數據挖掘與機台參數最佳化整合

Abstract

成品的品質優劣往往影響一家公司的利潤，所以如何有效地維持製程的穩定，是所有業者都必須克服的難題。在實務上，經常透過實驗設計方法來找到機台參數與控制項之間的關係，進而建立預測模型，以尋找不同情況下最佳的機台參數配置。然而此模型的複雜度會隨著機台參數與控制項的數目上升成指數型的成長，單使用實驗設計方法來評估，其精確度已不如以往，但重新建立一個新的實驗設計方法，對業者而言是個龐大的成本。因此藉由數據挖掘的方法來找尋過去調機模式的趨勢特徵，進而建立修補項來補足實驗設計方法未考慮到的機台狀況。除此之外，和以往用兩階段先後進行實驗設計方法和調機特徵評估不同，本研究建立一個混整數規劃模型能同時考量到兩種狀況，此模型在複雜度高的問題上更能考慮到多機台參數與多控制項的交互影響，進而得到更好的結果表現，有了修補項的測模型也能更精確地提供調機策略。

As the quality of finished products often plays an important role in one company’s profitability, the issue of the stability during the manufacturing process is crucial for all manufacturers. In practice, the design of experiment (DOE) is often used to determine the relationship between machine parameters and the controlled items. We then use the DOE results to build the forecast model to determine the best parameter setting. However, the complexity of the forecast model exponentially increases in the number of machine parameters and controlled items. The accuracy of the DOE method may not be accountable while rebuilding a new DOE method can be a huge cost for the manufacturer. Therefore, by adopting the data mining method, we can conduct the pattern mining on the data gathered from the past tuning mode. In this way, we can acquire more details which could not be measured before in the DOE method. In addition, we have to go through two stages to use the result of DOE method followed by the evaluation of tuning pattern. This study establishes a mixed-integer programming model, which simultaneously considers the result of DOE method and the evaluation of tuning pattern. Since this model can consider the interaction between multi-machine parameters and multiple controlled items, it performs better when solving problems of high complexity. Finally, the forecast model with supplementary items can also provide tuning strategies with better accuracy.