Repository logo
 

Machine learning approach to thermite weld defects detection and classification.

Loading...
Thumbnail Image

Date

2021

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

The defects formed during the thermite welding process between two sections of rails require the welded joints to be inspected for quality purpose. The commonly used non-destructive method for inspection is Radiography testing. However, the detection and classification of various defects from the generated radiography imagesremains a costly, lengthy and subjective process as it is purely conducted manually by trained experts. It has been shown that most rail breaks occur due to a crack that initiated from the weld joint defect that was not detected. To meet the requirements of the modern technologies, the development of an automated detection and classification model is significantly demanded by the railway industry. This work presents a method based on image processing and machine learning techniques to automatically detect and classify welding defects. Radiography images are first enhanced using the Contrast Limited Adaptive Histogram Equalisation method; thereafter, the Chan-Vese Active Contour Model is applied to the enhanced images to segment and extract the weld joint as the Region of Interest from the image background. A comparative investigation between the Local Binary Patterns descriptor and the Bag of Visual Words approach with Speeded Up Robust Features descriptor was carried out for extracting features in the weld joint images. The effectiveness of the aforementioned feature extractors was evaluated using the Support Vector Machines, K-Nearest Neighbours and Naive Bayes classifiers. This study’s experimental results showed that the Bag of Visual Words approach when used with the Support Vector Machines classifier, achieves the best overall classification accuracy of 94.66%. The proposed method can be expanded in other industries where Radiography testing is used as the inspection tool.

Description

Masters Degree. University of KwaZulu- Natal, Durban.

Keywords

Citation

DOI