Research On Equipment Remaining Useful Life Prediction Methods Based On Limited Failure Historical Data

The new generation of information technology such as Internet of Things,Cloud Computing,and Big Data as well as the new round of industrial revolution,result in the qualitative development of equipment lifecycle management.The collection of industrial big data including the performance and environment data captured by the distributed sensor networks,provides the necessary basic conditions for Prognostics and Health Management(PHM).As a key technology to ensure equipment safety and reliability,PHM can effectively monitor the operating state and predict the remaining useful life(RUL)of the equipment,so as to optimize operation,maintenance and assurance strategies.Accurate RUL prediction can provide effective information for decision makers to make maintenance plans in advance and optimize supply chain management,thus reduce unnecessary maintenance or replacement cost.In comparison with the physics-based methods,the data driven methods avoid the limitation of high complexity,high time consuming and expensive cost.It is more economical and applicable in modern industry where the iteration period of equipment renewal is gradually shortened.However,the failure historical data,i.e.the "run-to-failure" process data,are often very scarce in industry practice,that leads to a large degree of uncertainty in RUL prediction based on the limited failure historical data.The predicted results are often very different from the actual ones,and lose the effectiveness of optimizing operation,maintenance and guarantee strategies.This paper systematically analyzes two data-driven methods for RUL prediction based on limited failure historical data,and has made in-depth research on improving the accuracy of prediction.The main research contents are as follow:(1)For the similarity-based methods,this paper considers the effect of the reference samples' different degradation rates,and discusses how to quantitate such effect.Then a modified similarity-based method incorporating the impact of degradation rate has been proposed.Its effectiveness and accuracy are verified by a real-world case study.(2)For support vector regression(SVR)methods,this paper on the one hand considers the issue of the traditional SVR method that the linearity will increase with the increase of the sample set,which may lead to overfitting or underfitting.And on the other hand,this paper also considers the limitation that the diversity of degenerate trajectories may lead to an issue that a single SVR model may lose its practicability due to the pursuit of generalization.Therefore,a modified SVR method with health stage division,cluster sampling and similarity matching has been proposed in this paper.Its effectiveness and accuracy are verified by a simulation study and a real-world case study.