Modular and Compact Neural Network Framework for Internal Fault Detection in Generators Using Current Signature Data

One of the primary drawbacks of neural networks is that, despite usually offering superior accuracy compared to classical machine learning methods, their interpretability is often limited or even non-existent. It is easier to understand the decision-making process in support vector machine and k-nearest neighbors, for example. To contribute so that knowledge about the problem to be solved is directly incorporated into the method that seeks solutions, we propose this study. A neural network model based on decision trees is proposed to improve the prediction of a recurrent issue in power systems specifically designed to detect and classify electrical internal generator faults. Unlike conventional methods, the proposed model offers distinct advantages in terms of modularity, versatility, and interpretability, making it particularly suitable for specific applications. Validation results support these benefits, showing no evidence of a decline in prediction accuracy because of its tree-based structure.