An Image Processing Approach to Machine Fault Diagnosis Based on Visual Words Representation

Machine fault diagnosis and remaining service life prognosis provide the basis for condition-based maintenance, and is key to operational reliability. Accurate assessment of machine health requires effective analysis of vibration data, which is typically performed by examining the change in frequency components. One limitation associated with these methods is the empirical knowledge required for fault feature selection. This paper presents an image processing approach to automatically extract features from vibration signal, based on visual words representation. Specifically, a time-frequency image of vibration signal is obtained through wavelet transform, which is then used to extract "visual word" features for recognizing fault related patterns. The extracted features are subsequently fed into sparse representation-based classifier for classification. Evaluation using experimental bearing data confirmed the effectiveness of the developed method with a classification accuracy of 99.7%.