DC 欄位 |
值 |
語言 |
DC.contributor | 工業管理研究所 | zh_TW |
DC.creator | 郭倍豪 | zh_TW |
DC.creator | Bei-Hao Kuo | en_US |
dc.date.accessioned | 2020-7-3T07:39:07Z | |
dc.date.available | 2020-7-3T07:39:07Z | |
dc.date.issued | 2020 | |
dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=107426023 | |
dc.contributor.department | 工業管理研究所 | zh_TW |
DC.description | 國立中央大學 | zh_TW |
DC.description | National Central University | en_US |
dc.description.abstract | 因應軟硬體技術的發展,全世界的工業逐步利用物聯網、Big Data、Machine Learning等技術,從原先工業 3.0 的「自動化」邁向工業 4.0 的「智動化」,在新興的技術導入後,企業在實際生產中將有助於維持設備的高稼動率以及提升製程的穩定性。
本研究以探討工業 4.0中的預診斷與健康管理(Prognostics and Health Management, PHM)研究,提出一項基於長短期記憶神經網路(Long Short-Term Memory Neural Network, LSTM)的異常偵測方法對設備進行即時的診斷,從時間序列資料進行分析,首先透過LSTM中的記憶單元儲存過去輸入的健康狀態生產行為,在連續的時間點,預測輪軸特徵的數據走向,並透過比對預測值及實際值所產生的誤差,根據其機率密度函數設定異常閾值(Threshold)。日後在生產過程中,一旦誤差連續超出所設定之閾值界線,則進而預警設備即將出現故障,在設備因故停止運作前,提早進行機器的維護作業,以降低不預期的損壞所造成的突發損失,實現預測性維護。
本研究將所提方法應用於A企業中的塗佈機實際生產時所回饋的生產資料集,透過調整參數配置之間預測誤差的變化,找出最佳化的模型配置,之後利用最佳化模型於測試資料集中關注最早的預警時間以及預警的效能指標,實證結果闡明,基於LSTM的異常偵測方法在前後者皆有穩健的表現。 | zh_TW |
dc.description.abstract | In response to the development of software and hardware technologies, industries around the world are gradually using advanced technologies such as the Internet of Things, Big Data, and Machine Learning to achieve not only the "Automation" but also the "Intelligence" in the Industry 4.0 era. In the actual production, new technologies will help maintain a high utilization rate of equipment and improve the stability of manufacturing,
In this research, I explore the Prognostics and Health Management (PHM) study in Industry 4.0 and propose an anomaly detection method based on a Long Short-Term Memory Neural Network (LSTM) approach on time series data for assessing the equipment in real time. First, the memory unit in LSTM would be used to store past input health-state production behavior. Through continuous time points, algorithm will predict the data trend of the pattern. Then the error will be generated by comparing the predicted value with the actual value. Finally, setting the abnormal threshold based on its probability density function. In the future, once the error continuously exceeds the threshold in the production process, the user will receive the warning which the equipment is about to be malfunction. By assessing healthy state of the equipment in real time via anomaly detection, we can arrange early maintenance or replacement of parts before the equipment shutting down accidentally to achieve predictive maintenance.
The proposed method is applied to the production dataset feedbacked from the actual production of the coating machine from company A. First, finding the optimized model configuration by observing the difference of prediction error caused by different combination of model parameters. Second, using the optimized model investigates the earliest warning time and the indicators of predicted performance in the test dataset. Finally, the experimental results show that anomaly detection based on LSTM has robust performance on both earliest warning time and predicted performance. | en_US |
DC.subject | 智慧製造 | zh_TW |
DC.subject | 人工智慧 | zh_TW |
DC.subject | 預診斷與健康管理 | zh_TW |
DC.subject | 深度學習 | zh_TW |
DC.subject | 長短期記憶神經網路 | zh_TW |
DC.subject | 時間序列 | zh_TW |
DC.subject | Smart Manufacturing | en_US |
DC.subject | Artificial Intelligence | en_US |
DC.subject | Prognostics and Health Management | en_US |
DC.subject | Deep Learning | en_US |
DC.subject | Long Short-Term Memory Neural Network | en_US |
DC.subject | Time Series | en_US |
DC.title | 基於長短期記憶神經網路方法於預測性維護之研究—以塗佈機為例 | zh_TW |
dc.language.iso | zh-TW | zh-TW |
DC.title | Predictive Maintenance Based on A Long Short-Term Memory Neural Network Approach - A Case Study of Coating Machine | en_US |
DC.type | 博碩士論文 | zh_TW |
DC.type | thesis | en_US |
DC.publisher | National Central University | en_US |