This project will build fundamental tools to aid in the deployment and qualification of inspections involving ML in the processing chain through three complimentary but independent work packages. The first work package addresses the use of ML for the suppression of benign responses from component geometry in NDE measurements. The second work package concerns uncertainty in ML algorithms for NDE data. Given any input, an ML algorithm will produce an output, but the confidence and accuracy of that output are currently not well understood or quantified. We will develop techniques that can quantify uncertainty in NDE applications by giving an estimate of the error of output and flag when an algorithm is presented with data that is outside its defined domain of operation. The final work package will explore how ML algorithms for NDE can be made interpretable for end-users by producing a human-understandable description of how a result was reached. This is a key step in the path to qualification of ML-based approaches. The outcomes of this project will be fundamental developments in ML for NDE that will be generalisable to many techniques and will build a solid foundation on which future real world inspections may be developed.

Prof. Paul Wilcox, Ultrasonics and Non-Destructive Testing Group, University of Bristol; Prof. Anthony Croxford, Ultrasonics and Non-Destructive Testing Group, University of Bristol.

We will develop automated UT data analysis tools to improve the reliability of detection, sizing, and characterisation of defects which occur in high safety significance industrial plant. The project will develop capability for defects of critical industrial importance, targeting two species: Thermal Fatigue (a service-induced species) and Hydrogen cracking (a manufacturing/welding defect). Both species are known to occur and often materially impact plant availability/longevity. This project will also enable capability for other challenging defect species going forward, such as stress corrosion cracks.

Dr Stewart Haslinger, Department of Mathematical Sciences, University of Liverpool; Dr Daniel Colquitt, Applied Mathematics, University of Liverpool; Dr William Christian, Department of Mechanical, Materials and Aerospace Engineering, University of Liverpool; Professor Jason Ralph, Department of Electrical Engineering and Electronics, University of Liverpool; Mr. Michael Wright, Signal processing, University of Liverpool; Prof. Michael Lowe, Department of Mechanical Engineering, Imperial College London; Dr Peter Huthwaite, NDE Group, Imperial College London.

In this proposed research work the benefit of the inspection/monitoring mix for different scenarios of degradation and measurement capability will be explored. The amount by which the intervals can be stretched relies on the measurement errors in both the inspection and monitoring method and the spatial and temporal characteristics of the degradation. Therefore, the key challenge to address is to quantify the effect of measurement errors in the process so that probability of detecting and tracking damage can be quantified and spatial coverage and repeat scan intervals can be optimised. In order to do this we will first build a statistical model that simulates the spatio-temporal degradation, based on Kuniewski’s work [3]. We will use this model as ground truth and perturb it to simulate measurements containing errors that are expected in real measurements, e.g. spatial offsets and biases and random noise. Similarly, we will use estimates from experiments and previous work to impose trending errors on the data. Finally, we will use the simulated measurement results to analyse the performance of a pure inspection strategy, or a pure monitoring strategy and mixed strategies and inspection intervals that should be used to get the best results.

Dr. Frederic Cegla, NDE group, Mechanical Engineering Department, Imperial College London

Most engineering metals are microscopically polycrystalline, and their manufacturing processes result in the ever-present grain microstructures in the finished products. These include preferred crystallographic orientations (texture), and the so-called macrozones in titanium alloys. These microstructures have been found to be the culprit of many catastrophic failures (e.g. airplane motor explosions), but they are very difficult to detect during the manufacturing process or on the finished components. This project aims to develop an ultrasonic method for the characterisation of the microstructures. It is based on a recent advancement of an ultrasonic measurement of volumetric texture, and will focus on the following areas of work: 1) Automation of the immersion data acquisition equipment. This will expand the existing IP and significantly reduce the data acquisition time in the lab. 2) Physical understanding of shear wave attenuation, which can potentially enable the estimation of statistical geometries in 3D using the same immersion setup for texture measurement. 3) Interpretation of bulk attenuation and backscattering data to obtain microstructure information from normal-incident scans. 4) Exploration of the diffuse wave field methods to apply the characterisation capabilities on complex geometries.

Dr Bo Lan, Non-Destructive Evaluation Group, Department of Mechanical Engineering, Imperial College London

Producing realistic ultrasonic data is vital for a variety of NDE applications in industry: validating methods, testing trainee inspectors, evaluating new methods being developed and generating machine learning data. Numerical methods, such as Pogo (www.pogo.software), have shown strong promise for generating physically accurate data, but suffer from this data missing physical imperfections and noise, as well as often being computationally demanding to produce in the quantities needed, despite the ability to use high-speed hardware such as GPUs. This is a notable challenge for applications such as inspection qualification where many realisations are required across the parametric space of different defects, geometrical considerations etc. This project aims to develop techniques to enable ultrasonic NDE methods to produce necessary data quickly and reliably. The project has three objectives: 1) Develop methods to combine datasets together to generate accurate data at high speed, such as inserting a defect response into another dataset. 2) Develop interpolation techniques to enable finer sampling to be produced from coarse simulations. 3) Develop suitable sampling strategies for high-dimensional parametric space, by eliminating dependent parameters and enabling global performance to be assessed with few simulations.

Dr Peter Huthwaite, Department of Mechanical Engineering, Imperial College London