運用重組粒子群與LM法於倒傳遞類神經發展銑削表面糙度預測系統

The Development of a Surface Roughness Prediction System in Milling Operations by Implementing RPSO and LM in BPN

詹淳旭(Chun-Hsu Chan)

指導教授 ： 黃博滄

中原大學/電機資訊學院/工業與系統工程研究所/碩士(2013年)

https://doi.org/10.6840/cycu201300822

摘要

類神經網路發展至今已有許多成熟的模式被提出，其中以監督式的倒傳遞類神經使用最為普遍。傳統的倒傳遞類神經網路採用最陡坡降法(Steepest Descent Method)來訓練與更新權重值，具有以下缺點：(1)容易收斂至局部最小值(Local Minima)；(2)權重更新速率慢，學習時間長；(3)可能造成發散結果。

故本研究提出以粒子群演算法中的重組粒子群演算法作為前段的權重搜尋法，以達到全域搜尋，求得一組較好的權重值，再以Levenberg Marquardt(LM)法作為後段的權重搜尋法，以期使倒傳遞類神經網路的學習過程更有效率且讓結果不易陷入局部最小值。最後本研究將以銑削加工的表面粗糙度作為驗證資料，並透過田口方法將類神經網路與重組粒子群的參數進行參數配置，得到最佳的參數配置組合以建構重組粒子群LM倒傳遞類神經表面粗糙度預測系統。

為證明本研究所提出之方法的有效性、準確性與穩定性，將所發展之重組類子群LM倒傳遞類神經預測系統與LM倒傳遞類神經預測系統比較預測的準確性與穩定性，並使用t分配假設檢定與F分配假設檢定，比較兩種預測系統之顯著差異性，以驗證此預測系統之準確性與穩定性，最後再進行訓練資料最佳化測試。

透過假設檢定，其結果顯示本研究所提出之重組粒子群LM倒傳遞類神經預測系統在準確性與穩定性皆優於LM倒傳遞類神經預測系統；而在訓練資料最佳化測試部分，當訓練資料由原本的150筆縮減至80筆時，在顯著水準α = 0.05沒有顯著差異，故可將訓練資料縮減至80筆以來達到與150筆訓練資料相同的預測效果。

關鍵字

倒傳遞類神經網路 ； 重組粒子群 ； 表面粗糙度 ； 預測系統

並列摘要

An artificial neural network (ANN) was developed and has been widely applied for years. One of the most applications is the supervised back propagation neural network (BPN). A traditional BPN adopted the steepest descent method to train and adjust weight values between neurons. It has the following shortcomings: (1) apt to convergence to local minima (2) slowly updated weight values and had long learning time (3) it is possible to become divergence results if the learning speed value increases too much and too fast.

This study implements regrouping particle swarm optimization (RPSO) method to replace the steepest descent method at the beginning of training for BPN to achieve global search. It can prevent the weight values from local minima. And then use the Levenberg Marquardt (LM) method in posterior part to achieve a better result for BPN. Finally, this study applied the new model to develop a surface roughness monitoring system in milling operations and used Taguchi Method to obtain a set of training parameters from both ANN and RPSO to configure the parameters. We obtain the best parameter setting to construct Surface Roughness forecasting system of BPN based hybrid regrouping particle swarm optimization and LM (RPSOLM-BPN).

To prove the effectiveness, accuracy and stability of this study, we compare the accuracy and stability of the developed prediction results between RPSOLM-BPN and LM of Back-propagation neural networks (LM-BPN). A t-test and F-test are used to compare the differences between these two prediction systems. To verify the accuracy and stability of the prediction system, the optimization of the number of training data has been conducted.

The results show that RPSOLM-BPN prediction system which is proposed in this study has more accurate and stable results than that of the LM-BPN. Finally, in the study of optimization of training data set, there is no significant difference between 80 samples and 150 samples at significance level α = 0.05. As a result, training data can reduced to 80 samples, it can reach to the same prediction results as 150 training data samples.

並列關鍵字

Back Propagation Neural Networks ； Regrouping Particle Swarm Optimization ； Surface Roughness ； Prediction System

參考文獻

Huang, P. B., Chen, J. C., Lin, C. J., Lyu, P. H., & Lai, B. C. (2011). A Taguchi-neural-based in-process tool breakage monitoring system in end milling operations. Paper presented at the 3rd International Conference on Advanced Design and Manufacture, ADM2010, September 8, 2010 - September 10, 2010, Nottingham, United kingdom.

Cui, Z. H., Zeng, J. C., & Cai, X. J. (2004, June). A new stochastic particle swarm optimizer. Paper presented at the Evolutionary Computation, 2004. CEC2004. Congress on.

Abou-El-Hossein, K. A., & Yahya, Z. (2005). High-speed end-milling of AISI 304 stainless steels using new geometrically developed carbide inserts. Journal of Materials Processing Technology, 162–163(0), 596-602.

Ahilan, C., Kumanan, S., Sivakumaran, N., & Edwin Raja Dhas, J. (2013). Modeling and prediction of machining quality in CNC turning process using intelligent hybrid decision making tools. Applied Soft Computing, 13(3), 1543-1551.

Ak, R., Li, Y., Vitelli, V., Zio, E., López Droguett, E., & Magno Couto Jacinto, C. (2013). NSGA-II-trained neural network approach to the estimation of prediction intervals of scale deposition rate in oil & gas equipment. Expert Systems with Applications, 40(4), 1205-1212.