Modifying the Hilbert-Huang transform using the nonlinear entropy-based features for early fault detection of ball bearings

Abstract

Employing a powerful and adaptive method in time-frequency domain is one of the most important subjects in the nonlinear and non-stationary signal processing. The main objective of this paper is to modify Hilbert-Huang transform using the advantages of nonlinear entropy-based features in the time and frequency domain to reduce the noise effects. In addition, using appropriate entropy-based features will result in restriction the information redundancy and overcoming the need for dimension reduction, for the fault detection of a ball bearing system. To modify the Hilbert-Huang method, the effect of added noise on various types of nonlinear entropy-based features is investigated for each Intrinsic Mode Functions (IMFs) which is extracted by ensemble empirical mode decomposition algorithm. Considering the Approximate Entropy (ApEn) sensitivity to noise, an evaluation index is presented for selecting the proper amplitude of the added noise based on the ApEn and mutual information coefficient of the different IMFs. Subsequently, taking into account of the high capability of permutation entropy (PeEn), ApEn and Marginal Hilbert spectrum Entropy (MHE) in the signal characteristic, a threshold is determined for fault detection based on their values associating with the main IMF which has the highest value of mutual information coefficient. As a result, the entropy-based features related to the main IMF can be used for detection of any deviation from normal operation of ball bearings, regardless of the fault type. Finally, the results show the best rate of classification among the twelve experimental conditions involving normal and various type and severity of bearing faults which is achieved by employing Multi-Class Support Vector Machines (MCSVM) combined with PeEn and MHE values of the first tree IMFs as a composite input vector.