Prediction of strip steel product quality by combining small disturbance theory with INGO-GRNN

The quality control precision of strip rolling products is a crucial factor influencing the economic benefit and international com petitiveness of iron and steel enterprises. As the significant quality evaluation index of strip products, the precise control and measurement of thickness and flatness are of paramount importance in order to enhance the yield and use value of strip products. In order to address the issue of the limited fit degree between the current rolling mechanism model and machine learning algo rithm, a new algorithm based on Small Disturbance Theory (SDT) and Generalized Regression Neural Network (GRNN) has been proposed. The two-stage model framework of GRNN is used to predict the flatness and thickness quality of strip rolled products. Given the nonlinearity of the variables in actual rolling production, the small disturbance theory was introduced to linearize the deviation of the mechanism model caused by the perturbation amount and the adjustment amount. A linear error correction model with rolling force as the intermediate variable was established. The maximal relevance and minimal redundancy (mRMR) principle was employed to evaluate the importance of the rolling process feature parameters, and the feature subset of quality prediction with the most sufficient information and minimum data dimension was selected. Subsequently, the Improved Northern Goshawk Optimization (INGO) algorithm was employed to optimize the initial population distribution, after which the hyperparameters of the GRNN were optimized. This resulted in the establishment of a nonlinear error correction model with the prediction of flatness and thickness. To verify the predictive performance of the SDT-INGO-GRNN model, multiple indexes were calculated. The results demonstrate that the SDT-INGO GRNN model proposed in this paper exhibits accurate and stable prediction performance in the product quality prediction of the 1580 four-high hot tandem rolling line. The thickness prediction error is largely within 2 µm , and the flatness prediction error is largely within 0.5 I .