A Fast Adaptive S-Transform for Complex Quality Disturbance Feature Extraction

Abstract

This paper proposes a fast adaptive S-transform (FAST) to improve the time-frequency resolution and computational efficiency of power quality disturbances (PQDs) feature extraction. By directly controlling the standard deviation instead of other parameters, FAST can reduce the difficulty of optimizing time-frequency resolution. Based on the frequency spectrum of PQD signals, FAST only needs to calculate characteristic frequency points determined by maximum envelope curve, which can eliminate redundant calculation without losing effective feature information. In fact, the computational complexity of parameter optimization step is often higher than that of S-transform (ST) calculation step. To address this problem, a window matching spectrum (WMS) method is proposed to optimize the time-frequency resolution. Matching the effective window width with the main spectrum energy interval of signals, WMS determines the standard deviation without iterative calculation. Based on the time-frequency representation of FAST, four features are extracted as the feature vectors and applied to the support vector machine (SVM), probabilistic neural network (PNN), extreme learning machine (ELM), convolutional neural network (CNN), decision tree (DT-C4.5) and random forest (RF) classifiers. Classification results of the six classifiers show that FAST has better time-frequency resolution and lower computational complexity than that of generalized S-transform (GST) and ST. In addition, the FAST-ELM method has stronger noise immunity and better performance than other combination methods with the simulation signals and experimental signals.