| 摘要: | 隨著工業4.0的興起，製造業透過傳感器收集設備於生產流程上的資料，並利用這些資料建立模型及分析，促使他們可以朝向智慧工廠進行轉型。運用機器學習的方式分析收集數據達到預測設備故障或停機時間及原因，同時也希望透過制定有效的管理策略，來進行後續的設備維護。因此衍伸出預診斷與健康管理(PHM)。期望透過此模型使設備維護更有效率，為工廠提供更靈活、適應性更強的製造模式，不會因為設備臨時停機導致生產流程出現缺口。
本研究欲使用A公司所提供之塗佈機生產數據，運用預診斷與健康管理為架構改善該公司原有之維護策略。透過生成對抗網路(GAN)生成異常訊號，並於長短期記憶網路(LSTM)建置預測模型。依照異常徵兆分布狀況分別建置三種不同時間段模型比較模型間之績效，分別為5分鐘模型、15分鐘模型及25分鐘模型。最終選定15分鐘模型作為LSTM設備監控模型，判讀績效為準確率為97.56%、精確率為100%、召回率為93.33%及F_1-Score為96.55%。將其模型應用於塗佈機捲取B軸之運行狀態監控，LSTM監控模型在此次實驗中針對異常訊號會發出兩次異常預警報，第一次預警報提早約30分鐘發出，第二次預警報提早約1分鐘發出。因此，本研究針對這兩次預警報制訂一項判讀異常之規則，並制訂預診斷與健康管理維護策略於捲取B軸。
;With the emergence of Industry 4.0, the manufacturing industry has begun to collect data from machines in the production process through sensors. They use this data to transition towards smart factories, analyzing the collected data through machine learning to predict equipment failures or downtime and their causes. Simultaneously, there is a desire to develop effective management strategies for subsequent equipment maintenance. This has led to Prognostics and Health Management (PHM), aiming to make equipment maintenance more efficient. The goal is to provide manufacturers with a more flexible and adaptive manufacturing model while pursuing improved production quality, processes, and rapid responses to changes in production to meet customer demands. This approach ensures that temporary equipment shutdowns do not lead to gaps in the production process.
In this study, we aim to utilize production data from Company A′s coating machine and improve its existing maintenance strategy using the Prognostics and Health Management framework. We will apply Generative Adversarial Network (GAN) to generate abnormal signals and build a prediction model with Long Short-Term Memory (LSTM). According to the distribution of abnormal symptom, the research compares the performance of three different time-segment models, namely the 5-minute model, the 15-minute model, and the 25-minute model. This study selects the 15-minute model as the LSTM device monitoring model, achieving an accuracy rate of 97.56%, precision rate of 100%, recall rate of 93.33%, and F_1-Score of 96.55%.
Subsequently, this model is applied to monitor the operation status of the winder B in the coating machine, issuing two early warnings for anomalies. The first warning is issued approximately 30 minutes in advance, while the second warning is issued approximately 1 minute in advance. ased on these warnings, the research proposes a rule for judging anomalies and establishes a Prognostics and Health Management maintenance strategy for the winder B in the coating machine. |
