KAN-Koopman Based Rapid Detection Of Battery Thermal Anomalies With Diagnostics Guarantees

Researchers Sanchita Ghosh and Tanushree Roy have introduced a novel AI-driven framework, the KAN-Koopman detection algorithm, designed to rapidly identify thermal anomalies in batteries with formal diagnostic guarantees. Accepted for presentation at the 2026 American Control Conference, this work addresses critical challenges in battery safety by preventing catastrophic thermal failures through early diagnosis. The system tackles two major industry hurdles: aging-induced changes in battery models that render traditional model-based approaches ineffective, and the limited availability of training data that constrains pure machine learning solutions. By proposing an integrated structure, the researchers aim to provide reliable diagnostics where current methods fall short.

The technical innovation lies in combining two distinct mathematical approaches: a Kolmogorov-Arnold Network (KAN) for model-free, lightweight estimation of a battery's internal core temperature, and a Koopman operator learned in real-time using both the KAN's estimate and measured surface temperature to generate a diagnostic residual. This hybrid architecture allows the system to adapt online, overcoming model drift due to aging, while reducing dependence on large historical datasets. The authors have derived analytical conditions that provide verifiable guarantees on the detection scheme's performance. Simulation results demonstrate a significant reduction in detection time compared to a baseline Koopman-only algorithm, marking a substantial step toward safer, more reliable battery management systems for electric vehicles and grid storage.