Fault Feature Extraction Based On Local Tangent Space Alignment And Recognition Methods Research

Large rotating machines have the characteristics of complicated working conditions and are greatly influenced by external factors.And the vibration signals measured by acceleration sensors are usually with nonlinear and non-stationary characteristics.Therefore,the fault diagnosis based on vibration signals becomes more and more difficult.Fault feature extraction makes foundation of diagnosis and identification,so this paper takes important com-ponents-bearings and gears as research objects,and accomplishes the installa-tion and debugging of experiment platform.Thus the fault signals are meas-ured.For weak fault,an improved method based on minimum entropy decon-volution is applied to increase the SNR and complete the signal pretreatment.Combining LTSA,two main research directions are carried out:a de-nosing method of rolling bearing based on packing numbers-manifold learning and fault diagnosis methods based on local tangent space alignment and K-nearest neighbor classifier.The contents are as follows:(1)Fault feature extraction based on improved MED is studied.MED is applied to extract the feature of weak impact component.It can increase the SNR of original signal.However,under the situation of stronger noise the impact component will be lost by MED method.Concerned with the defi-ciency,apply intelligent optimization algorithm to minimum entropy decon-volution.Intelligent optimization algorithm is used as a tool to choose filter coeffients instead of objective function method(OFM)in MED with the maximum steepness value as the ultimate condition of filtering.Mainly two improved methods are proposed,one improved MED based on FOA and the other one based on PSO.Thus the interference of the strong background noise is eliminated and the fault feature signals could be extracted.Finally,results from low-speed bearing experiment showed that improved methods could ex-tract the fault features of the signals and they performed better than MED method.The reliability was verified.Comparing two improved methods,the improved MED based on FOA is better than that based on particle swarm op-timization at the aspect of non-linear de-nosing.The improved MED based on FOA is more suitable for early fault diagnosis of rolling bears.(2)A de-nosing method based on packing numbers-manifold learning is studied.It solves issues that actual engineering vibration signals are easily in-fluenced by complex working conditions.Firstly,the original signal contain-ing strong noise was decomposed into a series of intrinsic mode functions(IMFs)by complementary ensemble empirical mode decomposition(CEEMD).And the local tangent space alignment(LTSA)method was em-ployed to project IMFs set into low dimensional matrix.To determine reduc-tion dimension,packing numbers method was introduced.Then low dimen-sional components were inverse projected to the original high dimensional matrix by principal manifold reconstruction.After eliminating the noise,the reconstructed IMFs matrix was formed into a new signal.Thus Hilbert trans-form and FFT were carried on the signal.Experimental results show that the proposed methods both improve SNR and have good noise reduction effect compared with WT.At the same time,the method based packing numbers and phase space reconstruction is more accurate,but that based packing numbers and IMFs set is lack of precision.(3)The fault diagnosis methods based on local tangent space alignment and K-nearest neighbor classifier are studied.Aiming at the problem that the performance of local tangent space alignment(LTSA)is greatly influenced by nearest neighbor k,the fault diagnosis mode of LTSA and K-nearest neighbor classifier(KNN)based on clustering criterion is proposed.First,the vibration signal collected was used to construct high dimensional matrix,then the high dimensional matrix was standardized before dimensionality reduction.Ac-cording to clustering criterion,the nearest neighbor k in LTSA is chosen,and then the low dimensional feature vector of high dimensional matrix in the LTSA was extracted.Finally,the extracted low dimensional feature vectors were put into KNN to do fault pattern recognition.Rolling bearing and wind turbine fault diagnosis experimental systems are adopted to measure vibration signals of rolling bearings and gears at different states.Experimental results show that this method based on clustering criterion can effectively overcome the choice blindness of the optimal neighbor number k,and improve the preci-sion of dimension reduction and recognition rate of the fault diagnosis pattern.Compared with PCA,Laplacian eigenmaps(LE)and BP neural network-method,the method proposed is more suitable for bearing fault pattern recog-nition.