基于梯度加权类激活热力图的卷积神经网络故障诊断模型鲁棒性分析

Robustness Analysis of Convolutional Neural Network Fault Diagnosis Model Based on Gradient-weighted Class Activation Map

深度学习近年来在故障诊断领域受到广泛应用,但基于深度学习的故障诊断模型缺乏工程上的物理解释性,难以保证其故障诊断结果的稳定性。以轴承为例,建立了以小波时频图像为故障诊断依据的卷积神经网络模型(convolutional neural network,CNN),提出了一种基于梯度加权类激活热力图(gradient-weighted class activation map,Grad-CAM)的网络模型鲁棒性分析方法,并利用美国凯斯西储大学(Case Western Reserve University,CWRU)轴承数据集进行验证。首先,将故障直径轴承数据以不同方式混合并训练大、小多个模型。其次,利用Grad-CAM方法,建立时频区域与故障模式之间的联系。最后,利用其他工况下的轴承故障数据,以及含噪数据进行测试,并根据结果结合模型最注重的时频区域进行分析。结果表明,基于深度学习的轴承故障诊断模型在参数较少时更加注重低频区域,并能使其具有更好的鲁棒性。

Deep learning has been widely used in the field of fault diagnosis in recent years,but the fault diagnosis model based on deep learning lacks physical explanation in engineering,and it is difficult to ensure the stability of its fault diagnosis results.Taking bearings as an example,a convolutional neural network model(CNN)based on wavelet time-frequency images for fault diagnosis was established,and a gradient-weighted class activation map(Grad-CAM)based robust analysis method of the network model was proposed,which was verified by the Case Western Reserve University(CWRU)bearing data set.Firstly,the bearing data of fault diameter were mixed in different ways and multiple models of large and small were trained.Secondly,the Grad-CAM method was used to establish the relationship between the time-frequency region and the fault mode.Finally,the bearing fault data under other working conditions and the noisy data were used for testing,and the results were analyzed in combination with the time-frequency region that the model pays most attention to.The results show that the bearing fault diagnosis model based on deep learning pays more attention to the low-frequency region when there are fewer parameters and can make it more robust.