Hybrid Driven Learning for Joint Activity Detection and Channel Estimation in IRS-Assisted Massive Connectivity

• Published in: IEEE transactions on wireless communications 2024-09, Vol.23 (9), p.10834-10849
• Main Authors: Zheng, Shuntian, Wu, Sheng, Jia, Haoge, Jiang, Chunxiao, Kuang, Linling
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Data models; Deep learning; Hybrid driven; joint device activity detection and channel estimation; Machine learning; massive machine-type communication; Message passing; Noise levels; Performance degradation; Performance evaluation; Reconfigurable intelligent surfaces; Signal processing algorithms; Vectors; Wireless communication
• ISSN: 1536-1276; 1558-2248
• DOI: 10.1109/TWC.2024.3376381

We consider the uplink connectivity for massive machine-type communications (mMTC) assisted by intelligent reconfigurable surfaces (IRSs), where device activity detection (DAD) and channel estimation (CE) are challenging due to limited pilot sequences. Moreover, differentiation among device types causes channels to deviate from the assumed characteristics, leading to performance degradation of conventional compressive sensing (CS) algorithms. To this end, two innovative networks driven by the hybrid of model and data are proposed exploiting the iterative frameworks and deep neural networks. We first present a hybrid driven iterative shrinkage thresholding algorithm, dubbed HISTA-Net, where a dual attention network (DAN) is embedded within the iterations to adaptively suppress iterative noise and enhance sparsity properties. Subsequently, we encapsulate data driven network and the intrinsic channel matrix knowledge, and derive a hybrid driven approximate message passing network (HAMP-Net) to further improve the sparse recovery performance. Our experiments demonstrate that the proposed networks outperform existing CS methodologies and deep learning strategies in accuracy, convergence, and generalization ability. Remarkably, the proposed HAMP-Net reduces pilot overhead by 30%, and achieves an NMSE gain of 3 dB for signal-to-noise ratios exceeding 15 dB.