Remaining useful life prediction using a hybrid transfer learning-based adaptive Wiener process model

Because of the characteristics of uncertainty description and interpretability, Wiener process (WP) has found extensive application in the domain of forecasting remaining useful life (RUL). Nevertheless, most existing WP often require selecting the suitable deterioration function and drift coefficient types based on the deterioration characteristics of target sample, which greatly limits their universality and feasibility in practical engineering. In order to address this issue, a hybrid adaptive WP based on transfer learning is presented to dynamically model the deterioration process of products with different deterioration features. The Brownian motion-based drift coefficient is applied to improve the adaptive characteristics of WP on the time-variant deterioration rate. A transfer learning-based long short-term memory (LSTM) model is utilized to adaptively track the dynamic nonlinear characteristics. According to the notion of first arrival time, we have successfully derived the explicit formula for the probability density function, so that the uncertainty contained in predicted results can be directly characterized. By using two capacity datasets and one torque bar deterioration dataset exhibiting distinct deterioration features, comparative experiments with eight different existing models have proven the universality and superiority of our model in forecasting RUL.