Fault Diagnosis of Ultra-Supercritical Thermal Power Units Based on Improved ICEEMDAN and LeNet-5

Aiming at the problems of massive, high-dimensional, nonlinear, and strong noise data during operation, this article proposes a fault diagnosis method of ultra-supercritical (USC) thermal power units based on dual improved complete ensemble empirical mode decomposition with adaptive noise (IICEEMDAN) and improved LeNet-5. First, the raw data are decomposed into multiple intrinsic mode functions (IMFs) by using ICEEMDAN. Second, an effective IMF selective reconstruction method is proposed, and the reconstructed data are converted into a 2-D grayscale image as an input to the diagnostic model, which is able to improve the data stability and reduce the noise interference. Finally, the proposed improved deep learning method is used for fault diagnosis of a 1000-MW USC thermal power unit. The experimental results indicate that the proposed method has superiority in fault identification of USC thermal power units compared with the traditional LeNet-5 network, 1-D convolutional neural network (1-D CNN), BP, and SVM algorithms.