[4A5] Numerical simulations for crack analysis using machine learning

J Gaffney
University of Liverpool, UK

Many industries, including power generation, require non-invasive methods to ensure the structural integrity of key components. Ultrasonic testing is commonly utilised to identify defects and measure defect characteristics because this method is quick, generally reliable and non-invasive. Aspects of application of the method can be challenging in some circumstances, potentially reducing reliability.

Therefore, a tool to help practitioners quickly and accurately detect, size and characterise defects would be beneficial. Machine learning (ML) is an ideal candidate to help practitioners because it is well suited to interpreting large datasets (such as the output from time histories from a single crystal transducer). However, for an ML algorithm to give reliable predictions it must have accurate data to learn from. Because real defects in plant are not common, there are small numbers of training datasets from real defects. Artificial defects exist but these may not mirror real defects well under all circumstances. Additionally, the task of collecting training datasets for all probe and scanning combinations could be formidable.

In the project discussed here, finite element modelling (FEM) in two-dimensional space, informed by real defect data (obtained from samples provided by industrial partners and including geometric and roughness parameters) and realistic transducer characteristics (input signals, beamwidths, etc) are used to generate a synthetic dataset suitable for application to ML in the field.

[4A5] Numerical simulations for crack analysis using machine learning

J Gaffney University of Liverpool, UK

J Gaffney
University of Liverpool, UK