Reconstruction model is a powerful method for component condition monitoring and fault detection by considering the model prediction residuals. In this article, a new signal reconstruction modeling technique is proposed using support vector regression. Multiple indicators are calculated to recognize slight shift from normal condition, and detect the fault at an early stage. Input variables are selected based on correlation analysis and failure mode analysis. A sliding-time-window technique is employed to incorporate temporal information inherent in time-series data. Residuals between the observed signal and the reconstruction signal are utilized to indicate whether the desired quantity is different from its normal operation condition or not. Three statistical indicators (Deviation Index, Volatility Index and Significance Index) are defined to quantify the deviation level from normal condition to abnormal condition. Health index (HI) of a specific fault is derived from responsive statistical indicators, and the integral health index (integral-HI) of an entire component is composed of all individual health index. An experiment of real-life wind turbine high temperature fault detection scheme is studied. Results show that the proposed approach demonstrates improved performance in detecting wind turbine faults, and controlling false and missed alarms.

中文翻译：

---

基于机器学习重构模型的部件实时状态监测与故障检测

通过考虑模型预测残差，重建模型是一种用于组件状态监测和故障检测的强大方法。在本文中，使用支持向量回归提出了一种新的信号重建建模技术。计算多项指标，识别与正常情况的轻微偏差，并及早发现故障。基于相关分析和故障模式分析选择输入变量。采用滑动时间窗口技术来合并时间序列数据中固有的时间信息。观测信号和重建信号之间的残差用于指示期望量是否不同于其正常操作条件。三项统计指标（偏差指数、波动率指数和显着性指数）被定义为量化从正常情况到异常情况的偏差水平。特定故障的健康指数（HI）来源于响应性统计指标，整个部件的整体健康指数（integral-HI）由所有个体健康指数组成。研究了真实风电机组高温故障检测方案的实验。结果表明，所提出的方法在检测风力涡轮机故障以及控制误报和漏报方面具有改进的性能。研究了真实风电机组高温故障检测方案的实验。结果表明，所提出的方法在检测风力涡轮机故障以及控制误报和漏报方面具有改进的性能。研究了真实风电机组高温故障检测方案的实验。结果表明，所提出的方法在检测风力涡轮机故障以及控制误报和漏报方面具有改进的性能。