Objective There are a large number of ancient buildings with wooden structures in China. How to conveniently test and evaluate the physical and mechanical properties of ancient building normal timber is a rigid requirement for the daily protection, repair and safety assessment of ancient wood building structures. In this study, the machine learning algorithm was introduced to the knocking sound signal, and try to apply the convenient knocking method to the nondestructive testing of the physical and mechanical properties of ancient building wood.
Method The four-section ancient larch (Larix gmelinii) timber members dismantled from an ancient royal building in Beijing were used as raw materials to process clear specimens. Firstly, the influence of size and density of the wood specimen on the knocking sound signal was investigated. And the physical and mechanical property parameters such as density, modulus of rupture, modulus of elasticity and compressive strength parallel to grain of wood specimens were measured experimentally. Then, the Mel frequency spectral coefficient (MFSC) feature extraction was carried out on the knocking sound signal collected in the experiment. Taking the knocking sound MFSC feature as the input, and the physical and mechanical properties of the specimen as the output, a convolutional neural network (CNN) evaluating model for the physical and mechanical properties of ancient building timber member was constructed.
Result The size of the specimen had no effect on the knocking signal, and the dominant peak frequency of the knocking signal of the specimen with higher density was higher. The dropout layer had a significant impact on the performance of the model, and the fitting effect was the best when the dropout rate was 0.2. The established model had a good effect on the evaluation of physical and mechanical properties of ancient building wood, and the coefficient of determination between the evaluation value of density, modulus of rupture, modulus of elasticity and compressive strength parallel to grain and the real value reached 0.873, 0.819, 0.746 and 0.860, respectively.
Conclusion The CNN model constructed in this study based on the MFSC feature of knocking sound is feasible to detect and evaluate the physical and mechanical properties of timber in ancient buildings.