Limitations of genetic programming applied to incipient fault detection: SFRA as example

This document deals with the application of genetic programming to the fault detection task, specifically with the power transformer fault detection problem of incipient faults. To this end we use genetic programming to obtain an highly approximated model of the a power transformer. The sweep frequency response analysis test represents the response of the transformer to a discrete variable frequency stimuli. We have been able to obtain a highly precision model which improves the precision of a commercial PG system. This result would be good if we only needed to identify the system. However, for the fault detection task, we should be able to identify the components within the transformer to assert where the fault has taken place. This is because the SFRA test when an incipient fault is present are similar but different as the fault advance. The tree generated for the model after the fault is evolved from the tree defining the power transformer model before the fault. Both trees are similar but the evolution seems to take place in a very specific random place. There is no way we can relate such changes with the physical model of the transformer. This shows the limitations of genetic programming to deal with this task and calls for extensions to the genetic programming paradigm or the merge of paradigms in order to deal with such task.