An Improved Algorithm for Selecting IMF Components in Ensemble Empirical Mode Decomposition for Domain of Rub-Impact Fault Diagnosis

Alexander E Prosvirin,Jong-Myon Kim,M M Manjurul Islam

doi:10.1109/access.2019.2938367

Alexander E Prosvirin, Jong-Myon Kim + Show 1 more

Open Access

https://doi.org/10.1109/access.2019.2938367

Copy DOI

Abstract

A rubbing fault is a complex non-linear and non-stationary fault that frequently occurs in rotating machinery such as turbines. One of the most frequently applied signal processing techniques for the analysis of rub-impact faults in rotating machines is ensemble empirical mode decomposition (EEMD). Despite the advantages of using EEMD in analyzing non-linear and non-stationary signals, it is crucial to determine which of the extracted intrinsic mode functions (IMFs) carry the most valuable and significant information about the mechanical faults under investigation. In this paper, an improvement in the IMF selection technique is introduced, which is based on the recent ratio of degree-of-presence (DPR) to the Kullback-Leibler divergence (DPR/KLdiv). The number of selected IMFs in the DPR/KLdiv-based technique is subjective with a constant threshold, whereas we apply an adaptive thresholding technique to select the most meaningful IMFs that are relevant to a rubbing fault. The experimental results demonstrate that the proposed enhanced IMF selection algorithm allows for better signal denoising properties than the original technique while preserving significant features evidencing the presence of rubbing faults in rotating machinery.

Highlights

Turbines are one of the most critical rotating machines that are widely used in power plants
FUNCTION NORMALIZATION AND ADAPTIVE THRESHOLDING FOR IMPROVING THE intrinsic mode functions (IMFs) SELECTION PROCEDURE To reduce the influence of the growth of average DPR values on the objective function computation and to cope with the problems caused by the employment of the arbitrary threshold level for IMF selection, this paper introduces a modification of the original DPR/Kullback-Leibler divergence (KLdiv)-based approach
The results presented demonstrate that the proposed DPR/KLdiv-based IMF selection technique is capable of reducing the presence of high-frequency noise in the partially reconstructed signal while preserving the valuable information content, which consists of frequency harmonics that are considered evident features of rub-impact faults [27]–[29]

Summary

INTRODUCTION

Turbines are one of the most critical rotating machines that are widely used in power plants. A specific criterion was derived for selecting valuable IMFs that is based on the ratio between the ‘degree-of-presence’ (DPR) of information related to rubbing faults (frequency harmonics) and the Kullback-Leibler divergence (KLdiv) [26] information-based distance metric This approach demonstrates good capabilities in selecting informative IMF components for rub-impact fault diagnosis; the applied arbitrary thresholding level (e.g., 1) for creating a subset of valuable modes is a weak point of this methodology. As the intensity of rubbing increases, it becomes easier for noisy or non-informative IMF components to exceed the arbitrary threshold level and become a part of a subset containing the selected modes Another issue may appear while computing the objective function value itself.

BACKGROUND

EXPERIMENTAL RESULTS AND DISCUSSION

FAULT DIAGNOSIS PERFORMANCE ANALYSIS

CONCLUSION