Deep learning-based dynamic error correction and uncertainty estimation for digital twin-assisted fringe projection profilometry of rotating gears

TL;DR

This paper introduces a deep learning approach with uncertainty estimation for dynamic gear measurement using fringe projection profilometry, leveraging simulated datasets and transfer learning to improve accuracy and reliability over traditional methods.

Contribution

It develops a novel Concrete Dropout-Pixel wise Uncertainty Network with transfer learning for enhanced dynamic gear measurement and uncertainty estimation.

Findings

Significant accuracy improvements over traditional methods.

Effective use of simulated datasets for training.

Reliable pixel-level uncertainty estimation.

Abstract

This paper presents a deep learning-based method for dynamic gear measurement and uncertainty estimation. A twin-system proposed on the Unity platform is utilized to flexibly generate diverse simulated datasets. This effectively addresses the scarcity of real-world gear measurement data and facilitates verification of network performance.The designed Concrete Dropout-Pixel wise Uncertainty Network integrates the Concrete Dropout mechanism for pixel-level uncertainty estimation. Two lightweight layers are employed in the output layer to enhance the spatial continuity of prediction results.During network training, a transfer learning strategy is adopted: the model is first pretrained with a small amount of three-phase-shifting (3-PS) data, then fine-tuned on the target gear measurement dataset. Experimental results demonstrate that, compared with the traditional three-step phase-shifting (3-PS) method, the proposed approach achieves significant improvements in phase prediction accuracy, three-dimensional reconstruction accuracy, dynamic error correction capability, and uncertainty estimation reliability.This work provides a practical and efficient technical solution for fringe projection-based dynamic gear measurement.