Semi-supervised Bayesian classification of materials with impact-echo signals.

Area of Science:

• Materials Science and Engineering
• Non-Destructive Testing
• Signal Processing

Background:

• Internal material defects pose significant challenges in industrial applications.
• Accurate detection and classification of defects are crucial for quality control and safety.
• Existing methods may require extensive training data or lack flexibility.

Purpose of the Study:

• To develop and validate an impact-echo technique for detecting and identifying internal material defects.
• To classify materials based on defect status (homogeneous, single, multiple) and type (hole, crack).
• To evaluate the performance of a Bayesian classifier with Gaussian mixture models under limited supervision.

Main Methods:

• Impact-echo testing to generate wave propagation spectra from impacted specimens.
• Bayesian classification utilizing Gaussian mixture models for probability estimation.
• An extended expectation-maximization algorithm for parameter and class probability estimation.
• Testing with real aluminum alloy specimens.

Main Results:

• The proposed method effectively detects and classifies internal material defects.
• A harmonic average precision and recall of 92.38% was achieved with only 10% supervision.
• The algorithm demonstrated robust performance on real aluminum alloy samples.

Conclusions:

• The impact-echo technique combined with a Bayesian classifier offers a flexible and accurate approach for internal defect detection.
• The method is effective even with limited training data, reducing supervision requirements.
• Potential applications include industrial quality control and process optimization, such as in marble cutting.