Piecewise linear factor analysis by four layer neural networks and its application for modeling the partial discharge data

This paper presents the methodology of a nonlinear version of factor analysis by four layer feedforward neural networks and, as an example of its application, the result of modeling the structure of partial discharge data measured on a power cable. Here, the authors introduce the four layer auto associative memory with a reduced size of its second layer that learns identity mapping (the same pattern is used for both of the input data and the supervisory data for the network) and is used for data compression of the multivariate data, then they show that it is valid as a tool for so-called 'piecewise linear factor analysis'. They demonstrate the advantages of the piecewise linear factor analysis method over the current linear scheme regarding the modeling of the unknown structure of multivariate data such as electric pulse distribution data generated by simulated partial discharge.<>