Rolling bearing fault diagnosis method based on ISSPN

Abstract

The invention discloses a rolling bearing fault diagnosis method based on ISSPN, and relates to the technical field of rolling bearing fault detection, and the method comprises the steps: collecting bearing vibration acceleration signals in different health states, and constructing a rolling bearing fault data set; a rolling bearing fault data set is divided into different tasks through a meta-learning method, and each task is subdivided into a support set and a query set; for each task, constructing an initial prototype for each category through an improved semi-supervised prototype network on the basis of marked samples in the support set, and performing fine adjustment on the initial prototype by using unmarked samples to obtain a purified prototype; and on the basis of the purified prototype, samples in the query set are classified, back propagation is carried out through a total loss function, and updating of model parameters is completed. Therefore, by the adoption of the rolling bearing fault diagnosis method based on the ISSPN, rolling bearing fault diagnosis of small samples is achieved, the detection precision is improved, and meanwhile the complexity of the model is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rolling bearing fault detection, in particular to a rolling bearing fault diagnosis method based on ISSPN. BACKGROUND

[0002] In actual engineering scenarios, complete monitoring data samples are the basis for diagnosing mechanical equipment faults. However, in mechanical state monitoring, the collected data usually has limitations, so it is difficult to identify fault information. Due to the limited number of samples, the selected samples are difficult to well represent various situations in the actual operation of bearings, and there is a risk of sampling bias. In actual monitoring, the collected data only covers part of the common fault types, and there is a lack of sufficient data support for some rare but serious fault conditions, thereby affecting the generalization ability of the diagnosis model.

[0003] With the rapid development of modern intelligent technology, many deep learning models provide a large number of solutions for fault diagnosis, but there are still some defects. In general, deep learning models need a large amount of labeled training data to ensure convergence and good classification performance. However, due to the complexity and diversity of the actual working environment, it is still difficult to obtain rich data.

[0004] When the sample is extremely missing, the traditional transfer learning will inevitably lead to overfitting, because one or a few labeled instances are not enough to represent the data distribution of the new class. At the same time, the collected data is mostly unlabeled, which leads to more cost and time wasted in labeling the unlabeled samples one by one. This makes the deep learning method perform well in the big data scenario, but often faces the dilemma of insufficient data for small sample data. In addition, the transfer learning effect of the pre-trained model in small sample bearing fault diagnosis is not ideal, and a large amount of adjustment and optimization is needed for specific problems, which increases the difficulty of designing and implementing the diagnosis method. SUMMARY

[0005] The purpose of the present application is to provide a rolling bearing fault diagnosis method based on ISSPN, which can be applied to small sample rolling bearing fault diagnosis, improve the detection accuracy and reduce the complexity of the model.

[0006] To achieve the above purpose, the present application provides a rolling bearing fault diagnosis method based on ISSPN, comprising the following steps:

[0007] S1, collect bearing vibration acceleration signals under different health states, construct a rolling bearing fault data set, including labeled samples and unlabeled samples;

[0008] S2, dividing the rolling bearing fault dataset into different tasks by the N-way K-shot method, each task consisting of N classes, each class containing K samples; meanwhile, each task is subdivided into a support set and a query set;

[0009] S3, for each task, based on the labeled samples in the support set, constructing an initial prototype for each class by the improved semi-supervised prototype network, and continuously fine-tuning the initial prototype using unlabeled samples to obtain a purified prototype;

[0010] S4, classifying the samples in the query set based on the purified prototype, and completing the update of the model parameters through back propagation of the total loss function.

[0011] Preferably, in S3, the improved semi-supervised prototype network includes a convolution operation block composed of several groups of convolution blocks, batch normalization blocks, activation function ReLU, and pooling layers; the pooling layer is used for feature selection and dimension reduction operation.

[0012] Preferably, in S3, the improved semi-supervised prototype network introduces an effective channel attention mechanism after the convolution operation block, which is used to extract the feature vector of the sample and further construct the initial prototype of the sample, thereby enhancing the feature extraction capability of the small sample.

[0013] Preferably, in S3, continuously fine-tuning the initial prototype using unlabeled samples includes determining the probability of each unlabeled sample belonging to different classes, selecting the class corresponding to the maximum probability as the class to which the unlabeled sample belongs, and fine-tuning the initial prototype of the corresponding class using the unlabeled sample, as follows:

[0014] First, assign probabilities of labeled samples and unlabeled samples belonging to different classes, as follows:

[0015]

[0016] In the formula, S l , S u represent labeled samples and unlabeled samples, respectively; p i,d represents the probability of x l in the labeled sample S i belonging to class d, i.e. p(y=d|x i ∈S l ); represents the probability of x u in the unlabeled sample S i belonging to class d, i.e. p(y=d|x i ∈S u ); f e () represents a nonlinear feature mapping function; P d', where d' represents the prototype of different categories in the query set Q; E dist represents the Euclidean distance function;

[0017] Then, based on p i,d and the weight coefficients of the labeled samples and unlabeled samples in each category are calculated, and the weight coefficients are used to distribute the initial prototypes of each category to obtain the purified prototypes, as follows:

[0018]

[0019] In the formula, q i,d , are the weight coefficients of the labeled samples and unlabeled samples, respectively, and i p i,d , respectively represent the sum of the probabilities of all labeled samples and all unlabeled samples, represents the purified prototype.

[0020] Preferably, in S4, the total loss function is used to measure the distance between the sample prototypes of the query set and the purified prototypes, as follows:

[0021]

[0022] In the formula, L θ (Q) is the total loss function, T q is the total number of samples in the query set Q, (x i , y i ) is the sample in the query set Q, and p(y=y i |x i ) represents the sample probability of the query set.

[0023] Therefore, the rolling bearing fault diagnosis method based on the above-mentioned ISSPN has the following technical effects:

[0024] (1) The effective channel attention mechanism is introduced in the present application, so that the model can obtain more effective features under limited data, avoid the side effects caused by dimension reduction on channel attention prediction, and capture cross-channel information in an extremely lightweight manner.

[0025] (2) Based on the semi-supervised learning strategy, the present application adjusts the initial prototype using unlabeled data, which can effectively solve the position deviation problem caused by the limited labeled support data. Through the unlabeled samples, the prototype can be slightly guided, and the information can be better aggregated into the initial prototype, so that the effect is significantly improved, and the rolling bearing fault diagnosis under small samples is realized.

[0026] The technical solutions of the present application are described in further detail below with reference to the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flow chart of a rolling bearing fault diagnosis method based on ISSPN;

[0028] Figure 2 is a schematic diagram of a prototype network model in an embodiment of a rolling bearing fault diagnosis method based on ISSPN;

[0029] Figure 3 is a schematic diagram of a prototype purification in an embodiment of a rolling bearing fault diagnosis method based on ISSPN, wherein (a) is the construction of an initial prototype, and (b) is the prototype purification process;

[0030] Figure 4 is a schematic diagram of a COBs-ECA module framework in an embodiment of a rolling bearing fault diagnosis method based on ISSPN. DETAILED DESCRIPTION

[0031] The present application can be explained in further detail by the following examples, and the purpose of disclosing the present application is to protect all variations and improvements within the scope of the present application, and the present application is not limited to the following examples.

[0032] See Figure 1 , the present application provides a rolling bearing fault diagnosis method based on ISSPN, which comprises constructing an improved semi-supervised prototype network (Improved Semi-supervised Prototype Network, ISSPN), which is based on the meta-learning paradigm and combines the prototype network and the effective channel attention mechanism to map the training data into the feature space, further learns the similarity between two samples through distance metric, and maps them into different classes, and further utilizes the semi-supervised learning strategy to fully utilize the unlabeled data to realize the adjustment and purification of the data prototype, and finally realizes the correct clustering of various faults. The network structure of ISSPN involves three parts of prototype network, effective channel attention mechanism and prototype purification.

[0033] In this embodiment, bearing vibration acceleration signals under different health states are collected, containing multiple groups of 2048 data points generated by time series signals, i.e. i ∈R 1×2048 , a rolling bearing fault data set is constructed. Based on this data set, a typical meta-learning model, ProtoNet, is selected as the basis, and a feature function is used to map the rolling bearing fault data into different categories, and the similarity between each sample is learned through distance metric. See Figure 2, the prototype network follows a scenario training strategy, and divides the rolling bearing fault dataset into multiple tasks using the N-way K-shot method, each task consisting of N categories, each category containing K samples. In standard N-way K-shot classification, all query samples are forced to be assigned to one of the N known categories, but in actual applications, special cases often lead to misjudgment. Therefore, this embodiment introduces a rejection mechanism based on Euclidean distance. Specifically, for category d, the Euclidean distance between the support set samples and the corresponding prototype is calculated, and the mean μ d and standard deviation σ d are calculated d , and the rejection threshold τ d is set as τ d = μ e + 2σ d . In the inference stage, if the distance between the query sample and its nearest prototype is greater than the threshold, the sample is judged as "unknown class".

[0034] This embodiment selects N = 4, and the category set can be represented as {NC, IF, OF, BF}, NC is normal state, IF is inner ring fault, OF is outer ring fault, and BF is rolling element fault. In the time domain feature, NC signal shows stable low amplitude without sharp peak; IF signal shows sharp periodic pulse with slightly shorter interval; OF signal shows smaller pulse amplitude but more consistent periodicity; BF signal shows many "spike type" short period high energy disturbances, and irregular. In the frequency domain feature, NC signal has no obvious fault frequency component, only rotation frequency and system eigenfrequency; IF signal has obvious inner ring characteristic frequency (BPFI) and its harmonics; OF signal has prominent outer ring characteristic frequency (BPFO) and is affected by load; BF signal shows rolling element frequency (BSF) is relatively fuzzy, accompanied by modulation phenomenon and sideband. In addition, each task is further divided into support set S and query set Q. At the same time, different tasks are divided into meta-training set and meta-test set.

[0035] In the training process of the meta-learning model, a nonlinear feature mapping function f e (·) is learned by the model, which maps the prototypes of each category to a feature space, and in this feature space, prototypes from the same category are closer, while different prototypes are farther apart. Wherein, the prototype of each category is represented by the average value of the mapped feature vectors, and the calculation formula is:

[0036]

[0037] In the formula, P d is the prototype of category d, |∑S i | represents the total size of samples in category d, x i is a sample belonging to category d, y i and x i corresponding label.

[0038] The parameters e of the feature function are trained and updated by the sample of the query set Q under each task, and the classification of the sample is completed by measuring the Euclidean distance between the sample and all prototypes. The prediction probability of the sample belonging to the category d in the query set Q is:

[0039]

[0040] where P d' , d'∈{1,2,…,N} correspond to the prototypes of different categories in all query sets Q; E dist represents the Euclidean distance function, which is defined as:

[0041]

[0042] In the formula, o∈R Z represents the object to which the Euclidean distance function is applied, and subscripts 1 and 2 correspond to different objects, and subscript j represents the dimension.

[0043] In this process, the total loss function is:

[0044]

[0045] In the formula, T q is the total number of samples in the query set Q. By minimizing L θ , the model parameters e are updated.

[0046] Due to the small amount of sample data and the limited number of labeled samples, the initial prototype cannot correctly represent the cluster center. In order to obtain better prototypes, the initial prototypes are adjusted using unlabeled data, which can effectively solve the position deviation problem caused by the limited labeled support data. Through the unlabeled samples, the prototypes can be slightly guided, and the information can be better aggregated into the initial prototypes for prototype purification. Please refer to Figure 3 , the prototype purification process is as follows:

[0047] First, the support set S is divided into two parts, which are labeled samples S l and unlabeled samples S u , which are represented as:

[0048]

[0049] In the formula, s l , s u represent the number of labeled samples and unlabeled samples in the support set S, respectively.

[0050] Second, using the labeled samples S l in the support set, the COBs-ECA module is used to construct the initial prototype Pd ∈R Z , and mark the sample x i The probability p of belonging to category d i,d (i.e. p(y=d|x i ∈S l ))Assigned to:

[0051]

[0052] Where y represents sample x i The corresponding label.

[0053] Similarly, the unlabeled sample S u Medium x i The probability of belonging to category d (i.e. p(y=d|x i ∈S u ))Assigned to:

[0054] Used to measure the unlabeled samples S u Medium x i The similarity between different categories, select the category with the largest probability as the unlabeled sample S u Medium x i The category d to which it belongs is then fine-tuned on the initial prototype of the corresponding category d as follows:

[0055] Using the above p i,d and Define the weight coefficients of the two probabilities:

[0056]

[0057] Where q i,d 、 are the weight coefficients of labeled samples and unlabeled samples, Σ i p i,d 、 Represents the sum of the probabilities of all labeled samples and all unlabeled samples, respectively. Represents the purified prototype.

[0058] Then, for each category, the initial prototype is assigned according to the corresponding weight coefficient to obtain the probability label of the unlabeled prototype as follows:

[0059]

[0060] By using the above method, we can assign probability labels to unlabeled data. By repeating the above steps, we can assign corresponding probability labels to each category d. Then Assign to the initial prototype pd The unlabeled data is aggregated into the existing prototype to obtain more feature information, so that the model ISSPN is suitable for small sample rolling bearing fault diagnosis.

[0061] In addition, considering the problem of bearing fault diagnosis sample quantity, the embodiment introduces an effective channel attention (ECA) in the convolution operation block (COBs), that is, a COBs-ECA module. In the application process, the sample data in the support set (including labeled samples and unlabeled samples) and the query set are mapped to the feature space through the module, and then all channel features are combined to obtain the feature vector of the corresponding sample data after mapping f e (x i ), so as to enhance the feature extraction capability of the model for small samples, avoid information loss caused by dimension reduction, and effectively capture cross-channel interaction. Please refer to Figure 4 The convolution operation block (COBs) is composed of multiple types of COBs, and the COB includes a convolution block, a batch normalization block, an activation function ReLU, and a pooling layer. The batch normalization block accelerates the prototype network in the training process, reduces the displacement or disappearance of the internal covariance, avoids gradient explosion, and realizes the selection of features and the reduction of feature dimensions through the introduction of the pooling layer, thereby reducing the number of parameters and effectively avoiding the occurrence of overfitting. The ECA is established on the basis of squeeze-and-excitation attention (SE), uses global average pooling (GAP) to aggregate convolution features, and replaces the dimension reduction operation of the full connection (FC) in SE. The ECA is arranged after the COBs, and good results can be obtained through only a few parameters, which can significantly improve the performance of the model and has lower model complexity.

[0062] The initial prototype is adjusted by the unlabeled sample prototype to obtain a purified prototype, the samples in the query set Q are classified by using the purified prototype, and the model parameters are updated through the back propagation of the total loss function, so as to effectively solve the position deviation problem caused by the limited labeled support data, cope with the rolling bearing fault diagnosis problem under the unbalanced monitoring sample data, and ensure the accurate clustering of different rolling bearing faults. In one embodiment, the learning rate is dynamically adjusted by using the Adam algorithm at different stages and parameters, so that the model has high convergence speed and good stability.

[0063] Therefore, the rolling bearing fault diagnosis method based on the ISSPN can detect the rolling bearing fault with less bearing fault data, and is especially suitable for the unbalanced data.

[0064] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A rolling bearing fault diagnosis method based on ISSPN, characterized in that: The following steps are involved: S1. Collect bearing vibration acceleration signals under different health states and construct a rolling bearing fault dataset, including labeled samples and unlabeled samples; S2. The rolling bearing fault dataset is divided into different tasks using the N-way K-shot method. Each task consists of N categories, and each category contains K samples. At the same time, each task is subdivided into a support set and a query set. S3. For each task, based on the labeled samples in the support set, an improved semi-supervised prototype network (ISSPN) is used to construct an initial prototype for each category, and the unlabeled samples are used to fine-tune the initial prototype to obtain the purified prototype. Among them, the improved semi-supervised prototype network ISSPN includes a convolution operation block, and an effective channel attention mechanism is introduced after the convolution operation block; Continuously fine-tuning the initial prototype using unlabeled samples includes assigning probabilities that the labeled samples and the unlabeled samples belong to different categories, calculating weight coefficients of the labeled samples and the unlabeled samples in each category, and using the weight coefficients to assign initial prototypes to each category; S4. Based on the purified prototype, classify the samples in the query set and perform backpropagation through the total loss function to complete the update of the model parameters.

2. The rolling bearing fault diagnosis method based on ISSPN according to claim 1, characterized in that: In S2, the N-way K-shot method includes the introduction of a rejection mechanism based on Euclidean distance, which sets the rejection threshold by counting the Euclidean distance between the support set samples and the corresponding prototypes, and calculating the mean and standard deviation; In the inference stage, samples with a value greater than the rejection threshold are classified as unknown classes.

3. The rolling bearing fault diagnosis method based on ISSPN according to claim 1, characterized in that: In S3, the improved semi-supervised prototype network ISSPN includes a convolution operation block consisting of several groups of convolution blocks, a batch normalization block, an activation function ReLU and a pooling layer; the pooling layer is used for feature selection and dimensionality reduction operations.

4. The rolling bearing fault diagnosis method based on ISSPN according to claim 3 is characterized in that: In S3, the improved semi-supervised prototype network ISSPN introduces an effective channel attention mechanism after the convolution operation block to extract the feature vector of the sample, and then construct the initial prototype of the sample to enhance the ability to extract small sample features.

5. The rolling bearing fault diagnosis method based on ISSPN according to claim 1, characterized in that: In S3, the initial prototype is continuously fine-tuned using unlabeled samples, including determining the probability of each unlabeled sample belonging to different categories, selecting the category corresponding to the maximum probability as the category to which the unlabeled sample belongs, and using the unlabeled sample to fine-tune the initial prototype of the corresponding category, as follows: First, assign the probabilities that the labeled and unlabeled samples belong to different categories as follows: Where S l 、S u Represent the labeled samples and unlabeled samples respectively; p i,d Represents the labeled sample S l Medium x i The probability of belonging to category d, that is, p(y=d|x i ∈S l ); Represents the unlabeled sample S u Medium x i The probability of belonging to category d, that is, p(y=d|x i ∈S u );f e ( ) represents the nonlinear feature mapping function; P d' ,d`∈{1,2,…,N} represents the prototypes of different categories in the query set Q; E dist represents the Euclidean distance function; Then, based on p i,d and Calculate the weight coefficients of labeled samples and unlabeled samples in each category, and use the weight coefficients to distribute the initial prototypes of each category to obtain the purified prototypes as follows: Where q i,d 、 are the weight coefficients of labeled samples and unlabeled samples respectively, ∑ i p i,d 、 Represents the sum of the probabilities of all labeled samples and all unlabeled samples, respectively. Represents the purified prototype.

6. The rolling bearing fault diagnosis method based on ISSPN according to claim 1, characterized in that: In S4, the total loss function is used to measure the distance between the query set sample prototype and the purified prototype, as follows: Where, L θ (Q) is the total loss function, T q is the total number of samples in the query set Q, (x i ,y i )∈Q represents the sample in the query set Q, p(y=y i |x i ) represents the sample probability of the query set.

Images