Web-based Fault Composed Diagnostics Platform -A Bearing Case

[[abstract]]當診斷切削加工機台是否異常時，可採用多種感測來源與其分析方法，諸如振動、噪音、或溫度等來源以分析之。隨著分析診斷技術的發展，包含時域、RMS、或頻譜等的振動分析方法已廣為周知。但實仍缺乏一整合平台，以評估各診斷方法複合的效率性與有效性；且目前診斷工具亦多限於本機使用，缺少跨網路的遠端分析能力。 本研究開發出一網路化診斷平台系統，允許多使用者於遠端瀏覽器上傳其各種感測資料於資料庫，在架構彈性上，本平台可支援不同錯誤感測來源的彙整，進行不同特徵的匹配，分類方法的選取，與診斷參數的設定等。在方法多樣性上，於後端伺服器提供多種特徵萃取法，如時域、包絡譜、與小波轉換等，以處理感測資料為特徵向量作為後續診斷的輸入。於錯誤分類中，診斷技術亦提供多種類神經網路模型，如倒傳遞類神經、機率類神經、與徑向基類神經等類神經網路診斷模型，以分類其錯誤模式。各特徵的匹配，方法的選取，與參數的設定，均可為自訂。最後，於診斷呈現上，本系統提供各萃取方法特徵與診斷模型的精度比對分析，以方便評估診斷的有效性。 以機台軸承破壞模式診斷為例，當使用者上傳多次多種軸承破壞模式之振動資料後，包含軸承內軌、外軌、與滾珠等單一與複合模式，經不同特徵萃取、差異比對與模型訓練等，測試得知透過少數特定時域與小波特徵，配合倒傳遞類神經網路模型，可觀察到上述組合可獲得具最佳診斷效果。[[abstract]]Sensing sources such as vibration acoustics and temperature are adopted to diagnose equipment abnormality Followed the diagnostics development including time-domain RMS and frequency-domain the vibration analysis has been accepted as a popular diagnostics method However there lacked an integrated platform to assist users for evaluating the efficiency and effectiveness of various diagnostics and most of diagnostic tools also are limited in local side and are failed in remotely analysis This study developed a web-based diagnostics platform system which allows multi-users upload various sensing data to a database from browser and provides feature extraction methods e g time-domain envelope and wavelet to process the sensing data as feature vectors as the diagnostic inputs The system presents three neural network models Back-Propagation (BP) Probability (PNN) and Radius Basis Function (RBF) to classify the fault modes The feature mapping method selection and parameter setting can be defined by users Finally this system allows user to evaluate the diagnostics effectiveness by comparing the precisions of using various methods and models As a bearing failure modes case the different bearing failures including single and compound modes inner-race outer-race and rolling ball are studied After uploaded the vibration data extracted different features compared feature distances and trained models the system allows users to find that the best diagnostics combinations is the back-propagation neural-network with the specific time-domain and wavelet feature