基于ACON激活函数和卷积神经网络的滚动轴承故障诊断  

Fault Diagnosis for Rolling Bearings Based on ACON Activation Function and Convolutional Neural Network

作　　者：常志远 刘昌奎 李志农[1] 周世健 CHANG Zhiyuan;LIU Changkui;LI Zhinong;ZHOU Shijian(Key Laboratory of Nondestructive Testing Technology of Ministry of Education,Nanchang Hangkong University,Nanchang 330063,China;AECC Beijing lnstitute of Aeronautical Materials,Beijing 100095,China)

机构地区：[1]南昌航空大学无损检测技术教育部重点实验室,南昌330063 [2]中国航发北京航空材料研究院,北京100095

出　　处：《轴承》2024年第8期53-58,67,共7页Bearing

基　　金：国家自然科学基金资助项目(52075236);江西省自然科学基金重点项目(20212ACB202005)。

摘　　要：针对滚动轴承故障诊断任务的泛化问题,提出一种基于ACON激活函数和卷积神经网络(CNN)的故障诊断方法(ACON-CNN模型)。构造一种自适应激活因子,利用ACON激活函数的自适应激活特性增强整个卷积神经网络的自适应特征能力;同时构造一种基于稀疏结构的神经元簇,增加诊断模型的稳定性。对CWRU轴承数据集以及航空轴承数据集的试验结果表明:针对同一轴承不同采集端故障数据的训练与识别中,ACON-CNN模型具有比原始CNN,FFT-CNN,LSTM-CNN更好的识别效率和鲁棒性;在不同轴承样本数据集的迁移学习中,ACON激活函数和稀疏神经元簇的综合作用也使ACON-CNN模型获得了更好的泛化性能和识别效果。For the problem of generalization of rolling bearing fault diagnosis task,a fault diagnosis method(ACON-CNN model)is proposed based on ACON activation function and convolutional neural network(CNN).An adaptive activation factor is constructed,and the adaptive activation characteristics of ACON activation function is used to enhance the adaptive feature capability of whole CNN.At the same time,a sparse structure-based neural cluster is constructed to increase the stability of diagnosis model.The experimental results using CWRU bearing dataset and aviation bearing dataset show that the ACON-CNN model has better recognition efficiency and robustness than original CNN,FFT-CNN and LSTM-CNN in training and recognition of fault data from different collection ends for same bearing;the combined effect of ACON activation function and sparse neural clusters also enables ACON-CNN model to achieve better generalization performance and recognition effect in transfer learning of different bearing sample datasets.

关 键 词：滚动轴承 故障诊断 卷积神经网络 激活函数 深度学习 迁移学习 

分 类 号：TH133.33[机械工程—机械制造及自动化] TH113.3