基于注意力机制多尺度卷积神经网络的轴承故障诊断

Bearing Fault Diagnosis Based on Multi-Scale Convolutional Neural Network of Attention Mechanism

提出了基于注意力机制的多尺度卷积神经网络(Multi-scale and Attentive Convolutional Neural Network,MACNN)进行轴承故障分类,该模型以一维Resnet18网络结构为主体,卷积模块采用残差模块和空洞卷积并行方式以达到扩大感受野、避免特征信息丢失的目的,同时利用注意力机制可以自动提取有用特征的能力,将模型提取特征作为输入送入注意力机制模块,进一步提高模型故障分类能力。此外,采用边界平衡生成对抗网络(Boundary Equilibrium Generative Adversarial Networks,BEGAN)模型对故障数据增强,改变不平衡数据集的比例,增加数据集样本数量,降低MACNN模型的过拟合,提高诊断的准确率。在帕德博恩轴承数据集(Paderborn University Dataset,PU)上验证MACNN模型,实验结果表明,该模型在特征提取和故障分类方面都表现出了良好的性能,优于当前主流模型。

Bearing faults can seriously reduce industrial production efficiency and even endanger the safety of people's lives and property.Monitoring the operating conditions of bearings conducting fault diagnosis are of great significance for ensuring the safe operation of the production process.This paper proposes a Multi-Scale and Attentive Convolutional Neural Network(MACNN)based on attention mechanism for bearing faults classification.Firstly,onedimensional bearing signals are input into a convolution layer,followed by a maximum pooling layer to suppress noise and reduce information redundancy.Then,four MACNN modules are taken as input,each of which adopts a parallel network structure of ordinary convolution and void convolution.Under the premise without increasing model parameters,the model's receptive field is expanded to extract more fault features for improving the accuracy.In addition,the attention mechanism module is connected at the end of each MACNN model and the feature information is further extracted by using the ability of automatic extraction of important features.The average pooling layer is used in the network structure to prevent the overfitting and the full connection layer from the output of experimental classification results.Furthermore,the Boundary Equilibrium Generative Adversarial Networks(BEGAN)model is adopted to enhance fault data,change the proportion of unbalanced data sets,increase the number of dataset samples,reduce overfitting of MACNN model,and improve diagnostic accuracy.Finally,the experimental results on the Paderborn University Dataset show that MACNN can achieve better performance in feature extraction and fault classification,outperforming the state-of-the-art methods.