Fast Reconstruction Method of the Stress Field for the Steam Turbine Rotor Based on Deep Fully Convolutional Network

Abstract

Since it is difficult to directly measure the transient stress of a steam turbine rotor in operation, a rotor stress field reconstruction model based on the deep fully convolutional network for the startup process is proposed. The stress distribution in the rotor can be directly predicted based on the temperature of a few measurement points. First, the finite element model is used to accurately simulate the temperature and stress field of the rotor startup process, generating training data for the deep learning method. Next, data of only 15 temperature measurement points are arranged to predict the stress distribution in the critical area of the rotor surface, with the accuracy (R2-score) reaching 0.997. The time cost of the trained neural network model at a single case is 1.42 s in CPUs and 0.11 s in GPUs, shortened by 97.3% and 99.8% with comparison to finite element analysis, respectively. In addition, the influence of the number of temperature measurement points and the training size is discussed, verifying the stability of the model. With the advantages of fast calculation, high accuracy, and strong stability, the fast reconstruction model can effectively realize the stress prediction during startup processes, resulting in the possibility of a real-time diagnosis of rotor strength in operation.