Adaptive engineering-assisted deep learning for battery module health monitoring across dynamic operations

Accurately assessing the health degradation of battery modules is crucial for the stable operation of electric vehicles. However, most existing methods rely on ideal constant current charging data, which is not suitable for real-world dynamic conditions. To address this issue, this study proposes an adaptive engineering-assisted deep learning framework for estimating the state of health of battery modules under different dynamic conditions. The method comprises an adaptive voltage corrector with built-in Lowess smoothing, voltage descending, and correlation alignment components, along with a specifically developed convolutional neural network. Initially, 10 min discharge segments under different dynamic conditions are inputted into the adaptive voltage corrector and sequentially corrected by the three built-in components. The corrected voltages have similar degradation trends and smaller data divergence, which helps deep learning to extract potentially similar features under different dynamic conditions. Subsequently, a convolutional neural network is developed to establish a robust and reliable mapping between the corrected voltages and state of health. Finally, cross-validation results across four conditions show that the estimation results maintain an root mean square error around 1.0 % using 10-min discharge segments. The proposed method is expected to help deep learning process complex real-world data by introducing adaptive engineering.