Abstract

Graph-based fault diagnosis methods have shown strong capability in multi-sensor fusion by representing hidden inter- sensor  relationships  with  nodes  and  edges.  However,  two challenges   remain   in   complex   machinery   systems.   First, constructing  a  graph  that  fully  captures  system-level  fault characteristics is difficult, since fault features are subtle and widely  distributed  across  multiple  frequency  bands.  Second, sensor  signals  exhibit  varying  sensitivities  to  different  fault modes, and each fault mode shows distinct frequency-domain behaviors such as energy shifts or unique fault frequencies. To overcome these issues, this paper proposes a graph-based sensor 

fusion framework that jointly redesigns graph construction and graph representation. The proposed graph construction strategy decomposes sensor signals into multiple frequency bands, builds a band-specific subgraph for each band, and connects them to form a final graph that preserves band-specific fault features. 

The proposed representation network adopts a sensitivity-aware readout layer that captures sensor-wise importance, together with  an  energy-feature  embedding  mechanism  that  reflects energy-shift fault behaviors. The effectiveness of the proposed method is validated through its application to three different types of machinery: rotating machinery, a permanent magnet synchronous motor, and an urban air mobility system dataset. The  validation  results  show  that  the  proposed  method outperforms  conventional  and  state-of-the-art  graph-based models while offering enhanced explainability.