. This indicates that the data is sampled from the distribution of the actual fault characteristics. This represents a random interpolation between real fault features and generated features, and the calculation process is as follows: ,in, It is a random value in the interval [0, 1]. This indicates sampling from the interpolation distribution. Represents the gradient penalty coefficient. This indicates that the gradient is calculated over the interpolated features. This represents the expectation of all samples under the distribution.

[0035] Preferably, the attribute classification loss is defined as:

[0036]

[0037] In the formula, This represents the number of attributes in the semantic attribute matrix for unseen fault types, where p represents the attribute index. This represents the p-th attribute vector. Represents the distance metric function. This represents the generator's generation feature for the p-th attribute. , Let C represent a random noise vector, and let C represent the classifier. This represents the classifier parameters.

[0038] Preferably, the introduction of cyclic ranking loss includes inputting the attribute predicted by the attribute classifier into the generator to reconstruct features, and calculating the similarity of the reconstructed features among all categories of generated features;

[0039] The loss expression for the cyclic sorting is:

[0040]

[0041]

[0042] In the formula, This indicates the number of attributes in the semantic attribute matrix of unseen fault types. Indicates the attribute index. Indicates the first A vector of attributes For attribute vectors of other fault categories, The ranking loss term is used to constrain the classifier, making the features... The prediction and the first Attribute vectors The similarity between the generated features and other fault attributes is higher. , Indicates to Features after reconstruction Indicates the generator for the first... The generation features of each attribute, For generating features of other fault attributes, C represents the classifier. Indicates classifier parameters, This is the preset interval parameter.

[0043] This invention provides a generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning, which improves the model's ability to identify seen and unseen fault classes in generalized zero-shot fault diagnosis. Compared with existing technologies, this application not only effectively alleviates the problems of domain offset and semantic feature misalignment in generalized zero-shot fault diagnosis, but also significantly improves the diagnostic accuracy and balance in generalized zero-shot scenarios. Specific beneficial effects include:

[0044] 1. In the feature extraction stage, the designed category-aware sparse autoencoder utilizes sparse constraints and category-aware supervision to promote the consistent mapping between sample features and semantic attributes in the feature space, solve the problem of semantic feature misalignment, and thus extract deep features that are highly discriminative and aligned with semantic attributes.

[0045] 2. To alleviate the domain shift problem caused by the lack of unseen class samples, attribute-guided generative adversarial networks are used and cyclic ranking loss is introduced. High-quality features are directly generated in the feature space with the help of the semantic information of unseen classes to support subsequent diagnosis.

[0046] 3. A contrastive learning gating mechanism based on SINCERE loss optimization was constructed. Furthermore, by calculating the maximum similarity between the test sample and the prototype of the seen class, the accurate identification and triage diagnosis of the seen and unseen class samples can be achieved. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0048] Figure 1 This is a flowchart of the entire fault diagnosis process provided in the embodiments of the present invention.

[0049] Figure 2 This is a detailed network structure diagram of the category-aware sparse autoencoder in an embodiment of the present invention.

[0050] Figure 3 This is a detailed network structure diagram of the attribute-guided generative adversarial network in an embodiment of the present invention.

[0051] Figure 4 This is a process flow diagram of the semiconductor etching process used for verification in an embodiment of the present invention.

[0052] Figure 5This is the semantic attribute description matrix of the semiconductor etching process in this embodiment of the invention.

[0053] Figure 6 This is the maximum similarity probability density distribution diagram of the fault experiments in Group A of this invention.

[0054] Figure 7 This is a detailed confusion matrix diagram of the fault experiment in Group C of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] This invention discloses a generalized zero-shot fault diagnosis scheme that can effectively mitigate domain offset, achieve deep semantic alignment, and generate high-quality, diverse features. Specifically, it is a generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning.

[0057] In this embodiment, the specific diagnostic process includes:

[0058] The data to be diagnosed is input into a category-aware sparse autoencoder to extract test features;

[0059] Obtain the observed fault features, and based on the test features and the observed fault features, determine the fault category to which the test features belong by comparing and learning a gating model;

[0060] The contrastive learning gating model includes a contrastive learning unit and a gating unit. The contrastive learning unit is used to map test features and known class fault features to an optimization metric space of intra-class cohesion and inter-class repulsion. The gating unit is used to determine the visibility of sample faults by the maximum cosine similarity between the two.

[0061] In this application, the contrastive learning unit is trained by minimizing the SINCERE loss function based on the seen class fault features and the unseen class pseudo features; the seen class fault features are obtained by the category-aware sparse autoencoder based on the seen class fault samples and the seen class fault semantic attribute matrix; and the unseen class pseudo features are obtained by the generator based on the unseen class fault semantic attribute matrix.

[0062] Its overall process is as follows Figure 1 The execution process is illustrated below through specific examples.

[0063] Example 1:

[0064] In this embodiment, the category-aware sparse autoencoder, abbreviated as CSAE, includes an encoder, a decoder, and a category-aware branch; its network is as follows: Figure 2 As shown, the encoder maps the input samples to a low-dimensional latent feature space, and the decoder reconstructs the latent features back to the original data space.

[0065] In one implementation, a training set containing known fault samples is first constructed based on sensor data acquired from industrial processes; and semantic attributes of all possible fault categories are defined based on industrial mechanisms and expert experience, respectively constructing semantic attribute matrices for known faults and semantic attribute matrices for unseen faults. The attribute matrix is ​​a binary matrix, where columns represent fault categories and rows represent semantic attributes (such as "excessive pressure", "valve jamming", etc.).

[0066] Then, using the observed fault samples and the observed fault semantic attribute matrix, the reconstruction error loss, sparse constraint loss, and category-aware supervision loss are jointly used to train it:

[0067] In some implementations, by minimizing the reconstruction error and preserving the essential information of the data, the reconstruction error function between the CSAE input and output uses the mean squared error (MSE), which is defined as follows:

[0068]

[0069] In the formula, This represents the reconstruction error loss, where E represents the CSAE encoder and O represents the CSAE decoder. Indicates to The reconstructed output sample This represents the i-th known fault sample. This represents the number of training samples, and i represents the index of the training sample. This represents the i-th seen class attribute.

[0070] In some implementations, KL divergence is introduced as a sparsity constraint in the hidden layer of the encoder to suppress the activation of non-critical neurons and extract robust sparse features. Specifically, the sparsity constraint loss is defined as:

[0071]

[0072] In the formula, Represents the sparsity constraint loss. This indicates the number of neurons in the encoder's hidden layer. For the index of neurons, This represents the preset target sparse activation probability. Indicates the first The actual average activation probability of each neuron.

[0073] In some implementations, an attribute prediction branch (classifier) ​​is introduced at the encoder output to utilize the real fault attribute vector as a supervision signal. This forces the features extracted by the encoder to include semantic information that can distinguish fault categories, thereby achieving forced alignment between the feature space and the semantic space. Adding category awareness as a supervision constraint during training aims to measure the accuracy of the encoder's extracted features in predicting fault attributes. The loss is defined as follows:

[0074]

[0075] In the formula, This represents the category-aware supervision loss. This represents the true attribute of the j-th attribute of the i-th sample. This represents the predicted attribute of the j-th attribute of the i-th sample. The number of training samples is represented by , M represents the number of attributes contained in each fault sample, i is the training sample index, and j is the attribute index.

[0076] Preferably, this embodiment constructs the following overall loss function for CSAE based on the above-mentioned multi-constraint objectives:

[0077]

[0078] In the formula, and represents the weight coefficients of the sparse constraint loss and the category perception loss, respectively.

[0079] Example 2:

[0080] The contrastive learning unit is trained by minimizing the SINCERE loss function based on the features of the seen fault class and the pseudo features of the unseen fault class.

[0081] Among them, the features of the seen fault class are obtained by CSAE based on the seen fault class samples and the seen fault class semantic attribute matrix, and the pseudo features of the unseen fault class are obtained by generator based on the unseen fault class semantic attribute matrix.

[0082] In one implementation, the generator is used to output unseen class pseudo-features in the feature space based on the input random noise vector and the unseen class fault semantic attribute vector. In this application, it, together with the discriminator and the attribute classifier, forms an attribute-guided generative adversarial network, with the structure as follows: Figure 3 As shown.

[0083] The discriminator uses Wasserstein distance as a metric and introduces a gradient penalty mechanism to distinguish whether the input features are real, known class features extracted by CSAE or pseudo-features generated by the generator. Its optimization objective is:

[0084]

[0085] Furthermore, the adversarial loss function of the generator is obtained as follows:

[0086]

[0087] In the formula, (D, G) represents the objective function for optimizing the discriminator. The discriminator is trained by minimizing this loss. AGGAN stands for Attribute-Guided Generative Adversarial Network. (D, G) represents the adversarial loss of the generator, which is used to train the generator. adv represents the adversarial loss. This represents the discriminator's score for the input features. This represents the generated features output by the generator. Represents a random noise vector. Represents the standard normal distribution. This represents the semantic attribute matrix of unseen fault types. This represents the distribution of real fault feature data extracted by a category-aware sparse autoencoder. This indicates that sampling is performed from the distribution of the actual fault characteristic data. This represents a random interpolation between real fault features and generated features, and the calculation process is as follows: ,in, It is a random value in the interval [0, 1]. This indicates sampling from the interpolation distribution. Represents the gradient penalty coefficient. This indicates that the gradient is calculated over the interpolated features. This represents the expectation of all samples under the distribution.

[0088] In some preferred embodiments, an attribute classifier is introduced to predict attributes of the generated features, and an attribute classification loss is calculated to ensure that the generated features conform to a given semantic description. Specifically, the classifier first predicts the attributes, and then the loss between the predicted attributes and the true attributes is calculated, which is defined as:

[0089]

[0090] In the formula, This indicates the number of attributes in the semantic attribute matrix of unseen fault types. Indicates the attribute index. Indicates the first A vector of attributes Represents the distance metric function. Indicates the generator for the first... The generation features of each attribute, and C represents the classifier. Indicates classifier parameters, This represents a random noise vector.

[0091] Furthermore, to address the issue of insufficient diversity in generated features, a cyclic ranking loss is introduced into the attribute classifier. The generated features are fed back into the classifier to obtain predicted attributes, and then the predicted attributes are fed back into the generator to reconstruct features. The similarity ranking of the reconstructed features among all categories of generated features is calculated, forcing the model to generate highly distinctive and diverse features.

[0092] In one embodiment, for each generated feature, the classifier predicts its corresponding attribute:

[0093]

[0094] Indicates the generator for the first... The generation features of each attribute, Indicates the first p There are 3 predicted attribute vectors, where C represents the classifier. This represents the classifier parameters.

[0095] Then, the first p Predicted attribute vectors Input to the generator and obtain reconstructed features. , Indicates to Features after reconstruction This represents a random noise vector.

[0096] In addition to performing the classification task, the classifier is also used to calculate the similarity between the reconstructed features and the original generated features. The calculation process is as follows:

[0097]

[0098] Therefore, the definition of the loss for cyclic sorting is as follows:

[0099]

[0100] In the formula, This indicates the number of attributes in the semantic attribute matrix of unseen fault types. Indicates the attribute index. Indicates the first A vector of attributes For attribute vectors of other fault categories, The ranking loss term is used to constrain the classifier, making the features... The prediction and the first Attribute vectors The similarity between the generated features and other fault attributes is higher. , Indicates to Features after reconstruction Indicates the generator for the first... The generation features of each attribute, For generating features of other fault attributes, C represents the classifier. Indicates classifier parameters, This is the preset interval parameter.

[0101] As a preferred approach, the overall loss function of the attribute-guided generative adversarial network in this application is expressed as follows:

[0102]

[0103] in, and These are the weighting coefficients for balancing classification loss and cyclic ranking loss, respectively.

[0104] After training, the semantic attribute vectors of unseen classes are input into the generator to generate pseudo features of unseen class faults in batches, thus transforming the zero-shot problem into a supervised learning problem.

[0105] In one implementation, the contrastive learning unit is trained using a hybrid feature set consisting of known class fault features (from CSAE) and unseen class pseudo-features (from an attribute-guided generative adversarial network generator). Unlike traditional contrastive learning, this invention uses the SINCERE loss function for training. The SINCERE loss, through theoretical correction, eliminates the unnecessary repulsive force between similar samples in the feature space, making similar samples more compact and the boundaries of dissimilar samples clearer.

[0106] Specifically, the loss is calculated in batches for the mixed feature set, assuming that a batch contains There are 10 samples, and they are denoted as 10 samples. ,in For fault feature samples, This is the fault attribute vector. For any sample in the batch... ,definition For all items in the batch A set of positive samples of the same type Let be the set of all negative samples from different classes. The batch-based SINCERE loss function is defined as follows:

[0107]

[0108] In the formula, Indicates the number of samples within a batch. Indicates the sample index. For all samples in the batch A set of positive samples of the same type Indicates the first One sample, Indicates pairs of similar samples The contrast loss between them is defined as:

[0109]

[0110] In the formula, Represents the vector dot product. This is a temperature coefficient used to adjust the smoothness of the similarity distribution. The set of negative samples of all distinct classes.

[0111] By minimizing The model will bring similar fault features closer together and push different fault features further apart, thereby forming a clearly defined distribution of seen and unseen classes in the feature space, enhancing the effectiveness of gating discrimination.

[0112] Example 3:

[0113] The fault category to which the test feature belongs is determined by a comparative learning gating model; including:

[0114] The test features and observed fault features are mapped to the optimization metric space of intra-class cohesion and inter-class repulsion through contrastive learning units, resulting in test feature embedding vectors. Embedded vectors of known fault features ;

[0115] After normalizing all embedding vectors, the calculation is performed using a gating unit. With all The maximum cosine similarity between them is calculated as follows:

[0116]

[0117] In the formula, This represents the maximum cosine similarity between two vectors, where `max` indicates taking the maximum value. This represents the category index of a known type of fault, with a value range of [value range missing]. ∈{1,2,…,s}, where s represents the number of known fault categories. Represents the test feature embedding vector. This represents the embedding vector of known fault features. Indicates the first Embedded vectors of known fault features.

[0118] If the similarity of the test samples exceeds a preset threshold, the samples are determined to belong to the "seen category"; if it is below the threshold, the samples are determined to belong to the "unseen category".

[0119] Finally, based on the judgment result, the test features are input into the corresponding classifier (either the seen class classifier or the unseen class classifier) ​​for fault category identification. It should be noted that the seen class classifier is trained based on the seen class fault features output by the category-aware sparse autoencoder, while the unseen class classifier is trained based on the unseen class pseudo-features generated by the generator.

[0120] In this application, the test samples for the generalized zero-shot diagnostic task include both seen and unseen faults, thus requiring a comprehensive evaluation of the diagnostic results. The harmonic mean H can simultaneously measure the diagnostic accuracy for both seen and unseen faults, providing a comprehensive assessment of model performance and making it suitable for the generalized zero-shot diagnostic task. Therefore, H is chosen as one of the evaluation metrics. The definition of H is as follows:

[0121]

[0122] in, This indicates the diagnostic accuracy of known fault types. This represents the diagnostic accuracy for unseen fault types. The accuracy is defined as follows:

[0123]

[0124] in, This indicates the number of correctly diagnosed fault samples. This represents the total number of test failure samples.

[0125] Example 4:

[0126] The semiconductor etching process is used as an example for verification. Semiconductor etching is a typical nonlinear intermittent process and plays a crucial role in semiconductor manufacturing. Its process flow is as follows: Figure 4 The etching process utilizes BCl3 / Cl2 inductively coupled plasma (ICP) for etching. The SEP dataset was obtained using a Lam 9600 plasma etching tool. Process variables were sampled at 1-second intervals using a machine status sensor, ultimately selecting 19 key variables as monitoring signals. Eight different types of faults were used in this experiment, and their specific fault descriptions and semantic attributes are shown in Tables 1 and 2, respectively.

[0127] Table 1. Fault information in the semiconductor etching process

[0128]

[0129] Table 2. Description of fault semantic attributes in semiconductor etching process

[0130]

[0131] Faults are encoded based on their semantic attribute information, with 1 indicating that the fault possesses the attribute and 0 indicating that the fault does not possess the attribute. The final semantic attribute description matrix is ​​as follows: Figure 5 As shown.

[0132] To verify the performance of the generalized zero-shot fault diagnosis, six fault types were designated as seen faults, and the remaining two as unseen faults. Three experimental scenarios were set up with different combinations of these classifications, as shown in Table 3. For sample allocation, each fault type included 60 training samples and 30 test samples. In the SEP generalized zero-shot fault diagnosis experiment, the training set contained only seen fault data, while the test set contained data from both seen and unseen faults.

[0133] Table 3 Fault Group Classification in Semiconductor Etching Process

[0134]

[0135] During the model training phase, the batch size was set to 60, and 1500 training rounds were performed, all using the Adam optimizer with an initial learning rate of 0.001. Additionally, a grid search method was used to determine the values ​​of the six weight parameters in the proposed model; the search range and results are shown in Table 4.

[0136] Table 4. Grid Search and Results for Weight Parameters

[0137]

[0138] To visualize the effectiveness of the model, see Figure 6 The maximum similarity probability density distribution is shown for group A faults. Seen samples exhibit higher maximum similarity, primarily concentrated in the [0.6, 1.0] interval, while unseen samples are mainly distributed in the lower [0.4, 0.65] interval. Furthermore, the overlap between the two distribution curves is small, and the peaks are clearly separated, fully demonstrating the model's ability to distinguish between seen and unknown fault classes.

[0139] The intersection of the distribution curves of the seen and unseen classes was selected as the gate threshold. The specific gate thresholds and gate accuracy results for different groups are shown in Table 5. Among them, G... S G represents the accuracy of the contrastive learning gating mechanism for known fault types. U G represents the accuracy for unseen faults. H This represents the harmonic mean of the two values. The model demonstrated good gating accuracy across all three fault groups, further validating the effectiveness of contrastive learning gating. In fault group C, the PNB classifier achieved 100% accuracy for unseen fault classes. See also Figure 7The visualization results of the confusion matrix for Group C faults are shown, with the correct classification of unseen fault samples marked in the boxes.

[0140] Table 5. Accuracy (%) of contrastive learning gating on the SEP dataset

[0141]

[0142] The quality prediction model of this invention was compared with FDAT, ACGAN-FG, CVAE, and DSECMR-VAE. FDAT and ACGAN-FG were used for zero-shot fault diagnosis, while CVAE and DSECMR-VAE were applied to generalized zero-shot fault diagnosis. All models employed their proposed or experimentally validated optimal classifiers to obtain the best diagnostic performance. FDAT, DSECMR-VAE, and the proposed model used Probabilistic Naive Bayes (PNB), ACGAN-FG used a Multilayer Perceptron (MLP), and CVAE used a Linear Support Vector Machine (LSVM).

[0143] To verify the effectiveness of the model, a generalized zero-shot fault diagnosis experiment was conducted using the proposed model and four advanced models on different fault groups in the SEP dataset. The results are shown in Table 6, where S and U represent the diagnostic accuracy of seen and unseen faults, respectively, and H represents the harmonic mean of the two.

[0144] Table 6. Generalized zero-shot fault diagnosis accuracy of different models on the SEP dataset

[0145]

[0146] The experimental results show that the proposed model achieved the best harmonic mean (H) in all groups, improving accuracy by 11.55%, 11.86%, and 13.87% in fault groups A, B, and C, respectively, compared to the second-best performing DSECMR-VAE model. Furthermore, the proposed model improved both the accuracy (U) and accuracy (S) of diagnosed unseen faults. While the ACGAN-FG model achieved a high S value in group B, its U value was significantly lower, reflecting a severe overfitting problem for seen faults. In contrast, CSAEFG-CLG not only significantly outperformed the model in U value but also demonstrated a good level in S value. In unseen fault diagnosis, FDAT and ACGAN-FG were almost unable to effectively identify unseen faults and exhibited severe domain shift issues. The unseen fault features generated by CSAEFG-CLG possessed high quality and separability, making them effective for training the diagnostic model. Therefore, CSAEFG-CLG is effective and superior in industrial generalized zero-sample fault diagnosis scenarios.

[0147] This invention addresses the generalized zero-shot problem in industrial fault diagnosis, enabling accurate diagnosis of both seen and unseen faults simultaneously, even with a training set containing only seen faults. In the feature extraction stage, a designed category-aware sparse autoencoder promotes consistent mapping between sample features and semantic attributes in the feature space, thereby extracting highly discriminative deep features aligned with semantic attributes. To mitigate the domain shift problem caused by missing unseen samples, an attribute-guided generative adversarial network is introduced. This network directly generates high-quality features in the feature space using the semantic information of unseen classes to support subsequent diagnosis. Furthermore, a contrastive learning-based gating mechanism is constructed, calculating the maximum similarity between test samples and seen class prototypes to effectively distinguish and triage fault categories. Compared to existing state-of-the-art methods, this model significantly improves diagnostic accuracy while effectively mitigating domain shift and semantic feature misalignment issues.

[0148] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0149] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning, characterized in that, include: The data to be diagnosed is input into a category-aware sparse autoencoder to extract test features; Obtain observed fault characteristics, and based on the test characteristics and observed fault characteristics, determine the fault category to which the test characteristics belong by comparing and learning a gating model; The contrastive learning gating model includes a contrastive learning unit and a gating unit. The contrastive learning unit is used to map test features and known fault features to an optimization metric space of intra-class cohesion and inter-class repulsion. The gating unit is used to calculate the maximum cosine similarity between the two to determine the visibility of the sample fault. The contrastive learning unit is trained by minimizing the SINCERE loss function based on the seen fault features and the unseen pseudo features. The observed fault features are obtained by the category-aware sparse autoencoder based on observed fault samples and observed fault semantic attribute matrix; the unseen fault pseudo features are obtained by the generator based on unseen fault semantic attribute matrix. In training, the SINCERE loss function is minimized. The SINCERE loss function is: In the formula, Indicates the number of samples within a batch. Indicates the sample index. Let t represent the t-th sample. For all samples in the batch A set of positive samples of the same type Indicates the first One sample, Indicates pairs of similar samples The contrast loss between them is defined as: In the formula, Represents the vector dot product. This is a temperature coefficient used to adjust the smoothness of the similarity distribution. For the set of all negative samples of different classes, This represents the nth sample.

2. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 1, characterized in that, The category-aware sparse autoencoder includes an encoder, a decoder, and a category-aware branch.

3. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 1, characterized in that, The category-aware sparse autoencoder is trained based on known fault samples and the known fault semantic attribute matrix, and jointly employs reconstruction error loss, sparsity constraint loss, and category-aware supervision loss.

4. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 3, characterized in that, The sparse constraint loss introduces KL divergence based on the hidden layer of the encoder in the category-aware sparse autoencoder, and is defined as: In the formula, Represents the sparsity constraint loss. This indicates the number of neurons in the encoder's hidden layer. For the index of neurons, This represents the preset target sparse activation probability. Indicates the first The actual average activation probability of each neuron.

5. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 1, characterized in that, The gating unit is used to determine the visibility of a sample fault based on the maximum cosine similarity between the two, including: If the maximum cosine similarity exceeds a preset threshold, the sample is determined to belong to a seen class domain; if the maximum cosine similarity is lower than the preset threshold, the sample is determined to belong to an unseen class domain. Based on the judgment result, the test features are input into the corresponding classifier for fault category identification.

6. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 1, characterized in that, The generator, discriminator, and attribute classifier together form an attribute-guided generative adversarial network, which is trained based on the generator's adversarial loss, attribute classification loss, and / or cyclic ranking loss.

7. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 6, characterized in that, The adversarial loss of the generator is: The objective function for optimizing the discriminator is as follows: In the formula, (D, G) represents the adversarial loss of the generator. (D, G) represents the objective function of the discriminator. This represents the discriminator's score for the input features. This represents the generated features output by the generator. Represents a random noise vector. Represents the standard normal distribution. This represents the semantic attribute matrix of unseen fault types. This represents the distribution of real fault feature data extracted by a category-aware sparse autoencoder. This indicates that sampling is performed from the distribution of real fault characteristic data. This represents a random interpolation between real fault features and generated features, and the calculation process is as follows: ,in, For a random value in the interval [0, 1], This indicates sampling from the interpolation distribution. Represents the gradient penalty coefficient. This indicates that the gradient is calculated over the interpolated features. This represents the expectation of all samples under the distribution.

8. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 6, characterized in that, The attribute classification loss is defined as: In the formula, This indicates the number of attributes in the semantic attribute matrix of unseen fault types. Indicates the attribute index. Indicates the first A vector of attributes Represents the distance metric function. Indicates the generator for the first... The generation features of each attribute, and , Let C represent a random noise vector, and let C represent the classifier. This represents the classifier parameters.

9. The generalized zero-shot industrial fault diagnosis method based on feature generation and contrastive learning as described in claim 6, characterized in that, The cyclic ranking loss includes inputting the attribute predicted by the attribute classifier into the generator to reconstruct features, and calculating the similarity of the reconstructed features among all categories of generated features; The loss expression for the cyclic sorting is: In the formula, This indicates the number of attributes in the semantic attribute matrix of unseen fault types. Indicates the attribute index. Indicates the first A vector of attributes For attribute vectors of other fault categories, The ranking loss term is used to constrain the classifier, making the features... With the Attribute vectors The similarity is higher than that of other fault attributes in generating features. , Indicates to Features after reconstruction Indicates the generator for the first... The generation features of each attribute, For generating features of other fault attributes, C represents the classifier. Indicates classifier parameters, This is the preset interval parameter.

Images