Machine learning based approach for automatic defect detection and classification in adhesive joints

This study presents an automated technique combining ultrasonic pulse echo method with machine learning algorithms to detect and classify the depth of interface defects in adhesively bonded joints. After data preprocessing for machine learning and extracting 32 ultrasonic features, the binary and ternary datasets were established for “defect”-“no defect” and its depth classifications. The importance and classification accuracy of various feature subsets—initial, single interface, minimised, tree-based, recursive, sequential, and LDA—were explored. A support vector machine (SVM) model was trained on these datasets. For “defect” vs. “no defect” classification, the initial feature subset achieved over 90 % accuracy on train/test data and 83 % on unseen data. For the ternary dataset, depth classification accuracy on unseen data in recursive feature subset was 97 % for “depth 1,” 62 % for “depth 2,” and 91 % for “depth 3.” The obtained results demonstrate prediction accuracy and suitability of ML models for classifying defects and predicting their depths in adhesive bonds.