Remaining Useful Life Prediction Of Rolling Bearing Based On Feature Fusion And LSTM

Rolling bearing is the basic part of mechanical equipment,which plays an important role in rotating machinery.Therefore,condition monitoring and remaining useful life(RUL)prediction of rolling bearings are particularly important in industrial production.The progress of sensor technology,data storage and processing speed,as well as the rapid development of Internet of things and artificial intelligence,we are really brought into the era of industrial big data.In the context of industrial big data,the RUL prediction of rolling bearing based on data-driven methods has made great progress.This paper focuses on several key steps of traditional data-driven RUL prediction method,including feature extraction and selection,feature fusion,health state division and RUL prediction.The main research content is divided into the following three parts.(1)A feature selection method based on first and second selection is proposed.Firstly22 features are extracted from the vibration acceleration signal of bearings in time domain,frequency domain and time-frequency domain to form candidate feature set.The first selection of features is to select the sensitive subset which is sensitive to the degradation trend through the comprehensive index sorting.The second selection of features based on hierarchical clustering and mutual information can further eliminate redundant features from sensitive feature subset,and finally get the optimal subset.(2)A feature-based fusion method of adaptive feature fusion(AFF)and auto association kernel regression(AAKR)model is proposed to construct health indicator(HI).The performance of health indicator is a key factor that directly affects the prediction accuracy of RUL.The feature-based AFF-AAKR fusion method proposed in this paper is a combination and improvement of AFF and AAKR model.The superiority of the method in improving performance of HI is proved by experiment.Then,this paper uses the bottom-up segmentation algorithm to divide the health state of HI,and obtain the first failure point.HI is divided into healthy period and degradation period by this point,and the degradation period of HI is the input object of RUL prediction,so this point is also the starting point of RUL prediction and the First Prediction Time of RUL prediction is obtained.In this way,the RUL prediction error caused by randomly dividing healthy period and degradation period is avoided,and the accuracy of RUL prediction is further improved theoretically.(3)A RUL prediction method based on Long Short-Term Memory Neural Network(LSTM)and piecewise linear fitting HI is proposed.Under the dynamic condition and coupling of various failure modes,the LSTM with modeling advantages of long-term series and strong nonlinear processing ability to build RUL prediction model of rolling bearing,and to establish the mapping relationship between fitting HI and RUL.This paper does not insist on improving the type,structure or parameters of prediction model to improve the effect of RUL prediction,but studies the impact of three different inputs of prediction model on prediction results.Experiments show that the piecewise linear fitting HI used as the input of the prediction model can achieve more robust and accurate prediction results than using the original HI and the single feature of the original vibration signal as the input.