\r王太勇^{1, 2}，王廷虎¹，王 鹏¹，乔卉卉¹，徐明达\r¹\r
\r
AuthorsHTML:\r王太勇^{1, 2}，王廷虎¹，王 鹏¹，乔卉卉¹，徐明达\r¹\r
\r
AuthorsListE:\rWang Taiyong^{1, 2}，Wang Tinghu¹，Wang Peng¹，Qiao Huihui¹，Xu Mingda\r¹\r
\r
AuthorsHTMLE:\rWang Taiyong^{1, 2}，Wang Tinghu¹，Wang Peng¹，Qiao Huihui¹，Xu Mingda\r¹\r
\r
Unit:\r\r1. 天津大学机械工程学院，天津 300350；\r
\r\r2. 天津大学仁爱学院，天津 301636\r
\r
\r
Unit_EngLish:\r1. School of Mechanical Engineering，Tianjin University，Tianjin 300350，China；
2. Tianjin University Ren’ai College，Tianjin 301636，China\r
\r
Abstract_Chinese:\r\r状态监测与故障诊断是保证机械设备安全稳定运行的必要手段．本文提出一种基于注意力机制双向\rLSTM\r网络\r(\rABiLSTM\r)\r的深度学习框架用于机械设备智能故障诊断．首先，将传感器采集的设备原始数据进行预处理，并划分为训练样本集与测试样本集；其次，训练多个不同尺度的双向\rLSTM\r网络对原始时域信号进行特征提取，得到设备故障多尺度特征；再次，通过引入注意力机制，对不同双向\rLSTM\r网络提取特征的权重参数进行优化，筛选保留目标特征，滤除冗杂特征，以实现精准提取有效故障特征；最后，在输出端利用\rSoftmax\r分类器输出故障分类结果．通过利用发动机气缸振动实验数据和凯斯西储大学滚动轴承实验数据进行故障诊断实验，故障识别准确率均达到\r99\r%\r以上．实验结果表明，\rABiLSTM\r模型可以实现对原始时域信号的多尺度特征提取和故障诊断，通过与深度卷积网络\r(\rCNN\r)\r、深度去噪自编码器\r(\rDAE\r)\r和支持向量机\r(\rSVM\r)\r等方法进行对比，\rABiLSTM\r模型的故障识别性能优于各类常见模型．另外，通过利用凯斯西储大学滚动轴承在不同工况条件下的数据，对\rABiLSTM\r模型进行泛化性能实验，变工况样本的故障识别准确率仍然能够达到\r95\r%\r以上．\r\r
\r
Abstract_English:\r\rCondition monitoring and fault diagnosis are necessary means to ensure the safe and stable operation of mechanical equipment. A deep learning framework based on the attention-based bi-directional long and short-term memory\r(\rABiLSTM\r)\rnetwork is proposed for intelligent fault diagnosis of mechanical equipment. First\r，\rthe raw data collected by sensors were preprocessed and divided into the training and test sample sets. Second\r，\rbi-directional long and short-term memory\r(\rBiLSTM\r)\rnetworks of different scales were trained to extract multiscale data features from raw time-domain signals. Then\r，\rthe attention mechanism was introduced to optimize the weight parameters of different BiLSTM networks to extract the effective fault features accurately. Finally\r，\ra Softmax classifier was used to obtain the fault classification results. According to the experimental data of engine cylinder vibration and rolling bearing of Case Western Reserve University\r，\rthe accuracy of fault recognition is more than 99\r%\r. The experimental results show that the ABiLSTM model can extract multiscale features from the raw data and conduct fault diagnosis from raw time-domain signals. The fault recognition performance of the ABiLSTM model is superior to that of other common models\r，\rsuch as deep convolutional neural network\r，\rdenoizing autoencoder\r，\rand support vector machine. In addition\r，\raccording to the data of rolling bearing of Case Western Reserve University under different working conditions\r，\rthe accuracy of fault recognition can still reach more than 95\r%\r. The results of the generalization experiment show that the ABiLSTM model exhibits good fault recognition performance. The proposed ABiLSTM model can provide guidance for subsequent research and production practice.\r\r
\r
Keyword_Chinese:故障诊断；深度学习；双向长短期记忆网络；注意力机制\r

Keywords_English:fault diagnosis；deep learning；bi-directional long and short-term memory(BiLSTM) network；attention mechanism\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6468