An improved prototype network-based method for diagnosing bearing faults in the rocker arm spindle of a coal mining machine.

By using an improved CNN-Mamba feature extractor and a prototype network with Wasserstein consistency constraints, the problem of small-sample fault diagnosis of rocker arm spindle bearings in coal mining machines across working conditions was solved, achieving efficient and stable fault identification results.

CN123115995APending Publication Date: 2026-11-17XIAN COAL MINING MACHINERY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611131760.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to achieve efficient small-sample fault diagnosis of rocker arm spindle bearings under heavy load, impact, and variable operating conditions, especially lacking robustness in fault feature extraction and diagnosis across operating conditions.

Method used

An improved prototype network combining a CNN-Mamba feature extractor with Wasserstein distribution consistency constraints is adopted. Through overlapping convolutional embedding, local shock differential convolution, shock saliency gate and state compensation fusion module, local shock details are captured and long-range modeling is performed. At the same time, the Wasserstein consistency loss constraint supports the consistency of the predicted distribution of the support set and the query set, thereby improving the diagnostic accuracy and stability.

Benefits of technology

The accuracy of cross-condition fault diagnosis of the rocker arm spindle bearing of the coal mining machine was significantly improved under small sample conditions, reaching 88.92% in the 10-way 1-shot scenario and 95.82% in the 5-shot scenario, meeting engineering requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115995A_ABST
    Figure CN123115995A_ABST
Patent Text Reader

Abstract

This invention discloses an improved prototype network-based method for fault diagnosis of rocker arm spindle bearings in coal mining machines, belonging to the field of coal mine equipment health monitoring and small-sample fault diagnosis. First, a small-sample meta-task based on source-target operating conditions is constructed. Second, a CNN-Mamba feature extractor is designed, extracting local impact features through multi-scale differential convolution, guiding Mamba modeling of long-range state dependencies through impact saliency gates, and then outputting highly discriminative fault embeddings through state compensation fusion and impact-aware weighted convergence. Then, based on the prototype network classification, Wasserstein distance is introduced to constrain the consistency of the predicted distributions of the support set and the query set. Finally, a joint loss model is trained to achieve rapid adaptive diagnosis of a small number of samples under the target operating condition. This invention significantly improves the accuracy and robustness of fault diagnosis under small-sample conditions across operating conditions and is suitable for online health monitoring of rocker arm spindle bearings in coal mining machines.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of health monitoring of intelligent equipment in coal mines, fault diagnosis of rotating machinery and small sample learning, and particularly relates to an improved prototype network method for fault diagnosis of bearings of rocker arm spindles of coal mining machines, which is applicable to small sample fault diagnosis of bearings in rocker arm transmission systems of coal mining machines under heavy load, impact and variable working conditions coupled environments. Background Technology

[0002] The coal mining machine is the core production equipment in a fully mechanized mining face, and its operational reliability directly determines coal mining efficiency and mine production safety. As a key component for transmitting cutting power in the coal mining machine, the rocker arm transmission system operates under harsh conditions of heavy load, strong impact, and variable load for extended periods. Subjected to the combined effects of fluctuating coal and rock hardness, cutting impact, and coupled vibration of the transmission chain, the main shaft bearing is prone to wear, pitting, and fatigue spalling. If these faults are not identified in time, they can lead to reduced cutting efficiency, transmission system failure, unplanned equipment downtime, or even safety accidents. Therefore, accurate fault diagnosis of the rocker arm main shaft bearing in coal mining machines has significant engineering value for promoting intelligent operation and maintenance in coal mines.

[0003] In recent years, deep learning-based fault diagnosis methods based on vibration signals have become the mainstream technical approach, with models such as convolutional neural networks, recurrent neural networks, and Transformers widely used in rotating machinery fault diagnosis. However, these methods generally rely on a large number of labeled samples for training, while the fault conditions in coal mining machines are difficult to reproduce, the cost of fault sample collection is extremely high, and available labeled fault samples are very scarce. Meanwhile, operating condition fluctuations are a core factor limiting diagnostic performance: affected by changes in rotational speed and load, the vibration response of the same fault under different operating conditions exhibits significant distribution shifts, directly leading to a decrease in the model's cross-condition generalization ability. Existing cross-condition diagnosis methods mostly employ domain adaptive methods, but these generally require a large amount of target operating condition data for training to achieve distribution alignment. However, in actual engineering, labeled data for unknown target operating conditions is usually extremely scarce, failing to meet the application requirements of these methods.

[0004] For small-sample, cross-working-condition diagnostic scenarios, prototype networks based on meta-learning have become a research hotspot in this field due to their simple structure and high training efficiency. Prototype networks learn a general feature metric space by constructing a large number of small-sample meta-tasks to support the construction of category prototypes using set feature means, and achieve query sample classification through distance metrics, enabling rapid adaptation to new working conditions. However, when directly applied to the diagnosis of rocker arm spindle bearings in coal mining machines, two key shortcomings remain: First, feature extraction networks struggle to balance local impact details with long-term state correlations. The rocker arm vibration signal contains both high-frequency local impact mutations caused by bearing failures and long-term temporal correlations due to load fluctuations and transmission path coupling. Existing prototype networks mostly use ordinary convolutional backbones, which lack the ability to model long-term dependencies, while Transformer-type backbones suffer from high computational cost, overfitting to small samples, and insufficient stability in cross-condition feature representation.

[0005] Secondly, traditional prototype networks construct category prototypes solely based on feature mean, lacking consistency constraints on the predicted distributions of the support set and query set. Under the dual influence of operational bias and sample scarcity, the feature distribution of similar samples exhibits high dispersion, making it prone to matching bias when relying solely on mean prototypes, thus reducing diagnostic robustness.

[0006] In summary, existing technologies cannot simultaneously achieve accurate extraction of local impact features, efficient modeling of long-term time-dependent processes, and improved stability of prototype matching, and therefore cannot effectively solve the engineering challenge of small-sample fault diagnosis of rocker arm spindle bearings across working conditions in coal mining machines. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing an improved prototype network-based method for fault diagnosis of the rocker arm spindle bearing of a coal mining machine. By using CNN-Mamba collaborative feature extraction and Wasserstein distribution consistency constraints, the accuracy, stability, and engineering practicality of fault diagnosis under small sample cross-working conditions are improved.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: an improved prototype network method for diagnosing faults in the rocker arm spindle bearing of a coal mining machine, comprising the following steps: S1. Collect one-dimensional vibration signal data of the rocker arm spindle bearing of the coal mining machine under source and target working conditions. Perform fixed-length slicing, normalization and labeling on the vibration signal. Construct an N-way K-shot meta-task containing support set and query set according to the small sample meta-learning method.

[0009] S2. Construct a CNN-Mamba feature extractor. Input the vibration samples from the meta-task into this feature extractor. Then, pass them sequentially through an overlapping convolutional embedding module, a local impact differential convolution module, a Mamba-based impact saliency-guided module, a state-compensated fusion module, and an impact-aware weighted convergence module to output local-global collaborative sample-level fault embedding features.

[0010] S3. Take the mean value of the embedding features of each category of samples in the support set to obtain the category prototype, calculate the similarity between the embedding features of the query sample and each category prototype and obtain the fault category probability, and calculate the prototype classification loss.

[0011] S4. For the same fault category within the same meta-task, calculate the average prediction distribution of the support set and the query set respectively, and use the first-order Wasserstein distance to measure the difference between the two distributions to construct the Wasserstein consistency loss.

[0012] S5. The total training loss is obtained by weighted summation of the prototype classification loss and the Wasserstein consistency loss. The full parameters of the model are then updated by gradient descent to complete the meta-training stage.

[0013] S6. Target Working Condition Adaptation and Diagnosis: A small number of labeled samples under the target working condition are used as the target support set and input into the trained feature extractor to calculate the target category prototype; the samples to be diagnosed are used as the target query set and input into the model, and the final fault diagnosis result is output according to their similarity with the target prototype; the fault diagnosis result is that the fault types of the rocker arm main shaft bearing of the coal mining machine include normal state, inner ring fault, outer ring fault, and rolling element fault, and each type of fault contains subcategories with different defect sizes.

[0014] Further, in step S1, each meta-task is an N-way K-shot classification task; where N represents the number of fault categories in the current meta-task, and K represents the number of fault categories in the support set. The number of samples included in the query set Each category contains M samples that do not overlap with the support set.

[0015] Furthermore, the specific implementation of each module in step S2 is as follows: Overlapping convolutional embedding module: This module uses one-dimensional convolution with a stride smaller than the kernel size to embed the original vibration signal into blocks, compressing the sequence length while preserving continuous vibration information of adjacent segments. Learnable positional codes are then superimposed to inject temporal positional information into the sequence, resulting in the segment feature sequence Z. The specific formula is as follows:

[0016] in, This indicates the input vibration sample. Represents the overlapping convolutional embedding map. Indicates position code, This represents the embedded segment feature sequence; the overlapping convolution embedding compresses the sequence length while preserving the continuous vibration information between adjacent segments.

[0017] Local Impact Differential Convolution Module: Parallel one-dimensional convolutions of sizes 3, 5, and 7 are used to extract multi-scale local impact features; at the same time, first-order difference is performed on the features of adjacent segments to enhance the abrupt change features of fault impact; after concatenating the multi-scale features and the differential features, the local impact features are obtained by pointwise convolution.

[0018] Based on the Mamba shock saliency guidance module: local shock features and differential shock features are concatenated, and then a shock saliency gate is generated through pointwise convolution and sigmoid activation to highlight the feature weights of fault mutation segments. The specific formula is as follows:

[0019] in, Indicates channel splicing; This represents a one-dimensional pointwise convolution; express Activation function This indicates a segment used to enhance abrupt changes caused by bearing impact failure; Indicates localized impact characteristics; This represents the differential impact characteristic.

[0020] The impact saliency gate is multiplied element-wise with the local impact features to enhance the features of the fault impact abrupt segment. This enhanced feature is then input into the Mamba module for long-range state modeling, outputting the global state features. The specific formula is as follows:

[0021] in, This represents element-wise multiplication. This represents the global state characteristics output by Mamba. Mamba is a sequence modeling method based on a selective state-space model that can adaptively adjust the state according to the feature content of different sequence positions.

[0022] State compensation fusion module: This module uses global state features as long-range compensation terms for local impact features. An adaptive fusion gate dynamically controls the compensation intensity, enabling the model to retain local impact sensitivity while incorporating long-range operational state information, resulting in fused segment features. The specific formula is as follows:

[0023] in, Representation layer normalization, This represents element-wise multiplication. Indicates a fusion gate. Represents global state characteristics. Indicates local impact characteristics. This represents the difference between global state characteristics and local impact characteristics.

[0024] Impact-aware weighted convergence module: Calculates the contribution weight of each segment based on the impact saliency gate, and performs a weighted summation of the fused segment features to ensure a higher proportion of high-impact segments in the sample-level embedding; after L2 normalization, it outputs the final sample-level fault embedding features, with the specific formula as follows:

[0025] in, This represents the fusion feature of the nth segment. Indicates the weight of the corresponding segment. express Normalization.

[0026] Furthermore, the specific process of step S3 is as follows: S3.1. Take the mean of the embedded features of samples of the same fault category in the support set to obtain the category prototype:

[0027] in Indicates support set samples, This indicates that support for centralized categories belongs to the category. The sample set, Indicates feature extractor, These are the model parameters.

[0028] S3.2 Calculate the cosine similarity between the embedded features of the query sample and the prototypes of each category. The specific formula is as follows:

[0029] in For the query set sample, For fault category prototypes, This refers to the temperature parameter.

[0030] After scaling by the temperature coefficient, the probability distribution of the query sample belonging to each category is obtained through the Softmax function. The specific formula is as follows: .

[0031] S3.3. Based on the true labels and predicted probability distributions of the query samples, the prototype classification loss is calculated using cross-entropy loss. The specific formula is as follows:

[0032] Where N represents the number of fault categories in the current meta-task, and M represents the number of samples in the query set that do not overlap with the support set for each category.

[0033] Furthermore, in step S4, the Wasserstein consistency loss constrains the predicted probability distribution rather than feature domain alignment; by minimizing the predicted distribution distance between the support set and the query set of the same class, it reduces the prototype matching bias caused by the working condition offset and improves the stability of the class prototype's representation of the target sample. The specific process is as follows: S4.1, for the first [unclear] within the same meta-task For each type of fault, calculate the average prediction distribution of the support set samples. Average prediction distribution of the query set samples The specific formula is as follows: , in, The fault category is Support set, The fault category is query set, This represents the probability that a sample belongs to each fault category.

[0034] S4.2 Calculate the Wasserstein distance between the two classes of average prediction distributions. The mean Wasserstein distance for all classes constitutes the Wasserstein consistency loss.

[0035] in, The number of fault categories for the meta-task. This represents the first-order Wasserstein distance.

[0036] Furthermore, the total training loss in step S5 The calculation formula is:

[0037] in, Represents the prototype classification loss. This represents the Wasserstein consistency loss. This represents the consistency loss weighting coefficient, with a value ranging from 0.05 to 0.2.

[0038] Compared with the prior art, the present invention has the following advantages: 1. The CNN-Mamba feature extractor designed in this invention accurately captures the local impact details of bearing failures using multi-scale differential convolution, guides Mamba to focus on fault segments for long-term modeling using impact saliency gates, and then achieves local-global feature collaboration through adaptive compensation fusion. Compared with pure CNN or pure Mamba architectures, it retains the advantages of convolution in local feature extraction and achieves long-term temporal dependency modeling with low computational cost, significantly improving the discriminability and stability of fault features across operating conditions.

[0039] 2. This invention introduces Wasserstein distance to construct a support set-query set prediction distribution consistency loss, which constrains the distribution consistency of similar samples from the classification decision level, making up for the shortcomings of traditional prototype networks that only use the mean to represent the data; in scenarios with small samples and strong operational bias, it can effectively reduce the interference of distribution dispersion on prototype matching and improve diagnostic robustness.

[0040] 3. This invention combines CNN-Mamba feature extraction with Wasserstein distribution constraints, creating a synergistic gain effect: high-quality embedded features provide a reliable foundation for distribution constraints, while the distribution constraints further optimize the intra-class clustering of the feature space. Ablation experiments show that the combined approach significantly improves diagnostic accuracy compared to a single improvement scheme, achieving 88.92% accuracy in a 10-way 1-shot cross-condition scenario and 95.82% in a 5-shot scenario, meeting the engineering requirements for small-sample cross-condition diagnosis of coal mining machine rocker arm spindle bearings. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the CNN-Mamba feature extractor described in this invention.

[0042] Figure 2 This diagram illustrates the comparison of the accuracy of the method of this invention with existing comparative methods in fault diagnosis across small samples and operating conditions. Detailed Implementation

[0043] The present invention will be further described below.

[0044] The core idea of ​​the improved prototype network-based fault diagnosis method for the rocker arm spindle bearing of a coal mining machine is as follows: the joint feature representation of local impact and long-range state is achieved through the CNN-Mamba collaborative architecture, and the consistency of the predicted distribution of the support set and the query set is constrained by the Wasserstein consistency loss. The combination of the two improves the fault diagnosis performance of small samples across working conditions.

[0045] like Figure 1 As shown, the CNN-Mamba feature extractor takes a one-dimensional vibration sample as input. First, it generates a segment feature sequence through overlapping convolution embedding. Then, it extracts multi-scale impact features and differential mutation features through local impact differential convolution. Subsequently, it highlights the fault segment through the impact saliency gate and inputs it into the Mamba module to model long-range state dependence. Then, it adaptively fuses local and global features through the state compensation fusion module. Finally, it obtains the sample-level fault embedding features through impact-aware weighted convergence.

[0046] During model training, class prototypes are constructed using the support set embedding mean. The classification probability of the query sample is calculated using cosine similarity, yielding the prototype classification loss. Simultaneously, the Wasserstein distance between the predicted distributions of similar support sets and the query set is calculated, resulting in the consistency loss. The model parameters are then optimized end-to-end using the joint loss. During testing, a target support set is constructed using a small number of labeled samples from the target working condition. After calculating the target prototype, rapid classification of the sample to be diagnosed can be achieved.

[0047] Example 1: Small-sample fault diagnosis of rocker arm spindle bearing of coal mining machine across working conditions based on self-made dataset This embodiment uses a self-built fault simulation experimental platform to collect vibration data to verify the effectiveness of the method of the present invention.

[0048] Experimental Platform and Dataset: The experimental platform consists of a drive motor, transmission gearbox, magnetic powder brake, speed and torque sensor, and acceleration sensor. The NSK NJ204E cylindrical roller bearing was selected as the simulation object. Three typical fault categories were set: normal state, inner ring fault, outer ring fault, and rolling element fault. For each fault category, wire cutting was used to machine three defect widths: 0.2mm, 0.5mm, and 0.8mm, resulting in a total of 10 health state categories (1 normal category + 9 fault categories).

[0049] Three load conditions (C1, C2, and C3) of 0 Nm, 10 Nm, and 20 Nm were set to simulate a variable load environment; the vibration signal sampling frequency was 10 kHz, the length of each sample was 1024 points, and 200 samples were selected for each state under each condition.

[0050] Experimental setup: Six cross-condition diagnostic tasks were constructed: C1→C2, C1→C3, C2→C1, C2→C3, C3→C1, and C3→C2. Two small sample scenarios were set up: 10-way 1-shot and 10-way 5-shot. Each class of the query set contained 15 samples. During the testing phase, 1000 episodes were randomly constructed, and the average accuracy of 5 randomized repeated experiments was used as the evaluation metric.

[0051] Model parameters: The feature extractor output embedding dimension is 128, and the Wasserstein consistency loss weight λ = 0.1; the optimizer uses Adam with an initial learning rate of 1 × 10⁻⁶. -3 The training run consists of 1500 episodes.

[0052] Comparison Methods and Result Analysis: CNN, ProtoNet, RelationNet, MAML, ANIL, and CDTAPN were selected as comparison algorithms. The diagnostic accuracy comparison results are as follows: Figure 2 As shown.

[0053] Experimental results show that the method of the present invention achieves an average accuracy of 88.92% in the 10-way 1-shot scenario and 95.82% in the 10-way 5-shot scenario, both of which are significantly better than all the comparison methods.

[0054] To verify the effectiveness of each improved module, an ablation experiment was conducted in this embodiment, and the results are shown in the table below:

[0055] The ablation results show that both the CNN-Mamba feature extractor and the Wasserstein consistency loss contribute positively to the diagnostic performance. Among them, CNN-Mamba improves accuracy more significantly, reflecting the fundamental role of feature representation. The combination of the two further improves performance (i.e., the CNN-Mamba feature extractor enhances the model's joint expression ability for local shocks and long-range state dependencies, and the Wasserstein consistency constraint improves the matching stability of the support set prototype to the query sample. The combination of the two can significantly improve the fault diagnosis effect of small samples of the rocker arm spindle bearing of the coal mining machine across working conditions), proving that there is a synergistic gain effect between the modules.

[0056] In summary, the method of this invention can achieve high-precision cross-working-condition fault diagnosis of the rocker arm spindle bearing of a coal mining machine with very few labeled samples, and has good engineering application value.

Claims

1. A method of improving a prototype network's shearer rocker arm spindle bearing fault diagnosis, characterized in that, Includes the following steps: S1. Collect one-dimensional vibration signals of the rocker arm spindle bearing of the coal mining machine under source and target working conditions. After processing the signals, construct a meta-task containing support set and query set according to the small sample meta-learning method. S2. Construct a CNN-Mamba feature extractor. Input the vibration samples from the meta-task into the feature extractor. Then, pass them sequentially through an overlapping convolutional embedding module, a local impact differential convolution module, a Mamba-based impact saliency-guided module, a state compensation fusion module, and an impact-aware weighted convergence module to output local-global collaborative fault embedding features. S3. Calculate the category prototype of the fault embedding features of various support set samples, and calculate the query set fault category probability by the cosine similarity between the query set sample embedding features and each category prototype to obtain the prototype classification loss. S4. Construct Wasserstein consistency loss based on the category prediction distribution of the support set and query set, and impose distribution consistency constraints on the average prediction distribution of the support set and the average prediction distribution of the query set for the same fault category within the same meta-task. S5. The prototype classification loss from step S3 and the Wasserstein consistency loss from step S4 are weighted and summed to obtain the total training loss. The model parameters are then updated based on the total training loss to obtain the trained fault diagnosis model. S6. Input the trained fault diagnosis model with a small number of labeled samples under the target working condition as the target support set, and input the model with the sample to be diagnosed under the target working condition as the target query set. Output the fault diagnosis result of the rocker arm spindle bearing based on the similarity between the sample to be diagnosed and the target category prototype.

2. The method of claim 1, wherein, In step S1, each meta-task is an N-way K-shot classification task; where N represents the number of fault categories in the current meta-task, K represents the number of samples contained in the support set for each fault category, and each category in the query set contains M samples that do not overlap with the support set.

3. The method according to claim 1, characterized in that, In step S2, the processing procedure of the overlapping convolutional embedding module is as follows: The input vibration sample undergoes a one-dimensional overlapping convolutional mapping with a stride smaller than the kernel size. After superimposing learnable positional encodings, the segment feature sequence Z is obtained. The specific formula is: in, This indicates the input vibration sample. Represents the overlapping convolutional embedding map. Indicates position code, This represents the embedded segment feature sequence; the overlapping convolution embedding compresses the sequence length while preserving the continuous vibration information between adjacent segments.

4. The method according to claim 1, characterized in that, In step S2, the local impact differential convolution module processes the following steps: using three parallel one-dimensional convolution branches with kernel sizes of 3, 5, and 7, respectively, to extract local fault impact features at different time scales; simultaneously, performing first-order difference calculations on adjacent segment features to obtain differential impact features; and then concatenating the multi-scale local impact features and differential impact features through channels, followed by one-dimensional pointwise convolution to obtain the final local impact features.

5. The method according to claim 1, characterized in that, In step S2, the processing procedure based on the Mamba impact significance guidance module is as follows: S2.

1. The local impact features and differential impact features are concatenated into channels, and then the impact saliency gate is generated by one-dimensional pointwise convolution and the Sigmoid activation function. The specific formula is as follows: in, Indicates channel splicing; This represents a one-dimensional pointwise convolution; express Activation function This indicates a segment used to enhance abrupt changes caused by bearing impact failure; Indicates localized impact characteristics; Indicates the differential impact characteristics; S2.2 Multiply the impact saliency gate element-wise with the local impact features to enhance the features of the fault impact abrupt change segment. Then input the enhanced features into the Mamba module for long-range state modeling and output the global state features. The specific formula is as follows: in, This represents element-wise multiplication. This represents the global state characteristics output by Mamba.

6. The method according to claim 1, characterized in that, In step S2, the state compensation fusion module processes the following steps: It uses the global state features as the long-range state compensation term for the local impact features, calculates the difference between the two, controls the compensation intensity through an adaptive fusion gate, and outputs the fused segment features. The specific formula is as follows: in, Representation layer normalization, This represents element-wise multiplication. Indicates a fusion gate. Represents global state characteristics. Indicates local impact characteristics. This represents the difference between global state characteristics and local impact characteristics.

7. The method according to claim 1, characterized in that, In step S2, the processing procedure of the impact-sensing weighted convergence module is as follows: calculate the segment weights based on the impact response intensity of each segment, perform weighted summation on the fused segment features, and output sample-level fault embedding features after normalization. The specific formula is as follows: in, This represents the fusion feature of the nth segment. Indicates the weight of the corresponding segment. express Normalization.

8. The method according to claim 1, characterized in that, The specific process of step S3 is as follows: S3.

1. Take the mean of the embedded features of samples of the same fault category in the support set to obtain the category prototype: in Indicates support set samples, This indicates that support for centralized categories belongs to the category. The sample set, Indicates feature extractor, These are model parameters; S3.2 Calculate the cosine similarity between the embedded features of the query sample and the prototypes of each category, and obtain the probability distribution of the query sample belonging to each category through the Softmax function after scaling by the temperature coefficient. S3.

3. Based on the true labels and predicted probability distributions of the query samples, the prototype classification loss is calculated using cross-entropy loss.

9. The method according to claim 1, characterized in that, The specific process of step S4 is as follows: S4.1, for the first [unclear] within the same meta-task For each type of fault, calculate the average prediction distribution of the support set samples. Average prediction distribution of the query set samples ; S4.2 Calculate the Wasserstein distance between the two classes of average prediction distributions. The mean Wasserstein distance for all classes constitutes the Wasserstein consistency loss. in, The number of fault categories for the meta-task. This represents the first-order Wasserstein distance.

10. The method according to claim 1, characterized in that, Total training loss in step S5 The calculation formula is: in, Represents the prototype classification loss. This represents the Wasserstein consistency loss. This represents the consistency loss weighting coefficient, with a value ranging from 0.05 to 0.2.