Study On Method Of Manifold Learning For Mechanical Fault Diagnosis

The methods of mechanical fault diagnosis are various and the essence of which is a pattern recognition and classification problem.The key technology of mechanical fault diagnosis is the signal feature extraction and compression.Some of the existing analysis methods have great limitations for the nonlinear signal processing in the machine running.Manifold learning is a machine learning algorithms that has good effects for the nonlinear high-dimensional data processing.In the paper,manifold learning is applied to the mechanical fault diagnosis.The problem of high-dimensional feature compression of signal feature space can be better solved.The solution can provide a good basis for fault classification.The dimension reduction effects are tested utilizing the typical data sets.And according the analysis results,locally linear embedding(LLE)is selected as the focus in the study of the paper.Comprehensive parameters selection method based on the optimum classification results is proposed combining the signal processing characteristics of the mechanical fault diagnosis,to solve the problems of manifold learning application in mechanical fault diagnosis.The neighborhood factor k and the embedding dimension d can be simultaneously optimum selected utilizing the parameters selection method.Improved LLE method based on the local topology preservation is proposed to solve the problem of the feature compression and fault diagnosis for the newly added sample accurately and quickly.The method make the best of the retained information of local topology structure after the LLE dimensionality reduction of original feature space,and avoid the repeating calculation to all the data.The application's process and matters needing attention of improved LLE method are discussed in the mechanical fault diagnosis,and the K-Nearest Neighbor classifier(KNN)is chosen as the classifying rule of fault diagnosis after LLE feature compression.The improved LLE method is applied to the diagnosis of gearbox fault,diesel fuel injection fault and diesel mechanical fault in the paper.The feature space construction method based on the sub band energy is used to calculate the signal feature space for the running vibration signal of gearbox and diesel.And the sub band number is calculated by the minimumvariance among feature parameters in same fault type.The feature space construction method based on the oil pressure waveform parameters is used to calculate the signal feature space for the fuel injection pressure signal of diesel fuel system.Form the results of feature compression and fault diagnosis of gearbox vibration signal,diesel vibration signal,and the diesel fuel injection pressure signal,it can be seen that: the application results of improved LLE method in the paper is good for the feature compression and classification of mechanical fault signals.