| Summary: | 碩士 === 國立高雄第一科技大學 === 機械與自動化工程所 === 94 === Abstract
In this thesis, we adopt the experimental way to collect the revolution datas
of the base idle speed for the gasoline engine. In our experiments, we use the
intake valve clearance, the exhaust valve clearance and the ignition timing as the
input variables and the engine revolution as the output variable. Through the
training and examining of adaptive network-based fuzzy inference system
(ANFIS), we can carry out the construction of predicting model to provide the
available reference for maintenance operators'' adjustment the base idle speed of
the gasoline engine. It is available not only to save the adjusting time and labor,
but also reach the higher quality. In this thesis, for getting the better learning
effect, we firstly propose the larger-is-better performance for the engine
revolution as the quality feature. By using the Taguchi method, we can find out
the significant factor for the said predicting model. Following that, we increase
the levels of the mentioned-above significant factor to improve the learning
iii
effect of the predicting model. Then, we plan the training experiments and the
examining experiments for the predicting model, respectively. Therefore, we
obtain the training datas and the examining datas, respectively. The training
datas are applied to construct the predicting model for the base idle speed of the
engine revolution. In addition, the examining datas are employed to examine the
accuracy for the said predicting model. At the same time, in the ANFIS, the
types of membership function for the antecedent part will affect the accuracy of
the predicting model. We will study the effect of membership function included
4 types (triangle function, trapezoid function, bell-shape function and Gauss
function) on the accuracy of the predicting model, so that we can select the
better membership function of the antecedent part for the predicting model. In
the experimented example, we can find that the bell-shape membership function
can get the least error and its output values are for the said predicting model
much closer to the experimental values.
|
|---|
