Please use this identifier to cite or link to this item:
Title: Compressed Sensing for Open-ended Waveguide Non-Destructive Testing and Evaluation
Authors: Tang, Chaoqing
Issue Date: 2019
Publisher: Newcastle University
Abstract: Non-destructive testing and evaluation (NDT&E) systems using open-ended waveguide (OEW) suffer from critical challenges. In the sensing stage, data acquisition is time-consuming by raster scan, which is difficult for on-line detection. Sensing stage also disregards demand for the latter feature extraction process, leading to an excessive amount of data and processing overhead for feature extraction. In the feature extraction stage, efficient and robust defect region segmentation in the obtained image is challenging for a complex image background. Compressed sensing (CS) demonstrates impressive data compression ability in various applications using sparse models. How to develop CS models in OEW NDT&E that jointly consider sensing & processing for fast data acquisition, data compression, efficient and robust feature extraction is remaining challenges. This thesis develops integrated sensing-processing CS models to address the drawbacks in OEW NDT systems and carries out their case studies in low-energy impact damage detection for carbon fibre reinforced plastics (CFRP) materials. The major contributions are: (1) For the challenge of fast data acquisition, an online CS model is developed to offer faster data acquisition and reduce data amount without any hardware modification. The images obtained with OEW are usually smooth which can be sparsely represented with discrete cosine transform (DCT) basis. Based on this information, a customised 0/1 Bernoulli matrix for CS measurement is designed for downsampling. The full data is reconstructed with orthogonal matching pursuit algorithm using the downsampling data, DCT basis, and the customised 0/1 Bernoulli matrix. It is hard to determine the sampling pixel numbers for sparse reconstruction when lacking training data, to address this issue, an accumulated sampling and recovery process is developed in this CS model. The defect region can be extracted with the proposed histogram threshold edge detection (HTED) algorithm after each recovery, which forms an online process. A case study in impact damage detection on CFRP materials is carried out for validation. The results show that the data acquisition time is reduced by one order of magnitude while maintaining equivalent image quality and defect region as raster scan. (2) For the challenge of efficient data compression that considers the later feature extraction, a feature-supervised CS data acquisition method is proposed and evaluated. It reserves interested features while reducing the data amount. The frequencies which reveal the feature only occupy a small part of the frequency band, this method finds these sparse frequency range firstly to supervise the later sampling process. Subsequently, based on joint sparsity of neighbour frame and the extracted frequency band, an aligned spatial-spectrum sampling scheme is proposed. The scheme only samples interested frequency range for required features by using a customised 0/1 Bernoulli measurement matrix. The interested spectral-spatial data are reconstructed jointly, which has much faster speed than frame-by-frame methods. The proposed feature-supervised CS data acquisition is implemented and compared with raster scan and the traditional CS reconstruction in impact damage detection on CFRP materials. The results show that the data amount is reduced greatly without compromising feature quality, and the gain in reconstruction speed is improved linearly with the number of measurements. (3) Based on the above CS-based data acquisition methods, CS models are developed to directly detect defect from CS data rather than using the reconstructed full spatial data. This method is robust to texture background and more time-efficient that HTED algorithm. Firstly, based on the histogram is invariant to down-sampling using the customised 0/1 Bernoulli measurement matrix, a qualitative method which only gives binary judgement of defect is developed. High probability of detection and accuracy is achieved compared to other methods. Secondly, a new greedy algorithm of sparse orthogonal matching pursuit (spOMP)-based defect region segmentation method is developed to quantitatively extract the defect region, because the conventional sparse reconstruction algorithms cannot properly use the sparse character of correlation between the measurement matrix and CS data. The proposed algorithms are faster and more robust to interference than other algorithms.
Description: Ph. D. Thesis
Appears in Collections:School of Engineering

Files in This Item:
File Description SizeFormat 
Tang C 2019.pdfThesis8.2 MBAdobe PDFView/Open
dspacelicence.pdfLicence43.82 kBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.