Acoustic recognition of wood based on computer deep learning

Hejin Liu, Yongke Sun, Qiying Li, Kangkang Zhang, Ning Li, Changzhao Li, Yuxi Huang, Hao Lu, Yushan Yang, Jian Qiu,
Acoustic recognition of wood based on computer deep learning,
Industrial Crops and Products,
Volume 231,
2025,
121175,
ISSN 0926-6690,
https://doi.org/10.1016/j.indcrop.2025.121175.
(https://www.sciencedirect.com/science/article/pii/S0926669025007216)
Abstract: With the cross-disciplinary integration of wood science and computer technology, deep computer learning has been introduced into the field of wood recognition, which plays an important role in wood micro-recognition, DNA profile recognition, and so on. As an anisotropic material, wood has large differences in acoustic properties and structure in different directions, and the stress waveforms are complex and difficult to be analysed by human beings. Therefore, to address the difficulty of stress wave analysis, this study combines computer deep learning techniques to classify the acoustic properties of six types of wood. The sound of an iron pendulum continuously striking wood in the transverse, radial and chordal directions was recorded using a smartphone; the sound was segmented and then extracted by Fast Fourier Transform (FFT); Mel Frequency Cepstrum Coefficients (MFCC) to extract the sound features, thus visualising the original audio file as a Mel Spectrogram. Features with discriminative properties were extracted from the Mel spectrogram, and computer deep learning classification models: Residual Networks (ResNet), Convolutional neural Networks (ConvNeXt) were applied to achieve the classification of acoustic features of wood, with a classification accuracy of up to 99 %. It was found that the response frequency and response time of Mel spectra were regular, and the response frequency of Mel spectra of different tree species was positively correlated with the density, with high response frequency of the denser tree species and low response frequency of the less dense tree species. The high response time of chordal Mel spectra was lower than that of transverse and radial Mel spectra. The acoustic characteristics of wood of the same species are affected by moisture content and dimensions. Changing the moisture content and resonant dimensions of the wood will result in significant changes in the response frequency and response time of the wood's Mel spectrogram.
Keywords: Wood; Acoustic characterisation; Stress wave detection; Deep learning; Classification model