Machine learning-based digital twin development for real-time monitoring of mechanical processes

Abstract

Abstract. Machine learning (ML) and digital twin (DT) technology convergence hold revolutionary possibilities to real-time condition-based monitoring and predictive maintenance of complex mechanical systems. The current paper describes a new ML-based digital twin framework that incorporates a hybrid architecture of Long Short-Term Memory (LSTM) networks and a collection of Gradient Boosted Trees (GBT) to make the physical mechanical processes and their virtual counterparts to synchronize in high-fidelity and in real-time. The described framework is based on a multi-sensor fusion approach, which processes the vibration, temperature, acoustic emission, and torque as measured on a CNC milling testbed during changing cutting conditions. The digital twin can forecast tool wear development with a Root Mean Square Error (RMSE) of 4.3 µm and identify process states associated with an aberrant state with an F1-score of 0.964. Statistically significant predictive accuracy and latency improvement over standalone LSTM, Convolutional Neural Networks (CNN) and classical Support Vector Machine (SVM) baselines can be established through comparative evaluation. The modular architecture of the framework enables it to run on edge computing platforms with a latency of 11.7 ms to perform a real-time inference. These findings validate the suggested solution as an industry-level scalable solution to next-generation cyber-physical manufacturing systems.