CN-121256224-B - Few-sample bearing fault migration diagnosis method based on multi-working condition supervision contrast learning

CN121256224BCN 121256224 BCN121256224 BCN 121256224BCN-121256224-B

Abstract

The invention provides a few-sample bearing fault migration diagnosis method based on multi-working condition supervision, comparison and learning, which comprises the steps of taking vibration signals with a large number of labels under a plurality of different working conditions of a bearing as source domain data and taking vibration signals which are different from the working conditions of the source domain data and have only a small number of labels as target domain data; the method comprises the steps of obtaining feature representation and prediction probability distribution of each fault category according to a pre-constructed contrast learning classification model, source domain data and target domain data, respectively calculating supervision contrast loss and cross entropy loss, weighting and combining the supervision contrast loss and the cross entropy loss to obtain a total loss function, and obtaining a trained contrast learning training classification model through a self-adaptive learning rate optimization algorithm and back propagation updating model parameters according to the combined total loss function. The invention uses a plurality of different source domain data to extract features through an improved supervised contrast loss function, and optimizes model parameters in combination with a cross entropy loss function to diagnose the target domain of the scarce data.

Inventors

ZHANG YUE
CHEN XINYE
JIANG FEI
LAI JIE
XU CHENYANG

Assignees

东莞理工学院

Dates

Publication Date: 20260508
Application Date: 20250902

Claims (7)

1. The utility model provides a few sample bearing fault migration diagnostic method based on multi-condition supervision contrast study which characterized in that, the method includes: Taking at least two vibration signals of the bearing under different working conditions with a large amount of tag information as source domain data, and taking the vibration signals of the bearing under the working conditions with a small amount of tag information which are different from the working conditions of the source domain data as target domain data, wherein the tag information comprises fault information and working condition information; Inputting the labeled source domain data and target domain data into a feature extractor according to a pre-constructed initial comparison learning classification model to obtain feature representation; According to a global sample pair and local sample pair strategy pre-constructed for source domain data and target domain data, constructing positive and negative sample pairs based on feature representation, calculating supervision comparison loss, calculating cross entropy loss according to prediction probability distribution and real labels of each fault class, and performing weighted combination on the supervision comparison loss and the cross entropy loss to obtain a total loss function; according to the combined total loss function, updating model parameters through a self-adaptive learning rate optimization algorithm and back propagation to optimize the model, so that the loss value is reduced to be converged, and a trained contrast learning training classification model is obtained; inputting the unlabeled target domain data into a trained contrast learning training classification model to obtain a prediction result; the pre-constructed global sample pair and local sample pair strategy comprises the following steps: Taking similar fault samples of all working conditions as positive samples, and taking different fault samples of all working conditions as negative samples to construct a global sample pair; Taking the similar fault samples in the same working condition as positive samples, and taking all the similar fault samples in other working conditions as negative samples to construct local sample pairs; The total loss function comprises a supervision comparison loss function and a cross entropy loss function, wherein the supervision comparison loss function comprises a global comparison loss function and a local comparison loss function which are combined in a weighting manner; The global comparison loss function is used for calculating the feature similarity of similar fault samples and the feature difference of heterogeneous fault samples in all working conditions based on the global sample pairs; calculating the feature similarity of similar fault samples in the same working condition and the feature difference of similar fault samples in other working conditions based on the local sample pairs; The total loss function comprises a supervision contrast loss function and a cross entropy loss function, wherein L=lambda.L con +σ·L cro , L con represents the supervision contrast loss function, and L cro represents the cross entropy loss function; The formula of the supervision contrast loss function is as follows: , Wherein L con is an integral supervision contrast loss function, w represents the number of working conditions, namely the sum of the number of working conditions of a source domain and the number of working conditions of a target domain, alpha is a super parameter, the weight of the global contrast loss and the local contrast loss is determined, L global is a global contrast loss function, and L locak,k is a local contrast loss function.
2. The method of claim 1, wherein 50% -80% of the data in each working condition of the source domain data is divided as a training sample set, the rest is a test sample set, each working condition in the training sample comprises all fault types, K samples are randomly selected from each fault type of the target domain data as training sample data, and K is any one of 0,1, 2 or 4.
3. The method of claim 2, wherein after determining the training samples, the method further comprises: Dividing each continuous vibration signal in the training sample into a plurality of small fragments by adopting an overlapping sliding window method to obtain a corresponding number of samples; And (3) respectively carrying out standardization processing on each small section of vibration signal data after segmentation by adopting a standardization processing formula to obtain standardized sample data, wherein the method comprises the following steps of: Wherein x is the vibration signal data of each small segment after being divided, mu is the average value of the current data segment, sigma is the standard deviation of the current data segment, and z is the standardized data.
4. The method of claim 1, wherein the global contrast loss function L global is formulated as follows: Where sim (z i ,z p ) represents the cosine similarity ,sim(z i ,z p )=z i z p /||z i ||||z p ||;sim(z i ,z n ) between the eigenvector z i of anchor sample i and the eigenvector z p of sample p, and the cosine similarity between the eigenvector z i of anchor sample i and the eigenvector z n of sample n; feature vector z i representing anchor sample i and sample Feature vectors of (a) Cosine similarity between samples, τ represents a temperature parameter, pos i is a positive sample set of sample i, neg i is a negative sample set of sample i, and N total is the total number of samples; The formula of the local contrast loss function L local,k is as follows: Wherein k represents any working condition, N k represents the total number of samples of the working condition k, P (l) represents a positive sample set of samples l, all l represents a sample set of the same class in all working conditions of the samples l, sim (z l ,z p″ ) represents cosine similarity between a feature vector z l of the anchor sample l and a feature vector z p″ of the sample P ''; Feature vector z l representing anchor sample l and sample Feature vectors of (a) Cosine similarity between them; The calculation formula of the cross entropy loss function L cro is as follows: Wherein x i represents a predicted sample, y i represents a real label corresponding to the sample, and p (x i ) represents probability distribution output by the classifier, namely, the predicted probability distribution output by the contrast learning training classification model.
5. The method of any one of claims 1 to 4, wherein the contrast learning training classification model includes a convolutional neural network CNN module and a feature extractor of a bi-directional LSTM module capturing contextual information of the time series data and a classifier based on a multi-layer perceptron MLP.
6. A multiple-condition supervised contrast learning-based few-sample bearing fault migration diagnostic system, the system comprising: The data dividing module is used for taking at least two vibration signals of the bearing under different working conditions with a large amount of tag information as source domain data, and taking the vibration signals of the bearing under the working conditions with only a small amount of tag information different from the working conditions of the source domain data as target domain data, wherein the tag information comprises fault information and working condition information; The probability prediction module is used for inputting the labeled source domain data and target domain data into the feature extractor according to the pre-constructed initial comparison learning classification model to obtain feature representation; The loss function calculation module is used for constructing positive and negative sample pairs based on characteristic representation according to a global sample pair strategy and a local sample pair strategy which are constructed in advance for source domain data and target data, calculating supervision comparison loss, calculating cross entropy loss according to prediction probability distribution and real labels of each fault class, and carrying out weighted combination on the supervision comparison loss and the cross entropy loss to obtain a total loss function; the training module is used for updating model parameters through a self-adaptive learning rate optimization algorithm and back propagation according to the combined total loss function, optimizing the model, reducing the loss value to convergence, and obtaining a trained contrast learning training classification model; the diagnosis module is used for inputting the unlabeled target domain data into a trained contrast learning training classification model to obtain a prediction result; the pre-constructed global sample pair and local sample pair strategy comprises the following steps: Taking similar fault samples of all working conditions as positive samples, and taking different fault samples of all working conditions as negative samples to construct a global sample pair; Taking the similar fault samples in the same working condition as positive samples, and taking all the similar fault samples in other working conditions as negative samples to construct local sample pairs; The total loss function comprises a supervision comparison loss function and a cross entropy loss function, wherein the supervision comparison loss function comprises a global comparison loss function and a local comparison loss function which are combined in a weighting manner; The global comparison loss function is used for calculating the feature similarity of similar fault samples and the feature difference of heterogeneous fault samples in all working conditions based on the global sample pairs; calculating the feature similarity of similar fault samples in the same working condition and the feature difference of similar fault samples in other working conditions based on the local sample pairs; The total loss function comprises a supervision contrast loss function and a cross entropy loss function, wherein L=lambda.L con +σ·L cro , L con represents the supervision contrast loss function, and L cro represents the cross entropy loss function; The formula of the supervision contrast loss function is as follows: , Wherein L con is an integral supervision contrast loss function, w represents the number of working conditions, namely the sum of the number of working conditions of a source domain and the number of working conditions of a target domain, alpha is a super parameter, the weight of the global contrast loss and the local contrast loss is determined, L global is a global contrast loss function, and L locak,k is a local contrast loss function.
7. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 5.

Description

Few-sample bearing fault migration diagnosis method based on multi-working condition supervision contrast learning Technical Field The embodiment of the disclosure relates to the technical field of mechanical equipment fault diagnosis, in particular to a few-sample bearing fault migration diagnosis method, system and computer readable storage medium based on multi-working condition supervision and contrast learning. Background The bearing is a key component widely used in mechanical equipment, has the functions of reducing friction, supporting rotating components and the like, and is widely applied to the fields of high-speed railways, aerospace and the like. The working state of the bearing directly affects the normal operation of the equipment, and once the equipment fails, the mechanical production is forced to stop, so that economic loss is caused. Therefore, the research of the deep learning algorithm capable of efficiently and accurately diagnosing the health state of the bearing has important practical significance. The traditional bearing fault diagnosis method mainly relies on signal processing and feature extraction technology, relevant fault features are extracted through analysis of bearing vibration signals, and fault identification is performed through a classification algorithm. Such methods typically require a large number of marked training samples and most focus on fault diagnosis problems under a single operating condition. However, in actual operating situations, the device needs to operate under different operating conditions, and the training effect of the conventional fault diagnosis algorithm is significantly reduced for the working condition data with only a small number of marked samples. Therefore, conventional diagnostic algorithms are difficult to apply in practical industrial sites. In order to solve the problem of cross-working condition fault diagnosis, transfer learning is becoming an important technical means gradually. The core idea of the transfer learning is to transfer knowledge learned from the source working condition to the target working condition, so that the requirement of marking data on the target working condition is reduced. However, most migration methods only use a single working condition as a source domain, and most focus on global feature sharing between a single source working condition and a target working condition, lack effective adjustment of decision boundaries, and ignore possible local feature differences between different working conditions. This results in these methods being difficult to cope with the variability between conditions when dealing with cross-condition fault diagnosis, resulting in lower diagnostic accuracy. ‌ Disclosure of Invention An object of an embodiment of the present disclosure is to provide a method, a system and a computer-readable storage medium for diagnosing a few-sample bearing fault migration based on multi-working condition supervised contrast learning, so as to solve the foregoing problems in the prior art. In order to achieve the above objective, the technical solution adopted in the embodiments of the present disclosure is as follows: In one aspect, an embodiment of the present disclosure provides a method for diagnosing a fault migration of a bearing with a small sample based on multi-working condition supervision and contrast learning, where the method includes: Taking at least two vibration signals of the bearing under different working conditions with a large amount of tag information as source domain data, and taking the vibration signals of the bearing under the working conditions with a small amount of tag information which are different from the working conditions of the source domain data as target domain data, wherein the tag information comprises fault information and working condition information; Inputting the labeled source domain data and target domain data into a feature extractor according to a pre-constructed initial comparison learning classification model to obtain feature representation; According to a global sample pair and a local sample pair strategy which are pre-constructed on source domain data and target data, positive and negative sample pairs are constructed based on feature representation, supervision comparison loss is calculated, cross entropy loss is calculated according to prediction probability distribution and real labels of each fault class, and the supervision comparison loss and the cross entropy loss are combined in a weighting mode to form a total loss function; according to the combined total loss function, updating model parameters through a self-adaptive learning rate optimization algorithm and back propagation to optimize the model, so that the loss value is reduced to be converged, and a trained contrast learning training classification model is obtained; And inputting the unlabeled target domain data into a trained contrast learning training classification model to