Semi-supervised intelligent fault diagnosis method based on de-noising diffusion probability model and medium

Abstract

The invention discloses a de-noising diffusion probability model-based semi-supervised intelligent fault diagnosis method and a medium, and the method comprises the steps: collecting a vibration signal sample set of different fault types of mechanical equipment as a training data set which comprises marked and unmarked samples; a semi-supervised intelligent fault diagnosis model containing DDPM, distribution alignment and a mix module is constructed, the DDPM module carries out data enhancement on the training data set, the distribution alignment module predicts a pseudo label for an enhanced unmarked sample, and the mix module combines related samples to generate a new training data set; inputting the new training data set into the diagnosis model, and carrying out iterative training through a minimum comprehensive loss function until an iteration stopping condition is met, so as to obtain a trained diagnosis model; and finally, inputting the vibration signal into the trained intelligent fault diagnosis model to realize fault diagnosis of the mechanical equipment. The number of samples used for training can be effectively increased, the model training effect is improved, and the accuracy and efficiency of fault diagnosis are improved.

Description

Technical Field

[0001] The present application belongs to the technical field of fault diagnosis, and specifically relates to a semi-supervised intelligent fault diagnosis method and medium based on a denoising diffusion probability model. Background Art

[0002] With the continuous advancement of industrial technology and the rapid development of science and technology, rotating equipment is gradually developing in the direction of large-scale, high-precision and intelligent. However, this development trend also means that the reliability of equipment has become particularly important. Especially in harsh operating environments and long-term complex working scenarios, core rotating components such as bearings and gears are prone to various problems. Therefore, there is an urgent need for an efficient fault diagnosis method to perform status detection and fault diagnosis on various components of rotating equipment to ensure its stable operation.

[0003] With the continuous innovation and breakthrough development of deep learning technology, compared with traditional fault diagnosis methods, deep learning can autonomously learn valuable feature information. Fault identification through deep learning can significantly reduce the dependence on human experience, because the model can self-learn to extract data features and deeply explore the intrinsic relationship between these features and fault types, thereby providing scientific and accurate decision support for fault identification. However, the performance of deep learning-based diagnosis methods often depends on sufficient labeled samples. Semi-supervised learning (SSL) is between supervised learning and unsupervised learning. This method can improve the classification performance of the model by using only a small number of samples with category labels and a large number of samples without category labels. In recent years, more and more SSL fault diagnosis methods have been proposed to address the problem of limited number of labeled samples. Although subsequent SSL diagnosis methods have made many improvements in diagnostic technology and effectively enhanced the recognition ability, they have not deeply solved the problem of scarce sample data at the corresponding level. Due to the complexity of the operation process, these methods still face certain limitations in achieving efficient fault diagnosis. Among them, data augmentation is a key technology in SSL, which increases the diversity of the data set by generating new training samples, thereby improving the generalization ability of the model. However, they also have limitations. The training process of traditional models is usually very unstable and prone to mode collapse. In addition, the generation results are greatly affected by the input data, making it difficult to directly control the quality of the generated samples.

[0004] The diagnostic method based on the Denoising Diffusion Probabilistic Model (DDPM) uses a neural network to accurately predict noise, gradually remove noise, and generate samples that match the original data distribution. This process is relatively more stable and less prone to mode collapse. In addition, DDPM can flexibly adjust the noise level and the number of diffusion steps to control the quality and diversity of the generated samples. However, when a single Denoising Diffusion Probabilistic Model is used to simultaneously learn features of multi-category data, the generated dataset usually contains features from multiple categories. Using a single DDPM model to achieve multi-category data enhancement is a challenging task.

[0005] In complex production processes, the extreme environmental conditions and strong noise interference of industrial sites greatly reduce the quality of rotating equipment monitoring data. Coupled with the high cost of labeling, data scarcity will occur, making it very difficult to obtain sufficient and high-quality labeled data. Summary of the invention

[0006] The technical purpose of the present application is to provide a semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model to solve the technical problem that the data set generated by the current diagnosis method based on a single denoising diffusion probability model contains features of multiple categories but is not convenient for category labeling.

[0007] To achieve the above technical objectives, the present application provides a semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model, comprising:

[0008] Collect vibration signal sample sets of different fault types on mechanical equipment as training data sets; the vibration signal sample sets include labeled samples and unlabeled samples;

[0009] Construct DDPM module, distribution alignment module, mixup module and semi-supervised intelligent fault diagnosis model;

[0010] The DDPM module uses a denoising diffusion probability model to perform data enhancement on the training data set; the distribution alignment module applies a distribution alignment strategy to perform pseudo-label prediction on the unlabeled samples after data enhancement to generate effective predicted pseudo-label information; the mixup module performs a mixing operation on the labeled samples with data enhancement and the unlabeled samples with predicted pseudo-label information, and then combines the strongly enhanced unlabeled samples to obtain a new training data set;

[0011] Inputting the new training data set into the constructed semi-supervised intelligent fault diagnosis model to obtain the fault label probability, and iteratively updating the network parameter training of the semi-supervised intelligent fault diagnosis model by minimizing the comprehensive loss function until the iteration stop requirement is met, thereby obtaining a trained semi-supervised intelligent fault diagnosis model;

[0012] The vibration signal of the mechanical equipment to be diagnosed is input into the trained semi-supervised intelligent fault diagnosis model to perform fault diagnosis on the mechanical equipment.

[0013] Furthermore, the DDPM module uses a denoising diffusion probability model to perform data enhancement on the training data set, including:

[0014] In the forward process, the denoising diffusion probability model gradually adds preset Gaussian white noise to the labeled samples to obtain enhanced labeled samples, and adds strong Gaussian white noise and weak Gaussian white noise to the unlabeled samples to obtain strongly enhanced unlabeled samples and weakly enhanced unlabeled samples; in the reverse process, the Diffwave network architecture is used to obtain predicted noise based on the enhanced labeled samples, the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, and the predicted noise is gradually removed.

[0015] Furthermore, the distribution alignment module applies the distribution alignment strategy to perform pseudo-label prediction on the unlabeled samples after data enhancement to generate effective predicted pseudo-label information, including:

[0016] According to the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, the semi-supervised intelligent fault diagnosis model is used to obtain a moving average label probability distribution of the current unlabeled samples;

[0017] According to the enhanced labeled samples, the predicted label probability distribution is obtained using the semi-supervised intelligent fault diagnosis model, and each predicted label probability distribution is averaged to determine the average label probability distribution of the current batch of labeled samples;

[0018] Based on the moving average label probability distribution of the current unlabeled samples and the average label probability distribution of the current batch of labeled samples, the predicted label probability distribution of the weakly enhanced unlabeled samples is scaled using the semi-supervised intelligent fault diagnosis model;

[0019] The predicted label probability distribution after scaling of the weakly enhanced unlabeled samples is subjected to temperature sharpening to obtain the predicted pseudo-label information. Further, according to the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, the moving average label probability distribution of the current unlabeled samples is obtained using the semi-supervised intelligent fault diagnosis model, including:

[0020] According to the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, the label probability distribution predicted by each unlabeled sample is obtained by using the semi-supervised intelligent fault diagnosis model;

[0021] The predicted label probability distributions are averaged to obtain the moving average label probability distribution of the current unlabeled sample; the average of the average label probability distributions of the n+1 groups of unlabeled samples before the current position to the n+1 groups of unlabeled samples at the current position is calculated as the moving average label probability distribution of the current unlabeled sample.

[0022] Furthermore, the semi-supervised intelligent fault diagnosis model adopts a Vision Transformer model.

[0023] Furthermore, the expression for minimizing the comprehensive loss function is as follows:

[0024] L=L ml +λ α L mul +λ β L ul ;

[0025] Among them, L is the comprehensive loss, L ml is the cross entropy loss component of the labeled sample obtained after the mixup module processing, L mul is the cross entropy loss component of the unlabeled samples obtained after the mixup module, L ul To strongly enhance the cross entropy loss component of unlabeled samples, λ α Represents the weight coefficient for balancing the cross entropy loss component of unlabeled samples obtained after the mixup module, λ β Represents the weight coefficient for balancing the cross entropy loss component of strongly enhanced unlabeled samples.

[0026] Furthermore, the mixup module performs a mixing operation on the labeled samples with data enhancement and the unlabeled samples with predicted pseudo-label information, including:

[0027] Enhance the marked samples and weakly enhanced unlabeled samples Together with their label information, they are collected and then shuffled as follows:

[0028]

[0029] The Concat() function represents the concatenation operation, and the Shuffle() function represents the shuffle operation;

[0030] Determine enhanced labeled samples The weight W l and weakly enhanced unlabeled samples The weight W ul ;

[0031] The mixup method is applied to the enhanced labeled samples and the weakly enhanced unlabeled samples respectively. The specific expressions are as follows:

[0032]

[0033] Among them, X l′ is the labeled sample obtained after processing by the mixup module, X ul′ is the unlabeled sample obtained after the mixup module processing, X l is a labeled sample, x ul is an unlabeled sample.

[0034] In a second aspect, an embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a computer, the semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model provided in any possible implementation manner of the first aspect is executed.

[0035] Compared with the prior art, the semi-supervised intelligent fault diagnosis method based on the denoising diffusion probability model provided by the present application can process both labeled samples and unlabeled samples, effectively expanding the number of samples available for training, especially when the acquisition cost of labeled samples is high or the number is limited, the use of unlabeled samples can improve the model training effect. Through the data enhancement of the denoising diffusion probability model (DDPM), the model is exposed to more diverse sample features, reducing the risk of overfitting, enhancing the recognition and diagnosis capabilities of different fault types, and can make more accurate judgments even in the face of new fault manifestations that have not appeared in the training data set. The pseudo-labels generated by the distribution alignment module give effective information to the unlabeled samples, further enriching the label system of the training data, and the mixed processing of the samples by the mixup module allows the model to learn the transition relationship between different samples and labels, improving the adaptability and diagnostic accuracy of the semi-supervised intelligent fault diagnosis model to complex fault conditions. The semi-supervised intelligent fault diagnosis model obtained by iterative training of the overall method can quickly and accurately diagnose mechanical equipment faults in practical applications, which helps to timely discover and solve equipment faults, reduce equipment downtime and maintenance costs, and improve the operating reliability and production efficiency of mechanical equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic flow chart of a semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model provided in an embodiment;

[0037] Figure 2 Schematic diagram of the network structure for performing data enhancement on labeled samples and unlabeled samples using a denoising diffusion probability model (DDPM) in an embodiment. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0039] The embodiment provides a semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model, referring to Figure 1 ,include:

[0040] Step 1: Collect vibration signal sample sets of different fault types of equipment under a certain working condition as training data sets, including labeled samples (label information is known) and unlabeled samples (label information is unknown).

[0041] The labeled sample can be expressed as in represents the i-th group of labeled samples, N represents the length of each group of labeled sample data, Represents the label information of the i-th group of labeled samples, n l represents the number of groups of labeled samples. Unlabeled samples are represented by represents the unlabeled samples in group i, n ul Represents the number of groups of unlabeled samples.

[0042] Step 2: Construct DDPM module, distribution alignment module, mixup module and semi-supervised intelligent fault diagnosis model. The DDPM module uses the denoising diffusion probability model (DDPM model) to perform data enhancement on the training data set; the distribution alignment module applies the distribution alignment strategy to predict pseudo labels on the unlabeled samples after data enhancement to generate effective predicted pseudo label information; the mixup module mixes the labeled samples with data enhancement and the unlabeled samples with predicted pseudo label information, and then combines the strongly enhanced unlabeled samples to obtain a new training data set.

[0043] Step 3: Input the new training data set into the semi-supervised intelligent fault diagnosis model to obtain the fault label probability, and use the minimized comprehensive loss function to iteratively update the network parameter training of the semi-supervised intelligent fault diagnosis model until the iteration stop requirement is met, and the trained semi-supervised intelligent fault diagnosis model is obtained;

[0044] Step 4: Input the vibration signal of the mechanical equipment to be diagnosed into the trained semi-supervised intelligent fault diagnosis model to perform fault diagnosis on the mechanical equipment.

[0045] In the embodiment, the denoising diffusion probability model (DDPM model) is used to perform data enhancement on the labeled samples and the unlabeled samples. The migration network structure diagram of the knowledge distillation part implemented by DDPM is shown in FIG. Figure 2 As shown, it mainly includes two processes: forward process and reverse process.

[0046] The forward process involves gradually adding Gaussian noise to the original data until it is eventually transformed into Gaussian noise. That is, the variance is β t Gaussian noise is continuously added to the original data distribution x_0~q(x_0). This process can be expressed as:

[0047]

[0048] During the entire forward process, x can be sampled at any diffusion time step t t , the sampled x t The relationship with the original data x0 is as follows:

[0049]

[0050] Among them, α t =1-β t , ε~N(0,1). The inverse process is implemented by training a parameterized Markov chain model, in which each step relies on the noise prediction neural network to accurately predict the noise.

[0051] If the diffusion time step t is large enough, the generated data will lose its original information and become random Gaussian noise. t The forward process usually uses a predefined variation plan to fix the noise addition operation at each step to ensure the controllability and stability of the process. In the entire forward process, x can be sampled at any diffusion time step t. t .

[0052] The inverse process is achieved by training a parameterized Markov chain model, where each step relies on a noise prediction neural network (which can be a Diffwave network architecture) to accurately predict the noise. Specifically, the inverse process starts with completely random Gaussian noise data, and by gradually removing the noise, it eventually generates data samples that match the original data distribution.

[0053] The data expression of the inverse denoising process can be expressed as:

[0054]

[0055] Among them, x t-1represents the denoised data obtained in the t-1th step of the reverse process, ε θ (x t ,t) represents the noise predicted by the noise prediction neural network, z represents the standard Gaussian noise, σ t It represents the standard deviation of the prediction noise, and the formula is as follows:

[0056]

[0057] In the reverse process, Represents the noise level factor.

[0058] During the DDPM model training process, DDPM uses a noise prediction neural network to predict the mixed noise in each step. Then, by reducing the error between the expected noise and the actual noise, the generated data is ensured to be highly consistent with the real data distribution, so that the DDPM model can reconstruct the original data from the noise. In addition, in order to more efficiently process one-dimensional vibration signal data, some embodiments use the DiffWave network architecture as the noise prediction model of DDPM. The DiffWave network architecture is a noise prediction neural network based on a bidirectional dilated convolution structure, which has demonstrated strong performance in audio and signal generation.

[0059] In some embodiments, during the training process of the DDPM model, labeled samples from different categories are mixed and input into the DDPM model, so that it can learn the features of labeled samples across multiple categories. After the DDPM model training is completed, a certain degree of noise is added to the collected monitoring data. These noisy data are then used as input to the DDPM inverse process.

[0060] In order to facilitate the subsequent execution of the distribution alignment method, two levels of data enhancement operations are defined in some embodiments, namely weak data enhancement and strong data enhancement. The difference between the two lies in the level of noise added to the monitoring data. In the forward process, the denoising diffusion probability model gradually adds preset Gaussian white noise to the labeled samples to obtain enhanced labeled samples, and the denoising diffusion probability model adds strong Gaussian white noise and weak Gaussian white noise to the unlabeled samples to obtain strongly enhanced unlabeled samples and weakly enhanced unlabeled samples; in the reverse process, the Diffwave network architecture is used to obtain the predicted noise based on the enhanced labeled samples, strongly enhanced unlabeled samples and weakly enhanced unlabeled samples, and the predicted noise is gradually removed.

[0061] In some embodiments, 1 dB to 5 dB may be determined as strong Gaussian white noise, and 10 dB to 20 dB may be determined as weak Gaussian white noise.

[0062] As an example, in weak data enhancement, a small degree of noise is introduced to the monitoring data, that is, 20dB of Gaussian white noise is added, which can be expressed as In strong enhancement, a higher degree of noise is added to the monitoring data, which can be expressed as Where m=1, 2, 3, 4 means adding 1dB, 2dB, 3dB and 4dB Gaussian white noise respectively.

[0063] This application uses a denoising diffusion probability model to input labeled samples of various fault types into it. The model can adaptively learn the characteristic distribution of all fault types, which can be completed using a single model. Then, the monitoring data with Gaussian white noise added is used as the input of the inverse process of DDPM to generate enhanced data, thereby effectively expanding the dataset of labeled and unlabeled samples.

[0064] Combination Figure 1 In some embodiments, the specific steps of the distribution alignment strategy process are as follows:

[0065] The first step is to calculate the moving average label probability distribution of the current unlabeled samples. In the training process of the semi-supervised intelligent fault diagnosis model, the strongly enhanced unlabeled samples that have undergone strong data enhancement are (where m = 1, 2, 3, 4) and weakly enhanced unlabeled samples after weak data enhancement Input into the semi-supervised intelligent fault diagnosis model to obtain the predicted label probability distribution and The semi-supervised intelligent fault diagnosis model can use the Vision Transformer (ViT) model, which is a deep learning model based on the Transformer architecture. It successfully extends the Transformer model from the field of natural language processing (NLP) to the field of computer vision (CV). Based on these, the predicted label distribution results of multiple strong enhanced data and the predicted label distribution results of weak enhanced data for each unlabeled sample are averaged to obtain the average label probability distribution of the current unlabeled sample, that is, Then, calculate the average of the average label probability distribution of the n+1 groups of unlabeled samples from the current position to the current position, that is, As the moving average label probability distribution of the current unlabeled samples.

[0066] The second step is to calculate the average label probability distribution of the current batch of labeled samples. The enhanced labeled samples of the current batch are input into the semi-supervised intelligent fault diagnosis model to obtain the predicted label probability distribution, that is, By averaging these predicted label probability distributions, we get the average label probability distribution of the current batch of labeled samples, that is, Where i = 1, 2,…, B).

[0067] In the third step, based on the moving average label probability distribution of the current unlabeled samples and the average label probability distribution of the current batch of labeled samples, the predicted label probability distribution of the semi-supervised intelligent fault diagnosis model for the weakly enhanced unlabeled samples is scaled to

[0068] The fourth step is to perform temperature (T) sharpening on the scaled predicted label probability distribution of the weakly enhanced unlabeled samples to highlight the probability of labels predicted to belong to a specific category, thereby obtaining the predicted label information of the unlabeled samples. The expression is as follows:

[0069]

[0070] in, represents the expanded form of the scaled predicted label probability distribution of weakly enhanced unlabeled samples, K represents the total number of categories, Represents the expanded form of the predicted label probability distribution after temperature sharpening. and The relationship between them is as follows:

[0071]

[0072] The distribution alignment method used in this application is used to accurately predict the label information of unlabeled samples. The goal of this method is to achieve distribution alignment by adjusting the predicted label probability distribution of unlabeled samples to make it consistent with the true label probability distribution of labeled samples.

[0073] In the embodiment, the labeled samples with labels and the unlabeled samples with predicted labels are input into the mixup module for mixing operation. The specific steps of the mixing operation based on mixup are as follows:

[0074] Mixup generates new training samples by fusing the features and labels of two or more samples through linear interpolation. This operation can expand the diversity of the data set and reduce the risk of overfitting. To explain the mixup process, two sets of input samples and their corresponding label information are given, namely (x1, p1) and (x2, p2). The mathematical formula is as follows:

[0075] λ~Beta(a,a) (7)

[0076] λ′=Max(λ,1-λ) (8)

[0077] x′=λ′x1+(1-λ′)x2 (9)

[0078] p′=λ′p1+(1-λ′)p2 (10)

[0079] Where α is a hyperparameter, which ensures that the generated sample x′ is closer to the input sample x1 than the input sample x2, while ensuring that the generated sample label p ′ Consistent with the input sample label p1.

[0080] In order to make better use of existing data resources, the method proposed in this application uses the mixup technology to fully process labeled and unlabeled samples to achieve the task of data enrichment. Using the mixed-up processed data to train the model helps improve the generalization ability of the model and smooth the decision boundary. The specific steps are as follows:

[0081] First, the enhanced labeled samples and unlabeled samples Together with their label information, they are collected and then shuffled as follows:

[0082]

[0083] Then, the weight W is divided into two parts, namely W l and W ul , while ensuring that the number of data sets contained in each part is the same as the number of labeled samples and unlabeled samples The number of data sets in .

[0084] Finally, the mixup method is applied to the labeled samples and unlabeled samples respectively. The specific expressions are as follows:

[0085]

[0086] This application uses Mixup technology to enrich data, further improving the generalization and robustness of the model.

[0087] In some embodiments, the network parameters of the semi-supervised intelligent fault diagnosis model are updated by minimizing the comprehensive loss function, and a loss function corresponding to the labeled and unlabeled samples after mixup processing and the enhanced unlabeled samples is constructed. The specific steps of updating the network parameters are as follows:

[0088] The loss function of this method is mainly composed of three parts, all of which use cross entropy as the basic structure. These three parts are specially designed to constrain the labeled samples processed by mixup, the unlabeled samples processed by mixup, and the strongly enhanced unlabeled samples and their corresponding label information. The specific forms of these three parts are as follows:

[0089] For the labeled samples processed by mixup, cross entropy is used as the loss function to optimize the model parameters. The expression of this part is:

[0090]

[0091] in, Represents the labeled samples after mixup processing.

[0092] For unlabeled samples processed by mixup, cross entropy is also used as the loss function to optimize the model parameters. The expression of this part is:

[0093]

[0094] in Represents unlabeled samples processed by mixup.

[0095] For strongly enhanced unlabeled samples, combined with their predicted label information, the following optimization function is obtained:

[0096]

[0097] Finally, the comprehensive loss function of the semi-supervised intelligent fault diagnosis model in the method proposed in this application is as follows:

[0098] L=L ml +λ α L mul +λ β L ul (17)

[0099] In the formula, λ α and λ β represents the weight coefficient that balances these loss components. After error and experimentation, as an example, λ α and λ β The values ​​are set to 1.5 and 0.5 respectively.

[0100] The semi-supervised intelligent fault diagnosis method based on the denoising diffusion probability model provided in this embodiment first uses the denoising diffusion probability model to perform data enhancement on the training data set, wherein the monitoring data with strong background noise is progressively denoised to generate high-quality enhanced monitoring data; then, a distribution alignment strategy is applied to expand the predicted label probability distribution of unlabeled samples to align it with the distribution of labeled samples to effectively generate pseudo labels; in addition, the Mixup technology is introduced for data enrichment, which further improves the generalization ability and robustness of the model.

[0101] The semi-supervised learning fault diagnosis method based on the denoising diffusion probability model provided in the present application can be used for rotating equipment fault diagnosis tasks to improve fault diagnosis performance by utilizing existing data resources.

[0102] During the entire training process, the DDPM model update times are set to 200 times, and the batch size of the experimental bearing fault dataset is 20. The back propagation algorithm is used to update the parameters of the network, and the Adam optimizer is used to minimize the objective function. The network is implemented on a computer with an Intel Core i9-12900k CPU processor, clocked at 3.60GHz, equipped with 32GB of RAM, NVIDIAGeForce RTX 3060GPU and NVIDIA GeForce RTX 4090GPU, and the programming adopts the PyTorch 1.7.1 deep learning framework.

[0103] In order to verify the effectiveness of the method proposed in this application, 6 comparison methods were constructed. These comparison methods were subjected to 6 experiments in the following analysis. The details of these 6 comparison methods are as follows:

[0104] Comparative method 1: Both this method and the proposed method are based on the principle of consistency regularization. MixMatch predicts low entropy labels for unlabeled samples that have undergone data enhancement, and applies mixup technology to mix labeled and unlabeled samples, significantly reducing the need for labeled samples. It is worth noting that MixMatch was originally designed for two-dimensional image data, while the applicant's research focus is on one-dimensional vibration signal data. Therefore, in order to compare with the proposed method, the data enhancement technology based on DDPM and the network model of Vision Transformer (ViT) (implementing a semi-supervised intelligent fault diagnosis model) proposed in this application are used to replace the data enhancement and network model components in the comparative method.

[0105] Comparative Method 2: Both this method and the proposed method are built on the principle of consistency regularization. UDA enhances the performance of the model in SSL tasks by applying data augmentation techniques in supervised learning to unlabeled sample data. It is worth noting that UDA was also originally designed for two-dimensional image data. In order to compare with the proposed method, the same method as in Comparative Method 1 was adopted, that is, the data augmentation and network model components in the comparative method were replaced with the strong data augmentation technology based on DDPM and the VisionTransformer network model (implementing a semi-supervised intelligent fault diagnosis model) proposed in this application.

[0106] Comparative method 3: This method replaces the ViT network in the proposed method with CNN, while the other parts remain the same as the proposed method. By using this method, it can be verified that the ViT network has better global feature extraction capabilities compared to CNN.

[0107] Comparative method 4: This method only uses the DDPM-based data enhancement technology to process the labeled samples. It verifies the effectiveness of the proposed DDPM-based data enhancement method by updating the model parameters using these enhanced labeled samples.

[0108] Comparative method 5: This method specifically uses the DDPM-based data weak enhancement technology to process labeled samples. It verifies the effectiveness of the proposed DDPM-based data weak enhancement method by using these weakly enhanced labeled samples to update the model parameters.

[0109] Comparison method 6: This method only uses labeled samples to update model parameters and serves as the baseline for other comparison methods.

[0110] Table 1 shows the diagnosis results of the method proposed in this application and other comparative methods on the bearing dataset, and the following conclusions can be drawn:

[0111] 1) The method proposed in this application, comparative method 1 and comparative method 2 all formulate corresponding strategies for labeled and unlabeled samples, so the diagnostic results of these three methods are significantly improved compared with comparative method 4, comparative method 5 and comparative method 6.

[0112] 2) Compared with the method proposed in this application and comparative methods 1 and 2, the proposed method uses distribution alignment technology to predict labels for unlabeled samples, and integrates weak data enhancement strategy and strong data enhancement strategy based on diffusion model in data enhancement. As shown in the diagnosis results of Table 1, the proposed method achieves the highest average fault diagnosis accuracy and the smallest result fluctuation, which verifies that its diagnostic performance is significantly better than comparative methods 1 and 2.

[0113] 3) When comparing the method proposed in the embodiment of the present application (the semi-supervised intelligent fault diagnosis model adopts the ViT network model) with the comparative method 3, due to the self-attention mechanism of the ViT network model, it can deeply mine the global features in the monitoring data, thereby obtaining highly discriminative deep features. Compared with the ViT network model, CNN has difficulties in capturing feature correlations between different data segments due to the limitations of its local receptive field. In CNN, reducing the feature dimension by setting a large step size and using a maximum pooling layer may lead to the loss of key deep feature information and affect the distinguishability of the features. As shown in the diagnostic results in Table 1, the proposed method achieves a higher average diagnostic accuracy, with less result fluctuations and more stable diagnostic performance, which further verifies the effectiveness of the proposed method compared to the method using CNN.

[0114] 4) When comparing the diagnosis results of comparison method 4 and comparison method 5, since strong data augmentation can capture more data changes and further increase data diversity compared to weak data augmentation, it can be clearly seen from the results in Table 1 that updating the model parameters using strongly enhanced labeled samples significantly improves the accuracy of the diagnosis results, and the results fluctuate less.

[0115] 5) When the diagnostic results of comparison methods 4 and 5 are compared with those of comparison method 6, using only labeled samples to update model parameters often leads to lower diagnostic performance due to insufficient sample diversity. As shown in the diagnostic results in Table 1, updating model parameters using labeled samples processed with weak and strong data augmentation based on the diffusion model significantly improves the accuracy of the diagnostic results, demonstrating the effectiveness of the proposed data augmentation strategy based on the diffusion model.

[0116] Table 1 Diagnosis results of the method proposed in this application and the comparative method on the bearing dataset

[0117]

[0118] The embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a computer, the semi-supervised intelligent fault diagnosis method based on the denoising diffusion probability model provided in the above embodiment is executed.

[0119] The above is a detailed introduction to the semi-supervised intelligent fault diagnosis method and storage medium based on the denoising diffusion probability model provided by the present application. Specific examples are used in the present application to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the concept of the present application and should not be understood as limiting the scope of protection of the present application.

Claims

1. A semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model, characterized in that: include: Collect vibration signal sample sets of different fault types on mechanical equipment as training data sets; the vibration signal sample sets include labeled samples and unlabeled samples; Construct a DDPM module, a distribution alignment module, a mixup module, and a semi-supervised intelligent fault diagnosis model; perform data enhancement on the training data set using a denoising diffusion probability model through the DDPM module; perform pseudo-label prediction on the unlabeled samples after data enhancement using a distribution alignment strategy through the distribution alignment module to generate effective predicted pseudo-label information; perform a mixing operation on the labeled samples with data enhancement and the unlabeled samples with predicted pseudo-label information through the mixup module, and then obtain a new training data set by combining the strongly enhanced unlabeled samples; Inputting the new training data set into the semi-supervised intelligent fault diagnosis model to obtain the fault label probability, and iteratively updating the network parameter training of the semi-supervised intelligent fault diagnosis model by minimizing the comprehensive loss function until the iteration stop requirement is met, thereby obtaining a trained semi-supervised intelligent fault diagnosis model; The vibration signal of the mechanical equipment to be diagnosed is input into the trained semi-supervised intelligent fault diagnosis model to perform fault diagnosis on the mechanical equipment.

2. The semi-supervised intelligent fault diagnosis method based on denoising diffusion probability model according to claim 1 is characterized in that: Performing data enhancement on the training data set by using the DDPM module using a denoising diffusion probability model includes: In the forward process, the denoising diffusion probability model gradually adds preset Gaussian white noise to the labeled samples to obtain enhanced labeled samples, and adds strong Gaussian white noise and weak Gaussian white noise to the unlabeled samples to obtain strongly enhanced unlabeled samples and weakly enhanced unlabeled samples; In the reverse process, the prediction noise is obtained based on the enhanced labeled samples, the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples using the Diffwave network architecture, and the prediction noise is gradually removed.

3. The semi-supervised intelligent fault diagnosis method based on denoising diffusion probability model according to claim 2 is characterized in that: The distribution alignment module applies the distribution alignment strategy to perform pseudo label prediction on the unlabeled samples after data enhancement to generate effective predicted pseudo label information, including: According to the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, the semi-supervised intelligent fault diagnosis model is used to obtain a moving average label probability distribution of the current unlabeled samples; According to the enhanced labeled samples, the predicted label probability distribution is obtained using the semi-supervised intelligent fault diagnosis model, and each predicted label probability distribution is averaged to determine the average label probability distribution of the current batch of labeled samples; Based on the moving average label probability distribution of the current unlabeled samples and the average label probability distribution of the current batch of labeled samples, the predicted label probability distribution of the weakly enhanced unlabeled samples is scaled using the semi-supervised intelligent fault diagnosis model; Temperature sharpening is performed on the predicted label probability distribution after scaling of the weakly enhanced unlabeled sample to obtain the predicted pseudo-label information.

4. The semi-supervised intelligent fault diagnosis method based on denoising diffusion probability model according to claim 3 is characterized in that: According to the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, the moving average label probability distribution of the current unlabeled samples is obtained by using the semi-supervised intelligent fault diagnosis model, including: According to the strongly enhanced unlabeled samples and the weakly enhanced unlabeled samples, the label probability distribution predicted by each unlabeled sample is obtained by using the semi-supervised intelligent fault diagnosis model; The predicted label probability distributions are averaged to obtain the moving average label probability distribution of the current unlabeled sample; the average of the average label probability distributions of the n+1 groups of unlabeled samples before the current position to the n+1 groups of unlabeled samples at the current position is calculated as the moving average label probability distribution of the current unlabeled sample.

5. The semi-supervised intelligent fault diagnosis method based on denoising diffusion probability model according to claim 3 is characterized in that: The semi-supervised intelligent fault diagnosis model adopts the Vision Transformer model.

6. The semi-supervised intelligent fault diagnosis method based on denoising diffusion probability model according to claim 2 is characterized in that: The expression of minimizing the comprehensive loss function is as follows: L=L ml +λ α L mul +λ β L ul ; Among them, L is the comprehensive loss, L ml is the cross entropy loss component of the labeled samples obtained after the mixup module processing, L mul is the cross entropy loss component of the unlabeled samples obtained after the mixup module processing, L ul To strongly enhance the cross entropy loss component of unlabeled samples, λ α Represents the weight coefficient for balancing the cross entropy loss component of unlabeled samples obtained after the mixup module, λ β Represents the weight coefficient for balancing the cross entropy loss component of strongly enhanced unlabeled samples.

7. The semi-supervised intelligent fault diagnosis method based on denoising diffusion probability model according to claim 2 is characterized in that: The mixup module performs a mixing operation on the labeled samples with data enhancement and the unlabeled samples with predicted pseudo-label information, including: Enhance the marked samples and weakly enhanced unlabeled samples Together with their label information, they are collected and then shuffled as follows: The Concat() function represents the concatenation operation, and the Shuffle() function represents the shuffle operation; Determine enhanced labeled samples The weight W l and weakly enhanced unlabeled samples The weight W ul ; The mixup method is applied to the enhanced labeled samples and the weakly enhanced unlabeled samples respectively. The specific expressions are as follows: Among them, X l′ is the labeled sample obtained after processing by the mixup module, X ul′ is the unlabeled sample obtained after the mixup module processing, X l is a labeled sample, x ul is an unlabeled sample.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is run by a computer, the semi-supervised intelligent fault diagnosis method based on a denoising diffusion probability model as described in any one of claims 1 to 7 is executed.

Images