grind size, metal recovery, concentrate grade, etc.) inevitably becomes overwhelmed by the effect of the ever-changing feed, which can never be held ‘constant’. This is the largest and most important difference between experiments undertaken in a well-controlled laboratory setting, where small changes in a response have a much better chance of being observed, compared to a production environment, where many dominant variables that affect the targeted response are beyond the control of the experimenter. Another difficulty, which is probably less appreciated, is the inherent error associated with all experimental measurements. This stems from many sources, including the representativeness of samples obtained from the circuit, the precision of the instrumentation used to undertake the measurements, and operator skill and training. The implication is that even if the same operator was to undertake repeat measurements on a given sample, the resulting dataset will comprise a distribution of values. All measurements have error, which can never be eliminated, and therefore must be accounted for when undertaking rigorous data analysis. In mineral processing plants, these unfortunate truths conspire to camouflage the true effect of experiments that operators and production metallurgists must undertake to identify strategies that improve processing performance. Consider a copper concentrator processing ore of 1% Cu grade at 1000 t/h. A recovery improvement of 2% would be worth $11 M USD per year, assuming 8000 operating hours per year and a conservative copper price of $7000 USD per tonne. The problem is: how can this small change be detected in production data where recovery could vary by 10-20% (Figure 1)? Figure 48 – Actual metal recovery data from a typical concentrator, ranging in this example from 82-94% recovery. The answer lies in the application of mature, statistical concepts to design well-structured experiments that block out the effects of nuisance, uncontrollable variables and random error. The effect of the trial can then be identified with the use of statistical analysis that separates it from ‘noise’ in the data. The result is that, if an effect is real, it can still be observable, even in the midst of overwhelming variability in the data. Conversely, if the trial has had no effect, this will also be apparent. In either scenario, well-informed business decisions can then be made, where the risk of 80 82 84 86 88 90 92 94 96 0 10 20 30 40 50 60 Recovery (%) Day
RkJQdWJsaXNoZXIy MTM0Mzk2