Predicting sound propagation in the atmosphere using an artificial neural network

Abstract

The spectral content of acoustic signals is dramatically altered as sound propagates through the atmosphere due to refraction from vertical gradients in wind and temperature, ground interactions, and other effects. Parabolic equation (PE) techniques have been successfully used to calculate numerical solutions for many of these effects. The PE models generally produce accurate attenuation values, but execution time is excessive for many applications where near real-time results are required. To obtain sound level attenuation predictions more quickly, we developed an artificial neural network for a range of heights and source/receiver horizontal separations. The PE and boundary conditions were modified to obtain a nondimensional representation, which resulted in seven parameters required to specify all input combinations. This model version was then used to train the neural network, using a range of values for each parameter. The standard deviation of the errors (propagation model minus neural network simulations) was generally within 2 dB for the training data set containing approximately 15<th>000 cases. The smaller test data sets resulted in errors having standard deviations of approximately 1 dB. The neural network was a good predictor of the sound propagation model results except for small values of the nondimensional ground impedance parameter.