基于改进交替迁移学习的滚动轴承故障诊断算法

Rolling bearing fault diagnosis algorithm based on improved alternating transfer learning

针对不同型号滚动轴承监测信号之间特征分布差异大、故障数据样本少,导致轴承故障精度低的问题,提出了一种基于改进交替迁移学习的滚动轴承故障诊断算法。为了充分发挥卷积神经网络(convolutional neural network, CNN)对二维数据优秀的特征提取能力,首先将一维振动信号转化为二维图像,输入到深度卷积神经网络中学习;其次,为了减少源域与目标域数据间的特征分布差异,提出了改进的交替迁移学习(improved alternately transfer learning, IATL),通过交替计算域间的CORAL损失函数和最大均值差异(maximum mean discrepancy, MMD)损失函数,并反向传播更新各层网络权重与偏置参数,以实现变工况、跨轴承型号和小故障样本条件下轴承特征迁移适配;最后,在全连接层使用Softmax函数对目标域数据进行故障诊断。为了验证该算法的有效性,采用凯斯西储大学(Case Western Reserve University, CWRU)的滚动轴承数据集进行了迁移试验验证。结果表明,与仅计算CORAL损失函数和MMD损失函数等算法对比可知,该算法有效地减少了领域数据之间的特征分布差异,具有较高的故障分类准确率。

Here,aiming at problems of large differences in feature distribution among monitoring signals of different types of rolling bearing and less fault data samples to cause low bearing fault diagnosis accuracy,a rolling bearing fault diagnosis algorithm based on improved alternating transfer learning was proposed.Firstly,in order to fully utilize excellent feature extraction ability of convolutional neural network(CNN)for 2-D data,1-D vibration signals were converted into 2-D images and input into a deep CNN for learning.Secondly,in order to reduce differences in feature distribution between source domain data and target domain data,an improved alternating transfer learning(IATL)was proposed.By alternately calculating CORAL loss function and maximum mean difference(MMD)loss function among domains,and updating weights and bias parameters of network various layers through backpropagation,bearing features’transfer were realized under variable operating conditions,across bearing types and small fault sample conditions.Finally,within fully connected layer,Softmax function was used to do fault diagnosis of target domain data.To verify the effectiveness of the proposed algorithm,transfer test verification was conducted using the rolling bearing dataset of Case Western Reserve University,USA.The results showed that compared with other algorithms to only calculate CORAL loss function and MMD loss function,the proposed algorithm can effectively reduce feature distribution differences among domains’data,and have a higher fault classification accuracy.