[2A1] NDT machine learning for real-time inversion of locally anisotropic weld properties using in-process ultrasonic array measurements
R Pyle, K Tant, N Sweeney, E Nicholson, J Ludlam and C MacLeod
University of Strathclyde, UK
Many welds exhibit significant local anisotropy due to their elongated grain structure. Ultrasonic inspection of these anisotropic materials is challenging, as the variations in sound speed alter the direction of acoustic propagation within the component. This is detrimental to the quality of ultrasonic array imaging, as without accurate knowledge of the local anisotropic properties precise focusing cannot be achieved. A common approach to remedy this issue is to use lower frequencies (<3 MHz) and/or techniques that do not require focusing in reception, such as sectorial scanning and time-of-flight diffraction. However, to leverage the high spatial resolution ultrasonic array imaging can provide, knowledge of the local anisotropy of the weld is required.
This paper aims to improve ultrasonic array-based inspection of anisotropic welds by mapping local variations in anisotropy. This has been achieved previously with iterative solvers such as Markov chain Monte Carlo methods and genetic algorithms, but these involve repeated running of forward simulations, making them computationally intensive, precluding their use for in-process inspection. Real-time inversion is achieved in this paper by training a neural network to invert for the grain orientations in the weld using time-of-flight (TOF) measurements. The 64 × 64 TOF matrices are measured after each layer of weld deposition using two arrays, in tandem, positioned either side of the weld. The neural network is trained on data simulated using the anisotropic locally interpolated fast marching method (ALI-FMM). Once trained, the neural network is tested with both finite element and experimental data from a 316L stainless steel weld. Performance is measured in two ways. Firstly, by comparing the predicted and true grain orientations where ground truth is available, and secondly, by using the predicted orientations to calculate anisotropic travel time maps and observing the improvement in image quality compared to imaging with an assumption of isotropy.
This paper aims to improve ultrasonic array-based inspection of anisotropic welds by mapping local variations in anisotropy. This has been achieved previously with iterative solvers such as Markov chain Monte Carlo methods and genetic algorithms, but these involve repeated running of forward simulations, making them computationally intensive, precluding their use for in-process inspection. Real-time inversion is achieved in this paper by training a neural network to invert for the grain orientations in the weld using time-of-flight (TOF) measurements. The 64 × 64 TOF matrices are measured after each layer of weld deposition using two arrays, in tandem, positioned either side of the weld. The neural network is trained on data simulated using the anisotropic locally interpolated fast marching method (ALI-FMM). Once trained, the neural network is tested with both finite element and experimental data from a 316L stainless steel weld. Performance is measured in two ways. Firstly, by comparing the predicted and true grain orientations where ground truth is available, and secondly, by using the predicted orientations to calculate anisotropic travel time maps and observing the improvement in image quality compared to imaging with an assumption of isotropy.
