一种改进LeNet5结合LightGBM的滚动轴承故障诊断方法

Bearing fault diagnosis method based on improved LeNet5 and LightGBM

针对传统卷积神经网络在滚动轴承故障诊断方面容易发生的过拟合、精度不足等问题,提出了一种改进LeNet5结合轻量级梯度提升机(LightGBM)的滚动轴承智能故障诊断方法。将滚动轴承的时域信号堆叠为二维的灰度图作为输入,通过改进卷积神经网络LeNet5实现原始数据特征的自适应提取,引入批归一化操作以及全局平均池化代替全连接层降低计算量,将特征输入LightGBM分类器中。使用西储大学轴承数据集以及INV1618实验台变转速数据进行实验结果验证,将该模型其他机器学习算法进行对比,证明了其在准确率和鲁棒性上的优势。

Aiming at the overfitting and insufficient precision problem that were prone to occur in traditional convolutional neural networks in rolling bearing fault diagnosis.An improved LeNet5 combined with LightGBM intelligent fault diagnosis method was proposed.The time-domain signal of the rolling bearing was stacked into two-dimensional grayscale images,and fed into a convolutional neural network to realize the adaptive extraction of raw data features,and the extracted features was put into the LightGBM classifier to achieve accurate classification of rolling bearings under different working conditions.Case Western Reserve University bearing dataset and INV1618 test rig variable speed data were used to compare the model with other machine learning algorithms,and prove its advantages in accuracy and robustness.