Retired battery capacity screening based on deep learning with embedded feature smoothing under massive imbalanced data

Repurposing retired batteries is a pivotal solution to achieving carbon neutrality and optimizing resource allocation within the transportation and automotive industries. Accurate capacity estimation plays a definitive role in efficiently screening and reutilizing these retired batteries. However, the intricate and varied conditions of retired batteries in real-world applications can introduce challenges prominently characterized by the imbalanced properties of these massive and various batteries. Here, we present a capacity estimation method with adaptive feature engineering tailored to massive real-world battery data. First, a comprehensive feature base is established to identify optimal features for battery degradation level description. Then, an estimation model rooted in a modified ResNet-50 neural network is fortified by a unique feature distribution smooth technique to enhance learning efficacy within the challenging milieu of data imbalance. The proposed model can yield a test root-mean-square error of less than 0.2 Ah for a dataset encompassing over 30 million collected battery testing records. To the best of our knowledge, the developed model shows the first concerted effort to address the intricate task of capacity estimation with real-world massive imbalanced data for retired battery capacity screening applications.