Bear Fault Classification Based On Multilevel Neural Network

Bearings are one of the most important components in industrial production.However,the bearings usually operate under high temperature,high pressure,high speed and strong impact,resulting in frequent bearing failures.Unexpected failure of the bearing can result in slumps and even deaths of the entire mechanical system.Therefore,it is important to find the signs of bearing failure in advance and accurately judge the degree of bearing failure to ensure the safe and smooth operation of industrial production.In order to improve production efficiency and reduce production accidents,this paper proposes a multi-layer neural network rolling bearing fault diagnosis model integrated by improved perceptron,dynamic routing algorithm and stochastic optimization algorithm.Firstly,the original bearing vibration signal database is expanded by randomly intercepting the fault bearing vibration signal samples,and the bearing fault classification is performed on the randomly intercepted vibration signal samples according to the damaged position of the bearing,the diameter of the damaged point and the sampling position of the acceleration sensor,and then Each fault bearing vibration signal sample is marked with a corresponding fault category label.The bearing fault feature extraction network module is constructed by embedding the dropout algorithm in the multi-layer perceptron network,and then the scalar fault feature sequence extracted by the bearing fault feature module is rearranged into a vector form fault feature as the input vector of the primary capsule.The primary capsule input vector is transformed into the input vector of the advanced capsule by linear transformation to predict the bearing fault category.Furthermore,the dynamic consistency routing algorithm is used to measure the optimal coupling path of the bearing fault feature vector by measuring the uniform coupling coefficient between the input vector of the advanced capsule and the fault label vector.Then the transmission path of the bearing fault feature is determined according to the optimal coupling coefficient,and the fault category of the bearing is predicted.The prediction vector is the output vector of the advanced capsule.Furthermore,the objective function obtains the network prediction loss by comparing the consistency of the fault label with the prediction result of the fault diagnosis network,and the weight error of each layer is calculated by the back propagation algorithm,and then the Adam parameter optimization algorithm is used to update the bearing fault diagnosis network.The weighting parameter is used to select the optimal learning rate by the adaptive learning rate algorithm,so that the fault type prediction vector of the fault diagnosis model quickly approaches the label vector of the actual fault category of the bearing.Finally,based on the damaged position of the faulty bearing,the damaged diameter,and the sampling position of the accelerometer,the Case Western Reserve University fault bearing reference data is divided into five different types of bearing fault category labels: A,B,C,D,E.Then the neural network fault diagnosis model proposed in this paper is used to perform MATLAB numerical simulation experiments on each database.The average precision of the simulation experiments are 99.57%,98.84%,98.45%,99.31%,and 99.39%,respectively,while the simulation accuracy of the classical algorithm ADCNN is only 98.10%,the results show that the fault diagnosis model can achieve high-precision bearing fault diagnosis and classification.It is the main work of this paper to propose the bearing fault diagnosis network model and verify the effectiveness of the network model through different types of bearing fault vibration signals.It provides a new network model idea and verification method for the accurate classification of bearing faults.