Feature Representation Learning of Vibration-based Anomaly Detection for Rotation Machinery Condition Monitoring

Abstract

The present paper describes neural network (NN)/Gaussian mixture model (GMM) tandem connectionist anomaly detection to develop robust condition monitoring systems against environmental changes. The key to the success in generalizing anomaly detection systems is robust feature representation learning by effectively using normal-state and faulty-state data collected from non-target monitoring machines. Experimental comparisons conducted using vibration signals from actual wind turbine components demonstrated that NN/GMM tandem system developed using faulty-state data from non-target machines yielded significant improvements over the existing system, and that NN/GMM system developed using only normal-state data from target and non-target machines also performed better than the existing system.