Interpretable and Explainable Machine Learning for Ultrasonic Defect Sizing.

Area of Science:

• Materials Science
• Mechanical Engineering
• Computer Science

Background:

• Machine learning (ML) adoption in industrial nondestructive evaluation (NDE) is limited by the 'black box' nature of many algorithms.
• Interpretability and explainability are crucial for deploying ML in critical NDE applications.
• Ultrasonic testing is a key NDE technique, but extracting meaningful features for ML can be challenging.

Purpose of the Study:

• To introduce Gaussian Feature Approximation (GFA), a novel dimensionality reduction technique for ultrasonic NDE.
• To improve the interpretability and explainability of ML models used in ultrasonic defect sizing.
• To evaluate GFA's performance against other dimensionality reduction methods and raw data approaches.

Main Methods:

• Developed GFA by fitting a 2-D elliptical Gaussian function to ultrasonic images, extracting seven descriptive parameters.
• Utilized these GFA parameters as input features for a defect-sizing neural network.
• Compared GFA with 6 dB drop boxes, Principal Component Analysis (PCA), and Convolutional Neural Networks (CNNs) on raw images.

Main Results:

• GFA achieved defect-sizing accuracy close to raw image analysis, with only a 23% increase in root mean square error (RMSE).
• GFA reduced input data dimensionality by 96.5%, significantly outperforming PCA and 6 dB drop boxes in accuracy.
• Shapley Additive Explanations (SHAPs) confirmed that the GFA-based model aligns with traditional NDE sizing principles.

Conclusions:

• GFA offers a more interpretable and explainable alternative for ML in ultrasonic NDE compared to PCA or raw image inputs.
• The GFA method provides a robust feature extraction technique that balances dimensionality reduction with high sizing accuracy.
• This approach facilitates the practical implementation of ML for industrial applications like inline pipe inspection.