Performance Evaluation of Battery Abnormal Diagnostic Based on Autoencoder according to the Battery Management System Input Dataset

In this study, autoencoder was used to diagnose Li-ion battery fault type. Battery experiments for abnormal and normal were conducted to construct dataset for detection safety of lithium-ion batteries (LIBs). The dataset consisted of a total of 32 overcharge, overdischarge, short-circuit, and normal data. Among several sensor data (Voltage, Current, Temperature, etc.) collected by the battery management system (BMS), 1, 3, and 5 signal dataset were constructed by preprocessing BMS data. These were used as input data of autoencoder and abnormal or normal was diagnosed according to reconstruction error of encoder and decoder within autoencoder. Through errors in autoencoder-based restoration data and original data, it is diagnosed as abnormal if it is more than the threshold, and normal if it is within the threshold. The accuracy of the algorithm performance result was 84.38%, 90.63%, and 93.75%, respectively, and the difference in diagnostic performance was confirmed according to the input data set. In future work, these failure types are planned to be classified based on prior diagnosis.