Risks and prevention in field activities with transformers: statistical modeling of the probability of accidents and fatalities

Abstract

This study analyzes electrical risks during transformer testing, a critical aspect in electrical engineering to ensure the safety of both equipment and personnel. Predictive models such as Decision Tree and Logistic Regression were evaluated, showing limited performance (23% and 29.1% overall accuracy, respectively). These models presented significant difficulties in correctly classifying minority classes, evidencing problems associated with data imbalance and a lack of sophistication to capture complex patterns in challenging environments. Furthermore, a bias toward majority categories was observed, compromising their practical applicability in critical systems such as substation protection or internal fault analysis using advanced techniques such as Frequency Response Analysis (FRA). The discussion highlights the need to improve these models through strategies such as oversampling (SMOTE), selection of more relevant features, and the implementation of advanced algorithms such as XGBoost or artificial neural networks. From a practical perspective, the importance of integrating theoretical knowledge with advanced machine learning tools to anticipate risk patterns and mitigate operational failures is emphasized. Recommendations include addressing class imbalance, performing more in-depth variable analyses, and complementing models with interpretation techniques such as SHAP or LIME. Finally, it is concluded that a multidisciplinary approach combining physical testing with predictive models is crucial to optimize safety in electrical systems.