An orbit-based encoder–forecaster deep learning method for condition monitoring of large turbomachines

Abstract

Most data-driven methods used in condition monitoring and early faults alarming of large turbomachines need to manually extract features for modeling, which inevitably introduces subjectivity and uncertainties. In addition, conventional deep learning methods are able to automatically classify fault types using shaft orbit images constructed from the vibration data, but image-based deep learning methods have yet been investigated for the long-term condition monitoring of a turbomachine. This study presented a novel image-based encoder–forecaster deep learning model, OrbitLSTM, to address these issues by modeling the dynamic behaviors of orbit image sequences for condition monitoring of turbomachines. The model is constructed and trained by using orbit image series in the healthy operation status through convolutional long short-term memory network. Then the temporal evolution of orbit shapes is predicted by the trained OrbitLSTM model, and used as a reference to evaluate the machine status under unknown conditions. The early fault warning is triggered when the mean square error or the structural similarity index between measurements and predictions exceeds the predefined threshold. The proposed model seamlessly combines the CNN’s capability of automatic feature extraction from images and the LSTM’s advantage of long-term memory in historical sequence behaviors. The capability and accuracy of the proposed model for fault monitoring and early warning are demonstrated using both the simulation data and the raw vibration signals collected from sensors installed in different real-world industrial turbomachines. This study provides a new image-based deep learning method for automatic condition monitoring towards the predictive maintenance of large turbomachines in various industrial fields.