An Intelligent System for Bearing Fault Identification based on Gramian Multi-Resolution Dynamic Mode Decomposition

This work introduces a single sensor intelligent end-to-end framework for fault detection based on Gramian-multi-resolution dynamic mode decomposition (GMrDMD). Vibration signals are transformed using gram matrix followed by spatial temporal decomposition based on multi-resolution dynamic mode decomposition (MrDMD). The gram matrix converts the 1D data into time evolving snapshot matrix which retains the relation of signal with time. This forms the input to the MrDMD framework which decomposes the system dynamics into hierarchically evolving fast and slow modes capable of isolating the transient fault characteristics. To handle sensor and environmental noise a robust total least square DMD algorithm is applied at each level of MrDMD. The resultant mode matrix is color coded and fed as image to a convolutional neural network (CNN) for classification. The performance of the designed method is verified on University of Ottawa dataset which contains five type of fault vibration signal under four different time varying rotational speed condition. The results demonstrate that of the proposed data driven method is effectively able to distinguish between different fault characteristics with an accuracy of 96.83%.