Regression methods for improved lifespan modeling of low voltage machine insulation

This paper deals with the modeling of insulation material lifespan in a partial discharge regime under certain accelerated electrical stresses (voltage, frequency and temperature). An original model, relating the logarithm of the insulation lifespan, the logarithm of the electrical stress and an exponential form of the temperature, is considered. An estimation of the model parameters is performed using three methods: the design of experiments (DoE) method, the response surface method (RSM) and the multiple linear regression (MLR) method. The estimation is obtained on learning sets determined according to each method specification. The performance, in terms of estimation, of each of the three methods is evaluated on a test set composed of additional experiments. For economic reasons and flexibility, the learning and test sets are composed of experiments carried out on twisted pairs of wires covered by an insulator varnish. The ability of the DoE and the RSM methods to organize and to limit the number of experiments is confirmed. The MLR method, however, shows more flexibility with regard to the studied configurations. Thus, it offers an efficient solution when organization is not required or not possible. Moreover, the flexibility of MLR allows specific ranges for the factors to be explored. A local analysis of the estimation performance shows that very short and long lifespans cannot be simultaneously represented by the same model.