程建刚，毕凤荣，张立鹏，李鑫，杨晓，汤代杰.基于多重注意力-卷积神经网络-双向门控循环单元的机械故障诊断方法研究[J].内燃机工程,2021,42(4):77-83.
基于多重注意力-卷积神经网络-双向门控循环单元的机械故障诊断方法研究
Research on Mechanical Fault Diagnosis Method Based on Multiple Attention-Convolutional Neural Networks-Bidirectional Gated Recurrent Unit
DOI：10.13949/j.cnki.nrjgc.2021.04.011
关键词:注意力  故障诊断  多重注意力-卷积神经网络-双向门控循环单元（MA-CNN-BiGRU）  端到端
Key Words:attention  fault diagnosis  multiple attention-convolutional neural networks-bidirectional gated recurrent unit(MA-CNN-BiGRU)  end to end
基金项目:天津市自然科学基金项目（18JCYBJC20000）
作者单位
程建刚，毕凤荣，张立鹏，李鑫，杨晓，汤代杰 1.天津大学 内燃机燃烧学国家重点实验室,天津 300072；2.天津内燃机研究所，天津 300072 
摘要点击次数: 3080
全文下载次数: 1194
摘要:为解决传统机械故障诊断方法需要人工提取特征的不足，提出一种基于多重注意力-卷积神经网络-双向门控循环单元（multiple attention-convolutional neural networks bidirectional gated recurrent unit， MA-CNN-BiGRU）的端到端的故障诊断方法。首先将原始时域数据输入卷积神经网络（convolutional meural networks, CNN）进行初步特征提取并降维，然后将结果重组输入双向门控循环单元（bidirectional gated recurrent unit, BiGRU），可以有效地解决BiGRU对于过长序列数据处理困难的问题。采用美国凯斯西储大学开源轴承数据集进行训练，确定了最佳卷积层数和最佳样本长度约减比例分别为2和1/8。同时，通过在CNN和BiGRU中分别加入卷积注意力模块（convolutional block attention module, CBAM）和序列注意力模块（sequence attention module, SAM），进一步加强了模型对于关键信息的提取能力。最后实测柴油机故障振动信号试验表明：MA-CNN-BiGRU模型可以实现端到端的故障诊断，与变分模态分解(variational mode decomposition, VMD)〖CD*2〗核模糊C〖CD*2〗均值聚类算法（VMD-kernel fuzzy C-means clustering, VMD-KFCM）、VMD-反向传播神经网络（back propagation neural network, BPNN） 和一维CNN等方法进行对比，MA-CNN-BiGRU模型的故障诊断性能更优。
Abstract:To solve the shortcomings of traditional mechanical fault diagnosis methods that requires manual extraction of features, an end to end fault diagnosis method based on multiple attention-convolutional neural networks-bidirectional gated recurrent unit（MA-CNN-BiGRU） was proposed. First, the original time domain data was input into convolutional neural networks（CNN） to extract feature and reduce dimensionality preliminary, and then the results were reorganized into bidirectional gated recurrent unit（BiGRU）, which could effectively solve the problem that BiGRU was difficult to deal with long sequence data. The open-source bearing data set of Case Western Reserve University in the United States was used for training, and the optimal number of convolutional layers and the optimal sample length reduction ratio were determined to be 2 and 1/8 respectively. At the same time, by adding convolutional block attention module（CBAM） and sequence attention module（SAM） to CNN and BiGRU respectively, the models ability to capture key information is further strengthened. Finally, the actual measurement results of the diesel engine fault vibration signal experiment show that the MA-CNN-BiGRU model can achieve end-to-end fault diagnosis, and compared with the variational mode decomposition-kernel fuzzy C-means clustering algorithm（VMD-KFCM）, VMD-BP neural networks（VMD-BPNN） and one-dimensional CNN, the fault diagnosis performance of the MA-CNN-BiGRU is better.
查看全文  HTML   查看/发表评论