Machine-Learning Optimization of Rubber-Mixing and Curing Processes

Rubber manufacturing processing continues to rely on fixed curing times, operator guesswork and offline rheometer tests leading to over curing, energy loss and hardness variation from part to part. Conventional statistical and single-step neural networks cannot account for nonlinear mixing history, shear history and crosslink rate interactions such that the prediction accuracy degrades when humidity, dispersant of filler or batch temperature go out of historical ranges. A generic machine learning-based approach employing feature extraction for the mixer torque and temperature signals, followed by hybrid (GRNN as well as ANN) neural prediction with NSGA-II optimization is proposed to find out coupled-setpoints for different operations of mixing and curing. Experimental results show reductions of 12% in mixing energy, 15% in cure cycle time and 18% in scrap, with <3% average prediction error for tc90 and torque response. Result interpretation indicates a reduction in hardness deviation (±1.5 Shore A), modulus variation (±2.6%) and significant progress concerning stability, efficiency and quality compared to the state-of-The-Art production processes.