Wood hole-damage detection and classification via contact ultrasonic testing

Damage detection in wood materials has numerous applications in different industries, such as construction and forestry. Wood is generally a complex medium due to its orthotropic and random properties, which increases the difficulty of non-destructive damage testing. However, machine learning algorithms can be employed to overcome this problem. In this paper, hole-defect classification problems of two common types of wood materials, namely hard (marbau) and soft (pine) wood, are studied using a naive Bayes classification technique. To this end, the results of contact ultrasonic tests conducted on these types of woods in different directions, i.e. tangential and radial to the growth rings of wood, were investigated. The various states of the intact, small defect, and large defect of each type of wood were considered in the testing regime. It is known that contact ultrasonic tests are highly sensitive to different aspects of the test, such as the amount of couplant gel applied to surfaces, the amount of pressure applied to the transducer and receiver, and misalignment of the transducer and receiver. Therefore, 50 replicates of each test were implemented. First, an advanced signal decomposition algorithm termed Variational Mode Decomposition (VMD) was exploited to derive some features from the recorded ultrasonic signals. Then, the derived features were used in a set of classification problems using a naive Bayes classifier to classify the damage state of the specimens. Different types of naive Bayes classifiers, namely Gaussian and kernel, along with combinations of different types of features were employed to improve the results, ultimately achieving nearly 100% 10-fold cross-validation accuracy in all cases individually. However, when cases from different types of wood and direction of the tests were mixed, 93.6% 10-fold cross-validation accuracy was achieved for the classification problem based on the health state of the cases, using kernel naive Bayes classifier and a mixture of two types of features.