Research On Fault Diagnosis Of Ultra-low Speed Rolling Bearing Based On Acoustic Emission Technology

The running state of rolling bearing directly affects the performance and safety of the entire mechanical equipment as a mechanical component widely used in various of rotating machinery equipment.Unlike conventional bearings,the ultra-low speed rolling bearings used in large-scale heavy industrial machinery have complex structures and working environments,and often withstand large loads.It takes a lot of time and effort to repair and replace in case of failure,which leads to great economic losses.Therefore,it is of great significance to monitor the condition and early failure of ultra-low speed rolling bearings.Acoustic emission(AE)technology is a new type of dynamic detection technology with high sensitivity and sensitive to dynamic defect,which is widely used in the field of mechanical fault diagnosis.However,due to the non-stationarity,uncertainty and complexity of AE signal of the ultra-low speed rolling bearing,the extraction of the fault feature information from the AE signal becomes the key and difficult point of the diagnosis.Therefore,this thesis takes ultra-low speed rolling bearing as the research object,and pre-fabricates different types of defects(pitting and crack)on the inner ring and rolling element of rolling bearing by electric discharging machining and cutting process technology,respectively.The corresponding AE signals were collected on the simulation test bench and analyzed by different signal processing methods.Aiming at the feature extraction of ultra-low speed rolling bearing,the fault feature extraction method based on Compelete Ensemble Empirical Mode Decomposition With Adaptive Noise(CEEMDAN)-energy entropy and Improved Variational Mode Decomposition(IVMD)-sample entropy is studied respectively.For the fault pattern recognition and classification,the Deep belief network(DBN)model is established and the fault diagnosis of ultra-low speed rolling bearing is carried out.The Ensemble Empirical Mode Decomposition(EEMD)and CEEMDAN methods are applied to the AE signal processing of ultra-low speed rolling bearings.Compared with the EEMD method,the CEEMDAN method improves the completeness of the decomposition,and has better anti-modal mixing property.The sensitive Intrinsic Mode Function(IMF)component of CEEMDAN decomposition is selected by correlation coefficient and variance contribution rate,and its energy entropy is calculated as the feature vector of BP neural network for pattern recognition.The average accuracy is up to 94.13%,and the diagnosis effect is better.When the energyentropy of the sensitive IMF of EEMD is used as the characteristic parameter of BP neural network,the recognition accuracy is only 87.13%.The improved VMD sample entropy method is used to extract the feature of AE signals of ultra-low speed rolling bearings.The collected AE signal is decomposed by EMD,and the sensitive IMF component is extracted by correlation coefficient and variance contribution rate to reconstruct the signal.Then,the number of sensitive IMF components is used as the mode number of VMD to decompose the reconstructed signal and the sample entropy of IMF component is calculated as the feature vector of BP neural network for fault pattern recognition and classification.The recognition accuracy is as high as 94.27%.In addition,the improved VMD energy entropy and approximate entropy are used as the feature vectors of BP neural network respectively,and the recognition accuracy is 83.33% and 90.67%,which is lower than that of sample entropy.The DBN model is established to diagnose the fault of ultra-low speed rolling bearings.The number of training iterations of DBN model has a great influence on its classification performance.With the increase of iteration times,the recognition accuracy of DBN model increases first and then decreases.Too many iterations are not conducive to the improvement of recognition effect.The energy entropy of the first 9orders IMF of EEMD and CEEMDAN are used as the feature vectors of DBN pattern recognition classifier respectively for pattern recognition.The recognition accuracy of both models is more than 90%.The effectiveness of the DBN method in the fault diagnosis of ultra-low speed rolling bearings is verified.Among them,the average accuracy of CEEMDAN energy entropy as feature parameter is 99.33%,and the recognition accuracy of EEMD energy entropy as input vector of DBN model is only90.80%,which is obviously lower than that of CEEMDAN energy entropy as feature parameter,indicating that the extraction of fault features also has a great impact on the performance of DBN.Compared with the classical pattern recognition method-BP neural network,the deep structure of the DBN model can learn the feature information of the data set more fully,and the recognition accuracy is higher and the stability is better.Therefore,it can be applied to the AE diagnosis of ultra-low speed rolling bearing.