This study presents a data-driven framework for early detection of abnormal cells in lithium-ion battery modules. Incremental capacity analysis (ICA) is performed on current–voltage data to visualize cell degradation as ICA curves, which are then used as inputs to Autoencoder(AE) models for training and detection. Two 6S2P lithium-ion battery modules were tested: one with overcharged cells (positions 1 and 6) and the other with overdischarged cells (positions 1 and 6). Each cell underwent 100 charge– discharge cycles under constant current (CC) conditions to simulate progressive aging. From these, 80 ICA cycles from normal cells (positions 2–5) were used for training, and the rest for evaluation, ensuring the model only learned healthy behavior patterns. Four detection approaches were compared: ICA vectors, ICA images, Recurrence plot (RP) images, and RGB-based multi-channel time series encoding using Gramian angular field(GAF) and Markov transition field(MTF) images. Models were trained only on normal cell data, and anomalies were identified by increases in reconstruction error, such as Mean squared error (MSE). Results showed that, in terms of mean F1-score, RGB GAF–MTF delivered the best performance (0.841), followed by RP (0.775), whereas ICA-vector AE and ICA-image CNN lagged behind (0.673 and 0.504, respectively), underscoring the advantage of multi-channel time-series encodings for capturing subtle degradation.
Lee et al. (Fri,) studied this question.