A Hybrid Remaining Useful Life Prognosis Method Integrating Transformer Networks and Wiener process

Remaining useful life (RUL) prediction is the key part in prognostics and health management (PHM) which has proven effective for reliability strengthening, availability improving and cost saving. It is difficult to quantify the level of uncertainty of the RUL prediction model based on deep learning, and the stochastic process based model has some limitations in handing complex and mass data. Therefore, a Wiener process model for RUL prediction integrating Transformer neural network is proposed in this paper. Firstly, the historical data is filtered and smoothed, and the corresponding degradation trend in the processed historical data is extracted by the complete EEMD with adaptive noise (CEEMDAN) method. Based on the obtained degradation trend, the training data set, and test data set have been constructed through moving window skills to train a simplified Transformer neural network. An adaptive identification of the Wiener drift coefficient function is performed using the trained Transformer neural network. Then, the analytical probability density function of RUL based on the first passage time (FPT) has been derived. The proposed method has been illustrated by using the public Lithium-ion batteries capacity degradation data provided by NASA.