Deep Learning for Ultrasonic Crack Characterization in NDE.

Area of Science:

• Non-destructive evaluation
• Machine learning applications
• Computational modeling

Background:

• Machine learning (ML) offers advanced pattern recognition for improved non-destructive evaluation (NDE).
• A key challenge in applying ML to NDE is the limited availability of real-world defect data for training.
• Traditional methods like the 6-dB drop method have limitations in defect characterization accuracy.

Purpose of the Study:

• To demonstrate a hybrid simulation approach for training ML models for NDE.
• To develop and evaluate a convolutional neural network (CNN) for characterizing real defects in pipes.
• To compare the accuracy of the CNN approach against a standard image-based sizing technique.

Main Methods:

• Utilized an efficient, hybrid finite element (FE) and ray-based simulation to generate training data.
• Trained a convolutional neural network (CNN) for defect characterization using simulated inline pipe inspection data.
• Applied the 6-dB drop method as a benchmark for comparison.

Main Results:

• The CNN achieved significantly higher accuracy in crack characterization (length and angle) compared to the 6-dB drop method.
• CNN accuracy improved crack length prediction by nearly fourfold (±0.29 mm vs. ±1.1 mm).
• The CNN demonstrated robustness to variations in sound speed estimation, maintaining high accuracy.

Conclusions:

• A hybrid simulation-based ML approach can effectively train CNNs for accurate NDE defect characterization.
• Deep learning methods, particularly CNNs trained on simulated data, offer superior accuracy over traditional techniques.
• This methodology addresses data scarcity and enhances the reliability of defect analysis in industrial applications.