A Bearing Fault Diagnosis Method Based on Improved Mutual Dimensionless and Deep Learning

Under nonlinear and nonstationary dynamic conditions, the fault diagnosis methods based on multidimensional dimensionless indicators (MDIs) often cannot provide effective and accurate health monitoring in the fault diagnosis of petrochemical units. In view of the above problems, this article preprocesses the dynamic signal and reconstructs a new dimensionless indicator. The indicator combines complementary ensemble empirical mode decomposition (CEEMD) with MDI, named complementary ensemble multidimensionless indicators (CEMDIs). By using the sequential mapping method, the CEMDI processed data can be converted into Gramian angular fields (GAFs). In processing sparse data, the advantages of convolutional neural networks (CNNs) were used to identify different fault types. The method is validated using three datasets, motor bearing data provided by the Case Western Reserve University, multistage centrifugal fan data, and machinery failure prevention technology challenge data. Compared with the traditional dimensionless index method and the latest published dimensionless methods in the literature, the fault diagnosis methods based on CEMDI and CNN exhibit good performance in identifying fault types under different conditions, which verifies its effectiveness and superiority.