Abstract: To solve the problem of few samples and low recognition rate of the cracked state of wind turbine blades,a method for enhancing samples based on Generative Adversarial Networks (GAN) was proposed to improve the recognition rate.Taking the empirical risk classification model as an example,the imbalanced sample problem was deeply analyzed from the theoretical point of view.The GAN network model satisfying the generation and discrimination of cracking samples was designed,and the batch normalization was introduced to obtain samples with standard normal distribution,so that convergence of the network train process was accelerated.The proposed method was tested on 36 UCI benchmark data sets with neural network as the classifier and F1 value,Recall,Precision as the metrics.Results showed that the enhanced results were better.Furthermore,real wind turbine blade cracking data were used for experiments with logistic regression and neural network as classifiers,and the results showed that the enhanced results are improved by 13.88% and 8.20% respectively by comparing with original imbalanced samples.Compared with SMOTE and ADASYN algorithm,the two classifiers were improved by 74%,11% and 19%,23% respectively.

Key words: imbalanced sample, wind turbine blade, sample enhancement, generative adversarial networks, cracking identification

摘要： 针对风机叶片开裂状态样本少、识别率低的问题,提出基于生成式对抗网络(GAN)的开裂状态样本增强方法来提高识别率。以经验风险分类模型为对象,从理论角度对不平衡样本问题进行深入分析,设计了满足开裂样本生成与判别的GAN网络模型,引入批量归一化保障特征服从标准正态分布,加速网络训练过程收敛。以神经网络为分类器,以F1值、Recall、Precision为度量指标,在36个UCI基准数据集上对所提方法进行测试,结果表明增强后的结果更好。真实实验表明,以逻辑回归及神经网络为分类器,相比原始不平衡样本,增强后的结果分别提升13.88%,8.20%。与SMOTE算法对比,以上两种分类器的分类准确率分别提高74%和11%;与ADASYN算法对比,分类准确率分别提高19%和23%。

关键词: 不平衡样本, 风机叶片, 样本增强, 生成式对抗网络, 开裂识别