Remaining Useful Life Prediction of Rolling Bearings Based on Recurrent Neural Network

Qiang Cheng,Jing Li,Wei Ren,Aidong Deng,Yimeng Zhai

doi:10.32604/jai.2019.05817

Abstract

In order to acquire the degradation state of rolling bearings and achieve predictive maintenance, this paper proposed a novel Remaining Useful Life (RUL) prediction of rolling bearings based on Long Short Term Memory (LSTM) neural net-work. The method is divided into two parts: feature extraction and RUL prediction. Firstly, a large number of features are extracted from the original vibration signal. After correlation analysis, the features that can better reflect the degradation trend of rolling bearings are selected as input of prediction model. In the part of RUL prediction, LSTM that making full use of the network’s memory in time is used to improve the accuracy of RUL prediction. The proposed method is validated by life cycle experimental data of bearings, and the RUL prediction results of LSTM model are compared with Support Vector Regression (SVR) and Light Gradient Boosting Machine (LightGBM) models respectively. The results show that the proposed method is more suitable for RUL prediction of rolling bearings.

Highlights

Rolling bearings are the key parts to support and transfer torque, which have prompted its extensive us in rotating machinery, such as bearings in Wind Turbine Drive Train
Because the kurtosis characteristic whose feature values at early stage even exceed the values in later stage is more sensitive to impact, burrs can be observed in the initial stage of bearing operation, which will affect the Remaining Useful Life (RUL) prediction accuracy
In addition to No 3 bearing, the full life cycle samples of remaining bearings are taken as training sets to construct the RUL prediction model

Summary

Introduction

Rolling bearings are the key parts to support and transfer torque, which have prompted its extensive us in rotating machinery, such as bearings in Wind Turbine Drive Train. Degradation state assessment and RUL prediction can effectively prevent sudden failure of mechanical equipment, and maximize the use of the working capacity of key components, reduce maintenance costs and reduce unnecessary waste of resources. In recent years, it has become a research hotspot. This paper presents a method of rolling bearing RUL prediction based on LSTM, which mainly includes feature extraction and RUL prediction. The feature vectors which can better reflect the trend of bearing degradation are selected to be input of the prediction model. In addition to LSTM network model, SVR and LightGBM methods are used to compare and prove the effectiveness of the proposed method

Long short term memory

Experiments and result analysis

Feature extraction

RUL prediction

Findings

Conclusion