摘要
Most of existing data-driven studies on lithium-ion battery remaining useful life (RUL) prediction consider a large scope of cyclic data over the entire battery life. Yet, applications of these models can be hindered due to restricted availability of such data in reality. This paper thus aims to study the battery RUL prediction from a new angle, predicting RUL via data collected from a limited number of incomplete cycles, i.e., 10 cycles, at any ageing stage. An advanced deep learning framework, the attention-assisted temporal convolutional memory-augmented network (ATCMN), is developed to realize an accurate and rapid battery RUL prediction under such challenging problem setting. To build an informative input based on limited data, a three-dimensional tensor input structure is first designed to integrate 10-cycle raw battery data including the time, capacity, and temperature dimensions obtained from the partial discharge process. To process such high dimensional input, the ATCMN first develops an attention module to automate weighting different battery parameters, time steps, and ageing cycles in the input. A temporal convolution module coordinating the dilated causal convolution and point-wise convolution is next developed to learn latent spatial-temporal feature representation from the weighted input. A memoryaugmented module is further developed to enhance the latent feature representation through a reconstruction based on the historical information. Finally, the ATCMN employs a prediction module to derive nonlinear mappings from learned latent features to the battery RULs. A comprehensive computational study is conducted to verify the effectiveness of the ATCMN. Results report the higher accuracy and faster prediction of the ATCMN via benchmarking against a set of the state-of-the-art methods in the considered RUL predictions. Experimental results also show that the proposed ATCMN possesses better generalizability to different battery chemistries and operational conditions.
-
单位中山大学