With deeper mines comes increases in ore hardness, decreases in grade, and larger variations in material properties - optimized management of the grinding circuit has become more critical to maintain and increase throughput and minimize specific energy consumption whilst maintaining grinding effectiveness. The IntelliSense.io team was responsible for the following key tasks: reducing mill shutdowns through predictions of ball charge and liner states, reducing the frequency of SAG mill overloads, and increasing the controllability of the grinding circuit.
- Limited knowledge of changing ore properties lead to;
- Frequent mill overloads
- Reduced throughput
- The instrumentation on the mills is not sufficient to run the mills with optimum ball charge and knowledge of the liner state, leading to;
- Frequent mill stops for inspections
- Non-optimal running conditions.
- Traditional sensors cannot be used within the mill due to the nature of the balls
The IntelliSense.io Grinding Circuit Optimization application has been piloted at the Aktogay Plant of JSC "AK Altynalmas". A Digital Twin of the SAG mill was created which enables improved monitoring, performance prediction, and optimization to increase profitability.
A combined first principle and machine learning model virtually monitors and predicts ball charge to keep it in the optimal range for throughput and grinding effectiveness without having to stop the mill for inspections. A Material Model was created to provide visibility of streams that cannot directly be measured, providing increased control of the circuit for operators. An Overload Model predicts the profitability of an overload occurring in real-time, allowing early interventions to keep throughput consistently high.
Outputs of the material model include flow rates as well as the solid/liquid ratio in slurry streams. For example, the total solids flow rate to the ball mills (which cannot be measured directly) is one of the key deliverables. Additional outputs include an estimate of the outlet flow rate from the sump and other areas with broken sensors and estimates of the total process water added at various points along the comminution circuit which are not currently measured.
The overload predictions have two parts: Labelling and Overload Model. The Labelling is where historical data is analyzed to mark the overloads that we are trying to predict. The Overload Model is a machine learning model that is trained to classify live data in order to predict historic labels. The output of the model is a probability of an overload occurring within the next 20 minutes, which can trigger an alert.
The following key benefits will be realized through this application:
- Increased stability of the grinding circuit
- Reduced mill stoppages for liner and ball charge inspections
- Increased throughput