Reduced order model analysis to identify possible aerodynamic noise sources of small axial fan: POD and CNN



This paper reports computational analysis of location and strength of sound source of the noise generated by a small axial fan widely used as an air-cooling system. High-fidelity Navier-Stokes simulations with high-resolution compact scheme are conducted with an implicit Large Eddy Simulation (LES) method on a HPC system and the resultant large-scale data confirms existence of unsteady vortex structures and their interactions around the impellers, boss and casing of the fan. To identify location and strength of the sound sources, reduced order model analysis is conducted for the distribution of pressure fluctuations in space and time. Snapshot POD (Proper Orthogonal Decomposition) analysis both in time and in circumferential direction, together with conventional FFT analysis, identifies location and strength of the sound sources. In addition, Convolutional Neural Network (CNN) is attempted, which shows more physical mode decomposition and separates some of the important features shown in the snapshot POD analysis. The study shows that the two data-mining techniques considered here identify possible aerodynamic noise sources of the axial fan clearly in comparison to those in the previous studies.