Prior Information Aided Deep Learning Method for Grant-Free NOMA in mMTC

Data-Driven Compressed Sensing for Massive Wireless Access

“…Training a network with data of mixed settings would be less time-consuming, while its performance is usually limited. Compared with LISTA, LISTA-P [9] utilizes the errorchecking information as the prior information of the network and achieves outstanding performance.…”

“…1, different scenarios lead to different sparse structures in the massive access problem. For a BS equipped with a single antenna, the massive access problem is detecting active users and estimating their channel with one measurement [9], [11]. For a BS equipped with multiple antennas, we have multiple measurements and activity detection and data recovery is constructed as a row-sparse matrix recovery problem [7], [12], [14].…”

“…Another characteristic is considering the correlation embedded in the real and imaginary parts of the complex signal. As NNs cannot directly deal with complex signals, a common method is to flatten a complex signal into a real concatenated signal [9], [12], [13]. Yet, this neglects the correlation between the imaginary and real parts and results in some performance degeneration.…”

“…In [9], the checking mechanism is used to obtain the indices of successfully detected active users. This information is then used in the activity detection again, which corresponds to a sparse recovery problem with a known partial non-zero indices set.…”

“…Fig.3. Structures of deep-unfolding based networks, including LISTA[3], LISTA-P[9], and LISTA-H[11].…”