Considering the non-linearity existing in bearing vibration signals as well as the scarcity of fault samples, this paper presents a method for bearing health condition identification based on improved multi-scale entropy (IMSE) and support vector machines (SVMs). IMSE refers to the calculation of improved sample entropies (i.e. fuzzy sample entropies across a sequence of scales). Applying IMSE to mechanical vibration signals can take into account not only the non-linearity but also the interactions and coupling between mechanical components, thus providing much more information regarding the machine health condition than traditional single-scale entropy can be expected to. In engineering practice, the amount of fault samples is often limited for training a classifier, which thus decreases the performance of traditional classifiers like artificial neural networks (ANNs). SVMs are derived from statistical learning theory, which is different from the conventional statistical theory on which ANNs are based. SVMs provide a favourable solution to small sample-sized problems. In this study, IMSE and SVMs are employed as fault feature extractor and classifier, respectively. The experimental results verify that the proposed method has potential applications in bearing health condition identification.

The current paper presents a novel scheme for bearing fault diagnosis based on lifting wavelet packet transform (LWPT), sample entropy (SampEn), support vector machines (SVMs), and genetic algorithms (GAs). In the proposed scheme, bearing vibration signals were first decomposed into different frequency sub-bands through a four-level LWPT, resulting in a total of 31 node signal components throughout all layers of the LWPT decomposition tree. The SampEns of all 31 components were then calculated as an original feature pool to characterize the complexity of the bearing vibration signals within the corresponding frequency bands. For selecting the most informative features thus reducing the number of features, a GA was applied to simultaneously pick out the salient features and optimize the parameters of SVMs so as to avoid the curse of dimensionality, alleviate computational burden, and improve the subsequent classification. Experiments were conducted on an induction motor with respect to various bearing faults and a range of fault severities. As an example here, three salient features were selected from the original 31 features, with the dimension of features reduced dramatically. The selected three features were then presented into the optimized SVM to identify various bearing conditions. The scheme was compared with the widely used WPT-Energy method and with a neural network classifier with respect to feature extractions and fault classifications, respectively. The results are in favour of the proposed scheme in terms of the feature dimension, number of support vectors, robustness to the data partition, and classification rate.