| Summary: | 碩士 === 長庚大學 === 機械工程研究所 === 97 === Artificial neural networks (ANNs) are usually used as surrogate models to replace time-consuming numerical models due to ANN's quick solution for nonlinear problems. Because the training method used may affect the result of an ANN of given configuration, four training methods were compared and analyzed in this study. These methods are Levengerg-Marquardt (LM) method, genetic algorithm (GA), LM-GA, and disturbed LM method. The goal is to find a reliable method which gives the best ANN model in terms of prediction accuracy. Then, the ANN model is applied as a surrogate model in the fluid-film lubricated bearing analysis.
When the LM method is used as a training method (to get the network parameters), the result may be sensitive to the initial guesses of the network parameters used in the LM method. But, in general, the method can provide a satisfactory result. Since the range of the network parameters is very difficult to know in advance, the results of constructing ANN by using GA were usually very poor in the tested cases. In the case of using LM-GA method as the training method the result also shows little improvement over by using LM method alone. Finally, the disturbed LM method exhibits little benefit when the disturbing ratio is small. However, the result may be changed dramatically if a large disturbing ratio is applied. To overcome the pitfall resulted from poor initial guesses in using LM method, several runs of LM method using different initial guesses are suggested.
In this study, the numerical lubrication model in the optimization problem is replaced by the ANN model constructed by the LM method. The objective of the optimization is to optimize the stiffness of a fluid-film lubricated bearing. The optimization method is a genetic algorithm. The results show that the procedure can find the optimum bearing stiffness successfully. Similar approach can also be applied to other computationally intensive optimization problems.
|
|---|
