DyEdgeGAT: Dynamic Edge via Graph Attention for Early Fault Detection in IIoT Systems

Mengjie Zhao, Olga Fink

Abstract

In the Industrial Internet of Things (IIoT), condition monitoring sensor signals from complex systems often exhibit nonlinear and stochastic spatial-temporal dynamics under varying conditions. These complex dynamics make fault detection particularly challenging. While previous methods effectively model these dynamics, they often neglect the evolution of relationships between sensor signals. Undetected shifts in these relationships can lead to significant system failures. Furthermore, these methods frequently misidentify novel operating conditions as faults. Addressing these limitations, we propose DyEdgeGAT (Dynamic Edge via Graph Attention), a novel approach for early-stage fault detection in IIoT systems. DyEdgeGAT's primary innovation lies in a novel graph inference scheme for multivariate time series that tracks the evolution of relationships between time series, enabled by dynamic edge construction. Another key innovation of DyEdgeGAT is its ability to incorporate operating condition contexts into node dynamics modeling, enhancing its accuracy and robustness. We rigorously evaluated DyEdgeGAT using both a synthetic dataset, simulating varying levels of fault severity, and a real-world industrial-scale multiphase flow facility benchmark with diverse fault types under varying operating conditions and detection complexities. The results show that DyEdgeGAT significantly outperforms other baseline methods in fault detection, particularly in the early stages with low severity, and exhibits robust performance under novel operating conditions.

Introduction

In the Industrial Internet of Things (IIoT), detecting a fault at its earliest stage—known as an incipient fault—is critical for preventing catastrophic failures and extending component life. However, current monitoring systems face two major hurdles: the assumption of static relationships between sensors and the "novelty trap," where healthy but unseen operating conditions are misidentified as faults.

Comparison of DyEdgeGAT with existing approaches, highlighting its unique capability to model evolving relationships alongside spatial and temporal dynamics.

As shown in the comparison above, most existing Graph Neural Networks (GNNs) for time series assume fixed sensor relationships within a specific window. DyEdgeGAT addresses this by treating the "conversation" between sensors as a dynamic, evolving entity, allowing it to distinguish between a system that is failing and one that is simply operating in a new, healthy state.

Methodology

The DyEdgeGAT framework consists of five sophisticated components designed to capture the complex spatial-temporal patterns of industrial systems:

The five-component architecture of DyEdgeGAT, detailing the flow from raw sensor measurements to anomaly scoring through dynamic edge and node modeling.
  1. Dynamic Edge Construction: Uses a GATv2-based attention mechanism with temporal encoding to track "edge dynamics"—how relationship strengths between sensors fluctuate at every time step.
  2. Operating Condition-Aware Node Dynamics: Distinguishes between system-independent variables (control inputs) and system-dependent variables (sensor measurements). Control inputs serve as context to initialize node encoders, effectively informing the model of expected sensor behavior based on operator actions.
  3. Dynamic Interaction Modeling: Employs Graph Isomorphism Networks (GIN) on an aggregated temporal graph to capture complex spatial-temporal interdependencies.
  4. Reversed Signal Reconstruction: Reconstructs sensor signals in reverse order (Seq2Seq style) to better capture long-range causal dependencies that might be lost in standard forward passes.
  5. Temporal Topology-Informed Anomaly Scoring: Normalizes reconstruction error based on the node's importance (degree) in the dynamic graph, preventing naturally volatile sensors from triggering false alarms.

Experimental Evaluation

The model was evaluated on two distinct setups:

  • Synthetic Dataset: A custom dataset simulating control signals influencing measurement signals via nonlinear relationships. This allowed for precise control over fault severity by scaling signal interdependencies.
  • Pronto Benchmark Dataset: A real-world dataset from an industrial-scale multiphase flow facility. It includes 17 process variables and three types of induced faults (air leakage, air blockage, and diverted flow), as well as a "slugging" condition used to test robustness against novel operating regimes.

Key Results

DyEdgeGAT demonstrated superior performance in early fault detection. On the synthetic dataset, it maintained a high AUC (0.83) even at very low fault severities where traditional LSTM and FNN models struggled. In the Pronto dataset, it achieved an average AUC of 0.748 for the notoriously difficult "Diverted Flow" fault, where most baselines performed no better than random chance.

Crucially, the model showed high robustness to novelty. Using a custom ambiguity metric, DyEdgeGAT was far less likely to misidentify the "slugging" operating condition as a fault compared to high-performing baselines like Autoencoders (AE). Despite its complexity, the model remains computationally efficient with approximately 15k parameters, making it suitable for real-time edge deployment.

Conclusion and Future Work

DyEdgeGAT introduces a novel paradigm by focusing on edge dynamics as a primary indicator of system health. While the current model is purely data-driven, future work will explore the integration of physical priors and hierarchical structures to improve scalability for systems with thousands of sensors. Additionally, enhancing the model's performance in scenarios with limited control variable data remains a key area for development.

Summary updated

Abstract

In the Industrial Internet of Things (IIoT), condition monitoring sensor signals from complex systems often exhibit nonlinear and stochastic spatial-temporal dynamics under varying conditions. These complex dynamics make fault detection particularly challenging. While previous methods effectively model these dynamics, they often neglect the evolution of relationships between sensor signals. Undetected shifts in these relationships can lead to significant system failures. Furthermore, these methods frequently misidentify novel operating conditions as faults. Addressing these limitations, we propose DyEdgeGAT (Dynamic Edge via Graph Attention), a novel approach for early-stage fault detection in IIoT systems. DyEdgeGAT's primary innovation lies in a novel graph inference scheme for multivariate time series that tracks the evolution of relationships between time series, enabled by dynamic edge construction. Another key innovation of DyEdgeGAT is its ability to incorporate operating condition contexts into node dynamics modeling, enhancing its accuracy and robustness. We rigorously evaluated DyEdgeGAT using both a synthetic dataset, simulating varying levels of fault severity, and a real-world industrial-scale multiphase flow facility benchmark with diverse fault types under varying operating conditions and detection complexities. The results show that DyEdgeGAT significantly outperforms other baseline methods in fault detection, particularly in the early stages with low severity, and exhibits robust performance under novel operating conditions.
BibTeX
@article{zhao2024dyedgegat,
  author = {Mengjie Zhao and Olga Fink},
  title = {{DyEdgeGAT: Dynamic Edge via Graph Attention for Early Fault Detection in IIoT Systems}},
  journal = {IEEE Internet of Things Journal},
  year = {2024},
  doi = {10.1109/jiot.2024.3381002},
}
Figure