Early Warning of Multiple Battery Faults Based on Hybrid Deep Neural Networks and Transfer Learning

Battery fault early warning is essential for the safe operation of electric vehicles, yet early-stage faults remain highly concealed and difficult to detect. While traditional threshold methods only trigger alarms when parameters exceed fixed limits, they typically lack the predictive capability required for timely intervention. To address this, this paper proposes a robust multi-fault diagnosis method based on a hybrid deep neural network and transfer learning. Extensive real-world operational data are preprocessed through cleaning, sorting, and interpolation to ensure high data reliability. A hybrid architecture combining convolutional neural networks and long short-term memory networks is developed to extract spatiotemporal fault features, with transfer learning-based fine-tuning used to enhance generalization. By selecting inputs via Pearson correlation coefficients, the model outputs voltage and temperature indicators that reflect abnormal battery conditions. Based on the statistical distribution of prediction errors, multi-level thresholds are established to enable the online diagnosis of over/under-temperature, over/under-voltage, and low state-of-charge faults. Vehicle tests demonstrate that the method is robust under varying temperatures and driving behaviors. Compared to traditional threshold-based alarms, the proposed approach significantly improves early warning times by up to 6,950 seconds while effectively reducing the risks of missed and false detections.