Transfer learning in guided wave testing of pipes

Guided wave testing (GWT) is a non-destructive testing (NDT) technique extensively used for in-service testing of pipes that allows the inspection of tens of metres of pipe in either direction from a single sensor position. The aims are to identify and locate all physical features found along the pipe in the axial direction, and in particular the presence of defects, such as cracks or corrosion patches. However, the signals output by GWT of pipes are complex to interpret, making the quality of inspection highly dependent on the operator skills. Due to such signal complexities, at present there is a lack of automated procedures that can help operators in this task. Some of the recently developed machine learning (ML) algorithms are expected to possess the modelling capabilities required to address such a classification task, though they would typically need hundreds if not thousands of labelled input data for their training. This amount of experimental data is seldom available in the NDT field, particularly with regards to the damage cases. The main purpose of this article is to investigate whether, and how, it is possible to augment an available set of labelled experimental data with a synthetic dataset having characteristics that are similar but still distinct from the real ones. This is studied by training three different ML models with various combinations of actual and simulated data pertaining to GWT of pipes, the goal being the automated detection of reflections from pipe features within the inspection traces. The results demonstrate that when there is scarce availability of experimental data, substantial detection improvements can be achieved by pre-training the chosen ML model with synthetic data, before fine-tuning it on actual inspection data. In particular, the ML algorithm that is found to perform best for this task is a VGG-Net model, which is shown to yield false positive rates in the order of ∼1.5 to 4 % at the fixed true positive rate of 99.7 %.