基于1D-CNN和SWLSTM的风电轴承故障诊断方法

WIND TURBINE ROLLING BEARING FAULT DIAGNOSIS METHOD BASED ON 1D-CNN AND SWLSTM

针对风电机组滚动轴承故障特征微弱,对应的信号具有非线性、非平稳性并含有噪声干扰的问题,以及信号本身具有空间和时域信息的特点,提出一种基于一维卷积神经网络(One-Dimensional Convolutional Neural Network,1DCNN)和共享权重长短时记忆网络(Shared Weight Long Short-Term Memory Network,SWLSTM)进行空时融合的风电机组滚动轴承故障诊断的卷积共享权重记忆网络(Convolutional Shared Weight Long Short-Term Memory Network,CSWLSTM)。使用美国西储大学滚动轴承数据集进行验证,相较于具有相同结构的卷积长短时记忆网络(Convolutional Long Short-Term Memory Network,CLSTM)模型和卷积门控循环网络(Convolutional Gated Recurrent Unit Network,CGRU)模型,CSWLSTM模型在训练时间上分别降低了39.9%和19.0%,模型参数量分别降低了63.3%和53.4%。在测试集上使用的分类评价指标准确率分别提升了1.0%和1.5%,精确率分别提升了1.0%和1.7%,召回率分别提升了0.9%和1.0%。仿真实验结果表明,所提出的CSWLSTM模型在风电机组滚动轴承故障诊断方面具有较好的应用潜力。

Aiming at the subtle fault features of the wind turbines rolling bearing,the fault signal is nonlinear,nonstationary and contains noise interference,and the fault signal has the characteristics of space and time feature information,a space-time fusion convolutional shared weight long short-term memory network(CSWLSTM)model based on one-dimensional convolutional neural network(1D-CNN)and the shared weight long short-term memory network(SWLSTM)was proposed for wind turbine rolling bearing fault diagnosis.Using the Western Reserve University rolling bearing dataset for experiment,compared with the convolutional long short-term memory network(CLSTM)model and convolutional gated recurrent unit network(CGRU)model with the same structure,CSWLSTM model had a significant improvement in the convergence of the training dataset.The training time was reduced by 39.9%and 19.0%,respectively.The model parameters were reduced by 63.3%and 53.4%,respectively.The accuracy was increased by 1.0%and 1.5%,the precision rate was increased by 1.0%and 1.7%,and the recall rate was increased by 0.9%and 1.0%on the test dataset,respectively.The simulation experiment results show that the CSWLSTM model has good application potential in the wind turbine rolling bearing fault diagnosis.