Improved LeNet-5 Network for Equipment Fault Diagnosis of Ultra-supercritical Units

In order to improve the reliability of the power generation system of ultra-supercritical units, a fault diagnosis algorithm based on the improved LeNet-5 network is proposed to address the problems of difficult feature extraction, low accuracy and reliance on manual experience of traditional fault diagnosis methods. Firstly, multi-scale convolutional kernels in parallel are used to extract more details of the fault features. By using the improved inception V2 network and residual neural network, more complete and accurate features can be extracted while avoiding the degradation of the model due to too deep layers. Then a combination of
$1^{\ast}1$
convolution and global average pooling is used instead of partial fully connected layers, which greatly reduces the parameters of the model and prevents model overfitting. The test shows that the fault identification rate of this method can be 98.42%.