Abstract
Over the past decade, advances in sensing and information technologies have enabled industries to collect large amounts of data. Yet, decision-making often remains driven by the intuition of domain experts who rely on simplistic analyses and short-term considerations. This frequently leads to suboptimal decisions that fail to account for long-term effects, particularly in complex, interconnected systems. Current data-driven strategies typically focus on immediate objectives, overlooking relational structures and longer-term impacts. There is a growing need for more transparent, generalizable models that can simulate system behavior, reason about alternative future scenarios, and extrapolate to unseen conditions—capabilities that are essential for decision-making in Prognostics and Health Management (PHM). This research aims to advance reasoning and decision support in PHM through three novel contributions: (1) a physics-informed surrogate model for simulating rigid body interactions, enabling the exploration of "what-if" scenarios, (2) an object-centric visual reasoning model for dynamics prediction in sensor-limited environments, supporting visual inspection tasks, and (3) a neuro-symbolic framework for interpretable root-cause analysis in time series, improving diagnostic transparency and providing actionable insights.
Introduction
In industrial environments such as wind farms or manufacturing facilities, traditional Prognostics and Health Management (PHM) has primarily focused on predicting the Remaining Useful Life (RUL) or detecting anomalies. However, these "black-box" approaches often fail to provide the strategic depth required for complex decision-making. They lack the ability to perform counterfactual reasoning—answering "what-if" questions regarding alternative maintenance schedules or varying operational conditions. This research addresses the gap by developing models capable of simulating system behavior under unseen conditions and reasoning about long-term impacts through relational structures.
Methodology
The proposed framework integrates three distinct yet complementary approaches to enable robust neural counterfactual reasoning.
HOPNet: Physics-Informed Surrogate Modeling
A core contribution is the Higher-Order Physics-informed Network (HOPNet). Unlike standard graph-based models that rely on point-to-point connections, HOPNet utilizes Combinatorial Complexes (CCs). This topological representation allows for a hierarchical encoding of system components, including nodes, edges, faces, and entire objects. By implementing a message-passing framework that explicitly incorporates Newtonian principles, the model ensures physical consistency in energy and momentum transfer during simulations of rigid body interactions.
Object-Centric Visual Reasoning
To support dynamics prediction in environments where sensors are limited (e.g., drone-based inspections), an unsupervised vision-based model is employed. This model utilizes equivariant convolutional filters to disentangle object appearance from spatial location, facilitating stable tracking through occlusions and viewpoint changes.
Neuro-Symbolic Root-Cause Analysis
To bridge the gap between deep learning and human-interpretable logic, a neuro-symbolic framework is introduced. This component extracts high-level concepts from industrial time-series data and applies logical operators (AND, OR, NOT) to generate interpretable rules. This transparency is vital for safety-critical applications where engineers must understand the underlying cause of a predicted failure.
Results and Evaluation
The framework was validated using the MoVi-spheres, MoVi-A, and MoVi-B datasets for rigid body dynamics, as well as real-world industrial data from wind farms and hydropower plants.
The HOPNet model demonstrated significant improvements over state-of-the-art baselines, including MeshGraphNet and FIGNet. Key performance metrics indicated superior accuracy in predicting both translation and orientation of interacting objects. Furthermore, HOPNet achieved up to 50% longer stable rollout horizons, maintaining physical consistency over extended simulation periods.
Conclusion and Future Work
This research moves PHM beyond simple prediction toward a comprehensive simulation and reasoning toolset. By leveraging higher-order topologies and physics-informed inductive biases, the proposed models provide a more reliable foundation for "what-if" planning in industrial maintenance. Future work will focus on expanding these topological models to handle abstract interactions in sensor networks and improving visual disambiguation in highly cluttered environments.
Summary updated