Mechanised Harvesting of Standing Trees Using Contact Ultrasonic Testing and Machine Learning

The problem of hole-defect detection in standing trees is solved. To this end, the contact–ultrasonic device (Pundit PL-200) was employed to collect ultrasonic signals from testing some wood specimens both in the lab and some sites in Australia. The collected ultrasonic signals were then processed through the Variational Mode Decomposition algorithm to derive some features. In order to solve the classification problem more efficiently, the obtained characteristics were then analyzed through PCA to determine the most compelling features. Several machine learning algorithms and a one-dimensional convolutional neural network (1D-CNN) were employed to solve a set of classification problems based on data collected from (1) specimens with artificial defects in the lab and (2) billets with natural defects selected from trees harvested in sites of two states of WA and NSW, Australia. The results demonstrate the effectiveness of the proposed method for classifying wood materials based on their health state, where the accuracy result of 100% in the lab and at least 92.4% in the fields were achieved. The Fine Gaussian SVM was found to perform best on data collected from specimens in the lab and fields. It was also shown that 1D-CNN results were more reliable for solving the classification problem of standing trees in the fields. Keywords: Contact-ultrasonic testing, Machine learning, One dimensional CNN, Feature engineering, Variational mode decomposition