本研究探討了學習曲線在產品模組製造中的應用，以國防技術公司作為研究個案。個案公司專業從事國防科技武器和軍民兩用技術的開發、製造和服務。儘管經驗豐富，但該公司在數據分析方法上缺乏突破，本研究應用公司尚未使用過的「學習曲線」模型作為管理工具，以更瞭解產品製造期間的實際情況。 本研究採用學者Cobb-Douglas所提出的Multiplicative Power Model來定義多變數學習曲線，並將其應用於三個主要X、Y、Z產品來建立多變數和單變數學習曲線模型。數據經過對數轉換以進行線性迴歸分析。並將產品之線性迴歸結果經過各項驗證分析，包括線性迴歸模型之解釋力、顯著性檢驗、共線性診斷、異質變異診斷及是否符合常態性的假設。對於X產品和Y產品，多變數模型的表現優於單變數模型；而Z產品之資料經驗證為不適合分析(R^2判定係數< 0.5)。以線性迴歸導出之學習曲線模式，建構PQR達交率、產品批量生產量與品質管制之動態學習模式並加以應用。 本研究展示出學習曲線的應用在改善製造管理方面的潛力，學習曲線模型提供個案公司作為一個更合理有效的分析工具，成功建構了X和Y產品的學習曲線模型，對未來之產品製造具參考價值。本研究局限於資料敏感度和簡化假設，建議未來的研究可以添加更多變數，嘗試其他數據轉換，並突破當前的限制。;The thesis explores the application of learning curves in product modular manufacturing, specifically in the context of a defense technology company. Case company specializes in developing, manufacturing, and servicing defense technology weapons and military-civilian dual-use technologies. Despite its rich experience, the company lacks breakthrough in data analysis methods. The thesis is to apply the "learning curve" model, which the company has not used before, to better understand the actual situation during product manufacturing periods. The research uses the Cobb-Douglas Multiplicative Power Model to define learning curves, applying it to three main products (X, Y, Z) to construct multi-variable and single-variable learning curve models. Data is logarithmically transformed for linear regression analysis. The results of this study: (1) X and Y products showed good explanatory power (R^2>0.5) and passed various validations. (2) Multi-variable models performed better than single-variable models for X and Y. (3) Z product data was not suitable for analysis (R^2<0.5). This thesis demonstrates the potential of learning curve models in improving manufacturing management, providing company a more rational and effective analysis tool. The study successfully constructed dynamic learning curve models for X and Y products, offering better reference value for future product manufacturing. Application in PQR delivery rate, product batch production volume, and quality control. The research limited by data sensitivity and simplifying assumptions, hoping that future research could add more variables, try other data transformations, and break through current limitations.
