吳順德Wu, Shuen-De辛承宣Hsin, Cheng-Hsuan2019-09-032019-02-272019-09-032019http://etds.lib.ntnu.edu.tw/cgi-bin/gs32/gsweb.cgi?o=dstdcdr&s=id=%22G060573017H%22.&%22.id.&http://rportal.lib.ntnu.edu.tw:80/handle/20.500.12235/97059在工具機組成中，迴轉機械扮演了非常重要的角色，包括軸承、齒輪與轉軸等等零件，而這些零組件往往承受極高的負載，隨著系統運轉會漸漸出現缺陷，進而導致整個系統故障停機，此時就必須付出高昂的硬體及人力成本進行維修，影響加工品質及預訂的出貨時程，造成生產者極大的困擾，所以開發一套故障診斷系統來即時監測並進一步量化損壞情形就是極為重要的課題。 在此之前，迴轉機械系統中的故障診斷技術大都聚焦於錯誤診斷與預知保養上，本研究希望利用葛蘭傑因果分析做為異常偵測之演算法，分析迴轉機械之振動訊號並找出異常的源頭，達到「錯誤定位」與「根本原因診斷」的功效，同時也期待透過因果性分析多個零件間的交互關係，找出異常發生的前兆，達到預知保養的功效。 本研究將以資料驅動(data-driven)的方式來設計異常診斷系統之架構，可分為以下三部分：1.訊號擷取；2.特徵抽取；3.利用機器學習演算法進行異常偵測及判斷。在訊號擷取的部分，初期將會先利用IMS中心提供之軸承資料庫進行初步測試，並在之後架設一多軸承實驗平台，量測更多的實驗資料來加以佐證。在特徵抽取演算法的部分，本研究將會設計一系列實驗來驗證葛蘭傑因果分析是否能夠找到判斷根本原因的有效特徵，並從時域、頻域及時頻域等角度呈現其分析結果，確定其做為根本原因診斷的可行性後，期待未來能進一步利用支援向量描述等機器學習演算法進行訓練，達到自動化偵測及判定的功效。Rotary machines composed of bearings, gears and shafts play an important role in the machine tools. These components are often subjected to high loading during operations. Defects are then initiated, propagated, developed and finally cause machine breakdown. The quality and the estimated delivery time will then be affected, causing higher expense in maintenance and labor cost. Therefore, developing a diagnostic system capable of detecting anomaly instantly is important for both academic and industry field. In previous studies, anomaly detection in rotary machines usually focuses on "fault diagnosis" and "prognosis". In this study, we suggest a new method to identify the "fault location" and "root cause" by applying Spectral Granger Causality to detect the anomaly of a multi-bearing system. By using this method, we may find out some prognostic feature before faults occur through the cause-effect relationship analysis. The design of system is based on data-driven structure, including：Data Acquisition, Feature Extraction and Machine Learning. In data acquisition, the bearing data provided by the Center of Intelligent Maintenance Systems (IMS) and self-constructed multi-bearing platform is used to verify the proposed algorithm. In feature extraction, we design a series of experiments to test whether Granger Causality in time, frequency or time-frequency domain can be taken as the feature of fault localization. The third part: Machine learning is not conducted in this study. In the future, machine learning algorithms such as Support Vector Data Description can use the feature extracted by granger causality to detect the fault location intelligently.迴轉機械異常偵測錯誤定位根本原因診斷葛蘭傑因果分析時頻域因果性分析rotary machineanomaly detectionfault localizationroot cause diagnosisgranger causalitytime-frequency causality analysis