Abstract
Electric vehicles (EVs) and energy storage systems are prone to critical safety hazards associated with lithium-ion battery failures. The current fault detection tools have a fundamental trade-off: Model-based approaches cannot support aging-dependent dynamics, whereas deep learning methods generate too many false alarms and cannot estimate uncertainty of the predictions. To overcome these shortcomings, this paper proposes a novel hybrid framework integrating Adaptive Data-Driven Kalman Filtering (ADKF) with Transformer-based deep learning. The key innovation is a two-way, continuous feedback loop between the filter and the network. This is different from the earlier hybrid methods where the filter and the network work side by side without sharing the information, or the filter was used to prepare the network data in one direction only. The proposed transformer model learns adaptive state-space models from operational data, continuously refining the ADKF's predictions, while the ADKF simultaneously feeds its innovation sequence back to the transformer, providing optimal Bayesian filtering that reduces noise-induced false positives and quantifies prediction uncertainty, which is the principal shortcoming of transformer-only models. This two-way feedback loop is what allows the proposed transformer-based deep learning model to learn state-transition and observation mappings directly from data, supplying the Kalman filter with accurate, adaptive process models that evolve with battery aging rather than requiring manual re-tuning. When transformers detect faults, they can automatically adjust Kalman filter parameters, maintaining estimation accuracy during both normal and faulty operations. The proposed framework addresses five critical fault categories: two SOC-related anomaly types (drift, imbalance), SOH degradation (capacitor degradation), temperature faults (thermal anomalies), coupled multi-parameter faults, and sensor integrity issues. A comprehensive comparative study of F1-score, accuracy, error and time of fault detection with existing methods (CNN, LSTM, standalone transformer model) have been carried out to prove the efficacy of the proposed method. Compared with traditional method and other deep learning approaches, the proposed systems achieves 97.5% F1-score (10-fold cross-validated mean, 97.3% on the single held-out test split, 38% earlier fault detection, 55% fewer false positives, high-accuracy state estimation (SOC error <2%, SOH error <3%), 1.8x faster inference than ensemble and 67–71% faster adaptation to changing conditions. The proposed method not only achieves state-of-the-art performance, it also maintains computational feasibility for embedded Battery management systems by advancing safety and reliability in critical energy storage applications.