Gear Noise Diagnosis System for Automobile Transmission Using Artificial Intelligence (Effect of Number of Intermediate Layers on Judgement Performance)

The present paper describes a digitizing method for the measured gear noise and a construction of a neural network system for gear noise diagnosis.Gear noise emitted from automobile transmissions should be evaluated by gear noise experts. Although quietness performance estimates from measured noise levels of the transmissions on some production lines, the estimation must be very difficult. There is not a certain relationship between the measured noise levels and the evaluations by the gear noise experts. Therefore, the estimation should be severe. As a result, such an automatic gear noise diagnosis system must yield transmissions with over-quality.The present study deals with a new gear noise diagnosis system to which an artificial intelligence, that is, a neural network system is applied. The previous evaluations by the new gear noise diagnosis system were good when the statistical property of the teacher signals from which the neural network system learned was similar to that of population. This fact means that many teacher signals are necessary on the practical use.Proposed digitizing method of gear noise levels provided good evaluations of neural network system even when the statistical properties of the teacher signals were not similar to that of the population. In addition, a new method, “Moment method” for determining the construction of the neural network system was introduced instead of “Back Propagation Method”. The Moment Method contributed to the improvement of the system judgments.The neural network system constructed using the Moment Method brought good performance. And the number of intermediate layers in the neural network system could be small enough to obtain good performance. It was found that the Moment method provided good learning because of connecting weights update function. When the Moment method was used for determining the connection weights between neurons in the neural network system, the developed gear noise diagnosis system achieved high and stable correct answer ratio. And the number of intermediate layers in the neural network system was only one enough for obtaining good performance of the system. Four intermediate layers, which was the maximum in this paper, did not provide much good performance.Copyright © 2013 by ASME