The Library
Deep learning for defect detection on sheet metal stamped parts
Tools
Singh, Aru Ranjan (2023) Deep learning for defect detection on sheet metal stamped parts. PhD thesis, University of Warwick.
PDF
WRAP_Theses_Singh_A_R_2023.pdf - Submitted Version Embargoed item. Restricted access to Repository staff only until 12 January 2026. Contact author directly, specifying your specific needs. - Requires a PDF viewer. Download (154Mb) |
Official URL: http://webcat.warwick.ac.uk/record=b3985990~S15
Abstract
Sheet metal stamping is widely used for high-volume production. However, it is susceptible to the occurrence of defects. These defects have far-reaching implications, such as compromised product quality, increased costs, and potential safety hazards. Therefore, developing an accurate automated defect detection method is vital. This thesis addresses the specific challenges observed during split defect detection in prototype stamping components, aiming to advance the deep learning (DL) based inspection in sheet metal stamping.
The study begins by evaluating the effectiveness of DL models for classifying split defects and suggests a route toward the creation of a reliable vision-based DL model for stamping defect inspection. The visibility of defects in captured images is crucial for successful vision-based defect detection. In real stamping environments, the shiny metallic surfaces interact with overhead electric lights, sunlight, and skylights, leading to unpredictable specular reflections. To address this challenge, the thesis introduces a method that combines DL models with high-dynamic-range (HDR) imaging. This approach outperforms traditional imaging-based models, achieving enhanced accuracy while reducing false-positive predictions.
The low failure rates of stamping parts result in limited data for training DL models. This thesis introduces a method for generating a synthetic dataset of sheet metal stamping split defects to train DL models. This approach combines physics-based simulation and real defect texturing. By leveraging finite element simulation, plausible split locations are determined using a forming limit curve, and defect features are generated by mapping fine details of real splits. Incorporating synthetic images improves the detection performance and achieves a similar accuracy with significantly fewer real samples.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | T Technology > TS Manufactures | ||||
Library of Congress Subject Headings (LCSH): | Sheet-metal, Sheet-metal work, Metal stamping, Deep learning (Machine learning) | ||||
Official Date: | August 2023 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Warwick Manufacturing Group | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Debattista, Kurt | ||||
Format of File: | |||||
Extent: | xv, 178 pages : illustrations | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |