Real-time battery safety diagnosis via Siamese convolutional neural network combined with online passive electrochemical impedance spectroscopy pattern extraction based on driving data and short-time Fourier transform

Electrochemical impedance spectroscopy (EIS) provides rich information on battery internal electrochemical processes, but conventional EIS requires dedicated excitation and external instruments, limiting on-board and real-time use. This paper proposes an online passive electrochemical impedance spectroscopy (PEIS) extraction method that estimates an impedance-like spectrum from naturally occurring current/voltage fluctuations during vehicle operation using a short-time Fourier transform (STFT), without injecting additional AC signals or interrupting operation. The STFT parameters are selected via a quantitative trade-off study between frequency resolution and PEIS–EIS fitting deviation. PEIS is validated against reference laboratory EIS measured at matched SOC and temperature points, showing a maximum mean fitting error below 6% for the real impedance at selected frequencies. Finally, the extracted PEIS patterns are used as electrochemical features for real-time safety screening based on similarity evaluation. To classify the internal state using this PEIS pattern, it is converted to image for deep learning. Based on this, we propose a new methodology to learn a Siamese convolutional neural network (SCNN) algorithm on the converted PEIS pattern and to effectively diagnose the state of the battery through similarity evaluation. The proposed framework enables practical, instrument-free extraction of electrochemical signatures from standard BMS data, supporting scalable safety monitoring for EV batteries.