CN-122016308-A - Bearing fault diagnosis method under data unbalance and variable speed environment by speed robust feature learning and minority class generation

CN122016308ACN 122016308 ACN122016308 ACN 122016308ACN-122016308-A

Abstract

The invention relates to a bearing fault diagnosis method for speed robust feature learning and minority class generation in a data imbalance and variable speed environment, which belongs to the technical field of bearing fault diagnosis and comprises the following steps of S1, mapping an original vibration signal to a potential space with unchanged state by using an encoder based on variable mutual information VIB, realizing fault key feature extraction, S2, generating a sample with discrimination through potential space heavy parameterization based on a feature level sample enhancement mechanism, S3, constructing a multi-task loss function, cooperatively realizing fault feature distinguishability, consistency of generated features and original features and decoupling of running state information, and S4, diagnosing bearing faults.

Inventors

QIN YI
ZHAO LIJUAN
BIE QIN
MAO YONGFANG
CHEN RENXIANG

Assignees

重庆大学

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. A bearing fault diagnosis method under the conditions of data unbalance and speed change by speed robust feature learning and minority class generation is characterized by comprising the following steps: s1, mapping an original vibration signal to a potential space with unchanged state by using an encoder based on variable mutual information VIB to realize fault key feature extraction; s2, generating a sample with discriminant force through potential space heavy parameterization based on a feature level sample enhancement mechanism; S3, constructing a multi-task loss function, and cooperatively realizing the distinguishability of fault features, the consistency of generated features and original features and the decoupling of running state information; And S4, diagnosing bearing faults.
2. The method for diagnosing bearing faults under the conditions of unbalanced data and variable speed according to the speed robust feature learning and minority class generation of claim 1 is characterized in that in the step S1, VIB introduces variation approximation to convert a mutual information objective function into a form which can be optimized by a deep neural network; In the VIB process, the conditions are distributed Parameterized by a neural network, it is generally assumed that it obeys a gaussian distribution: Wherein the method comprises the steps of And Respectively representing the average value and standard deviation of the encoder network output; the mutual information between the input and the feature is expressed as: Due to edge distribution Is difficult to obtain, and VIB adopts simple prior distribution Approximation is performed to obtain an upper bound variation estimate of mutual information: The mutual information between the features and the tags is expressed as: Wherein the method comprises the steps of Is label entropy by introducing discrimination model To approximate the condition distribution Obtaining mutual information variation lower bound: Finally, the optimization objective of the variation mutual information VIB is expressed as: 。
3. The method for diagnosing bearing faults in a data imbalance and variable speed environment by speed robust feature learning and minority class generation according to claim 1, wherein in step S1, the original vibration signal is mapped to a potential space with unchanged state by using an encoder based on variable mutual information VIB, so as to realize fault key feature extraction, and the method comprises the following steps: Using a VIB-based encoder to convert the original vibration signal Mapping to potential space, the encoder extracts a compact representation and generates VIB parameters, including the mean And variance of And describing Gaussian distribution in the state-invariant hidden space.
4. The method for diagnosing bearing faults in a data imbalance and variable speed environment by speed robust feature learning and minority class generation as claimed in claim 3, wherein step S1 specifically comprises the following steps: s11, obtaining compact global vectors through three continuous convolution layers : S12, vector quantity Projection to mean value through linear mapping layer And logarithmic variance : Wherein the method comprises the steps of 、、 And The weight and bias of the linear layer respectively; Representation of Is a dimension of (c).
5. The method for diagnosing bearing faults in a data imbalance and variable speed environment by speed robust feature learning and minority class generation according to claim 1, wherein step S2 is based on a feature level sample enhancement mechanism, generating samples with discrimination through potential space heavy parameterization, and comprises the following steps: sampling from the gaussian distribution by VIB re-parameterization to obtain fault-related features ; Simultaneous generation of redundant feature related representations For capturing interference information and for imposing decoupling constraints to prevent infiltration of operating state information ; Parameters of VIB Storing to construct potential feature libraries of each fault class; And generating new distinguishing features through a re-parameterization process based on the potential feature library.
6. The method for diagnosing bearing faults in a data imbalance and variable speed environment by speed robust feature learning and minority class generation according to claim 5, wherein step S2 specifically comprises the steps of: S21: from gaussian distribution by reparameterization techniques of VIB And (3) sampling: S22, classifying each fault All training samples For storage in a feature library: Wherein the method comprises the steps of As a total number of classes of failure, Is of fault type Training samples of (a); s23, calculating the number of various samples and comparing the number with the maximum class scale Comparing to determine minority classes, for each minority class Calculate the number of missing samples By from corresponding class libraries Random sampling Generating synthetic features and applying VIB heavy parameterization techniques: Wherein the method comprises the steps of Representing slave categories Feature library of (a) And, in addition, Representing the current category Sample set A synthetic feature index generated in (a); s24, combining the features With the original features in the current lot Splicing to form enhanced feature set : Wherein the method comprises the steps of Representing the total number of enhanced samples; S25, the enhanced condition-invariant features Then input fault classifier, which is implemented by using fully-connected network The classifier outputs a predicted logic value for each sample : S26, at the same time, by Obtaining a redundant characteristic related to rotational speed or external disturbances : Training is performed by explicitly modeling state-related components And (3) with Applying decoupling constraint to prevent state information from penetrating 。
7. The method for bearing fault diagnosis in a data imbalance and shift environment for speed robust feature learning and minority class generation of claim 1, wherein the multitasking loss function comprises: Applying orthogonality constraints to ensure And (3) with Inter-effective decoupling: The loss of orthogonality is: At the position of Introduces a loss of supervision contrast for a batch Sample of Wherein For the corresponding class label, pairs of Euclidean distances between all samples are first calculated: The positive and negative sample pairs are defined as: Is provided with And Indicating the number of positive and negative pairs, respectively, in the batch, the loss of supervision contrast is expressed as: Wherein the method comprises the steps of Is an interval parameter; applying multi-scale maximum mean difference MMD loss to ensure reconstructed potential features Features relating to original faults Maintaining alignment: wherein the kernel function Defined as a multi-scale radial basis function: finally, cross entropy classification loss function is adopted to monitor fault classification prediction: The total multitasking loss function contains four terms: 。
8. An electronic device comprising a memory and a processor; the memory is used for storing a computer program; The processor, when executing the computer program, is configured to implement the method for bearing fault diagnosis in a data imbalance and variable speed environment according to any one of claims 1-7.
9. A computer readable storage medium, wherein a computer program is stored on the storage medium, which when executed by a processor, implements the method for bearing fault diagnosis in a data imbalance and shift environment according to any one of claims 1-7.
10. A computer program product comprising a computer program which when executed by a processor implements the method for bearing fault diagnosis in a data imbalance and shift environment for speed robust feature learning and minority class generation as claimed in any one of claims 1 to 7.

Description

Bearing fault diagnosis method under data unbalance and variable speed environment by speed robust feature learning and minority class generation Technical Field The invention belongs to the technical field of bearing fault diagnosis, and relates to a bearing fault diagnosis method under the conditions of unbalanced data and variable speed by speed robust feature learning and minority generation. Background As a critical component of a rotating machine, the health of the rolling bearing directly affects the operational reliability of the device. Therefore, timely and accurate state monitoring and fault diagnosis have important significance for guaranteeing the safe operation of equipment and prolonging the service life. Vibration signals contain rich information of the health state of equipment, and an intelligent fault diagnosis method based on vibration has achieved remarkable results. In practical engineering application, the equipment is operated under a variable speed working condition, and distribution characteristics of vibration signals in different operation states can be changed obviously, so that difficulty in fault diagnosis is increased. Meanwhile, fault samples in actual production often have the problem of data unbalance, and data of certain fault types are relatively scarce, so that a diagnosis model is difficult to fully learn few types of characteristics. Therefore, the research on the problem of data unbalance under the variable speed working condition has important theoretical and practical significance. The method for solving the data unbalance is mainly divided into three types, namely a data hierarchy method, a model hierarchy method and a weight strategy. Among the data-level techniques, the synthetic minority class oversampling technique (SMOTE) is the most representative, which mitigates the class imbalance problem by interpolating minority class instances in feature space to generate new samples. In addition, data enhancement may be performed based on a simulated kinetic model or generation of the countermeasure network (GANs). But these methods are mainly applicable to constant rotational speed or fixed conditions. At the model level, jia et al propose a Depth Normalized Convolutional Neural Network (DNCNN) to solve the problem of fault diagnosis under rolling bearing imbalance sample conditions. Wu et al introduced a model independent framework (MAF) designed specifically for class imbalance fault diagnosis. Weight strategies such as focus loss (FCL) and label distribution aware margin loss (LDL) also often alleviate the class imbalance problem by adjusting the loss function. Although the above approach achieves some effect in alleviating data imbalance, most research has focused on constant rotational speed or single operating conditions. Under variable speed working conditions, vibration signals often exhibit significant non-stationarity and cross-working condition distribution shifts, resulting in difficulty in adapting to different operating conditions by methods that rely on sample generation or static distribution assumptions. Furthermore, during variable speed conditions, the sample distribution for different fault types and their corresponding conditions is often more unbalanced. The number of the minority samples is limited, and the characteristic distribution also deviates significantly along with the change of the rotating speed. The existing method still faces significant challenges when simultaneously coping with working condition changes and class unbalance, and is difficult to ensure diagnostic performance. Disclosure of Invention In view of the above, an object of the present invention is to provide a bearing fault diagnosis method for speed robust feature learning and minority class generation in a data imbalance and variable speed environment. In order to achieve the above purpose, the present invention provides the following technical solutions: a bearing fault diagnosis method for speed robust feature learning and minority class generation in a data imbalance and variable speed environment comprises the following steps: s1, mapping an original vibration signal to a potential space with unchanged state by using an encoder based on variable mutual information VIB to realize fault key feature extraction; s2, generating a sample with discriminant force through potential space heavy parameterization based on a feature level sample enhancement mechanism; S3, constructing a multi-task loss function, and cooperatively realizing the distinguishability of fault features, the consistency of generated features and original features and the decoupling of running state information; And S4, diagnosing bearing faults. In step S1, VIB introduces variation approximation, and converts the mutual information objective function into a form which can be optimized through a deep neural network; In the VIB process, the conditions are distributed Parameterized by a neural network, it