With the acceleration of the development of intelligent and informatization of military equipment, lithium batteries, as one of the key energy supply sources of artillery equipment, have a reliability that greatly affects the combat effectiveness of artillery equipment. Accurate prediction of the health status of lithium batteries has become a key technology for improving artillery combat effectiveness and upgrading the intelligent system. In order to improve the accuracy of lithium battery degradation process prediction, we propose a multi-task fusion prediction method based on multi-scale features. This method uses battery capacity and internal resistance as multi-task prediction goals, and uses a pseudo-twin architecture to downsample the original one-dimensional data and the two-dimensional data obtained by wavelet transformation to capture potential long-term and short-term degradation characteristics; then uses the cross-attention mechanism to fuse the downsampled one-dimensional and two-dimensional time series information, and inputs the CNN-GRU combination network to achieve joint prediction of lithium battery capacity and internal resistance degradation trajectory, effectively overcoming the problem of data singularity and enhancing the robustness of the model. The root mean square error of the predicted value and the true value of the experiment on the public data set CALCE is 0.0269, and the model is fine-tuned on the collected artillery battery data through transfer learning, with an RMSE value of 0.0800. The experimental results fully verify that the proposed method significantly improves the prediction accuracy of the battery degradation process, provides new solutions for the integration of inference training and self-learning processes, and lays a solid foundation for improving the reliability and intelligence level of equipment.