利用類神經網路與線性迴歸進行成本預測之研究-以印刷電路板產業為例

Using Artificial Neural Network and Linear Regression Approach to Predict Cost in PCB Industry

楊金聲(Chin-Sheng Yang)

指導教授 ： 皮世明

中原大學/商學院/資訊管理研究所/碩士(2005年)

https://doi.org/10.6840/cycu200500367

摘要

在導入ERP熱浪風潮後，使我們瞭解系統沒有所謂的好壞，只有合不合身。當今市面上有很多知名的ERP系統，如Oracle、SAP系統，其中上線成功並幫企業做好營運管理的成功案例很多。然而綜觀整個PCB產業，為其ERP量身訂做的系統很少，而提供適合PCB產業的成本計算與成本預測的完整功能的幾乎沒有，有些企業已經成功導入ERP系統，但是所使用的都是目前在適合組裝業的標準化系統，但因原本就非為PCB產業所設計，所以在成本結算與預測上，績效不彰。
既然績效不彰，如何正確快速的計算PCB的成本，進而克服PCB企業成本預估的瓶頸，更精準快速的成本預估以助相關決策支援。本研究希望透過類神經網路與線性迴歸來探討PCB產業成本預測的可行方法，而且產品型態日新月異，成本也依據產品的改變或製程的更新或原物料的進貨成本而變動，如何才能縮小成本預測誤差？希望藉此提供PCB企業建構成本預測系統的參考，或能藉此能對PCB企業主或資訊主管在選擇與導入成本預測系統能有所幫助。
本論文以ABC成本制的原則，主要鎖定成本最高的六種成本分類「壓合物料」、「防焊物料」、「金手指物料」、「鑽孔變動製造費用」、「成型變動製造費用」與「測試變動製造費用」深入探討，結果發現類神經網路在預測的表現上準確且穩定，而線性迴歸因為擁有能解釋變數的能力，有助於改善預估準確性的資訊搜集。而且在本研究的範圍內使用線性迴歸配合分類迴歸樹，可以得到如類神經網路一樣的準確度，並能提供解釋變數的能力。
在區分距今期間長短為測試基準的測試中，「金手指物料」的預測準確度在不同的時間區段有明顯的不同，及在迴歸檢定的係數變化，說明了成本動因的影響程度因為時間的變化而改變，說明了選擇正確的訓練週期有助於預測的準確性。

關鍵字

線性迴歸 ； 類神經網路 ； 成本預測 ； 印刷電路板 ； 分類迴歸樹

並列摘要

Since implement ERP becoming widely popular, make us understand that we do not need the so-called best system, what we need is fit. Nowadays there are many famous ERP systems on the market, such as Oracle, SAP…and so on, among them there are a lot of successful cases to reach the standard grade and helping enterprises to do a good management job. But take a broad view of the whole PCB industry, there are few systems made according to its ERP's production, and can not offering suitable and full function for cost calculation and cost predict , some enterprises have already succeeded in implementing ERP system, but what use is all to suit the standardized system of the package trade at assembly industries, but the script is not designed by PCB industry , so on calculation and predicting cost , the performance is not well apparently .
Since the performance is not well , how to calculation cost correctly and quickly in PCB, and then overcoming the bottleneck of the entreprise estimating cost in PCB industry, in advance to help cost estimated more perfect and fast for support of relevant decision. This research hopes to probe into the feasible track law that the cost predicted of PCB industry through Artificial Neural Network and Linear Regression. But products style changed with each passing day, according to change of products , the process also make renewal , and raw material also change , and purchasing cost of supplies change too. How could narrow the error of the cost prediction ? Hope can offer PCB enterprises to build and construct the cost predict systematically reference by this, or can be helpful for choosing and implementing PCB enterprise's cost predicting system by this.
Follow the principle of ABC thesis, fix six kinds of highest cost class to process predicting , that is ‘MLB Lam. Material’, ’Solder Resist Material’, ’Golden Finger Material’, ’Drilling Var. Over Head’ ,’Routing Var. Over Head’, ’E.Test Var. Over Head’. Deepen the discussion the behavior, and finded that the predicted of Artificial Neural Network was being accurately and steadily finally, and Linear Regression is ability to explain the parameter , contribute to improving estimating accuracy and to collect the information of estimating. And use Linear Regression to cooperate to Classify and Regression Tree in the range of this research, can get the result as accurate as Artificial Neural Network , and also offer the ability to explain parameter .
Base on distinguishing distance from today test, there is obvious difference in different time sectors in the prediction accuracy on ’Golden Finger Material’, and the coefficients of Linear Regression is coming back to changes , prove that the influence degree of cost driven factor changes because of change of time, have stated that choosing the correct training time cycle is useful to accuracy predicting.

並列關鍵字

PCB ； Classify and Regression Tree ； Linear Regression ； Artificial Neural Network ； Cost Predicted

參考文獻

29. 連惟謙，應用資料分析技術進行顧客流失與顧客價值之研究，中原大學資訊管理研究所未出版碩士論文，民93年。

5. 朱國珍，信評式債券型基金之評價-通用迴歸類神經網路模式應用，臺中健康暨管理學院經營管理研究所未出版碩士論文，民93年。

12. 呂奇傑，演化式類神經網路分類技術於資料探勘上之應用，輔仁大學應用統計學研究所未出版碩士論文，民89年。

15. 李炎燊，多層印刷電路板製程中組裝與加工次序之整合分析與評估，元智大學工業工程研究所未出版碩士論文，民86年。

22. 唐永成，印刷電路板製造資料模型建構與成本分析，元智大學工業工程與管理學研究所未出版碩士論文，民91年。