Bearing fault diagnosis method and device based on deep multi-view element learning and medium

Abstract

The invention discloses a bearing fault diagnosis method and device based on deep multi-view element learning and a medium, and relates to the technical field of bearing fault diagnosis, and the method comprises the steps: carrying out the pre-training of a feature extractor through training data; inputting the label data into a pre-trained feature extractor to generate an original prototype; inputting the label data in the support set into the pre-trained feature extractor to generate a vector median; generating a clustering center according to the label-free data; updating the original prototype according to the vector median and the clustering center to generate a reference prototype; calculating cross entropy loss based on the tag data according to the tag data in the query set, and optimizing the feature extractor according to the cross entropy loss; inputting to-be-diagnosed data into the optimized feature extractor to generate a target prototype; and comparing the target prototype with the reference prototype to generate a fault diagnosis result. According to the method, small label data and sufficient label-free data are adopted to carry out fine adjustment on the model, and the model performance can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method, device and medium based on deep multi-view meta-learning. Background Art

[0002] Rotating bearings are crucial in industrial production. Their long-term high-intensity operation is prone to cause failures, affecting safety and efficiency. With technological advances, bearing fault diagnosis technology has developed rapidly, including vibration analysis and machine learning, which effectively improves diagnostic accuracy. Research on bearing fault diagnosis is of great significance for preventing accidents, optimizing maintenance strategies and promoting production technology progress.

[0003] With the rapid development of computer and sensor technology, data-driven deep learning bearing fault diagnosis methods have gradually replaced traditional machine learning and shallow learning methods based on manual feature extraction, and have been widely used in various fields. In some of the latest studies, deep learning-based methods have achieved excellent performance in diagnostic accuracy and noise resistance. However, in the field of rolling bearing fault diagnosis, these deep learning models all rely on large-scale labeled data sets for training. In actual industrial production environments, due to the complexity of operating conditions and potential high risks, it is extremely difficult to collect fault samples that can be used for model training. Therefore, it is very necessary to study a model that can obtain representative knowledge from a small number of training samples but does not overfit to existing samples.

[0004] In order to solve this kind of problem encountered in deep learning, few-shot learning has been reported in recent studies. The purpose of few-shot learning is to learn from a small number of labeled samples to automatically classify a large number of unlabeled samples. In recent years, there have been many explorations of few-shot learning methods, including twin networks, graph neural networks, relational networks, meta-learning, etc.

[0005] However, existing small sample learning methods often only focus on the contribution of the features of the original signal from a single perspective to solving the small sample problem, and ignore the study of the features of the original signal from multiple perspectives on the small sample problem, and still cannot meet the needs of users. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a bearing fault diagnosis method, equipment and medium based on deep multi-perspective meta-learning, which can use smaller labeled data and sufficient unlabeled data to fine-tune the model, and can effectively improve the model performance.

[0007] In order to solve the above technical problems, the present invention provides a bearing fault diagnosis method based on deep multi-view meta-learning, comprising: pre-training a feature extractor using training data, the training data including labeled data and unlabeled data, the labeled data including a support set and a query set; inputting the labeled data into the pre-trained feature extractor to generate an original prototype; inputting the labeled data in the support set into the pre-trained feature extractor to generate a vector median based on the labeled data; generating a cluster center based on the unlabeled data according to the unlabeled data; updating the original prototype according to the vector median and the cluster center to generate a benchmark prototype; calculating the cross entropy loss based on the labeled data according to the labeled data in the query set, and optimizing the feature extractor according to the cross entropy loss; inputting the data to be diagnosed into the optimized feature extractor to generate a target prototype; comparing the target prototype with the benchmark prototype to generate a fault diagnosis result.

[0008] As an improvement of the above scheme, the step of pre-training the feature extractor using training data includes: obtaining training data, and inputting the training data into a time domain feature extractor to generate an initial time domain feature vector; converting the training data into initial frequency domain data, and inputting the initial frequency domain data into a frequency domain feature extractor to generate an initial frequency domain feature vector; splicing and fusing the initial time domain feature vector and the initial frequency domain feature vector to generate an initial multi-view fusion feature; calculating the reconstruction loss and the regularization loss respectively according to the initial multi-view fusion feature; and updating the network parameters of the time domain feature extractor and the frequency domain feature extractor according to the reconstruction loss and the regularization term loss.

[0009] As an improvement of the above scheme, the step of inputting label data into a pre-trained feature extractor to generate an original prototype includes: obtaining label data, and inputting the label data into a pre-trained time domain feature extractor to generate a first time domain feature vector; converting the label data into first frequency domain data, and inputting the first frequency domain data into a pre-trained frequency domain feature extractor to generate a first frequency domain feature vector; splicing and fusing the first time domain feature vector and the first frequency domain feature vector to generate a first multi-view fusion feature; and generating the original prototype based on the first multi-view fusion feature.

[0010] As an improvement of the above scheme, the step of inputting the label data in the support set into the pre-trained feature extractor to generate a vector median based on the label data includes: obtaining the label data in the support set, and inputting the label data in the support set into the pre-trained time domain feature extractor to generate a second time domain feature vector; converting the label data in the support set into second frequency domain data, and inputting the second frequency domain data into the pre-trained frequency domain feature extractor to generate a second frequency domain feature vector; splicing and fusing the second time domain feature vector and the second frequency domain feature vector to generate a second multi-view fusion feature; and generating a vector median based on the label data according to the second multi-view fusion feature.

[0011] As an improvement of the above scheme, the step of generating cluster centers based on the unlabeled data according to the unlabeled data includes: using the original prototype as the initial cluster center; and using the K-Means algorithm to generate cluster centers based on the unlabeled data.

[0012] As an improvement of the above scheme, the step of updating the original prototype according to the vector median and the cluster center to generate a benchmark prototype includes: calculating the mean of the vector median and the original prototype to fine-tune the original prototype; calculating the mean of the cluster center and the fine-tuned original prototype to refine the fine-tuned original prototype; and using the refined original prototype as the benchmark prototype.

[0013] As an improvement of the above scheme, the step of updating the original prototype according to the vector median and the cluster center to generate a reference prototype includes: according to the formula , calculate the benchmark prototype; where, For the Class The benchmark prototype of the dimensions, For the Class The original prototype of the dimension, For the Class The median of the vector of dimensions, For the Class The cluster centers of the dimensions, is the adjustment rate.

[0014] As an improvement of the above scheme, the steps of calculating the cross entropy loss based on the label data according to the label data in the query set and optimizing the feature extractor according to the cross entropy loss include: obtaining the label data in the query set, and inputting the label data in the query set into the pre-trained time domain feature extractor to generate a third time domain feature vector; converting the label data in the query set into third frequency domain data, and inputting the third frequency domain data into the pre-trained frequency domain feature extractor to generate a third frequency domain feature vector; splicing and fusing the third time domain feature vector and the third frequency domain feature vector to generate a third multi-view fusion feature; calculating a soft assignment based on the label data according to the third multi-view fusion feature; calculating the cross entropy loss based on the label data according to the soft assignment, the third multi-view fusion feature and the true label of the label data; and optimizing the network parameters of the time domain feature extractor and the frequency domain feature extractor according to the cross entropy loss.

[0015] Correspondingly, the present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the above-mentioned bearing fault diagnosis method based on deep multi-perspective meta-learning when executing the computer program.

[0016] Correspondingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the above-mentioned bearing fault diagnosis method based on deep multi-view meta-learning are implemented.

[0017] The implementation of the present invention has the following beneficial effects: The bearing fault diagnosis method based on deep multi-view meta-learning of the present invention can make full use of the multi-view features of the data and unlabeled data, solve the small sample problem and make full use of the large amount of unlabeled data in the actual production environment, thereby improving the diagnosis accuracy and convergence speed of the model under the condition of sparse labeled samples. Specifically: The present invention proposes a multi-view deep feature extractor, which is composed of two variational encoders with the same structure connected in parallel, and can respectively extract time domain and frequency domain features from the original signal, improve the model accuracy, have better classification effect for boundary samples, and have stronger anti-noise performance; The present invention integrates a self-attention mechanism on the feature extractor to form a feature-layer self-attention mechanism variational encoder. The feature-layer self-attention mechanism is beneficial to minimizing the intra-class compactness (locality) to maximize the inter-class separability (non-locality), thereby improving the convergence speed and noise resistance of the model, and achieving the effect of improving the model performance. The present invention proposes a prototype adjustment strategy based on semi-supervised multi-view meta-learning. The class prototype is first constructed using the multi-view features of labeled data, and then the clustering centers of the multi-view features of unlabeled data are iterated on the class prototype. The clustering centers are less affected by label noise, have higher accuracy, and converge faster. When adjusting the prototype, the dependence on labeled data can be reduced, the features of unlabeled data can be learned more effectively, and the risk of overfitting of sparse samples can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of a first embodiment of a bearing fault diagnosis method based on deep multi-view meta-learning of the present invention; Figure 2 It is a schematic diagram of the framework of the feature extractor in the present invention; Figure 3 It is a schematic diagram of the accuracy of the ablation experiment test in the present invention; Figure 4 It is a schematic diagram of the accuracy of testing different noise combinations on the CWRU dataset in the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] See also Figure 1 , Figure 1 The flowchart of the first embodiment of the bearing fault diagnosis method based on deep multi-view meta-learning of the present invention is shown, which includes: S101, pre-training a feature extractor using training data; It should be noted that the training data includes labeled data and unlabeled data; Accordingly, the labeled data can be divided into a support set and a query set in the form of "N way K shot"; that is, N categories are randomly sampled in the labeled data, and each category is guaranteed to contain K samples to form a support set; at the same time, Q samples are randomly sampled from the remaining labeled data to form a query set; in addition, all unlabeled data can also be divided into part of the support set.

[0021] Different from the prior art, the feature extractor in the present invention includes an encoder module (i.e., a variational encoder), a self-attention module, and a reparameterization module.

[0022] VAE (Variational Encoder) is a generative model that combines the idea of ​​autoencoder and probabilistic inference, which is used to learn the potential representation of data and generate new data. Compared with AE (autoencoder), VAE inputs the specific distribution of latent variables in the decoder part instead of data points. At the same time, the present invention introduces a regularization term to make the distribution generated by the encoder close to the normal distribution, the covariance matrix is ​​close to the unit matrix, and the mean is 0. The principle of VAE and the improvement of the present invention to it will be introduced below: First, according to amortized variational inference, the encoder is equivalent to an approximate distribution , the decoder is equivalent to the likelihood distribution , from which we can construct an ELBO (evidence lower bound):

[0023] in, represents the reconstruction error, Represents an approximate distribution and the latent variable prior distribution The KL divergence of is set to a unit Gaussian distribution, i.e. .because Part of it needs to be based on its mean and variance in the latent variable z Random sampling is performed in this process, which cannot be back-propagated, so reparameterization is introduced:

[0024] This means that any Gaussian distribution can be expressed as the sum of the mean and standard deviation multiplied by a standard Gaussian distribution. In summary, the loss function of VAE can be obtained from ELBO:

[0025] in, L 1 is the reconstruction loss, L 2 is the regularization loss, and is a hyperparameter for adjusting the loss weight.

[0026] like Figure 2 As shown, based on VAE, the present invention introduces a self-attention mechanism in the feature layer to improve the sampling quality of latent variables and help the model better learn the structure and pattern in the data, so that the model can dynamically adjust the generation position of the latent variables according to the information obtained in training, thereby improving the efficiency and accuracy of model training.

[0027] Specifically, the description of the feature extractor is shown in Table 1 below: Table 1

[0028] in, K Indicates the minimum batch size; Relu The activation function is expressed as:

[0029] Softmax The activation function is expressed as:

[0030] The self-attention mechanism needs to be constructed first , , 3 vectors, Q and K The full connection of the input data is obtained:

[0031] in: The feature vector representing the input; Represents the fully connected layer function; That is, layer normalization, represents the normalization layer function; represents the random dropout neuron function, which aims to reduce the risk of overfitting.

[0032] Then, Q and K After the dot product, divide by scaled The value (i.e. the square root of the hidden layer dimension) is used to obtain the energy value, as shown below:

[0033] in, Indicates the size of the input dimension.

[0034] Finally, Energy enter Softmax The function projects it into the interval [0, 1] to get the attention score, and then compares the attention score with V Multiply and then K Add them together to get the weighted value output.

[0035]

[0036] Further, the step of pre-training the feature extractor using the training data includes: (1) Obtain training data and input the training data into the time domain feature extractor to generate an initial time domain feature vector; It should be noted that the training data (labeled data and unlabeled data) in the present invention is generally bearing vibration information.

[0037] (2) converting the training data into initial frequency domain data, and inputting the initial frequency domain data into a frequency domain feature extractor to generate an initial frequency domain feature vector; It should be noted that the structure of the time domain feature extractor and the frequency domain feature extractor in the present invention is the same.

[0038] In addition to the bearing vibration information In addition to inputting the time domain feature extractor to generate the initial time domain feature vector, the bearing vibration information Perform FFT transformation to obtain the amplitude tensor in the spectrum ; Then, the magnitude tensor The corresponding frequency tensor By performing point multiplication, symmetrical frequency domain information can be generated Y , as another view:

[0039] in: represents the frequency tensor; Represents the magnitude tensor after FFT transformation.

[0040] (3) Concatenate and fuse the initial time domain feature vector and the initial frequency domain feature vector to generate the initial multi-view fusion feature; Therefore, the present invention introduces multi-view feature fusion technology to extract differentiated feature representations under the same feature extraction framework for different viewing angles of the same data set.

[0041] (4) Calculate the reconstruction loss and regularization loss respectively based on the initial multi-view fusion features; Accordingly, according to the formula Calculate the reconstruction loss L 1 , and according to the formula Calculate the regularization loss L 2 .

[0042] The purpose of the above formula is to convert the input variable and the model output mean The squared difference between represents the negative log-likelihood, i.e., the reconstruction loss.

[0043] in: Represented as a given latent variable z , reconstruct the input variables The logarithm of the generation probability of ; Indicates that given the hidden variable z In the case of The conditional probability distribution of ; Represents input variables and the model output mean The square difference between .

[0044] (5) Update the network parameters of the time domain feature extractor and the frequency domain feature extractor based on the reconstruction loss and the regularization loss.

[0045] It can be seen from the above that the features from different views are consistent and complementary to each other. The present invention integrates different views to play the functions of mutual verification and deviation correction in the feature space, thereby correcting the offset and error of a single view in the feature space, improving the clustering performance, and providing a richer and more accurate feature information basis for the model.

[0046] S102, inputting the label data into the pre-trained feature extractor to generate an original prototype; Furthermore, the steps of inputting the label data into the pre-trained feature extractor to generate the original prototype include: (1) Obtaining label data, and inputting the label data into a pre-trained time domain feature extractor to generate a first time domain feature vector; (2) converting the label data into first frequency domain data, and inputting the first frequency domain data into a pre-trained frequency domain feature extractor to generate a first frequency domain feature vector; (3) concatenating and fusing the first time domain feature vector and the first frequency domain feature vector to generate a first multi-view fusion feature; (4) Generate an original prototype based on the first multi-view fusion features.

[0047] It should be noted that the mean of the first multi-view fusion features of each class can be used as the original prototype of the corresponding class, that is, The original prototype It can be expressed as:

[0048] in: For Category A collection of; For collection The amount of data in To generate a classified prototype, we should input the data with the same label into the feature extractor and average the features of the same category. Indicates the label of this type of data. If the fault class is 10, then The value is 0 ~ 9; For label data The first multi-view fusion feature.

[0049] S103, inputting the label data in the support set into the pre-trained feature extractor to generate a vector median based on the label data; Further, the step of inputting the label data in the support set into the pre-trained feature extractor to generate a vector median based on the label data includes: (1) Obtaining label data in the support set, and inputting the label data in the support set into the pre-trained time domain feature extractor to generate a second time domain feature vector; That is, the N×K label data in the support set are input into the pre-trained time domain feature extractor to generate a second time domain feature vector.

[0050] (2) converting the label data in the support set into second frequency domain data, and inputting the second frequency domain data into the pre-trained frequency domain feature extractor to generate a second frequency domain feature vector; That is, the N×K label data in the support set are converted into second frequency domain data, and the second frequency domain data is input into the pre-trained frequency domain feature extractor to generate a second frequency domain feature vector.

[0051] (3) concatenating and fusing the second time domain feature vector and the second frequency domain feature vector to generate a second multi-view fusion feature; (4) Generate a vector median based on the label data according to the second multi-view fusion feature.

[0052] Specifically, the mean value of the second multi-view fusion features in each dimension may be used as the median value of the vector in each dimension.

[0053] S104, generating a cluster center based on the unlabeled data according to the unlabeled data; Further, the step of generating a cluster center based on the unlabeled data according to the unlabeled data includes: (1) Taking the original prototype as the initial cluster center; (2) Use the K-Means algorithm to generate cluster centers based on unlabeled data.

[0054] S105, updating the original prototype according to the vector median and the cluster center to generate a benchmark prototype; Further, the steps of updating the original prototype according to the vector median and the cluster center to generate the benchmark prototype include: (1) Calculate the mean of the vector median and the original prototype to fine-tune the original prototype; (2) Calculate the mean of the cluster center and the fine-tuned original prototype to refine the fine-tuned original prototype; The cluster center is combined with the fine-tuned original prototype, and the mean of each dimension is calculated to further update the original prototype.

[0055] Specifically, the benchmark prototype can be calculated according to the following formula:

[0056]

[0057] in: For the Class Benchmark prototypes in 3 dimensions; For the Class The original prototype of the dimension; For the Class The median of the vector of dimensions, The second multi-view fusion feature obtained after the N label data in the class is input into the encoder F is The mean vector obtained by averaging the dimensions; For the Class The cluster centers of the dimensions; is the adjustment rate of the cluster center to the original prototype, generally taken as =N / 10, the smaller N is, The smaller.

[0058] (3) The refined original prototype is used as the baseline prototype.

[0059] S106, calculating a cross entropy loss based on the label data according to the label data in the query set, and optimizing the feature extractor according to the cross entropy loss; Further, the steps of calculating the cross entropy loss based on the label data according to the label data in the query set and optimizing the feature extractor according to the cross entropy loss include: (1) Obtaining label data in the query set, and inputting the label data in the query set into the pre-trained time domain feature extractor to generate a third time domain feature vector; (2) converting the label data in the query set into third frequency domain data, and inputting the third frequency domain data into the pre-trained frequency domain feature extractor to generate a third frequency domain feature vector; (3) concatenating and fusing the third time domain feature vector and the third frequency domain feature vector to generate a third multi-view fusion feature; (4) Calculate the soft assignment based on the label data according to the third multi-view fusion feature:

[0060] in: Soft Assignment is the assignment matrix; P A tensor consisting of all primitive prototypes; Represents the original prototype belonging to this category, that is, if the query set The data labels are , it belongs to kind; Represents the feature domain of the query set data and the Original Prototype The Euclidean distance between represents all original prototypes, and the entire denominator represents the sum of the distances between the query set data and all original prototypes (there are 10 original prototypes for 10 classification problems); A tensor input representing the inverse of the Euclidean distance between the query set and each original prototype. softmax function.

[0061] (5) Calculate the cross entropy loss based on the label data according to the soft assignment, the third multi-view fusion feature and the true label of the label data:

[0062] in: L CE represents the cross entropy loss; K Indicates the total number of categories; The first one-shot feature vector k The true label of each element; Represents the feature vector extracted by the feature extractor.

[0063] (6) Optimize the network parameters of the time domain feature extractor and the frequency domain feature extractor based on the cross entropy loss.

[0064] It should be noted that the main training strategy of the present invention is semi-supervised deep multi-view clustering and original learning operation, that is, a pre-trained feature extractor is used to generate an original prototype from a small amount of labeled data, and then a self-supervised clustering algorithm is used with unlabeled data to correct the deviation of the original prototype.

[0065] S107, inputting the data to be diagnosed into the optimized feature extractor to generate a target prototype; S108, comparing the target prototype with the reference prototype to generate a fault diagnosis result.

[0066] Therefore, the bearing fault diagnosis method based on deep multi-view meta-learning in the present invention can make full use of the multi-view features of the data and unlabeled data, thereby improving the accuracy of small sample learning and the speed of model convergence.

[0067] In addition, the second embodiment of the bearing fault diagnosis method based on deep multi-view meta-learning of the present invention includes: S201, initialize the network parameters of the feature extractor through Kaiming , , initialize the prototype through Xavier ; S202, for each step, execute: S203, label data and unlabeled data Composition of training data , in the training data Random sampling in S204, the randomly sampled training data Convert to frequency domain data , the two data are input into the feature extractor respectively and ; S205, calculate the reconstruction loss: -> ; S206, calculate the regularization loss: -> ; S207, ; S208, updating the network parameters of the feature extractor , ; S209, end; S210, label data and frequency domain data Input pre-trained feature extractor , , generate the original prototype ; S211, for each step, execute: S212, in the label data Randomly sample N categories, each category contains K samples, forming a support set ; Similarly, randomly sample Q samples to form a query set ; S213, from the support set Randomly sample label data , label data Convert to frequency domain data , label data and frequency domain data Input feature extractor and , and obtain the feature vector and ; S214, generate the median of the vector of label data of N classes ; S215, Generate unlabeled data using K-Means The cluster center ; S216, using the median of a vector , original prototype and cluster centers , forming a benchmark prototype ; S217, Using QuerySets and frequency domain data Calculating cross entropy loss ; S218, update the network parameters of the feature extractor , ; S219, end; S220, inputting the data to be diagnosed into the optimized feature extractor to generate a target prototype; S221, comparing the target prototype with the reference prototype to generate a fault diagnosis result.

[0068] Therefore, the present invention includes two feature extractors to extract features from the original signal and the frequency domain signal, and adopts a metric-based method as a classifier; an encoder with good feature extraction capability is obtained in pre-training; in fine-tuning learning, only the encoder obtained after pre-training and the original prototype obtained by the labeled data are needed, and a smaller labeled data set (1%-5%) and more sufficient unlabeled data (30%) are used for fine-tuning, so that the model can obtain satisfactory performance.

[0069] The present invention is further described in detail below in conjunction with specific comparative experiments: 1. Dataset Description The bearing data for the comparative experiment comes from the CWRU bearing fault dataset, PCIU bearing fault dataset, and IMS bearing fault dataset. The details are as follows.

[0070] CWRU: This data set is obtained from the Fault Bearing Data Center of Case Western Reserve University, with a collection frequency of 12KHz and 1797rpm selected; the data set contains four bearing states: healthy state (N), external fault (OF), ball fault (BF) and internal fault (IF); each fault type includes three damage sizes of 7, 14 and 21mils. The present invention uses three types of fault data and one type of healthy data for all the above speeds, and uses four state data points as diagnostic targets; in order to construct the bearing fault data sample, a window with a length of 1200 and a step size of 100 is used to slice the data, each sample contains 1200 data points, and each classification consists of 1000 samples; the CWRU bearing fault data set is shown in Table 2 below: Table 2

[0071] PCIU: This data set is collected by the Key Laboratory of Fault Diagnosis of Guangdong University of Petrochemical Technology. NSK1310 rolling bearing is used as the experimental platform, and CTCaccelerometer is used as the sensor. PCIU has six fault types, four of which correspond to normal state (N), inner ring fault (IF), outer ring fault (OF) and roller fault (RF). The above six fault data states are sliced ​​using a window with a length of 1024 and a step size of 80. The PCIU bearing fault data set is shown in Table 3 below: Table 3

[0072] IMS: This data set is collected by the University of Cincinnati and records the entire life cycle data of the bearing. IMS has four fault types, including outer ring fault (ORF), inner ring fault (RF), ball fault (BF) and normal state (N). The above four fault state data are sliced, and the data point length is set to 1200 and the step length is 120. The IMS bearing fault data set is shown in Table 4 below: Table 4

[0073] Accordingly, the model hyperparameter settings are shown in Table 5 below: Table 5

[0074] 2. Comparative Test In order to verify the superiority of the present invention, several currently popular deep learning methods are selected, and their performance on the above CWRU, PCIU and IMS data sets is compared with that of the present invention. The specific comparison method is as follows: CNN: A common supervised learning method with the same network structure as the feature extractor of ProtoNet.

[0075] VAE: A supervised learning method, the decoder part is reduced to 10 dimensions and then input Softmax The function gets the classification result.

[0076] ProtoNet: A meta-learning strategy that uses a feature extractor to extract feature vectors and then uses a metric-based classifier for classification.

[0077] PSSPN: A prototype network with convolutional neural network as feature extractor, which is a semi-supervised few-shot learning method for pseudo-labeling unlabeled data.

[0078] MFFML: A meta-learning method based on multi-feature fusion. The frequency domain signal is obtained by Fourier transform of the original data. During the model training process, the time domain and frequency domain features under different loads are extracted and fused, and a multi-task meta-training and meta-testing method is established to realize small sample bearing fault diagnosis.

[0079] SSMN: A prototype network with a convolutional neural network as a feature extractor and an SE attention mechanism module. It is a semi-supervised few-shot learning method that uses pseudo-label operations for prototype refinement.

[0080] The experiment was conducted using the data division method in Table 2 and the model hyperparameter setting method in Table 5 (note that the training samples, unlabeled samples, and test samples in Table 2 do not overlap). The experiment was trained using the 5-way 10-shot small sample training method. The average accuracy (%) of CWRU is shown in Table 6 below: Table 6

[0081] The PCIU dataset was divided using the data division method in Table 3, and the model hyperparameter setting method in Table 5 was used for the experiment. The experiment was trained using the 3-way 10-shot few shot training method. The average accuracy (%) of PCIU is shown in Table 7 below: Table 7

[0082] The IMS dataset was divided using the data division method in Table 4, and the model hyperparameter setting method in Table 5 was used for the experiment. The experiment was trained using a 2-way 10-shot small sample training method. The average accuracy (%) of IMS is shown in Table 8 below: Table 8

[0083] It can be seen from Table 6, Table 7 and Table 8 that the average accuracy of the present invention is the highest.

[0084] 3. Ablation Experiment and Analysis The solution proposed in this paper includes three strategies: multi-view feature fusion (MVFF), feature layer self-attention mechanism (FLSA), and semi-supervised prototype refinement (SSPR). This paper designs a series of ablation experiments to control variables to analyze the effect of each strategy on small sample learning. Among them, the accuracy (%) of each method in the ablation experiment is shown in Table 9 below: Table 9

[0085] As can be seen from Table 9, the use of MVFF+PN achieves significantly better performance than the other three combinations (PN, FLSA+PN, SSPR+PN), which shows that the MVFF strategy has a significant effect on small sample learning. In addition, under different task settings, FLSA+PN and SSPR+PN both effectively improve the accuracy of diagnosis, verifying the effectiveness of SSPR and FLSA.

[0086] Among the combinations of MVFF+FLSA+PN, FLSA+SSPR+PN, MVFF+SSPR+PN, and MVFF+FLSA+SSPR+PN, the combination of MVFF and SSPR makes the model performance optimal, which shows that the coordination of the two strategies is the main factor in improving the accuracy, and the accuracy of MVFF+FLSA+PN and FLSA+SSPR+PN is also improved to a certain extent. Finally, the MVFF+FLSA+SSPR+PN of the present invention achieves the best accuracy in all tasks, which fully reflects the superiority of the present invention.

[0087] Excluding the excellent performance of MVFF in terms of accuracy, the effects of FLSA and SSPR on model convergence speed and accuracy are analyzed. Figure 3 It can be seen that FLSA has a significant effect on improving the convergence speed, which verifies the role of the feature layer attention mechanism in improving the model training efficiency. In addition, SSPR not only improves the convergence speed of the model but also improves the accuracy of the model by introducing unlabeled data for prototype refinement. With the cooperation of FLSA and SSPR, a higher accuracy rate can be obtained at a higher training efficiency.

[0088] 4. Anti-noise experiment In order to verify the noise resistance of the present invention, additive Gaussian noise SNR values ​​were designed to be set to 0, 3, 6, and 9 dB for input. The CWRU dataset was selected as the data set, and the 10 way 5 shot training method was selected. It was compared with CNN, VAE, Protonet, PSSPN, SSMN, and MFFML in the comparative test. The experiment was repeated 10 times, and the average value of the accuracy was selected.

[0089] Combination Figure 4 It can be seen that the present invention has stronger robustness than the method in the comparative experiment, which proves that the feature extractor of the present invention has effective anti-noise capability.

[0090] V. Results Analysis According to the above experiments, the semi-supervised multi-view clustering and meta-learning method proposed in the present invention, which integrates the features of different views of the same sample, effectively improves the convergence speed and accuracy of the model, and also achieves a certain effect in noise resistance. The present invention is not only applicable to VAE composed of fully connected layers, but also to other models as feature extractors, so it is possible to be deployed in large-scale practical applications.

[0091] Correspondingly, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the above-mentioned bearing fault diagnosis method based on deep multi-perspective meta-learning when executing the computer program.

[0092] At the same time, the present invention also discloses a computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the above-mentioned bearing fault diagnosis method based on deep multi-view meta-learning are implemented.

[0093] In summary, the bearing fault diagnosis method based on deep multi-view meta-learning of the present invention can make full use of the multi-view features of the data and unlabeled data, solve the small sample problem, and make full use of the large amount of unlabeled data in the actual production environment, thereby improving the diagnosis accuracy and convergence speed of the model under the condition of sparse labeled samples. Specifically: The present invention proposes a multi-view deep feature extractor, which is composed of two variational encoders with the same structure connected in parallel, and can respectively extract time domain and frequency domain features from the original signal, improve the model accuracy, have better classification effect for boundary samples, and have stronger anti-noise performance; The present invention integrates a self-attention mechanism on the feature extractor to form a feature-layer self-attention mechanism variational encoder. The feature-layer self-attention mechanism is beneficial to minimizing the intra-class compactness (locality) to maximize the inter-class separability (non-locality), thereby improving the convergence speed and noise resistance of the model, and achieving the effect of improving the model performance. The present invention proposes a prototype adjustment strategy based on semi-supervised multi-view meta-learning. The class prototype is first constructed using the multi-view features of labeled data, and then the clustering centers of the multi-view features of unlabeled data are iterated on the class prototype. The clustering centers are less affected by label noise, have higher accuracy, and converge faster. When adjusting the prototype, the dependence on labeled data can be reduced, the features of unlabeled data can be learned more effectively, and the risk of overfitting of sparse samples can be reduced.

[0094] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A bearing fault diagnosis method based on deep multi-view meta-learning, characterized in that: include: Pre-training a feature extractor using training data, wherein the training data includes labeled data and unlabeled data, and the labeled data includes a support set and a query set; Feed the labeled data into the pre-trained feature extractor to generate the original prototype; Input the labeled data in the support set into the pre-trained feature extractor to generate a vector median based on the labeled data; Generating a cluster center based on the unlabeled data according to the unlabeled data; Updating the original prototype according to the vector median and the cluster center to generate a benchmark prototype; Calculating a cross entropy loss based on the label data according to the label data in the query set, and optimizing the feature extractor according to the cross entropy loss; Input the data to be diagnosed into the optimized feature extractor to generate a target prototype; The target prototype is compared with the reference prototype to generate a fault diagnosis result.

2. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1 is characterized in that: The step of pre-training the feature extractor using training data comprises: Acquire training data, and input the training data into a time domain feature extractor to generate an initial time domain feature vector; Converting the training data into initial frequency domain data, and inputting the initial frequency domain data into a frequency domain feature extractor to generate an initial frequency domain feature vector; Concatenating and fusing the initial time domain feature vector and the initial frequency domain feature vector to generate an initial multi-view fusion feature; Reconstruction loss and regularization loss are calculated respectively according to the initial multi-view fusion features; According to the reconstruction loss and the regularization term loss, the network parameters of the time domain feature extractor and the frequency domain feature extractor are updated.

3. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1, characterized in that: The step of inputting the label data into the pre-trained feature extractor to generate the original prototype includes: Acquire label data, and input the label data into a pre-trained time domain feature extractor to generate a first time domain feature vector; Converting the label data into first frequency domain data, and inputting the first frequency domain data into a pre-trained frequency domain feature extractor to generate a first frequency domain feature vector; Concatenating and fusing the first time domain feature vector and the first frequency domain feature vector to generate a first multi-view fusion feature; An original prototype is generated according to the first multi-view fusion features.

4. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1, characterized in that: The step of inputting the label data in the support set into the pre-trained feature extractor to generate a vector median based on the label data comprises: Acquire label data in a support set, and input the label data in the support set into a pre-trained time-domain feature extractor to generate a second time-domain feature vector; Converting the label data in the support set into second frequency domain data, and inputting the second frequency domain data into a pre-trained frequency domain feature extractor to generate a second frequency domain feature vector; Concatenating and fusing the second time domain feature vector and the second frequency domain feature vector to generate a second multi-view fusion feature; A vector median based on label data is generated according to the second multi-view fusion feature.

5. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1, characterized in that: The step of generating a cluster center based on the unlabeled data according to the unlabeled data comprises: Taking the original prototype as the initial cluster center; The K-Means algorithm is used to generate cluster centers based on unlabeled data.

6. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1, characterized in that: The step of updating the original prototype according to the vector median and the cluster center to generate a reference prototype comprises: Calculating the mean of the median of the vector and the original prototype to fine-tune the original prototype; Calculating the mean of the cluster center and the fine-tuned original prototype to refine the fine-tuned original prototype; The refined original prototype is used as the benchmark prototype.

7. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1, characterized in that: The step of updating the original prototype according to the vector median and the cluster center to generate a reference prototype comprises: According to the formula , computing benchmark prototypes; in, For the Class The benchmark prototype of the dimensions, For the Class The original prototype of the dimension, For the Class The median of the vector of dimensions, For the Class The cluster center of the dimension, is the adjustment rate.

8. The bearing fault diagnosis method based on deep multi-view meta-learning according to claim 1, characterized in that: The step of calculating a cross entropy loss based on the label data according to the label data in the query set and optimizing the feature extractor according to the cross entropy loss includes: Obtaining label data in a query set, and inputting the label data in the query set into a pre-trained time-domain feature extractor to generate a third time-domain feature vector; Converting the label data in the query set into third frequency domain data, and inputting the third frequency domain data into a pre-trained frequency domain feature extractor to generate a third frequency domain feature vector; Concatenating and fusing the third time domain feature vector and the third frequency domain feature vector to generate a third multi-view fusion feature; Calculate a soft assignment based on label data according to the third multi-view fusion feature; Calculate a cross entropy loss based on the label data according to the soft assignment, the third multi-view fusion feature and the true label of the label data; The network parameters of the time domain feature extractor and the frequency domain feature extractor are optimized according to the cross entropy loss.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the bearing fault diagnosis method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the bearing fault diagnosis method according to any one of claims 1 to 8 are implemented.

Images