Scholarlink

In nuclear power plants (NPPs), the time series data following a fault at different power levels exhibit non-independent and identically distributed (non-IID) characteristics. Due to the predominance of operational scenarios at 100% power level, data from other power levels are relatively scarce and difficult to obtain. When transferring a diagnostic model from the source domain (100% power level) to target domains with different power levels using transfer learning, the insufficient data in the target domain and the non-IID nature of the data make effective domain adaptation challenging. Consequently, the performance of the transferred diagnostic model deteriorates. This study aims to address this issue by introducing adaptive batch normalization (AdaBN) into long short-term memory (LSTM) networks through three strategies: applying AdaBN within LSTM layers, between LSTM layers, and collaboratively both within and between LSTM layers. Experimental results demonstrate that incorporating AdaBN significantly enhances the adaptability of transfer learning models. In particular, the collaborative application of AdaBN within and between LSTM layers effectively reduces the feature distribution discrepancy between the source and target domains.

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 Temporal Transfer Learning Model with Adaptive Batch Normalization for Fault Diagnosis of Nuclear Power Plants Under Multiple Power Levels

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wenhao Jia,
  • Zhanguo Ma,
  • Long Tian,
  • Dihao Zheng,
  • Ziyang Cui

摘要

In nuclear power plants (NPPs), the time series data following a fault at different power levels exhibit non-independent and identically distributed (non-IID) characteristics. Due to the predominance of operational scenarios at 100% power level, data from other power levels are relatively scarce and difficult to obtain. When transferring a diagnostic model from the source domain (100% power level) to target domains with different power levels using transfer learning, the insufficient data in the target domain and the non-IID nature of the data make effective domain adaptation challenging. Consequently, the performance of the transferred diagnostic model deteriorates. This study aims to address this issue by introducing adaptive batch normalization (AdaBN) into long short-term memory (LSTM) networks through three strategies: applying AdaBN within LSTM layers, between LSTM layers, and collaboratively both within and between LSTM layers. Experimental results demonstrate that incorporating AdaBN significantly enhances the adaptability of transfer learning models. In particular, the collaborative application of AdaBN within and between LSTM layers effectively reduces the feature distribution discrepancy between the source and target domains.