Distributed TinyML on Resource-Constrained IoT Sensor Networks

The transformative integration of IoT (Internet of Things) devices across various industrial platforms, product solutions, and smart systems, etc., has been revolutionizing the ways that we live. While sophisticated smart devices are prevalent in Industrial IoT, we often encounter devices with limited computational resources and other constraints, which may restrict the ability to perform complex data analysis and decision-making operations. In this paper, we present a novel IoT machine learning framework for devices with limited resources - the Distributed TinyML Sensor Network (DTSN) framework. Our goal is to create intelligent, effective, and automated data computations in these devices for the sensor networks that connect to the edges of IoT systems. We have designed a set of function calls for enabling the distributed deployments of neural network models across multiple resource-constraint sensing devices in DTSN. It results in facilitating autonomous data analysis and decision-making while reducing reliance on Cloud services. With the popular Bluetooth technology, Bluetooth mesh networks are utilized for inter-device communications and support dynamic memory management without compromising model precision. Our model offers on-device model training, fast deployment, and provides inferences at an IoT gateway node. The experiment results indicate that the DTSN achieves high accuracy in both regression and classification tasks. It demonstrates the feasibility of training and inference on embedded devices. In conclusion, the DTSN framework provides a new method for deploying neural network models across multiple IoT devices, thus significantly enhancing the system intelligence and autonomy.