基于CEEMDAN和CNN-TSA-GRU的滚动轴承故障识别方法研究

Research on Rolling Bearing Fault Identification Method Based on CEEMDAN and CNN-TSA-GRU

为避免复杂噪声对滚动轴承智能诊断模型的准确率干扰,提出一种基于完全集合经验模态分解CEEMDAN和深度时间自注意力卷积网络CNN-TSA的滚动轴承故障识别模型。该模型首先采用CEEMDAN将信号分解为若干固有模态函数(intrinsic mode function,IMF)分量,利用光谱放大因子SAF指标自适应筛选最优高信噪比分量;其次采用改进时间自注意力机制对数据分配权重并采用卷积神经网络CNN提取空间特征,弱化冗余特征信息,保留目标特征;最后利用门控循环单元GRU提取样本数据时间特征,使得网络得到更充分的学习,提高模型鲁棒性。经试验数据验证:所提出的深度学习智能故障识别模型故障识别准确率达到98.87%;对比一维CNN和CNN-LSTM模型,识别准确率分别提高9.15%和8.86%,验证了该模型的有效性和优越性。

In order to avoid the interference of complex noise on the accuracy of intelligent diagnosis model of rolling bearing,a rolling bearing fault recognition model based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)and convolutional neural network-time self-attention(CNN-TSA)was proposed.Firstly,CEEMDAN was used to decompose the signal into several intrinsic mode function(IMF)components,and spectral amplification factor(SAF)was used to self-adaptively select the optimal high signal-to-noise ratio components.Secondly,an improved temporal self-attention mechanism was used to assign weight to the data and CNN was used to extract spatial features,weakening redundant feature information and retaining target features.Finally,gated recurrent unit(GRU)was used to extract the time features of sample data,so that the network can be more fully learned and the robustness of the model can be improved.The results show that the proposed deep learning intelligent fault recognition model can achieve rolling bearing fault recognition with an accuracy of 98.87%.Comparing the 1D CNN and CNN-LSTM models,the recognition accuracy can be improved by 9.15%and 8.86%,respectively,which verifies the effectiveness and superiority of the model.