Photoacoustic signal to image based convolutional neural network for defect detection

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Material defect detection is crucial for ensuring structural integrity and product quality.
• Traditional non-destructive testing methods can be limited in sensitivity and scope.
• Photoacoustic (PA) imaging offers a promising technique for subsurface defect visualization.

Purpose of the Study:

• To develop a novel photoacoustic (PA) signal to image conversion based convolutional neural network (CNN) model for automated material defect detection.
• To create a low-cost computer-aided PA triggering and acquisition device for data collection.
• To evaluate the model's performance across different materials and evaluation strategies.

Main Methods:

• A low-cost computer-aided PA triggering and acquisition device was developed.
• PA signals from defected and intact materials (aluminum, iron, wood, plastic) were collected.
• Variational mode decomposition was used for PA signal to image conversion and feature extraction.
• A lightweight CNN was trained and tested on the converted grayscale PA images for defect classification.

Main Results:

• The proposed CNN model achieved high accuracy in defect detection.
• Within-class (material-dependent) evaluation yielded a mean accuracy of 0.977 (up to 1.0).
• All-class (material-independent) evaluation resulted in a mean accuracy of 0.942 (up to 0.955).

Conclusions:

• The developed PA signal to image conversion CNN model is effective for material defect detection.
• The model demonstrates robust performance across diverse materials, both independently and collectively.
• This approach offers a cost-effective and accurate solution for non-destructive material inspection.