Date of Award

2019

Publication Type

Master Thesis

Degree Name

M.Sc.

Department

Computer Science

First Advisor

D. Wu

Second Advisor

R. Maev

Keywords

applied AI, artificial intelligence, computer vision, deep learning, machine learning, spot weld

Rights

info:eu-repo/semantics/openAccess

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Abstract

Resistance spot welding is a widely used process for joining metals using electrically generated heat or Joule heating. It is one of the most commonly used techniques in automotive industry to weld sheet metals in order to form a car body. Although, industrial robots are used as automated spot welders in massive scale in the industries, the weld quality inspection process still requires human involvement to decide if a weld should be passed as acceptable or not. Not only it is a tedious and error- prone job, but also it costs industries lots of time and money. Therefore, making this process automated and real-time will have high significance in spot welding as well as the field of Non-destructive Testing (NDT). Research team in Institute of Diagnostic Imaging Research (IDIR) have developed technology to obtain grey-scale 2D images called ultrasonic b-scans in real-time during production in order to visualize the weld development with respect to time. They have demonstrated that by extracting and interpreting relevant patterns from these b-scans, weld quality can be determined accurately. However, current works combining conventional image and signal processing techniques are unable to extract those patterns from a wide variety of weld shapes with production-level satisfaction. Therefore, in this thesis, we propose to apply SSD, a single-shot multi-box detection based deep convolutional neural network framework for real-time embedded detection of components of cross-sectional weld shape from ultrasonic b-scans and interpret them to numeric parameters which are used as features to classify welds as good, bad or acceptable in real-time. Our proposed model has showed significant improvement in deciding weld quality compared to existing methods when tested on real industry facility.

Share

COinS