Classifier fusion for vibrational NDT of complex metallic turbine blades

Parts with geometrical complexity bring significant challenges in nondestructive testing (NDT). The Process Compensated Resonance Testing (PCRT) method has recently shown promising results for inspecting complex-shaped metallic parts. PCRT is a broadband vibrational testing procedure that relies on the extraction of resonant frequencies coupled to advanced learning methods. Once a suitable classifier is determined, it is then applied to unknown test samples in order to classify them as healthy/defected. The target of this work is to upgrade the PCRT with an advanced classifier to increase the classification performance. For this purpose, first the inclusion of the Q-factors to the available PCRT feature set i.e., only frequencies, is investigated and then, a novel classifier fusion based on Dempster-Shafer theory of evidence (DST) has been proposed to combine several constituent models. The constituent models are selected to be adaptively boosted NNs (ABNNs) trained by using different numbers of features. The proposed algorithm (ABNN + DST) is applied to polycrystalline Nickel alloy first-stage turbine blades with complex geometry. The results indicate that the proposed DST-based fusion algorithm increase the classification accuracy from 93.5% to 96.5%.