Preprocessing of information is an essential step for the effective design of machine learning applications. Feature construction and selection are powerful techniques used for this aim. In this paper, a feature selection and construction approach is presented for the detection of wind turbine generator heating faults. Data were collected from Supervisory Control and Data Acquisition (SCADA) system of a wind turbine. The original features directly collected from the data collection system consist of wind characteristics, operational data, temperature measurements and status information. In addition to these original features, new features were created in the feature construction step to obtain information that can be more powerful indications of the faults. After the construction of new features, a hybrid feature selection technique was implemented to find out the most relevant features in the overall set to increase the classification accuracy and decrease the computational burden. Feature selection step consists of filter and wrapper-based parts. Filter based feature selection was applied to exclude the features which are non-discriminative and wrapper-based method was used to determine the final features considering the redundancies and mutual relations amongst them. Artificial Neural Networks were used both in the detection phase and as the induction algorithm of the wrapper-based feature selection part. The results show that, the proposed approach contributes to the fault detection system to be more reliable especially in terms of reducing the number of false fault alarms.

翻译：信息预处理是机器学习应用有效设计中的关键步骤。特征构造与选择是实现该目标的强大技术。本文提出了一种用于风电机组发电机发热故障检测的特征选择与构造方法。数据采集自某风电机组的监控与数据采集系统。直接从数据采集系统获取的原始特征包括风特性、运行数据、温度测量值和状态信息。除这些原始特征外，在特征构造步骤中还生成了新特征，以获取更能有效指示故障的信息。生成新特征后，采用混合特征选择技术从整体特征集中筛选最相关特征，以提高分类精度并降低计算负担。特征选择步骤包含基于过滤器和基于封装器的两部分：基于过滤器的特征选择用于排除非判别性特征，而基于封装器的方法则通过考虑特征间的冗余性和相互关系来确定最终特征。人工神经网络既用于故障检测阶段，也作为基于封装器的特征选择部分的归纳算法。结果表明，所提方法有助于提高故障检测系统的可靠性，尤其在减少误报警次数方面效果显著。