產品的出產良率代表著一個產品最終的價值，而半導體產品製造過程複雜且繁多，產品瑕疵在一道道的製程下累積，因此在如此多的製造過程中設置檢測點已必不可少。每一次產品進行檢測，無法即時進入下一階段製程產生了時間成本，目前檢測人員透過機台掃描資訊進行判讀，於產品中進行隨機抽樣，往往無法檢測到有問題的關鍵瑕疵，因此若能對那一種機台進行解讀，得到產品目前狀況，便能夠有效改善產品之良率，減少量測時間。 本論文以上述為出發點，透過國內某工廠之實際產品檢測數據，針對產品瑕疵與非瑕疵資料的不平衡情況，以人工智慧堆疊模型建立一產品瑕疵預測模型，透過不同機器學習模型的組合，以期能提升檢測速度與準確度。 ;The yield rate of a product is a crucial factor in determining its final value, since the semiconductor manufacturing process is complex and varied. Product defects accumulate over different processes, making it essential to set up detection points in the manufacturing processes. However, the product won′t be able to enter the next process while detecting, which leads to a huge time cost. Currently, fetching the data of the product from scanning machines, and randomly sampling the defects, couldn′t solve the problem of the key product defects detection effectively. Therefore, analyzing the data from the inspection machine, and find out the current state of the product, will improve the yield of the product effectively, and reduce the measurement time. Base on the basis mentioned about, this thesis uses artificial intelligence stacking model to establish a product defect prediction model based on the actual product detection data from a domestic factory, aiming at the imbalance between product defect and non-defect data, and through the combination of different machine learning models, in order to improve the detection speed and accuracy.
