ResNet-MFII-PN for imbalanced small-sample lithium-ion battery fault diagnosis with a physics-based model generative real-word dataset

Accurate fault diagnosis of lithium-ion batteries (LIBs) is critical for ensuring the operational safety and reliability of electric vehicles (EVs). However, most existing diagnostic methods rely on large-scale fault datasets, which are often unavailable in real-world applications due to data scarcity and class imbalance. To address this limitation, this study proposes a novel ResNet-MFII-PN framework, integrating a Multi-Factor Imbalance Index (MFII) and a Prototypical Network (PN) anomaly detector to residual neural network (ResNet) for multi-fault diagnosis under imbalanced, small-sample conditions. The MFII mechanism dynamically evaluates feature variations and class imbalance during training to enhance minority-class fault recognition, while the PN detector constructs normal-state prototypes for distance-based fault detection, improving recall without increasing false alarms. More importantly, to validate our approach, a comprehensive real-world LIB fault dataset is established, comprising voltage, temperature, and state-of-charge (SOC) signals across diverse fault types and operating conditions, with over 80 % fault-labeled samples. Experimental results demonstrate that the proposed method achieves a minority-class accuracy of 96.23 % and an overall accuracy of 96.07 %, outperforming conventional approaches such as LSTM and Transformer models. The framework's robustness in handling imbalanced data is further verified, establishing a new benchmark for practical LIB fault diagnosis. This work provides a reliable solution for early fault detection in safety-critical applications.