Data-Driven Fault Detection for Lithium-Ion Battery Packs via Behavioral System Representation

Fault detection and diagnosis of lithium-ion batteries have been of intense investigation in energy systems, but most applicable methods rely on precise and complicated mechanistic models, which are nontrivial to establish in practice. The recently emerging behavioral system theory yields a new model-free representation of dynamical systems using only a single input-output trajectory. This enables us to develop a new data-driven fault detection scheme for lithium-ion battery packs, which effectively captures battery dynamics from data and bypasses the cumbersome work of building and calibrating first-principle battery models. By applying Willems’ fundamental lemma to the innovation form of stochastic systems, our method attains a purely data-driven yet approximated realization of Kalman filter-based output predictor, which inspires a new data-driven residual generator for fault detection. Compared with generic data-driven residual generators such as stable kernel representation, our method performs more closely to a Kalman filter-based residual generator and thus better handles uncertainty including stochastic disturbance and noise. Simulations on a three-cell battery pack demonstrate the effectiveness of our method and its outperformance over existing data-driven fault detection design.