摘要
现有基于卷积神经网络(convolutional neural network,CNN)的柴油机故障诊断方法易过拟合,网络收敛速度较慢、处理小样本数据时诊断精度低,针对以上问题,提出了一种基于改进CNN的“端到端”柴油机故障诊断方法。该方法在CNN架构上,采用指数线性单元(exponential linear units,ELU)作为激活函数及小批量训练方法加速模型收敛,用全局平均池化(global average pooling,GAP)代替全连接层以降低过拟合风险。基于台架试验的诊断结果表明:所提方法进行柴油机典型故障诊断的精度达到99.18%;与未改进模型及现有基于CNN的柴油机故障诊断算法相比,该方法在处理小样本数据集时仍保持最高识别精度。
Aiming at the problems of slow model convergence and low diagnosis accuracy when processing small sample data based on the existing convolutional neural network(CNN)diesel engine fault diagnosis method,a more effective method based on an improved CNN was proposed.In the convolutional neural network architecture,exponential linear units(ELU)were used as the activation function,the small batch training method accelerated the model convergence,and the global average pooling(GAP)replaced the fully connected layer to reduce the risk of overfitting.The experimental data analysis shows that the accuracy of the method proposed for the diesel engine typical fault diagnosis reaches 99.18%;compared with the unimproved model and the existing CNN-based diesel engine fault diagnosis algorithm,this method still maintains the highest accuracy when dealing with small sample data sets.
作者
张俊红
孙诗跃
朱小龙
周启迪
戴胡伟
林杰威
ZHANG Junhong;SUN Shiyue;ZHU Xiaolong;ZHOU Qidi;DAI Huwei;LIN Jiewei(State Key Laboratory of Combustion for Internal Combustion Engines,Tianjin University,Tianjin 300072,China;Renai College,Tianjin University,Tianjin 301636,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2022年第6期139-146,共8页
Journal of Vibration and Shock
基金
内燃机可靠性国家重点实验室开放课题(skler-202009)。
