Design and Development of Artificial Intelligence (Al)-Based Desicion Support System For Manufacturing Applications

• Main Author: Lim , Chee Peng
• Format: Monograph
• Published: Universiti Sains Malaysia 2016
• Subjects: TK1-9971 Electrical engineering. Electronics. Nuclear engineering
• Online Access: http://eprints.usm.my/37361/
• Institution: Universiti Sains Malaysia

In this report, the research on welding defect detection and classification using radiograph images is presented. The first part of the report describes work on collection of digital radiograph images while the second part covers work on image processing and analysis using the collected images. The radiograph images from the Fuji DynamIX DynaView Workstation are custom-exported with the help of the NDT specialist. The collection of interpreted images is diverted from radiograph images captured using the old X-ray tube {Tube A) to the new X-ray tube (Tube B). Tube B images are needed to evaluate the performance of the developed defect detection algorithm under different radiography conditions. However, the total number of requested images remains approximately the same so that no extra workload is imposed to the NDT specialist. In the image processing stage, a flaw map, as described in the previous report, is used. Six welding defect types, namely Porosity{POR), Drop Through{DT), and Lack of Fusion{LOF), Lack of Penetration{LOP), Linear Indication{LI) and Undercut{UC), have been investigated. DT is detected using the background subtraction technique along with some heuristic rules as described in the previous report. For other detects, a set of image features including shape and texture information is extracted to characterize the welding defect flaws at the regions of interest (ROl). For POR, a series of sub-regions are further segmented in order to better represent the characteristics of POR at different locations in the ROl. To perform classification of the welding defects, an artificial intelligence (AI) technique, i.e., the Fuzzy ARTMAP (FAM) neural network, is applied. A series of experiments has been conducted by using the sample images collected from Tubes A and B. The overall performance is around 73% for accuracy, sensitivity, and specificity for both CF6-80 Connector Weld and Cover Weld programs. The only exception is that the sensitivity rate of the Connector Weld program stands around 63%. Further work will focus on ascertaining the stability of the FAM network in defect classification, as well as on improving the overall performance of the defect detection algorithms developed in this project. Ill