Multi-Domain Extreme Learning Machine for Bearing Failure Detection Based on Variational Modal Decomposition and Approximate Cyclic Correntropy

Abstract

Rolling bearings are critical in industrial mining machinery. Due to strong Gaussian noise, frequent random shocks, and disordered loads in industrial settings, it is usually difficult to detect weak fault symptoms in vibration signals from a bearing. To detect incipient bearing faults, this paper proposes a new multi-domain kernel extreme learning machine (MKELM) based on variational modal decomposition (VMD) and a cyclic correntropy function. A normalized approximation algorithm for a cyclic correntropy function (NACCF) was first built to suppress the impulsive background noise. This approach is suitable for machine learning. To eliminate the Gaussian noise effectively, genetic mutation particle swarm optimization (GMPSO) with cyclic information entropy (CIE) was used to optimize the VMD parameters. The CIE was created as a fitness function in GMPSO to search for the best hyperparameters. It can be used to select effective intrinsic mode functions (IMFs) to reconstruct denoised signals. Then, statistical functions based on NACCF were used to extract the cyclic frequency-domain characteristics of the denoised signal, and the singular values of the IMFs were obtained as time-domain features of the signal. Finally, the multi-dimensional features from the two domains were input into MKELM to classify the health of the bearing. Experimental studies were carried out to investigate the proposed method in bearing fault detection and identification. The results demonstrated the effectiveness of the proposed method in motor-bearing failure detection and its robustness to noise when analyzing bearing vibration signals under different working loads.