Knowledge-Driven Signal Detector for Uplink Transmission in IoT Networks With Unknown Channel Models

In this article, an uplink signal detection problem is considered for Internet of Things (IoT) networks. Owing to the imperfections of IoT devices, including I/Q imbalance and amplifier nonlinearity, exact end-to-end channel models and accurate channel state information (CSI) are typically unavailable at the receiver, which obstructs the application of traditional model-based signal detection algorithms. A consensus has been reached recently that deep learning (DL) is a promising tool to cope with this problem. However, for the IoT scenarios under consideration, devices typically transmit data using short packets with few pilot symbols, the amount of which is insufficient for each device to individually train a detector. In order to combat the data scarcity barrier and enable few-shot learning, a novel training paradigm is proposed where pilot symbols from different devices are aggregated in an intelligent manner to train a universal signal detector. Specifically, this article devises a knowledge-driven signal detector architecture following the modular design methodology typically used in classical communication system receivers. Under this framework, three neural networks (NNs), a signal classifier, a channel feature extractor, and a signal feature extractor are created to form decision statistics and produce estimates of the transmitted symbols. Furthermore, borrowing ideas from domain adaptation, a novel component referred to as a link discriminator is integrated into the architecture to improve its generalizability. The proposed signal detector exploits pilot symbols from various IoT devices to train a universal detector that can be applied to different channel conditions without retraining, including those not seen in the training phase. Simulation results verify the superiority of the proposed knowledge-driven detector compared with existing solutions in the sense that it enjoys higher detection accuracy and can be well trained with less data.