Data-Informed Method Selection for Predictive Maintenance: A Structured Framework Beyond Default AI Adoption

This seminar introduces a five-phase maintenance flowchart that makes the choice between non-data-driven (RCM/CBM/physics/rules) and data-driven (traditional methods & ML/AI) methods explicit and auditable for maintenance strategy adaptation in power systems. The framework gates on requirements (availability, safety, interpretability, latency) and data-worthiness (relevance, sufficiency, labels, drift) before any modeling, and it is operationalized by a Data-Attribute→Method mapping. Evidence from four case studies (AI4I, BatteryLife, EVBattery, Ellevio) shows why AI must not be the default: a shallow decision tree can be optimal in governed tabular settings; sequence models are justified only when temporal structure and supervision support them; and, when observability is weak, non-data-driven policies are the correct choice. I’ll conclude with two thesis directions: reliability-aware BESS optimization (integrating degradation, maintainability, and service readiness) and distribution-level capacity value (ELCC) under network constraints.