Using multi-layer perceptrons to predict vehicle pass-by noise

All new vehicle designs have to pass a legislative, noise emission test - the 'pass-by noise' test. In the highly competitive automotive industry, it is important to predict the test result early in the design process, rather than waiting until a prototype is built. Engineers can 'guess' test results about as well as the best, although inadequate, analytical models. They achieve this by using experience and their knowledge of acoustics and of the vehicle's design. Neural networks should also be capable of pass-by noise prediction, learning from the results of previous tests. This paper describes a neural network approach to the problem. First, expert knowledge is used to select vehicle design and test parameters to present as inputs to a multi-layer perceptron. Since data is scarce, the problem is broken down into two stages, vehicle performance and pass-by noise. The two trained networks are evaluated and their performance discussed