In the field of nondestructive evaluation, accurate characterization of defects is required for the assessment of defect severity. Defect characterization is studied in this paper through the use of the ultrasonic scattering matrix, which can be extracted from the array measurements. Defects that have different shapes are classified into different defect classes, and this essentially allows us to distinguish between crack-like defects and volumetric voids. Principal component analysis (PCA) is used for feature extraction, and several representational principal component subsets are found through exhaustive searching in which quadratic discriminant analysis (QDA) and support vector machine (SVM) are used as the pattern classifiers. Instead of choosing a single optimal classifier, the best classifier is dynamically selected for different measurements by estimating the local classifier accuracy. The proposed approach is validated in simulation and experiments. In simulation, the depths (lengths of the minor axes) of 4441 out of 4636 test samples are measured accurately, and the measurement errors (with respect to the defect size) are below 10%. Arbitrarily shaped rough volumetric defects are identified as ellipses, which are reasonably good matches in shape to the original defects. Experimentally, six subwavelength scatterers are characterized and sized to within 0.14λ.
|Number of pages||15|
|Journal||IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control|
|Publication status||Published - 1 Dec 2015|
- Feature extraction
- Principal component analysis
- Support vector machines