Bearing-fault diagnosis using non-local means algorithm and empirical mode decomposition-based feature extraction and two-stage feature selection

Bearing-fault-diagnosis problem can be conceived as a pattern recognition problem, which includes three main phases: feature extraction, feature selection and feature classification. Thus, to improve the performance of the whole bearing-fault-diagnosis system, the performance of each phase must be improved. The aim of this study is threefold. First, in the feature extraction step, a new feature extraction technique based on non-local-means de-noising and empirical mode decomposition is developed to more accurately obtain fault-characteristic information. Second, in the feature selection phase, a novel two-stage feature selection, hybrid distance evaluation technique (DET)–particle swarm optimisation (PSO), is proposed by combining DET and PSO to select the superior combining feature subset that discriminates well among classes. Third, in the classification phase, a comparison among three types of popular classifiers: K-nearest neighbours, probabilistic neural network and support-vector machine is done to figure out the sensitivity of each classifier corresponding to the irrelevant and redundant features and the curse of dimensionality; then, find out a most suitable classifier incorporating with feature selection phase. The experimental results for the vibration signal of the bearing are shown to verify the effectiveness of the proposed fault-diagnosis scheme.