Under complex working conditions, the fault characteristics of the power battery of new energy vehicle (NEV) are weak, the data distribution is quite different, and the labeled samples are scarce. In this paper, a fault diagnosis model integrating multi-source monitoring data and transfer learning is proposed. This model constructs a multi-source data fusion framework that concurrently extracts voltage, temperature, and operating condition features using convolutional neural networks (CNN), graph convolutional networks (GCN), and long short-term memory networks (LSTM). It incorporates an attention mechanism to achieve adaptive weighted fusion. Furthermore, a domain adversarial transfer learning module is designed to learn domain-invariant features, mitigating distribution shifts between the source domain (richly labeled) and target domain (sparsely labeled). The experiment was conducted using NASA and real-vehicle datasets, comparing traditional Support Vector Machines (SVM), single-source CNN, and classic MMD (Maximum Mean Discrepancy) transfer learning methods. Results demonstrate that the proposed MS-TL model achieves 93.7% accuracy and an F1-Score of 90.8% in the target domain, outperforming the best baseline by 5.3 and 6.3 percentage points respectively. T-SNE visualization proves that it can effectively align cross-domain features and maintain high sensitivity in early weak fault identification. The research provides a new idea of high robustness and low labeling dependence for predictive maintenance of power batteries.
Cai et al. (Sun,) studied this question.