Open set fault diagnosis method combining generative data enhancement and uncertainty measurement

Abstract

The invention discloses an open set fault diagnosis method combining generative data enhancement and uncertainty measurement, and aims to solve the problems that a traditional closed set fault diagnosis model is high in misjudgment rate and insufficient in generalization when facing unknown fault types. Existing methods generally depend on sufficient samples of known fault categories, and unknown fault modes are difficult to recognize. Therefore, a VAE-GMM generation model based on Gaussian mixture distribution is constructed, an auxiliary negative sample separated from a known category is generated by adopting a low probability density sampling strategy, and the sensitivity of the model to an abnormal sample is optimized in combination with comparative learning. Further, uncertainty quantization is carried out on the prediction result through an EDL classifier, and a threshold judgment criterion is established to distinguish known / unknown fault types. Compared with the prior art, the method has the advantages that through dual-module cooperation of generative data enhancement and uncertainty measurement, the diagnosis robustness in an open set scene is remarkably improved, and meanwhile, the dependence on unknown fault prior data is reduced.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis of key components of mechanical equipment, and in particular to an open set fault diagnosis method combining generative data enhancement and uncertainty measurement. Background Art

[0002] Reliable fault diagnosis is crucial to ensuring the long-term stable operation of complex mechanical systems. In recent years, intelligent fault diagnosis methods have been widely used to identify and locate known fault modes. Due to the superior nonlinear mapping capabilities of deep learning technology, reliance on expert knowledge has been significantly reduced, enabling end-to-end diagnosis of equipment faults. However, most existing methods are based on the "closed set assumption," assuming that data in the source and target domains have the same fault categories. This ignores the existence of potential unknown fault modes in actual industrial applications. In addition, due to safety restrictions or high reproduction costs, it is difficult to obtain sufficient samples in experiments for certain known fault modes (such as elevator overspeed and traction rope breakage). This can lead to false positives or false negatives for unknown faults, thereby increasing the risk of equipment operation. Therefore, it is crucial to introduce open set fault diagnosis methods that can effectively handle unknown fault modes. Summary of the Invention

[0003] This invention aims to address the limitations of existing intelligent fault diagnosis methods when faced with unknown fault modes, by providing a novel approach that enables reliable fault diagnosis in an open-set environment. This approach aims to improve the model's accuracy for identifying known fault types while simultaneously enabling the model to effectively identify and reject unknown fault modes, thereby reducing false positives and missed negatives and ensuring the safe and stable operation of complex mechanical systems.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] An open set fault diagnosis method combining generative data augmentation and uncertainty measurement includes the following steps:

[0006] S1: Collect multi-operational data of mechanical equipment and process vibration signals to construct a training sample set;

[0007] S2: Construct a VAE-GMM generative model based on mixed Gaussian priors;

[0008] S3: Based on the low probability density sampling strategy, the above model is used to generate out-of-distribution negative samples that are separated from the known categories;

[0009] S4: Use generated negative samples to perform comparative learning with known category samples to optimize the model's ability to distinguish abnormal samples;

[0010] S5: Design an EDL classifier to predict the failure mode of input samples and quantify the uncertainty of the prediction results;

[0011] S6: Establish a discrimination threshold criterion based on uncertainty, judge high uncertainty samples as unknown fault types, and output accurate classification results of known faults.

[0012] In step S2, the VAE-GMM generates a model, in which the encoder maps the input data to the latent space, and the decoder reconstructs the latent space features into data. When constructing the latent coding of the latent space, the single Gaussian prior of the traditional VAE model is replaced by a mixed Gaussian prior through iterative training of the fused GMM model. In addition, when training the mixed Gaussian variational encoder network, a distance correlation constraint is imposed to ensure that the separation degree of the samples in the latent space and the sample space is correlated.

[0013] In step S2, the low probability density sampling strategy performs inverse sampling based on the probability density function of the Gaussian mixture model. By uniformly sampling in the latent space and rejecting points in high-density areas, the sampling of points separated from known class samples in the latent space is achieved, and then auxiliary negative sample data separated from the known class is generated through the decoder.

[0014] In step S3, the low probability density sampling strategy performs inverse sampling based on the probability density function of the Gaussian mixture model. By uniformly sampling in the latent space and rejecting points in high-density areas, the sampling of points separated from known class samples in the latent space is achieved, and then auxiliary negative sample data separated from the known class is generated through the decoder.

[0015] In step S4, the contrastive learning strategy adopts InfoNCE loss to optimize the network parameters by minimizing the distance between known class samples and their corresponding positive samples (including samples of the same category and enhanced samples) and maximizing the distance between known class samples and the auxiliary negative samples generated by S2.

[0016] In step S5, the EDL classifier is constructed based on the Dirichlet distribution and DST evidence theory. The non-negative output is calculated through the ReLU activation function, which is regarded as the evidence vector of each fault category and the Dirichlet distribution parameters and EDL loss function are calculated to perform fault mode prediction and uncertainty quantification.

[0017] The uncertainty quantification discrimination threshold described in step S6 is based on the distribution characteristics of the uncertainty measure of the training set prediction, and its 95% quantile is used as the judgment threshold of the unknown category to realize the category decision boundary division in the open set recognition task; when the uncertainty measure value of the sample to be tested calculated by S5 is greater than the discrimination threshold, it is judged as an unknown fault type; if the uncertainty measure is lower than the discrimination threshold, the fault type is judged according to the model output category prediction probability.

[0018] The beneficial effects of the present invention are:

[0019] This paper proposes an open-set fault diagnosis method that combines generative data augmentation with uncertainty measurement. By deeply integrating generative data augmentation with uncertainty quantification, it effectively overcomes the limitations of traditional closed-set diagnosis methods. Specifically:

[0020] 1. Enhanced recognition of unknown faults: By building a VAE-GMM generative model and adopting a low-probability density sampling strategy, we can proactively generate high-quality, out-of-distribution auxiliary negative samples that are separate from the distribution of known fault patterns. Combined with contrastive learning, this significantly improves the model's ability to distinguish between known faults and unknown abnormal patterns, effectively reducing the risk of misclassifying unknown faults as known ones.

[0021] 2. Improved the reliability and credibility of diagnostic results: The introduction of an EDL-based uncertainty quantification mechanism enables the model to not only predict the fault category but also provide a credibility measure for the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of model training and prediction according to an embodiment of the present invention;

[0024] Figure 3 Schematic diagram of the negative sample generation process according to an embodiment of the present invention; DETAILED DESCRIPTION

[0025] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0026] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, the main steps of the method provided by the present invention are as follows:

[0027] S1: Collect multi-operational data of mechanical equipment and process vibration signals to construct a training sample set;

[0028] S2: Construct a VAE-GMM generative model based on mixed Gaussian priors;

[0029] As a preferred embodiment of the present invention, refer to Figure 2, the VAE-GMM generation model structure includes:

[0030] Feature encoder f φ : The vibration signal is mapped into the latent space through a multi-layer convolutional network to extract the low-dimensional latent code z;

[0031] Reconstruction decoder g θ : Use deconvolution and upsampling operations to remap the hidden code z in the latent space to the sample space to generate a reconstructed signal with the same dimension as the original signal;

[0032] GMM Fitting Module: This module fits the Gaussian mixture distribution of the latent code z. It sets the parameters of the Gaussian mixture distribution of known class samples in the latent space and incorporates them into the model loss function. Through model training, the Gaussian mixture distribution fitting model of known class samples in the latent space is iteratively updated to enable sampling of known class samples in the latent space. Then, by sampling regions with low probability density, auxiliary negative sample latent codes with a high degree of separation from known class samples are obtained.

[0033] Loss calculation module: The overall loss term of the model includes the separation correlation loss L corr , signal reconstruction loss L recon and distribution difference loss L dist .

[0034] The separation correlation loss L corr The calculation method is: the Euclidean distance d(z1,z2) of the sample in the latent space and the Euclidean distance d(g θ (z1),g θ (z2)) calculates the correlation loss and constrains the correlation of the distance between the training samples in the latent space and the sample space, so that the separation of the latent code in the latent space can reflect its separation in the sample space, thereby ensuring that the latent code z that is separated from the known class in the subsequent latent space sampling can still maintain a high degree of separation from the known class samples in the sample space after being reconstructed by the decoder. Specifically expressed as:

[0035]

[0036] Signal reconstruction loss L recon , distribution difference loss L dist Through the derivation of the Evidence Lower Bound (ELBO) of VAE and the introduction of Gaussian Mixture Model (GMM) prior, ELBO is expressed as:

[0037]

[0038] The first term of ELBO represents the signal reconstruction loss L recon ,use Represents sample x iThe corresponding reconstructed signal generated by this model is N, which represents the total number of training samples; the signal reconstruction loss L recon Expand to get:

[0039]

[0040] The second term of ELBO represents the distribution difference loss L dist , by approximating the variational posterior q φ (c, z|x) and the KL divergence calculation of the Gaussian mixture model prior p(c, z); According to the definition of KL divergence and Bayesian theorem, the distribution difference loss L dist Expanding and simplifying gives:

[0041]

[0042] Among them, the former realizes the alignment of the latent code with each cluster center, and for each cluster c, the variational posterior q is calculated φ (z|x,c) and the KL divergence of the corresponding Gaussian prior, and the clustering probability q output by the encoder φ (c|x) weighted average; the latter term measures the consistency of the clustering probability with the clustering mixing coefficient prior.

[0043] The final overall loss is the sum of all loss items, which is expressed as follows:

[0044] L=L recon +L dist +L corr (5)

[0045] S3: Based on the low probability density sampling strategy, the above model is used to generate out-of-distribution negative samples that are separated from the known categories;

[0046] As a preferred embodiment of the present invention, in step S3, the low probability density sampling process specifically includes the following steps:

[0047] S31: Pre-training the generative model of known class samples. Use the generative model built in S2 to pre-train the known class sample data. The model uses Gaussian mixture distribution as the prior distribution of the latent space, obtains the training loss based on the loss calculation module, and updates the encoder and decoder parameters (φ, θ) and GMM parameters simultaneously through gradient descent.

[0048] S32: Determine the distribution of known class samples in the latent space. After the training of the mixed Gaussian autoencoder VAE-GMM generation model is completed, the distribution of known class data in the latent space can be obtained by using the learned mixed Gaussian distribution. express:

[0049]

[0050] Where N() represents the Gaussian distribution of a given class c

[0051] Set the specified threshold τ, the probability density of a given sampling point z When the threshold is exceeded, it is considered as the hidden code Z of the known class sample in the hidden space ID , auxiliary positive sample signals can be generated through the decoder, and the specific sampling process is expressed as:

[0052]

[0053] S33: Generate hidden code of auxiliary negative samples. For the generation of auxiliary negative samples, we first use a low probability density sampling strategy to obtain the hidden code from the uniform distribution U in the latent space. z Sampling multiple points in the , and then based on the learned mixed Gaussian distribution Filter out points with high probability density and select points with probability density lower than the specified threshold τ as the hidden code for auxiliary negative samples. The specific sampling process is expressed as:

[0054] Z OOD ={z~U z |M(z)<τ} (8)

[0055] S34: Decoder reconstructs and generates auxiliary negative samples.

[0056] Generate auxiliary negative samples X through decoder reconstruction OOD =g θ (Z OOD ).

[0057] S4: Use generated negative samples to perform comparative learning with known category samples to optimize the model's ability to distinguish abnormal samples;

[0058] As a preferred embodiment of the present invention, in step S4, the comparative learning process specifically includes the following steps:

[0059] S41: Construct contrastive learning batches. In each training iteration, a batch of samples {x ID}, and sample a batch of samples {x OOD}, together form a mixed small batch data B = {x ID}∪{x OOD}.

[0060] S42: Extract sample feature representation. Input all samples in the mixed batch B constructed in step S41 into the encoder model f that needs to be optimized again. φ , obtain the feature representation of all samples.

[0061] S43: Define contrastive learning sample pairs.

[0062] Select each known category sample in the mixed batch B of positive and negative samples in turn As an anchor point, and define its corresponding positive and negative samples;

[0063] Among them, the positive samples are selected from other samples of the same category in the mixed batch B The goal is to bring similar samples closer in the latent space; negative samples are auxiliary negative samples x generated in the batch OOD , the goal is to keep dissimilar samples away from each other in the latent space.

[0064] S44: Calculate the contrast loss function. The contrast loss function InfoNCE Loss is used to measure and optimize the sample distribution in the latent space. and its characteristics The loss function maximizes the distance between the anchor point and the positive sample feature z pos The cosine similarity of the anchor point and all negative sample features z is minimized neg The cosine similarity between the two is used to “pull the positive samples closer and push the negative samples further away.” The temperature coefficient κ is introduced to adjust the discrimination. The loss function can be expressed as:

[0065]

[0066] S45: Update the encoder model parameters. Calculate the average contrast loss of all anchor samples in a batch The model parameters are updated through back propagation and gradient descent. Steps S41 to S45 are repeated until the model performance meets the requirements, enabling it to generate feature representations that can better distinguish known class samples from abnormal samples, thereby improving the ability to distinguish abnormal samples.

[0067] S5: Design an EDL classifier to predict the failure mode of input samples and quantify the uncertainty of the prediction results;

[0068] As a preferred embodiment of the present invention, in step S5, the quantitative model prediction is uncertain, and constructing a credible prediction interval includes the following steps:

[0069] S51: Construct the EDL classifier architecture. Use the encoder model optimized in step S4 as the feature extractor. Connect multiple fully connected layers to form the Evidence Generation Distribution Network classifier. The final output layer of the network has K neurons (K is the number of known failure modes) and uses the ReLU activation function to ensure that the output value is non-negative, representing the evidence e for each category. That is, for the input sample x, the output of the classifier is the evidence vector where e=[e1,e2,...,eK ], and e K ≥0.

[0070] S52: Calculate the parameters of the Dirichlet distribution. Generate network h based on evidence φ The output evidence vector e is used to calculate the corresponding Dirichlet distribution parameter α. The relationship between parameter α and evidence e is defined as:

[0071] α k =e k +1 (10)

[0072] Among them, α k is the parameter of the Dirichlet distribution corresponding to the kth category, e k is the evidence of the kth category output by the network. This ensures that all α k ≥1. Therefore, for the input sample x i , whose prediction results are given by a Dirichlet distribution D(p i |α i ) is used to represent the i =[α i,1 ,...,α i,K ].

[0073] S53: Train the EDL classifier. Freeze the encoder f φ Parameters, use known class sample data to train the EDL classifier By minimizing the EDL loss function L EDL To update the model parameters, the loss function usually includes:

[0074] Dirichlet distribution fitting loss: Combined with the mean square error loss to minimize the prediction error and variance, specifically expressed as:

[0075]

[0076] in, is the Dirichlet intensity, is the expected probability of the kth class.

[0077] KL divergence regularization term: In order to avoid the complete coordination between the evidence and the predicted probability output by the model, a penalty is imposed on the prediction mode that does not show any uncertainty at all. The KL divergence regularization term is introduced:

[0078]

[0079] Among them, 1 is a vector of all 1s, D(p i |1) is the uniform Dirichlet distribution, which represents the case with maximum uncertainty, It is a parameter that filters out the corresponding items in the label that are not 0, ensuring that the evidence of the category corresponding to the label remains unchanged, while other evidence tends to be highly uncertain.

[0080]

[0081] Among them, λ t =min(1.0,t / T) is the annealing coefficient, t is the current training round number, and T is the preset annealing period.

[0082] S54: Perform failure mode prediction and uncertainty quantification. For a new input sample x new , through the trained model, its evidence vector Dirichlet parameter α = e + 1 and Dirichlet intensity Then, calculate the expected probability of each class The category with the highest expected probability is selected as the predicted failure mode. At the same time, the overall uncertainty u of the prediction is calculated according to the subjective logic theory:

[0083]

[0084] S6: Establish a discrimination threshold criterion based on uncertainty, judge high uncertainty samples as unknown fault types, and output accurate classification results of known faults.

[0085] As a preferred embodiment of the present invention, the discrimination threshold of the uncertainty quantification described in step S6 is based on the distribution characteristics of the training set prediction uncertainty measure, and its 95% quantile is used as the discrimination threshold of the unknown category to achieve the category decision boundary division in the open set recognition task; when the uncertainty measure value calculated by S5 of the test sample is greater than the discrimination threshold, it is judged to be an unknown fault type; if the uncertainty measure is lower than the discrimination threshold, the fault type is judged according to the model output category prediction probability.

[0086] To evaluate the proposed network model, the CWRU bearing fault dataset was selected to simulate fault diagnosis in the presence of unknown faults. The CWRU dataset contains single-point faults in the inner race, outer race, and sphere. Each fault type has three fault sizes: 0.007, 0.014, and 0.028 inches. Class labels are defined as 1-9, with the normal bearing class label set to 0. Samples from different classes are then treated as unknown classes, making them inaccessible during training. Five open-set fault diagnosis tasks are defined, with the specific settings shown in Table 1:

[0087] Table 1 Open set diagnosis task settings based on the CWRU dataset

[0088]

[0089] Select EVT and OpenMax for comparative experiments, select AUROC and FPR 95 The indicator evaluates the model's ability to identify unknown classes. AUROC is the area under the ROC curve, which is used to evaluate the model's comprehensive performance in distinguishing positive and negative categories at different thresholds. A larger value indicates a stronger classification ability. 95 When the recall rate of true positive examples (known classes) reaches 95%, the false positive rate at which the model mistakenly classifies negative examples (unknown classes) as positive. Lower values ​​indicate stronger unknown class rejection. Furthermore, the AcD metric is defined to represent the accuracy of the model's predictions for the remaining samples after rejecting unknown class samples. The specific results are shown in Table 2:

[0090] Table 2 Open set fault diagnosis results

[0091]

[0092] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, can make equivalent replacements or changes based on the technical solution and inventive concept of the present invention, which should be within the scope of protection of the present invention.

Claims

1. An open set fault diagnosis method combining generative data augmentation and uncertainty measurement, characterized in that: The specific steps include: S1: Collect multi-operational data of mechanical equipment and process vibration signals to construct a training sample set; S2: Construct a VAE-GMM generative model based on mixed Gaussian priors; S3: Based on the low probability density sampling strategy, the model built in S2 generates out-of-distribution negative samples that are separated from the known categories; S4: Use generated negative samples to perform comparative learning with known category samples to optimize the model's ability to distinguish abnormal samples; S5: Design an EDL classifier to predict the failure mode of input samples and quantify the uncertainty of the prediction results; S6: Establish a discrimination threshold criterion based on uncertainty, judge high uncertainty samples as unknown fault types, and output accurate classification results of known faults.

2. The open set fault diagnosis method combining generative data augmentation and uncertainty measurement according to claim 1, characterized in that: The VAE-GMM generation model in step S2 includes a feature encoder f φ , reconstruct decoder g θ , GMM fitting module and loss calculation module; The feature encoder f φ Specifically: the vibration signal is mapped into the latent space through a multi-layer convolutional network to extract the low-dimensional latent code z; The reconstruction decoder g θ Specifically, deconvolution and upsampling operations are used to remap the latent code z in the latent space to the sample space to generate a reconstructed signal with the same dimension as the original vibration signal; The GMM fitting module specifically updates the mixed Gaussian distribution fitting model of known class samples in the latent space through model training iteration, thereby realizing the sampling of known class samples in the latent space; The loss calculation module is specifically as follows: the overall loss term of the model includes the separation correlation loss L corr , signal reconstruction loss L recon and distribution difference loss L dist .

3. The open set fault diagnosis method combining generative data augmentation and uncertainty measurement according to claim 2, characterized in that: The separation correlation loss L corr The calculation method is: the Euclidean distance d(z1,z2) of the sample in the latent space and the Euclidean distance d(g θ (z1),g θ (z2)) Calculate the correlation loss, which is specifically expressed as: The signal reconstruction loss L recon , distribution difference loss L dist It is derived from the Evidence Lower Bound (ELBO) of VAE. After introducing the Gaussian mixture model prior, ELBO is expressed as Among them, c represents category information, and x represents input samples; The first term of ELBO represents the signal reconstruction loss L recon ,use Represents sample x i The corresponding reconstructed signal generated by this model is N, which represents the total number of training samples; the signal reconstruction loss L recon Expand to get: The second term of ELBO represents the distribution difference loss L dist , by approximating the variational posterior q φ (c, z|x) and the KL divergence calculation of the mixed Gaussian model prior p(c, z); According to the definition of KL divergence and Bayesian theorem, the distribution difference loss L dist Expanding and simplifying gives: The overall loss term of the model is the sum of the loss terms, which is expressed as follows: L=L recon +L dist +L corr 。 4. The open set fault diagnosis method combining generative data augmentation and uncertainty measurement according to claim 1, characterized in that: The step S3 specifically includes the following steps: S31: Pre-training generative models for known class samples; Pre-training the known class sample data using the generative model constructed in step S2; the model uses a Gaussian mixture distribution as the prior distribution of the latent space, obtains the training loss based on the loss calculation module, and simultaneously updates the feature encoder and reconstruction decoder parameters, as well as the GMM parameters, through gradient descent; S32: Determine the distribution of known class samples in the latent space; After the VAE-GMM generative model is trained, the distribution of known class data in the latent space uses the learned mixed Gaussian distribution. express: in, Represents the Gaussian sub-distribution corresponding to a given class c; Set the specified threshold τ, when the probability density of a given sampling point z When it exceeds the threshold, it is regarded as the hidden code Z of the known class sample in the hidden space ID , auxiliary positive sample signals can be generated through the decoder, and the specific sampling process is expressed as: S33: Generate hidden codes of auxiliary negative samples through low probability density sampling strategy; First, from the uniform distribution U in the latent space z Sampling multiple points in the , and then based on the learned mixed Gaussian distribution Filter out points with high probability density and select points with probability density lower than the specified threshold τ as the hidden code for auxiliary negative samples. The specific sampling process is expressed as: WITH OOD ={z~U z |M(z)<τ} S34: Generate auxiliary negative samples X through decoder reconstruction OOD =g θ (Z OOD ).

5. The open set fault diagnosis method combining generative data augmentation and uncertainty measurement according to claim 1, characterized in that: The comparative learning process of step S4 specifically includes the following steps: S41: Construct contrastive learning batches; In each training iteration, a batch of samples {x ID }, and sample a batch of samples {x OOD }, together form a small batch of positive and negative sample mixed data B = {x ID }∪{x OOD }; S42: extract sample feature representation; Input all samples in the mixed batch B of positive and negative samples constructed in step S41 into the encoder model f φ , obtain the feature representation of all samples; S43: define contrastive learning sample pairs; Select each known category sample in the mixed batch B of positive and negative samples in turn As an anchor point, and define its corresponding positive and negative samples; Among them, the positive samples are selected from other samples of the same category in the mixed batch B (k≠i); negative samples are selected as auxiliary negative samples x in mixed batch B OOD ; S44: Calculate the contrast loss function; use the contrast loss function InfoNCE Loss to measure and optimize the sample distribution in the latent space; S45: Update encoder model parameters; Calculate the average contrast loss of all anchor samples in a batch Update model parameters through backpropagation and gradient descent.

6. The open set fault diagnosis method combining generative data augmentation and uncertainty measurement according to claim 1, characterized in that: The EDL classifier in step S5 is constructed based on the Dirichlet distribution and DST evidence theory. The EDL classifier calculates a non-negative output through the ReLU activation function, regards the output as the evidence vector for each fault category, calculates the relevant parameters of the Dirichlet distribution based on the evidence vector, and implements classifier training based on the constructed EDL loss function. Finally, the trained EDL classifier is used to predict the fault mode and quantify the uncertainty of the input sample.

7. The open set fault diagnosis method combining generative data augmentation and uncertainty measurement according to claim 1, characterized in that: The step S6 is specifically as follows: Based on the distribution characteristics of the uncertainty measure of the training set prediction, its 95% quantile is used as the judgment threshold of the unknown category to realize the category decision boundary division in the open set recognition task; when the uncertainty measure value of the test sample calculated in step S5 is greater than the discrimination threshold, the test sample is judged to be an unknown fault type; if the uncertainty measure is lower than the discrimination threshold, the test sample is judged to be a known fault type, and the fault type is determined according to the category prediction probability output by the model.

Images