Adaptive Online-Learning Digital Twin for Edge-Based Prediction of Process Stability in Thin-Wall Machining
TL;DR
An edge-deployable digital twin framework that incorporates online machine learning for real-time monitoring of process stability and surface quality and enables continuous model adaptation to non-stationary machining dynamics, thereby eliminating the need for complete model retraining.
Abstract
The manufacturing industry continues to face challenges in preventing chatter-induced defects and maintaining surface quality during machining of thin-walled components. This paper presents an edge-deployable digital twin framework that incorporates online machine learning for real-time monitoring of process stability and surface quality. The proposed approach enables continuous model adaptation to non-stationary machining dynamics, thereby eliminating the need for complete model retraining. The framework deploys Adaptive Random Forest classifiers and regressors trained on nonlinear dynamic features of cutting force and vibration signals gathered during Al2014 alloy thin-wall milling experiments. The results indicate that the Adaptive Random Forest provides a prediction accuracy over 86% in chatter detection tasks and high coefficient of determination (R ${}^{2}$ ) scores above 0.92 in surface roughness estimation tasks. These models also deliver low inference latencies and high online adaptation speeds, making them appropriate for implementing adaptive digital twins. The proposed architecture was effectively deployed on a resource-constrained edge platform. The deployment results indicate potential for real-time, modular, and scalable implementation of adaptive digital twins in dynamic manufacturing settings.