Reliable Battery Terminal Voltage Collapse Detection Using Supervised Machine Learning Approaches

In this study, image feature extraction techniques based on principal component analysis (PCA) and QR factorization were employed to represent the measured terminal voltage of several batteries in terms of a few dominant features instead of raw data. In previous studies, directly measured terminal voltage was used to create the feature space by inputting raw sampled data to a supervised classifier. This was ineffective because it leads to false alarms owing to data overlap and requires a relatively considerable processing time. However, the proposed methods are useful for building a reduced and distinguishable feature space, where a classifier can quickly process and separate regions with small labeling errors. Therefore, different supervised machine learning classifiers, namely neural networks (NNs), k-nearest neighbors (kNNs), and support vector machines (SVMs), were trained based on the dominant features to distinguish between the safe and failure regions of battery units. Interestingly, the proposed technique requires no knowledge of the model parameters and state of charge (SOC) in the training and testing phases, although the SOC is necessary in the data labeling stage. The results showed that the nonlinear classifiers represented by NNs and kNNs demonstrated slightly better detection performance than the linear SVMs.