Comparative study on the suitability of feature extraction techniques for tungsten inclusion and hotspot detection from weld thermographs for on-line weld monitoring

Abstract

Welding is the most commonly used technique for joining metals in industries. In spite of various technological advances defects do occur in welds. Post non-destructive testing (NDT) techniques assess the quality of weld after completion of welding process. Monitoring and controlling weld parameters during welding can avoid the defect or if the defect is already intolerable the welding process can be stopped there to save time and money. It is thus necessary to develop an automated on-line welding system to make the correct decision. Weld thermographs are acquired on-line with IR camera. Effective feature extraction algorithms are to be developed to isolate and quantify the defect features from thermographs. This paper compares the effectiveness and suitability of three different feature extraction algorithms namely discontinuity based detection (conventional), region growing and Euclidean distance based color image segmentation developed for on-line monitoring and control. Tungsten inclusion and different depths of penetration thermographs are the database considered for defect feature extraction. Online weld monitoring necessitates a standardized feature extraction technique that works well irrespective of the size and shape of defect. Hence, comparison is based on the accuracy of the results, parameter independency and image independency. It is found that feature extraction by Euclidean distance based segmentation is best suited for on-line weld monitoring as it is parameter independent and can be standardized for a defect.