基于通道注意力机制与多尺度减法轻量化网络的滚动轴承故障诊断  

Fault Diagnosis of Rolling Bearing Based on Channel Attention Mechanism and Multi-scale Subtraction Lightweight Neural Network

作　　者：章力 邓艾东[1,2] 王敏[1,2] 卞文彬 张宇剑 ZHANG Li;DENG Aidong;WANG Min;BIAN Wenbin;ZHANG Yujian(National Engineering Research Center of Power Generation Control and Safety,Southeast University,Nanjing 210096,Jiangsu Province,China;School of Energy and Environment,Southeast University,Nanjing 210096,Jiangsu Province,China)

机构地区：[1]东南大学大型发电装备安全运行与智能测控国家工程研究中心,江苏南京210096 [2]东南大学能源与环境学院,江苏南京210096

出　　处：《动力工程学报》2025年第4期571-581,共11页Journal of Chinese Society of Power Engineering

基　　金：江苏省碳达峰碳中和科技创新专项资金资助项目(BA2022214);中央高校基本科研业务费专项资金资助项目(2242023K30011、2242023K30058)。

摘　　要：针对传统多尺度卷积神经网络模型存在的特征定位不精确、训练时间长、抗噪性能差等问题,提出了一种基于通道注意力机制与多尺度减法轻量化网络的滚动轴承故障诊断模型。首先,将滚动轴承的一维振动信号转换为二维灰度图作为输入,丰富特征信息;同时,构建多尺度减法神经网络模型,关注层级差异;其次,引入轻量化模块,减少内存访问;然后,结合通道注意力机制,调整特征权重;最后,将故障样本输入到网络模型中,实现精确分类。利用风电机组传动系统模拟实验台采集的样本数据进行诊断任务。结果表明:该故障诊断模型能够有效克服传统多尺度卷积神经网络模型网络层数多、参数量大所带来的弊端,能够充分关注各层级之间的差异信息,减少冗余信息的提取,精确定位故障特征,缩短模型训练时间,在恒定工况、变工况和强噪声工况下都具有较高的诊断精度.Aiming at the problems of inaccurate feature location,long training time,and poor anti-noise performance in traditional multi-scale convolutional neural network models,a fault diagnosis model for rolling bearings based on channel attention mechanism and multi-scale subtraction lightweight neural network was proposed.First,the one-dimensional vibration signal of rolling bearing was converted into a two-dimensional grayscale images as input to enrich feature information.Meanwhile,a multi-scale subtraction neural network model was constructed to focus on level differences.Secondly,a lightweight module was introduced to reduce memory access.Then,the feature weights were adjusted by combining with the channel attention mechanism.Finally,the fault samples were input into the network model to achieve accurate classification.The diagnosis tasks were conducted using the sample data collected from the wind turbine transmission system simulation test bench.Results show that this model has high diagnostic accuracy under constant,variable working conditions and strong noise working conditions,which can overcome the drawbacks caused by the large number of network layers and parameters in traditional multi-scale convolutional neural network models.It can fully focus on the difference information between the levels,reduce the extraction of redundant information,locate fault characteristics accurately,and shorten the model training time.It has high diagnostic accuracy under constant working condition,variable working conditions and strong noise conditions.

关 键 词：滚动轴承 故障诊断 多尺度减法神经网络 轻量化模块 通道注意力机制 变工况 

分 类 号：TK83[动力工程及工程热物理—流体机械及工程]