An Unsupervised Double-Layer DBN Based Intelligent Diagnosis Method for Bearing Faults
The operating environment of large rolling bearing equipment is complex and variable,and previous di-agnostic methods established using pattern recognition are usually difficult to effectively solve problems such as data containing noise,incompleteness and lack of labels.Therefore,an unsupervised double-layer deep belief network(DBN)method for intelligent classification and diagnosis of rolling bearing faults is proposed.The method makes use of layer-by-layer greedy learning of DBN to mine the feature information related to faults and input to the classifier.Outliers in the unknown data are identified by an adaptive fuzzy C-mean clustering algorithm.If the density of outliers is low,they are judged to be noisy and this is used to eliminate noise interference in the classification process;If the density of outliers is high,they are judged to be a new class and are merged into the fault knowledge base.The Bayes-ian classifier method is then applied to the secondary DBN network to enable unsupervised learning of fault damage classes.This method is validated using data from the rolling bearing experimental platform at Western Reserve Univer-sity,and the conclusions show that the accurate classification of fault types and damage classes can be accomplished well with some intelligence in the presence of noise and incomplete data.
Deep belief networkRolling bearingIncomplete dataBayesian classifier
NETL
NSTL
万方数据
