Automated detection of structural resonances using neural networks (vor Ort)
* Presenting author
Abstract:
Considering the high number of sensors used in acoustic investigations of the vehicle interior or on component test benches (e.g. 20 microphones in ISO3745), the development of methods that minimize the analysis effort is of great interest. The current state of the art for resonance detection in rotordynamic systems focuses on peak detection in peak-hold or averaged spectra based on a Campbell diagram. In this paper, two new concepts for resonance detection are presented. The first concept follows the conventional approach. First, image processing methods are applied to the Campbell diagram, followed by a newly developed dimensional reduction method. The resonances are then detected very accurately using peak detection, achieving a true positive rate (tpr) of more than 96 %. The second concept is based on a convolutional neural network (CNN). The CNN is trained using over 85.000 resonances as training data. The trained neural network shows a tpr of almost 90 %. The extent to which tonally excited structural resonances can be distinguished from broadband noise excitation is also investigated. It is shown that a CNN can be trained to provide not only detection but also classification of resonances with a detection rate of 85.7 % for unseen data.
