Learning-Based Optimization of Copper Converter Predictive Maintenance

Abstract:
In the presented work, the development of a hybrid intelligent system for Condition-Based Maintenance (CBM) of Peirce-Smith Converter (PSC) is considered as an integral part of a plant-wide operational system. PSC is considered as a nonlinear, stochastic, time-varying and with a big uncertainty. Thus Neural Networks (NN) for modeling the degradation of each tuyer is accepted. Data reconciliation is performed in order to reduce the gross errors. A method for the candidate-tuyer for a possible blocking is proposed. The quantity of blocked tuyers and the blocking time is determined at the higher hierarchical level as a trade-off between learning procedure, Case-Based Reasoning Model Predictive Control and a Threshold-Based Maintenance. To estimate the minimal residual size of the PSC refractory lining during the final series of the process cycle, a NN-based approach is proposed using signals from the hottest spots image processing and pattern recognition. Some representative experimental and stimulation results are presented.