Uncertainty Quantification for Deep Learning in Ultrasonic Crack Characterization.

Area of Science:

• Nondestructive Evaluation (NDE)
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) shows promise for human-level data analysis in NDE.
• Quantifying prediction uncertainty in DL for NDE is underdeveloped but essential for inspection qualification and trust.
• This study focuses on uncertainty quantification (UQ) for deep learning models in crack sizing for inline pipe inspection.

Purpose of the Study:

• To demonstrate optimal uncertainty quantification (UQ) methods for deep learning in crack sizing for inline pipe inspection.
• To evaluate the performance of deep ensembles and Monte Carlo dropout for UQ in this context.

Main Methods:

• A convolutional neural network (CNN) was employed to size surface-breaking defects using plane wave imaging (PWI) data.
• Two UQ methods, deep ensembles and Monte Carlo dropout, were implemented and compared.
• The CNN was trained on PWI images of defects simulated using a hybrid finite element/ray-based model.

Main Results:

• Monte Carlo dropout exhibited poor UQ, with limited separation between in-distribution and out-of-distribution data and a weak correlation (R=0.84) between experimental RMSE and uncertainty.
• Deep ensembles demonstrated improved calibration (R=0.95) and anomaly detection compared to Monte Carlo dropout.
• Enhancements like spectral normalization and residual connections further improved deep ensemble calibration (R=0.98) and out-of-distribution detection.

Conclusions:

• Deep ensembles offer superior uncertainty quantification for deep learning-based crack sizing in NDE compared to Monte Carlo dropout.
• Optimized deep ensemble architectures enhance the reliability of NDE inspections by providing trustworthy uncertainty estimates.
• This research advances the application of trustworthy AI in critical infrastructure monitoring.