Nowadays, physical complexity, wide dimensions and big data volume have become pivotal keywords in the Structural Health Monitoring (SHM) field, thus posing important challenges in the design of resilient and effective SHM systems. Indeed, spanning from the civil to the industrial field, the urge for sensor network optimization is claimed, aiming at (i) reducing the total amount of information that should be transmitted over the network, (ii) increasing the responsiveness to subtle changes in the structural behaviour and (iii) lowering the latency in the data-to-user transfer process.In this context, the Tiny Machine Learning (TinyML) paradigm, i.e. the implementation of very thin and low power neural networks running on edge devices, has very recently pioneered outstanding solutions to jointly cope with all the above-mentioned issues. Hence, TinyML could bring a radical shift of perspective, moving from cloud-based data analytics, which is usually performed on remote servers in a time and energy-consuming manner, to sensor-near data inference, empowered to smart sensors in charge of processing information in a streaming fashion.Among the multiple tasks involved in the SHM-chain, TinyML founds its natural application to damage detection and classification. In line with this observation, the main contribution of this work is to present the practical embodiment of TinyML architectures on resource-constrained devices, targeted on the health assessment of vibrating structures. The latter comprises all those application domains which can be thoroughly described by frequency-related quantities, conventionally extracted by means of operational modal analysis strategies.More in detail, the health assessment goal is tackled as a binary classification problem, by tracking over time the evolution of frequency-based parameters. Such an approach yielded to the design of two different neural networks, which require natural frequencies of vibration as inputs and provide the corresponding damage status (i.e. healthy/unhealthy) as output. To this end, besides the standard feedforward Autoassociative Neural Network (ANN), the very recent Once Class Classifier Neural Network (OCCNN), which relies on the cascade of an adversarial generator block followed by a dense and fully-connected neural network, were considered. Furthermore, the models were designed to take into consideration the effects of environmental variability.The neural network machines have been initially coded in the Python TensorFlow programming environment; then, their distilled versions have been obtained after conversion to TensorFlow Lite so as to be compliant with embedded devices. Finally, the sought models have been ported on the Arduino Nano 33 BLE Sense platform and validated with experimental data from the Z24 bridge use case, reaching an average accuracy and precision of 92% and 91%, respectively.

Enhancing Vibration-Based Structural Health Monitoring via Edge Computing: A Tiny Machine Learning Perspective

Paper Proceedings: https://asmedigitalcollection.asme.org/qnde/QNDE2021/volume/85529

ASME 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation (QNDE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.