Research On Fault Diagnosis Of Rolling Bearing Based On Wavelet Packet And Machine Learning

As one of the key components of industrial equipment,the failure of rolling bearing in service will have a great impact on the performance and utilization of the equipment,and even lead to safety accidents.Therefore,it is necessary to diagnose the rolling bearing fault effectively and accurately.The complexity of rolling bearing signal with fault brings some difficulties to the feature selection used to judge the fault.Although there are many fault feature selection methods at present,in the traditional fault feature selection methods,due to the limitations caused by the influence of manual selection of parameters,the selection process of fault representation falls into a saddle point and cannot achieve the global optimization,which ultimately affects the accuracy of bearing fault diagnosis.Aiming at the above problems in fault feature selection,and in order to further improve the efficiency and accuracy of bearing fault diagnosis,this paper combines particle swarm optimization support vector machine,BP neural network algorithm and wavelet packet to optimize the selection process of fault features and diagnose,and uses CNN algorithm to diagnose the two-dimensional image representation of signal.The results show that the combination of machine learning algorithm and signal processing method can effectively solve the limitations caused by external artificial factors,reduce the uncertainty representation in the signal,and improve the efficiency and accuracy of bearing fault diagnosis.1.In order to obtain more accurate and effective fault features,the advantages of wavelet packet analysis in time domain signal processing are used to carry out multi-scale orthogonal decomposition of healthy vibration acceleration signals and vibration acceleration signals containing typical bearing faults such as inner ring,rolling element and outer ring,so as to obtain multi-layer decomposition signals with independent frequency bands,and then construct the decomposed fault sensitive feature information.2.In view of the disadvantage that the traditional BP neural network algorithm is easy to fall into the saddle point and difficult to achieve the global optimization,the weight parameters of BP neural network are optimized by using particle swarm optimization(PSO)algorithm,and the sensitive characteristics of bearing fault dynamic signal are input into the optimized BP neural network for diagnosis.Compared with other algorithms,the diagnosis results verify the effectiveness of this method.3.Aiming at the problems existing in the widely used supervised algorithm support vector machine(SVM),PSO algorithm is combined into parameter pairs to optimize the penalty factor C and Gaussian kernel parameter g,and a multi classifier model is constructed and trained.The diagnostic accuracy of different kernel functions and kernel parameters is verified,and the influence of the selection of model parameters on the diagnostic results is analyzed.The sensitive features are used to diagnose the fault,and compared with other algorithms.The diagnosis results verify the practicability of this method.4.The convolution neural network(CNN)architecture under the image representation of two-dimensional vibration time domain signal is built,and the data set of fault image representation is constructed as the input of CNN network.The le-net5 model is used to diagnose the bearing fault,compare the models under different combination optimization,and analyze the diagnosis effect under different errors and losses.The diagnosis results show that the algorithm has good performance and broad application prospects.