Research On Pattern Recognition Of Mechanical Fault Based On Feature Relevance

In machinery equipment failure pattern recognition, there are certain kinds of internal or external relevance between characteristic features during the process of extracting features from samples. Usually, the traditional feature extraction and classification methods have weakened this relevance and regard the features as just numbers, without considering the deep meaning of the features relevance. In fact, extracting and analyzing the relevance can help determine the intrinsical model type, strengthen the clustering degree of the same class samples and eliminate the redundant features, which can contribute to prove the recognition rate.According to the relevance information of fault features, this paper analyzed the significance of relevance to equipment fault pattern recognition. Taking the spatial structure relevance and statistical relevance of features as a starting point, this paper proposed two mechanical fault pattern recognition methods based on features relevance.(1) Equipment fault pattern recognition method based on layered texture of the time-frequency spectrum and support tensor machine: this method can extract the spatial structure relevance of the texture figures from the time-frequency spectrum, and optimize the clustering performance. Meanwhile, texture matrix are used by the support tensor machine to do learning and prediction, without ignoring the inherent relationships between texture matrix features.(2) Equipment fault pattern recognition method based on recurrence quantification analysis and V-VPMCD(Voted-Variable Predictive Model Based Class Discrimination): this method takes the advantages of recurrence quantification analysis in analyzing nonlinear and non-stationary signal. It can provide high stability of performance when there's little quantity of samples and the quality is also poor. It takes statistical relevance between features as a basis for classification and optimizes the variable predictive model based class discrimination method by voting, which improving the stability and recognition rate of the traditional algorithm.In this paper, two recognition experiments have been carried out mainly about different kinds and levels of rolling bearing faults and unstable states of sliding bearing oil film. The results proved that the methods of this paper have higher recognition rate and better overall performance than traditional methods. Specifically, layered texture extraction combined with support tensor machine performs better than overall/layered texture extraction combined with SVM/BP, recurrence quantification analysis combined with V-VPMCD performs better than recurrence quantification analysis combined with VPMCD/SVM/BP.