Sound classification and power consumption to sound intensity relation as a tool for wood machining monitoring

Abstract

Non-contact process monitoring could be a powerful tool to prevent tool misuse, detect wood species, detect tool dullness and reduce electrical energy consumption—all of which could reduce production costs. The aim of this study is to identify recognizable patterns in the sound signals produced during the circular sawing of two different wood species—beech (Fagus moesiaca) and fir (Abies alba)—and to classify them in order to obtain an intelligent machining process capable of recognizing the wood species being machined. These two wood species were selected for this study due to their morphological, physical and mechanical differences. The cutting power was also recorded during the process and measured indirectly via the motor power used. A sound signal can easily be converted into an image (spectrogram), which is suitable as a data basis for the deep learning process. Several neural networks were used to classify the sounds. In order to prepare the raw audio signal for machine learning using image recognition, it was processed in several steps. The relationship between the audio and the recorded cutting power was also investigated and found to be strongly correlated, but only for audio frequencies up to 4500 Hz. Based on the results and further analysis, the classification accuracy for wood species identification varied between 98% for MobileNetV2 and 94% for the InceptionV3 deep learning network.