Comparison with Current Control
The realization of a control algorithm involves a number of practices including programming and integration with production environment . This step can be unexpectedly long on account of technical resources. In every case, it is possible to make comparisons with the current control used and somehow determine if the new strategy is able to detect or prevent undesirable deviations in the process. At the time of writing this article, the control algorithm was not yet implanted in production environment, but we performed a simulation of what control actions would be taken for every minute of operation.
In Fig. 9 we plotted the results of a comparison over one work day in the plant. The light yellow line with circles at the top refers to the calcine temperature; the dark line with triangles is the actual control actions taken, and the dark blue line with circles at the bottom refers to the ore load. The gray lines represent optimal the interval for the control that was determined via the rules found by the decision tree or general association rules algorithm.
Some interesting and worth noting cases are discussed as follows. During the preheating phase, the lining may heat up very quickly while the material that remained in the kiln takes longer time to heat, causing sudden variations in temperature profile. Moreover, during preheat the ore load and rotation is low, leading to a long response time of the calcine temperature, and because of that, we take into account only ore loads greater than 50 T/h. In this figure, although the lining heated up quickly, the material that came afterwards was not sufficiently heated, as the kiln rotated faster. The algorithm suggested for that state (ore load around 50 T/h) one extra ton for coal dosing (3.5 T/h instead of 2.5 T/h).
The numbers in the figure indicates some other cases:
Fig. 9 Comparison of this data analysis approach with current control
- 1. The calcine temperature was in a rise trend and the algorithm suggested a lower level of fuel
- 2. Same case of 1
- 3. Forecasting a fall in the calcine temperature, suggested an additional fuel load
- 4. Although the calc. temperature was falling, the algorithm caught signs that the temperature was about to rise again, and suggested a lower level of fuel after having suggested a higher level
- 5. To prevent temperature rise, suggested a lower level
- 6. Some event may have cause calcine temperature to fall in the next hours and the control suggested a high load of fuel.
It is worth noting that this simulation should be dynamic, i.e. if the control actions taken by operators were the same of the suggested ones, the next suggestions would surely be different, provided that the observable variables would have changed. This may explain sudden and abrupt variations in the coal dosing suggestions. Since this algorithm is purely based on statistics, it is very likely that the kiln eventually matches a rule, which would suggest a very different coal dosing level than the previous or current level. Fuzzy decision trees  would certainly prevent such sudden changes in suggestions, as the bounds between rules would be soft and gradual instead of crisp hard limiting. It should also be considered that this algorithm is based on real data, events that actually happened in the plant, so new events that can affect calcine temperature are not detected. In any case, actions that an experienced operator or an expert in kilns would take are likely to be represented by these rules; therefore, new staff will have access to this knowledge by means of this system.