Classification of Rigid Rotor Faults Using Time Domain Features Extracted from Multiple Vibration Sensors

Abstract

Rotating machinery is the most common in industry. To avoid financial losses and catastrophic failures, accurate identification of rotor faults is crucial. Common rotor faults include unbalance (UF), misalignment (MF), bent (BF), eccentricity (EF) and cocked rotor (CF).  Each fault is addressed through distinctive maintenance technique, and thus inaccurate identification of these faults may introduce additional problems in the machinery.  Vibration-based predictive maintenance is very effective method to monitor the condition of machinery. Problem arises when traditional vibration analysis methods do not provide clear picture of the rotor faults. To address the issue, this research presents a fault diagnostic model, which employ supervised learning-based pattern recognition (PR) method using time domain statistical features (TD).  The TD features are extracted from vibration signals acquired from multiple accelerometers to capture radial and axial vibrations simultaneously. Difference of mechanical forces, exhibited by these faults on the multiple axes, provides very informative fault related TD features. Salient features are selected with the help of Decision tree (DT) to be utilized by Support Vector Machine (SVM). The proposed model provides very accurate classification of the faults. Moreover, the presented model identifies maximum number of rotor faults reported so far.  The model provides classification accuracy of 98%, and outperforms the previously presented methods for the problem at hand.