Fault Pattern Classiifcation Of Gearbox Based On Local Linear Embedded

The incipient fault signals are weak, its characteristics are not significant to be easilysubmerged in strong noise background, which makes the classic fault diagnosis method basedon signal analysis can not effectively extract the fault information. At the same time, manytypes of failures in high dimension and complexity of the conventional signal is a majorproblem faced by the fault diagnosis method. Therefore, the intelligent fault diagnosis methodof research work has an important significance.Linear discriminant analysis (LDA) is one of the supervised feature extractionmethods, which is a linear classifier essentially. It can be used for data preprocessing orfault diagnosis directly. However, its application is limited in fault areas due to itslinearity. In recent years, manifold learning is a new research focus in pattern recognitionfield, its nature is to study and found the low-dimensional smooth manifold embedded inhigher dimensional space based on a limited number of discrete samples, which revealsdata intrinsic structures. In this paper, we introduced two representatives of global andlocal manifold learning algorithm. Manifold algorithms are unsupervised learning, whichcan not take full advantage of the sample label information. A new nonlineardimensionality reduction method, named local linear discriminant embedding (LLDE),combined linear discriminant analysis (LDA) with local linear embedding (LLE) isproposed. Classic data sets are used to verify the effectiveness of this algorithm.Experiments of gear faults are conducted on transmission testing platform. Wecollected gear faults under normal, pitting and spalling states. Select the bestcharacteristics to indicate fault information as the input of the proposed method in order toachieve the best result of transmission’s failure pattern classification. The originalalgorithmand this improved method is applied to gearbox fault diagnosis, results show thatthe proposed supervised local linear discriminant embedding method is more effective andmore stable than the orginnal one. It is successfully applied to faults pattern classificationof transmission.Besides that, the bearing datasets from Case Western Reserve University were alsoutilized to verify the effectiveness of the LLDEC method. There were four conditions in thesesignals, the normal data, the inner-race fault, the outer-race fault and rolling element faultsignals correspondingly, which cannot be separated from each other using spectrum analysisand time statistical analysis. Experiments results demonstrated that the proposed method iseffective and correct in bearing fault prognosis.