A few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance
By combining domain knowledge and semi-supervised learning, utilizing wavelet packet decomposition and information entropy to extract features, and integrating deep learning networks and XGBoost classifiers, the problem of data scarcity and complex operating conditions in bearing fault diagnosis is solved, achieving efficient fault localization and diagnosis with few samples.
Patent Information
- Application Number
- CN202310341999.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-03
- Publication Date
- 2025-12-09
- Estimated Expiration
- 2043-04-03
AI Technical Summary
In bearing fault diagnosis, existing deep learning models rely on a large amount of data for training, and are difficult to apply effectively in engineering practice due to data scarcity and complex operating conditions, especially in cases with few samples, where efficient diagnosis is difficult to achieve.
A domain knowledge-guided semi-supervised bearing fault diagnosis method is adopted, which combines wavelet packet decomposition and information entropy to extract physical features, integrates virtual features from deep learning networks, uses an improved attention mechanism and XGBoost classifier, and performs few-shot fault localization through a semi-supervised self-training strategy.
It improves classification accuracy and model applicability under complex working conditions, reduces computational requirements, is applicable to various working conditions and a small number of labeled samples, and achieves efficient fault diagnosis.
Smart Images
Figure CN116304836B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of mechanical part fault positioning, and particularly relates to a few-sample semi-supervised bearing fault diagnosis method based on field knowledge guidance, and is applied to bearing fault positioning in few-sample multi-working-condition. BACKGROUND
[0002] In the modern industrial tide, power machines represented by rotary machines are gradually developing towards precision, complexity and automation. As a core component in mechanical transmission systems, rolling bearings play a crucial role in various fields such as wind power generation, transportation, aerospace, etc. In actual industrial production and daily mechanical operation, the operation of bearings affects production efficiency, production safety, operation stability, etc. Fault prediction and health management technology links the monitoring data of the machine with the health status, and focuses on the health status of the machine. The diagnosis model automatically judges the fault status by analyzing the signal. The fault diagnosis technology based on deep learning relies on a large amount of data for learning, but in actual engineering, fault data is scarce, and the amount of data used for model training is limited, so the actual application ability of the model is restricted. In the few-sample fault diagnosis task, generally starting from the network ability or data enhancement two angles. However, in the actual industry, the collected signal is affected by the working condition and working environment, and the distribution of the data is inconsistent, which not only requires a lot of training time, but also has high requirements for data processing. Directly applying deep learning algorithm to engineering practice is limited by the complexity of the scene and the scarcity of data, and it is difficult to get the expected effect. On the basis of the previous research in a certain field, the deep learning network is established, that is, the process of integrating field knowledge into data-driven model to create a model with physical common sense to provide guidance for learning and more information, which can greatly improve the effect of the model.
[0003] Therefore, the ideal bearing fault positioning model suitable for engineering practical tasks should have the following characteristics: the method should not be high in computing power requirement and have high computing efficiency; the method should be suitable for various working environments; the method should be able to diagnose in the case of only a small amount of samples with correct labels. SUMMARY
[0004] To solve the above technical problems, the application provides a few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance, which is based on the domain knowledge embedding idea, improves the feature engineering based on the deep learning network to extract virtual features, adopts a feature extraction strategy of fusing physical features and general features, and realizes feature highlighting through mutual complementation of the two features. Meanwhile, the Meta-ACON new technology is introduced to improve the XGBoost classifier, and before the feature vector is input into the classification layer, the attention mechanism is used to weight the features, so that the classification accuracy is improved. Finally, the semi-supervised-self-training training strategy is introduced, and a semi-supervised bearing fault diagnosis model under the condition of few samples is established.
[0005] The application is realized through the following technical schemes:
[0006] A few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance comprises the following steps:
[0007] Step 1, data acquisition and segmentation. The vibration signal is acquired from the test bench or actual working parts by using a vibration sensor. The original signal is segmented and sampled by using a sliding window.
[0008] Step 2, physical feature extraction based on domain knowledge embedding. The signal is decomposed by using a wavelet packet algorithm, and the energy information of the signals at each layer after decomposition is extracted based on information entropy and energy value to constitute domain knowledge features.
[0009] Step 3, virtual feature extraction based on a deep learning network. The virtual features are extracted by using a deep convolutional neural network, and a feature extraction module containing three convolutional layers, pooling layers and normalization layers is established.
[0010] Step 4, feature fusion and fault positioning classification. After the above features are normalized, the feature set is spliced and input into the improved attention mechanism layer, the features are transformed and weighted. XGBoost is used to replace the Softmax classification layer to realize fault positioning.
[0011] Step 5, semi-supervised training strategy. When performing classification training, the semi-supervised-self-training training strategy is adopted. 30% to 50% of the labels are randomly discarded, the classifier is updated through iteration, and few-shot semi-supervised fault positioning is realized.
[0012] Specifically, step 1 is specifically as follows:
[0013] The sliding window is set to sample; it is assumed that the sample window length is a, the sample signal is extracted from i th The total length of the signal is N, and the number of samples n can be obtained:
[0014] The specific method of domain knowledge feature extraction in step 2 is as follows:
[0015] The sample x is decomposed orthogonally as follows: i
[0016]
[0017] Where P is the approximation wavelet packet function, D is the detail wavelet packet function, and x(t) is the frequency band signal. j x(t) is the low frequency band signal. j And ψ jk (t) is the approximation wavelet packet function, and the coefficients and have the following values:
[0018]
[0019] h is the recursive coefficient. Through formulas (1) and (2), the wavelet packet decomposition process is to decompose the signal into multiple wavelet packets. If the original signal is decomposed into j layers, 2 j
[0020] After wavelet packet decomposition of the signal, the decomposition sequence of the jth layer is S jk (k=0~2 j -1), and the energy value on each scale is calculated; that is, the wavelet packet decomposition process is equivalent to the energy decomposition process, and the energy sequence is marked as E1, E2,..., E k .
[0021] E i =|D k (i)| 2 (3)
[0022] Where D k (i) is the signal component of each wavelet packet reconstruction coefficient.
[0023] The random variable X is used to represent the state characteristics of the uncertain system, and the probability p j of taking value x j is:
[0024] p j = P{X=x j} (4)
[0025] The information I j obtained by a certain result of X is represented as:
[0026]
[0027] The information entropy H(X) of X is represented as:
[0028]
[0029] The total power E of the signal at a certain scale is represented as the sum of the energy of each node; let:
[0030] ε jk (i) = E i / E (7)
[0031] ∑ k ε jk = 1 (8)
[0032] ε jk is the information ratio occupied by each layer of signal, then the wavelet packet energy entropy H jk is:
[0033]
[0034] where N is the length of the signal, H jk is the kth wavelet packet energy entropy of the signal.
[0035] In step 3, the virtual feature extraction based on deep learning network, the general feature is improved from CNN network, including convolution layer, pooling layer, activation function and full connection layer; 1D CNN network is used in time series signal processing, the output is shown in formula (10):
[0036] where k w and b w are weights, and x is input;
[0037]
[0038] The pooling layer is used for feature selection, as shown in formula (11), to reduce parameters and speed up the convergence process.
[0039]
[0040] Data regularization is realized through Batchnorm layer, and the output is y i , which improves the generalization ability and solves the problem of inconsistent internal variables.
[0041] GELu is a nonlinear activation function that is a random regular way that meets expectations.
[0042]
[0043] Three convolutional layers are set, and the pooling layer is connected with the full connection layer, and the output of the full connection layer is:
[0044] gq = GELu(p3w q +b q ) (13)
[0045] where g q is the output of the fully connected layer, w q and b q are the weight matrix and bias of the layer.
[0046] Step 4: Feature fusion and fault location classification, the specific method is:
[0047] The domain knowledge features and general features are spliced and added to the attention block;
[0048] After feature splicing, the intermediate matrix is generated by inputting the convolution layer and Batchnorm layer; as shown in equation (14):
[0049] h = δ (F g [cat(x CNN ,x Prior )]) (14)
[0050] δ is the Meta-ACON activation function, and F g represents the attention mechanism function;
[0051] Then h is the output of the attention mechanism, and h is divided into two parts x' and other parts; to make full use of the original signal information of the prior feature part and the critical pulse spectrum part in the feature, x' is mapped through the function to make it consistent with X Prior The number of channels is consistent, and the convolution layer and Sigmoid are connected to output y c :
[0052] g = σ [F (f x′ )] (15)
[0053]
[0054] y c is connected to the XGBoost classification to obtain the classification result.
[0055] Step 5: Semi-supervised-self-training training strategy, the steps are as follows:
[0056] Step 1: The sample set contains labels, randomly select data and discard labels to obtain a new data set; divide the data set according to the proportion of the training set and the test set, and train the classifier in the data set containing labels;
[0057] Step2: Use the trained classifier to predict the labels of unlabeled data, set the threshold, and select the data with higher probability as "pseudo labels";
[0058] Step3: Combine the data set containing "pseudo labels" and "true labels" for secondary training of the classifier;
[0059] Step4: Use the secondary trained classifier to classify the data set containing "true labels", and set the evaluation standard to evaluate the classifier at this time;
[0060] If there are still unlabeled labels, repeat steps 2-4 until there are no unlabeled data sets.
[0061] During the training iteration, F1 is used as the indicator of the effect of the classifier:
[0062]
[0063] Where precision is the precision, and recall is the recall.
[0064] The advantages of the present application are:
[0065] (1) The present application is inspired by knowledge embedding and knowledge discovery, and based on wavelet packet decomposition and information entropy, a domain knowledge guided improved feature engineering is established, which reduces the training time of the network for virtual feature selection, compared with the simple physical feature engineering of the domain and the feature engineering based on deep learning network, the network can obtain a feature vector less affected by the environment, and the classification accuracy under complex working conditions is improved;
[0066] (2) The input part of the classifier is improved, and an improved attention mechanism module based on Meta-ACON is used instead of the feature scoring and feature selection part of XGBoost, which is combined with the XGBoost classifier for fault classification.
[0067] (3) The present application uses a semi-supervised-self-training training strategy to iteratively improve the classifier, making it suitable for industrial practice. Multiple data sets are used to verify the model, which verifies the universality and robustness of the model under complex working conditions and on multiple workbenches. BRIEF DESCRIPTION OF DRAWINGS
[0068] Figure 1 is the semi-supervised bearing fault diagnosis method flowchart based on domain knowledge guidance of the present application;
[0069] Figure 2 is a feature fusion algorithm schematic diagram based on an improved attention mechanism of the present application;
[0070] Figure 3is a schematic diagram of the semi-supervised training strategy of the present application;
[0071] Figure 4 is an experimental device diagram in an example of the present application;
[0072] Figure 5 is a comparison chart of experimental results of the present application;
[0073] Figure 6 is a flowchart of the present application. DETAILED DESCRIPTION
[0074] In order to more clearly present the technical solutions of the present application and their superiorities, the present application will be described in detail in combination with the accompanying drawings and experimental examples. In addition, the experimental examples used prove only for the description of the technical solutions of the present application, but the present application is not limited to the examples used.
[0075] As shown in the present application, Figure 1 In order to ensure the continuity of the sample in the time dimension and make the most of the limited data, a sliding window is used to sample the data collected in a period of time. The sample set x i is a certain sample containing n data points, and N is the total number of samples. In order to simulate the phenomenon of labeled sample scarcity in engineering practice, the sample is divided into labeled part X L and unlabeled part X U during training.
[0076] In the feature engineering part, first, according to the energy distribution characteristics of the rotating part failure, the energy feature is combined with the wavelet packet decomposition and the information entropy algorithm to propose a wavelet packet tree-energy entropy feature algorithm as the prior knowledge embedding part, marked as X Prior . Then an unsupervised CNN neural network is used to extract virtual features. The general features extracted by the deep learning network often do not contain the information possessed by the prior features, marked as X CNN . The prior knowledge part features and the general features are complementary to each other to obtain the fusion feature set F i ={f1,f2,f3,...f 15 ,f 16}. A fault discriminator composed of attention mechanism and XGBoost classifier is connected with the feature layer, and an intelligent diagnosis algorithm is proposed.
[0077] As shown in the flowchart, Figure 6 the method process and principle are described as follows:
[0078] Step 1, data segmentation. In order to make the most of the sample and prevent the CNN network from extracting irrelevant fault features, a sliding window is set for sampling. Assuming that the sample window length is a, from i thThe sample number n can be obtained as follows: the sample signal (i sample point position) is extracted, the length of the overlapping interval is b, and the total number of signals is N.
[0079] Step 2, field knowledge feature extraction.
[0080] The wavelet packet theory is improved on the basis of the wavelet theory, and can simultaneously use a high-pass filter and a low-pass filter in a signal, simultaneously process an approximate signal and a high-frequency band detail signal, and effectively ensure signal integrity. The sample x i is orthogonally decomposed as follows:
[0081]
[0082] where P j x(t) is a frequency band signal, D j x(t) is a low-frequency band signal, and ψ jk (t) is an approximate wavelet packet function, and the coefficients and have the following values:
[0083]
[0084] h is a recursive coefficient. Through formulas (1) and (2), the wavelet packet decomposition process is to decompose a signal into multiple wavelet packets. If the original signal is decomposed into j layers, 2 j wavelet packets can be obtained, each wavelet packet represents a component of the original signal in different frequency bands, and each wavelet packet has the same width and the number of adjacent decomposition layers. With the increase of the number of decomposition layers, the description of the original signal is more and more detailed.
[0085] After the signal is decomposed by the wavelet packet, the decomposition sequence of the jth layer after decomposition is S jk (k=0~2 j -1), and the energy value on each scale is calculated. That is, the wavelet packet decomposition process is equivalent to the energy decomposition process, and the energy sequence is marked as E1, E2,..., E k .
[0086] E i =|D k (i)| 2 (3)
[0087] where D k (i) is a signal component of each wavelet packet reconstruction coefficient.
[0088] A random variable X is used to represent the state characteristics of an uncertain system, and the probability that the value X takes x j is p j .
[0089] p j = P{X = x j} (4)
[0090] Information I obtained by a certain result of X j can be expressed as:
[0091]
[0092] Then the information entropy H(X) of X can be expressed as:
[0093]
[0094] In a certain scale of the signal, the total power E of the signal represents the energy sum of each node. Assuming:
[0095] ε jk (i) = E i / E (7)
[0096] ∑ k ε jk = 1 (8)
[0097] ε jk is the information ratio occupied by each layer signal, then the wavelet packet energy entropy H jk is:
[0098]
[0099] Where N is the length of the signal, H jk is the kth wavelet packet energy entropy of the signal.
[0100] Step 3, virtual feature extraction based on CNN.
[0101] The general feature extraction part is mainly improved from the CNN network, including convolution layer, pooling layer, activation function and full connection layer. In time series signal processing, 1D CNN network is adopted, and the output is shown in formula (10):
[0102] Where k w and b w are weights, and x is input
[0103]
[0104] The pooling layer is used for feature screening, as shown in formula (10), to reduce parameters and speed up the convergence process.
[0105]
[0106] Data regularization is realized through the Batchnorm layer, and the output is y i, improve the generalization ability and solve the problem of internal variable inconsistency.
[0107] GELu(Gaussian error linear units) as a nonlinear activation function is a random regular way as expected.
[0108]
[0109] Three layers of convolution and pooling layers are set to connect with the full connection layer, and the output of the full connection layer is:
[0110] g q = GELu(p3w q +b q ) (13)
[0111] where g q is the output of the full connection layer, w q and b q are the weight matrix and bias of the layer
[0112] Step 4, feature fusion and fault location classification based on improved attention mechanism.
[0113] As shown in Figure 2 , the prior features and general features are spliced and added to the attention block. In order to obtain the spatial and positional relationship between the prior features and the general features, the relationship between the two groups of vectors is established, and the intermediate matrix is generated after the convolution layer and Batchnorm layer are input after the feature splicing. As shown in equation (14):
[0114] h = δ(F g [cat(x CNN ,x Prior )]) (14)
[0115] δ is the Meta-ACON activation function, and F g represents the attention mechanism function.
[0116] Then h is the output of the attention mechanism, and h is divided into two parts x' and other parts; In order to make full use of the original signal information of the prior feature part and the critical pulse spectrum part in the feature, x' is mapped through the function to make it consistent with X Prior channel number, and connect the convolution layer and Sigmoid, output y c :
[0117] g = σ[F(f x′ )] (15)
[0118]
[0119] yc Then, Flatten and L2-Norm are connected to XGBoost for classification to obtain the classification results.
[0120] Step 5: Training strategy based on semi-supervised self-training.
[0121] In practical engineering applications, not only is there a scarcity of specific samples, but also the phenomenon of missing labels. Manually labeling the sample set is not only costly in terms of manpower, resources, and finances, but also carries the risk of incorrect labeling. Therefore, semi-supervised training methods are used during training, such as... Figure 3 As shown, the training scheme steps used in this invention are as follows:
[0122] Step 1: Since the sample set contains labels, randomly select data and discard the labels to obtain a new dataset. Divide the dataset proportionally into training and test sets, and train the classifier on the labeled dataset.
[0123] Step 2: Use the trained classifier to predict the labels of the unlabeled data, set a threshold, and select the data with higher probability as "pseudo-labels";
[0124] Step 3: Combine the datasets containing "pseudo-labels" and "real labels" for secondary training of the classifier;
[0125] Step 4: Use the retrained classifier to perform classification prediction on the dataset containing "true labels", and set evaluation criteria to evaluate the classifier at this point.
[0126] If there are still unlabeled data points, repeat Steps 2 through 4 until there are no more unlabeled datasets.
[0127] The above training scheme is an incremental algorithm. It trains the original classifier using a small number of labeled samples, and then iterates through the training process until certain conditions are met, resulting in a classifier with good classification performance. During the training iterations, F1 score is used as the metric for classifier performance.
[0128]
[0129] Where precision is the accuracy rate and recall is the recall rate, this invention uses bearing failure data from a comprehensive experimental platform for simulating mechanical failures in a high-speed train safety laboratory for practical verification. The experimental platform structure is as follows: Figure 4The test bearing model is MBER-12K, the number of rollers is 8, and the ball diameter is 0.3125 mm. The fault diameter of the fault bearing is 1 mm and 2 mm, and the damage is formed on the roller, the inner ring and the outer ring to form a single-point fault and a composite fault by using the electric spark point erosion technology. The motor speed is 2000 r / min, and the sampling frequency is 32 kHz. According to the fault position, the samples are classified into inner ring fault, outer ring fault, roller fault, mixed fault and healthy state. For small sample and sample label loss, experiments are arranged for verification, and compared with other methods.
[0130] In the small sample fault diagnosis classification task, the sample is defined according to the sample proportion or sample number. In the verification of the application, not more than 50 samples of each class are defined as a small sample experiment. The effectiveness of the model is compared when the number of samples of each class is 30, 50 and 100. At the same time, other classifiers are compared.
[0131] GP is a feature that contains general features and prior features, G only contains general features, and P represents prior features. Five-fold cross-validation is used, and the results are shown in Table 1. Prior features play a leading role in the CSU data set based on general features. The two features complement each other to effectively improve the diagnosis accuracy. Compared with SVM, KNN and RandomForest as a classifier, the classification accuracy of XGBoost is higher, and when the number of samples of each class is 30, the classification accuracy can be more than 85. It can be proved that the feature engineering based on prior knowledge embedding contributes to the classification accuracy and the superiority of XGBoost compared with other classifiers. The classification ability of the method proposed in the application and the existing deep learning network is compared, especially in the case of extremely small number of samples. Each fault sample is set to 5, 30, 50 and 100, and the comparison experiments are performed with MLP, ResNet and Alexnet network. As shown in Table 1, when the number of training samples of each class is only 5, the method of the application can also achieve an accuracy of more than 70%. The CNN-based model selects the same structure of CNN network for feature extraction in the backbone network part. Although the effect is slightly better than the method in some tasks, the accuracy is about 2% higher than the method in the application, but the effect can be stable after training for more than 50 epochs, the learning efficiency is not high, and the time-consuming is long. And in each data set, ResNet, Alexnet and MLP do not perform stably, and the generalization ability is inferior to the method proposed in the application. Figure 5
[0132] Table 1: Results of small sample fault positioning experiment
[0133]
[0134]
[0135] In the above fault diagnosis scheme, a semi-supervised bearing fault training strategy is added to verify the applicability of the method proposed in the present application under the condition of small number of model learning samples and partial sample label loss. A part of the labels in the training set is randomly lost, and the loss label proportion is 30% to 50%. The accuracy and F1 score of each class of samples with 30, 50 and 100 samples are verified respectively.
[0136] The results are shown in Table 2. In the semi-supervised training task, the method proposed in the last section has an accuracy decrease of 5% to 10%, but it can still maintain a high accuracy in the small sample classification task. In the experimental data set, two fault specifications are set for the experimental task, and the maximum test set proportion is 92.5%, and the accuracy can reach at least 88.44%, and the F1 index can reach 89.14%. Compared with the other three machine learning classification methods, it can be proved that it has better performance in accuracy and stability. In the semi-supervised training iteration process, the convergence speed is fast, and the iteration times are 8 to 15. In the task of small sample semi-supervised bearing fault diagnosis, the method provided in the present application has good robustness and generalization performance in training accuracy and speed.
[0137] Table 2 Semi-supervised small sample fault positioning experimental results
[0138]
[0139] The above experiments prove that the method has advantages in training speed, accuracy and stability under complex working conditions, and reduces the requirement for computing power, and is suitable for early bearing fault location and bearing health monitoring tasks.
Claims
1. A few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance, characterized in that, Comprise the following steps: Step 1, data acquisition and segmentation; utilize vibration sensor to obtain vibration signal from test bench or actual working parts; adopt sliding window to sample from original signal; Step 2, physical feature extraction based on domain knowledge embedding, adopt wavelet packet algorithm to decompose signal, extract energy information of each layer signal after decomposition based on information entropy and energy value to constitute domain knowledge feature; Step 3, virtual feature extraction based on deep learning network; utilize deep convolutional neural network to extract virtual feature, establish a feature extraction module containing three convolution layers, pooling layers and normalization layers; Step 4, feature fusion and fault positioning classification; after normalization processing of above features, splice feature set and input improved attention mechanism layer, transform and weight features; use XGBoost to replace Softmax classification layer to realize fault positioning; The specific method is: Splice domain knowledge feature and general feature and add attention block; After feature splicing, input convolution layer and Batchnorm layer to generate intermediate matrix; as shown in formula (14): h = δ(F g [cat(x CNN ,x Prior )]) (14) delta is a Meta-ACON activation function, F g represents an attention mechanism function; Then h is the attention mechanism output, h is divided into two parts x' and other parts; in order to make full use of the original signal information of the prior feature part, and the critical pulse spectrum part in the feature, x' is mapped through the function to make it consistent with X Prior The number of channels remains the same, and the convolution layer and Sigmoid are connected, and the output y c : g = σ[F(f x′ )] (15) y c Post-access Flatten and L2-Norm, with XGBoost classification to get the classification result; Step 5, semi-supervised training strategy; when classifying and training, adopt semi-supervised-self-training training strategy; randomly discard 30% to 50% labels, update classifier through iteration to realize few-sample-semi-supervised fault positioning.
2. The few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance according to claim 1, characterized in that, The specific method of step 1 is: A sliding window is set for sampling; assuming a sample window length of a, a sample signal is extracted from i th The sample number n can be obtained as follows: a sample point position i, an overlapping interval length b, and a total signal length N.
3. The few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance according to claim 2, characterized in that, In step 2, the specific method of domain knowledge feature extraction is: On the sample x i Performing orthogonal decomposition: where P j x(t) is the band signal, D j x(t) is the low band signal, and ψ jk (t) is the approximation wavelet packet function, the coefficients and have the values: h is a recursive coefficient The process of wavelet packet decomposition is to decompose the signal into multiple wavelet packets by formula (1), (2). If the original signal is decomposed into j layers, 2j wavelet packets are obtained, each of which represents the composition of the original signal in different frequency bands, and each wavelet packet has the same width and the number of adjacent decomposition layers. j The signal is decomposed by wavelet packet, and the decomposition sequence of the jth layer after decomposition is S jk (k = 0~2 j -1), the energy value on each scale is calculated; that is, the wavelet packet decomposition process is equivalent to the energy decomposition process, and the energy sequence is marked as E1, E2,..., E k ; E i =|D k (i)| 2 (3) where D k (i) is the signal component of each wavelet packet reconstruction coefficient; A random variable X is used to represent the state characteristics of the uncertain system, and the value of X is x j with a probability p j is: p j = P{X = x j} (4) a certain result of X, information I j is represented as: The information entropy H(X) of X is expressed as: On a certain scale of signal, the total power E of signal is expressed as the energy sum of each node; Let: ε jk (i) = E i / E (7) ∑ k ε jk =1 (8) ε jk The information ratio occupied by each layer signal is denoted as H, and the wavelet packet energy entropy H is jk is: where N is the signal length, H jk is the kth wavelet packet energy entropy of the signal.
4. The few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance according to claim 3, characterized in that, In step 3, virtual feature extraction, general feature is improved from CNN network, containing convolution layer, pooling layer, activation function and full connection layer; adopt 1D CNN network in time series signal processing, output as shown in formula (10): where k w and b w are weights, x is the input; The pooling layer is used for feature screening, as shown in formula (11), to reduce parameters and accelerate convergence process; Data regularization is achieved by Batchnorm layer, and the output is y i , which improves the generalization ability and solves the inconsistency of internal variables. GELu is used as nonlinear activation function, which is a random regular way in line with expectation; Three convolution layers are set to be connected with full connection layer, and the output of full connection layer is: g q = GELu(p3w q +b q ) (13) where g q is the output of the fully connected layer, w q and b q are the weight matrix and bias of the layer.
5. The few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance according to claim 1, characterized in that, Step 5, training strategy based on semi-supervised-self-training, steps are as follows: Step1: all samples in the sample set contain labels, randomly select data to discard labels to obtain new data set; divide data set according to training set and test set in proportion, and train classifier in data set containing labels; Step2: use the trained classifier to predict labels of unmarked data, set threshold, and select data with higher probability as "pseudo label"; Step3: combine data set containing "pseudo label" and "true label" to perform secondary training of classifier; Step4: use the classifier after secondary training to predict the classification of data set containing "true label", and set evaluation standard to evaluate the classifier at this time; If there are still unmarked labels, repeat steps 2-4 until there are no unmarked data sets.
6. The few-shot semi-supervised bearing fault diagnosis method based on domain knowledge guidance according to claim 5, characterized in that, During training iteration, F1 is used as the index of classifier effect: Wherein, precision is the precision rate, and recall is the recall rate.