Towards Transparent Operations and Sustainment: A Conceptual Framework for Causal Interpretable Machine Learning Models for System Health Prognostics and Maintenance

A significant portion of system life cycle costs incur during operations and sustainment (O&S). Additionally, most O&S decisions are made in a reactionary manner, resulting in prolonged system downtime, capability gaps, and supply chain issues. While health prognostics and management (PHM) techniques can enable proactive and data-driven O&S decisions, these greatly rely on black box methods that lack transparency and are fundamentally based on statistical correlations instead of causality. Thus, PHM does not allow to investigate why observed relationships are occurring about a system’s health state. To that end, this paper proposes a conceptual framework for infusing causal interpretability into PHM by (i) preparing the system, either via architecting it accordingly during development or, for legacy systems, via smart postproduction design change decisions, and (ii) integrating traditional PHM algorithms with intrinsic explainability models that incorporate model-based systems engineering (MBSE) models. We suggest that PHM algorithms with intrinsic interpretability can learn from causal formal and functional relationships engrained in MBSE models, which then can be used to explain the causality behind a system’s degradation. While at its infancy, the ideas discussed in this paper can potentially enable more proactive and intelligent decision-making for O&S of systems and portfolios.