A Tiny Convolutional Neural Network driven by System Identification for Vibration Anomaly Detection at the Extreme Edge

Vibration data analysis is the driving tool for the Structural Health Monitoring (SHM) of structures in the dynamic regime, i.e., structures showing important oscillatory behaviours, which largely dominate the transportation back-bone: from terrestrial/aerial vehicles (e.g., trains, aircraft, etc.) to the supporting infrastructures (e.g., bridges, viaducts, etc.). Outstanding opportunities have recently been disclosed in the field of Intelligent Transportation Systems (ITS) by the advent of sensor-near processing functionalities, eventually empowered by Artificial Intelligence (AI). The latter allow for the extraction of damage-sensitive features at the extreme edge, without the need of transmitting long time series over the monitoring network. In this work, we explore for the first time a novel anomaly detection workflow for on-sensor vibration diagnostics, which combines the unique advantages of embedded System Identification (eSysId) as a data compression strategy with the computational/energy advantages of Tiny Machine Learning (TinyML). Experimental results conducted on a representative SHM dataset demonstrate that the proposed pipeline can achieve high classification scores (above 90%) for the health assessment of the well-known Z24 bridge. In particular, the minimal inference time (less than 44 ms) and power consumption performed while running on three different general-purpose microprocessors make it a promising solution for the development of the next generation of SHM-oriented ITS.