Lightweight ResNet with Multi-scale Attention Fusion for Fault Diagnosis of Accelerator Power Systems

As a critical experimental facility, the particle accelerator requires high reliability, stability, and continuity during operation. Traditional fault detection methods rely on signal analysis and shutdown inspections, which are timeconsuming and dependent on expert knowledge. Existing machine learning approaches struggle with high-dimensional complex features and often perform poorly with limited data. This paper proposes a lightweight fault diagnosis model that integrates a multi-scale convolutional encoder and a gated attention fusion mechanism to efficiently identify complex multi-class faults in accelerator power systems. Based on multifeature time-series signals, the model extracts local and global features across different temporal scales using multi-scale convolutions, applies gated attention for adaptive feature fusion, and employs a lightweight ResNet backbone to capture deep discriminative features while maintaining low computational cost. The experiments are conducted using real-world data collected from the spallation neutron source at oak ridge national laboratory, and evaluated under a small-sample setting. Results show that the proposed model achieves high accuracy and F1-score with a small parameter size, significantly outperforming conventional CNN and RNN architectures. It demonstrates excellent fault identification performance and strong deployment potential for edge diagnostics in industrial equipment.