Intelligent Image Recognition for SAG Mill Grinding Media

Company: Compañía Electro Metalúrgica S.A. (Elecmetal) — Developer: IO-link
Funding: CORFO Innova Chile — Grant no. 16ITE1-70889 (120k USD)
Role: Solution Architect & Developer

Context

SAG (Semi-Autogenous Grinding) mills are central to copper ore processing in Chile’s large-scale mining operations. These mills are continuously fed with Steel Grinding Media (MMS) — steel balls of varying diameter and hardness — at an average rate of 30 tonnes per day, in campaigns lasting 4 to 6 months. The MMS represent the second-largest operational cost item in grinding, after energy, which itself is directly affected by MMS efficiency.

Despite this economic significance, MMS monitoring today is conducted in aggregate terms only: tonnes consumed per day, correlated with historical productivity indices. There is no practical, real-time, quantitative system capable of answering the questions that actually drive operational decisions:

• How many individual MMS units pass through the mill per day?
• What is the average residence time of an MMS inside the mill?
• What condition do MMS exit in — intact, fractured, or deformed?
• What is the exit diameter of each ball?
• Can worn MMS be reused in downstream grinding stages?
• Are there correlations between mill operating parameters and MMS output quality?

Solution

This project developed a machine vision prototype that automatically detects, classifies, and measures MMS exiting the SAG mill through the screen (harnero) area, operating 24/7 without human intervention and transmitting real-time KPIs to a cloud-based web dashboard.

Detection & Classification Algorithm

The algorithm pipeline processes video frames through four sequential stages:

1. Template correlation: For each frame, a non-parametric correlation against a set of MMS reference templates is computed pixel by pixel over a reduced transverse region of interest (ROI). This region-of-interest constraint reduces computational load while preserving detection accuracy in real time. Areas of maximum correlation are flagged as candidates.

2. Neural network classification: For each candidate region, concentric pixel strips (rays) are extracted from the center point at 0°–360°, concatenated into a single N-dimensional vector, and fed into a one-hidden-layer neural network that outputs: intact ball, fractured ball, oval ball, or deformed ball. This encoding abstracts away background noise and overlapping structures.

3. Trajectory analysis: Candidate regions are tracked across successive frames using a covariance analysis and distance-based clustering. A linear regression fit on each track determines whether the object’s trajectory is consistent with natural harnero movement, discarding false positives. A “string algorithm” reduces combinatorial complexity in point-to-track assignment.

4. Classification & diameter estimation: Confirmed detections are classified by ball type and their diameter is estimated, contributing to a cumulative mass-of-steel register over time.

Distributed Camera Architecture

Due to the conflicting requirements of detection (real-time, lower resolution) and classification/measurement (high resolution, latency-tolerant), the system splits processing across two camera groups:

• Detection cameras: dedicated to real-time MMS detection, transmit (x,y) coordinates of confirmed candidates
• Classification/measurement cameras: receive coordinates, capture high-resolution crops, run the classification and diameter pipeline

A central field PC coordinates both groups, formats results, and transmits them to the remote server via Internet (or cellular backup). A web application hosted on the cloud server renders real-time KPI dashboards accessible from any device.

Software Stack

Hardware

Field hardware was revised mid-project based on findings at Minera Centinela: the initially proposed distributed Raspberry Pi architecture was replaced by an industrial PC with industrial cameras, given the hostile mine environment (dust, vibration, outdoor exposure) and limited connectivity/power infrastructure at the harnero installation point.

Project Progress (at reporting date)

Neural network training required field video captures at Minera Centinela under varying lighting, camera positions, and zoom levels — a step that proved critical given the algorithm’s sensitivity to site-specific visual conditions. Based on this finding, the project moved Stage 4 (field validation) earlier in the timeline, conducting field visits in parallel with development to build a representative training image set.