Research On Multi-objective Fault Feature Extraction And State Assessment Of Rolling Element Bearing

As a key component of rotating machines,rolling element bearing plays an irreplaceable role in the equipments such as locomotives,aero engines,etc.In many cases,it operates with high speed,high load or prolonged time.Thus,some localized defects as pitting,spalling may generate on the contact surfaces.These faults will lead to abnormal vibrations which will reduce the working accuracy and even cause catastrophic accidents.So,detection of the bearing failures as early as possible is an important and meaningful subject for the maintenance.However,due to the complex interference and heavy background noise,these faults are difficult to be identified directly.To meet the above problem,this paper presents a series of qualitative and quantitative fault diagnosis methods based on multi-objective feature extraction which take impulsiveness and cyclostationarity into consideration simultaneously.The details are listed as follows:(1)Based on spectral kurtosis and frequency domain correlated kurtosis(FDCK),a fast filtering method is proposed for the envelope analysis.Firstly,binary wavelet packet transform(WPT)is utilized to decompose the signal into subbands.To obtain the correct relationship between the node and frequency band,an ordering process is conducted to solve the frequency folding problem due to down sampling.Correlated kurtosis of the squared envelope spectrum i.e.the FDCK is utilized to generate the kurtogram,in which the maximum value can well indicate the optimal resonance band.Moreover,with different inputs of the period in correlated kurtosis,the proposed method can be applied for the compound faults detection.Several cases of simulated and experimental bearing fault signals are used to evaluate the immunity of the proposed method to strong noises.The improved performance has also been compared with three previous developed methods.(2)The fast filtering method has a high computational efficiency but also easily results in segmentation of the informative frequency band.In addition,contribution of the bearing fault frequency harmonics to the FDCK will decrease as the order increase.Considering the above drawbacks,a new dynamic Bayesian wavelet demodulation method based on the normalized FDCK and complex Morlet wavelet is further proposed.By maximizing the ratio of correlated kurtosis to kurtosis of the squared envelope spectrum,the general particle filter is employed to estimate posterior probability density function of the wavelet parameters,then an optimal wavelet fitering can be preformed to extract the bearing fault features.The performance of this method is assessed using the same previous signals,and the results demonstrate its superiority than the fast filtering method.(3)Calculation of the correlated kurtosis requires accurate bearing fault frequency as prior knowledge,so the measurement error of the rotating speed will seriously affect its effectiveness.To overcome this shortcoming,the single-objective optimal filtering method is further extended,and a multi-objective dynamic Bayesian wavelet demodulation method is proposed.The time domain negentropy is utilized to characterize the impulsive features and the frequency domain negentropy is utilized to characterize the cyclostationary ones.With the non-dominated sorting,the Pareto set of these two aspects is calculated for updating the posterior wavelet parameters distribution in the Bayesian estimation.Then,the impulsiveness and cyclostationairty can be well synthesized and balanced in extracting the repetitive transients.Its performance is examined by both simulated and experimental signals,some comparisons with peer methods are also conducted to show its robustness in challenging different noises and interferences.(4)Previous studies focus on qualitative diagnosis,whereas the quantitative can make more sense for the maintenance strategy.Under the framework of multi-objective fault feature extraction,a novel methodology based on Infogram and hidden conditional random field(HCRF)is proposed to address the on-line performance degradation assessment of the bearing.Vibration signals from the bearing is firstly processed by the WPT-based SE infogram and SES infogram,the sub-band negentropies are collected as features to characterize the bearing condition.With a learning algorithm,a HCRF model can be trained to best fit the features.Then the current bearing state can be quantitatively assessed by the log-likelihood vallue.The reliability of the proposed method is inspected by two test-to-failure bearing data sets which are collected under different conditions.The experimental results show its effectiveness and robustness than the hidden Markov model-based and conditional random field model-based methods.