Intelligent fault diagnosis for broken rotor bar using wavelet packet signature analysis

• Main Author: Zolfaghari, Sahar
• Format: Thesis
• Language: English
• Published: 2016
• Online Access: http://psasir.upm.edu.my/id/eprint/66162/1/FK%202016%2080%20IR.pdf; http://psasir.upm.edu.my/id/eprint/66162/
• Institution: Universiti Putra Malaysia

Induction motors are one of the extensively used machines in many industries due to their high reliability and simple structure. However, owing to the high stresses that happen during operation, induction motors are subjected to unavoidable failures. Among numerous inevitable burdens happening in different part of induction machines, rotor faults are considerable priority as they cause precipitate deterioration and, secondary failures that lead to an unexpected shutdown and result in time-consuming and expensive maintenance. Therefore, intelligent fault diagnosis of induction machines is an ongoing research topic because of the complexity of the issue as well as progress in signal processing. As a sensitive signal processing wavelet-based analysis is implemented and some difficulties like, lack of frequency localization, selection of best basis, and fault index are addressed in this study. Intelligent methods have concerted on sensing precise failure modes and recommending intelligent maintenance decisions based on the signatures collected through signal processing. Therefore, an advanced signal processing must be considered to derive the fault signature accomplish with a powerful decision-making technique. In this thesis, intelligent fault detection and severity classification of broken rotor bars in induction motor is carried out using the secondary data of stator current. The stator current was decomposed using wavelet packet decomposition. Then, the most precious faulty sub bands were identified after spectrum analysis. Next step to assist the most relevant feature extraction was the definition of mother wavelet function. In order to alleviate the time-variant characteristics of the wavelet packet transform coefficients, statistical parameters of wavelet packet coefficients are calculated. Some combinations of features extracted from wavelet packet signature analysis could design neural network trained, cross validated and tested input vector to not only elucidate the faultless from faulty condition, but also classify the number of broken rotor bars. As an effective signal processing, the time-scale characteristic of wavelet packet transform fused with the frequency resolution of Fast Fourier Transform named as wavelet packet signature analysis. This transformation technique is suitable for locating certain frequency components of a signal superimposed to fundamental frequency and associated with broken rotor bars. Then, the practically identical mother wavelet, db44, was selected based on standard deviation of wavelet packet coefficients. To make an intelligent decision without the presence of expert, in this research simple multi-layer perceptron NN-based fault classifier is proposed for fault diagnosis which is inexpensive, reliable, and non-invasive by employing best combination of wavelet statistical parameter after a simple feature selection technique as the input vector. The fault detection and classification algorithm is carried out under the unknown dataset and the off-line testing results with 98.8% classification accuracy indicate good reliability of the proposed method in identifying broken rotor bars severity.