Recurrence plot image and GoogLeNet based historical abuse backtrace for li-ion batteries

Abstract

Failures of Li-ion batteries (LiBs) are generally ascribed to the side reactions of internal damages, which are usually induced by abuses like mechanical destruction, electrical overload, or thermal overheating. In this setting, a systematic scheme is proposed to backtrack the historical abuses suffered by LiBs based on the Recurrence Plot (RCP) and Convolutional Neural Network (CNN). First, the voltage correlation of the cells in a series pack is quantified to reflect their state consistency. Then, resorting to the RCP transformation, the time series of voltage correlation is picturized as images that deliver the indicative textures of cross-temporal autocorrelation. In order to analyze the abuse characterizing ability of the RCP images, an unsupervised classification test is conducted on the images using hierarchical clustering, and, as expected, the images show favorable separability and damage correspondences, Finally, leveraging the power of CNN in the extraction and fusion of multiscale features, the multilayer GoogLeNet is introduced to process the RCP images, thereby providing qualitative and graded judgments on the historical abuses of LiBs. Moreover, various abusive operations are inflicted on LiB cells to acquire a realistic dataset. Experimental verification suggests that the proposed scheme can provide accurate and reliable recognition and estimation results on abuse type and intensity, with accuracy rates up to 77.3 % and 75.4 %, respectively.