Timber moisture detection using wavelet packet decomposition and convolutional neural network

Abstract

As timber structures are vulnerable to degradation due to the tendency to trap moisture, the present study proposed a new percussion-based method to replace the existing constant contact between structures and sensors. A total of two approaches have been proposed to automated detect the moisture content (MC) of timber: (a) the random forest classifier (machine learning-based) was employed to classify the wavelet packet decomposition (WPD) features extracted from excitation-induced sound signals (WPD + RF); and (b) the 2D-CNN framework (deep learning-based) was employed to classify the Mel frequency cepstral coefficient (MFCC) features extracted from excitation-induced sound signals (MFCC + 2DCNN). The proposed automatic detection methods are covered from 1D time-domain signal classification to 2D image classification. To verify the effectiveness of both two approaches, an experimental study was conducted. The MC of two types of timber specimens (i.e. softwood and hardwood) was gradually increased from 0% to 60% with 10% increments. The change of MC of timber material caused different material properties, resulting in a measurable differential in forced vibration among the various specimens used. The results demonstrated that MFCC + 2DCC outperformed the RF + WPD in MC classification of timber material. Overall, the percussion-based method proposed in this study can provide an outstanding classification performance.