Propeller noise localization and background noise separation using isogeometric analysis and neural networks

This study is based on experimental data of the E1619 propeller obtained in the large cavitation tunnel at National Taiwan Ocean University. A sound source identification method that combines frequency–domain analysis with database-based inverse estimation is proposed to effectively separate propeller noise from background noise signals. By analyzing the spectral distributions at different rotational speeds, background noise components that remain unchanged with speed variations are identified and removed. Furthermore, by comparing the acoustic signals before and after propeller installation, the dominant sound signals emitted by the propeller within specific frequency bands are extracted for source localization analysis. The principle of vector superposition is then employed to further isolate propeller and other mechanical noise from the total signal. In terms of deep neural network and database construction, the study adopts isogeometric analysis (IGA) to compute the frequency–domain sound field of the cavitation tunnel as training data. Multi-frequency responses are used to enhance the generalizability of the deep neural network model. The sound source locations are uniformly distributed throughout the test section of the tunnel, and a deep neural network architecture is used for inverse localization, significantly improving the accuracy and flexibility of the sound source identification.