基于声成像与卷积神经网络的轴承故障诊断方法及其可解释性研究

An acoustic fault diagnosis method of rolling bearings based on acoustic imaging and convolutional neural network

常用的振动诊断技术一般采用接触式测量,在测量受限的场合具有一定的局限性。该研究提出一种具有非接触测量优势的基于声成像与卷积神经网络的滚动轴承声学故障诊断方法。首先,利用传声器阵列获取滚动轴承辐射的空间声场;然后,用波叠加法进行声成像,重建后的声像能够描述声场的空间分布信息;最后,建立卷积神经网络(convolutional neural network,CNN),使用不同轴承运行状态下的声像样本对CNN模型进行训练用于故障诊断。同时,针对深度学习模型的诊断结果缺乏可解释性的问题,采用梯度加权类激活图(gradient-weighted class activation map,Grad-CAM)算法对卷积神经网络在基于声像的轴承故障诊断中的可解释性进行了研究。轴承试验台的声阵列数据验证了所提方法的有效性及优越性。

Contact measurementis generally used in the common vibration diagnosis techniques,which has certain limitations in situations where measurement is limited.In this paper,a rolling bearing acoustic fault diagnosis method based on acoustic imaging and convolutional neural network with the advantage of non-contact measurement was proposed.First,the spatial acoustic field radiated by the rolling bearing was obtained by using microphone array;then,acoustic imaging was performed by wave superposition method,and the reconstructed acoustic image can describe the spatial distribution information of the acoustic field;finally,a convolutional neural network(CNN)was established,which was trained for fault diagnosis using the acoustic image samples of different bearing operating states.Meanwhile,to address the problem of lack of interpretability of diagnostic results of deep learning models,this paper investigates the interpretability of convolutional neural networks in acoustic image-based bearing fault diagnosis using the gradient-weighted class activation map(Grad-CAM)algorithm.The acoustic array data from the bearing experimental bench verifies the effectiveness and superiority of the proposed method.