Convolutional Deep Unfolding Network for Asynchronous Active User Detection and Channel Estimation

Grant-free random access has garnered considerable attention in recent years. However, the absence of scheduling means user signals might arrive asynchronously at the base station. Fortunately, sporadic activity and restricted latency transform active user detection and channel estimation boils down into the sparse linear inverse problem of compressive sensing, which can be addressed by deep unfolding networks, e.g., Learned Iterative Shrinkage Thresholding Algorithm (LISTA). Nonetheless, as the maximum user delay rises, the number of training parameters escalates exponentially. To address this issue, we analyze the structural characteristics of asynchronous sensing matrices and design a convolutional network to incorporate partial structures of LISTA, ensuring that the number of training parameters does not grow with increased maximum delay. An autoencoder network for pilot optimization is investigated to fur-ther enhance the performance. Experimental results verify that the convolutional network-assisted LISTA with pilot optimization can achieve enhanced performance.