Smart Objects

Abstract

Growing momentum in embedded systems and the wide use of sensors in everyday life, have motivated significantly, novel research in Internet of Things (IoT) systems and on-device Machine Learning (TinyML) processing. However, limitations in the energy stock and the computational capabilities of resource-scarce devices prevent the implementation of complex ML algorithms in IoT devices, which typically have limited computing power, small memory, and generate large amounts of data. This paper, aims to research and exploit the TinyML emerging technology for embedding intelligence in low-power devices, towards next generation IoT paradigm and smart sensing, in the context of SHM. In particular, the purpose is to provide integrated SHM functionality in plastic structures and thus make them "conscious" and self-explanatory (smart objects), by being able to localize any occurring impacts on the structure. We implement and benchmark Random Forest and Shallow Neural Network models on Arduino NANO 33 BLE, using an experimental dataset of piezoelectric sensor measurements concerning impact events in a thin plastic plate. The classification and model footprint results, 98.71% - 8KB and 95.35% - 12KB of accuracy and flash memory size for each model respectively, are very promising and constitute a solid baseline for motivating our concept.