Noise forecasting: A machine-learning and probabilistic approach

Abstract

Forecasting the transmission of transient noise is a challenge since several sources of uncertainty exist: source and receiver positions, meteorology, and boundary conditions. These sources of uncertainty are considered to be model parameters. Experimental observations of noise, such as peak sound pressure level, or C-weighted sound exposure level, are data parameters with their attendant sources of uncertainty. Forward models, relating model parameters to the data parameters, are also imprecise. We quantify all of these sources of uncertainty by a probabilistic approach. Probability density functions quantify a priori knowledge of model parameters, measurement errors, and forward model errors as states of information. A conjunction of these states of information is used to generate the joint probability distribution of model and data parameters. Given a forecast of model parameters, say, from a numerical weather prediction model, the joint probability distribution is marginalized in order to forecast the noise field. In this study, we examine the feasibility of this approach using, instead of numerical weather predictions, point measurements of meteorological observations and peak sound pressure level collected during a long-range sound propagation experiment. Furthermore, we examine different types of forward models based on machine learning.