Toward Electric Vehicle Safe Operation: Unsupervised Learning-Enabled Multiple Fault Diagnosis in Lithium-ion Battery Systems

How to quantify, classify and locate faults are the highlights and difficulties for securing electric vehicle battery systems. This paper presents a diagnostic scheme specifically designed for identifying multi- and micro-faults in series-connected battery systems, employing dynamic time warping (DTW) and unsupervised learning techniques. First, we conduct cycling tests by injecting multiple faults to simulate potential failures during battery operation. Subsequently, open circuit voltage is online identified by using an equivalent circuit model combined with a recursive least square algorithm. Based on the identification results, DTW distance and mean differential voltage are calculated and extracted as fault features. Finally, a bi-Gaussian mixture model is applied for accurate and effective classification and identification of multiple faults. This research offers a significant contribution toward enhancing fault analysis in battery systems, enabling improved security and reliability.