Transmitter-Oriented Dual-Mode SWIPT with Deep-Learning-Based Adaptive Mode Switching for IoT Sensor Networks

In this article, we propose a dual-mode simultaneous wireless information and power transfer (SWIPT) system with a deep-learning-based adaptive mode switching (MS) algorithm to exploit both advantages of single-tone and multitone SWIPT. For self-powering of low-energy Internet-of-Things (IoT) devices, a duty-cycling operation is used with nonlinear energy harvesting. For this, we employ a new energy-assisted single-tone modulation which simplifies the receiver structure for information decoding. Considering the symbol-error rate performance, we formulate an adaptive MS problem to maximize the achievable rate under the energy-causality constraint by adjusting the MS threshold. To relieve the computational burden of the receiver, we introduce asymmetric processing for adaptive MS, for which the transmitter adapts the communication mode based on the feedback from the receiver. We invoke deep learning for adaptive MS at the transmitter that iteratively updates the MS threshold in a long-term scale via deep long short-term memory (LSTM) recurrent neural network (RNN) while deciding on the communication mode and modulation index in a short-term scale. We demonstrate the achievable rate improvement under an energy-neutral operation while providing interesting insights into designing the adaptive MS algorithm for the dual-mode SWIPT system.