copper-flotation-cell

Summary

Rovjok developed and deployed a virtual sensors system to improve the reliability and accuracy of grade monitoring in a large-scale copper flotation circuit. The sensors were built using machine learning modelling, using data from multiple surrounding equipment sensors to provide accurate predictions of copper grade for material pumped to the tailings management facility. Rovjok deployed a fully automated, near real time cloud-based solution, which delivered results back into the on-site operational system. The virtual sensors provided full redundancy when the physical stream analyser equipment was offline, and in many cases was more accurate than the physical measurements. This double benefit enabled the customer to make better operational decisions to always keep flotation performance high.

The Client

The client operation is one of the largest copper mines in the world by processing capacity, located in Latin America.

The copper flotation circuit is a critical stage in the processing flowsheet, as it is responsible for the recovery of copper from the ore. Ore is concentrated via froth flotation and collected off the top of the flotation tanks. Any unrecovered copper that reaches the end of the flotation cells is sent to the tailings as waste, so maximising the flotation recovery is a crucial component of improving a mining operation’s overall financial performance. Any lack of understanding, by the operators, as to the current recovery performance leads to uncertainty in true recovery and sub-optimal circuit performance. Online sampling devices are commonly used to monitor copper grade at various points within the circuit, most importantly the flotation feed, within the flotation cells, and prior to the tailings feed.

For a simple circuit, the recovery can be calculated as:

R = 100C(F — T)/(F(C — T))

where,
R = Recovery (%),
C = Final copper concentrate assay (%),
F = Flotation feed assay (%),
T = Tailings assay (%).

Figure 1 shows a simplified view of the recovery circuit at the client operation for single rougher train with two banks of cells. The project focussed on the rougher section of the flotation circuit shown in the figure, as this is the first section downstream of the mill cyclones and represents where the majority of copper is recovered.

The Challenge

The operation used an online x-ray fluorescence (XRF) multi-stream analyser (MSA) system to track copper grades in near real-time throughout the flotation circuit. The locations of some key MSA measurements are denoted in Figure 1. This system received regular samples from the flotation cells and provided copper (and other element) grade measurements every 5 minutes, giving operators an up-to-date view of the recovery performance. Adjustments to pH levels through reagents are made to maintain high recovery levels, depending on the coarseness and mineralogical characteristics of the incoming ore slurry. Separately to the MSA system, during each 12-hour shift, slurry samples are collected, which are then sent to a sampling laboratory for accurate grade assay measurements. These are then used to crosscheck the historic performance of the MSA system over the previous shift, and if any new calibration of the XRF device needs to be performed.

flotation circuit rougher stage - flow chart
Figure 1: overview of the flotation circuit rougher stage. Slurry enters from the mill cyclones into the two surge tanks on the left hand side, and is then pumped through four parallel rougher banks (only one train of two rougher banks shown in figure). Copper not recovered by each of the rougher banks is sent to the tailings management facility (TMF), shown on the right of the flowsheet. The virtual sensors were developed for the RT-1 and RT-2 physical sensors (with two more for the second flotation train).

However, the MSA system was unable to operate as reliably as needed. Frequent downtime in the operational environment (e.g. sensor malfunction, material blockages, x-ray source replacement) meant that appreciable proportions of operator time were spent with no live understanding of the copper grade and recovery performance. When the MSA system was offline, it simply returned the last available value. Therefore, providing operators with more reliable visibility was the main goal of the project, with a secondary goal to improve upon the relative accuracy of the MSA system compared to the laboratory assay results.

The sensors recording the copper grade sent to the four tailings pumps were the first set of targets to be virtualised, as these were the most unreliable and had the highest measurement uncertainty.

The solution

Rovjok’s solution was to develop a virtual-MSA system: a suite of machine learning virtual sensors, which utilised surrounding sensor datapoints and provided regular predicted copper grade values alongside the physical MSA system’s measurements.

15-minute sampled timeseries data from many surrounding physical sensors and equipment, including slurry feed pumps, density measurements, and copper feed grades from a separate high reliability sensor. The training dataset covered a period from January 2021 to May 2022. An extensive, iterative data science process was then performed, covering the following stages:

  1. cleaning the input data, removing periods of poor operating conditions based upon pump flowrate thresholds specific to the circuit,
  2. understanding the time-lag of the input datapoints relative to the position of the MSA measurement in the processing flow, and adjusting the relative time axes accordingly,
  3. feature engineering to define the set of best performing input parameters for predicting the copper grade,
  4. regression model training for different standardisation pre-processing and ML algorithms,
  5. model performance analysis on a test dataset.

Once the set of ML models were created with acceptable accuracy to the client, these were setup within a fully automated data pipeline. Data was streamed from the on-premise PI System environment, to Azure cloud-based ML prediction generation infrastructure, and back to the on-premise PI System for real-time ingestion on operator dashboards. An overview of the full system is shown in Figure 2.

Figure 2: VMSA system architecture. Data is received from on-premise PI System operational data system, sent to the cloud-based ML flow, with results returned to PI system. The cloud environment used numerous Azure systems, including SQL databases to store ETL settings, Azure Data Factory to handle the ETL, data lake storage, and Azure Functions to call Azure-hosted ML models to generate the predictions.

Results

The VMSA system is able to track the copper grade sent to all four tailings pumps with 100% reliability, so long as the flotation circuit itself is operating. This presents an important boost to operator visibility compared to the MSA system, and alleviates the issues of maintenance or unplanned downtime of the physical MSA machine.

The VMSA system therefore keeps operators informed as to the copper levels being sent to tailings and therefore of the recovery performance. This allows continued informed operational decision-making. An additional bonus was the ability to produce models that ‘measure’ the copper grade at a higher accuracy than the MSA system, lowering the uncertainly on the inferred recovery and therefore more truly representing the actual intrinsic performance of the circuit at a given moment in time. More accurate tracking of the intrinsic recovery lowers the discrepancy that has to be factored during end of period metallurgical accounting reconciliation. For the client, a single typical day where the MSA system is unavailable and a 5% deviation between ‘last-known’ and intrinsic recovery rate, corresponds to around $370,000 of invisible revenue (at current copper prices), which needs to be captured during reconciliation.

Lab MSA VMSA
Figure 3: dashboard tracking the averaged performance of the MSA and VMSA systems compared to the laboratory assay measurements taken for each 12 hour shift.

The development of the VMSA has revolutionised our approach to real-time optimisation of plant operations. Beyond the tangible benefits it has delivered in terms of recovery gains and cost reductions (on reagents and energy), the development journey is where most of the value materialised; it has deepened our understanding of the recovery circuit’s dynamics and intricate interactions, it has refined our data management practices and expanded our data processing capabilities, it has fostered effective cross-functional teamwork by bringing together experts from different disciplines and, finally, it has paved the way for future data-driven initiatives across our organisation. Rovjok’s successful implementation of the VMSA has reinforced the importance of the digital transformation within our industry, and will be a springboard for further innovation and optimisation throughout our business.

Frederic Wouters, Process Plant Technical Manager, FQM Cobre Panama

Conclusions

The virtual sensors ‘VMSA’ system developed and deployed by Rovjok at a large-scale copper mine has provided a reliable monitoring capability for copper grade and recovery tracking independent of the physical MSA system. This has given the flotation circuit operators continual oversight into flotation performance even during sensor downtime events in the MSA system, allowing for better decision-making at all times to maintain high recoveries, given the ore conditions received from the grinding circuit.

If you would like to learn more about how Rovjok can utilise advanced data science methods to help optimise your operation, read about our other solutions, or get in touch.