A threshold based governance approach to balancing reliability and sustainability in predictive maintenance

The transition to Industry 5.0 requires industrial systems that are not only reliable but also sustainable. Predictive maintenance (PdM), a cornerstone of such systems, typically optimizes for technical accuracy alone, treating sustainability impacts—energy consumption, cost, and carbon emissions—as external considerations rather than integrated objectives. Maintenance practitioners lack practical tools to answer strategic questions such as: “What is the carbon cost of improving failure detection by 5%?” This paper proposes a threshold-based governance framework that quantifies the relationship between technical performance (F1-score) and sustainability metrics without requiring model retraining. Using a Random Forest classifier on the NASA C-MAPSS turbofan dataset as proof-of-concept, 30 classification thresholds (0.3–0.9) generate a Pareto frontier of non-dominated solutions. A utility function with strategic priority coefficients (λ) translates high-level goals into specific operating points. The analysis reveals a dual-regime trade-off: an initial “win-win” region (F1: 0.458→0.614) where improving reliability is associated with reduced environmental impact, followed by a regime of diminishing returns (F1: 0.614→0.695) where marginal reliability gains correspond to increasingly rapid increases in sustainability costs. The observed inflection point (F1 ≈ 0.614) provides an evidence-based stopping rule for optimization. Cross-validation, bootstrapping, and sensitivity analyses were performed to assess the robustness of the dual-regime structure across different data splits and model choices. The qualitative trade-off structure was partially preserved; full results are provided in Supplementary Materials. The framework closes a practical governance gap by making the sustainability-reliability trade-off visible, quantifiable, and navigable, equipping practitioners to align PdM operations with strategic ESG goals without complex algorithmic changes. Limitations include single dataset, one model class, and illustrative impact values requiring domain-specific calibration.