Scholarlink

<p>In this study, we propose a computationally efficient and accurate feature-engineered Machine Learning (ML) model for anomaly detection in Pressurized Water Reactor (PWR)-type Nuclear Power Plants (NPPs) to address the challenge of accurate and interpretable anomaly detection under data scarcity. Unlike existing approaches relying on large datasets and complex deep learning models, this work explores whether feature-driven models can achieve superior performance with limited data and improved interpretability. We evaluate several Deep Learning (DL) and ML models under multiple dimensionality-reduction strategies. Principal Component Analysis (PCA) preserves approximately 90% of the variance using eight components but provides limited class separability, achieving only 94.78% accuracy with all principal components. In contrast, feature selection based on Random Forest importance scores significantly improves model performance. Among the evaluated models, including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Transformer architectures, the Multi-Layer Perceptron (MLP) delivers the best results. The MLP achieves 94.297±1.323% accuracy over 10 independent runs using all features and 97.99% using only the top seven selected features, outperforming previously reported results on the same dataset and reducing training time per epoch from 46.79 to 40.41&#xa0;s, with memory usage decreasing to 23.61% of CPU and 32.41% of GPU consumption. Our findings demonstrate that a carefully designed, feature-engineered approach can outperform more complex models while improving computational efficiency and interpretability, making it highly suitable for deployment in safety-critical nuclear monitoring environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Comparative Study of Deep Learning Architectures for Anomaly Detection in PWRs: Achieving Optimal Accuracy, Efficiency, and Interpretability with a Feature-Engineered Multi-layer Perceptron

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Ihtesham Ibn Malek,
  • S. M. Shohorab Hossain Hasib,
  • Shakhawat Parvez

摘要

In this study, we propose a computationally efficient and accurate feature-engineered Machine Learning (ML) model for anomaly detection in Pressurized Water Reactor (PWR)-type Nuclear Power Plants (NPPs) to address the challenge of accurate and interpretable anomaly detection under data scarcity. Unlike existing approaches relying on large datasets and complex deep learning models, this work explores whether feature-driven models can achieve superior performance with limited data and improved interpretability. We evaluate several Deep Learning (DL) and ML models under multiple dimensionality-reduction strategies. Principal Component Analysis (PCA) preserves approximately 90% of the variance using eight components but provides limited class separability, achieving only 94.78% accuracy with all principal components. In contrast, feature selection based on Random Forest importance scores significantly improves model performance. Among the evaluated models, including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Transformer architectures, the Multi-Layer Perceptron (MLP) delivers the best results. The MLP achieves 94.297±1.323% accuracy over 10 independent runs using all features and 97.99% using only the top seven selected features, outperforming previously reported results on the same dataset and reducing training time per epoch from 46.79 to 40.41 s, with memory usage decreasing to 23.61% of CPU and 32.41% of GPU consumption. Our findings demonstrate that a carefully designed, feature-engineered approach can outperform more complex models while improving computational efficiency and interpretability, making it highly suitable for deployment in safety-critical nuclear monitoring environments.