On-Load Tap-Changer Mechanical Fault Diagnosis Method Based on CEEMDAN Sample Entropy and Improved Ensemble Probabilistic Neural Network

Abstract

The vibration signals of on-load tap-changer (OLTC) contain a rich of operating status information and will effectively diagnose the mechanical fault of OLTC. For the purpose of improving the level of OLTC diagnosis in mechanical condition, this study used the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) sample entropy (SampEn) combined with K-L divergence as a vibration signal. Meanwhile, an improved ensemble probabilistic neural network was used for mechanical condition. Then the OLTC mechanical vibration signals under different conditions were measured by experiments. The original vibration signals were decomposed into IMF components with different frequency distributions by CEEMDAN, and then calculate K-L divergence between them. Next, calculate the sample entropy of selected IMF component as the vibration signal feature vector. At the same time, construct a probabilistic neural network (PNN) and optimize the smooth factor. Then the optimized PNN and other weak classifiers were combined as the base classifier of bootstrap aggregating (bagging) algorithm, which greatly improves the classification accuracy of PNN. The final experimental results prove that the improved model can exhibit a high diagnostic efficiency and accuracy rate, which can effectively extract mechanical characteristics and generate some meaningful help for the research of other mechanical fault diagnosis.