Multi-Task Deep Learning With Transpose Processing Pipeline for Joint Modulation Classification and DOA Estimation

Joint automatic modulation classification (AMC) and direction of arrival (DOA) estimation are critical for signal intelligence, but multi-task learning solutions are often hindered by performance-degrading feature misalignment between the two tasks. This paper proposes TP-Net, a novel multi-task DL model featuring a transpose processing pipeline for simultaneous AMC and DOA estimation. TP-Net employs a Y-shaped architecture with shared convolutional blocks, attention mechanisms, and task-specific branches. As the key innovation, the transpose processing pipeline strategically rearranges input data to dynamically align extracted features for optimally learning both AMC and DOA tasks. Evaluated on a synthetic dataset, TP-Net demonstrates robust performance by achieving high classification accuracy over 99% at and precise DOA estimation with RMSE less than 0.5° at 6 dB, accordingly outperforming existing joint task models while maintaining low computational complexity.