一维改进LeNet-5及机械故障诊断应用

Improved LeNet-5 Convolutional Neural Network and Application on Mechanical Fault Diagnosis

针对往复机械设备故障诊断容易受到转速波动影响和深度网络诊断模型复杂、鲁棒性差的问题,提出了一种一维改进型LeNet-5的机械故障诊断方法,并对比分析了滑动窗和阶次采样的数据样本构造方法的效果。在经典模型LeNet-5基础上构建了结构简单紧凑的一维卷积神经网络诊断模型,模型仅包含了两个卷积模块、单一全连接层和输出层,卷积模块结合批规范化层和Relu层,提高训练速度和网络泛化能力,利用重叠池化窗口和随机失活来缓解网络出现过拟合现象。利用凯斯西储大学开源轴承数据集进行验证,12种工况下的故障识别率能够达到99.82%。针对往复机械的转速波动性影响,采用阶次采样的数据样本构建方法,提高网络模型的训练样本数据质量,柴油机阶次采样条件下可以实现小样本条件下取得良好的训练效果。

Aiming at the problems that the fault diagnosis of reciprocating machinery equipment is easily affected by speed fluctuations and the deep network diagnosis model is complicated and poor in robustness, a one-dimensional improved LeNet-5 mechanical fault diagnosis method was proposed, and the effects of sliding window and order sampling data sample construction method are compared and analyzed. Based on the classic model LeNet-5, a simple and compact one-dimensional convolutional neural network diagnosis model was constructed, which only contains two convolution modules, a single fully connected layer and an output layer. And its convolution module combines the batch normalization layer and Relu layers to improve training speed and network generalization ability, and use overlapping pooling windows and random inactivation to alleviate network overfitting. Using the open source bearing data set of Case Western Reserve University to verify, the fault detection accuracy under 12 working conditions can reach 99.82%. Aiming at the influence of speed fluctuation of reciprocating machinery, the method of constructing data samples of order sampling is adopted to improve the quality of training sample data of the network model. Under the condition of order sampling of diesel engine, good training results can be achieved under the condition of small samples.