A methodology for real-time health monitoring and management optimization of Substation switching equipment using interpretable machine learning

Substation (SEMD), focusing on the 500 kV and 220 kV sectors, which are configured in oneand-a-half circuit breaker schemes associated with two transformers. The objective is to evaluate improvements to current preventive maintenance policies and establish the conceptual groundwork for transitioning to condition-based maintenance through the estimation of an asset health index. The datasets comprise historical corrective, preventive, operational, and component‑status records for primary equipment. By integrating these data sources, the actual forced downtime of switching equipment over its lifecycle is derived, enabling a more precise assessment of equipment condition and operational performance. Two machine learning methods are evaluated: the long short-term memory (LSTM) neural network and the random forest (RF) ensemble model. Both algorithms are trained, followed by hyperparameter optimization and performance comparison. Evaluation based on the F1‑score and confusion matrices shows that the RF model performs best for the disconnector dataset, whereas the

LSTM model achieves superior results for the circuit breaker dataset. In each case, the selected model is applied to new data to estimate potential failure events, and a health index is calculated.

Interpretability analyses using Explainable Artificial Intelligence (XAI) techniques are also conducted. The results support the enhancement of existing preventive maintenance policies and establish the foundational concepts for implementing condition-based maintenance. The proposed methodology demonstrates applicability across different algorithms. Finally, the analysis reveals distinct yet complementary learning patterns among the models, indicating that hybrid multimodal fusion architectures should be a primary focus of future research.