Estimación del contenido de cobre y razón hierro sílice en la escoria de un convertidor flash usando datos de proceso y deep learning.
No Thumbnail Available
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
El presente trabajo aborda el desarrollo de una herramienta de monitoreo que permita el seguimiento de la ley de cobre (Cu) y la razón hierro-sílice (Fe/SiO2) en la escoria de un horno de conversión flash. Esta problemática surge de la dependencia actual de métodos de análisis químicos, invasivos y con desfase temporal, los cuales impiden un monitoreo en tiempo real, no invasivo y en línea.
Para solucionar esta problemática se propone un sistema no invasivo basado en el uso de datos de proceso y modelos de Deep Learning programados en Python. La metodología consistió en la construcción de dos conjuntos de datos: uno endógeno, basado en el pasado de las variables objetivo y otro exógeno, compuesto por variables de proceso operacional, simulando así un escenario industrial real donde no siempre se dispone de mediciones químicas pasadas. Esto permitió evaluar la capacidad de los modelos en ambos contextos.
Se implementaron y compararon tres arquitecturas predictivas: un modelo de regresión lineal (que sirvió como línea base), una red neuronal recurrente LSTM (Long Short-Term Memory) y un Temporal Fusion Transformer (TFT). El desempeño se evaluó utilizando como métrica principal Mean Absolute Scaled Error (MASE), y las métricas como MAE, sesgo de pronóstico (FB), coeficiente de correlación de Pearson, R2 e información mutua normalizada (NMI).
Los resultados obtenidos muestran que la regresión lineal constituye una base sólida para capturar tendencias generales, pero resulta insuficiente frente a la alta variabilidad de las series. El modelo LSTM alcanzó el mejor desempeño global, demostrando robustez en la estimación tanto del cobre como de la razón hierro-sílice en el conjunto endógeno. Por su parte, el TFT presentó un mejor comportamiento en el conjunto exógeno, al explotar relaciones no lineales en las variables de proceso, aunque en el conjunto endógeno exhibió dificultades de generalización.
En conclusión, este trabajo analiza la factibilidad de construir un sistema de monitoreo no invasivo para la industria del cobre, con proyección a su integración en sistemas de control en línea. Este enfoque contribuye a la optimización de los procesos pirometalúrgicos, ofreciendo beneficios en seguridad, reducción de costos y monitoreo operacional.
This work addresses the development of a monitoring tool that enables the tracking of copper grade (Cu) and the iron–silica ratio (Fe/SiO2) in the slag of a flash converting furnace. This problem arises from the current reliance on chemical analysis methods, which are invasive and delayed, preventing real-time, non-invasive, and online monitoring. To solve this issue, a non-invasive system is proposed based on the use of process data and Deep Learning models programmed in Python. The methodology consisted of constructing two datasets: an endogenous one, based on the past values of the target variables, and an exogenous one, composed of operational process variables, thus simulating a real industrial scenario where past chemical measurements are not always available. This approach allowed evaluating the capability of the models in both contexts. Three predictive architectures were implemented and compared: a linear regression model (serving as the baseline), a recurrent neural network LSTM (Long Short-Term Memory), and a Temporal Fusion Transformer (TFT). Performance was assessed using Mean Absolute Scaled Error (MASE) as the main metric, along with MAE, forecast bias (FB), Pearson’s correlation coefficient, R2, and normalized mutual information (NMI). The results show that linear regression provides a solid basis for capturing general trends, but is insufficient when facing the high variability of the series. The LSTM model achieved the best overall performance, demonstrating robustness in estimating both copper grade and the iron–silica ratio in the endogenous dataset. Meanwhile, the TFT exhibited better behavior in the exogenous dataset by exploiting non-linear relationships among process variables, although it showed generalization difficulties in the endogenous case. In conclusion, this work analyzes the feasibility of building a non-invasive monitoring system for the copper industry, with the potential to be integrated into online control systems. This approach contributes to the optimization of pyrometallurgical processes, offering benefits in safety, cost reduction, and operational monitoring.
This work addresses the development of a monitoring tool that enables the tracking of copper grade (Cu) and the iron–silica ratio (Fe/SiO2) in the slag of a flash converting furnace. This problem arises from the current reliance on chemical analysis methods, which are invasive and delayed, preventing real-time, non-invasive, and online monitoring. To solve this issue, a non-invasive system is proposed based on the use of process data and Deep Learning models programmed in Python. The methodology consisted of constructing two datasets: an endogenous one, based on the past values of the target variables, and an exogenous one, composed of operational process variables, thus simulating a real industrial scenario where past chemical measurements are not always available. This approach allowed evaluating the capability of the models in both contexts. Three predictive architectures were implemented and compared: a linear regression model (serving as the baseline), a recurrent neural network LSTM (Long Short-Term Memory), and a Temporal Fusion Transformer (TFT). Performance was assessed using Mean Absolute Scaled Error (MASE) as the main metric, along with MAE, forecast bias (FB), Pearson’s correlation coefficient, R2, and normalized mutual information (NMI). The results show that linear regression provides a solid basis for capturing general trends, but is insufficient when facing the high variability of the series. The LSTM model achieved the best overall performance, demonstrating robustness in estimating both copper grade and the iron–silica ratio in the endogenous dataset. Meanwhile, the TFT exhibited better behavior in the exogenous dataset by exploiting non-linear relationships among process variables, although it showed generalization difficulties in the endogenous case. In conclusion, this work analyzes the feasibility of building a non-invasive monitoring system for the copper industry, with the potential to be integrated into online control systems. This approach contributes to the optimization of pyrometallurgical processes, offering benefits in safety, cost reduction, and operational monitoring.
Description
Tesis presentada para optar al título de Ingeniero/a Civil en Telecomunicaciones.
Keywords
Procesamiento de datos, Inteligencia artificial Aplicaciones industriales