Research On Fault Diagnosis Of Rolling Bearing Based On Manifold Learning Algorithm

Aiming at the obviously nonlinear and non-stationary characteristics of the vibration signal detected when the modern complex mechanical equipment breaks down, taking rolling bearing as the research object, the paper uses the vibration acceleration signal collected by the sensor as a basis for the rolling bearing state recognition, then reduce noise, extract feature, build a high dimensional feature space and recognize state. It introduces Laplacian Eigenmap of which clustering effect is better in the manifold learning method into the field of rolling bearing fault diagnosis, and makes a system research on noise reduction, feature extraction and state recognition of the nonlinear and non-stationary state signal of bearing fault.It makes a research from the different angles of keeping point features of the high-dimensional data, divides the current main manifold learning method into two kinds of methods of local and global features to keep, obtains different dimension reduction results from different ways to keep through studies he classical nonlinear high dimensional crimp data set Swiss Roll and Swiss Hole with an "empty" defects, and from the analysis of 2 d figure result discovers that LE algorithm which stress on keeping neighborhood relationship has a very good clustering effect for the similar samples, and is very suitable for recognition of fault samples.It introduces LE algorithm into fault diagnosis field, building up a high dimensional feature space by characteristic quantity of time domain gained through extracting and conversing the nonlinear signal of the simulation rolling bearing fault and frequency band energy ratio decomposed by wavelet packet, uses the advantage of LE algorithm for similar samples in high dimensional feature space clustering to identify two classes of samples of the normal and the faults, and proves the effectiveness and superiority of LE method in realizing feature extraction and sample classification of dimension reduction results through the comparison of PCA and MDS. In order to further proves the effectiveness of LE algorithm in rolling bearing fault identification, the paper analyzes the measurement data of rolling bearing fault, and does experimental verification by making a classification recognition of two groups of fault sample through four different fault types of rolling bearings and four different damage condition on the roller. It uses the method of 3 d visualization to show the low dimensional results extracted, and use three parameters of within-distance Sb and between-class scatter Sw in pattern recognition and the average recognition rate of different samples as evaluation index, and proves that LE has the advantages of good clustering, small class spacing, and high sample recognition rate in dimension reduction results, and that LE can be effectively applied in rolling bearing fault identification.In order to further prove the applicability of LE method, the paper uses a set of test experiments to conduct a test. It uses the sample of the first group experiments of rolling bearing fault type recognition as the training sample, and then selects two kinds of state samples with 50 groups in each as the test sample, does the verification and realizes the anticipated effect:Similar fault test sample well gathers in the training sample of similar samples, and the new sample separately gathers in a new location.