基于CNN-GRU网络的轴承故障检测算法

Bearing fault detection algorithm based on CNN-GRU network

传统的卷积神经网络(Convolutional Neural Networks,CNN)采用的激活函数sigmoid,tanh计算量大,容易出现梯度消失和过拟合现象,ReLU函数容易导致均值偏移,且在轴承故障诊断中,故障数据一般为时序信号。为了解决CNN网络中激活函数存在的问题,同时提高对故障数据的时序特征提取能力,将CNN与门控循环单元(Gated Recurrent Unit,GRU)相结合,提出了一种基于eLU激活函数的CNN-GRU网络模型,用于轴承故障诊断。用凯斯西储大学轴承故障数据集对该网络进行验证,预测准确率达到99.93%,相较于其他算法更具优越性。

Traditional convolutional neural network adopts sigmoid function or tanh function as activation function,which cause gradient disappears and over-fitting.And ReLU function often leads to mean shift.In rolling bearing fault diagnosis,fault data is usually time series signals.In order to solve the problem about activation function and improve the ability of extract the timing features,a CNN-GRU(Gated Recurrent Unit,GRU)model based on ELU activation function is proposed for bearing fault diagnosis by combining CNN with GRU.The network is verified by the bearing fault data set of Case Western Reserve University.The prediction accuracy reaches 99.93%,which is superior to other algorithms.