基于深度置信网络的风电机组主轴承故障诊断方法研究

摘要

为解决目前传统风电机组主轴故障诊断过程中因主轴结构复杂、信号非线性变化和机械大数据等因素引起的故障诊断困难问题,提出了一种高效、准确的风电机组主轴承故障诊断方法.利用深度置信网络强大的特征分层提取和泛化能力优势,结合Python语言基于TensorFlow学习框架,实现了高效准确的风电机组主轴承故障诊断.采集风电机组主轴轴转动频率,内、外圈故障频率,滚动体频率和保持架频率,对数据预处理并划分测试集和训练集,同时进行归一化处理.构建深度置信网络DBN诊断网络模型,确定网络层数、学习率、各层节点数等参数.输入训练样本逐层无监督训练达到局部参数最优,反向微调使整体性能最优并用测试集数据进行验证.试验结果表明:在网络参数合适且训练集和测试集相同的情况下,采用深度置信网络的风电机组主轴承故障诊断方法的准确率高达86.18%,同比优于传统支持向量机、人工神经网络故障诊断方法.%In order to solve the difficulty of fault diagnosis in traditional wind turbine spindle fault diagnosis methods due to complex spindle structure,signal nonlinearity and large mechanical data,an efficient and accurate fault diagnosis method for main bearing of wind turbines is proposed.By using the advantage of deep belief network,i.e.,the hierarchical feature extraction and generalization capabilities,and combining with Python language based on TensorFlow learning framework,the efficient and accurate fault diagnosis for main bearing of wind turbine is achieved.Firstly,the rotation frequency of the spindle shaft of the wind turbine,the frequency of the inner and outer ring faults,the frequency of the rolling body,and the frequency of the holder are collected,preprocessed,and divided into test set and the training set,then the normalization processing is performed.Secondly, deep belief network(DBN)diagnosis network model is constructed,and the parameters such as number of network layer,learning rate,and number of nodes at each layer,are determined.Finally,the training samples are input to the layer-by-layer unsupervised training to achieve the optimal local parameters.The reverse fine-tuning optimizes the overall performance and the test set data are used for verification.The test results show that when the appropriate network parameters are appropriate and training set and test set are the same,the accuracy of main bearing fault diagnosis method for wind turbines with deep belief network is up to 86.18%, which is better than that of traditional support vector machine and artificial neural network.