Generalized Likelihood-Ratio Enabled Machine Learning for UE Detection over Grant-free SCMA

In this work, we consider an uplink grant-free sparse coded multiple access (GF-SCMA) system that superimposes the transmissions from up to J user equipments (UEs) onto K resource elements (REs). A critical issue for GF-SCMA is that data retrieval is performed without the knowledge on the activeness of each UE. The detection accuracy of UE statuses thus becomes a dominant factor in data retrieval performance. At this background, a generalized likelihood ratio (GLR) enabled convolutional neural network (CNN) scheme is proposed for UE status detection. In order to reduce detection complexity, K parallel CNN structures are used, each of which decides the status of the UE associated with a respective RE. Mismatched decisions are resolved by the soft outputs from each CNN classifier. Simulation results show that the proposed GLR-enabled parallel CNNs with soft mismatch resolution can achieve 10-5 detection error probability at moderate to high SNRs, regardless of time correlation characteristic of channel gains, when the Rayleigh amplitude distortion of the channel can be removed by a sophisticate power control mechanism.