A single weld defect in a safety-critical engineering structure has the potential to incur high monetary costs,damage to the environment or loss of human life. This makes comprehensive non-destructive internal and externalinspection of these welds essential. For non-destructive internal inspection the ultrasonic phased array supportsa number of methods for producing a cross sectional image at a fixed location. Full coverage of the weld requiresa sequence of images to be taken along the full length, each image at a unique incremental step. If the weldhas a geometrically regular structure, such as that corresponding to a long linear section or the circumferenceof a pipe, automation becomes possible and data is now provided for post processing and auditing. Particularlyin a production process this may provide many thousands of images a day, all of which must be manuallyexamined by a qualified inspector. Presented in this paper is an approach for rapid identification of anomaliesin sequences of ultrasonic sector images taken at equally spaced index points. The proposed method is basedon robust principal component analysis (PCA). An assumption is that most sectors are anomaly free and havea statistically similar geometrical structure. Unsupervised multivariate statistical analysis is now performed toyield an initial low dimensional principal subspace representing the variation of the common weld background.Using the Mahalanobis distance outliers, observations with extreme variations and likely to correspond to sectorscans containing anomalies, are removed from the reference set. This ensures a robust PCA-based reference modelfor weld background, against which a sectorial scan is identified as defect free or not. Using a comprehensiveset of sector scan data acquired from test blocks, containing different types and sizes of weld defects at differentlocations and orientations, the paper concludes that PCA has potential for anomaly detection in this context.Although trimming improves the accuracy of the system eigenvectors, it is shown that greater accuracy of thelow rank subspace is possible through principal component pursuit (PCP). This is evident by an almost 100%anomaly detection rate with a false alarm rate of well below 10%.
展开▼