MTL-CNET: An Advanced Integrated Sensing and Communication Framework Utilizing Complex Neural Networks and Multi-Task Learning

Integrated sensing and communication (ISAC) is anticipated as a “game-changer” for the sixth generation systems by enhancing both efficiency and reliability. However, critical challenges remain, including performance degradation due to the separate treatment of sensing and communication, ineffective processing of complex-valued wireless signals, and limited adaptability to dynamic wireless environments. To address these limitations, this paper explores orthogonal delay-Doppler division multiplexing (ODDM)-enabled ISAC with passive sensing (ISAC-PS) systems and proposes a novel ISAC framework, called MTL-CNET, which leverages multi-task learning and complex neural networks. MTL-CNET can facilitate sensing and communication tasks concurrently while improving robustness to environmental variations. Specifically, we develop a novel three-dimensional complex convolutional neural network (3D-CCNN) approach to process complex-valued ODDM signals in a singular architecture, as opposed to two separate networks in conventional methods. The 3D-CCNN effectively captures the characteristics of real and imaginary components of signals and the intrinsic relationship between them for ISAC-PS tasks. Moreover, we propose a multi-task learning framework with inter-layer information exchange (MTL-IE), which achieves simultaneous sensing and communication enhancement by using the property of MTL and information exchange design. Additionally, a novel transfer-learning-based self-updating scheme is proposed to enhance resilience to environmental fluctuations with minimized updating frequency and effort. Experimental results demonstrate that MTL-CNET outperforms state-of-the-art methods in sensing accuracy, communication reliability and environmental robustness.