365 0

Acoustic Source Localization via Predicting Propagated Waves on Complex Systems using Time Reversal and Deep Neural Networks

Acoustic Source Localization via Predicting Propagated Waves on Complex Systems using Time Reversal and Deep Neural Networks
Yunsang Kwak
Alternative Author(s)
Issue Date
A novel method for the localization to identify acoustic sources was proposed by utilizing time reversal for dispersive waves and deep neural networks based on the wave prediction. The structural vibrations in complex structures with multi-linked paths were predicted using the wave approach after considering dispersive propagation characteristics. In order to consider linked conditions between multiple paths, the vibrations in the complex structures were analyzed as waves coupled both in transverse and torsional directions. The coupling effects on dispersive vibrations was verified by simple beam models. The numerical procedure for time reversal was proposed to identify the impact location in the complex structures having arbitrary paths. The proposed method was applied for experiments in an actual vehicle structure. The locations of the rattle sources were identified from the focused point of flexural vibrations analyzed by the proposed numerical time reversal procedure. Given the coupled wave approach, the proposed method was applied for localization of dispersive signals in arbitrary complex structures. The deep convolutional neural networks (DCNNs) were proposed through the feature construction for the localization and the simulative learning technique. The features for detecting the location of sources were extracted by performing a cross-cepstral analysis and image-mapping process. Groups of complex cross-cepstrums were calculated by using waves measured by three closely spaced transducers. The proposed feature construction allows to classifying the source locations regardless of spectral density of the sources. Each group was transformed into red, green, and blue (RGB) channels by pixel mapping. The image patterns were influenced by the source location. A simulative learning technique was proposed in this study and presented to train the DCNN without repetitive experiments. In order to generate the learning data for the DCNN, the propagated waves were predicted for various source locations and conditions. The method was verified by performing experiments in an anechoic room. The mapped images of the measured acoustic waves were classified by using the DCNN to detect the location of the acoustic sources. The sources were accurately determined by using only small microphones irrespective of the type of acoustic source and with reduced effects from the background noise. The proposed DCNN for localization of the sources on the complex structures was presented. The DCNN for the complex structures was achieved through the coupled wave prediction and feature construction for the flexural waves. The features for the DCNN were comprised by utilizing measured vibrations. The modified cepstral analysis using envelops of the spectrums was presented in order to extract the features for the source locations. The data for training of the DCNN were generated by changing conditions of sources and structures in the coupled wave prediction. The four changed conditions were corresponded to location and spectral density of sources, thickness and width of structures. The networks were trained by the generated data, and the source locations were identified for 150 classes. The proposed networks were verified by the vibration measurements on the actual vehicle. The locations of sources in the complex structure were predicted through the networks with only simple computational costs. The deep neural network with the simulative learning technique and the feature constructions provides the optimal performance for localization of sources on the complex structures even under sparse data for training. Given the wave approach for the structures, the proposed method makes it possible to easily adopt various complex systems.
Appears in Collections:
Files in This Item:
There are no files associated with this item.
RIS (EndNote)
XLS (Excel)


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.