Groundbreaking Nondestructive Evaluation Technique Developed for Metallic Components
A pioneering study has unveiled a cutting-edge automatic and dependable nondestructive evaluation (NDE) technique, revolutionizing the quantification of subsurface defects in metallic components. Through the innovative integration of a non-contact laser ultrasonic technique with a sophisticated machine learning (ML) algorithm, researchers have achieved remarkable strides in simultaneously measuring the width and depth of subsurface defects.
The research, led by an expert team, involved twenty-two specimen designs and fabrications with diverse subsurface defect dimensions. Researchers established a comprehensive essential dataset for developing the technique by leveraging twelve distinctive features, including time-domain parameters and wavelet energy features extracted from the laser-generated Rayleigh ultrasonic waves.
Key to the success of this groundbreaking approach was implementing principal component analysis (PCA) as a dimensionality reduction method, optimizing the recognition algorithm, and significantly enhancing detection accuracy. Researchers explored three prominent ML models in NDE—adaptive boosting (AdaBoost), extreme gradient boosting (XGBboost), and support vector machine (SVM)—combined with PCA to detect both the width and depth of subsurface defects.
Among the models tested, the PCA-XGBoost model emerged as the standout performer, achieving an exceptional recognition rate of 98.48%. This remarkable accuracy underscores the efficacy of the approach in analyzing laser-ultrasonic signals. Crucially, unlike previous studies that relied on simulated data or limited experimental datasets, the proposed model was trained and evaluated using real-world experimental data covering diverse classification labels, rendering it more adaptable and reliable for practical applications.
This groundbreaking research represents a significant leap forward in the field of nondestructive evaluation. You can have access to the full article through the link below: