| 宋凯,黄盟,尤健,等.基于改进残差卷积网络的柴油机故障诊断方法[J].内燃机工程,2023,44(5):66-73. |
| 基于改进残差卷积网络的柴油机故障诊断方法 |
| Diesel Engine Fault Diagnosis Method Based on Improved Residual Convolution Network |
| DOI：10.13949/j.cnki.nrjgc.2023.05.009 |
| 关键词:柴油机 故障诊断 卷积神经网络 残差结构 |
| Key Words:diesel engine fault diagnosis convolutional neural network（CNN） residual module |
| 基金项目:河北省高等学校科学技术研究项目（QN2022159）；省部共建电工装备可靠性与智能化国家重点实验室2021年度开放课题项目（EERIPD2021008） |
|
| 摘要点击次数: 659 |
| 全文下载次数: 443 |
| 摘要:针对基于卷积神经网络（convolutional neural network， CNN）的柴油机故障诊断方法在训练样本匮乏时易过拟合、诊断准确率低的问题，提出一种基于改进残差卷积网络的“端到端”柴油机故障诊断方法。该方法采用连续可微指数线性单元（continuously differentiable exponential linear units， CELU）作为CNN激活函数并采取小批次训练方法，提高模型提取特征能力的同时加速其收敛；在模型中加入残差结构将深层网络提取到的抽象特征与表层特征相融合，避免深层网络导致的特征信息丢失与梯度消失问题。经柴油机故障模拟试验验证，该方法在仅使用20个样本进行训练时，能实现95.5%的故障诊断准确率；与CNN相比，该方法在不同类型及规模的训练集下，故障诊断准确率均有显著提升。 |
| Abstract:Aiming at the problem that the diesel engine fault diagnosis method based on convolutional neural network（CNN） tended to over-fit and the diagnatic accuracy was low when the samples were scarce， an “end-to-end” diesel engine fault diagnosis method based on an improved residual convolution network was proposed. The continuously differentiable exponential linear units（CELU） were used as the activation function of CNN and a small batch training method was adopted to improve the ability of feature extraction and accelerate model convergence. The residual module was added to the model to integrate the abstract features extracted from the deep network with the surface features， avoiding the loss of feature information and gradient loss caused by the deep network. The diesel engine fault simulation experiment shows that the method can achieve 95.5% fault diagnosis accuracy with only 20 samples used for model training. Compared with CNN， the method can significantly improve the accuracy of fault diagnosis under different types and scales of training sets. |
|
查看全文 HTML
查看/发表评论 下载PDF阅读器 |