MDA-JITL model for on-line mechanical property prediction

Mechanical performance prediction is the key to the transformation and upgrading of steel enterprises to intelligent manufacturing. Due to time-varying manufacturing data, the traditional prediction model of mechanical properties of hot-rolled strip may cause performance degradation or even failure in its use. An MDA-JITL model was thus proposed to handle the modeling problem of complex time-varying data. Relevant parameters were first chosen and normalized. Then, a distance measurement method combining the importance of data attributes and time characteristics was designed to select the most suitable samples for on-line local modeling. After that, using the chosen dataset, a linear local model was created to predict target sample. Finally, an uncertainty evaluation method was designed to evaluate the uncertainty of prediction results. Furthermore, the appropriate dataset partition and off-line simulation experiment scheme were created based on the peculiarities of hot-rolling production. The suggested model performs much better than the classic global model when applied to actual production data from a steel plant. The stability of its prediction accuracy is demonstrated in a simulation prediction for up to five months. Moreover, there is a high link between the uncertainty evaluation metrics and the prediction error of the model, reducing the field sampling rate by 30% in industrial applications in the latest year.