State Prediction And Fault Identification Of Gearbox Based On Gaussian Process

Gearboxes,as the basic equipment for manufacturing,are widely used and are critical to industrial safety production.Effective monitoring and management of the service status of the gearbox contributes to the safe and economical production of industrial production.Therefore,the prediction of the service state of the gearbox and the fault identification and positioning are conducive to improving the level of industrial intelligent production.However,the service environment of the gearbox system is harsh,and the fault information is coupled with each other.Therefore,the state monitoring and fault identification of the gearbox system are faced with great challenges.Based on this,this paper studies the sensitive degradation feature extraction,degradation trend prediction and fault identification.The content is as follows:(1)The key to accurately evaluate and predict the service state of gearbox is to extract the sensitive characteristic parameters.At first,this paper analyzes some characteristic parameters commonly used in industrial monitoring.In view of the lack of sensitivity of conventional feature quantities to early damage,based on the analysis of the current situation of time-frequency analysis methods,combined with the superiority of complex wavelet transform,put forward structural similarity index Wigner-Ville Distribution Complex Wavelet Structural Similarity,WVD-CWSS.Through the bearing life test data,quantitative analysis of the value of the characteristics,RMS,margin,frequency domain variance and STFT time-frequency structure similarity evaluation index(STFT-CWSS),The results show that WVD-CWSS is more sensitive to the early failure of bearing.(2)State prediction is to judge the future degradation trend by learning the degradation law under the current service state.The Gaussian process prediction result has probabilistic significance,has certain uncertainty,and has good adaptability to small sample and complex stochastic processes,its prediction results have probabilistic significance,uncertainty,can good adaptability to small sample and complex stochastic processes.Compared with neural network,its parameter setting is simple and it can adaptively learn hyperparameters.In view of the WVD-CWSS sensitivity to the early damage,Gaussian process regression is used to predict the variation trend of WVD-CWSS feature quantity.Finally,compared with the BP neural network prediction results,it is proved that the method can predict the tendency of bearing degradation.(3)For the monitoring process of the gearbox system,it is difficult to realize the fault modes by single sensor and single feature quantity.In order to ensure the completeness of the information,usually,multiple feature parameters are fused for pattern recognition.But the constructed high-dimensional feature set often exists a large number of coupling,redundant information.The Gaussian Process Hidden Variable Model(GPLVM)algorithm can eliminate redundant information for high-dimensional feature sets.Through the experimental data,by comparing with PCA,LE,Isomap manifold space mapping algorithms,the result proves that GPLVM is beneficial to the identification of different operating modes.