Improving Ultrasonic Imaging Using Machine Learning

Research Problem

“Ultrasonic testing is a crucial non-destructive testing & evaluation technique employed around the world. Ultrasonic testing can involve the testing of highly important, safety critical components, therefore, highlighting the need for the highest quality of ultrasonic testing techniques. Ultrasonic imaging plays an important part in this, imaging algorithms designed to image within large varieties materials & components are typically high quality (e.g. TFM), however, some historically difficult to image materials such as welds (which have a degree of anisotropy) can prove more difficult to image accurately. Machine learning can offer an avenue to improve current imaging algorithms via improved accuracy, defect detection/characterisation & automation [1][2][4], offer avenues for usage within additive manufacturing environment & for in-process inspection through real-time analysis of ultrasonic NDE data and offer a level of explainability to AI/ML within industry [5] among other benefits [3].”

Aims & Objectives

Ultrasonic inspection is a critical process in manufacturing. However, dealing with challenging materials, such as anisotropic materials, can prove to be difficult. My project looks to explore using machine learning methods as a potential solution to problems related to this by improving the imaging of historically “difficult” materials such as welds, using deep neural networks to characterize and predict material properties, improving imaging for materials with complex geometry, improving explainable AI/ML within industry & working on improving model training times through various techniques.

Additionally, my project also focuses around trying to improve the machine learning process itself – coming up with solutions that could aid in the timeliness of models & addressing issues related to data unavailability, providing more efficient applications of machine learning within NDT.

Current Progress

Currently I am working on improving weld imaging by adapting usable experimental data for usage with simulated data. This data will be trained on a neural network. The aim is to more accurately predict microstructures within the weld material, leading to higher quality images within a weld. 

I am also working on improving model training times through a variety of methods, including parallelization. As well as expanding the above current research into more complex weld geometries, with the aim being to improve imaging of welds with more complex geometry. This work assumes in-process inspection.

 

Bibliography:

[1] Singh, J., Tant, K., Curtis, A. et al. Real-time super-resolution mapping of locally anisotropic grain orientations for ultrasonic non-destructive evaluation of crystalline material. Neural Comput & Applic 34, 4993–5010 (2022). https://doi.org/10.1007/s00521-021-06670-8

[2] Singh, J.; Tant, K.; Mulholland, A.; MacLeod, C. Deep Learning Based Inversion of Locally Anisotropic Weld Properties from Ultrasonic Array Data. Appl. Sci. 2022, 12, 532. https://doi.org/10.3390/app12020532 

[3] Pyle, R. (2023). Application of machine learning to ultrasonic nondestructive evaluation. [PDF]. https://pure.strath.ac.uk/ws/portalfiles/portal/157277681/Pyle_Bristol_2023_Application_of_Machine

[4] R. J. Pyle, R. L. T. Bevan, R. R. Hughes, R. K. Rachev, A. A. S. Ali and P. D. Wilcox, “Deep Learning for Ultrasonic Crack Characterization in NDE,” in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 68, no. 5, pp. 1854-1865, May 2021, doi: 10.1109/TUFFC.2020.3045847.

[5] R. J. Pyle, R. R. Hughes, A. A. S. Ali and P. D. Wilcox, “Uncertainty Quantification for Deep Learning in Ultrasonic Crack Characterization,” in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 69, no. 7, pp. 2339-2351, July 2022, doi: 10.1109/TUFFC.2022.3176926.