| الوصف: |
This research has been motivated by the fact that present road traffic noise prediction models have not improved significantly since their development in the 1970s and 1980s, although road traffic noise nuisance is a significant and growing issue in Australia and elsewhere. This thesis reviews the nature of road traffic noise, its measurement, and interpretation of noise levels in terms of noise nuisance. It then examines the principal noise propagation influences that are described by road traffic noise prediction models such as ST AM SON and TNOISE, and outlines how these quasi-empirical models produce noise level predictions. Present road traffic noise prediction models are essentially pattern recognition tools, but while they perform satisfactorily for very simple situations, accurate noise prediction in more complex situations is beyond their ability. However, artificial intelligence pattern recognition tools have proven their power and usefulness in a variety of applications in recent years, and this thesis examines the hypothesis that a neural network approach to predicting road traffic noise offers a way to move forward in noise impact assessment. A simple two-layer feed-forward neural network architecture is found to be able to easily mimic present road traffic noise prediction models, with tangent-sigmoidal transfer functions specified for the input layer of 20-30 neurons, and a linear transfer function specified for the single output neuron. A priori rescaling of input values to roughly match the requirements of the transfer function facilitates the neural network training using a backpropagation algorithm with momentum and adaptive learning. Ways of avoiding the problem of overfitting are discussed. A case study based on a 1993 noise impact assessment project is presented that demonstrates that a neural network can easily be trained from fairly limited field data to satisfactorily predict road traffic noise in site-specific situations, and the case study was one in which a model such as STAMSON or TNOISE is not able to perform well. The effort and expertise needed for this exercise is comparable to an air emission dispersion modelling exercise, a conclusion that should prove of great interest to road and environment authorities. The thesis then proposes a strategy whereby grid-based neural networks can be developed to enable road traffic noise prediction in complex situations. The methodology is explained with the aid of a barrier adjustment calculation. The development of such a model for a sitespecific situation is quite straightforward, but there is also clear potential to develop a generic 2-dimension modelling capability. The basic approach to this parallels the modelling strategy of present noise prediction models, but with reference sound levels and adjustments referred to a grid, and determined using neural networks. |
