在當前全球供應鏈高度波動與不確定性加劇的背景下，製造業面臨自然災害、地緣政治風險、運輸瓶頸及國際貿易摩擦等多重挑戰，尤以高度客製化生產模式之影響尤為顯著。傳統品質管制與監測方法已難以有效因應原物料價格波動、能源及人力成本上升以及技術快速迭代所帶來之營運壓力。隨著 2024 年被視為台灣 AI 法規元年，能源消耗與電力負荷議題亦受到產業界廣泛關注，促使企業積極投入超高壓變電設施建置，以降低潛在缺電風險。本研究聚焦於台灣機電工程製造產業，選取高階客製化電力變壓器為實證對象，旨在探討於小批量、數據稀疏情境下，如何運用先進深度學習技術提升成品品質預測之準確性與穩健性。針對小批量學習易產生高變異性及過擬合之問題，本研究設計兩種混合模型比較第一模型為結合生成對抗網路（Generative Adversarial Network, GAN）與 AdaBoost 集成學習之資料增強型混合模型，用以強化少量樣本之代表性與多樣性；第二個模型為結合 t 分佈隨機鄰域嵌入法（t-distributed Stochastic Neighbor Embedding, t-SNE）與 AdaBoost 之非線性降維型混合模型，藉由保留資料局部結構以提升學習效能並降低高維度資料帶來的複雜性。實證結果顯示，所提出之 t-SNE–AdaBoost 混合模型在處理高度客製化且樣本有限之品質預測任務時，較傳統基準模型可提升近 20% 之預測準確率，證實資料結構提取結合集成學習技術之可行性與優勢。同時，GAN–AdaBoost 混合模型亦驗證了虛擬樣本生成對擴充原始數據及緩解小批量學習限制具有正面效果。本研究之成果可為智慧製造領域中品質預測系統之開發與應用提供實證依據與技術參考，對於提升製造彈性、強化產品一致性及降低營運成本，具備實務應用與學術貢獻之價值。

In the context of increasing global supply chain volatility and uncertainty, the manufacturing industry is facing numerous challenges, including natural disasters, geopolitical risks, transportation bottlenecks, and international trade frictions. These challenges are particularly pronounced in highly customized production environments. Traditional quality control and monitoring methods have proven inadequate in effectively addressing the operational pressures arising from fluctuations in raw material prices, rising energy and labor costs, and rapid technological advancements. With 2024 anticipated to be the inaugural year for AI regulatory frameworks in Taiwan, issues related to energy consumption and power load have garnered significant attention from the industry. This has prompted enterprises to actively invest in the construction of ultra-high-voltage substation facilities to mitigate potential power shortages. This study focuses on the electromechanical engineering manufacturing sector in Taiwan, specifically examining high-end customized power transformers as the empirical subject. The research investigates how advanced deep learning techniques can enhance the accuracy and robustness of finished product quality predictions under conditions of small-batch production and sparse data. To address the high variability and overfitting issues commonly encountered in small-sample learning, this research designs and compares two hybrid models. The first model integrates a Generative Adversarial Network (GAN) with Adaptive Boosting (AdaBoost) as a hybrid data augmentation approach, aiming to enhance the representativeness and diversity of limited samples. The second model combines t-distributed Stochastic Neighbor Embedding (t-SNE) with AdaBoost as a nonlinear dimensionality reduction hybrid model, aiming to preserve local data structures and improve learning performance while reducing the complexity associated with high-dimensional data. Experimental results demonstrate that the proposed t-SNE–AdaBoost hybrid model, when applied to quality prediction tasks for highly customized products with limited samples, improves prediction accuracy by nearly 20% compared to the baseline model. This validates the feasibility and advantages of integrating data structure extraction with ensemble learning. Meanwhile, the GAN–AdaBoost hybrid model further confirms the positive impact of virtual sample generation in augmenting original data and mitigating the challenges associated with small-sample learning. The findings of this study provide empirical evidence and a technical reference for the development and implementation of quality prediction systems in smart manufacturing. They offer practical value by enhancing manufacturing flexibility, improving product consistency, and reducing operational costs, thereby contributing valuable insights to both academia and industry.