Abstract: Rolling bearing is a key component of mechanical equipment.The instability of its vibration signal and the limitation of single domain features increase the difficulty of bearing fault diagnosis in some extent.On this basis,a bearing fault diagnosis technology based on deep learning and multi-domain decision fusion was proposed.The S transform and recurrence plot transform were used to extend the vibration signal from one-dimensional time domain to two-dimensional time-frequency domain and spatial domain.Then,to adapt the diagnosis model to the lack of fault data,a micro-convolutional neural network with better generalization ability and adaptability was built to learn and extract multi-domain features of the signal,and the network parameters were as low as 6 orders of magnitude,which could be trained and classify fault data efficiently.Finally,D-S evidence theory was introduced to fuse the single domain diagnosis results.The proposed method achieved an average diagnostic accuracy of 99.84% for nine types of bearing faults in the Case Western Reserve University dataset.

Key words: rolling bearing, micro-convolutional neural network, multi-domain fusion, fault diagnosis

摘要： 鉴于滚动轴承振动信号的不平稳性及单一信息域特征的局限性在一定程度上增加了故障诊断难度,提出一种基于深度学习和多域决策融合的轴承故障诊断技术。采用S变换和递归图变换技术将振动信号从一维时域扩展至二维时频域和空间域;为使诊断模型适应故障数据稀缺的现状,构建泛化性和自适应性较好的微型卷积神经网络,学习提取信号的多域特征,并使网络参数低至6个数量级,可实现快速训练和故障诊断;最后引入D-S证据理论对单域诊断结果进行融合。所提方法对凯斯西储大学数据集的9类轴承故障的平均诊断准确率达到99.84%。

关键词: 滚动轴承, 微型卷积神经网络, 多域融合, 故障诊断