Fault diagnosis method based on self-supervised single-domain generalization under unknown equipment conditions

Abstract

The present invention discloses a fault diagnosis method based on self-supervised single-domain generalization under unknown working conditions of equipment. Through a self-supervised comparative learning framework, vibration signal features are extracted, which solves the problem of insufficient generalization ability of existing diagnostic models in complex working conditions. The method includes the following steps: obtaining bearing vibration signal data. Preprocessing the vibration signal, and dividing the training set and the test set. Constructing a data enhancement library. Constructing a self-supervised learning framework to solve the instance classification problem with comparison as the goal. Performing data enhancement on the vibration signal. Using momentum comparison to train the model, obtain a feature extractor and fine-tune it. Applying the trained model to classify new vibration signals to obtain their health status types. The present invention reduces the cost of obtaining fault labels and improves the applicability of diagnostic models in complex working conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis and self-supervised learning, and specifically relates to an intelligent bearing fault diagnosis method based on a self-supervised single-domain generalization framework under unknown operating conditions of equipment. Background Art

[0002] In today's industrial production and equipment operation, fault detection is becoming increasingly important. Traditional fault detection methods usually rely on manually designed features and rules, which have problems such as insufficient generalization ability and high demand for large amounts of labeled data, limiting their application in complex systems.

[0003] At present, fault diagnosis methods based on deep learning have been widely used in the fields of industrial production and equipment operation. These methods usually automatically extract data features and perform fault detection by constructing deep neural networks, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory artificial neural networks (LSTM). These models can automatically learn complex patterns and features from a large amount of sensor data, improving the automation and detection accuracy of fault detection.

[0004] Typical deep learning fault detection methods are generally based on supervised learning frameworks and require a large amount of labeled data for training. In these methods, the network optimizes model parameters by maximizing a certain performance indicator (such as accuracy) so that it can accurately identify faults in given test data. Although this method alleviates the work of feature engineering to a certain extent, it still has several obvious disadvantages.

[0005] First, these supervised learning methods require a large amount of labeled data, which means that a lot of data needs to be manually annotated. This not only takes a lot of manpower and time, but also in many cases, due to the difficulty in obtaining labeled data, it is impossible to obtain sufficient labeled data in all industrial scenarios.

[0006] Second, although deep learning models usually perform well on training data, they often perform poorly when encountering previously unseen fault types due to their reliance on labeled data. This is because the models lack adaptability to new fault types and have difficulty generalizing to unseen situations.

[0007] Third, the performance of the model in practical applications is sometimes unsatisfactory, especially when faced with complex and changing working conditions. In this case, the model may not be able to stably identify and detect faults, resulting in insufficient robustness in practical applications.

[0008] Finally, a large amount of industrial data is usually unlabeled, which is not fully utilized in traditional methods. Unlabeled data is a valuable information resource, and failure to effectively use this data will lead to waste and reduce the potential of the model. Summary of the invention

[0009] Purpose of the invention: To solve the above problems, the present invention proposes a fault diagnosis method based on self-supervised single-domain generalization under unknown working conditions of equipment. In the experimental scenario of real production, intelligent fault detection is performed by comparing and learning the self-supervised model, so as to solve the problems of insufficient generalization ability and high demand for a large amount of labeled data in traditional fault detection methods.

[0010] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is:

[0011] The fault diagnosis method based on self-supervised single-domain generalization under unknown working conditions of equipment is characterized by comprising the following steps:

[0012] The first step is to obtain the bearing vibration signal data, preprocess the vibration signal, and divide the training set and test set;

[0013] The second step is to build a data enhancement library to randomly enhance the input original vibration signal;

[0014] The third step is to learn the representation by maximizing the similarity between two different enhancement results of the same data point and minimizing the similarity between the data point and other data points, thus building a contrastive learning self-supervised model.

[0015] The fourth step is to use the data after data enhancement in the second step as the model input to train the comparative learning self-supervised model, so that the model can extract the invariant features of the vibration signal and obtain the feature extractor;

[0016] The fifth step is to use the trained model to classify the vibration signals of the sampled equipment under the new operating conditions to obtain its health status type.

[0017] Furthermore, overlapping sampling is performed on the original vibration signal, that is, the original vibration signal is divided into data of the same length using a sliding window method to construct a training set for the network; the signal is divided into different domains according to different working conditions to obtain training sets of several different domains.

[0018] Furthermore, a data enhancement library with noise addition, quantization, envelope spectrum signal enhancement, drift, inversion and Mixup is constructed to enhance the original vibration signal input. Perform random enhancement, including:

[0019] Add random noise to the vibration signal, the formula is as follows:

[0020] ,

[0021] in, is the signal after adding random noise, is a random noise signal;

[0022] The vibration signal is quantified as follows:

[0023] ,

[0024] in, is the quantized signal, is the quantization step size. The round function converts the continuous signal Divide by the quantization step size The result is rounded to the nearest integer;

[0025] The envelope of the vibration signal is extracted using Hilbert transform to enhance the envelope spectrum signal. The formula is as follows:

[0026] ,

[0027] ,

[0028] in, is the Hilbert transform of the vibration signal, is the envelope of the signal, is the original signal In time The value at Indicates the current time point and the integration variable time point The distance between

[0029] Introduce random drift into the vibration signal to simulate the changes in the real environment. The formula is as follows:

[0030] ,

[0031] in, is the signal after adding drift, is a time-varying noise signal, is a random drift signal that varies with time;

[0032] The vibration signal data is time-reversed, and the formula is as follows:

[0033] ,

[0034] in, is the time-reversed signal, is the maximum time length of the signal;

[0035] Mixup randomly selects two original vibration signals and their corresponding fault type labels, and uses linear interpolation to generate a new signal and the corresponding fault type label to obtain the enhanced signal. The formula is as follows:

[0036] ,

[0037] in, It is the signal enhanced by Mixup. It's a signal The corresponding fault type label, and are two original signals randomly selected from the input signal, and Signal and The corresponding label, is the mixing coefficient sampled from the beta distribution, , is a parameter that controls the shape of the beta distribution.

[0038] Furthermore, we learn representations by maximizing the similarity between two different enhancement results of the same data point while minimizing the similarity between the data point and other data points, and build a contrastive learning self-supervised model framework as follows:

[0039] After different data augmentation, the same batch of samples are encoded by the momentum encoder as a feature extractor. The momentum encoder consists of two networks with shared parameters. Specifically:

[0040] Construct two networks with the same structure, respectively denoted as and ,Each network consists of three convolutional layers, each of which is followed by a maximum pooling layer, and finally connected to a fully connected layer and an output layer;

[0041] The parameters of the two networks are recorded as and , updated by:

[0042] ,

[0043] in, is the momentum coefficient, and the parameter Update via back-propagation;

[0044] Given a batch in the training set ,in is the size of each batch during training, express The samples;

[0045] Randomly select two different data enhancement methods in the data enhancement library to enhance the signal data and obtain two sets of enhanced signals and ,in and They are and The Enhanced samples;

[0046] and Through the network and The momentum encoder is constructed, and the and ,in It is Enhanced samples Via momentum encoder The encoded feature vector, It is Enhanced samples Via momentum encoder The encoded feature vector;

[0047] The contrast loss function is:

[0048] ,

[0049] in, is the number of samples in the same batch, is the temperature coefficient, and Constitute a positive sample pair, and Constitute a negative sample pair, represents the inner product;

[0050] By minimizing the loss function of formula (9), the cosine similarity between positive samples is reduced and the similarity between negative samples is increased, so that the network Distinguish positive and negative samples and update using back propagation ,according to Update again , and repeat the process until the training is completed.

[0051] Furthermore, the KL divergence between the likelihoods of two augmented batches is introduced as an invariance loss:

[0052] ,

[0053] in, represents the difference between the feature representation distributions of the same data sample under two enhancement strategies, , is Case The probability distribution of , is Case The probability distribution of

[0054] The feature representation of the data is learned by optimizing the following loss:

[0055] ,

[0056] Among them, the loss function It consists of two parts: one is the contrast loss, which is used to measure the similarity between the feature representations of the same sample under different data enhancement strategies; the other is the invariance loss, which is used to measure the difference between the feature representation distributions under two enhancement strategies. is the weight of the invariance loss.

[0057] Furthermore, during the training process, the multicrop method is used to process the data and perform multiple random crops on the input data as follows:

[0058] Two cropping windows W1 and W2 are defined. The size of window W1 is set to 70% to 100% of the original vibration signal to capture global features. The size of window W2 is set to 10% to 30% of the original vibration signal to capture local features.

[0059] Two windows are randomly applied to the original vibration signal to obtain slice data of different sizes after cropping, and the slice data are randomly enhanced using a data enhancement method in a data enhancement library;

[0060] In the training stage, the data that has undergone the second step of data augmentation and multicrop data augmentation is used as input to train the contrastive learning model, so that the momentum encoder as a feature extractor can extract the invariance features of the vibration signal.

[0061] Furthermore, the model training is divided into two steps: pre-training and adjustment. The process is as follows:

[0062] In the pre-training stage, an unlabeled vibration signal under a working condition is used, and the data enhancement library and multicrop multi-cropping constructed in the second step are used for data enhancement as the input of the comparative learning self-supervised model. Finally, a feature extractor is trained. , extract the essential features of the vibration signal;

[0063] In the adjustment phase, directly load the feature extractor trained by the self-supervised model through contrastive learning in the pre-training phase , freeze all parameters , and then add three fully connected layers and a classification head to build an end-to-end model. The parameters of the fully connected layers and the classification head are optimized using the same labeled training set as in the pre-training stage to complete the adjustment of the comparative learning self-supervised model in the pre-training stage.

[0064] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0065] The method of the present invention first uses a comparative learning method, and the model can automatically learn the key features in the data without human intervention, thereby improving the efficiency and accuracy of the fault detection system. By comparing the similarities and differences between data samples to learn the representation of data, the model can better understand the inherent structure and characteristics of the data, thereby improving the generalization ability and robustness of the model, and achieving more accurate and efficient fault identification and prediction. This method extracts vibration signal features through a self-supervised comparative learning framework, solving the problem of insufficient generalization ability of existing diagnostic models in complex working conditions. Self-supervised comparative learning and multi-cropped data enhancement effectively solve this problem, and the momentum encoder improves the stability of the model.

[0066] Through self-supervised pre-training, the model can learn the common features and structures of data from unlabeled data, so as to better distinguish different fault modes and improve the adaptability and generalization ability of the system. It can use a large amount of unlabeled data for training without manual labeling, which reduces the cost of data labeling and increases the scalability of model training. This method can also effectively handle noise and changes in data, improve the robustness and generalization ability of the model, and make the fault detection system more robust and reliable.

[0067] Through the method of the present invention, the intelligent fault detection system can better adapt to new fault types, improve the adaptability and generalization ability of the system, and provide a more reliable fault detection solution for industrial production. The application of this technology is expected to improve the reliability and operating efficiency of equipment, reduce production stoppages and maintenance costs caused by faults, and bring significant economic and efficiency improvements to industrial production. This method can also better cope with new fault types, adapt to complex and changeable working conditions, and provide a more reliable fault detection solution for industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart of the bearing fault diagnosis method of the present invention. DETAILED DESCRIPTION

[0069] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0070] The fault diagnosis method based on self-supervised single-domain generalization under unknown working conditions of the equipment described in the present invention proposes a new contrastive learning method, and uses the momentum update method and the multi-cropping method and the data enhancement library to build and train the model. Multi-cropping technology is mainly used for image data. The present invention innovatively applies the multi-cropping method to one-dimensional vibration signals, while capturing the global and local features of the signal. The data enhancement library is constructed by six data enhancement methods, including Mixup, a method applied to one-dimensional vibration signal data. The contrastive learning framework is improved, and richer contrast data is constructed in combination with the multi-cropping method. The momentum update encoder is applied to reduce training fluctuations, and the performance is more stable when dealing with variable working condition fault diagnosis problems.

[0071] The process of the present invention is as follows Figure 1 As shown, the following steps are included:

[0072] Step 1: Get the original vibration signal.

[0073] Time domain vibration signal As input, It is Type of fault The signal of each working condition contains data points, and They are the total number of categories and total number of working conditions of vibration signals respectively; is the label of the fault category.

[0074] Step 2: Build a training set.

[0075] Overlap sampling is performed on the original vibration signal, that is, the sliding window method is used to divide the original vibration signal into data of the same length to construct the network training set. In order to complete the domain generalization task, the signal is divided into different domains according to different working conditions. ,get training sets from different domains.

[0076] Step 3: Build a data augmentation library.

[0077] First, a data enhancement library with functions such as adding noise, quantization, envelope spectrum signal enhancement, drift, inversion and Mixup is designed to randomly enhance the input signal for subsequent contrastive learning.

[0078] Add random noise to the vibration signal to increase the robustness and generalization ability of the data:

[0079] ,

[0080] in, is the signal after adding random noise, is the original signal, is a random noise signal.

[0081] Quantify the vibration signal and reduce the accuracy of the data to reduce the computational complexity of the model:

[0082] ,

[0083] in, is the quantized signal, is the original signal, is the quantization step size. The round function converts the continuous signal Divide by the quantization step size The result is rounded to the nearest integer.

[0084] Envelope spectrum signal enhancement uses Hilbert transform to extract the envelope of the original signal, thereby reducing the impact of high-frequency noise and improving the feature representation and analysis of the signal:

[0085] ,

[0086] ,

[0087] in, is the original signal, is the Hilbert transform of the vibration signal, is the envelope of the signal, is the original signal In time The value at Indicates the current time point and the integration variable time point The distance between.

[0088] Introducing random drift into the original vibration signal to simulate changes in the real environment and increase the diversity of the data:

[0089] ,

[0090] in, is the signal after adding drift, is the original signal, is a time-varying noise signal, is a random drift signal that changes with time.

[0091] Time-reverse the original vibration signal data:

[0092] ,

[0093] in, is the time-reversed signal, is the original signal, is the maximum time length of the signal.

[0094] Mixup randomly selects two original vibration signals and their corresponding fault type labels, and uses linear interpolation to generate a new signal and the corresponding fault type label to obtain the enhanced signal:

[0095] ,

[0096] in, It is the new signal after Mixup enhancement. It's a signal The corresponding fault type label, and are two original signals randomly selected from the input signal, and Signal and The corresponding label, is the mixing coefficient sampled from a beta distribution, usually , is a parameter that controls the shape of the beta distribution.

[0097] Step 4: Build a contrastive learning self-supervised model framework.

[0098] The instance classification problem is solved with a contrastive objective, where representations are learned by maximizing the similarity between two different augmentation results of the same data point (positive) while minimizing the similarity between this data point and other data points (negative).

[0099] After different data enhancements, the same batch of samples are encoded by encoders. The present invention uses a momentum encoder. Compared with the traditional contrastive learning method, it consists of two networks with shared parameters. Using momentum encoding contrastive learning can effectively improve the convergence speed of the model, enhance the robustness of the model and improve the sample utilization rate. The specific method is to construct two networks with the same structure, denoted as and , each network consists of three convolutional layers, each convolutional layer is followed by a maximum pooling layer, and finally connected to a fully connected layer and an output layer. The parameters of the two networks are respectively denoted as and , updated by:

[0100] ,

[0101] in, is the momentum coefficient, and the parameter Updated via back-propagation.

[0102] Given a batch in the training set ,in is the size of each batch during training, express The Randomly select two different data enhancement methods from the data enhancement library to enhance the signal data, and obtain two sets of signals after data enhancement and , and They are and The An enhanced sample. and Through the network and The momentum encoder (also a feature extractor) is constructed to obtain and ,in It is Enhanced samples Via momentum encoder The encoded feature vector, It is Enhanced samples Via momentum encoder The encoded feature vector is then substituted into the contrast loss function:

[0103] ,

[0104] in, is the number of samples in the same batch, is the temperature coefficient, and is a pair of positive samples, and Divide The other samples of are negative samples, that is, and Constitute a positive sample pair, and Constitute a negative sample pair, represents the inner product; the numerator part Weighed and The larger the value, the more similar the two are. For all possible Normalization is performed so that the sum of all possibilities is equal to 1, which is equivalent to a distribution.

[0105] By minimizing the loss function of formula (9), the cosine similarity between positive samples can be reduced and the similarity between negative samples can be increased, making the network Able to distinguish positive and negative samples and update using back propagation ,according to Update again , and repeat the process until the training is completed.

[0106] In order to optimize the invariant prediction, the KL divergence between the likelihoods of two augmented batches is introduced as the invariance loss:

[0107] ,

[0108] in, represents the difference between the feature representation distributions of the same data sample under two enhancement strategies, , Calculated in Case The probability distribution of , Calculated in Case The invariance loss forces the feature representation to and Maintain consistent similarity structure under positive and negative samples.

[0109] Finally, the feature representation of the data is learned by optimizing the following loss:

[0110] ,

[0111] Among them, the loss function It consists of two parts: one is the contrast loss, which is used to measure the similarity between the feature representations of the same sample under different data enhancement strategies; the other is the invariance loss, which is used to measure the difference between the feature representation distributions under two enhancement strategies. is the weight of the invariance loss.

[0112] Step 5: multicrop multi-crop data enhancement.

[0113] In the training process of the fourth step, the present invention uses the multicrop method to further process the data, and increases the diversity of the data by performing multiple random crops on the input data, thereby improving the generalization ability of the model.

[0114] Two cropping windows are defined: a large window W1 and a small window W2. The size of the large window W1 is set to 70% to 100% of the original vibration signal, and the size of the small window W2 is set to 10% to 30% of the original vibration signal. The large window W1 is used to capture global features in a larger range, while the small window W2 is used to capture local features in a smaller range. This allows features of different sizes to be taken into account at the same time, thereby providing more comprehensive information for training.

[0115] The large and small windows are randomly applied to the original vibration signal to obtain slice data of different sizes after cropping. The slice data are randomly enhanced using the data enhancement method in the data enhancement library. In the pre-training stage, the data after the data enhancement in the third and fifth steps are used as input to train the comparative learning model, so that the momentum encoder as a feature extractor can extract the invariant features of the vibration signal.

[0116] Step 6: Train the model.

[0117] The model training is divided into two steps: pre-training and fine-tuning. In the pre-training stage, only unlabeled vibration signals under one working condition are used, and the third and fifth steps are used for data enhancement as the input of the comparative learning model. Finally, a feature extractor is trained. , which can extract the essential features of the vibration signal. In the fine-tuning stage, the feature extractor trained by the contrast learning model in the pre-training stage is directly loaded. , freeze all parameters , and then added three fully connected layers and a classification head to build an end-to-end simple model. The parameters of the fully connected layers and the classification head were optimized using the same labeled training set as the pre-training stage to complete the fine-tuning of the contrast learning model in the pre-training stage.

[0118] Step 7: Troubleshooting.

[0119] The vibration data of the equipment under new operating conditions is sampled and input into the network model to output its health status type.

Claims

1. A fault diagnosis method based on self-supervised single-domain generalization under unknown equipment working conditions, characterized by: The following steps are involved: The first step is to obtain the bearing vibration signal data, preprocess the vibration signal, and divide the training set and test set; The second step is to build a data enhancement library to randomly enhance the input original vibration signal; The third step is to learn the representation by maximizing the similarity between two different enhancement results of the same data point and minimizing the similarity between the data point and other data points, and build a contrastive learning self-supervised model as follows: After different data augmentation, the same batch of samples are encoded by the momentum encoder as a feature extractor. The momentum encoder consists of two networks with shared parameters. Specifically: Construct two networks with the same structure, denoted as f q and f k ,Each network consists of three convolutional layers, each of which is followed by a maximum pooling layer, and finally connected to a fully connected layer and an output layer; The parameters of the two networks are denoted as θ q and θ k , updated by: i k ←mθ k +(1-m)θ q (8), Among them, m is the momentum coefficient, and the parameter θ q Update via back-propagation; Given a batch X in the training set, n }, where n is the size of each batch during training, x i represents the i-th sample in X; Randomly select two different data enhancement methods in the data enhancement library to enhance the signal data and obtain two sets of enhanced signals and in and X q and X k The i-th enhanced sample in ; X q and X k Through the network f q and f k The momentum encoder is constructed, and the and in is the i-th enhanced sample After the momentum encoder f q The encoded feature vector, is the i-th enhanced sample After the momentum encoder f k The encoded feature vector; The contrast loss function is: Where S is the number of samples in the same batch, τ is the temperature coefficient, and Constitute a positive sample pair, and Constitute a negative sample pair, · represents the inner product; By minimizing the loss function of formula (9), the cosine similarity between positive samples is reduced and the similarity between negative samples is increased, so that the network f q Distinguish positive and negative samples and use back propagation to update θ q , according to θ q Update θ again k , repeat the process until the training is completed; The fourth step is to use the data after data enhancement in the second step as the model input to train the comparative learning self-supervised model, so that the model can extract the invariant features of the vibration signal and obtain the feature extractor; The fifth step is to use the trained model to classify the vibration signals of the sampled equipment under the new operating conditions to obtain its health status type.

2. The fault diagnosis method according to claim 1, characterized in that: The original vibration signal is overlapped and sampled, that is, the sliding window method is used to divide the original vibration signal into data of the same length to construct a training set for the network; the signal is divided into different domains according to different working conditions to obtain several training sets in different domains.

3. The fault diagnosis method according to claim 1, characterized in that: Construct a data enhancement library with noise addition, quantization, envelope spectrum signal enhancement, drift, inversion and Mixup to randomly enhance the input original vibration signal x(t), including: Add random noise to the vibration signal, the formula is as follows: x noisy (t)=x(t)+∈(t)(1), Among them, x noisy (t) is the signal after adding random noise, ∈(t) is the random noise signal; The vibration signal is quantified as follows: Among them, x quantized (t) is the quantized signal, Q is the quantization step size, and the round function rounds the result of dividing the continuous signal x(t) by the quantization step size Q to the nearest integer; The envelope of the vibration signal is extracted using Hilbert transform to enhance the envelope spectrum signal. The formula is as follows: in, is the Hilbert transform of the vibration signal, E(t) is the envelope of the signal, x(τ) is the value of the original signal x(t) at time τ, and t-τ represents the distance between the current time point t and the integral variable time point τ; Introduce random drift into the vibration signal to simulate the changes in the real environment. The formula is as follows: x drifted (t)=x(t)+noise(t)+drift(t) (5), Among them, x drifted (t) is the signal after adding drift, noise(t) is the noise signal that varies with time, and drift(t) is the random drift signal that varies with time; The vibration signal data is time-reversed, and the formula is as follows: x reversed (t)=x(T-t) (6), Among them, x reversed (t) is the time-reversed signal, and T is the maximum time length of the signal; Mixup randomly selects two original vibration signals and their corresponding fault type labels, and uses linear interpolation to generate a new signal and the corresponding fault type label to obtain the enhanced signal. The formula is as follows: in, It is the signal enhanced by Mixup. It's a signal The corresponding fault type label, x i and x j are two original signals randomly selected from the input signal, y i and j The signal x is i and x j The corresponding labels are, λ is the mixing coefficient sampled from the β distribution, λ~Beta(α,α), and α is the parameter controlling the shape of the β distribution.

4. The fault diagnosis method according to claim 1, characterized in that: The KL divergence between the likelihoods of two augmented batches is introduced as an invariance loss: in, represents the difference between the feature representation distributions of the same data sample under two enhancement strategies, is Case The probability distribution of is Case The probability distribution of The feature representation of the data is learned by optimizing the following loss: Among them, the loss function It consists of two parts: one is the contrast loss, which is used to measure the similarity between the feature representations of the same sample under different data enhancement strategies; the other is the invariance loss, which is used to measure the difference between the feature representation distributions under two enhancement strategies. β is the weight of the invariance loss.

5. The fault diagnosis method according to claim 1, characterized in that: During the training process, the multicrop method is used to process the data and perform multiple random cropping on the input data as follows: Two cropping windows W1 and W2 are defined. The size of window W1 is set to 70% to 100% of the original vibration signal to capture global features. The size of window W2 is set to 10% to 30% of the original vibration signal to capture local features. Two windows are randomly applied to the original vibration signal to obtain slice data of different sizes after cropping, and the slice data are randomly enhanced using a data enhancement method in a data enhancement library; In the training stage, the data that has undergone the second step of data augmentation and multicrop data augmentation is used as input to train the contrastive learning model, so that the momentum encoder as a feature extractor can extract the invariance features of the vibration signal.

6. The fault diagnosis method according to claim 5, characterized in that: The training of the model is divided into two steps: pre-training and adjustment. The process is as follows: In the pre-training stage, an unlabeled vibration signal under a working condition is used, and the data enhancement library and multicrop multi-cropping constructed in the second step are used for data enhancement as the input of the comparative learning self-supervised model. Finally, a feature extractor f is trained. k , extract the essential features of the vibration signal; In the adjustment phase, the feature extractor f trained by the self-supervised model through contrastive learning in the pre-training phase is directly loaded. k , freeze all parameters θ k , and then add three fully connected layers and a classification head to build an end-to-end model. The parameters of the fully connected layers and the classification head are optimized using the same labeled training set as in the pre-training stage to complete the adjustment of the comparative learning self-supervised model in the pre-training stage.

Images