基于多通道融合多尺度自适应残差学习的行星齿轮箱故障诊断研究

Fault diagnosis of planetary gearbox based on multi-channel fusion and multi-scale adaptive residual learning

针对风电机组行星齿轮箱振动激励源多、故障诊断精度低的问题,提出了一种基于多通道融合多尺度动态自适应残差学习(MC-MSDARL)的行星齿轮箱故障诊断方法。首先,采用多尺度动态自适应卷积神经网络(MSDAC)对不同尺度卷积核权重进行了动态调整,自适应提取了单通道数据的局部和全局特征;其次,通过将MSDAC与残差学习结合,提升了模型的学习能力;最后,采用MC-MSDAR将多通道数据的多尺度特征进行了融合,输入到SoftMax层,实现了故障识别与分类。研究结果表明:基于MC-MSDAR的方法进行行星齿轮箱故障诊断的准确率为97%,验证了该方法的有效性;通过与其他深度学习方法进行对比,该方法具有更好的泛化能力。

Aiming at the problems of low accuracy of fault diagnosis for the planetary gearbox of wind turbine and its multiple sources of vibration excitation,a method used to diagnosis the detect of the planetary gearbox was proposed,it was based on multi-channel fusion and multi-scale dynamic adaptive residual learning(MC-MSDARL).Firstly,the proposed multi-scale dynamic adaptive convolution neural networks(MSDAC)was used to dynamically adjust the weights of convolution kernels at different scales to adaptively extract the local and global intrinsic features of single channel data.Secondly,in order to improve the learning ability of the model,the method combined MSDAC with residual learning.Finally,MC-MSDAR was used to fuse the multi-scale features of multi-channel data into a feature vector,and then it was inputted to SoftMax layer to achieve the identification and classification of the fault which was in the planetary gearbox.The research results show that the accuracy of fault diagnosis of planetary gearbox based on MC-MSDAR is 97%,which verifies the effectiveness of this method.When the results of MC-MSDAR is compared with the results implemented by other deep learning methods,the proposed MC-MSDAR has a better performance on generalization ability than other deep learning methods.