Attention-Driven Long-Time Integration for Weak Maneuvering Targets in Integrated Sensing and Communication System

Integrated sensing and communication base stations are widely used for low-altitude target detection. However, detecting weak and maneuvering targets remains challenging due to range and Doppler migration during long time coherent integration, especially when targets exhibit multiple motion modes. Traditional methods, such as constructing motion compensation filters via parameter search or employing time-frequency analysis, face difficulties in balancing computational complexity with integration efficiency. More importantly, during long integration periods, targets may exhibit multiple motion modes, whereas most existing algorithms assume a single motion pattern, limiting their effectiveness in such cases. Inspired by the short-time Fourier transform, we propose an AI-driven long-time coherent integration (LTCI) module method based on an attention-enhanced gated recurrent unit network. It features a motion segmentation module to capture inter-pulse dependencies and a block-based motion compensation module to generate segment-wise phase compensation matrices. After phase correction, the echoes over the entire integration period are coherently accumulated through the LTCI. The proposed method avoids exhaustive parameter search and adapts to complex motion patterns. For a target exhibiting four motion segments during the accumulation period, the proposed method achieves an integration gain approximately 3 dB higher than that of the classical generalized Radon Fourier transform method. Both simulations and real-world experiments demonstrate its superiority over conventional approaches in terms of integration gain and detection probability.