Scholarlink

<p>Wind turbine (WT) anomaly detection (AD) models often experience performance degradation under time-varying operating conditions. While incremental learning (IL) can address this issue, error accumulation caused by false negatives may result in over-fitting to noise during online updating, and time-triggered updating strategies lead to extra computational cost. The scientific aim of the work is to address the performance degradation of the AD models under the error accumulation and the computational cost brought by time-triggered updating. The subject of the research was to obtain a novel WT AD method based on the multivariate state estimation technique (MSET) and a robust event-triggered IL strategy. This study extends the existing research in constructing an instance-based AD model with MSET to alleviate catastrophic forgetting in IL, and designing a robust event-triggered IL strategy including a blocking module for potential false negatives and an event-triggered module based on relative entropy, which screens valid incremental data and triggers adaptive model updating, respectively. Two real-world WT datasets, including blade icing and gearbox faults, are used in the experiments. Results show that the blocking module enhances the detection accuracy significantly compared to not processing potential false negatives. The event-triggered module reduces computing time by over 50% compared to the time-triggered strategy, while maintaining high accuracy. Cross-algorithm experiments demonstrate the generalization of the proposed IL strategy, and the hyperparameter experiments are also provided to guide practical implementation.</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

Anomaly Detection of Wind Turbines Based on MSET and Robust Event-Triggered Incremental Learning

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Ziqi Wang,
  • Feng Yan,
  • Yibo Geng,
  • Zhuo Shen

摘要

Wind turbine (WT) anomaly detection (AD) models often experience performance degradation under time-varying operating conditions. While incremental learning (IL) can address this issue, error accumulation caused by false negatives may result in over-fitting to noise during online updating, and time-triggered updating strategies lead to extra computational cost. The scientific aim of the work is to address the performance degradation of the AD models under the error accumulation and the computational cost brought by time-triggered updating. The subject of the research was to obtain a novel WT AD method based on the multivariate state estimation technique (MSET) and a robust event-triggered IL strategy. This study extends the existing research in constructing an instance-based AD model with MSET to alleviate catastrophic forgetting in IL, and designing a robust event-triggered IL strategy including a blocking module for potential false negatives and an event-triggered module based on relative entropy, which screens valid incremental data and triggers adaptive model updating, respectively. Two real-world WT datasets, including blade icing and gearbox faults, are used in the experiments. Results show that the blocking module enhances the detection accuracy significantly compared to not processing potential false negatives. The event-triggered module reduces computing time by over 50% compared to the time-triggered strategy, while maintaining high accuracy. Cross-algorithm experiments demonstrate the generalization of the proposed IL strategy, and the hyperparameter experiments are also provided to guide practical implementation.