Determination of disturbance patterns using fault recorder data and regression trees

This paper proposes a new method to train a regression tree for disturbance classification based on long term records considering the most probable operational conditions obtained by historical data of the analyzed system. Unlike the usual, deterministic approach of simulating a set of predetermined conditions varying fault resistance and position, the decision making process is now supported by a decision tree that was trained taking into account a set of probable initial conditions and disturbance parameters. This makes the classifier more accurate to identify the most common shapes of records for particular classes of disturbance, resulting in an appropriate balance between the number of records being generated and sensitivity of the classifier to a particular type of perturbation. The database for training was created by a large set of simulations, generated randomly by the Monte Carlo method until a target value for the confidence interval in the classification of the variables is reached. Results obtained showed that simple indices, together with a comprehensive set of rules, lead to a highly accurate classifier and highlight relations between input quantities and operational conditions with their effects on the nature of the corresponding disturbance.