Ph.D. Proposal

Shraddha Meda Sheshadri

(Faculty Advisor: Professor Dimitri Mavris)

"A Heterogenous Graph-Based Framework for Aircraft Systems Prognostics and Health Management"

Tuesday, May 26

2:00 p.m.

Weber, CoVE

Abstract: 

Aircraft maintenance costs and safety concerns continue to rise due to increasing fleet complexity, aging aircraft, and the operational impact of unplanned failures and diagnostic uncertainty. Aviation has historically relied on scheduled maintenance, where components are serviced at fixed intervals regardless of their actual condition. With the growth of onboard sensing and monitoring capabilities, the industry has shifted towards condition-based maintenance, continuously assessing system health through sensor data to detect and diagnose faults as they emerge. The next evolution of this paradigm is prognostics: leveraging operational sensor data to predict degradation trajectories and estimate Remaining Useful Life (RUL). This enables maintenance actions to be performed proactively before failures occur. A vast body of research has developed prognostic techniques spanning physics-based, data-driven, and hybrid approaches, achieving strong prediction accuracy on benchmark datasets. However, modern aircraft operate as tightly coupled systems in which degradation in one subsystem can propagate through physical interactions to affect others, and existing methods largely model components independently, failing to capture these interdependent degradation dynamics. This limits both prediction robustness and interpretability, preventing component-level reasoning required for operational maintenance decision support.

Graph-based representations offer a natural solution; fault trees and reliability block diagrams have long been used in industry to model system-level interactions. However, these approaches are static, they encode fixed structural relationships but cannot incorporate the continuous stream of sensor data available on modern aircraft and adapt to evolving degradation dynamics. Graph Neural Networks (GNNs) extend graph-based modeling into the temporal and data-driven domain by jointly learning system interactions and degradation behavior. GNN prediction effectiveness is heavily dependent on the underlying graph structure. Existing approaches often rely on statistical sensor correlations, treat all nodes as semantically identical regardless of whether they represent sensors or physical components, and use static or unconstrained edge relationships that fail to capture evolving fault propagation pathways.

This thesis proposes a Heterogeneous Graph Neural Network (H-GNN) framework for aircraft systems health management that addresses these limitations through a physically grounded and interpretable architecture. The framework encodes known physical system architecture as a learnable graph, distinguishing sensors and physical components through a typed heterogeneous node schema, and dynamically adapting edge weights to capture evolving component influence during degradation progression. To support component-level RUL estimation and maintenance decision making, the framework proposes jointly modeling heterogeneous temporal degradation dynamics using an intervention-aware reset mechanism for maintenance simulation while simultaneously producing aleatoric and epistemic uncertainty estimates for each component. The proposed system operates entirely on existing onboard sensor data, requires no hardware modifications, and can generalize across aircraft configurations by updating the physical system prior. This work advances aircraft prognostics from accurate but opaque predictions toward interpretable, uncertainty-aware health management suitable for next-generation maintenance decision systems.

Committee:

Dr. Dimitri Mavris (advisor), School of Aerospace Engineering
Dr. Kyriakos G. Vamvoudakis, School of Aerospace Engineering
Dr. Kai A. James, School of Aerospace Engineering
Dr. Evan Harrison, School of Aerospace Engineering