在工業4.0 時代，製造業為提升生產效率與產品品質，導入智慧製造以強化競爭力，透過整合物聯網 (IoT)、雲端運算(Cloud Computing) 及大數據分析(Big Data Analytics)，使企業能夠即時監控生產設備狀態、優化生產流程，並提升決策精準度。傳統的設備維護方式，往往難以有效預防突發故障，導致非計畫性停機 (Unplanned Downtime)，進而影響生產排程，甚至造成重大經濟損失。為解決此問題，企業開始採用數據驅動的智慧維護策略，透過感測器持續收集設備運行數據，並結合人工智慧(AI)與深度學習(Deep Learning)來進行數據分析與識別異常，應用於設備故障預測與診斷，以提升維護效率。 本研究以A公司塗佈機的歷史運行數據為基礎，採用預診斷與健康管理為維護策略。Transformer模型已被廣泛運用於各類序列任務中，因此本研究選用基於Transformer架構的PatchTST模型進行異常預測，目標為降低生產設備的意外停機風險並減少維護成本。最終模型於測試資料上達到 F1 Score 為 58.43%、Precision 為 53.04%、Recall 為 65.03% 的表現，雖未達到理想準確度，結果仍顯示Transformer在異常偵測任務上具備一定的潛在應用性，值得後續進一步優化與研究。 ;In the era of Industry 4.0, the manufacturing sector strives to enhance production efficiency and product quality by adopting smart manufacturing to strengthen its competitiveness. Through the integration of the Internet of Things (IoT), cloud computing, and big data analytics, enterprises can monitor equipment status in real time, optimize production processes, and improve decision-making accuracy. Traditional maintenance strategies often struggle to effectively prevent unexpected equipment failures, leading to unplanned downtime that disrupts production schedules and causes significant economic losses. To address this issue, companies are turning to data-driven intelligent maintenance strategies, continuously collecting operational data from sensors and applying artificial intelligence (AI) and deep learning techniques for anomaly detection and fault prediction, thereby improving maintenance efficiency. This study is based on historical operation data from Company A′s coating machine, adopting predictive diagnostics and health management as the maintenance strategy. Given the widespread application of Transformer-based models in various sequence tasks, this research employs the PatchTST model, built on the Transformer architecture, to perform anomaly prediction with the goal of reducing unexpected equipment downtime and minimizing maintenance costs. The final model achieved an F1 Score of 58.43%, Precision of 53.04%, and Recall of 65.03% on the test dataset. Although the accuracy did not reach an ideal level, the results suggest that the Transformer architecture holds potential for anomaly detection tasks and merits further optimization and investigation.
