Using Neural Networks to Identify Squeaks and Rattles in Vehicles

Previous works have developed a methodology to characterize in a unique form annoying noises (squeaks and rattles) that appear inside of automotive vehicles. Thus, for each characterized noise, its origin and alternatives of eliminating it, or to make it more tolerable, are known. Using this methodology, data bases with the squeaks and rattles that typically appear in vehicles have been created.

This paper describes the work done towards the development of a tool, based on neural networks, that determines if an squeak or rattle present in a vehicle corresponds to any of the noises already characterized and stored in the database. Using Matlab and experimental results, different types and structures of neural networks have been evaluated in a systematic form. Preliminarily, it was found that for this specific application the best topology is a net with 100 inputs, one hidden layer of 50 neurons with sigmoid active functions and four outputs neurons with linear active functions. Additionally, it was found that the best training method is the gradient descent back-propagation method with a learning rate of 0.05.