Laser ultrasonics and machine learning for automatic defect detection in metallic components

This paper develops an automatic and reliable nondestructive evaluation (NDE) technique that enables quantification of the width and depth of subsurface defects of metallic components simultaneously by using non-contact laser ultrasonic technique and identified machine learning (ML) algorithm. Twenty-two specimens with various subsurface defect dimensions are designed and fabricated for laser ultrasonic experiments, and a total of 220 labeled laser ultrasonic signals are obtained for training and verifying ML models. Twelve features, including four time-domain features (maximum, minimum, peak-to-peak, and |Neg|/Pos value of the laser generated Rayleigh ultrasonic waves) and eight wavelet energy features, are identified and extracted as sensitive feature vectors for establishing the dataset. The principal component analysis (PCA) is implemented as dimensionality reduction method of feature vectors to optimize the recognition algorithm and improve the detection accuracy. Three widely used ML models in NDE, adaptive boosting (Adaboost), extreme gradient boosting (XGBboost), and support vector machine (SVM), combined with the PCA are proposed and compared for detecting both the width and depth of subsurface defects. The PCA-XGBoost achieves the highest recognition rate of 98.48%, and is therefore identified as the most effective approach for analyzing laser-ultrasonic signals. Unlike published reports, the proposed model is trained and evaluated with experimental data covered various classification labels, which is more adaptive and reliable in practical application than the models established using simulated data or limited experimental data. In other applications, as long as sufficient laser ultrasonic data with regards to various defect properties (dimensions, orientations, locations, shapes, etc.) can be acquired, the developed approach can realize accurate detection of corresponding defects.