基于GST与改进CNN的滚动轴承智能故障诊断

Intelligent fault diagnosis of rolling bearings based on GST and improved CNN

针对滚动轴承传统故障诊断方法需要先验知识以及人工提取特征导致故障识别错误率高的问题,提出一种基于广义S变换(Generalized S transform,GST)和改进卷积神经网络(Convolutional Neural Network,CNN)的滚动轴承智能故障诊断方法。使用GST将一维振动信号转换为特征信息更加丰富的时频图,更加全面提取滚动轴承的故障特征信息。通过加入弹性斜率和高斯分布的神经元噪声,提出改进的激活函数EReLTanh(Elastic Rectified Linear Tanh,EReLTanh),并基于EReLTanh激活函数构建CNN。将得到的时频图进行压缩和归一化处理,生成时频图数据集并划分数据集。利用时频图数据集训练改进CNN,实现滚动轴承的智能故障诊断。使用自制实验平台采集不同种类滚动轴承故障数据,利用t-SNE进行全连接层特征降维可视化,结果表明:使用EReLTanh激活函数的CNN模型能够将不同故障样本的特征进行准确的分类,达到故障识别要求,同时使用该数据利用S变换、小波变换、GST并结合改进CNN和未改进CNN进行对比,提出的方法准确率得到提升。通过分析和对比实验可得出结论,利用GST和改进CNN的滚动轴承智能故障诊断方法能够在实际工程中更加简单方便地判断出故障类型及损伤程度,满足实际工程的需求。

Aiming at the problem that traditional fault diagnosis methods of rolling bearings require prior knowledge and manual feature extraction lead to high error rate of fault recognition,a Generalized S Transform(GST)and improved Convolutional Neural Network(CNN)intelligent fault diagnosis method for rolling bearings.Firstly,GST was used to transform one-dimensional vibration signals into time-frequency diagrams with richer characteristic information,so fault characteristic information of rolling bearings can be more comprehensively extracted.Secondly,an improved Elastic Rectified Linear Tanh(EReLTanh)activation function was proposed by adding neural noise with Elastic slope and Gaussian distribution,and CNN was constructed based on EReLTanh activation function.Thirdly,time-frequency diagramwas compressed and normalized,and the time-frequency diagrams data set was generated and divided.Finally,CNN was trained and improved by using time-frequency diagrams data set to realize intelligent fault diagnosis of rolling bearings.The fault data of different kinds of rolling bearings were collected by self-made experimental platform,and t-SNE was used to carry out dimensionality reduction visualization of features of the full-connection layer.The results show that the CNN model using EReLTanh activation function can accurately classify the features of different fault samples and meet the requirements of fault identification.At the same time,S transform,wavelet transform and GST were used to compare the proposed method with the improved CNN and the unimproved CNN.The accuracy of the proposed method was improved.Through analysis and comparative experiments,it can be concluded that the rolling bearing intelligent fault diagnosis method based on GST and improved CNN can judge the fault type and damage degree more simply and conveniently in practical engineering,so as to meet the needs of practical engineering.