CN-122020269-A - Robot joint flexible bearing fault diagnosis method based on physical priori drive

CN122020269ACN 122020269 ACN122020269 ACN 122020269ACN-122020269-A

Abstract

The invention relates to a robot joint flexible bearing fault diagnosis method based on physical priori driving, which comprises the steps of collecting robot joint flexible bearing vibration signals under different working conditions, constructing a data set comprising a labeled source domain sample set and a label-free target domain sample set, training a domain anti-convolution neural network to obtain a diagnosis model, inputting a target sample to be diagnosed into the diagnosis model which is completed by training, outputting a flexible bearing fault type identification result, introducing the domain anti-convolution neural network into a nuclear selective fusion attention module to adaptively fuse multi-scale feature responses, enhancing stability and discrimination of fault features, constructing a sample instance graph, and enhancing structural relation modeling capacity among samples by combining the multi-receptive field convolution network, and simultaneously introducing physical priori consistency constraints based on fault feature frequency and harmonic neighborhood energy thereof. The invention improves the accuracy, stability and physical interpretability of bearing fault diagnosis under variable working conditions.

Inventors

Zhuang jichao
MAI MINGHUI
LIANG ZHONGWEI
CHEN XIN
CAO YONGJUN
ZHANG CAIXIA
HAN SHOULEI

Assignees

广州大学

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (10)

1. A fault diagnosis method of a robot joint flexible bearing based on physical priori driving is characterized by comprising the steps of collecting vibration signals of the robot joint flexible bearing under different working conditions, constructing a data set comprising a labeled source domain sample set and a label-free target domain sample set, training a domain contrast graph convolution neural network to obtain a diagnosis model, inputting a target sample to be diagnosed into the trained diagnosis model, and outputting a fault type identification result of the flexible bearing; the training of the domain countermeasure convolution neural network comprises the following steps: After preprocessing the flexible bearing vibration signal of the data set, inputting an improved convolution feature extraction network to obtain a fault characterization feature map, wherein a kernel selective fusion attention module is introduced into the improved convolution feature extraction network, and the fault characterization feature map is taken as input to output an enhanced feature representation; Mapping the enhanced feature representation into sample node features, and constructing a sample instance graph according to the similarity among samples by using a graph construction layer; Inputting the structural enhancement feature representation into a classifier to obtain model prediction distribution of a target domain sample, and calculating source domain classification loss; inputting the structural enhancement feature representation into a domain discriminator, carrying out domain discrimination on a sample, and calculating domain alignment loss; Constructing physical prior distribution corresponding to the target domain sample according to fault characteristic frequency in an envelope power spectrum of the target domain sample and energy distribution in a harmonic neighborhood of the fault characteristic frequency; And updating parameters of the domain contrast graph convolutional neural network by using the weighted total loss of the physical consistency loss and other losses to obtain a trained diagnosis model.
2. The method of claim 1, wherein the calculating of the physical consistency loss comprises: , In the formula, In the event of a loss of physical consistency, For the mathematical expectation of the target domain samples, Representing a sample Is used to determine the gating weight of the (c), For the Kullback-Leibler divergence, For a physical a priori distribution, The distribution is predicted for the model.
3. The method of claim 2, wherein the step of adjusting the weight of the physical consistency loss using a training strategy that delays the linear boost when calculating the total loss comprises: , In the formula, Is the total loss; 、、 Respectively classifying loss of a source domain, domain alignment loss and structure alignment loss; And Is a trade-off coefficient; for the weight of the physical consistency loss dynamically changing along with the training round, the calculation formula is as follows: , In the formula, For the current training round of time, For the round in which the physical consistency constraint starts to take effect, For a continuous round of linear weight increase, Is the maximum weight.
4. The method of claim 1, wherein constructing the physical prior distribution corresponding to the target domain samples comprises: Carrying out band-pass filtering on the sample x, obtaining an envelope signal through Hilbert transformation, obtaining an envelope spectrum through fast Fourier transformation, and further obtaining an envelope power spectrum; defining a center frequency for fault characteristic frequencies of each type of basic fault mechanism, and constructing upper and lower bounds of a harmonic neighborhood window; accumulating the energy in each harmonic neighbor to obtain the physical evidence score of the sample x on each basic fault mechanism: , In the formula, For the sample In the basic failure mechanism A physical evidence score on the corresponding harmonic neighborhood; is the first Envelope power spectrum values at the individual frequency points; represents the first An indication function of whether the frequency point falls in the h-order harmonic neighborhood window of the t-th fault characteristic frequency ,H、 The total harmonic total order and the total fundamental failure mechanism are respectively; introduction of category mechanism indication matrix C is the total number of fault categories, and the original mechanism score is obtained: , In the formula, For sample x to belong to The original mechanism score of the class fault; Represent the first Class faults contain the underlying fault mechanism , Represent the first Class failures do not include the underlying failure mechanism ; Normalizing the original mechanism score to obtain a physical prior distribution: , In the formula, For sample x to belong to Physical prior distribution of fault-like; For adjusting the temperature parameters introduced by the prior distribution smoothness, And j is a fault category index.
5. The method according to claim 4, wherein the calculation of the indication function comprises , Wherein: is the first The actual frequency of the frequency points; , Wherein, the Is the t-th fault characteristic frequency The center frequency of the order harmonic wave, 、 The upper bound and the lower bound of the corresponding harmonic neighborhood window are respectively; for determining relative width coefficient of harmonic neighborhood window relative to center frequency Bandwidth size of (a); Is the sampling frequency.
6. The method of claim 5, wherein the defining of the center frequency comprises: , Wherein, the , Is the characteristic frequency of the t-type fault.
7. The method of claim 1, wherein the structural alignment loss is calculated based on a maximum mean difference, comprising: , In the formula, In order to achieve a loss of alignment of the structure, Respectively a labeled source domain sample set and an unlabeled target domain sample set; Representing a nonlinear mapping function; 、 respectively the first Individual source domain samples And the first Individual target domain samples Is a structural enhancement feature representation of (1); representing regenerated kernel Hilbert space Is the square norm of (a).
8. A system for performing the method of any of claims 1-7, comprising: The acquisition module is used for acquiring vibration signals of the robot joint flexible bearing under different working conditions and constructing a data set comprising a labeled source domain sample set and a non-labeled target domain sample set; the training module is used for training the domain countermeasure graph convolution neural network by utilizing the data set to obtain a diagnosis model; the training of the domain countermeasure convolution neural network comprises the following steps: After preprocessing the flexible bearing vibration signal of the data set, inputting an improved convolution feature extraction network to obtain a fault characterization feature map, wherein a kernel selective fusion attention module is introduced into the improved convolution feature extraction network, and the fault characterization feature map is taken as input to output an enhanced feature representation; Mapping the enhanced feature representation into sample node features, and constructing a sample instance graph according to the similarity among samples by using a graph construction layer; Inputting the structural enhancement feature representation into a classifier to obtain model prediction distribution of a target domain sample, and calculating source domain classification loss; inputting the structural enhancement feature representation into a domain discriminator, carrying out domain discrimination on a sample, and calculating domain alignment loss; Constructing physical prior distribution corresponding to the target domain sample according to fault characteristic frequency in an envelope power spectrum of the target domain sample and energy distribution in a harmonic neighborhood of the fault characteristic frequency; Updating parameters of the domain contrast graph convolutional neural network by using the weighted total loss of the physical consistency loss and other losses to obtain a trained diagnosis model; the identification module is used for inputting the target sample to be diagnosed into the trained diagnosis model and outputting the identification result of the fault class of the flexible bearing.
9. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method according to any one of claims 1 to 7.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when loaded by a processor, is able to carry out the steps of the method according to any one of claims 1 to 7.

Description

Robot joint flexible bearing fault diagnosis method based on physical priori drive Technical Field The invention relates to the technical field of bearing state monitoring and intelligent fault diagnosis, in particular to a robot joint flexible bearing fault diagnosis method based on physical priori driving. Background The flexible bearing is a key basic component in the robot joint module, and the running state of the flexible bearing is directly related to the safety, reliability and maintenance cost of the equipment. In an actual industrial scene, the flexible bearing usually works under complex working conditions of continuous change of rotating speed, load and environmental conditions, and the vibration signal easily shows obvious non-stationarity and distribution drift. Once the flexible bearing breaks down, the running performance of the robot can be reduced, and a shutdown accident can be caused, so that larger economic loss is caused. Therefore, flexible bearing state monitoring and fault diagnosis under variable working conditions are always important research directions in the field of intelligent operation and maintenance. The convolutional neural network and other methods can directly learn fault characterization features from the original vibration signals, and good effects are achieved under fixed working conditions or light distribution offset scenes. However, when the working condition changes greatly, the statistical distribution of the vibration signals, the time-frequency scale characteristics and the similarity relationship among samples are obviously changed, so that the problems of characteristic distortion, decision boundary deviation, generalization capability reduction and the like of the existing model are easy to occur. In order to solve the diagnosis problem of the target working condition under the condition of lacking a label sample and even completely no label, an unsupervised domain self-adaptive method is gradually introduced into a bearing cross-working condition fault diagnosis task. Most of the existing methods realize the characteristic alignment of the source domain and the target domain through distribution matching or countermeasure learning, but the consistency of global statistical distribution is usually focused, and the utilization of the structure relationship among samples and the topology mode in the class is still insufficient. To this end, the graph-convolution network is further used for cross-condition diagnostics to promote structure propagation capability by explicitly modeling inter-sample relationships. However, the existing graph convolution diagnosis method still has two defects, namely, on one hand, the graph structure generates node characteristics which are highly dependent on convolutional network extraction, and under the variable working condition, the characteristics are easily influenced by scale fluctuation and local mode drift, so that similarity measurement is unstable and adjacency relation is distorted, and on the other hand, the existing optimization process mainly depends on data driving constraints such as classification supervision, domain alignment, structure alignment and the like, and lacks explicit physical consistency constraints directly related to bearing failure mechanisms, so that the defects still exist in diagnosis reliability and physical interpretability. In addition, the bearing fault characteristic frequency and the energy response thereof in the envelope spectrum contain physical information closely related to the fault part, and can provide important priori basis for fault identification. How to integrate the failure mechanism information into the model training process in a stable and controllable manner under the condition that the target domain is not labeled, and to perform collaborative optimization with cross-domain alignment and graph structure modeling is still a problem to be solved in the current technology. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a robot joint flexible bearing fault diagnosis method based on physical priori driving, which solves the problems of insufficient characteristic characterization stability, insufficient structural relation modeling among samples and lack of explicit physical constraint related to a fault mechanism in the prior art in the variable working condition bearing fault diagnosis, and limited fault recognition precision, diagnosis stability and result interpretation under the cross working condition. The technical scheme adopted by the invention is as follows: The invention provides a robot joint flexible bearing fault diagnosis method based on physical priori driving, which comprises the steps of collecting robot joint flexible bearing vibration signals under different working conditions, constructing a data set comprising a labeled source domain sample set and a label-free target domain sample set, training a domain contrast