Ensuring the safety of nuclear power plants (NPPs) and preventing radioactive material leakage is critical. This requires continuous operational monitoring and timely anomaly detection to implement appropriate countermeasures. With advancements in modeling and computational power, data-driven methods trained on simulation data have become the primary research direction for anomaly detection and fault diagnosis in NPPs. Inspired by the success of the Transformer architecture in fields such as natural language processing, we propose a novel approach that employs a lightweight self-attention-based encoder to process time-series data, enabling the extraction of feature-space embeddings. Subsequently, feature representation for each fault condition is compressed into a class center based on intra-class similarity, significantly reducing the reliance on large quantities of training samples. By focusing on learning distinguishable embeddings and class centers in the feature space, rather than memorizing numerous individual samples, our method is suitable for scenarios with limited data. We evaluate the proposed method on eight typical faults within the modular high-temperature gas-cooled reactor (MHTGR), and the experimental results demonstrate that our approach outperforms several baseline models.