基于自适应升维和卷积自注意力的轴承故障诊断网络

Bearing fault diagnosis network based on adaptive dimension-increasing and convolutional self-attention

深度学习网络是一种端到端的黑箱模型,对网络进行可解释分析可以更深入理解网络内部运行机理以便合理优化网络结构和调参。目前基于Transformer网络的轴承故障诊断模型必须借助时频分析等方法对时域信号进行升维,使其转换为时频图等二维图像以便于可解释分析,这种方法存在着升维过程中参数固定、网络参数量过大和网络可解释性较差等缺点。针对以上问题提出融合网络内自适应升维方法、卷积自注意力模块(convolutional self-attention module)以及中间层类激活热力图(mid layer class activation map, ML-CAM)的卷积自注意力自适应升维网络(convolutional self-attention adaptive dimension-increasing network, CSADI-Net)。卷积自注意力模块使用卷积层获取特征图的查询(query, Q)、键(key, K)和值(value, V),极大地减少了可训练参数量;网络内自适应升维方法通过内部特征图拼接等操作将升维过程与网络训练融合,使其具有自适应参数调节的能力;中间层类激活热力图可以将网络内升维方法所得二维特征图上各个部位的关注度以热力图的形式展现,是一种直观的可视化可解释分析方法。此外,对CSADI-Net和ML-CAM进行了测试,CSADI-Net在凯斯西储大学轴承数据集上的准确率可以达到97.32±0.12%并且可以完全分类大连大学实测轴承故障数据集,同时使用ML-CAM对CSADI-Net在两数据集各样本上绘制了类激活热力图,解释了网络运行机理,证实CSADI-Net具有高准确率,高抗噪性能,可解释性强等优点。

Deep learning networks are an end-to-end black box model,and network interpretable analysis can provide a deeper understanding of network internal operating mechanism to reasonably optimize network structure and adjust parameters.At present,the bearing fault diagnosis model based on Transformer network must use methods of time-frequency analysis,etc.to do dimension increasing for time-domain signal and convert it into a 2-D image of time-frequency map,etc.for interpretable analysis.This method has drawbacks of fixed parameters,large number of network parameters,and poor network interpretability in dimension increasing process.Here,aiming at above problems,a convolutional self-attention adaptive dimension-increasing network(CSADI-Net)integrating adaptive dimension increasing method in network,convolutional self-attention module and mid layer class activation map(ML-CAM)was proposed.The convolutional self-attention module could use convolutional layers to obtain query(Q),key(K)and value(V)of feature map to greatly reduce the number of trainable parameters.The adaptive dimension increasing method in network could integrate dimension increasing process with network training through internal feature maps splicing,etc.to make it have the ability to adaptively adjust parameters.ML-CAM could display attention levels of various parts on 2-D feature map obtained with the network dimension increasing method in the form of a heatmap,it could be an intuitive,visual and interpretable analysis method.In addition,CSADI-Net and ML-CAM were tested.It was shown that the accuracy of CSADI-Net on the bearing dataset of Case Western Reserve University can reach 97.32±0.12%,it can fully classify the actually measured bearing fault dataset of Dalian University;at the same time,ML-CAM is used to draw class activation heatmaps for CSADI-Net on various samples of the two datasets to interpret the network operation mechanism,and confirm CSADI-Net having advantages of high accuracy,high anti-noise ability and strong interpretability.