Improvement of Generalization Performance for Timber Health Monitoring using Machine Learning

In studying damage detection in timber using the Timber Health Monitoring system, we have succeeded in classifying the positions of the weight of the timber by using vibration waveforms with machine learning. In this study, we investigated the generalization performance of the system, which is indispensable for practical applications. Previous studies have yet to confirm this type of performance. We prepared 90 timber pieces as we expected that the system's performance would be improved if more timbers were learned. We divided the pieces into nine classes, representing no damage and damage to eight different positions, respectively. A piezoelectric sensor was attached to the pieces to acquire their vibration waveforms. The waveforms were divided into training and evaluation data, and a neural network (NN) was used to learn the training data and classify the evaluation data. As a result, we found that the NN was able to classify the positions of the damage or no damage with up to 83.8% accuracy, even for unlearned timber pieces. This demonstrated good generalization performance in the Timber Health Monitoring system.