Research On Fault Feature Extraction And Pattern Recognition Method Of Rolling Bearings

The complexity,variability and uncertainty of rolling bearing fault make its diagnosis more difficult.The key process that determines the diagnosis performance includes feature extraction and fault recognition.The quality of feature extraction will directly determine the level of diagnosis accuracy,and building a reasonable recognition model is also of great significance.In this paper,from the aspect of fault identification model construction,aiming at the data samples of various fault types,different fault locations,different fault degrees and different load mixtures of rolling bearing,the optimized kernel extreme learning machine(KELM)model and the improved deep belief network(IDBN)model are constructed respectively to realize the intelligent identification of complex condition mixtures of rolling bearing.The main work and research contents of this thesis are as follows:(1)Aiming at the problem of fault feature extraction of rolling bearing,statistical analysis,FFT and VMD methods are used to extract the features of rolling bearing signals in time domain,frequency domain and time-frequency domain respectively,and a high-dimensional feature vector which can fully reflect the inherent attributes of the original signal.Compared with the traditional single domain feature recognition method,it has higher recognition accuracy.In order to eliminate the high-dimensional fault characteristics of rolling bearing in order to improve the execution speed of the algorithm,the Laplace Score(LS)algorithm is used to automatically select the features of the high-dimensional feature matrix.Compared with the original feature and Fisher Score(FS)feature selection,LS algorithm can effectively screen out the sensitive features and obtain higher recognition accuracy.(2)Aiming at the problem of time-consuming and easy to fall into extreme value of KELM parameter intelligent optimization,Simulated Annealing Particle swarm optimization(SAPSO)was used to optimize the penalty coefficient C and kernel parameter g,and the low-dimensional feature of automatic selection was input into KELM classifier to realize the intelligent recognition of complex working conditions of rolling bearing,as well as the diagnostic performance of ELM,PNN and SVM classifiers Compared with KELM,it has faster recognition speed,higher recognition accuracy and better robustness,which is more in line with the actual engineering conditions of intelligent recognition.(3)Combined with the advantages of DBN in feature extraction and processing of high-dimensional and nonlinear data,a sparse regularized IDBN model based on continuous Gaussian distribution is proposed.In order to solve the problem that IDBN needs a lot of training data,the sliding window method is used to increase training data samples.In order to solve the problem of low training efficiency,the compressed sensing theory(CS)is used for data preprocessing.Finally,a fault identification method based on CS and IDBN model is proposed to realize the fault intelligent identification of rolling bearing under different fault types,different fault degrees and different loads Compared with the traditional intelligent recognition method,the CS-DBN model has higher recognition accuracy.