Cassels, Bryan (2018) Weld Defect Detection using Ultrasonic Phased Arrays. Doctoral thesis, University of Central Lancashire.
Preview |
PDF (Thesis document)
- Submitted Version
Available under License Creative Commons Attribution Non-commercial Share Alike. 9MB |
Abstract
Traditional ultrasonic test methods, based on the manipulation of an ultrasonic probe by an experienced inspector are beginning to be replaced by Automated Ultrasonic Testing (AUT). Although largely limited to regular structures the integration of phased arrays and computer controlled mechanical manipulators allows AUT to provide fast regular and repeatable data acquisitions for off-line inspection and future auditing. Although this automation provides comprehensive data its inspection remains subject to time consuming manual interpretation. The objective of this thesis is to investigate a method of automating the inspection process. To this end the emphasis is on detecting regions of a weld that do not comply to the normal, anomaly free, structure. These regions being highlighted for manual sentencing.
Experimental data is available as sets of sequential images. These are as a set of sector scans from a contact phased array probe and secondly as a set of images created by a Total Focusing Method (TFM) algorithm using immersion probes. In both cases the simplest approach to segmentation is to directly threshold images. This is found not to produce reliable discrimination and is particularly ineffective in the presence of a continuous major feature such as a front wall. An alternative multivariate technique is Principal Component Analysis (PCA). This was found to have far greater discrimination.
A primary limitation of PCA is its susceptibility to outliers. Although an outlier indicates an anomaly their presence also prevents an accurate estimation of the anomaly free background. This may, in turn, lead to the masking of smaller anomalies preventing their detection. Two approaches at obtaining a robust estimate of an anomaly free representation of background statistics are investigated. The first trims the data set by determining the Mahalanobis squared distance of each observation to the original data set's centroid. Assuming a normal distribution then the Mahalanobis distance behaves as a chi-squared distribution. In this work a proposed cut-off of 0.05 is used throughout to identify and remove outliers.
The second approach is to obtain a background representation by Principal Component Pursuit (PCP). Here the background is modelled by a low-rank matrix. A drawback of this approach is the computational time required. However, as with trimming, once an accurate background representation is established it may be used any number of times on data acquired using the same set up.
Repository Staff Only: item control page