Identification of defects from ECT signals using linear discriminant function

Statistical pattern recognition method is applied to eddy current testing (ECT) based defect identification in this work. Spatially distributed ECT signals are converted into a multi-dimensional space vector using data embedding method and the Fisher's discriminant is then applied to find the linear projection of the multi-dimensional data that best distinguish defect and non-defect signals. A set of potential defects are clearly distinguished from welding noise by utilizing the discriminant function derived from supervised learning. Furthermore, the correlation between the probability of detection and the dimension of constructed vector space is investigated. Both the false and miss detection probability decrease with the increase of vector dimension. The reliability of a discriminant method is significantly enhanced by increasing the dimension of vector space.