Ultrasonic rough defect characterisation using deep learning method

Abstract

Rough crack sizing is an important topic in the non-destructive evaluation (NDE) region. Accurate characteristics of defects are important for determining the structural integrity of safety-critical engineering components used in power plants and jet engines. Most existing ultrasonic non-destructive evaluation (NDE) methods are developed for smooth defect detection and characterisation. However, the irregular surface of the defect usually introduces the diffuse wave field into the ultrasonic measurements, causing an increase in the uncertainties of the traditional methods in defect severity assessment. Hence, one of the main motivations of this thesis is to study the effects of surface roughness on the accuracy of traditional ultrasonic defect characterisation methods. In addition, more efficient and accurate rough defect and surface characterisation methods are developed to enhance the reliability of ultrasonic testing in rough crack characterisation. The main content of this thesis can be divided into three contributions. The first contribution is implementing the traditional Time-of-Flight Diffraction (ToFD) method for sizing rough defects. Previous studies on the ToFD technique have primarily concentrated on theoretical and experimental investigations of smooth defects. However, rough defects produce more complicated tip-diffracted patterns than smooth defects, mainly due to additional scattered wave fields from mode conversion and propagation of surface waves along the rough surface. High-fidelity finite element modelling and stochastic Monte Carlo methods are employed to gain physical and statistical insights into how both the incident beam angle and the degree of roughness affect the planar defects scattering fields. The results demonstrate that roughness may cause larger diffraction amplitude values at different angles, which leads to increased uncertainties when sizing. The ToFD sizing methods, including envelope peak detection and autocorrelation approaches, are implemented to estimate the size of rough cracks. The second contribution is developing a self-attention deep learning method to interpret the ToFD A-scan signals for sizing rough defects. A high-fidelity finite-element (FE) simulation software, Pogo is used to generate the synthetic datasets for training and testing the deep learning model. Besides, the transfer learning (TL) method is used to fine-tune the deep learning model trained by the Gaussian rough defects to boost the performance of characterising realistic thermal fatigue rough defects. To demonstrate the accuracy of the proposed method, the crack characterisation results are compared with those obtained using the conventional Hilbert peak-to-peak sizing method. The results show that the deep learning approach significantly reduces uncertainty and error in characterising rough defects compared to traditional ToFD measurement techniques. The third contribution introduces a methodology to recover the morphology of a complex rough surface from ultrasonic pulse-echo measurements with an array of equidistant sensors using the one-dimensional convolution neural network (1DCNN). The neural network is trained on datasets simulated from high-fidelity finite element simulations for surfaces with a range of roughness parameters and is tested on both numerical and real experimental data. To evaluate the effectiveness of our method, we compared the reconstruction results from the deep learning approach with those obtained from conventional ultrasonic array imaging techniques. Unlike traditional methods that require numerous sensors (e.g., 128, 64, or 32), our deep learning-based method uses fewer sensors while maintaining accuracy.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Fan SHI (Supervisor)