Our system currently accounts for between 20 and 25 variables. Whilst this is quantity can be handled by an average user, it would be interesting to try to reduce this number. Such a reduction would provide information about the variables responsible for the most noise, and would make our system easier to use.
Natalia Genaro García
Scientists from the University of Granada have developed software capable of accurately forecasting future noise levels within urban environments. The system, designed by researchers from Granada’s Departments of Civil Engineering, Applied Physics and Computer Sciences and Artificial Intelligence, is able to predict both the frequency and type of noise that a neighbourhood is likely to generate in the future.
The software, which uses neural networks to assess data such as street types, road conditions, average vehicle speeds and the presence of road works, has been found to be 95 per cent reliable. After analysing noise data collected from Granada in 2007, the researchers plan to validate their software by conducting tests in new cities.
Natalia Genaro García, one of the study’s authors, told ScienceOmega.com
more about the advantages of this new, noise-predicting system…
What were the challenges of applying soft computing methods to urban noise assessment?
Basically, the challenge was to improve upon existing methods that use mathematical models to predict noise. These methods were not obtaining accurate results and there were practically no contributions relating to soft computing and urban noise. Granada is a very noisy city, so in collaboration with other departments, our research group thought that it would be interesting to apply soft computing to this issue.
Why will it be advantageous to reduce the number of variables required to produce accurate forecasts?
Our system currently accounts for between 20 and 25 variables. Whilst this is quantity can be handled by an average user, it would be interesting to try to reduce this number. Such a reduction would provide information about the variables responsible for the most noise, and would make our system easier to use. Subsequent research suggests that 10 to 15 variables will be sufficient to accurately predict noise levels.
What are the major differences between this type of system and the traditional mathematical models used to predict noise levels?
Conventional mathematical models don’t tend to use more than five variables, so they are not as powerful as our system. What’s more, as most of them are based on data obtained from particular cities, the models tend to fit these specific localities. When you come to test such models in new cities, the results are not very good.
Traditional models are mostly based on traffic speed and the number of passing vehicles. When tested in streets with little traffic, the predictions that they make are completely incorrect. As our system is not over-reliant on these variables, it doesn’t make the same mistakes.
Our system is more accurate because it uses a model based on artificial neural networks (ANNs). ANNs try to emulate the ways in which the human brain works. Our software is therefore an intelligent system that can learn from examples in a similar manner to humans. Once the system has learned, it can solve new instances of the problem.
What hardware is required to operate this software?
Just a normal computer.
Is specialist knowledge required to operate this system?
Not at all. The user only needs to know the values of the different variables.