Recurrence rate spectrograms for impact localization in wood

Characteristics like cellular grain structure, inhomogeneous density, aging, and altering environmental conditions (moisture, temperature) give wood highly anisotropic viscoelastic properties and non-linear vibrational wave propagation properties. Nonlinearity limits the use of linear methods, such as modal analysis and parameter identification via transfer functions. Acoustic localization of natural damage to wood, like crack growth, is of general interest in structural health monitoring of timber structures. Time-difference of arrival or energy attenuation is commonly used for localization, which are prone to boundary reflections or require the frequency response function. Recent advancements in machine learning-based classification of non-linear signals can achieve a much higher accuracy when recurrence rate-based spectrograms are used compared relative to conventional short-time Fourier transforms, especially in the presence of noise. Hence, in this work, multi-sensor measurements of impulse induced vibration in wood beams are classified by their distance to the excitation, based on their time series, avoiding a priori knowledge of a transfer function for the localization. The machine learning model is trained across various widths and thicknesses of samples, giving a localization estimate independent of beam dimensions. This research will contribute to early detection of damage in the field of vibration-based structural health monitoring of wood.