Study On Fault Detection And Prediction Based On Multi-dimensional Time-frequency Feature Extraction

In the process of actual production, long-term fatigue wear of some mechanicalequipment or their parts will cause faults, and may have a big influence on normal operationof the whole system. It is very necessary to detect and predict faults of equipment and partstimely. In this paper, the wavelet time-frequency analysis was combined with statisticalanalysis. Multi-dimensional feature vetors of the observed sequence, which were composed ofwavelet Shannon entropy and wavelet packet bi-spectrum would be extracted to build theprediction model then. The main work included two aspects as below.The first was multi-dimensional features extraction. The faults detection and predictionof equipment or system were based on features extraction. Multi-dimensional featuresextraction could reflecte the features of signals from different aspects. Wavelet Shannonentropy and wavelet packet bi-spectrum were extracted. Simulation was completed with thesignals of rolling bearings in different working states, including normal state, outer-racer faultand rolling element fault. Simulation results have shown that the method of multi-dimensionalfeatures extraction could show features of signals roundly, so with the method, different kindsof faults could be recognized accurately.The second was faults detection and prediction. The prediction model of Least SquaresSupport Vector Machines (LS-SVM) was built, and then combined it with thresholdestimation which was based on the center of gravity entropy. According to the method, thefaults of sequences could be monitored on time. Simulation was completed with the toolmaterial KC5010on aviation engine. Simulation results have shown that prediction withLS-SVM model could get high prediction accuracy, and the combination of thresholddetection and fault prediction could find out faults in time.