本研究聚焦於高密度電子製程中的機台差異對鑽孔品質潛在影響，並提出一套基於振動訊號分析的監測方法，作為健康狀態判斷依據。考量目前多數產線仍依賴人工巡檢與固定保養週期，對連續高速加工中早期異常的掌握有限，本論文以實際部署的傳統鑽孔機與雷射鑽孔機為實驗對象。方法上，特別以振動衰減係數（Decay Coefficient）作為描述系統能量耗散與動態穩定性指標，搭配線性回歸與隨機森林回歸模型進行品質預測與特徵排序，並結合 SHAP 解釋法（SHapley Additive exPlanations）提升模型可解釋性。 結果顯示，線性回歸模型對製程能力指標 CPK 具高度預測力：傳統鑽孔機在空轉與加工階段的振動變異與 CPK 呈顯著線性關聯，振動變異對 CPK 的解釋力達 97.26 %。雷射鑽孔機則經 SHAP 分析辨識出多組具顯著交互效應的參數組合，驗證特徵排序機制有效性。與僅以 OA 值設定閾值的傳統方法相比，所提架構能更精細掌握機台狀態，並具部署彈性。當異常樣本有限時，此方法仍能即時提供可解釋的健康監測與品質預測，具備高變動製程中的穩定性與預測維護價值。;This thesis explores how machine variability affects drilling quality in high-density electronic manufacturing processes and presents a vibration-based monitoring approach to assess machine health. In modern industrial settings, production lines typically depend on manual inspections and fixed maintenance schedules, which offer limited responsiveness to early-stage anomalies during continuous, high-speed operations. To overcome this limitation, the study introduces a cross-machine vibration monitoring framework that incorporates both conventional mechanical drilling machines and laser drilling machines operating in real production environments. The methodology centers on the use of the vibration decay coefficient as an indicator of energy dissipation and dynamic stability within the system. Quality prediction and feature ranking are performed using linear regression and random forest regression models, while the SHAP (SHapley Additive exPlanations) framework is incorporated to enhance model interpretability. The results demonstrate that the linear regression model exhibits strong predictive capability for the process capability index (CPK). In conventional drilling machines, a significant linear relationship was observed between vibration variability during both idle and machining stages and the CPK, with vibration variability explaining up to 97.26% of its variation. For laser drilling machines, SHAP analysis identified multiple parameter combinations with notable interaction effects, confirming the effectiveness of the proposed feature-ranking mechanism. Compared to traditional threshold-based approaches that rely solely on OA values, the proposed framework offers more precise detection of machine condition changes and greater flexibility for deployment. Even under conditions with limited abnormal samples, the method is capable of providing real-time, interpretable health monitoring and quality prediction, demonstrating both stability and practical value for predictive maintenance in high-variability manufacturing environments.
