Axial Flux Permanent Magnet Synchronous Generators (AFPMSGs) are critical in high power-density applications, making effective condition monitoring and fault diagnosis essential for operational continuity. Among various failure modes, interturn short-circuit (ITSC) faults are notoriously difficult to diagnose early. These defects are characterised by extremely weak initial signatures that are hard to detect, yet they tend to escalate rapidly into catastrophic failures. To overcome this early detection bottleneck, we introduce a Dual-Stream Graph Neural Network (DSGNN) fusion architecture. This framework integrates a convolutional feature extractor with two parallel Graph Neural Networks, enabling simultaneous modelling of spatial correlations among multi-source sensor channels and the temporal dependencies within the signals. An attention mechanism adaptively fuses these learned heterogeneous representations. Tested on an AFPMSG experimental platform, the DSGNN framework achieved a diagnostic accuracy of 89.26% for short-circuits as low as 1 to 3 turns in a single coil, significantly surpassing baseline models. This multi-source, structured data fusion framework provides a critical technical reference for detecting early mechanical and electrical defects in rotating machinery.
Gao et al. (Mon,) studied this question.