Fault Diagnosis Method for Nuclear Power Primary Loop Systems Based on Inception-ResNet-GEMA

The safety of the nuclear power plant’s primary circuit directly impacts the stable operation of the facility and personnel safety, making fault diagnosis within this system of significant research and practical value.To address the challenges of multi-scale feature complexity and fault type overlap in nuclear power primary loop system fault diagnosis, this paper proposes a fault diagnosis method that integrates the Inception-ResNet architecture with an enhanced Exponential Moving Average (EMA) attention mechanism. This method first utilizes the parallel convolutional layers of the Inception network to extract multi-scale fault features. It then enhances nonlinear feature extraction capabilities through the residual connection mechanism of ResNet, effectively capturing deep-level fault information. Subsequently, an improved EMA attention module is introduced to dynamically optimize feature weights without significantly increasing parameter complexity. This approach markedly enhances the saliency of key features while suppressing redundant information. To optimize the model training process, a StepLR learning rate scheduling strategy accelerates convergence. The model’s generalization capability and robustness are validated and evaluated using experimental datasets. Simulation results demonstrate that this method achieves a classification accuracy of 0.985 in multi-class fault diagnosis tasks for nuclear power primary circuits, verifying its high-precision diagnostic performance within complex nuclear power systems.