| As a key component in rotating machinery and equipment,with the advantages of high precision,high load,compact structure and reliable quality,rolling bearings are widely used in transportation,electric power,aviation and other fields.However,as rolling bearings are prone to failures during operation,the study of fault detection and diagnosis methods for rolling bearings plays a vital role in reducing maintenance costs,improving productivity and preventing system failures.The key point and the difficulty of rolling bearing diagnosis is effective feature extraction from rolling bearing vibration signal data,feature fusion as well as pattern recognition.In this paper,the multi-feature fusion method called IE-ASD combining Information Entropy andαStable Distribution parameters was proposed in order to diagnose the fault state of rolling bearings more accurately.The clustering performance of feature vectors constructed by various methods was explored based on vibration signals.In addition,this paper built a machine learning model for PSO-SVM,which applied of feature fusion method based on typical correlation analysis to fused IE and ASD features,thus performing fault diagnosis of rolling bearings.The effectiveness and noise immunity of the multi-feature fusion method were analyzed by experiments.The main contents of this paper are as follows:(1 The structure of rolling bearings,common types of faults,fault monitoring techniques and analysis of the characteristic frequencies of rolling bearing faults were outlined.The theoretical formulas for the fault frequencies of rolling bearings were presented at each position.The CWRU bearing data used in this paper and the data obtained from experiments were introduced.(2 Based on Information Entropy theory,the four entropies(singular spectrum entropy in time domain,power spectrum entropy in frequency domain,wavelet space characteristic spectrum entropy and wavelet energy spectrum entropy in time-frequency domain of the vibration signal were extracted.The four parametersα,β,δ,γof the ASD were estimated based on the empirical eigenfunction method,and the eigenvectors were constructed by serial combination.Principal Component Analysis was performed on the feature vectors.According to the cumulative contribution rates to obtain the feature scatter plots under different damage degrees,the three principal components with the highest contribution rates were selected as the projection directions of the feature vectors.The scatter plots of the feature vectors constructed only by single IE and single ASD were compared to analyze the clustering effect of the feature vectors.(3 Based on the principle of Particle Swarm Optimization algorithm and Support Vector Machine,the machine learning model of PSO-SVM is constructed.The Particle Swarm algorithm was used to optimize the kernel parametersσ~2 and penalty factors C of SVM.Aiming at comparing the classification accuracies of the three methods under different damage levels of rolling bearings,the feature vector sets constructed by each of the IE-ASD,ASD,and IE methods were input into the model for training tests.Confusion matrix plots of the diagnostic results for each dataset were drawn to analyze the specific classification identification of the samples when testing the model.(4 The noise immunity performance of the proposed IE-ASD method was analyzed and explored by manually adding a series of Gaussian white noise with Signal to Noise Ratio of-6,-4,-2,0,2,4,6,8,and 10 d B to the algorithm.The typical correlation analysis was applied to the Canonical Correlation Analysis of IE entropy values and ASD parameters to obtain the fused features.The fused eight-dimensional feature vectors or four-dimensional feature vectors were obtained by concatenating or summing between the two types of feature vectors respectively.Then,the PSO-SVM model was inputted for training and testing. |
