A Multi-Task Learning Framework for Signal Detection and Modulation Recognition

Reliable signal detection and automatic modulation recognition (AMR) in non-cooperative scenarios remain major challenges, particularly under low signal-to-noise ratio (SNR) conditions where performance degradation is severe. On one hand existing approaches lack robustness for practical deployment, mainly due to their limited ability to extract discriminative features in noisy environments; on the other hand, they often rely on computationally expensive time-frequency transformations or inefficient sequential architectures, which are susceptible to error propagation. To address these limitations, this paper introduces a unified multi-task learning framework that leverages the selective state space model (Mamba) for joint signal detection and modulation classification. The framework first employs an Overlapping Patch Splitter block to preserve structural continuity in the raw input signal, followed by the Mamba Encoder Block to efficiently extract deep temporal features and form a rich representation. Extensive experiments conducted on a simulated Low-Probability-of-Intercept (LPI) radar dataset demonstrate that the proposed frame-work achieves superior recognition accuracy and computational efficiency compared to the existing methods, with the average probability of detection (PD) of 88.69 % and recognition accuracy (ACC) of$\mathbf{8 2. 8 4 \%}$.