CN-122021287-A - Bearing spalling size prediction method based on manifold self-adaptive migration network

CN122021287ACN 122021287 ACN122021287 ACN 122021287ACN-122021287-A

Abstract

A bearing spalling size prediction method based on manifold self-adaptive migration network comprises the steps of constructing a two-stage multi-scale manifold self-adaptive migration network model for training, training a multi-scale residual expansion convolution network by utilizing dynamic simulation full life cycle data to learn a mapping relation between vibration signals and spalling sizes to obtain a pre-training feature extractor, inputting a real bearing vibration signal serving as a target domain into the network model together with a source domain simulation signal for counterlearning, distinguishing whether an input feature comes from a source domain or a target domain through a domain discriminator, synchronously applying dynamic direction consistency constraint, wherein the constraint requires that the source domain features extracted in the domain self-adaptive stage keep consistent with the evolution direction of corresponding source domain features extracted by the pre-training feature extractor in a high-dimensional space, and training the network model by optimizing a joint loss function. The problem of bearing life cycle data is not enough, model training difficulty is solved.

Inventors

JIANG CHENXING
Yang Cangjia
WANG XI
WANG ZHIZHENG
CHEN YUXUAN
CAO HAIFENG
TANG ZHILONG
SONG ZHENGKAI

Assignees

厦门大学

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (10)

1. The bearing flaking size prediction method based on the manifold self-adaptive migration network is characterized by comprising the following steps of: S10, establishing a dynamic model of a rotor-bearing system, simulating vibration response under different peeling sizes, and generating a bearing vibration simulation signal covering the whole life cycle; s20, collecting a bearing vibration real signal, and carrying out signal preprocessing matched with a simulation signal on the real signal; s30, constructing a two-stage multi-scale manifold self-adaptive migration network model for training: A pre-training stage, namely training a multi-scale residual error expansion convolution network by taking a bearing vibration simulation signal as a source domain sample to learn the mapping relation between the vibration signal and the peeling size so as to obtain a pre-training feature extractor; A domain self-adaptation stage, namely taking the bearing vibration real signal preprocessed in the step S20 as a target domain, inputting the target domain and a source domain simulation signal into a network model together for countermeasure learning, distinguishing whether input features come from a source domain or a target domain through a domain discriminator, and synchronously applying dynamic direction consistency constraint, wherein the constraint requires that the source domain features extracted in the domain self-adaptation stage keep consistent with the evolution directions of the corresponding source domain features extracted by a pre-training feature extractor in a high-dimensional space; S40, inputting a bearing vibration signal by using the trained network model, and predicting the corresponding bearing peeling size.
2. The method for predicting bearing flaking dimension based on manifold adaptive migration network of claim 1 wherein said phase two dynamic direction consistency constraint is implemented by computing cosine similarity of feature vectors extracted in the domain adaptation phase of source domain samples and corresponding feature vectors extracted by a pre-trained feature extractor.
3. A method for predicting bearing flaking dimension based on manifold adaptive migration network as claimed in claim 2, wherein in the pre-training stage, the fixed features obtained by fixing the source domain sample features extracted by the pre-training feature extractor are defined as The feature vector iteratively extracted in the domain adaptation stage is as follows The loss function of the dynamic direction consistency constraint is expressed as: ; In the formula, A loss function representing a dynamic direction consistency constraint, Expressed as the amount of source domain samples, For uncertainty-based adaptive weights, & represents a vector dot product; Representing the L2 norm.
4. The method for predicting the peeling dimension of a bearing based on a manifold adaptive migration network according to claim 3, wherein the adaptive weights Dynamically adjusting the predictive entropy value of the current sample according to the domain discriminator, namely reducing the weight when the discriminator cannot distinguish the sample domain To avoid overfitting to the aligned features or noise.
5. The method for predicting bearing flaking dimension based on manifold adaptive migration network of claim 3 wherein the countering loss in countering learning is represented by: ; In the formula, Representing the countermeasures loss function, For the target domain sample size, Is a domain label (source domain is 0, target domain is 1), For the feature vector output by the feature extractor, Is a domain arbiter.
6. The method for predicting the bearing flaking dimension based on the manifold adaptive migration network of claim 5, wherein the multi-scale residual expansion convolution network enables the network to capture high-frequency transient impact local features related to defects and low-frequency trend global features related to degradation processes in a vibration signal in parallel in a single layer by setting an exponentially increasing expansion rate.
7. A method for predicting bearing flaking dimension based on manifold adaptive migration network as claimed in claim 6, wherein in the pre-training stage, regression prediction loss function is used As a mean square error of source domain prediction, it is defined as: ; In the formula, For the regression predictor, for inputting the extracted high-dimensional vector, outputting the predicted peeling size, For the i-th source domain sample, For the corresponding true scale-off size, G f is a feature extractor for extracting high-dimensional feature vectors from the input signal.
8. The method for predicting the bearing flaking dimension based on the manifold adaptive migration network of claim 7, wherein the total loss function L is: ; In the formula, 、 The weight coefficients for the countermeasures and consistency losses, respectively.
9. The method for predicting the bearing flaking dimension based on the manifold adaptive migration network of claim 1, wherein the step S10 comprises the following steps: S11, establishing a rotor-bearing system dynamics model based on a finite element method, wherein the contact force between a bearing rolling body and a raceway adopts nonlinear Hertz contact theory modeling; S12, in a rotor-bearing system dynamics model, simulating stress abrupt change and unloading processes when the rolling bodies enter and exit the spalling pits by introducing a time-varying displacement excitation function; s13, generating a simulation signal sequence reflecting vibration responses under different peeling sizes by continuously changing peeling pit size parameters in the time-varying displacement excitation function.
10. The method for predicting the bearing flaking dimension based on the manifold adaptive migration network of claim 1, wherein the preprocessing in step S20 comprises the following steps: S21, carrying out band-pass filtering on the simulation signal and the real signal, and extracting an envelope spectrum of the impact component by using Hilbert transformation so as to highlight periodic impact components caused by defects; S22, according to the rotating speed pulse signal, aiming at the rotating speed variable working condition, carrying out order tracking by utilizing the rotating speed pulse signal, resampling a time domain non-stationary signal to an angular domain through an interpolation algorithm, and eliminating frequency blurring caused by rotating speed fluctuation; S23, carrying out phase locking segmentation on the angular domain signals according to the theoretical characteristic period of the rolling body passing through the defects, which is calculated according to the geometric parameters of the bearing, so as to ensure that each signal segment contains a complete impact event.

Description

Bearing spalling size prediction method based on manifold self-adaptive migration network Technical Field The invention relates to the technical field of bearing health management and intelligent fault diagnosis, in particular to a bearing spalling size prediction method based on a manifold self-adaptive migration network. Background The rolling bearing is used as a core supporting component of the rotary machine, and the running state of the rolling bearing directly determines the stability and the service life of the whole mechanical system. According to industry statistics, more than half of the rotating machine failures originate from rolling bearings, with fatigue spalling of the raceway surfaces being the predominant failure mode. Existing life prediction methods are roughly divided into three categories, physical models, statistical models, and data-driven models. The physical model is usually based on a fatigue crack propagation theory (such as a Paris-Erdogan model and a Paris formula), and can accurately describe a crack development process under certain conditions, but the physical model is difficult to completely accord with the actual working condition due to uncertainty factors such as lubrication, load impact, rotation speed fluctuation, noise interference and the like in the actual working condition. Statistical models such as Wiener processes (Wiener processes, also called brownian motion) and Gamma process methods can describe degradation uncertainty through probability distribution, but require stronger expert priori knowledge, and are difficult to apply in engineering. With the development of artificial intelligence and sensor technology, data driven methods have gradually become a research hotspot. The method has the advantage that effective features can be automatically extracted directly from original signals such as vibration and the like. For example, CNN can capture local patterns, and RNN/LSTM can model time series dependencies. However, such methods typically rely on a large amount of "full life cycle" degradation data as a training set, whereas in practical industrial environments, the equipment is often not run to complete failure for safety reasons, and thus data for severe degradation phases is difficult to obtain. In addition, the data distribution acquired under different working conditions (the difference of rotating speed and load) is different, so that the training of the traditional deep learning model is not matched with the application scene, and the prediction accuracy is greatly reduced. Therefore, how to solve the two key problems of insufficient data and cross-domain distribution difference has become a bottleneck problem to be broken through in the bearing life prediction field. Disclosure of Invention The invention aims to provide a bearing peeling size prediction method based on a manifold self-adaptive migration network, which solves the technical problem of how to solve the problem that the bearing peeling size is difficult to predict through a model accurately due to insufficient data of the whole life cycle of a bearing. To achieve the above object, the solution of the present invention is: a bearing flaking size prediction method based on manifold self-adaptive migration network comprises the following steps: S10, establishing a dynamic model of a rotor-bearing system, simulating vibration response under different peeling sizes, and generating a bearing vibration simulation signal covering the whole life cycle; s20, collecting a bearing vibration real signal, and carrying out signal preprocessing matched with a simulation signal on the real signal; s30, constructing a two-stage multi-scale manifold self-adaptive migration network model for training: A pre-training stage, namely training a multi-scale residual error expansion convolution network by taking a bearing vibration simulation signal as a source domain sample to learn the mapping relation between the vibration signal and the peeling size so as to obtain a pre-training feature extractor; A domain self-adaptation stage, namely taking the bearing vibration real signal preprocessed in the step S20 as a target domain, inputting the target domain and a source domain simulation signal into a network model together for countermeasure learning, distinguishing whether input features come from a source domain or a target domain through a domain discriminator, and synchronously applying dynamic direction consistency constraint, wherein the constraint requires that the source domain features extracted in the domain self-adaptation stage keep consistent with the evolution directions of the corresponding source domain features extracted by a pre-training feature extractor in a high-dimensional space; S40, inputting a bearing vibration signal by using the trained network model, and predicting the corresponding bearing peeling size. Further, the dynamic direction consistency constraint of the second stage is realiz