Neural Primary Discharge Control

Application

Primary crusher discharge rate control in mining operations is a complex problem that involves multiple variables and uncertainties. There is a substantial transfer delay between the control variable (apron feeder speed) and the process variable (weightometer), which makes modelling the process essential.

While empirical models have been successfully applied for apron feeder control, data-driven methods offer greater adaptability, accuracy and robustness that is more suitable for highly variable, multivariable control.

Variability in apron feeder feedrate originates from:

• Changing ore properties: different ores have different hardness, size distributions, and other properties that affect the crusher’s efficiency and, consequently, its discharge rate.
• Mining from different locations: ore from different sections of the mine can vary significantly in quality and characteristics.
• ROM bin geometry: the geometry of the Run-of-Mine (ROM) bins can cause transient changes in the feedrate.
• Tipping material: during the feeding process, the manner and rate at which material is tipped into the crusher can create fluctuations in the feedrate.

Because of the above factors, creating an accurate model to predict and control the feedrate is a complex task. The model must account for all these variables and be robust enough to handle fluctuations and uncertainties. This generally involves a combination of data-driven techniques, utilising machine learning, coupled with more traditional control theory approaches. The model must not only be accurate but also capable of running in real time, providing timely feedback for control actions. This often requires highly optimised computational techniques. The model’s accuracy is only as good as the data fed into it — incorrect or noisy sensor readings can lead to incorrect predictions and poor control actions, so ensuring high-quality, reliable data acquisition systems is crucial.

Solution: Time Series Forecasting for Neural Control

Time series forecasting can play a significant role in solving the problem of apron feeder control by providing accurate predictions of material flow rates. By analysing historical data and using forecasting techniques, engineers can make informed decisions to optimise the operation of apron feeders, or the control can be automated based on the predictions. Since there are multiple sensors and other data sources, AI-based models can efficiently bring together this data and create accurate predictions.

The model estimates near-term discharge from feeder speed, measured flow and the other signals available at the site. The model type is selected against the process dynamics and deployment hardware; a CNN is one possible implementation, not the product definition. The important output is a validated forecast with a known horizon, confidence and fallback response.

The forecast can also be used to compare alternative control strategies in a staged or simulated environment. This gives the site a way to test response quality, constraint handling and fallback behaviour before a strategy is approved for live use.

The same forecast can support predictive alarms when its lead time and error profile are understood. Alarm thresholds, operator response and any automatic action are defined against the site’s operating procedures.