Abstract: As the essential nuclear measurement equipment in new generation nuclear power plants, the self-powered neutron detector (SPND) plays a crucial role in ensuring the safe operation of reactors. The existing fault detection methods focus on time-domain analysis to build data-driven models, without leveraging the spatial coupling relationship of neutron flux in the reactor core. Therefore, an in-core SPND fault detection and isolation method integrating spatial-temporal information is proposed. First, the spatial-temporal graph data for SPND fault detection are established by combining SPND data with the layout of detector components within the reactor. Then, a real-time SPND fault detection model is designed using the graph convolution network-gate recurrent unit (GCN-GRU) and fault isolation (FI) strategy. Finally, using historical data and simulated fault samples from a pressurized water reactor, case analysis demonstrates that the method effectively fuses the spatial-temporal joint information of the overall SPNDs to reconstruct the current signals of individual SPNDs. The method can accurately detect and isolate faulty SPNDs, which exhibits higher accuracy and universality.