Abstract: To improve the reliability and safety of mechanical equipment,it is imperative to carry out fault identification on bearings.However,when the number of training samples is insufficient,the accuracy of existing fault identification models will drop significantly.At the same time,the noise interference and load variation during the operation of the bearing make its fault identification face significant difficulties and challenges.Aiming at the above problems,a credible multi-scale quadratic attention convolutional neural network model was proposed,which adopted a multi-scale wide convolution kernel suitable for bearing vibration signals by fully considering the idea of the feature pyramid firstly.Then,a small convolution kernel was used in the subsequent feature extraction stage,and quadratic neuron including attention mechanism was introduced at this stage.In the multi-scale feature fusion stage,the outputs of the model were transformed into a Dirichlet distribution,and then fused using DS evidence theory to achieve credible classification.The experimental results showed that the model had excellent generalization ability and robustness,and its fault identification performance was superior to other comparison models under various complex operating conditions when samples were scarce,showing highly competitive fault identification results.To improve the reliability and safety of mechanical equipment,it is imperative to carry out fault identification on bearings.However,when the number of training samples is insufficient,the accuracy of existing fault identification models will drop significantly.At the same time,the noise interference and load variation during the operation of the bearing make its fault identification face significant difficulties and challenges.Aiming at the above problems,a credible multi-scale quadratic attention convolutional neural network model was proposed,which adopted a multi-scale wide convolution kernel suitable for bearing vibration signals by fully considering the idea of the feature pyramid firstly.Then,a small convolution kernel was used in the subsequent feature extraction stage,and quadratic neuron including attention mechanism was introduced at this stage.In the multi-scale feature fusion stage,the outputs of the model were transformed into a Dirichlet distribution,and then fused using DS evidence theory to achieve credible classification.The experimental results showed that the model had excellent generalization ability and robustness,and its fault identification performance was superior to other comparison models under various complex operating conditions when samples were scarce,showing highly competitive fault identification results.

Key words: bearing fault identification, quadratic attention convolution, multi-scale learning, credible classification

摘要： 为提高机械装备的可靠性与安全性,对轴承进行故障识别势在必行。然而当训练样本量缺乏时,现有故障识别模型的精度会大幅下降,同时轴承运行过程中的噪声干扰和负载变动,使其故障识别面临显著困难与挑战。针对上述问题本文提出了一种可信多尺度二次注意力卷积神经网络模型,该模型在充分考虑特征金字塔的思想上,首先采用适用于轴承振动信号的多尺度宽卷积核,其次在后续特征提取阶段采用小卷积核,并在此阶段引入了包含注意力机制的二次神经元,最后在多尺度特征融合阶段通过将模型的输出转化为狄利克雷分布,再利用DS证据理论进行融合,达到可信分类。实验结果表明该模型具有优异的泛化能力和鲁棒性,在各种样本缺乏时的复杂工况下,其故障识别性能均优于其他对比模型,表现出极具竞争力的故障识别结果。

关键词: 轴承故障识别, 二次注意力卷积, 多尺度学习, 可信分类