基于注意力机制的三流卷积神经网络在碰摩故障检测中的应用  

Application of three-stream convolutional neural network based on attention mechanism in rubbing fault detection

作　　者：康晓凤[1] 鲍蓉[1] 施汉琴 厉丹[1] KANG Xiaofeng;BAO Rong;SHI Hanqin;LI Dan(School of Information Engineering(School of Big Data),Xuzhou University of Technology,Xuzhou 221018,China)

机构地区：[1]徐州工程学院信息工程学院(大数据学院),江苏徐州221018

出　　处：《扬州大学学报（自然科学版）》2023年第2期29-33,58,共6页Journal of Yangzhou University：Natural Science Edition

基　　金：国家自然科学基金资助项目(62102344);江苏省“333高层次人才培养工程”资助项目;徐州科技计划资助项目(KC21303)。

摘　　要：为提高机械碰摩故障的早期诊断率,提出一种基于残差连接和注意力机制的三流卷积神经网络检测系统.系统中的每一流均基于卷积神经网络结构,分别从短时傅里叶变换频谱、梅尔频率倒谱系数和节奏图中提取特征,然后利用注意力机制获取最有效信息.实验结果表明,该模型的整体性能良好,其平衡正确率与传统分类方法和单流卷积神经网络相比均有较大提高.In order to detect and identify the early rub-impact fault,a three-stream convolutional neural network detection system based on residual connection and attention mechanism is proposed for mechanical fault detection.Each stream of the model is based on the convolutional recurrent neural network structure framework,and features are extracted from short-time Fourier transform spectrum,Mel-frequency cepstral coefficient and rhythm map respectively,and then the attention mechanism is used to obtain the most effective information.Experimental results show that the overall performance of the proposed model is good,and the balance accuracy is much improved compared with the traditional classification method and the single-stream convolutional neural network.

关 键 词：卷积神经网络 注意力机制 声发射 故障检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]