Automatic Detection and Classification of Natural Weld Defects Using Alternating Magneto-Optical Imaging and...

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Detecting natural defects in welded components is challenging.
• Magneto-optical (MO) imaging, leveraging the Faraday effect, offers a potential solution for nondestructive testing.
• Existing machine learning models show limitations in accurately classifying specific weld defects like cracks and gas pores.

Purpose of the Study:

• To develop and evaluate an MO imaging system for nondestructive detection of natural weld defects.
• To compare the performance of various image preprocessing and machine learning models for defect classification.
• To establish a highly accurate classification model for natural weld defects using MO images.

Main Methods:

• An MO imaging system excited by an alternating magnetic field was established.
• Image preprocessing techniques including Gaussian, bilateral, and median filtering were applied.
• Feature extraction using Principal Component Analysis (PCA) fed into Support Vector Machine (SVM) and Backpropagation (BP) neural network models.
• Convolutional Neural Network (CNN) and ResNet50 models were developed and optimized for MO image classification.

Main Results:

• The ResNet50 model achieved an overall classification accuracy of 99% for natural weld defects.
• ResNet50 demonstrated improved accuracy by 7.4% over PCA-SVM and 1.8% over CNN.
• Classification accuracy for gas pores specifically increased by 10% (vs. PCA-SVM) and 4% (vs. CNN).

Conclusions:

• The developed ResNet50 model effectively and accurately classifies natural weld defects from MO images.
• ResNet50 significantly enhances the detection and identification capabilities for challenging weld defects.
• This advanced MO imaging and AI approach holds promise for improved quality control in welded structures.