This project extends prior work on hybrid fault diagnosis for production machinery, in which physically grounded domain knowledge (order-domain, time-domain, and frequency-domain features) was integrated with a 1D-CNN-BiLSTM network to improve diagnostic robustness under variable operating speeds (published in IOP Conf. Ser.: Mater. Sci. Eng., vol. 1342, article 012065). Building on that published hybrid framework, this project develops a heterogeneous graph neural network (GNN) architecture in which multi-sensor vibration measurement channels and process-context information (e.g., conveyor speed) are represented as nodes in a heterogeneous graph. Attention-based message passing (GATConv, implemented in PyTorch Geometric) is used to learn cross-sensor relationships, with a soft physics-informed bias constraining attention weights toward physically meaningful sensor relationships. Perturbation-based faithfulness validation is used to assess whether the learned attention weights correspond to physically interpretable relationships, extending the interpretability analysis of the earlier hybrid framework. The project is validated using vibration data collected from a pharmaceutical carton labelling machine under multiple operating speeds and fault conditions. Training and evaluation of the graph-based models require GPU resources for iterative architecture development, cross-validation, and hyperparameter search. This work forms part of a PhD thesis on integrated production data systems for fault diagnosis. Main supervisor: Andreas Archenti: archenti@kth.se KTH Department of Production Engineering