國立虎尾科技大學 |

Language: English

Back

探討採樣策略對檢驗預測模型之影響 = = The Impact of S...

鄭宇倫

探討採樣策略對檢驗預測模型之影響 = = The Impact of Sampling Strategies on Inspection Predictive Models /

Record Type:	Language materials, printed : Monograph/item
Title/Author:	探討採樣策略對檢驗預測模型之影響 =/ 鄭宇倫.
Reminder of title:	The Impact of Sampling Strategies on Inspection Predictive Models /
remainder title:	The Impact of Sampling Strategies on Inspection Predictive Models.
Author:	鄭宇倫
Published:	雲林縣 :國立虎尾科技大學 , : 民113.08.,
Description:	[9], 76面 :圖, 表 ; : 30公分.;
Notes:	指導教授: 李英聯.
Subject:	Confusion Matrix. -
Online resource:	電子資源

探討採樣策略對檢驗預測模型之影響 = = The Impact of Sampling Strategies on Inspection Predictive Models /
鄭宇倫

探討採樣策略對檢驗預測模型之影響 =The Impact of Sampling Strategies on Inspection Predictive Models /The Impact of Sampling Strategies on Inspection Predictive Models.鄭宇倫. - 初版. - 雲林縣 :國立虎尾科技大學 ,民113.08. - [9], 76面 :圖, 表 ;30公分.

指導教授: 李英聯.

碩士論文--國立虎尾科技大學工業管理系工業工程與管理碩士班.

含參考書目.

自動光學檢測是電子產業常用檢測儀器，決定產品最終品質關鍵，檢測中常有漏報或誤檢情形，需再次複檢，此情形為產業所不樂見，改善瑕疵檢測是產業重視一大議題。而瑕疵品即早發現，能節省不必要的成本，故不斷的改善自動光學檢測是解決問題的關鍵。收集數據常發生數據不平衡問題，導致機器學習無法有效分類，採樣方法是解決類別不平衡常用方法，目的於使類別數據能趨於平衡，而資料的不同採樣，所訓練出模型會有不同結果。本研究探討不同比例及不同採樣方法，嘗試解決不平衡情況，數據集為表面貼裝技術產線不平衡資料集。本研究採用不同資料分布模擬不同的不平衡情況，使用隨機降採樣、隨機過採樣及隨機過採樣結合隨機降採樣，三種不同採樣透過隨機森林得出分類效果比較何種採樣方法適合解決此問題，並用該數據集不同的不平衡問題是否有效。而驗證其數據集前，本研究使用單因子重複量數變異數分析，檢測模型效度，並以Cohen's Kappa判斷模型信度，證明模型的效果和穩定度。研究成果發現，隨機過採樣結合隨機降採樣優於隨機過採樣優於隨機降採樣，防止漏檢及誤報，而漏檢比誤報更為嚴重，本研究選擇隨機降採樣為模型，原因於隨機降採樣有效地降低漏檢情況發生而誤報的情況也還在可控制的範圍內，該數據集不同不平衡結果本研究模型是有效，不過數據集本身瑕疵樣本極少，造成不平衡比例越高，漏檢率有提升現象，說明不同不平衡問題會因原始樣本數多寡影響分類結果。.

(平裝)Subjects--Topical Terms:

1449068
Confusion Matrix.

探討採樣策略對檢驗預測模型之影響 = = The Impact of Sampling Strategies on Inspection Predictive Models /
LDR:05155cam a2200241 i 4500 001 1129962
008 241015s2024 ch ak erm 000 0 chi d
035 $a (THES)112NYPI0030024
040 $a NFU $b chi $c NFU $e CCR
041 0 # $a chi $b chi $b eng
084 $a 008.169M $b 8732 113 $2 ncsclt
100 1 $a 鄭宇倫 $q (Yu-Lun Cheng) $3 921105
245 1 0 $a 探討採樣策略對檢驗預測模型之影響 = $b The Impact of Sampling Strategies on Inspection Predictive Models / $c 鄭宇倫.
246 1 1 $a The Impact of Sampling Strategies on Inspection Predictive Models.
250 $a 初版.
260 # $a 雲林縣 : $b 國立虎尾科技大學 , $c 民113.08.
300 $a [9], 76面 : $b 圖, 表 ; $c 30公分.
500 $a 指導教授: 李英聯.
500 $a 學年度: 112.
502 $a 碩士論文--國立虎尾科技大學工業管理系工業工程與管理碩士班.
504 $a 含參考書目.
520 3 $a 自動光學檢測是電子產業常用檢測儀器，決定產品最終品質關鍵，檢測中常有漏報或誤檢情形，需再次複檢，此情形為產業所不樂見，改善瑕疵檢測是產業重視一大議題。而瑕疵品即早發現，能節省不必要的成本，故不斷的改善自動光學檢測是解決問題的關鍵。收集數據常發生數據不平衡問題，導致機器學習無法有效分類，採樣方法是解決類別不平衡常用方法，目的於使類別數據能趨於平衡，而資料的不同採樣，所訓練出模型會有不同結果。本研究探討不同比例及不同採樣方法，嘗試解決不平衡情況，數據集為表面貼裝技術產線不平衡資料集。本研究採用不同資料分布模擬不同的不平衡情況，使用隨機降採樣、隨機過採樣及隨機過採樣結合隨機降採樣，三種不同採樣透過隨機森林得出分類效果比較何種採樣方法適合解決此問題，並用該數據集不同的不平衡問題是否有效。而驗證其數據集前，本研究使用單因子重複量數變異數分析，檢測模型效度，並以Cohen's Kappa判斷模型信度，證明模型的效果和穩定度。研究成果發現，隨機過採樣結合隨機降採樣優於隨機過採樣優於隨機降採樣，防止漏檢及誤報，而漏檢比誤報更為嚴重，本研究選擇隨機降採樣為模型，原因於隨機降採樣有效地降低漏檢情況發生而誤報的情況也還在可控制的範圍內，該數據集不同不平衡結果本研究模型是有效，不過數據集本身瑕疵樣本極少，造成不平衡比例越高，漏檢率有提升現象，說明不同不平衡問題會因原始樣本數多寡影響分類結果。.
520 3 $a Automatic optical inspection (AOI) is a commonly used inspection instrument in the electronics industry. It is the key to determining the final quality of the products. During inspection, there are often omissions or false detections, which require re-inspection. This situation is not welcomed by the industry. Improving defect detection is an important issue that the industry attaches great importance to. Early detection of defective products can save unnecessary costs, so continuous improvement of AOI is the key to solving the problem. Data imbalance often occurs when collecting data, resulting in the inability of machine learning to effectively classify. The sampling method is a common method to solve the category imbalance. The purpose is to make the category data tend to be balanced. Different sampling of the data will produce different results for the trained model. This study explores different proportions and different sampling methods to try to solve the imbalance situation. The data set is the surface-mounted technology production line imbalance data set. This study uses different data distributions to simulate different imbalance situations, using Random Under-Sampling (RUS), Random Over-Sampling (ROS), and ROS combined with RUS. Three different samplings are used to obtain classification effects through random forests to compare which sampling method is suitable to solve this problem. The question and whether the different imbalance problems are valid with this data set. Before validating its data set, this study used one-way repeated measures ANOVA to test the validity of the model, and used Cohen's Kappa to judge the reliability of the model to prove the effect and stability of the model. The research results show that ROS combined with RUS is better than ROS than RUS in preventing missed detections and false alarms. Missed detections are more serious than false alarms. This study chose RUS as the model because RUS can effectively reduce the occurrence of missed detections and the number of false positives is still within a controllable range. The different results of the imbalance of this data set are effective. This research model is effective, but the data set itself has very few defective samples, resulting in an imbalanced ratio. The higher the value, the higher the missed detection rate, indicating that different imbalance problems will affect the classification results due to the number of original samples..
563 $a (平裝)
650 # 4 $a Confusion Matrix. $3 1449068
650 # 4 $a Random Forest. $3 1125891
650 # 4 $a Defect Detection. $3 1419456
650 # 4 $a Class Imbalance. $3 1449067
650 # 4 $a Random Under-Sampling. $3 1449066
650 # 4 $a 混淆矩陣. $3 1407504
650 # 4 $a 隨機森林. $3 1125894
650 # 4 $a 瑕疵檢測. $3 1099910
650 # 4 $a 類別不平衡. $3 1449065
650 # 4 $a 隨機降採樣. $3 1449064
856 7 # $u https://handle.ncl.edu.tw/11296/v5v6us $z 電子資源 $2 http