1. Application of machine learning to ultrasonic nondestructive evaluation
- Author
-
Pyle, Richard, Wilcox, Paul, and Hughes, Robert
- Subjects
Ultrasonics ,NDT ,Machine Learning - Abstract
Machine learning (ML) techniques have the potential to provide automated data analysis for nondestructive evaluation (NDE) applications with human-level accuracy. This is of great value as the data gathered by NDE inspection is, increasingly, large and complex, making manual data analysis expensive, slow and sensitive to operator variability. However, there are three major barriers to the application of ML models to industrial NDE: sourcing useful training data, choosing informative features, and building trust in the model's predictions. This thesis investigates how these barriers can be overcome by deep learning with simulated training sets, domain adaptation, uncertainty quantification, and improved interpretability. An example NDE use case is considered: defect sizing for ultrasonic inline pipe inspection. An inspection configuration is devised to closely match the conditions found in inline inspection of oil pipelines, resulting in ultrasonic plane wave images of surface breaking defects. These ultrasonic images are used as input to ML models to predict the size of the defects. A convolutional neural network (CNN) is trained to size defects, using a simulated data set, and applied to previously unseen experimental data. As the CNN takes ultrasonic images as input there is no need to manually select informative features. The CNN is compared to a traditional NDE sizing method, 6 dB drop, and demonstrates significantly better sizing accuracy. Further sizing accuracy improvements are achieved through the inclusion of a small amount of experimental data in the training procedure. This additional training data is included with the aim of reducing the effect that differences in simulated and experimental data have on sizing performance. An adversarial-based domain adaptation technique is found to be the optimal way to leverage small amounts of experimental training data. Building trust in the prediction of ML models is essential for qualifying them for use in NDE industry. Uncertainty quantification (UQ) is a significant part of this, as it is essential to the decision making for any automated data analysis. This thesis investigates two modern UQ techniques, finding deep ensembles to be an effective way to quantify the uncertainty of sizing predictions. Further trust is built by improving the interpretability and explainability of ML for NDE. This is achieved with a novel dimensionality reduction method: Gaussian feature approximation (GFA). GFA involves fitting a 2D gaussian to an ultrasonic image and storing the resulting seven parameters that describe it. These parameters can be used as input features for a ML model. As individual GFA features are meaningful to a human (unlike pixel intensities) the resulting model is implicitly more interpretable than one trained on raw images. Shapley additive explanations are used to indicate how each feature contributes to a crack size prediction. The results presented in this thesis indicate that it is possible to use ML to achieve automated data analysis for real-world industrial NDE applications.
- Published
- 2023