Reliability Prediction of Near-Isothermal Rolling of TiAl Alloy Based on Five Neural Network Models

Abstract

The near-isothermal rolling process has the characteristics of multi-variable and strong coupling, and the industrial conditions change constantly during the actual rolling process. It is difficult to consider the influence of various factors in industrial sites using theoretical derivation, and the compensation coefficient is difficult to accurately determine. The neural network model compensates for the difficulty in determining the compensation coefficient of the theoretical model. The neural network can be trained in advance through historical data, the trained network can be applied to industrial sites for prediction, and previous training errors can be compensated for through online learning using real-time data collected on site. But it requires a large amount of effective historical data, so this research uses a combination of production data from a controllable two-roll rolling mill and finite element simulation to provide training data support for the neural network. Five trained neural networks are used for prediction, and the results are compared with industrial site data, verifying the reliability and accuracy of genetic algorithm optimized neural network prediction. We successfully solved the problem of low control accuracy of TiAl alloy outlet thickness during near-isothermal rolling process.