Research On Bearing Fault Diagnosis Based On CPA-CYCBD And Dense Capsule Neural Network

Rolling bearings,as one of the most commonly used parts in rotating machinery,will be damaged due to the poor working environment of rotating machinery and other factors,which will lead to work stoppage and production,and even lead to significant property losses or casualties.Therefore,it is very necessary to carry out state monitoring and fault diagnosis research on bearings.Intelligent fault diagnosis,as the most popular fault diagnosis direction in recent years,has been widely studied by scientific researchers for its advantages of high fault diagnosis accuracy and no need to know the fault characteristic frequency in advance.In this dissertation,the formula for calculating the characteristic frequency of rolling bearings in single point fault was derived firstly.Then,the important theoretical basis of neural network was introduced,and the formula of neural network parameter gradient descent training method was derived.Aiming at the problem that the vibration signal in the strong noise background is directly imported into the fault recognition model and the fault diagnosis effect is not strong,the noise reduction algorithm under the strong noise background was studied.Maximum Second-order Cyclostationarity Blind Deconvolution(CYCBD)was studied.Aiming at the problem of length selection of inverse filter in CYCBD,Carnivorous Plant Algorithm(CPA)was used for optimization,and Weighted Permutation Entropy(WPE)was studied.WPE was used as a fitness function in CPA,and the fitness threshold was determined through simulation signal experimental research.In order to test the noise reduction performance of the proposed algorithm,experiments were carried out under the simulation signal under strong background noise and the multi-point fault bearing vibration signal under the accelerated life experiment.The results show that the noise component of the signal after noise reduction is slightly less than that before noise reduction.Signal processing and other methods need to know the fault characteristic frequency of test signals in advance.For multi-point fault or compound fault scenarios,the fault characteristic frequency is often difficult to be solved by mechanism,so a data-driven fault diagnosis method was studied.Dense convolutional neural Networks(Densenet)and Capsule neural Networks(Cap Net)were studied,and Dense Capsule neural network(DCN)was proposed.DCN has the classification ability of strong feature extraction ability.There is no need to use other feature extraction methods to improve its fault recognition ability.According to the fault diagnosis scenario,a novel fault diagnosis process was proposed by combining CPA-CYCBD and DCN.In order to test the performance of the proposed process,experiments were carried out on two groups of public datasets.The influence of fault diameter,load and composite fault was considered during the experiment.The results show that the proposed method has good fault diagnosis performance on both groups of datasets.Compared with the methods applied to the same dataset in recent three years,the results show that the proposed method has certain advantages in the accuracy of fault identification and experimental design.