Considering the increase in the deployment of wind energy conversion systems, improving the coexistence between wind turbines and wildlife with an efficient method for blade impact assessment is of primary importance. Automated blade-impacts on potential wildlife monitoring can support the development and operations of wind farms. A substantial challenge is represented by the typical case of impacts vibrations signature embedded in the operational vibrations of the wind turbine. A heterogeneous multisensor system for automatic eagle detection and deterrent, including an automatic blade-event detection module, was developed providing the necessary field data. An automated blade event detection system, based on support vector machine, a form of machine learning, was developed and tested. Training of the algorithm was performed using features extracted from vibration signals and energy distribution graphs obtained from numerical simulations of blade impacts. Performance of the method, evaluated using numerical simulations at different levels of signal-to-noise ratios, relative to artificial impacts, showed the best results when trained using combined raw vibration signal and time marginal integration graphs, exhibiting an overall accuracy of 93% at . The proposed model was tailored for improving specificity (i.e., false negative error), a critical aspect for endangered species events. Performance of the trained algorithm evaluating field data exhibited an improvement in impact detection from a visually identifiable rate of 42% to true positive prediction rate of 75%. The system could perform, with appropriate training, diverse functions as components health monitoring or lighting strike automatic monitoring.

中文翻译：

---

基于支持向量机的风力机事件检测

考虑到风能转换系统部署的增加，通过有效的叶片影响评估方法改善风力涡轮机和野生动物之间的共存至关重要。对潜在野生动物监测的自动化叶片影响可以支持风电场的开发和运营。嵌入在风力涡轮机运行振动中的冲击振动特征的典型案例代表了一个重大挑战。开发了一种用于自动鹰检测和威慑的异构多传感器系统，包括自动叶片事件检测模块，可提供必要的现场数据。开发并测试了基于支持向量机（一种机器学习形式）的自动叶片事件检测系统。算法的训练是使用从振动信号中提取的特征和从叶片冲击的数值模拟中获得的能量分布图进行的。该方法的性能，使用不同信噪比水平的数值模拟评估，相对于人工影响，当使用组合的原始振动信号和时间边际积分图进行训练时，显示出最佳结果，表现出整体93% 的准确率。提议的模型是为提高特异性（即假阴性错误）而定制的，这是濒危物种事件的一个关键方面。评估现场数据的训练算法的性能表现出影响检测的改进，从视觉识别率 42% 到 75% 的真阳性预测率。通过适当的培训，该系统可以执行多种功能，如组件健康监测或雷击自动监测。