A knowledge-based fault diagnosis method for rolling bearings without fault sample training

Rolling bearings are indispensable components of many engineering machinery, especially rotating machinery. If rolling bearing faults are not diagnosed promptly, it may cause huge economic losses. Bearing fault diagnosis can avoid catastrophic accidents, ensure the reliability of equipment operation, and reduce maintenance costs. Existing intelligent bearing fault diagnosis methods have fast diagnosis speeds and excellent fault recognition capabilities, which is not feasible for most important mechanical devices because of the difficulty in obtaining fault samples for training. To tackle this problem, a two-stage bearing fault diagnosis method without fault sample training based on fault feature knowledge is proposed. In the first stage, a fault detection vector is constructed based on signal statistical indicators. The Mahalanobis distance of the feature vector between online signals and historical normal signals serves for anomaly detection. In the second stage, based on the bearing fault knowledge, envelope spectrum fault indicators are proposed to form diagnosis vectors. By calculating the similarity between the diagnosis vector and the present fault label, the probability of different fault types will be obtained. Three experimental analyses show that the method is effective in detecting early faults and achieves high fault identification accuracy. The above results advantageously prove that the method can be used for fault diagnosis without fault sample training, and has the possibility of practical application.