Dual-Sparse Transformer Based Deep Collaborative Device Activity Detection for Massive GF-RA in Satellite IoT

Given the heterogeneous access requirements of massive devices in satellite Internet of Things (IoT), we propose a deep learning-assisted collaborative device activity detection scheme for grant-free random access (GF-RA). In the proposed scheme, the device activation probability is not required to be known in advance, since it can be accurately predicted by means of the designed sparsity estimation module. By leveraging the current activity information to optimize the network structure, we further develop a dual-sparse transformer architecture for the efficient identification of active devices, which can significantly improve the model generalization capability while reducing the computational redundancy. Simulation results reveal the feasibility of our scheme in large-scale GF-RA scenarios, and demonstrate its remarkable performance superiority in terms of successful detection probability over the state-of-the-art ones.