Online Equipment Fault Diagnosis
Ongoing

Physics-Informed Hypergraph Networks for Compound Fault Diagnosis and Traceability

Mapping physical topology into hypergraph priors, combined with a physics-informed multi-task network and uncertainty-driven cascaded inference, achieves 99.55% compound-fault diagnosis accuracy

Physics-Informed Network Hypergraph Learning Compound Fault Uncertainty Quantification
Physics-Informed Hypergraph Networks for Compound Fault Diagnosis and Traceability

Project Overview

High-order dynamic coupling in complex electromechanical systems makes features hard to disentangle; flat diagnostic pipelines lack mechanistic and cascaded reasoning, and black-box inference offers no risk measure. This project proposes a protocol that maps physical topology into hypergraph priors, explicitly representing high-order physical coupling across multiple components.

Research Objectives

1

Mapping bond graphs into high-order physical hypergraph priors

2

A multi-task diagnostic network regularized for physical consistency

3

An uncertainty-driven cascaded inference strategy

Methodology

A physics-informed hypergraph multi-task network aggregates interaction features under physical-consistency regularization; sample-level uncertainty dynamically governs diagnostic depth, balancing accuracy against inference efficiency.

Technical Approach: How It Works

  1. 1

    Physical Topology to Hypergraph Prior

    A protocol maps "bond graphs to high-order physical hypergraphs": the energy-flow topology of the electromechanical system is explicitly translated into a hypergraph structure, where a single hyperedge represents one group of multi-component coupling relations.

  2. 2

    Physics-Informed Hypergraph Multi-Task Network

    Message passing over the hypergraph aggregates interaction features, with physical-consistency regularization keeping the learned representation aligned with the system's physical state; multi-task heads output component-level and system-level diagnoses simultaneously.

  3. 3

    Uncertainty Quantification

    Each diagnostic output carries an uncertainty estimate quantifying "how confident the model is," providing a risk measure for safety-critical decisions.

  4. 4

    Uncertainty-Driven Cascaded Inference

    Easy samples take a lightweight fast path; only high-uncertainty samples are routed to deeper, fine-grained diagnosis — balancing accuracy and inference efficiency at a routing rate below 50%.

Figures: Methods & Results

Physics-informed hypergraph multi-task network architecture (Reliability Engineering & System Safety, 2026)
Physics-informed hypergraph multi-task network architecture (Reliability Engineering & System Safety, 2026)
t-SNE manifold of the latent feature space: feature distribution closely aligns with physical state, precisely isolating boundary high-uncertainty samples
t-SNE manifold of the latent feature space: feature distribution closely aligns with physical state, precisely isolating boundary high-uncertainty samples

Key Results

Compound-fault diagnosis accuracy of 99.55%

Cascaded strategy achieves selective inference at a routing rate below 50%

Feature manifolds align closely with physical states, precisely isolating high-uncertainty samples

Related Publications

Uncertainty-informed cascaded diagnosis of compound faults in electromechanical systems via a physics-informed hypergraph framework, Reliability Engineering & System Safety, 2026

Online Equipment Fault Diagnosis

Explore More Research Directions

Back to Online Equipment Fault Diagnosis All Research Areas