Physics-Constrained Multi-Task Learning for Intra-Pulse Modulation Recognition and Parameter Estimation of Radar Signals

Background In complex electromagnetic environments, intra-pulse modulation recognition of radar signals faces significant challenges, particularly in terms of poor adaptability to low signal-to-noise ratios (SNR) and low accuracy in joint multi-parameter estimation.

Purpose To address these issues, this paper proposes a radar signal intra-pulse modulation recognition and parameter estimation method based on physics-constrained multi-task learning.

Methods First, a feature extraction backbone network is constructed based on the attention mechanism. By extracting physical features in parallel, such as time-domain rise time and energy compactness, as prior constraints, the network is guided to focus on the essential attributes of the signal to enhance noise robustness and interpretability. Second, a multi-task shared encoding layer and a hierarchical feature utilization architecture are designed to jointly optimize the intra-pulse modulation classification and pulse parameter estimation tasks. Furthermore, a signal time-frequency reconstruction decoder is introduced to improve the completeness of the features. Finally, a dynamically weighted loss function is utilized to balance the multi-task learning process.

Results Simulation results demonstrate that under low SNR conditions of -10 dB and -5 dB, the recognition rates for five types of typical radar signals reach 88.57% and 99.11%, respectively. Additionally, the parameter estimation errors are significantly lower than those of traditional methods.

Conclusions The proposed method effectively improves joint recognition performance in complex environments, demonstrating high potential for engineering applications.