Fault Diagnosis of Rolling Bearings Based on SURF algorithm

This paper proposed a new method for fault diagnosis of rolling bearings based on SURF (Speeded-Up Robust Features) algorithm, where two-dimension signal is used. Different from other classical 1-d signal processed methods, the proposed method transforms the 1-dimensional vibration signals into images, then image processed methods are utilized to analyze the image signal so as to reach the goal of faulty classification. Images transformed from vibration signals often have special texture features and each faulty category’s texture varies. SURF is a computer vision algorithm improved from SIFT (Scale Invariant Feature Transform) algorithm, and it can more efficiently extract local features through the texture of the image. Firstly, normalized time domain vibration signals were converted into gray-scale images. Then the mean filter was employed to complete the image pre-processing. Secondly, local features were extracted from the images by using SURF algorithm. Through the mean-shift clustering algorithm, extracted features were clustered to form a texture dictionary. Finally, features extracted from testing signals were compared with the texture dictionary to determine the corresponding faulty category. To validate the proposed method, several comparative experiments between SURF and SIFT-based algorithm have been carried out. The experimental results indicate that the proposed method outperforms the existing SIFT descriptor in terms of classification accuracy and computation cost for bearing fault diagnosis. Key-Words: fault diagnosis, rolling bearing, SURF, feature vector, mean-shift clustering, computation cost