# Graph neural networks and belief rule base collaborative modeling for automated and interpretable fault diagnosis in proton exchange membrane fuel cells

## TL;DR

This paper introduces a new method combining graph neural networks and belief rule bases to improve fault diagnosis in fuel cells, making it more accurate and easier to understand.

## Contribution

The paper introduces a GNN-BRB framework that reduces reliance on expert input and enhances interpretability in fault diagnosis.

## Key findings

- The proposed GNN-BRB framework significantly reduces manual expert input in fault diagnosis.
- The method achieves superior diagnostic performance on real-world PEMFC data.
- Ablation studies confirm the effectiveness of each model component.

## Abstract

Proton exchange membrane fuel cells (PEMFC) are critical for clean energy conversion, but their reliability is severely compromised by complex faults, creating a pressing need for accurate and interpretable diagnostic methods. While the Belief Rule Base (BRB) provides a transparent reasoning framework, its practical deployment faces two fundamental challenges: the “combinatorial explosion” of rules with increasing system complexity, and a heavy reliance on domain experts to provide precise quantitative parameters. To address these issues, this paper proposes a novel GNN-BRB framework that synergistically integrates Graph Neural Networks (GNN) with BRB. Our solution introduces two key innovations: an exponential ordered weighting operator to systematically convert qualitative expert rankings into quantitative confidence parameters, and a GNN-based mechanism that models the BRB rule base as a graph to automatically generate initial parameters through information propagation among semantically related rules. Experimental results on a real-world PEMFC fault diagnosis case demonstrate that the proposed method significantly reduces dependency on manual expert input while achieving superior diagnostic performance. Ablation studies further validate the contribution of each model component. This work establishes a new paradigm for developing automated, highly accurate, and interpretable fault diagnosis systems for complex engineering applications.