Lithium-ion battery-powered trucks, as an emerging category in the field of heavy-duty electric commercial vehicles, their in-depth application is of great significance for China to accelerate the implementation of the “dual carbon” goals in the transportation sector. However, heavy-duty electric commercial vehicles have the problem of false battery status alarms during high-current charging, which leads to unnecessary shutdowns, low accuracy in fault diagnosis and low operational efficiency. To address the above issues, this paper proposes a battery fault detection method based on the fusion of convolutional neural network (CNN), long short-term memory network (LSTM), and dynamic autoencoder (DYAD), which is used to solve the problems of battery state changes and cumulative fault effects. The encoder-decoder architecture is adopted. The spatial features of battery data are extracted by CNN, the dynamic evolution law of time series is captured by LSTM network, and the nonlinear characteristics of the battery system are processed with DYAD information entropy, thereby achieving deep learning of complex fault modes. The model architecture effectively extracts spatio-temporal features through the integration of CNN and LSTM, introduces multimodal decoupling technology to monitor key feature errors in real time, and combines global interpretability analysis and potential spatial visualization, significantly enhancing the credibility and transparency of the model. The experimental results show that this method achieves an area under the receiver operating characteristic curve (AUROC) of 88.7% on the real dataset, which is 26.2% higher than that of the graph dynamic network (GDN). By reconstructing error analysis, false alarms caused by high-current charging events have been significantly reduced, and early fault detection has been achieved.