Electrochemical impedance spectroscopy (EIS) separates physicochemical effects across frequencies and enables nonintrusive diagnosis. However, one-dimensional spectra and equivalent-circuit parameters are insufficient for early anomaly detection under high-temperature operation. Cyclewise EIS residuals relative to a reference are derived and encoded into two complementary images. Gramian angular field (GAF) capturing the global angular structure and a dynamic time warping (DTW) cumulative cost matrix encoding nonlinear shape and timing differences. These images are fused as a multi-chanel input to a weight-sharing Siamese convolutional neural network (SCNN), which determines anomalies from the learned similarity score. Experiments under ambient and high-temperature conditions show that multi-channel image representation detects anomalies more reliably than raw data or single-channel images. Combining angle-based and distance-based information improves the sensitivity to subtle degradation and supports timely risk identification in battery management systems (BMSs).
Lee et al. (2026) studied this question.