Résumé
In the last decade, with the increasing requirement for the quality of equipment, several effortshave been devoted to the development of new real-time detection technics for weld jointsquality in many industrial fields such as the nuclear, chemical and aeronautical. Welding is acomplicated process, which is often affected by the welding process and environmentaluncertainties, and it is easy to produce welding defects such as overlap, pore, spatter, andincomplete fusion. In fact, the welding quality control is carried out by a traditional destructiveand non-destructive methods such as macrographs, ultrasonic testing and x-ray detecting.However, these methods have some limitations: ray detection can easily result in side-effectson the human body and ultrasonic testing is susceptible to the location, orientation and shapeof the defect. In addition, the in-situ observation of the weld pool dynamics using non-intrusiveinstruments will give information about the stability and consequently evaluate the real-timewelding quality. In this study, the objective is to analyze by using several sensors (arc voltage,current and high-speed cameras), the influence of two welding positions (horizontal, flat) onthe weld pool behavior for the GTAW process. Therefore, the setup configuration according tothe full penetration of weld joint, two cameras are fixed on the front and back side of the workpiece. In the other hand, an algorithm based on computer vision and image processing isdeveloped including four steps: image acquisition, image preprocessing, weld pool featuresextraction and classification. The weld joints were classified into four classes, flat andhorizontal position with two welding speeds. in the second part of the algorithm, differentmachine learning methods (KNN, Random Forest…) are tested for detecting and classifyingweld joints on one hand, and on the other hand they are compared by analyzing performancescores (accuracy, time of calculations…). Actually, image processing and machine learningcombination is trained with extracted features to predict the classification results with a high-level perspective to realize the real-time intelligent identification of weld surface defects.