Real-time impending lithium battery terminal voltage collapse detection via dynamic mode decomposition and optimal Kalman filtering algorithms

In this study, several lithium battery terminal voltage collapse detection methods were proposed to protect physical battery units from permanent damage resulting from over-discharge scenarios. The proposed technique does not require prior knowledge of sophisticated battery models, battery current measurement, state of charge estimation, and Jacobian evaluation. In addition, it is free of training data and labeling issues. This was possible through the utilization of dynamic mode decomposition (DMD) and Kalman filtering algorithms (i.e., Kalman filter (KF), H-infinity filter (H-$\infty$), and unscented Kalman filter (UKF)). Moreover, the aforementioned design was verified, tested, and compared with techniques from relevant studies such as the battery model-based DMD, Lyapunov stability and supervised machine learning methods, specifically principal component analysis (PCA) and support vector machine (SVM) algorithms, using experimental and simulation results. The preliminary technique was further improved by incorporating QR-pivoting and maximum likelihood estimation (MLE) techniques into the battery failure detection structure. Consequently, both detection accuracy and computation time were significantly enhanced, with the improved version demonstrating satisfactory performance compared to other techniques. Based on the first experimental dataset, the proposed preliminary methods, represented by DMD-KF, DMD-H-$\infty$, and DMD-UKF, showed accuracy as 91.3%, 91.04%, and 91.58%, respectively. In contrast, the enhanced detection approaches, represented by QR-DMD-MLE-KF, QR-DMD-MLE-H-$\infty$, and QR-DMD-MLE-UKF, showed 97.15%, 97.15%, and 97.15%, respectively. The PCA-SVM, Lyapunov stability, and battery model-based DMD approaches showed 93.48%, 90.33%, and 89.67%, respectively.