Performance Comparisons Between MFDFA and EMD Using Neural Network and Support Vector Machine

Abstract

Multifractal detrended fluctuation analysis (MFDFA) is a powerful tool for discovering dynamics hidden in complex data. Similarly, empirical mode decomposition (EMD) is a typical method for time-frequency analysis of nonstationary data. Many works have applied MFDFA and EMD to get insight into nature of machine vibration signals. However, comparisons of performance of MFDFA and EMD in feature extraction have rarely been performed so far. To fill this gap, this paper benchmarked the performance of MFDFA against EMD by neural network (NN) and support vector machine (SVM) using a group of gearbox vibration signals containing gear faults in different types and severity levels. In this way, five characteristic parameters were obtained using MFDFA. For comparison, an eleven-dimension feature vector was acquired using EMD. Following this, either of NN and SVM served to distinguish between gearbox conditions characterized by the features extracted above. The results indicated that MFDFA is comparable to EMD in discrimination between gearbox conditions. In addition, this paper demonstrated that the integration of MFDFA and SVM is promising in fault diagnosis of gearboxes.