Characterizing residual stress on pipework weld repairs using machine learning

Nuclear power plant systems are constructed by joining pressure vessels and piping
components using modern welding processes. Welded structures may also be subjected to weld repair either
during fabrication to mitigate manufacturing defects, or during service to maintain the original design life, or
to provide life extension. The PhD project will review the latest developments in data mining and machine learning
methods, identify baseline parameters controlling residual stresses at weld repairs, collect high quality
training data using data mining methods, develop/optimise a suitable machine learning tool, train the tool
using the mined data, validate the outputs against independent measurements, and create residual stress
characterisation tool with a user-friendly interface for engineers in industry.
A major limitation in using ANN for prediction of residual stress in weldments is the shortage of high quality
measurement data for training. The innovation and challenge of the proposed PhD project will be to exploit
the opportunity of combining measurement results and synthetic data from validated simulations in order
to predict/estimate residual stress profiles in families of repair weldments (that lie within the training data
parameter envelope).