结合CWT和LightweightNet的滚动轴承实时故障诊断方法

Real time fault diagnosis method of rolling bearing based on CWT and LightweightNet

针对普通的深度学习算法用于轴承故诊断分类时计算量大、消耗成本高的问题,提出一种结合连续小波变换和轻量级神经网络的滚动轴承实时故障诊断方法。首先,使用Morlet母小波函数对轴承振动加速度数据进行连续小波变换,提取出时频域特征并将一维信号转换成二维图片;然后,结合分组卷积、通道混洗、倒残差结构等轻量级神经网络设计元素设计一个轻量级卷积神经网络LightweightNet用于时频图片的故障分类,LightweightNet网络在保证具有足够特征提取能力的同时还具有轻量级特点。使用凯斯西储大学轴承故障数据集进行实验表明,本方法相比于其他使用经典轻量级神经网络的方法具有更少的参数、最高的准确率和更快的诊断速度,基本可以实现滚动轴承的实时故障诊断,且在内存消耗与模型存储占用空间方面远小于其他同类方法。

In order to solve the problem of large computation and high cost when common deep learning algorithm is applied to bearing fault diagnosis and classification,a real-time rolling bearing fault diagnosis method combining continuous wavelet transform and lightweight neural network is proposed in this paper.Firstly,the Morlet mother wavelet function is used to carry out continuous wavelet transform on the bearing vibration acceleration data,extracting the time-frequency domain features and converting the one-dimensional signals into two-dimensional images.Then,Lightweight-Net,a lightweight convolutional neural network,is designed for time-frequency image fault classification by combining lightweight neural network design elements such as group convolution,channel shuffle and inverted residual structure.LightweightNet not only ensures sufficient feature extraction ability,but also has lightweight characteristics.The bearing failure experiment data sets from Case Western Reserve University show that,compared with other methods using classic lightweight neural network,this method has less parameters,the highest degree of accuracy and faster diagnosis speed,the real-time fault diagnosis of rolling bearing can be achieved basically,far less than other similar methods in memory consumption and model storage space.