Por: Jean Campos, Manuel Berrospi y Jossy Atoche, Minera Chinaco Perú. 

Trabajo ganador en Operaciones Mineras y Gestión de Activos en el Foro TIS de PERUMIN 37.

Abstract

This study describes the development, training, and operational implementation of an anomaly detection model based on convolutional autoencoders, applied to the overland conveyor system of the Toromocho mining project. This 5 km-long conveyor is a critical asset in the production chain, connecting the primary crushing area with the concentrator plant’s stockpile. Given the system’s complexity -with more than 370 sensors integrated into the PI System-, a modular and scalable monitoring architecture was designed to enable early detection of operational deviations across six functional systems: pulleys, master CSTs, slave CSTs, winch, braking system, and operational variables.

The model was trained using representative data from normal operating conditions, collected between September and December 2024. One-hour moving windows with a 30-second sampling frequency were used, and data were normalized with a MinMaxScaler. Anomaly detection was based on the autoencoder’s reconstruction error, classifying as anomalous any observations exceeding three standard deviations from the mean.

For field implementation, a dedicated analytics server was deployed to execute the six models every 30 seconds. The results are stored in an SQL database and integrated into the PI System for visualization. The model outputs a health index ranging from 0 to 100% for each component, system, and the conveyor as a whole, classified into three levels: normal, warning, and critical.

During the first quarter of operation, the model enabled early detection of events such as oil contamination in CST, hydraulic system leaks, and accumulator depressurizations; in some cases, more than three hours in advance. These results validate the effectiveness of the proposed approach and demonstrate its potential to optimize maintenance planning, reduce unplanned failures, and advance toward predictive and intelligent mining.