Lithium Battery Terminal Voltage Collapse Detection via Kalman Filtering and Machine Learning Approaches

Abstract

A low self-discharge rate, memoryless effect, and high energy density are the key features that make lithium batteries sustainable for unmanned aerial vehicle (UAV) applications which motivated recent works related to batteries, where UAV is important tool in navigation, exploration, firefighting, and other applications. This study focuses on detecting battery failure in the form of terminal voltage collapse using Kalman filtering and machine learning approaches. In the Kalman filtering approach, state estimation techniques were employed to determine the state of charge (SOC) and model output that utilized to detect battery failure when the battery is about to die. In the machine learning approach, this work analyzed unsupervised machine learning approaches to distinguish between safe and failure regions. Accordingly, this study requires no knowledge of SOC owing to utilizing clustering techniques, such as the Gaussian mixture model and k-means, to determine the labels, where the feature space was built based on principal component analysis (PCA). In addition, PCA feature space was incorporated with several classification methods to detect battery failure. Consequently, the proposed approach achieved considerably similar results to the supervised machine learning approach. Therefore, the proposed approach is advantageous because it does not require information on SOC to protect batteries from over-discharge scenarios that could lead to permanent damage.

Keywords

• Battery management systems (BMSs)
• principal component analysis (PCA)
• neural networks (NNs)
• k-nearest neighbors (kNNs)
• support vector machines (SVMs)