A federated learning-based method for predicting steel mechanical properties in the hot rolling mill process

In the hot rolling mill process(HRMP), as important quality indicators, the steel mechanical properties should be precisely modelled and predicted. The paper is focused on such a prediction method development that considers multiple steel production line’s cooperation. First of all, a general regression neural network(GRNN)with attention mechanism is used to develop the base prediction model in each HRMP line; then, a federated learning framework is used to achieve a global GRNN model training and optimizing for different HRMP lines. The proposed methods are verified using real HRMP data from three production lines. It is shown that the proposed method performs better than the other methods, and can improve the prediction of yield strength(YS) by 4 to 7 Mpa, tensile strength(TS) by 2 to 4 Mpa and elongation(EL) by 0.3% to 1%.