旋轉機械線上監控與異常辨識系統

Abstract

軸承是旋轉機械中很重要的關鍵零組件，當軸承出現異常時往往會使機台出現無法預期的損壞，因此過去十年來有許多學界與業界的相關人員投入軸承異常辨識的相關研究。本論文提出一個完整的軸承異常辨識流程，能準確區分不同型態的軸承錯誤，此流程又可分為三個部分，首先是提出數種不同計算亂度的方法從機台的振動訊號中抽出相關的特徵，其次是透過特徵選取的流程選出最有用的特徵，最後再以選出的特徵及一對一支持向量機建立辨識模型。 本論文主要的貢獻在於： 1. 我們介紹數種以計算亂度為基礎的特徵抽取演算法，包括多尺度熵、多尺度排序熵、多頻帶頻譜熵，此外，我們還提出組合多尺度分析方法，可以有效提升多尺度分析的性能。 2. 我們介紹了Fisher score、Mahalanobis distance兩種特徵選取方法，透過本論文設計的流程可以選出最佳化的特徵，實驗結果顯示，以最佳化的特徵所訓練的模型，辨識能力會遠遠高於未使用的結果。 本論文實驗的資料使用Case Western Reserve University (CWRU)的軸承振動訊號，實驗中設計了19種不同的情況用來驗證系統的辨識能力。實驗的結果顯示，本論文所提出的辨識流程對於辨識軸承異常的種類有非常高的辨識率。Bearings are the most frequently used component in a rotary machine. Bearing failure could lead to unpredictable productivity loss for production facilities. Therefore, the fault diagnosis of bearing has attracted significant attention from the research and engineering community over the past decades. In this dissertation, a bearing fault diagnosis algorithm was proposed to identify the types of bearing fault. This proposed algorithm is decomposed into three key phases. Firstly, the defect-related features were extracted from the vibrational signal by using several entropy-based algorithms. Secondly, the optimal feature set is obtained by the feature-selection algorithm. Finally, a classifier model is trained by using the optimal feature set and the one-against-one support vector machine. The main contributions of this dissertation can be summarized as follows: 1. We introduce several entropy-based algorithms including, multiscale entropy (MSE), multiscale permutation entropy (MPE) and multiband spectral entropy (MBSE) to extract defect-related feature hidden in the measured signal. Furthermore, a composite algorithm for MSE and MPE, named composite MSE (CMSE) and composite MPE (CMPE) were proposed to improve the performance of the feature extraction. 2. We introduce two feature selection algorithms including, Fisher score and Mahalanobis distance to select the sensitive features from the original feature set. An optimal feature set can be obtained by using the propose feature selection algorithm. Experiments show classifier model trained by optimal feature set is the most effective for the fault diagnosis of bearings.

In the simulations presented in this dissertation, the vibration signal datasets of bearing from Case Western Reserve University (CWRU) are utilized. Nineteen experiments are designed to evaluate the capability of the proposed fault diagnosis system. Experimental results demonstrate that the proposed system provides a significantly higher accuracy of prediction for the classification of bearing faults.