Research on fault diagnosis of supercharged boiler with limited data based on few-shot learning

Abstract

The safety of the supercharged boiler affects the normal operation of the steam power system, while its fault samples are few and contain large noise in reality. Therefore, we propose a few-shot fault diagnosis framework for supercharged boilers based on Siamese Neural Network(SNN). The variable analysis and two screening processes are introduced to train the model efficiently. The results show that when the number of training samples is 30 and the noise is −4 dB, the accuracy of Five-shot method is 45.17%, 26.56%, 19.31%, 18.32% and 8.43% higher than that of K-Nearest Neighbor (KNN), Multi-Layer Perceptron (MLP), Random Forest (RF), Support Vector Machine (SVM) and Convolutional Neural Network (CNN), respectively. When the number of training samples is 60, the accuracy difference between Five-shot and its main competitor CNN under the proportion of 30%, 20% and 10% new categories are 4.53%, 5.72% and 4.12%, respectively. When all 75 samples from different thermal systems are used for training, the accuracy of Five-shot method can reach 85% with the help of support set. The proposed few-shot fault diagnosis framework and variable screening method can be used as the preferred scheme for supercharged boilers fault diagnosis with limited fault data.