Mechanical fault diagnosis model based on feature evaluation and neural networks

To overcome blindness of subjective selecting dimensionless indicators as sensitive features without any experience, a novel mechanical fault diagnosis model based on feature evaluation and radial basis function (RBF) networks is proposed, where the original signals are decomposed via wavelet packet and empirical mode decomposition (EMD) respectively, and the dimensionless indicators in time domain are extracted from the original signals and each decomposed signal to construct the combined features. Furthermore, a feature evaluation method is applied to calculate evaluation factors of the combined features, and the corresponding sensitive features are selected according to the evaluation factors and input into the RBF networks to automatically identify different conditions of mechanical equipment. The experiments of rolling bearings fault diagnosis are carried out to test the performance of this model. The results demonstrate that the model integrating wavelet packet, EMD, feature evaluation method and RBF networks enables to precisely extract fault information, and select sensitive ones from a large number of features to correctly and rapidly diagnose the mechanical faults. This model is also employed to classify heavy oil catalytic cracking set under 3 conditions. The results show it can reduce the networks scale, increase the classification accuracy, and enhance the robustness.