Induction Motor Fault Diagnosis Using Second Generation Wavelet And Ensemble Classifiers

As one of the most important parts manufaturing systems, induction motors have been widely used in amonst all industries. However, faults can occure at various parts in the induction motor, which usually lead to severe consequences such as performance deterioration, costly downtime and production delay. In order to satisfy the increasingly high demand for productivity, quantity and production flexibility, condition monitoring and fault diagnosis of the induction morot has been paid much more attention than ever before. In recent years, the rapid development of the nonstationary signal processing and information fusion technologies has opened up potential opportunities for induction motor fault diagnosis.The main research tasks of this thesis are as follows:1) The shortcomings of the original second generation wavelet transform (SGWT) is analysed, and a space-adaptive redundant SGWT (ARSGWT) is proposed for analyzing signals measured from induction motors. Different from the original SGWT, in ARSGWT the split operation is skiped to eliminate frequency aliasing. In addition, different prediction operators and the update operators are chosen for each sample point in the signal. The failure information hidden in the signals is then effectively extracted with ARSGWT.2) The proposed ARSGWT is applied to extracting energy features at some specific frequency components from both the stator current signal and the vibration signals. These features are collected to form a high-dimensional feature vector. Then an effective and practical multiple class feature selection (MCFS) approach is introduced to select representative ones from the feature vector and used for induction motor fault pattern recognition.3) Information fusion technology is introduced into induction motor fault diagnosis. Based on decision fusion, an adaptive weighted voting multiple classifier fusion algorithm is proposed. The features extracted from different signals are used as input to the corresponding fault pattern recognition classifier. Then clustering analysis is applied to both the test samples and training samples to determine the weight of each classifier for final diagnosis result.4) Experimental studies performed on the induction motor test system are designed to verify the effectiveness of the proposed signal processing method. The MCFS is proved to be capable of selecting the best feature combination as compared to some other popular feature selection methods based on manifold learning. The presented classifier fusion algorithm is also verified to be able to improve the diagnosis accuracy.