Neural Network Modeling of In-Vehicle Noises with Different Roadway Roughness

Noise may cause adverse effects on human health. However, there is less attention paid to the quantification of traffic on in-vehicle noise for drivers as well as for passengers. This paper identified a neural network model to characterize the in-vehicle sound level along roadway segments from pavement roughness and vehicle activity information. An on-road test was conducted along a roadway in El Paso, Texas, United States, resulting 22,267 data pairs. The identified neural network was with one hidden layer of eight neurons, which weights were calibrated from part of the data pairs, while the rests were for testing. The resulted root mean squared error was 1.58 dB, the predicted and measured in-vehicle sound levels were highly correlated (R = 0.95), and the autocorrelation of prediction errors was very close to 0.0. The estimation error in exposure rate for hazardous sound levels (greater than 85 dB) was only 3.33%.