Scholarlink

Detecting anomalies in wind turbine is crucial for extending equipment lifespan, reducing operational costs, and ensuring system reliability. However, due to the complexity of influencing factors, current methods struggle with precise identification, leading to room for improvement in detection accuracy. Hence, this paper proposes an anomaly detection approach based on error-aware Markov blanket and dual attention contrastive representation learning. Firstly, error-aware Markov blanket is used for feature selection from extensive operational states of wind turbine, uncovering primary causal relationships and effectively reducing dimensions. Building on this, the method utilizes contrastive learning with dual attention mechanism to obtain feature representations from both global and local views. It maximizes the consistency of feature representations from normal samples and the differentiation of feature representations from abnormal samples, thereby enhancing anomaly detection capability while reducing feature quantity. The accuracy and effectiveness of the proposed approach are demonstrated by real operational data from Irish wind turbine.

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 for Wind Turbine Based on Contrastive Learning and Causal Feature Selection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jie Song,
  • Yan Yang,
  • Yi Wu,
  • Ying Zhang,
  • Chang Liu,
  • Fujin Zhong,
  • Mengyuan Zhang,
  • Zhiwei Zhang

摘要

Detecting anomalies in wind turbine is crucial for extending equipment lifespan, reducing operational costs, and ensuring system reliability. However, due to the complexity of influencing factors, current methods struggle with precise identification, leading to room for improvement in detection accuracy. Hence, this paper proposes an anomaly detection approach based on error-aware Markov blanket and dual attention contrastive representation learning. Firstly, error-aware Markov blanket is used for feature selection from extensive operational states of wind turbine, uncovering primary causal relationships and effectively reducing dimensions. Building on this, the method utilizes contrastive learning with dual attention mechanism to obtain feature representations from both global and local views. It maximizes the consistency of feature representations from normal samples and the differentiation of feature representations from abnormal samples, thereby enhancing anomaly detection capability while reducing feature quantity. The accuracy and effectiveness of the proposed approach are demonstrated by real operational data from Irish wind turbine.