汇报时间：2019年4月15日（周一）16:00~18:00
汇报地点：yl23455永利中3-2321
汇报人：贠红光
短访交流简介：
本次出国访学前往加拿大英属哥伦比亚大学大学Zheng Liu老师课题组。研究课题为风力发电机组叶片的结冰检测。主要工作包括调研机器学习理论中迁移学习的研究现状，调研数据驱动的风电叶片结冰检测方法的研究现状。针对目前迁移学习理论中的不足之处，提出了一种自适应的迁移学习方法，并整理成一篇文章，已投稿IEEE Access期刊（SCI索引，三年平均IF2.68）。
汇报内容：
本次短访的研究课题为风力发电机组的叶片结冰检测。本文实际是以风电结冰检测作为载体，提出了一种自适应的迁移学习方法。迁移学习（Transfer learning）是机器学习中的一个分支，学习迁移是指一种学习对另一种学习的影响，或习得的经验对完成其他活动的影响。以神经网络为例，迁移学习旨在将一个经大量数据训练好的模型应用到其他学习任务中并得到优秀的结果。本文提出一种自适应的迁移学习方法，该方法能显著提高叶片结冰检测的Recall,并能提高整体模型的性能。
英文摘要
An adaptive approach for ice detection in wind turbine based on inductive transfer learning.
Ice-accretion on wind turbine blades will cause power degradation and threaten the system operating safety. The use of machine learning method offers a promising solution for blade ice detection. However, collecting a complete data set to train a separate model for each individual unit is costly and impossible. This paper proposes an adaptive inductive transfer learning method (ITL) to address this problem for blade icing detection. The inductive transfer learning aims to improve the detection performance by transferring knowledge from a well-established source. Generally, the inductive transfer learning requires a large amount of data from source domain and a small quantity of data from the target domain for training. As there is a distribution divergence for the source and target domain, most instance-transfer based learning methods can potentially reduce the prediction error by lowering the negative transfer or strengthening the “good” instance in the source domain. With the assumption that both the training and testing data in target domain are from the same distribution, the source-domain model prediction with the target training and testing data conforms to the same error distribution. An auxiliary classifier is then adopted to correct the prediction error with the training data. Finally, the source-domain model and auxiliary classifier are combined into a complete model. Experimental results on field ice-accretion data in wind turbine demonstrate that the proposed adaptive ITL framework can significantly improve the performance of the basic instance-transfer based model and is superior to the state-of-art inductive transfer learning methods.
