Duty cycling and wake-up protocols

This paper treats IoT deployments in which IoT devices (IoTDs) rely solely on EH to sense and transmit information about events/alarms to a base station (BS). The objective is to effectively manage the duty cycling of the IoTDs to prolong battery life and maximize the relevant data sent to the BS. Ths BS can also wake up specific IoTDs if extra information about an event is needed upon initial detection. We propose a K-nearest neighbors (KNN)-based duty cycling management to optimize energy efficiency and detection accuracy by considering spatial correlations among IoTDs’ activity and their EH process. We evaluate machine learning approaches, including reinforcement learning (RL) and decision transformers (DT), to maximize information captured from events while managing energy consumption.

Figures on the left show the mean number of IoTDs with lowenergy availability per TTI and the mean battery level of the IoTDs per TTI. These capture the IoTD availability for sensing and transmission. These performance metrics stay around the same level for both the random duty cycling and the genieaided benchmarks. Meanwhile, the mean number of IoTDs with low-energy availability decreases with the device density for the KNN-based and RL-based proposals from around 67% for 10 IoTDs to 8%-13% for 250 IoTDs. Similarly, the mean IoTD battery level per TTI increases from around 10 units to 72-86 units in high-density scenarios. For the optimized duty cycling proposal, the number of devices with low-energy availability fluctuates around 30%- 45% while the mean battery level is around 31-54 units. Moreover, for the proposal based on DT, the IoTDs with low energy availability also decrease in a 50%, and the mean battery is always above 40 units per IoTD, increasing to almost 88 units per IoTD for high-density scenarios.

The top figure on the left shows this wrong activation of IoTDs while the corresponding energy level wasted due to the previous situations is shown in bottom figure on the left. The genie-aided benchmark does not waste any energy since it always makes the best decision at first. Note that the energy wasted per IoTD decreases with the device density, but also more wrong activations occur. Notably, the optimized duty cycling leads to around 17 wrong activations per event for high-density scenarios while the RL-based proposal leads to 9.5 wrong IoTD activations pursuing a higher amount of information per event while the KNN-based proposal leads to less than 6 wrong activation. Using the DT-based proposal, the number of wrong IoTDs activations per event decreases by 50%, outperforming the other proposed methods and lowering the energy wasted per event below 1.5 units per IoTD despite the device density, as shown in bottom figure on the left).