Balancing Performance and Fairness in Explainable AI for Anomaly Detection in Distributed Power Plants Monitoring

Ensemble machine learning methods significantly outperformed traditional approaches for detecting equipment failures in diesel generators, with LightGBM achieving an F1-score of 0.99, demonstrating that more sophisticated AI models can be more reliable for critical infrastructure monitoring.

The framework successfully minimized bias across different regional clusters, achieving a Disparate Impact Ratio of approximately 0.95, which shows that AI systems can be designed to treat different geographic regions fairly rather than systematically disadvantaging some areas over others.

SHAP explainability revealed that fuel consumption rate and daily runtime were the most important factors driving anomaly predictions, meaning operators received clear, actionable insights into why the AI flagged potential problems rather than receiving unexplained "black box" recommendations.

Addressing data imbalance through specialized resampling techniques was critical to the model's success, as real-world monitoring systems often have far fewer equipment failures than normal operations; this highlights that responsible AI requires careful data preparation to avoid misleading predictions.

The study demonstrates that performance, interpretability, and fairness are not mutually exclusive goals—by intentionally designing for all three simultaneously, the researchers created an AI system suitable for real-world deployment in critical infrastructure where stakeholders need to understand, trust, and equitably benefit from automated decision-making.