Intelligent Fault Diagnosis Method Integrated Beluga Whale Optimization-based VMD and LSSVM with WSET-CNN for Rolling Bearing

As a key component of various machinery, rolling bearing’s running reliability performance has a significant impact on the whole system. However, it will become an extra challenge when facing few-shot data and severe noise interference. Recently, the Beluga Whale optimization (BWO) algorithm has been presented, which has caused widespread concerns due to its powerful optimization capability and fast convergence characteristic. The variational modal decomposition (VMD) has been proven by numerous scientific experiments to be a powerful tool for noise reduction; the wavelet synchronous extraction transformation (WSET) exhibits strong time-frequency aggregation performance. In addition, the CNN model has significant superiority in automatically extracting hidden features, and the LSSVM algorithm is highly effective in handling small sample pattern recognition problems. Integrating all of the advantages of the above-mentioned techniques, a novel intelligent fault diagnosis scheme using BWO-based VMD and LSSVM with WSET-CNN is introduced. This method includes three parts, which are BWO-based VMD for signal preprocessing, feature extraction based on WSET-CNN architecture, and fault identification using BWO-optimized LSSVM. At first, the VMD can extract various intrinsic mode components (IMFs) contained in the original signals so as to achieve noise reduction after selecting the optimal IMF. However, the effect is greatly influenced by its parameters. Therefore, during the signal preprocessing phase, the BWO algorithm is implemented to optimize key parameters of variation mode decomposition to achieve optimal noise-reduction results. Second, time-frequency characteristics are captured using the WSET technique during the feature-extraction stage, and subsequently, the time-frequency spectrum images are utilized as an input matrix for the CNN model to fulfill the feature mapping and dimension reduction through convolution and pooling operations, automatically extracting the hidden characteristics vectors. In the final stage, the BWO algorithm was once again utilized to search for the best parameters for the LSSVM model so that the classification results could be achieved. In the final section, a fault diagnosis experiment for bearing is conducted and the analysis results are obtained, from which it can be concluded that the newly-developed methodology has superiority with high accuracy even in small-sized sample intelligent fault diagnosis.