A Study of The Prediction Model by Integrating Fuzzy Theory and Back-propagation Neural Networks

• Main Authors: GAO, PEI-RU, 高佩如
• Other Authors: HOU, TUNG-HSU
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/60226219690286994189

碩士 === 國立雲林科技大學 === 工業工程與管理系 === 104 === We often explore the causation lying in issue by using data analysis in our life. Structured data is divided into two categories, namely continuous data and discrete data. In this study, discrete data was the IC assembly wire bonding process parameter of Taguchi method, while continuous data was compiled from the data collected at Douliou air quality monitoring station. The purpose of this study is first, to apply Fuzzy theory integrated with Back-Propagation Neural Network to predict the discrete data and continuous data, second, to identify which of the following models, ie.Back-Propagation Neural Network (BPNN), Adaptive Neuro-Fuzzy Inference System (ANFIS) and Fuzzy Back-Propagation Neural Network (FBPN) would help provide more accurate data of discrete data and continuous data, and third, to make a summary of the appropriate data type when data was predicted by using BPNN, FBPN and ANFIS. This study finds that discrete data Taguchi process outputs predicted by FBPN and BPNN had high accuracy of prediction. Mean absolute percentage error was less than 10%. However, discrete data Taguchi process outputs predicted by ANFIS had excellent accuracy of prediction, but its prediction output value was a single value. Continuous data air quality of PM10 concentrations predicted by FBPN and BPNN, their the mean absolute error percentages were between 20% to 50%. The obtained prediction mode accuracy of BPNN and FBPN was reasonable. The air quality of PM2.5 concentrations predicted by FBPN and BPNN, their mean absolute percentage error were higher than 50%. The air quality of PM10 and PM2.5 concentrations predicted by ANFIS, its mean absolute percentage error was also higher than 50%, and predicted output was a single number value. Three kinds of network prediction model utilized to predict the air quality of particle matter concentrations were unable to achieve high accuracy. The study suggests that continuous air quality data should include more attributes from the open database for further study, then the three prediction models were tested to verify their prediction ability.