A long-tailed distribution data fault diagnosis method and device and a storage medium

Abstract

The application relates to the technical field of fault diagnosis, in particular to a long-tail distribution data fault diagnosis method, device and equipment and a computer storage medium. The long-tail distribution data fault diagnosis method based on a learnable data enhancement network comprises the following steps: a set of transformations is generated, the transformations are different from each other while keeping similar semantic information with original samples; the set of transformations is used as positive samples of supervised contrast representation learning, the representation ability of supervised contrast learning is improved, a feature vector of the original sample is obtained through supervised contrast learning, and then a classifier is inputted to perform fault diagnosis. Compared with other methods, the long-tail distribution data fault diagnosis method has higher accuracy and better performance.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method, apparatus, equipment, and computer storage medium for diagnosing faults in long-tailed distributed data. Background Technology

[0002] With the widespread application of artificial intelligence (AI) technology, modern manufacturing processes are moving towards intelligentization. Through information interaction and connection, various links in a factory can be interconnected, thereby achieving intelligent manufacturing based on automated production. In the process of building smart factories, equipment failure prediction and health management are key issues. If potential hazards cannot be predicted and handled in a timely manner, normal industrial production processes may be affected, leading to equipment damage, decreased product quality, and even personal injury or death. Therefore, ensuring the safety and reliability of modern production systems has become an urgent issue, not only concerning the safety of people's lives and property but also contributing to enhancing the market competitiveness and economic benefits of enterprises.

[0003] Fault diagnosis technology for complex industrial processes can detect faults in a timely manner, ensuring the safe operation of the system. However, modern industrial production processes are extremely complex, involving numerous processing stages, making it difficult and impractical to establish accurate mathematical models and perform fault diagnosis. With the advancement of information technology, real-time production data can be stored. Therefore, extracting key information from massive amounts of historical data and applying it to fault diagnosis of complex industrial processes has become a significant research area. In the context of smart factories, the development and application of data-driven fault diagnosis technology has become crucial. Among these, researching fault diagnosis for complex industrial processes with limited fault samples is a significant milestone towards the practical application of intelligent fault diagnosis technology.

[0004] However, a common challenge in practice is the low frequency of failures. Therefore, in real-world applications, it's more common to find a large number of normal samples and a very small number of faulty samples. This situation is often referred to as a long-tailed distribution in statistics. Fault diagnosis using long-tailed data requires addressing both imbalanced learning and few-shot learning. The extreme imbalance between normal and outlier data poses significant challenges to model detection; trained models may easily fail to effectively identify tail categories, leading to a sharp drop in detection rates. The main challenge in this situation lies in learning discriminative representations when outlier data is extremely scarce. Therefore, researching fault diagnosis for complex industrial processes under conditions of extremely scarce fault samples is of great significance. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problem of low accuracy of fault diagnosis in the prior art.

[0006] To address the aforementioned technical problems, this invention provides a method for diagnosing faults in long-tailed distributed data, comprising: Preprocess the raw data; The preprocessed raw data is input into a set of transformation networks trained under the supervision of deterministic contrastive learning loss to obtain multiple augmented data. The multiple augmented data are used as positive samples for supervised contrastive learning to train a pre-built autoencoder. The preprocessed raw data is input into the trained autoencoder to obtain the target representation vector; The target representation vector is input into a pre-trained classifier to obtain the fault diagnosis result.

[0007] Preferably, the training process of the transform network includes: Construct a learnable data augmentation network that includes a set of transformation networks and an autoencoder; The learnable data augmentation network is trained under supervised instruction using deterministic contrastive learning loss; Prune the trained, learnable data augmentation network, retaining the transformed network portion.

[0008] Preferably, the process further includes, before training the pre-built autoencoder: The distribution of positive and negative samples is balanced using the progressive balancing sampling technique.

[0009] Preferably, before inputting the preprocessed raw data into the trained autoencoder, the method further includes: The gradient descent algorithm is used to optimize the trained set of transformation networks and the trained autoencoder.

[0010] Preferably, the deterministic contrastive learning loss is expressed as:

[0011]

[0012]

[0013]

[0014] Where n is the number of transform networks, Let represent the cosine similarity between the i-th augmented data sample and the original data sample. Let represent the cosine similarity between the i-th augmented data sample and the j-th augmented data sample. This represents the i-th augmented data sample. Let z represent the j-th augmented data sample, and z(·) represent the feature representation of a sample in the embedding space.

[0015] Preferably, the loss function for supervised contrastive learning is:

[0016] in, To augment the sample set index, For the sample Feature representation in the embedding space For temperature coefficient, For the sample set to the sample The supplement, yes All positive sample index sets, It is the number of its corresponding elements. This represents the sample label with index p. This represents the sample label with index i.

[0017] Preferably, the classifier is trained using the cross-entropy loss function.

[0018] The present invention also provides a fault diagnosis device for long-tailed distributed data, comprising: The preprocessing module is used to preprocess the raw data; The data augmentation module is used to input preprocessed raw data into a set of transformation networks trained under the supervision of deterministic contrastive learning loss to obtain multiple augmented data. The supervised contrastive learning module is used to train a pre-built autoencoder by using the multiple augmented data as positive samples for supervised contrastive learning. The target feature acquisition module is used to input the preprocessed raw data into the trained autoencoder to obtain the target representation vector. The fault diagnosis module is used to input the target representation vector into a pre-trained classifier to obtain the fault diagnosis result.

[0019] The present invention also provides a fault diagnosis device for long-tailed distributed data, comprising: Memory, used to store computer programs; A processor is used to implement the steps of the above-described long-tailed distributed data fault diagnosis method when executing the computer program.

[0020] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described long-tailed distributed data fault diagnosis method.

[0021] The technical solution of the present invention has the following advantages compared with the prior art: The fault diagnosis method for long-tailed distributed data based on learnable data augmentation networks described in this invention first generates a set of transformations. These transformations maintain semantic information similar to the original samples while also having differences from each other. This set of transformations is used as positive samples for supervised contrastive representation learning to improve the representational power of supervised contrastive learning. The feature vector of the original sample is obtained through supervised contrastive learning and then input into a classifier for fault diagnosis. Compared with other methods, the fault diagnosis method for long-tailed distributed data of this invention has higher accuracy and better performance. Attached Figure Description

[0022] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the implementation of a long-tailed distributed data fault diagnosis method provided by the present invention. Figure 2 It is a diagram of a learnable data augmentation network structure; Figure 3 This is a schematic diagram illustrating the effect of deterministic contrastive learning loss; Figure 4 This is a schematic diagram of progressive equilibrium sampling; Figure 5 These are t-SNE visualizations for 5-shot, 10-shot, 15-shot, and 20-shot scenarios. Detailed Implementation

[0023] The core of this invention is to provide a method, apparatus, device, and computer storage medium for diagnosing faults in long-tailed distributed data, which improves diagnostic accuracy and performance.

[0024] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely 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.

[0025] Please refer to Figure 1. Figure 1 The flowchart illustrates the implementation of a long-tailed distributed data fault diagnosis method provided by this invention; the specific operation steps are as follows: S101: Preprocess the raw data; Before being fed into the neural network model, the offline collected data is cleaned and preprocessed, and normalized to... The data in the online phase should undergo the same preprocessing operations as in the offline phase to achieve model matching.

[0026] S102: Input the preprocessed raw data into a set of transformation networks trained under the supervision of deterministic contrastive learning loss to obtain multiple augmented data; Obtaining well-enhanced data from the original data to serve as positive samples for the next step of supervised contrastive learning helps improve the representational power of supervised contrastive learning.

[0027] S103: Use the multiple augmented data as positive samples for supervised contrastive learning to train the pre-constructed autoencoder; S104: Input the preprocessed raw data into the trained autoencoder to obtain the target representation vector; S105: Input the target representation vector into a pre-trained classifier to obtain the fault diagnosis result.

[0028] In this stage, a good classifier is learned through the cross-entropy loss function, which serves as the final output of the entire fault diagnosis framework.

[0029] Based on the above embodiments, this embodiment will provide a detailed description of step S102: The success of supervised contrastive learning owes much to the development of data augmentation techniques. However, traditional data augmentation methods are generally based on manual design, such as random pruning, splicing, and masking. This is not only inefficient but also difficult, and most data augmentation techniques are not suitable for sequential data, which poses challenges to using contrastive learning methods for fault diagnosis.

[0030] We propose a Learnable Augmentation Network (LAN) for sequence-type data based on deterministic contrastive loss. This process comprises two components: a series of transformation networks and an autoencoder. We jointly train these two components end-to-end, transforming each sample through the transformation networks to obtain semantic information similar to the original sample, termed "views." Details of the process are as follows: Figure 2 As shown.

[0031] Suppose we have Sample and Transformation functions For any sample The deterministic contrastive loss is expressed as follows:

[0032]

[0033]

[0034]

[0035] Where n is the number of transform networks, Let represent the cosine similarity between the i-th augmented data sample and the original data sample. Let represent the cosine similarity between the i-th augmented data sample and the j-th augmented data sample. This represents the i-th augmented data sample. Let z represent the j-th augmented data sample, and z(·) represent the feature representation of a sample in the embedding space.

[0036] Here, all transformation functions are fitted using neural networks, meaning they all have learnable parameters and can be optimized using gradient descent. The main idea of ​​deterministic contrastive learning loss is to generate a series of data that are semantically similar to the original data, but at the same time, the generated data should have as much semantic information as possible that is different from each other. Its intuitive understanding is as follows: Figure 3 We still define the original data. For anchor points.

[0037] Prune the trained, learnable data augmentation network, retaining the transformed network portion.

[0038] Based on the above embodiments, this embodiment will provide a detailed description of step S103: Self-supervised contrastive learning belongs to the category of self-supervised learning methods. Self-supervised learning attempts to extract feature representations z from inputs using large amounts of unlabeled data. The basic idea of ​​this work is as follows: An anchor point is placed in the embedding space... And a "positive sample" They cluster together and anchor the points to many "negative samples". Separate. Typically, positive samples are generated through data augmentation. The basic idea of ​​self-supervised contrastive learning is to learn how to make the representation z(x) of a sample point x in the embedding space satisfy the following relationship:

[0039] In the formula express and The distance can be Euclidean distance, Mahalanobis distance, cosine distance, etc. Suppose we have a batch of randomly sampled... Each sample and its label pair For each input sample Both can generate two augmented samples using data augmentation techniques. and Record the corresponding tag as , to obtain a by A batch consisting of sample label pairs Self-supervised contrastive learning can then be represented as:

[0040] in To augment the sample set index, For the sample The index corresponding to the positive sample. For the sample Feature representations in the embedding space are typically obtained using an autoencoder. For temperature coefficient, For the sample set to the sample The complement, i.e. For indexes Assuming it is an anchor index, then the index For the corresponding positive sample index, others The index is a negative sample index. It is worth noting that the self-supervised contrastive learning method discussed above does not utilize label information.

[0041] This invention uses supervised contrastive learning to leverage existing data labeling information for better long-tail classification of imbalanced diagnostic problems. Its main idea is to bring the representations of samples from the same class closer together, while distancing the representations of samples from different classes. The corresponding loss function is:

[0042] in yes All positive sample index sets, This represents the number of its corresponding elements. Then, gradient descent can be used to optimize the parameters of the entire network.

[0043] Based on the above embodiments, the step of training the pre-built autoencoder further includes: The progressive balancing sampling technique balances the distribution of positive and negative samples, effectively alleviating class imbalance during training and suppressing overfitting. During the network training phase, this invention uses progressive balancing sampling to balance the data distribution. Progressive balancing sampling combines the advantages of two common strategies: instance balancing sampling and class balancing sampling. These two basic sampling strategies can be represented in a unified form:

[0044] in For the first The number of samples in each class. This represents the total number of sample categories. For the first The sampling probability of a class of samples. When When it is class-balanced sampling; when This is instance balanced sampling. In the former, instance balanced sampling (IBS), each sample has an equal probability of being selected. Meanwhile, the... The probability of a sample being selected is directly proportional to its number of samples. This leads to a higher probability of samples from the majority class being selected, potentially causing the model to overfit the majority class. Class Balanced Sampling (CBS) overcomes this drawback. In CBS, the probability of a sample from each class being selected is equal during training, resulting in an equal number of samples from different classes in each batch. This ensures that samples from each class are adequately trained, preventing some classes from being ignored during training. This study employs Progressively Balanced Sampling (PBS), combining the advantages of both. In the first few iterations of Progressively Balanced Sampling, it approximates Instance Balanced Sampling, while in the last few iterations, it approximates Class Balanced Sampling. Its... Category after the second iteration The sampling probability is:

[0045] in The total number of training iterations is [number], and the working principle is as follows: Figure 4 .set up ,when This indicates balanced sampling of instances, where normal samples are sampled much more likely than faulty samples; when This indicates balanced sampling, where the sampling probability is the same for each class. As... From 0 to Progressive balanced sampling has gradually transitioned from instance balanced sampling to class balanced sampling, which means that it combines the advantages of both instance-based and class-based balanced sampling, thereby ensuring the collection rate of faulty samples and suppressing the overfitting problem.

[0046] Based on the above embodiments, this embodiment verifies the beneficial effects of the present invention through specific experiments: During the data augmentation network learning phase, we used one-dimensional convolutions, with each layer containing 3 residual blocks, and then performed data normalization using the Instance Normalization (IN) strategy. Convolutional layers were also used at the top and bottom layers of the entire network, configured with a kernel size of 3, a stride of 1, a bias of 0, and an output dimension equal to the dimension of each data point. The number of transformation networks was set to 4.

[0047] In the representation learning phase, we designed a six-layer autoencoder. Each layer of the encoder had 52, 20, and 10 neurons, while each layer of the decoder had 10, 20, and 52 neurons. We set the maximum number of iterations to 2000. After training the autoencoder, we used its encoder portion as the raw feature extractor. Then, we further trained the feature extractor using a supervised contrastive loss function, setting the maximum number of iterations to 500.

[0048] During the classifier learning phase, we employed a two-layer learnable linear classifier. The first layer had 10 neurons, while the second layer had the number of neurons corresponding to the number of fault categories. We set the number of iterations for training the linear classifier to 100 and the number of iterations for obtaining the scaling parameter to 20. The learning rate was set to 0.001 for all training processes, and the number of samples in each batch was set to 75.

[0049] The algorithm proposed in this invention was validated on the Tennessee-Eastman process dataset. The Tennessee-Eastman process is a typical chemical process, and its corresponding dataset is a standard validation dataset currently used for fault diagnosis of complex industrial processes. It is described below: The Tennessee-Eastman process was created in 1993 by Downs and Vogel, chemical engineers at Eastman Chemical Company in the United States. It is based on a real-world chemical reaction process developed by the company. This process serves as a standardized test platform for evaluating troubleshooting methods for complex industrial processes. The process involves eight components: A, B, C, D, E, F, G, and H. Reactants A, C, D, and E, along with inert gas B, are added to the reactor to produce products G and H, with F as a byproduct. The specific chemical reactions are shown below:

[0050]

[0051]

[0052]

[0053] The Tennessee-Eastman process is a typical complex chemical industrial process. It comprises multiple modules and variables, and its internal reactions are highly complex. Due to this complexity, it is widely used as a standard test platform for troubleshooting methods in complex industrial processes. Data on this process is available at http: / / depts.washington.edu / control / LARRY / TE / download.html. In specific research, corresponding Simulink models can also be downloaded according to specific task scenarios to generate the required data.

[0054] This dataset is widely used by researchers in the field of fault diagnosis for complex industrial processes. The dataset contains a total of 21 types of faults. It involves 52 variables, including temperature and pressure, whose real-time values ​​are measured by sensors. Detailed descriptions of the 11 operational variables can be found in Table 1. For the directly downloadable dataset, under normal circumstances, the training dataset contains 500 samples, and the test dataset contains 960 samples. Under fault conditions, the training dataset contains 480 samples for each fault, and the test dataset contains 800 samples.

[0055] Table 1. Operational variables of the Tennessee-Eastman process

[0056] This invention randomly selected six faults from the Tennessee-Eastman process—fault 1, fault 2, fault 4, fault 6, fault 7, and fault 8—for fault diagnosis of long-tailed distributed data. In this scenario, sufficient normal samples can be collected. However, for all other fault categories, only a very small number of samples can be collected; therefore, all fault categories are considered tail categories. Table 2 shows the fault diagnosis results in this scenario. Experiments were conducted with four scenarios: 5-shot, 10-shot, 15-shot, and 20-shot. 5-shot, 10-shot, 15-shot, and 20-shot represent that each tail fault category contains only 5, 10, 15, and 20 samples, respectively. Data augmentation strategies used included Gaussian blur, random pruning, random masking, random flipping, and augmented data generated by the learnable data augmentation network of this invention. It can be seen that the method of this invention achieves optimal performance in all cases. For example, in the case of 10-shot, the method of the present invention increases efficiency by 28.4%, 42.6%, 17.6%, and 11.8% respectively compared to other methods.

[0057] Table 2 Experimental Results (Accuracy %)

[0058] Furthermore, it can be seen that, thanks to the powerful representation learning capabilities of supervised contrastive learning, linear classifiers are sufficient to achieve the best classification performance. The t-SNE visualization results are shown below. Figure 5 As can be seen, supervised contrastive learning can produce discriminative representations between categories, which will be very helpful for fault diagnosis of long-tailed distributed data downstream.

[0059] This invention also provides a fault diagnosis device for long-tailed distributed data, the specific device including: Preprocessing module 100 is used to preprocess the raw data; The data augmentation module 200 is used to input the preprocessed raw data into a set of transformation networks trained under the supervision of deterministic contrastive learning loss to obtain multiple augmented data. The supervised contrastive learning module 300 is used to train a pre-built autoencoder by using the multiple augmented data as positive samples for supervised contrastive learning. The target feature acquisition module 400 is used to input the preprocessed raw data into the trained autoencoder to obtain the target representation vector. The fault diagnosis module 500 is used to input the target representation vector into a pre-trained classifier to obtain the fault diagnosis result.

[0060] The long-tailed distributed data fault diagnosis device of this embodiment is used to implement the aforementioned long-tailed distributed data fault diagnosis method. Therefore, the specific implementation of the long-tailed distributed data fault diagnosis device can be found in the previous embodiment section of the long-tailed distributed data fault diagnosis method. For example, the preprocessing module 100, data augmentation module 200, supervised contrastive learning module 300, target feature acquisition module 400, and fault diagnosis module 500 are respectively used to implement steps S101, S102, S103, S104, and S105 in the above-mentioned long-tailed distributed data fault diagnosis method. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0061] A specific embodiment of the present invention also provides a long-tailed distributed data fault diagnosis device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described long-tailed distributed data fault diagnosis method.

[0062] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described long-tailed distributed data fault diagnosis method.

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for fault diagnosis of long-tailed distributed data, characterized in that, include: Preprocess the raw data; The preprocessed raw data is input into a set of transformation networks trained under the supervision of deterministic contrastive learning loss to obtain multiple augmented data. The training process of the transformation network includes: Construct a learnable data augmentation network that includes a set of transformation networks and an autoencoder; The learnable data augmentation network is trained under supervised conditions using a deterministic contrastive learning loss, which is expressed as: Where n is the number of transform networks, Let represent the cosine similarity between the i-th augmented data sample and the original data sample. Let represent the cosine similarity between the i-th augmented data sample and the j-th augmented data sample. This represents the i-th augmented data sample. Let z represent the j-th augmented data sample, and z(·) represent the feature representation of a sample in the embedding space; Prune the trained, learnable data augmentation network, retaining the transformed network portion; The multiple augmented data sets are used as positive samples for supervised contrastive learning to train a pre-constructed autoencoder. The loss function for supervised contrastive learning is: in, To augment the sample set index, For the sample Feature representation in the embedding space For temperature coefficient, For the sample set to the sample The supplement, yes All positive sample index sets, It is the number of its corresponding elements. This represents the sample label with index p. This represents the sample label with index i. The process of training the pre-built autoencoder also includes: balancing the distribution of positive and negative samples using a progressive balancing sampling technique; The preprocessed raw data is input into the trained autoencoder to obtain the target representation vector; The target representation vector is input into a pre-trained classifier to obtain the fault diagnosis result.

2. The method for fault diagnosis of long-tailed distributed data according to claim 1, characterized in that, Before inputting the preprocessed raw data into the trained autoencoder, the following steps are also included: The gradient descent algorithm is used to optimize the trained set of transformation networks and the trained autoencoder.

3. The method for fault diagnosis of long-tailed distributed data according to claim 1, characterized in that, The classifier is trained using the cross-entropy loss function.

4. A fault diagnosis device for long-tailed distributed data, characterized in that, The steps for implementing the long-tailed distributed data fault diagnosis method according to any one of claims 1 to 3 include: The preprocessing module is used to preprocess the raw data; The data augmentation module is used to input preprocessed raw data into a set of transformation networks trained under the supervision of deterministic contrastive learning loss, to obtain multiple augmented data. The training process of the transformation networks includes: Construct a learnable data augmentation network that includes a set of transformation networks and an autoencoder; The learnable data augmentation network is trained under supervised conditions using a deterministic contrastive learning loss, which is expressed as: Where n is the number of transform networks, Let represent the cosine similarity between the i-th augmented data sample and the original data sample. Let represent the cosine similarity between the i-th augmented data sample and the j-th augmented data sample. This represents the i-th augmented data sample. Let z represent the j-th augmented data sample, and z(·) represent the feature representation of a sample in the embedding space; Prune the trained, learnable data augmentation network, retaining the transformed network portion; The supervised contrastive learning module is used to train a pre-built autoencoder by using the multiple augmented data as positive samples for supervised contrastive learning. The loss function for supervised contrastive learning is: in, To augment the sample set index, For the sample Feature representation in the embedding space For temperature coefficient, For the sample set to the sample The supplement, yes All positive sample index sets, It is the number of its corresponding elements. This represents the sample label with index p. This represents the sample label with index i. The process of training the pre-built autoencoder also includes: balancing the distribution of positive and negative samples using a progressive balancing sampling technique; The target feature acquisition module is used to input the preprocessed raw data into the trained autoencoder to obtain the target representation vector. The fault diagnosis module is used to input the target representation vector into a pre-trained classifier to obtain the fault diagnosis result.

5. A fault diagnosis device for long-tailed distributed data, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the long-tailed distributed data fault diagnosis method as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the long-tailed distributed data fault diagnosis method as described in any one of claims 1 to 3.

Images