Fault Diagnosis Of Rolling Bearing Based On Adaptive VMD And Optimized Neural Network

Rolling bearing is one of the basic parts in rotating machinery system,and its good operating condition determines the performance and safety of the mechanical system.However,due to the production process,operating environment and many other factors will cause the bearing failure,so it is necessary to effectively monitor and diagnose the running state of the rolling bearing.With the birth of bearing fault diagnosis technology,the disadvantages of regular shutdown detection and maintenance were eliminated,reliability and production efficiency of rotating machinery were improved.In this paper,the vibration signals of rolling bearings in different operating states are taken as the research object.Aiming at some problems that need to be solved in the current fault diagnosis technology.For example,when the variational modal decomposition method is used to process the bearing signal,it is necessary to set the modal number of the bearing signal;and the training process of BP neural network is easy to stop.The research is carried out from two directions of signal processing and pattern recognition.The main research contents are as follows:Firstly,the significance of the state of rolling bearing to industrial production is introduced,and the technical means commonly used in fault diagnosis of rolling bearing and their advantages and disadvantages were expounded.In this paper,the basic principle of variational modal decomposition algorithm(VMD)is analyzed in depth.Aiming at the limitation that VMD needs to preset the number of modes when analyzing the bearing signal,a strategy is proposed to determine the optimal number of IMF in the variational modal decomposition algorithm based on the relationship between adaptive threshold and spectrum extremum.In this paper,it is called adaptive variational modal decomposition,which is referred to as AVMD.Through the analysis of simulation signals and rolling bearing signals,it is verified that AVMD can effectively decompose complex multi-band signals and there is no modal aliasing.In addition,in the process of processing the fault signal of rolling bearing,AVMD can effectively separate the impact characteristics that are overwhelmed by the noise factor,and can obtain the rotation frequency and fault characteristic frequency.Secondly,Because the BP neural network is prone to premature convergence during training,and the current better improvement strategy is to introduce intelligent optimization algorithms.However,most intelligent algorithms have the common defect that the learning direction is single and the training process is easy to stop.So this paper puts forward an improved intelligent optimization algorithm called the differential evolution with repulsive behavior algorithm,referred to as RBDE.The core idea of the RBDE is that the offspring nolonger simply learn from the optimal individuals,and this mechanism increases the diversity of learning directions of the population.There are two unique learning mechanisms in RBDE algorithm.The one is that RBDE selects two parent individuals as the source of repulsion and generates two kinds of repulsive forces in different directions,the repulsive forces will push the offspring to explore the optimal position.The second is to take the historical gradient between the two parent individuals as the learning direction of the offspring.The CEC2017 test function proves that RBDE has an excellent ability to solve complex or random problems,and it is significantly better than some algorithms,such as particle swarm optimization.Therefore,RBDE algorithm is applied to optimize BP neural network(It is called RBDE-BP neural network.),which will help BP neural network to quickly converge to the global best advantage,and pave the way for subsequent multi-type bearing fault identification.Finally,The bearing signals in seven operating states of western reserve university were studied.After all kinds of signals are processed by AVMD,the kurticity index is used to reconstruct the signals.The multi-scale permutation entropy of the reconstructed signals is extracted as the feature vector of the bearings,and the training and testing of the RBDE-BP neural network are completed by using these feature vectors.The results show that the multi-scale permutation entropy values of various types of bearing fault signals extracted through kurtosis are separable.In addition,the number of iterations and training errors of the BP neural network model optimized based on RBDE algorithm in fault identification of seven kinds of bearings are significantly reduced,and the diagnosis rate can reach 98.57%.