| 摘要: | 在半自動化生產環境中，直接人員工時需求受到產能變動、設備狀態與人力管理因素影響，若僅依賴經驗法則或歷史平均進行規劃，容易造成工時預測偏誤，進而影響生產效率與營運穩定性。此外，實務資料往往來自不同管理或作業層級，其資料生成頻率與分析單位不一致，使工時需求預測在模型建構與結果穩定性上面臨挑戰，特別是在可用樣本有限的情境下。
本研究以某LED封裝製造公司的半自動化生產線為研究對象，整合生產、人力與設備等多來源資料，提出一套適用於資料顆粒度不一致情境的人員工時需求預測架構。研究採比較式設計，先建構單層模型作為基準，再提出分層模型，其中第一層以站別層級資料建立線性迴歸模型，第二層則以月度層級管理變數校正預測殘差。
為因應小樣本特性，本研究結合主成分分析法與自助法檢驗模型穩定性，並以判定係數與均方根誤差作為評估指標。實證結果顯示，分層模型在預測精準度與穩定性上皆顯著優於單層模型，且在多次重抽樣下能維持一致表現。此外，透過移動平均法推估未來輸入變數，使模型具備跨月度預測能力。
在模型解讀方面，敏感度分析結果顯示站別結構差異為影響工時需求的主要因素，而設備叫修次數在連續型變數中具有顯著影響，並可辨識具有管理意義的轉折區間。整體而言，本研究所提出的分層建模策略，能有效處理資料顆粒度不一致與小樣本問題，並提供具實務價值的人力規劃與改善決策依據。;In semi-automated manufacturing environments, direct labor hour demand is influenced by production variability, equipment conditions, and workforce management factors. When planning relies solely on experiential rules or historical averages, biased labor hour forecasts are likely to occur, thereby affecting production efficiency and operational stability. Moreover, practical manufacturing data are often generated across different managerial and operational levels, resulting in heterogeneous data granularity and inconsistent analytical units. These characteristics pose additional challenges to labor hour demand forecasting, particularly under limited sample conditions.
This study investigates a semi-automated production line of an LED packaging manufacturer and integrates multi-source production, manpower, and equipment data to develop a labor hour demand forecasting framework suitable for heterogeneous data granularity. A comparative research design is adopted by first constructing a single-layer model as a baseline, followed by a proposed two-stage ensemble learning framework. In the proposed framework, the first stage employs station-level data to build a linear regression model that captures operational differences among workstations, while the second stage utilizes month-level managerial variables to correct the prediction residuals.
To address the small-sample characteristic, principal component analysis and bootstrapping are employed to evaluate model stability, with model performance assessed using the coefficient of determination and root mean squared error. Empirical results indicate that the proposed two-stage ensemble model significantly outperforms the single-layer baseline model in terms of both predictive accuracy and stability, and maintains consistent performance across repeated resampling. In addition, a moving average approach is applied to estimate future input variables, enabling multi-period labor hour forecasting.
From a model interpretation perspective, sensitivity analysis reveals that structural differences among workstations are the primary drivers of labor hour demand, while machine repair frequency exhibits a significant impact among continuous variables and helps identify management-relevant turning point ranges. Overall, the proposed two-stage ensemble learning approach, which follows a hierarchical modeling rationale, effectively addresses heterogeneous data granularity and small-sample challenges, and provides practical insights for workforce planning and decision-making for operational improvement. |
