Advanced Analytics for the Secondary and Tertiary Crushing System, El Teniente Division - Development and Innovation Management

In this project, an integrated system with the operation was developed to improve the characterization of the secondary and tertiary crushing plant, as well as to predict and recommend situations to improve the performance of the process. For this purpose, a first part was developed that allows characterizing the different stages of the process and estimating variables of which there are no associated sensors in the plant. Later, a system of recommendations was implemented based on different optimization criteria of the plant operation, to improve the performance according to relevant KPIs. Finally, a connection with Teniente Codelco’s data was created to perform an integration with the existing system, as well as with several interfaces for communication with the user and the administrator of the system.

Firstly, a thorough study of the plant’s variables was carried out, including an analysis of the existing relationships between them, seen both at the process and associated expert knowledge, as well as at the level of the data and its temporal relationship. Based on this, a pre-processing step of the data was carried out, in order to have the most polished information for modeling purposes. Along the same line, characteristic plant operation modes were verified, which are related to the availability of equipments and the needs of the plant at a given time.

Subsequently, a system was developed for the on-line estimation of different variables of the secondary and tertiary crushing process. Particularly, the ore flow in different process belts, several of which do not have an associated measurement sensor. The estimation made of the variables is based on a Bayesian probabilistic model, where the different available measurements are mixed, along with causal relationships among the variables that characterize the behavior of different parts of the system, in order to update the values that we have of both the observed variables and unavailable variables (virtual sensors). Thus, this modeling works as a digital twin of the process, allowing to verify both the behavior in specific parts, as well as through relevant indicators about the overall performance of the system.

Finally, a system was implemented that allows recommendations to be made for potential situations of improvement in the performance of the entire plant process. According to a characterization of the current situation of the process, the system generate possible future scenarios based on the available historical data. Then, integrating the information of the scenarios together with machine learning predictions of interest variables, the system performs an optimization of the subsequent performance of the plant, for different general criteria (e.g. increase in production, energy efficiency, among others). Subsequently, according to an optimal future scenario (considering relevant system indicators, as well as operational criteria and restrictions), a recommendation is generated to the operator, maintaining or modifying the values of the manipulated variables. These recommendations are stored in the system, so that different users can consult them and obtain descriptive statistics of the effect of the generated recommendations.