A long short-term memory neural network based Wiener process model for remaining useful life prediction

• A Wiener process-based degradation model is proposed to obtain remaining useful life prediction. • Model universality among different types of degradation data is considered. • An uncertain description of remaining useful life prediction is obtained. • Model parameters are updated online by transfer learning. • Accurate remaining useful life predictions are acquired. An unsuitable type of degradation trend function in the Wiener process-based degradation model will negatively influence its performance when calculating remaining useful life (RUL) predictions. To solve this problem, we propose a Wiener process-based degradation model that can adaptively learn the degradation trend in different degradation data, which avoids the selection of degradation trend function. First, based on the degradation trends extracted by empirical mode decomposition, a long short-term memory (LSTM) neural network is trained and used as the degradation trend function of a Wiener process-based degradation model. Then, transfer learning is used to update the parameters of the LSTM neural network online. Concurrently, the diffusion coefficient of the Wiener process-based degradation model is obtained via maximum likelihood estimation. Finally, using the concept of first hitting time, the analytical formulation to the probability density function of RUL can be derived in a closed form. Two numerical examples are presented to demonstrate the implementation and the achieved parameter estimation accuracy of the proposed model. In addition, a real battery dataset is used to demonstrate the superior performance of the proposed model against previous Wiener process-based degradation models in RUL prediction.