Energy efficient Firmware Over The Air Update for TinyML models in LoRaWAN agricultural networks

Abstract

Current agricultural practices fail to feed everyone correctly while being harmful to the environment and highly sensitive to climate change. Therefore, new and modern agriculture needs to be developed and overcome numerous challenges. Artificial Intelligence (AI) is one of the tools widely used in this new type of Agriculture called Precision Agriculture (PA) or Smart Farming (SF). Thanks to the Internet of Things (IoT) technologies, AI algorithms are fed with a vast amount of data to provide valuable insights for farmers, such as weather prediction, pest development detection, irrigation management, etc. AI algorithms are often executed in cloud servers, thus requiring IoT devices to offload their data to process. This creates privacy, latency, and security issues, but mostly, it requires a large quantity of energy for transmission. To overcome those issues, recent research brought new tools like Tiny Machine Learning (TinyML) to perform AI directly on the IoT devices and break free from the cloud. Despite promising results, such Smart devices cannot train new models on their constrained hardware and therefore need frequent updating to increase the model's accuracy over time, regarding the specific environment where the sensor is deployed. In the Agricultural domain, sensor devices are numerous and usually spread over vast geographical areas while running on battery. For this reason, Farms Wireless Sensor Network (WSN) use mostly Low Power Wide Area Network like LoRaWAN to communicate. Therefore Firmware Update Over the Air (FUOTA) is required. In this context, this paper proposes a study of the FUOTA process for a TinyML model using LoRaWAN in a specific agricultural scenario. A TinyML sensor prototype was built to evaluate the feasibility of FUOTA for tinyML devices using LoRaWAN. The system's energy consumption and Packet delivery ratio are then analyzed in a simulator with different network scenarios.