University of Bristol

Supervisors: Dr Jie Zhang, Professor Tony Mulholland

Background

There has been recent work using machine learning and S-matrix methods to characterise a defect that resides in the bulk of a component. The challenge is to separate the scattering signals from the defect and those from the impedance mismatches at the grain interfaces in the heterogeneous host material. When a defect resides on a corroded and uneven surface then the situation becomes quite different and more difficult.  This is because of extra challenges from reduced inspection angular coverage to avoid the specular reflections from the surface, possible wave paths through anisotropic weld region, the scattering from the uneven surface of unknown topography, and possible corrosion zone near the surface which has some possible porosity and therefore could be shown as some highly reflective point scatterers, the ultrasound scattering behavior of which is like that of the crack tip.

The aim of the project is to identify ultrasound signatures of a surface breaking defect from its surrounding scattering signals from the corroded and uneven surface that the defect resides on. This will be achieved by the developed artificial intelligence (AI) based ultrasonic array systems to analyse array data. The project will involve the use of experimental and physical-modelling-based datasets to train machine learning algorithms such as convolutional neural networks, generative adversarial networks, and neural differential to create a real-time NDT technology.  The physical-modelling-based datasets will be built using the developed simulation tools, which includes an uneven surface topography model, a corrosion model and a finite element wave scattering model. One target application is inspection of welds, and these materials are heterogeneous, locally anisotropic, and highly scattering.  An attenuation model will therefore be required for the physical model and the project will develop one based on stochastic differential equations and a diffusion approximation theorem. The use of AI systems to interrogate large NDT datasets is a rapidly increasing research area and this project will provide the student with an excellent foundation for a future career in AI assisted NDT inspection.

This project will provide more accurate information about surface breaking defects for structural integrity assessment and hence enable more accurate assessment of the health of structures and predict their remaining life. This project will also provide a further step towards the industrial uptake of the emerging technique for surface breaking defect characterization, particularly in ultrasonic testing applications in vessels and pipes used in nuclear power plants.

Objectives

  • Develop simulation tools to predict the scattering from surface breaking defects residing on an uneven and corroded surface.
  • Build up the array data base of S-matrices of surface breaking defects residing on an uneven and corroded surface.
  • Explore and develop a range of ultrasonic array imaging methods that can be used for separating image features from surface breaking defects and corroded and uneven surfaces.
  • Develop S-matrix method that can be used for separating ultrasound signal features from surface breaking defects and corroded and uneven surfaces.
  • Develop machine learning techniques to identify surface breaking defects developed on corroded and uneven surfaces using their features in their ultrasound images and S-matrices.
  • Produce an experimental procedure of inspection of surface breaking defects on corroded and uneven surfaces using ultrasonic arrays in real industrial components.

References

Singh, J., Tant, K.M.M., Mulholland, A.J. and MacLeod “Deep learning based inversion of locally anisotropic weld properties from ultrasonic array data”, Appl. Sci., 12(2), 532-552, (2022).