Fast bearing fault diagnosis of rolling element using Lévy Moth-Flame optimization algorithm and Naive Bayes

Mechanical fault diagnosis is a technology that monitors the state of the machine during operation, determines its overall or partial normality or abnormality, determines the fault and its cause, and can predict the development trend of the fault. Due to the improvement of the automation and integration of the current mechanical system, the maintenance time and cost have gradually increased [7, 14, 21, 47], causing greater economic losses [2, 3, 19, 34]. Therefore, the importance of the fault detection system in the daily maintenance system is becoming stronger [4, 15, 24, 50]. The fault diagnosis method can be divided into time-frequency analysis method and pattern recognition method. The traditional timefrequency analysis method identifies the fault based on the energy change of each frequency band of the vibration signal when the fault occurs. Commonly used methods include Wavelet transform (WT) [37], Empirical mode decomposition (EMD) [11], ensemble empirical mode decomposition (EEMD) [23], and etc [1, 48]. WT can effectively extract the characteristics of nonlinear transient vibration timefrequency signals, but when processing complex vibration signals, different basis functions need to be selected to obtain the best results, and there is no uniform standard for parameter selection. EMD has adaptive signal processing capabilities, but it has disadvantages such as edge effect and modal mixing [14]. The existing pattern recognition methods mainly include ELM [28], SVM [7, 48], ANFIS [22, 46], and etc. ELM has fast learning speed and requires few training samples, and can realize fast main bearing fault diagnosis, but its stability is relatively weak. SVM can efficiently solve high-dimensional nonlinear decision-making problems, but it is difficult to select kernel parameters and sample parameters, and is significantly affected by fault samples. However, for the pattern recognition method, due to the high cost of obtaining mechanical system failure vibration experimental data and the limited degree of failure, which makes the existing methods have limited recognition accuracy and even misrecognition problems. Machine learning related methods have certain advantages in small sample classification. However, the original signal is distorted and non-linear, and is often buried in a large amount of background noise and other interference. Therefore, higher requirements are put forward for signal preprocessing and feature extraction. Meanwhile, with different varieties of features given specific effectiveness, it is feasible to apply compound features for eliminating information loss. However, the feature combination will introduce new proper information and Keywords