In the safety monitoring systems of nuclear power plants, ex-core neutron detectors are essential. Existing fault detection methods for these detectors primarily focus on extracting temporal features and using fixed thresholds to identify faults. These methods do not fully leverage the spatial coupling relationships between detectors and lack flexibility. To address these limitations, this paper introduces a spatial-temporal dynamic detection model (STDDM) for fault detection in ex-core neutron detectors. The model comprises three components: a temporal convolutional network (TCN), a graph convolutional network (GCN), and dynamic thresholds. By combining the TCN and GCN, the model captures implicit spatial-temporal relationships between detectors to reconstruct detector signals. The residuals between the reconstructed and actual signals are then computed, with dynamic thresholds set based on the mean residual and the overall standard deviation of residuals across the reactor. This approach allows the model to adapt to varying reactor operating conditions. Tested with real data from a nuclear power plant, the STDDM not only provides accurate real-time signal reconstruction but also exhibits strong fault tolerance under various fault conditions, proving its effectiveness and practicality for fault detection in excore neutron detectors.