TFT-MTL: A Time–Frequency Transformer with Multi-Task Learning for Radar Working Mode Recognition in Complex Multipath Environment

Abstract

Radar working mode recognition is crucial for characterizing the functional state of multifunction radar systems. However, in complex electromagnetic environments, multipath propagation and noise distort time-frequency representations, exacerbating feature aliasing and weakening the robustness of recognition. To address this issue, this paper models multipath channels and constructs a dataset of wavelet-based multiscale time-frequency maps. Based on this, the Time-Frequency Transformer with Multi-Task Learning (TFT-MTL) framework is proposed. This method employs a dual-axis time-frequency Transformer to jointly model frequency-dimensional dependencies and temporal structure, while introducing multi-task learning to collaboratively optimize working mode classification and physical parameters regression. Experimental results show that the proposed method achieves high recognition accuracy across different Signal-to-Noise Ratio (SNR) and various interference conditions. It is effective for recognizing multifunctional radar working modes in complex environments.