Deployment of Tiny Machine Learning models on a low-end computing platform for vibration-based diagnostics

This thesis explores the practical implementation of Machine Learning (ML) models on embedded systems for vibration-based diagnostics in Structural Health Monitoring (SHM) applications. The proposed approach utilizes a One-Class Classifier Neural Network (OCCNN) model, trained to distinguish between healthy and anomalous conditions. Overall, this thesis establishes the feasibility of implementing OCCNN for vibration-based diagnostics on resource-constrained embedded systems. The research outcomes hold significant implications for enhancing safety, reducing maintenance costs, and improving reliability in diverse structural applications. As Tiny ML continues to advance, this work opens new opportunities for practical implementations in real-world scenarios. The OCCNN model is trained on a combination of healthy and anomalous data and tested with experimental data from the two bridge use cases. The successful deployment of the OCCNN model on the Arduino Nano 33 BLE SENSE Rev2 board demonstrates its applicability in resource-constrained embedded systems. Evaluation metrics, including accuracy, precision, recall, and F1 score, validate the OCCNN model's effectiveness in vibration-based diagnostics, offering promising results for SHM applications.