Chapter 7 Conclusions and Future Work
7.1 Conclusions
The IMs are commonly used in various industrial facilities, and a reliable condition monitoring system can be used to diagnose the IM fault at its early stage, so as to prevent malfunction of the driven machinery, and to improve productivity. Several IM fault detection techniques have been proposed in the literature for IM health condition monitoring, but each has its own merits and limitations. It still remains a challenging task to accurately recognize the IM fault due to reasons such as insignificant fault features under some load conditions. An intelligent monitoring system was developed in the thesis to provide more reliable IM fault diagnostics. The developed intelligent monitor consists of two modules: feature extraction and decision-making. Feature extraction is a process to extract fault-related representative features from stator current signals. The decision- making module consists of pattern classification, system state prediction and the fuzzy inference system. The diagnostic information is retrieved by mapping fault features to IM health condition categories using the pattern classification technique. The prognostic information is provided by estimating the future states of fault features (indices) using the system state prediction technique. Finally, the fuzzy inference system is utilized to integrate the diagnostic information and prognostic information, in order to provide a more accurate assessment of the IM health conditions.
A novel fault detection technique, the spectrum synch technique, was proposed in Chapter 2 to extract representative features associated with the early IM defects from the IM stator current spectrum. The local bands of the IM fault frequency components are synchronized to form a fault information spectrum, in which the fault features are enhanced and the unrelated high amplitude frequency components are mitigated. A central kurtosis method is proposed to effectively extract useful information from the fault information spectrum to generate an IM fault index. The effectiveness of the proposed spectrum synch technique was demonstrated through experimental tests on IM with broken rotor bars and IM with pitted bearing outer race. The experiments were conducted under different operating conditions (i.e., different supply frequencies and different load conditions). Test results showed the superiority of the proposed spectrum synch technique over the commonly used techniques in the frequency domain, the power spectral density, and the envelope analysis.
A new pattern classification technique, the selective boosting classifier, was developed in Chapter 3 to categorize the fault indices from selected fault detection techniques in order to diagnose the IM fault. The traditional boosting techniques suffer from the overfitting problem, which can degrade the classification accuracy. The proposed selective boosting, sBoost, classifier can adaptively process noisy data based on the noise level of each sample, so as to enhance the classification performance. An error correction mechanism was also employed to further improve the classification accuracy by examining the class label distribution in the neighborhood of each sample. The effectiveness of the sBoost classifier was verified by using 12 benchmark examples from the literature. Test results showed the proposed sBoost classifier to be an effective classification tool. It could categorize the patterns with different characteristics effectively and accurately.
A novel system state prediction technique, pBoost predictor, was proposed in Chapter 4 to forecast the future states of the IM health conditions. A base learner is adaptively incorporated into the ensemble at each step to improve the performance of the ensemble predictor, and the resultant ensemble predictor outperforms all the base learners. Each base learner addresses a particular data distribution, which is updated based on the performance of the ensemble at the previous step. If the base learner is relatively strong or has relatively good performance (e.g., AR predictor), the ensemble predictor is prone to suffering from the overfitting problem. An advanced sample weight regulation mechanism was suggested to reduce the overfitting problem. The effectiveness of the proposed pBoost predictor was verified by simulation tests and a real-world application. The test results showed that the pBoost predictor was an effective tool to conduct system state prediction. The prognostic information can be used to further improve the reliability of the IM health condition monitoring.
A knowledge-based evolving fuzzy neural network (i.e., eFNN) predictor was developed in Chapter 5 to predict future state of the system. In the proposed eFNN, the linear properties of the data are modeled by a vector autoregressive-moving-average (i.e., VARMA) system and the nonlinear properties of the data are characterized by evolving NN system. A novel clustering technique was proposed to adjust the structure of evolving NN system, in order to capture the characteristics of the input data more accurately. Since the linear properties of the data are filtered out by VARMA, less structured information exists in the remaining data. Thus, fewer clusters are generated by the evolving
NN system, and the simplified structure of the predictor can further facilitate reasoning and training operations.
An integrated monitoring system was developed in Chapter 6 to synthesize the information from both sBoost classifier and pBoost predictor for IM health condition monitoring. In the proposed integrated monitor, a confidence-rate-based reasoning mechanism is proposed to address uncertain IM fault diagnostic decisions, in order to improve the accuracy of fault diagnosis. The effectiveness of the developed integrated monitor was verified by experimental tests corresponding to the common IM faults (e.g., IM broken rotor bar fault and IM bearing outer race defects) under different load conditions (i.e., light-load, medium-load and heavy-load). The test results demonstrated that the proposed integrated monitoring system is a reliable IM fault diagnosis tool, and it can provide a more accurate assessment of IM health conditions.