Deep Learning Metrology in Industrial 4.0

Yang, Ru

doi:https://doi.org/10.21985/n2-rtbm-k251

Work

Deep Learning Metrology in Industrial 4.0

Public

Download PDF

In recent years, we have seen the embryo of Industry 4.0 which has been promoting manufacturing processes towards the future with better efficiency, higher accuracy and better reliability. However, manufacturing precision has been restricted by the precision of metrology and material characterization. In other words, one can only manufacture parts as precisely as one can measure them. Thus, this research intends to leverage the deep learning techniques, one of the enablers in Industry 4.0, for developing new computer vision-based metrology for intelligent manufacturing. Conventionally, machine learning has already been adopted as an effective tool in computer vision metrology, such as wafer quality inspection, material defect detection, scanning electron microscope (SEM) image de-noising, etc. However, current vision-based metrology methods rely extensively on the existing machine vision tasks, which are not specifically designed for manufacturing. This work presents my effort in bridging the fundamental knowledge in manufacturing metrology and deep learning techniques, thereby innovating vision-based metrology to achieve higher accuracy and efficiency as well as unprecedented capabilities with economically efficient hardware.This work addresses two immediate challenges in modern manufacturing: in-process 3D measurement and vision-based material characterization. Integrating characterization into manufacturing pipeline has been beneficial in many aspects including but not limited to: realizing in-situ characterization of manufactured products, cost reduction, and accelerating product iteration. Material characterization addresses another important and basic need of most manufacturing processes, which is understanding the material properties. Vision-based material characterization has the advantage of being non-invasive and fast implementation. Its small footprint also makes it valuable to broader audience. Chapter 1 of this thesis will focus on introducing the background of current 3D measurement and material characterization processes, as well as their limitations. Chapter 1 will also briefly discuss the potentiality of implementing deep learning technologies in those fields. In Chapter 2, I will introduce my effort in developing a surface 3D reconstruction with deep-learning-based point-light photometric stereo (DPPS). DPPS takes multiple images of a surface under different illuminations as inputs and uses a convolutional neural network (CNN) to recover the surface normal and height maps based on a data-driven reflectance model. As an extension of traditional photometric stereo, the parallel illumination assumption is replaced with a more realistic point-light source. The proposed approach will work with metallic surfaces with unknown surface roughness within the calibrated range from specular to diffusive. By adopting a deep learning model, the reconstruction of 3D geometry can be reduced to tens of milliseconds, making the in-process 3D measurement possible. In Chapter 3, I will discuss the deep-learning vision-based approach to improve the strain measurement and predict material properties. Instead of extracting deformation through digital image correlation (DIC), I propose to use two CNNs to predict the displacement and strain fields separately. This Deep DIC method not only exhibits high accuracy and robustness but also shows potential to real-time strain estimation. In Chapter 4, an extension of Deep DIC is present which is able to predict the discontinuous displacement field that is often encountered in CFRP (carbon fibre reinforced plastics) materials. Chapter 5 of this thesis briefly introduces my feature research work that falls under the theme of this thesis. Chapter 6 summarizes the contributions of the thesis.

Creator