Constructing a health indicator based on long short-term memory and using an extreme inflection point with a slope model to enhance monotonicity

摘要

Traditional health indicator (HI) construction methods usually require manual feature selection and fusion, and different models are selected for combination; therefore, a large amount of expert experience and signal-processing techniques is required. In addition, there is a problem of smoothness and monotonicity in the overall angle of the acquired HI curve which can affect the difficulty of predicting the remaining useful (RUL). It is usually necessary to select an appropriate model and parameter settings to delete abnormal points, eliminate the oscillation, and achieve monotonic enhancement. In order to solve these problems, an HI extraction model based on long short-term memory (LSTM) in deep learning and an extreme inflection point with a slope (ES) is proposed. First, to reduce the dependence on prior knowledge and artificial experience, the proposed model uses LSTM to learn the frequency domain data directly and then constructs a preliminary HI from the learnt data through nonlinear mapping. Moreover, an ES model, based on an exponential function, is proposed to eliminate violent oscillation and improve the overall monotonicity of the HI without any prior knowledge for parameter selection. Simultaneously, the proposed method lays a good foundation for prediction of the remaining useful life. Finally, the proposed model is validated on the rolling bearing data. And to prove the superiority of the proposed model, the paper compares other traditional HI extraction models.