Vibration Monitoring And Fault Diagnosis Of High Voltage Circuit Breakers

Mechanical failures of high-voltage circuit breaker (HVCB) are frequently occurring on site of power system, which bring economic losses and impact on power system stability. Because vibration signals contain plentiful condition information and can be obtained invasively, fault diagnosis via vibration monitoring is a potent technique for mechanical condition estimation of HVCBs, which is competent to detect and prevent impending failures in advance. Though domestic and foreign scholars studied extensively in this subject, few significant breakthroughs were achieved. Vibration monitoring and fault diagnosis are still among the frontiers of mechanical condition maintenance of HVCBs. Effective feature extraction and accurate diagnoses are main difficulties in this field. It is essential to study intensively and comprehensively the vibration monitoring and fault diagnosis method of HVCBs.In this dissertation, vibration signals of HVCBs are acquired through simulated experiments. Time-frequency analysis scenario and denoising method are studied. The time-frequency characteristics of HVCB vibration signals are analyzed. Feature extraction approaches based on the time-frequency characteristics are proposed. And fault diagnosis scheme is investigated using support vector machine (SVM) to verify the performance of fault identification via the feature of vibration signals. Main researches and conclusions are as follows:1. A virtual instrumentation based experimental system is established. And vibration samples are collected via simulated experiments carried out on a12kV vacuum circuit breaker (VCB) and a126kV gas insulated switchgear (GIS).2. A time-frequency analysis scenario combining wavelet packet decomposition (WPD) with Hilbert transform (HT) is proposed via comparison time-frequency resolution between WPD and ensemble empirical mode decomposition (EEMD). The scenario provides precisely the instantaneous amplitude and the instantaneous frequency of a non-stationary signal. A novel wavelet packet denoising method associated soft threshold with hard threshold is presented according to the apriori knowledge of the noise as well. The new denoising method reduces noise effectively but retains abrupt information in the signal moderately.3. And the time-frequency distributions of vibration signals from distinct positions of one HVCB are compared, and the discrepancy between time-frequency characteristics of12kV HVCB and that of126kV GIS are discussed. The relationship between moving parts and the time zone of the vibration signal are investigated too. As a result, a time-frequency plane division measure is constructed based on marginal spectrum and instantaneous energy density level. This measure provides a basis for quantification of time-frequency characteristics, which takes into account both frequency characteristics and energy characteristics of the vibration.4. Grounded on the division of time-frequency plane, amplitude envelopes of the time-frequency subarea are obtained via WPD and HT and are utilized to constitute entropy vector of the vibration signal. This process affords a WPD time-frequency entropy approach to quantify the high dimensional time-frequency characteristics of a vibration signal into a low dimensional feature vector. To lessen the calculation cost, zero-phase digital filter (ZPDF) method is deployed as a substitute for WPD and formed a ZPDF time-frequency entropy approach for feature extraction. Meanwhile, a segment ZPDF is invented, which enhanced the transition process performance of the normal ZPDF. Comparisons of both types of feature vectors from normal condition and that from fault conditions including high friction in solenoid, loose flexible connection, defective insulating tie, and broken rod pin exhibit good repeatability in identical conditions but evident discrimination between different conditions.5. Employing WPD and ZPDF entropy vectors respectively as input samples, SVM are introduced for fault diagnosis. On condition that samples of multi fault patterns can be gain, SVM-based multi-class classifiers are applied. The structure type, the kernel function and the classification strategy of the optimal SVM are determined by cross validation (CV). The verification by independent test data gives a high total accuracy of84%via ZPDF entropy vector. The results indicate the feasibility of fault diagnosis scheme by entropy vector and multi-class SVM. In the case of adversity of samples of fault patterns, one-class SVM is experimented with. A total accuracy of80%via ZPDF entropy vector is achieved with lower accuracy in normal condition. One-class SVM provides a viable alternative for primitive research and sample accumulation of HVCBs vibration monitoring and fault diagnosis.