Weak Fault Feature Extraction Method for RV Reducer Based on CPO-MNAD and LCPSO-BP Neural Network

This thesis introduces a fault feature extraction method that integrates CPO-optimized MNAD with the LCPSO-BP neural network, targeting the problems of insufficient fault diagnosis accuracy and the difficulty in distinguishing among different fault types in RV reducer within practical applications. To enhance the feature expression capability of fault signals, two critical parameters of the MNAD approach are initially optimized using CPO. The failure signal was decomposed modally using the EEMD method to extract multiscale characteristics. The multiscale fuzzy entropy was subsequently calculated to quantify the signal’s complexity and uncertainty, hence enhancing the description of tiny fluctuations within the signal. The data is subsequently input into the optimized LCPSO-BP neural network for training and classification, after the enhancement of the BP neural network by Levy flight and Chebyshev chaotic mapping techniques in particle swarm optimization. The experimental results indicate that the method exhibits substantial robustness against noise interference and markedly enhances classification accuracy in the fault recognition of RV reducers. It significantly enhances the precision in identifying fault types such as tooth wear, surface faults, root cracks, and others. The recognition accuracy improves by 17.5%, 19.17%, and 17.51%, respectively, in comparison to CYCBD, MOMEDA, and MCKD. Furthermore, in the context of optimizing MNAD, the CPO algorithm outperforms GWO and PSO, demonstrating superior accuracy, expedited convergence, and enhanced resilience.