CLASSIFICATION AND PREDICTION ANALYSIS FOR WELD BEAD SURFACE QUALITY USING FEATURE EXTRACTION AND MAHALANOBIS-TAGUCHI SYSTEM

This research develops an innovative algorithmic scoring system designed for the accurate assessment of welding bead samples in UNITEN’s welding laboratory. Leveraging image feature extraction in combination with the Mahalanobis-Taguchi System (MTS) implemented in MATLAB software, the study constructs a tailored camera jig to ensure consistent image data capture. A comprehensive threshold analysis improves classification accuracy, minimizing the risk of misclassification. The results reveal that while the K-means clustering method outperforms the Variable Bin Width method across several performance metrics, including an accuracy of 86.36% and a high specificity of 94.5%, the method’s recall rate of 50.49% indicates room for improvement in identifying true positives. The balanced F1 score of 51.07% highlights the trade-off between precision and recall. This research developed a strong basis in educational evaluation methodologies by introducing a framework for simplifying the way in assessing student’s welding sample, ultimately aiming to enhance the instructor’s focus in providing training for UNITEN students, ensuring they build strong foundational welding skills.