Top of page
ABSTRACT
This paper details development of a Fuzzy Logic Controller at Carajás Mine and its implementation within Comdale/X - an off-line Expert System Development Tool. The controller provides advice about operating two 80m diameter slime thickeners. Creation of the fuzzy sets and rules involved a Knowledge Engineer and Expert-Operator working together for a period of about 8 hours. System refinement and addition of adaptive techniques took another day's work. The controller is based on a heuristic model using the expertise of one of CVRD's most-experienced operators.
1. BACKGROUND
Companhia Vale do Rio Doce operate one of the world's largest iron ore mines at Serra dos Carajás, Parauapebas, Para State, Brasil, some 550 km south of Belem. The mine currently produces over 40,000,000 tonnes of ore per year at ~ 67 %Fe. Ore is crushed, screened and milled to produce lump ore and sinter feed which are shipped by rail 889 km to the port of Saõ Luis do Maranhão.
In developing the Carajás project, CVRD has taken great care to protect Amazon eco-systems. Of the 412,000 hectares of native rainforest under the mine's direct control, less than 1.6% is taken up by the mine, plant and residential townsite. In all, CVRD is responsible for 1.2 million hectares of rainforest as well as providing support to 320,000 hectares of ecological reserves belonging to IBAMA - Brasil's Environmental Protection Agency. The forest extends right up to the edge of all mine development and much work is expended to preserve the natural environment of this beautiful, yet sensitive region.
As part of this undertaking, all waste slimes from the plant are thickened in two 80m diameter thickeners prior to being stored in a tailings dam nearby. These thickeners serve to recycle precious water for use in the processing plant so as to reduce the fresh-water requirements of the mine. In operating these units, the underflow density target is 65 %solids with an overflow density as low as possible (typically below 0.3 % and never above 1.0 %).
The thickeners are subject to considerable variations in feed flowrate and solids content as different parts of the plant shut-down or start-up due to maintenance or other reasons. Ore changes also affect the operation in terms of fines content and the presence of Al2O3 material (clay minerals). The underflow density is controlled manually by adjusting the flowrate of the underflow pulp. The flowrate is controlled by variable speed pumps.
Proper operation of the thickeners is essential to maintain plant performance. Water requirements must be met and any delays in dewatering can affect plant production. In analysing thickener data, it was apparent that certain operators were more skilled than others at running the thickeners. Accordingly, it was believed that operations could be improved through use of a standardized control procedure. Fuzzy Logic was selected as a rapid way to build such a system by using the expertise of one of the mill's most experienced operator.
Fuzzy Systems provide superior adaptation than does conventional PID or Supervisory Control. Set-points are elastic within the controller and respond to external knowledge. There is also inherent reduction in control action whenever a set-point is approached. Rapid prototyping was also judged to be an advantage of a Fuzzy Systems approach
2. FUZZY CONTROL SYSTEM
Two basic control loops have been established: one to control underflow density by varying the underflow pulp flowrate; the second to regulate overflow turbidity by adjusting the addition rate of a flocculant reagent. The rules-of-thumb which make up the knowledge base were derived from the advice of an experienced plant operator.
These rules were placed into two Fuzzy-Associative Memory (FAM) controllers within the Comdale/X off-line Expert System environment. Creation of the fuzzy sets and rules involved a Knowledge Engineer and Expert-Operator working together for a period of about 8 hours. Refinement of the system and incorporation of adaptive techniques required another day of work. Testing was conducted during the first Quarter of 1995.
For simplicity, Correlation-Product Inferencing and Weighted-Average Defuzzification were chosen to convert each variable and its rate of change into the desired output for each FAM module. Figure 1 depicts the overall structure of these two interacting FAMs. Dynamic adaptation using meta-rule over-ride is used to overcome interaction effects.
The system has been implemented in an off-line environment on a PC located in the process metallurgist's office. Each hour the plant operator calls the office for advice after recording the current thickener data. Mean residence time of material in these thickeners is approximately 15 hours, however, control system response is typically about 1 hour as the solids pass through the process much faster than the water and slimes. The controller is fed the current and past data readings to advise on new set-points for the underflow flowrate and flocculant addition. These recommendations are passed on to the plant operators for implementation.
Figure 1. FAM modules for Thickener Control.
2.1 Underflow Density Control
The rules used in the U/F Density controller are given in Table 1. Seven fuzzy expressions are used to characterize U/F density while five terms describe the change in density between readings. The U/F flowrate change is defined by seven output fuzzy sets.
Figure 2 shows the set definitions in current use. These have evolved from testing the controller against an Artificial Neural Network model of the thickener. The model was built using 4 inputs (time interval, feed flowrate, U/F flowrate, Al2O3 content), 5 hidden nodes and 1 output node (U/F density) in a 3-layer back-propagation network with bias. With 182 data patterns, the network was trained to ~ 8% average error (14% max) in about 10,000 iterations using a Cumulative Delta Rule on a 486-33 PC. It is planned to use this network to tune the controller regularly. The network will be expanded to model the overflow density as a function of flocculant addition and flowrate conditions.
Table 1. U/F Flowrate Fuzzy Associative Map
Several meta-rules are used to over-ride the U/F FAM Controller and prevent the thickener from becoming unstable. These rules include the following:
These meta-rules are currently under study. They are only important under conditions where the thickeners have become unstable - a situation that is avoided at all costs and, hence, difficult to examine.
2.2 Overflow Turbidity Control
Development of the O/F Turbidity Controller is focused on the addition rate of flocculant. This reagent causes solids to settle faster generally yielding a low O/F turbidity and helping to increase the U/F density. At very high additions, the flocs formed can retain water leading to difficulties in attaining the required U/F density, but this is rare. Power draw on the thickener rakes must also be watched closely under these conditions as the solids bed can form very quickly leading to increased torque requirements.
The FAM mapping for this controller is given in Table 2. Direct output of flocculant addition rate is generated rather than a change in flocculant addition. This reduces cycling of the addition rate.
Table 2. Flocculant Fuzzy Associative Map
The fuzzy sets for this controller are depicted in Figure 3. These are the original set definitions and may be changed once studies are completed. Their formulation was done during a series of descussions with the plant "Expert". During controller testing several of the sets were modified and a number of output rules were adjusted to adjacent linguistic expressions.
The system's goal is to maintain the O/F density as low as possible while using as little flocculant as possible. Several tests are being conducted to establish the utility of replacing at least a part of the flocculant addition with lime. The ability to detect clay content may determine the viability of this idea.
Similar to the U/F Density system, several "meta-rules" override the FAM controller when certain extreme conditions are met:
Controller output is adjusted by clay content based on the concept that high clay in the ore requires more flocculant while a low level needs less reagent. Fuzzy sets describing clay content give a scaling factor to adjust the O/F FAM defuzzification process.
Clay content, at the moment, is inferred since reliable data on Al2O3 content is unavailable to the operator and often, is not reported by the assay lab for up to 12 hours. Sometimes the mine switches ore between assay-reporting periods preventing useful application of the knowledge. A metallurgist can estimate Al2O3 content based on experience and current operating conditions. If turbidity and past assays are "low", the ore is considered "good" allowing flocculant to be reduced. On the other hand, if turbidity and past assays are "high" indicating "high" Al2O3 content, then the ore is considered "poor" leading to increased flocculant.
Changing ore conditions causes changes in thickener response particularly with regard to turbidity. Consideration is being given to developing an on-line Al2O3 analyser.
3. USER INTERFACE
The system operates within the Comdale/X environment and can interface directly with the ANN model when required. Controller input uses the form shown in Figure 4.
All data are available from thickener instruments: U/F density and flowrate are measured directly in the discharge line while O/F turbidity is monitored using a thief sampler located approximately
0.5 m below the O/F weir. Actually, several samplers at different depths allow an operator to quantitatively determine the solids bed-level within the thickener. It may be possible to modify this sampling system to input bed-level to the controller.
Figure 5. Underflow Flowrate Output Report.
The three data points ( U/F Density, U/F Flowrate and O/F Turbidity) are input to the system every hour by the plant metallurgist following communication with the plant operator. The data for the two thickeners are averaged as the total flowrate is distributed equally to each thickener.
The U/F FAM output is shown in Figure 5. A User can see quickly which DoBs are currently in use by the system and can understand how the recommended output is derived. In the example given, the density is above the setpoint but is declining. So, the controller recommends a small drop in the U/F flowrate to return the thickener to steady operation. An additional input planned for the future will account for input flowrate changes from various parts of the overall process.
Figure 6. Flocculant Addition Output Report.
The flocculant addition output report is shown in Figure 6. In the situation presented here, the current O/F turbidity has decreased from 0.9 to 0.7. The clay content is about normal at 0.6 % Al2O3 so a scaling factor of 1.04 is applied to the output defuzzification. The important rule output is "above-normal" so the recommended setting is 7.3 m3/hr.
In both controllers, the User has the right to over-ride the system and decide on different setting. A series of tests are now underway to verify the approaches being recommended by the system and to note situations where the operator uses an alternate value. These cases will help to derive new meta-rules, to adjust current fuzzy set definitions or to derive new FAM rules. Adjustment of the controller is relatively easy using an MS/Windows Menu-driven Editor.
Eventually, the system will be used to implement the decisions directly by transferring the knowledge base into a real-time SCADA environment such as ProcessVision. We are presently evaluating the ability of this controller to return the thickener to stable conditions following an upset. Current work with the ANN model suggests that some upsets produce significant oscillatory behavior in the controller while other disturbances indicate steady-state can be achieved in about 1 hour. Closer control provides cleaner recycle water and higher water recovery rates.
4. CONCLUSION
Development time for this system was remarkably swift - it is believed no other control philosophy that provides immediate results can match the power of a Fuzzy Logic-based system. Use of AI tools such as Fuzzy Logic and ANN models has proven successful.
ACKNOWLEDGMENT
The authors wish to thank CVRD for permission to publish this paper. The enthusiasm of our Expert - Luiz Augusto Mapa, is gratefully acknowledged.
Return to the publications index