CAAV Talk by Thore Hertrampf
Impact localization in timber through machine learning classification using recurrence rate spectrograms of noisy vibration signals.
Speaker: Thore Hertrampf, Master Student in Mechatronics, Dynamics Group, University of Technology Hamburg, Germany,
Hosted By: Associate Professor Sebastian Oberst
Title: Impact Localization in Timber through Machine Learning Classification using Recurrence Rate Spectrograms of noisy Vibration Signals
Abstract: In the face of climate change, wood as a construction material is gaining popularity as a renewable alternative to cement and steel. Sustainability is only achieved through long-term use by storing carbon, while forests regrow. Reduced longevity is therefore not only an economical, but also ecological problem. Common hazards causing wood degradation are fire; moisture; mechanical stress; and wood eating insects. The latter traditionally requires intensive manual inspection, and its control often involves application of chemical pesticides harming the animals and nature.
The development of a non-destructive, automatic, acoustic insect deterrent device requires detection and localization of their interaction with the wood. Characteristics like cellular grain structure, ageing, inhomogeneous density and moisture give wood highly anisotropic viscoelastic properties and non-linear vibrational wave propagation, limiting the feasibility of modal analysis and identification of transfer functions, which are often the foundation for localization in structures.
Preliminary experiments show that Recurrence Plot based spectrograms of six nonlinear signals with complex periodic or chaotic behaviour can be classified using the convolutional neural network (CNN) ResNet-50 with accuracy of 75% when proper embedding is applied, compared to below 17% for periodogram spectrograms, in a challenging signal to noise scenario of 0.1.
In a first case study, this technique is applied to classify the vibration signature in wood beams according to their relative distance to the impact excitation induced by a modal hammer. Various CNNs of different complexities are compared and their models are generalized by augmentation of training data, to achieve robust classification independent of measurement position and beam dimensions.
This thesis contributes towards automatic insect detection and vibration-based structural health monitoring in general.
