基于低频激励下深度特征融合的锂离子电池微过充诊断

Minor overcharge (MOC) seriously affects the safe use of lithium-ion batteries. However, it is very insidious and difficult to detect and quantify in practical scenarios. To this end, a MOC diagnosis framework based on deep feature fusion under low-frequency excitation (LFE) is proposed. First, the LFE sequences are determined based on offline tests and the voltage response sequences (VRS) are obtained for different MOC degrees. Second, global morphological features of VRS are extracted by the minimally randomized convolution kernel transform and local incremental features of VRS are extracted based on domain knowledge. Finally, a deep learning network is built to introduce the cross-attention mechanism to fuse the multiscale features for localization and degree quantification of MOC. Experimental results indicate that the proposed method can detect various degrees of MOC in the fastest 120 s across the full state of charge range, with detection rate, detection accuracy, and quantization root mean square error of 98.17%, 98.56%, and 2.18%, respectively. Its specificity, robustness, and generalizability have also been validated. Importantly, the algorithm has been deployed in the battery management system and proven to work efficiently. This provides a practical and effective approach for MOC diagnosis.