Lithium plating compromises battery safety during fast charging, yet real-time detection remains challenging due to the entanglement of natural polarization and anomalous impedance drops. Here, we propose a physics-informed, data-driven framework for online plating detection using dynamic electrochemical impedance spectroscopy. We utilize the complex Morlet wavelet transform (CMWT) to robustly extract the 1 Hz real impedance ( Z 1 Hz ), mitigating conventional boundary artifacts. A gated recurrent unit (GRU) network then continuously tracks the impedance evolution. The optimized GRU achieves an R 2 of 0.99 and a root mean square error of 0.23% under standard conditions. Reducing parameters by about 30% and memory by about 15% relative to traditional recurrent architectures, it provides a pragmatic inference engine for resource-constrained battery management systems. Crucially, we introduce a phase-adaptive differential thresholding strategy based on the macroscopic impedance derivative (▽ Z ). Integrating a downsampled computational interval with an RMSE-bounded late-stage threshold, the framework physically decouples normal activation polarization from genuine structural collapse. This approach deliberately prioritizes false-positive suppression, ensuring reliable lithium plating detection across diverse thermodynamic environments without relying on rigid absolute baselines. • A CMWT–GRU framework enables real-time lithium plating onset detection under non-stationary charging. • CMWT outperforms FFT in extracting smooth and robust 1 Hz impedance features for plating identification. • Bayesian-optimized GRU achieves >98% accuracy with lower memory cost, suitable for embedded BMS.
Li et al. (Fri,) studied this question.