Combined use of partial least-squares regression and neural network for residual life estimation of large generator stator insulation

Residual life estimation of insulation material is always a difficult problem in electrical insulation condition assessment. This paper describes the development of hybrid framework- based systems to estimate the residual life of large generator stator winding insulation using routinely measured non- destructive diagnostic data. Nonlinearity of the stator winding insulation degraded process makes it difficult to establish the relationship between the residual life and the nondestructive parameters by diagnostic data. By using the universal approximation property of neural networks, the PLS modelling method was used to generalize a nonlinear framework. RBF networks were introduced to the PLS modelling in this study, termed ANN- PLS hybrid modelling. A nonlinear modelling task was therefore decomposed into linear outer relations and simple nonlinear inner relations, which was performed by a number of single- input - single- output networks. Since only a small- size network was trained at one time, the over- parametrized problem of the direct neural network approach was circumvented even when the training data were very sparse. As an application, the proposed ANN- PLS hybrid soft computing method was compared with the standard linear PLS and the multiple linear regression. Test results demonstrated that the hybrid approach had better prediction ability than traditional estimation methods.