Machine Learning Based Classification of Welded Components

Welding is one of the most common approaches employed for fusing metals. However, welding errors are usually encountered during the process due to external factors. The detection and classification of these welding faults are of great importance for the reliability of the weld and the materials which were welded. Traditionally, welding error detection is performed through visual inspection carried out by inspectors or quality control personnel, which is an error-prone and slow process. In this paper, an image processing and machine learning based algorithm is proposed to automatically detect welding defects. The detection of component irregularities and classification is essential in quality control in manufacturing processes. Machine learning algorithms are widely employed for various applications as it reduces the precious human visualization time to classify the welded components manually and errors due this, which cannot be completely eliminated. The proposed algorithm also allows acceptance sampling or 100
^%
inspection of the components as the speed of the classification process is within a few milliseconds with higher accuracy. The presented components are classified into Grades - A, B, C, and D with images obtained in visible spectrum. Experimental results prove that random forest classifier provided an accuracy of 82.8% compared to 77.6
^%
of decision trees. Therefore, the welding evaluation process is made effective through machine learning based algorithms. The advantage is that automating the evaluation process makes it quicker and provides an unbiased grading.