Classification study for using acoustic-ultrasonics to detect internal decay in glulam beams

Bayes, k-nearest neighbor (KNN), and neural network classifiers were used to study the decay detection efficiency of acousto-ultrasonics (AU). Brown-rotted Douglas-fir glulam beams removed from service were measured by using through-transmission AU. Single and multiple sets of AU signal features included velocity, attenuation, shape, and frequency content. Although all of the AU signal features were sensitive to decay, they were also affected by natural characteristics of wood. However, it was possible to improve the detection efficiency by using multiple signal feature sets in classification analysis. A 79% efficiency was achieved with the neural network classifier for detecting small levels of decay (10% of the cross section) and a 68% overall correct classification for different degrees of decay when using three or four signal features as inputs. The results of the Bayes and KNN classifiers were quite similar, with 79% KNN and 75% Bayes detection efficiency for small levels of decay, and 67% overall. Received 22 January 1999