Aiming at the time-series dependence and dynamic characteristics in radar target attitude angle prediction, this paper proposes a neural network model integrating multi-scale convolution and deep bidirectional LSTM. A time-series dataset is constructed from polarized RCS signals, and local-global features are extracted using parallel multi-scale convolutions (Progressive Odd-Kernel Multi-Scale Convolution). After additive fusion, these features are fed stepwise into a four-layer bidirectional LSTM network to model long-term dependencies. Finally, the three-dimensional attitude angles (radar attitude angle, missile axis azimuth angle, missile axis elevation angle) are output through the fully connected layer. The small-batch RMSE of the multi-scale fusion network is optimized from 1.41 to 0.13, and the validation RMSE is reduced to 0.42. The prediction accuracies reach 97.66% (radar attitude angle), 95.44% (missile axis azimuth angle), and 98.93% (missile axis elevation angle), which are significantly better than those of traditional single-modal models. The results show that the model converges stably on the validation set. This method achieves high-precision dynamic prediction of radar target attitude angles through deep fusion of spatiotemporal features and can provide key technical support for trajectory tracking and landing point prediction in air defense systems.