Adaptive Internal Short Circuit Detection Using Spatial–Temporal Graph Attention Network and Improved Deep Echo State Network

Internal short circuit (ISC) detection is vital for alleviating thermal runaway and improving the safety of electric vehicles (EVs). Most of the existing ISC detection techniques rely on a static threshold-based mechanism to differentiate healthy batteries from faulty batteries, resulting in high false alarms and missed detections. To overcome this issue, an integrated model combining the Spatial–Temporal Graph Attention Network (STGAT) and the Improved Deep Echo State Network (IDESN) is proposed for adaptive threshold residual analysis-based ISC fault detection. The proposed STGAT-IDESN identifies faults through three stages: adaptive voltage prediction, adaptive threshold estimation, and a threshold-based fault identification stage. Battery characteristics are utilized by STGAT for predicting voltage, while driving behavior, historical battery voltage, and temperature are used by IDESN for predicting the threshold. A sliding window is employed to generate an adaptive threshold, which is then compared with the residuals between the predicted voltage and actual voltage to accurately detect ISCs. The proposed model is simulated in Python 3 using detection accuracy, false alarm rate (FAR), and detection time as evaluation metrics. The simulation results exhibited improved detection accuracy, reduced false alarms, and shorter detection time. The proposed model achieved higher ISC detection accuracy than state-of-the-art techniques, including LSTM, MSA, 1DVCNN, and MDM-Bi-LSTM, by 4%, 3.8%, 3%, and 2%, respectively.