利用倒傳遞類神經網路探討深孔鑽孔壁表面特性之預測模式

A Predictive Model on the Surface Characteristics of Deep-hole Drilling Explored by Back-propagation Neural Network

陳昱瑋(Yu-Wei Chen)

指導教授 ： 簡文通

國立屏東科技大學/工學院/機械工程系所/碩士(2010年)

https://doi.org/10.6346/NPUST.2010.00073

摘要

本研究之目的在探討中碳鋼S45C槍鑽深孔加工之表面特性並建立其表面粗糙度預測模式。首先利用全因子法規劃實驗參數的配置，以深孔加工後之孔壁粗糙度及刀腹磨耗為目標品質作探討。相關實驗數據並用做為架構以倒傳遞類神經網路為理論基礎之孔壁粗糙度預測模式。整體實驗規劃分為三部分，第一部分經由前置實驗結果選定影響槍鑽深孔加工最基本的兩個參數，分別為轉速與進給速度，各有5個水準，依照全因子法配置深孔加工參數組合進行實驗。第二部份則是搭配倒傳遞類神經網路以第一部份的25組實驗數據，提供倒傳遞類神經網路架構之預測模式所需的練習範例及回想範例，再利用田口法之平均數分析找出最佳化網路參數組合。第三部分則是配置9組不包含回想範例及訓練範例之深孔加工參數進行實驗以驗證預測模式之準確性。結果顯示在最佳實驗組中，分別得到孔壁粗糙度(Ra)0.46μm，及刀腹磨耗(VB)量為0.13mm等為最佳深孔加工品質。此外，經由驗證實驗得知利用倒傳遞類神經網路所架構之孔壁粗糙度(Ra)預測模式的平均誤差為6.64%，顯示此模式具有良好的預測能力。本研究的過程及結果對於槍鑽深孔加工中碳鋼S45C之加工品質及結果預測可提供實質上的協助及參考。

關鍵字

深孔加工 ； 槍鑽 ； 倒傳遞類神經網路 ； 預測模式

並列摘要

The purpose of this study is to investigate the characteristics for the deep-hole drilling of carbon steel S45C in a gun drilling process and construct a predictive model for predicting its surface roughness. Firstly, experiments are arranged by an all-factor method to conduct deep-hole drilling. After experiments have been performed, the workpiece surface roughness and the tool flank wear are selected as single quality objective, respectively the experimental results can be further used for constructing a deep-hole surface roughness predictive model that is based on a back-propagation neutral network algorithm. The overall experimental procedure is divided into three parts. In the first part, according to the preliminary tests there are two deep-hole drilling parameters that showed stronger influence on drilling quality have been selected; namely, spindle speed and feed rate, each with 5 levels. The experiments were conducted based on the arrangement of deep-hole drilling parameters by all-factor method. In the second part, the former experimental results were used for training patterns and recalling patterns of the predictive model which was constructed base on a back-propagation neutral network. The optimum network parameters were attained by the average mean analysis of Taguchi method. In the third part, there are nine sets of verifying experiments without including any experiment of the training patterns or the recalling patterns have been conducted to validate the accuracy of this predictive model. It has shown that a 0.46μm for the best surface roughness(Ra) and a 0.13mm for the best flank wear in the experimental group, Moreover, the results of the verifying experiments showed that a mean error as 6.64% was found when the predictive values were compared. It indicates the predictive model developed base on a back-propagation neural network has good predicting ability for the surface roughness(Ra) in a deep-hole drilling process with gun drill. The processes and results in this study provide substantial assistance and reference efficiently for a gun-drill in deep-hole drilling carbon steel S45C.

並列關鍵字

Deep-hole drilling ； Gun drill ； Back-propagation neutral network ； Predictive Model

參考文獻

[5] 張逸祥，2005，槍鑽深孔加工表面品質最佳化設計，碩士論文，大同大學，機械工程研究所，台北。

[15] Katsuki A., Sakuma K., Tabuchi K., Onikura H., Akiyoshi H. and Nakamuta Y., 1987, “The Influence of Tool Geometry on Axial Hole Deviation in Deep Drilling : Comparison of Single- and Multi-Edge Tools”, JSME International Journal, Vol. 30, No. 265, pp. 1167-1174.

[16] Katsuki A., Onikura H., Sakuma K., Chen T. and Murakami Y., 1992, “The Influence of Workpiece Geometry on Axial Hole Deviation in Deep Drilling”, JSME International Journal Series III, Vol.35, No.1, pp. 160-167.

[17] Gao C. H., Cheng K. and Kirkwood D., 2000, “The Investigation on the Machining Process of BTA Deep Hole Drilling”, Journal of MaterialsProcessing Technology, pp. 222-227.

[18] Choudhury S. K., Jain V. K. and Ranma Rao Ch. V. V., 1999, “On-line monitoring of tool wear in turning using a neural network,” International Journal of Machine Tools & Manufacture, Vol. 39, pp. 489-504.