基于空洞卷积神经网络与注意力机制GRU的滚动轴承故障诊断

针对传统故障诊断方法需要人工提取特征的不足,以及大数据下滚动轴承故障振动信号自适应特征提取与智能诊断问题,利用空洞卷积神经网络(DCNN)可以在不增加计算量的基础上兼顾不同尺度空间特征的能力、门控循环单元(GRU)善于从动态变化的序列数据中学习到时间上的关联性的能力,提出了一种将DCNN、注意力机制和GRU多路径融合的端到端故障诊断方法。首先利用DCNN从原始数据中自动提取时序信号特征,然后将注意力机制(Attention)的GRU通路和DCNN通路进行融合,最后将提取到的特征融合之后送入分类层进行分类。试验结果表明,所提方法的诊断准确率平均为98.75%,高于比较方法,更加适用于滚动轴承故障诊断。

Fault diagnosis of rolling bearing based on DCNN and attention GRU algorithm

Aiming at the shortcomings of traditional fault diagnosis methods requiring manual feature extraction and the problems of adaptive feature extraction and intelligent diagnosis of rolling bearing fault vibration signals under big data, an end-to-end fault diagnosis method was proposed based on multi-path fusion of dilated convolutional neural network(DCNN), attention mechanism, and gated recurrent unit(GRU) by taking advantage of the ability of DCNN that could take into account spatial features at different scales without increasing computational effort and the ability of GRU to learn temporal correlations from dynamically changing sequence data. First, DCNN was used to automatically extract timing signal features from the original data. Then the GRU channel of the attention mechanism and the DCNN channel were fused, and finally the extracted features were fused and sent to the classification layer for classification. Experimental results show that the diagnosis accuracy of the proposed method is 98.75% on average, which is higher than the comparison methods and more suitable for rolling bearing fault diagnosis.