AI-assisted high-precision dynamic error calibration for on-machine measurement systems

Description

In precision optics manufacturing, transferring components between multiple machines introduces uncertainties due to datum changes. On-machine surface measurement (OMSM) integrates a non-contact profiler directly with the manufacturing platform, overcoming this bottleneck. However, dynamic measurement errors, arising from both the host machine and the integrated subsystem, significantly degrade the measurement accuracy. Traditional offline system identification methods struggle to capture the overall system response using a generalised model under time-varying dynamics and uncertain model orders. This study proposes a novel AI-assisted calibration approach based on deep learning to enable adaptive error compensation in OMSM systems. Specifically, a regression model using Long Short-Term Memory (LSTM) networks is developed to learn the complex temporal dependencies and nonlinear dynamics inherent in the system. The experimental results demonstrate that the LSTM-based calibration achieves comparable accuracy to traditional methods while offering greater adaptability, automation, and scalability. In a practical test on an off-axis parabolic (OAP) surface, the model significantly reduced the root mean square (RMS) of measurement error from 1.9073 µm to 0.4324 µm by compensating for dynamic errors caused by vibrations and fluctuating conditions. These findings highlight the potential of LSTM networks to improve the efficiency and robustness of OMSM systems, contributing to the development of intelligent and adaptive error calibration frameworks for smart manufacturing.

Period	11 Jun 2025
Event title	euspen’s 25th International Conference & Exhibition
Event type	Conference
Location	Zaragoza, SpainShow on map
Degree of Recognition	International