针对引风机轴承变工况运行时,传统深度学习方法特征学习能力不足的问题,本文提出了一种基于改进CNN的引风机变工况轴承故障诊断方法。该方法首先利用多尺度卷积对原始特征图进行特征提取,然后利用通道注意力机制对多尺度特征进行自适应筛选；之后构建稠密连接残差模块结构,以实现模型快速收敛和特征重用目的；最后以SoftMax交叉熵作为损失函数,利用Adam优化实现对轴承健康状态的识别。试验结果表明,在模拟试验台变转速轴承数据集上,该方法的平均准确率达到99.68%；在美国凯斯西储大学变工况轴承数据集上,平均准确率达到97.10%；与常用的机器学习模型进行对比实验,验证了该方法在变工况下具有更好的诊断性能。

Aiming at the problem of insufficient feature learning capability of traditional deep learning methods when the induced draft fan bearings operate under variable operating conditions, this paper proposes an improved CNN-based fault diagnosis method for induced draft fan bearings under variable operating conditions. The method firstly uses multi-scale convolution to extract features from the original feature map, and then uses the channel attention mechanism to adaptively screen the multi-scale features; after that, it constructs the structure of densely connected residual module to achieve the purpose of rapid convergence of the model and feature reuse; finally, it takes the SoftMax cross entropy as the loss function, and uses the Adam optimization to achieve the identification of the bearing"s health state. The experimental results show that the average accuracy of the method reaches 99.68% on the variable speed bearing dataset of the simulation test bench, and 97.10% on the variable operating condition bearing dataset of Case Western Reserve University, U.S.A. Comparison experiments with commonly used machine learning models verify that the method has better diagnostic performance under variable operating conditions.

国家重点研发计划（2022YFB4100403）