为解决风电齿轮箱状态监测数据样本量较少，特征指标间存在相互干扰且具有非线性难以分类等问题，本文提出了一种基于主成分分析结合支持向量机的风电齿轮箱故障诊断方法。首先，采用主成分分析法（PCA）对原始数据进行降维，做出第1，2主成分二维图及前3个主成分三维图，表明PCA对监测状态数据具有一定的分类效果。其次，提取累计贡献率80%以上的前5个主成分作为数据集。最后,采用支持向量机（SVM）比较4种不同核函数的诊断准确度，并加入噪声验证。分析结果表明：径向基核函数构建的支持向量机总体分类精度达到97%，准确率最高；在含噪的情况下，线性核函数与径向基核函数分类精度达到94%；与MLP神经网络进行对比发现，支持向量机更适应小样本分析且测试精度较高。实例分析表明，主成分分析结合支持向量机有较好的分类效果，适用于风电齿轮箱故障诊断的工程应用。

In order to solve many difficulties in the process of data analysis of wind turbine gearbox condition monitoring,such as small sample size,mutual interference and nonlinearity difficult to classify between characteristic indexes, this paper proposed a fault diagnosis method of wind turbine gearbox based on principal component analysis (PCA) and support vector machine(SVM). Firstly,PCA was used to reduce the dimension of the original data. The first and the second principal components were made in two dimensional maps and the first three principal components were made in the three dimensional maps,indicating that PCA had a certain classification effect on the monitoring condition data. Secondly,the first five principal components with cumulative contribution rate of more than 80% were extracted as new data sets. Finally,SVM was used to compare the diagnostic accuracy of four different kernel functions,and then added noise to verify. The analysis results show that the overall classification accuracy of radial basic function (RBF) kernel function based on SVM is 97%,which is the highest. In the case of noise,the classification accuracy of linear kernel function and RBF kernel function is 94%. Comparing with MLP neural network,it is shown that the SVM is more suitable for small sample size analysis and has higher test accuracy. The example analysis shows that PCA combined with SVM has a better classification effect and is suitable for engineering application of fault diagnosis of wind turbine gearbox.

TK221

广东省科技计划项目（2020B1212070022）;广州机械科学研究院有限公司博士后专项（17300065）