Automatic Quantification of Subsurface Defects by Analyzing Laser Ultrasonic Signals Using Convolutional Neural...

Area of Science:

• Materials Science
• Non-Destructive Testing
• Machine Learning

Background:

• Conventional machine learning for ultrasonic signal analysis requires manual feature extraction, which is unreliable and ineffective for defect detection.
• Accurate quantification of subsurface defect width is crucial for structural integrity assessment.

Purpose of the Study:

• To develop a novel, automated approach for accurately detecting and quantifying the width of subsurface defects using laser-generated ultrasonic signals.
• To overcome the limitations of manual feature selection in traditional machine learning methods for defect analysis.

Main Methods:

• Laser-generated ultrasonic signals were converted into scalograms (images) using wavelet transform.
• A pre-trained convolutional neural network (CNN) was used to automatically extract features from these scalograms for defect width quantification.
• An experimentally validated numerical model generated signals for training and validating the CNN model.

Main Results:

• The proposed method achieved a prediction accuracy of 98.5% on the validation set.
• The algorithm demonstrated 100% prediction accuracy on four experimental data points.
• The approach successfully quantified subsurface defect widths, avoiding manual feature engineering.

Conclusions:

• The combined wavelet transform and CNN approach is a feasible and reliable method for quantifying subsurface defect widths.
• This technique offers a universal solution for various defect detection tasks, including location and shape analysis.
• The automated feature extraction significantly enhances the reliability and effectiveness of defect detection using ultrasonic signals.