Top of page
Edgardo I. CHILVIETa,b; John A. MEECHa; Andrew L. MULARa;
Hans RAABEc,d; John MITCHELLc
a The University of British Columbia,
Department of Mining and Mineral Process Engineering, Vancouver,
B.C., Canada
b Now with Fluor Daniel Wright, Vancouver, B.C., Canada
c Highland Valley Copper Mine, Logan Lake, B.C., Canada
d Now with Lakefield Research, Lakefield, ON, Canada
This paper describes the fundamentals of qualitative model construction and implementation within ProcessVision, a real-time intelligent SCADA software package. Details of the application of the system to assist in interpreting the operation of a semiautogenous grinding circuit at Highland Valley Copper mine are given.
The main objective of this project is to develop and implement an on-line Intelligent Supervisory System (ISS) that employs a qualitative approach to model and supervise the operation of a complex process.
The developed ISS has been applied to the C-line grinding circuit at Highland Valley Copper (HVC) to assist in automatic operation. The initial focus of this system is to monitor and detect tonnage restrictions that affect circuit production.
The system is designed to provide on-line advice via an intelligent SCADA interface to the shift supervisors in the control room regarding the occurrence of circuit restrictions. At this stage the system does not intervene directly with the process. Appropriate actions are left to the discretion of plant operators.
Periodic reports summarize diagnostic results; these are used to assess the extent and frequency of production-limiting bottlenecks. Changes in operating philosophy are expected to derive from this analysis. On-line information, together with reports summaries, can reduce the decision-making time. This can lead to removal of a restriction fast enough to reduce tonnage losses. Average losses due to grinding circuit restrictions range up to 5000 tonnes per week.
This paper gives a brief description of the circuit and the analysis of tonnage restrictions. The concept of Qualitative Modeling (QM) using Fuzzy Logic to input or output data is described. The paper concludes on the ability of the QM approach to improve plant operations and address tonnage losses.
Feed for C-line is reclaimed from a stockpile by five variablespeed feeders (see Figure 1). Primary grinding is achieved by a SemiAutogenous Grinding (SAG) mill (43'x16') equipped with grate discharge. The mill feeds two grizzlies where the undersize is split out. Oversize returns to the SAG mill, while the undersize is sent to two ball mill discharge pumps. Mill power derives from two variable-speed 4700 KW motors. Feed tonnage ranges from 1200 to 1800 tph.
Figure 1- C Grinding Line at HVC.
Secondary grinding occurs in two 16.5'x27' ball mills operated in closed circuit with a cluster of ten 760 mm cyclones. Cyclone overflow discharges by gravity to the flotation plant where copper minerals are extracted.
A Bailey Distributed Control System interfaces with the grinding plant. The global objective of the control strategy is to maximize tonnage at maximum power draw. Fresh feed is adjusted by one of the two main control loops: SAG mill power draw or SAG bearing pressure. A switching logic program determines which loop manipulate the fresh feed at any given time. There is also a second algorithm to adjust the speed of the SAG mill according to the current rate. All the other major process variables (water addition, cyclone conditions, etc.) are also monitored and adjusted by the control system.
A tonnage restriction exists when the tonnage fed to the mill cannot be maintained at, or increased above a specific value due to the presence of certain operating conditions. Tonnage restrictions prevent the control system from achieving maximum tonnage so a tonnage loss is attributed to their occurrence.
The most frequent type of restriction occurs when the tonnage cannot be maintained at a specific value. Two cases are distinguished:
The second type of restriction is when the tonnage has reached the absolute maximum that can be delivered to Cline. Tonnage is limited at this level even though the SAG mill could still handle more ore. This is generally known as a "soft ore" restriction.
Raabe developed a heuristic procedure to evaluate the occurrence of restrictions, to determine possible causes, and to estimate the tonnage lost in each case [1]. The procedure is performed manually on a weekly basis requiring intensive examination of trend graphs as well as information from the control room.
Much of the procedure is based on subjective judgment; hence, different people will obtain different results. As well, the knowledge is only available well after the restrictions have occurred. It was considered that an online system to perform this analysis automatically could help to standardize the procedure and decrease the frequency and duration of such restrictions. The application thus, attempts to mimic how experienced plant metallurgists think about and interpret circuit upsets rather than modeling data extracted from the plant directly.
Qualitative modeling is one of the central aspects in the design of the ISS. It provides the mechanisms required by the ISS to handle the kind of subjective judgment and approximate reasoning found in the diagnosis of tonnage restrictions.
A pseudo-qualitative approach to qualitative modeling proposed by Cifuentes is employed on this application [2]. Under this approach, a qualitative model (QM) is defined as a mathematical model with a hybrid structure capable of handling both numerical and qualitative information.
Inputs to a model are processed mathematically to obtain a predicted output, expressed in either qualitative or quantitative format. Fuzzification and defuzzification algorithms convert the input/output information of the model as required for processing by the model or for presentation to the user.
A variable is qualitatively represented by a set of fuzzy terms that describe both its sign and dynamic trend. A degree of belief (DoB), attached to each term, is used to indicate the uncertainty or appropriateness of using a particular word to describe the variable.
Essentially, a QM is an approximate representation of the behavior of a process; it captures only relevant aspects of that behavior. A major advantage of qualitative modeling is that a model can be constructed with information obtained directly from heuristic knowledge of the process. If the only information available about the behavior of a process indicates that "the output is proportional to the input", we could write a QM as
QM1: output = K . input (%)
with K = 1 as a default value.
QM1, although mathematically inaccurate, can be successfully employed in approximate reasoning, i.e., when we are interested only in a rough estimation of input/output behavior. Also, if needed, QM1 can be tuned rapidly when new information about the process behavior becomes available
With this pseudo-qualitative approach, a QM could also be as accurate as a mathematical model (if such information is available). In fact, an existing mathematical model of a process can be represented and integrated with the QM and other elements of the system to increase the utility of such models.
The ISS runs on a PC interfaced to the Bailey DCS currently in operation at HVC (See Figure 2). The ISS accesses the Bailey database to obtain on-line information about the operation of Cline. The data are used by the ISS reasoning system to detect and monitor restrictions that affect the operation of the process.
Figure 2 - Implementation structure of the ISS.
The ISS has been implemented using the software package ProcessVision from Comdale Technologies, Toronto, Ontario. ProcessVision is a powerful tool for real-time applications ranging from data acquisition to supervisory control. Version 5.3 is a multi-tasking development toolkit in which all sub-tasks are assigned to separate modules; data transfer, message transfer, alarming, scheduling events, trend analysis, knowledge processing, etc.
A file with required process data is created on the HVC-PC and transferred from this computer to the UBC-PC via an RS232 serial link. A communication routine on the UBC-PC updates the ProcessVision database with the information received. Direct access to the Bailey network will eventually be setup once the HVCPC has upgraded its operating system.
ProcessVision is a modular knowledgebased software package for real-time applications ranging from data acquisition to supervisory control [3]. Running on a standard PC under the QNX 4.2 operating system from Quantum Software of Kanata, Ontario. PV5.3 provides facilities to interface with existing control systems, field I/O, and PLCs.
Information in ProcessVision is represented by keyword triplets of the form:
object.attribute.value :DoB
or
{class}.attribute.value :DoB
where DoB is the degree of belief (assigned by the user or estimated by the system) in the knowledge represented by the keyword triplet.
In this format, a developer can conceptualize knowledge in an objectoriented manner. This approach provides a structured manipulation and organization of objects by type or "classes" of objects. A piece of equipment or a process variable can be considered as an object represented by the object token of a keyword triplet. Specific characteristics of an object, such as the area or height of a tank, are represented by the attribute token; quantification of these attributes is executed by the value token of the keyword triplet. An example of a keyword triplet is
tank.area.small DoB: 85
This triplet indicates that "the area of the tank is small", and that the degree of belief associated with this piece of information is 85%.
Comdale/C, a real-time expert system, is the module that enables ProcessVision to handle symbolic information. This module deals with heuristics about the behavior of a process as well as information captured from the expertise of human operators. Comdale/C can reason with this knowledge to make decisions on the best actions to be taken or to advise the operator on the qualitative state of the process.
Other modules that comprise ProcessVision include a historical database, a network administrator, an alarm administrator, an explanation facility, and a graphical user interface. The modularity of a ProcessVision application together with the features of each individual module allow for a system to be designed to mimic the heterogeneous functions used by a human to reason about a complex process.
Some examples of the knowledge represented by keyword triplets in this application are as follows:
A complete classification of all the major tonnage restrictions that affect the operation of C-line grinding, and possible causes of these restrictions were obtained during discussions with HVC personnel. This information along with the heuristic procedures involved in the diagnosis of tonnage restrictions were incorporated into the ISS. This heuristic knowledge was implemented using IFTHEN rules and "procedures" (a structured knowledge representation element offered by ProcessVision).
The ISS updates its knowledge base with information read from the process database at regular time intervals. The reasoning system uses this information together with heuristic rules to detect the occurrence of tonnage restrictions, determine potential causes and estimate tonnage losses.
The control system automatically compensates for losses resulting from brief restrictions: an initial drop in tonnage is compensated by a sharp increase in tonnage when the restriction is removed. Hence, loss in tonnage is calculated only when the restriction affects operation for a prolonged time period.
Estimation of tonnage loss is based on measured and predicted values of the SAG mill fresh feed. The predicted value is assumed to be a level processable by the mill provided that the restriction had not occurred. For some restrictions, the predicted value is obtained from an analysis of tonnage trends before the occurrence of the restriction. For other restrictions, the predicted value derives from appropriate qualitative models of the process.
The ISS provides on-line information associated with the supervision of the process and findings from its restriction diagnosis. This output includes type of restriction detected, classification of restriction cause, estimate of tonnage lost, and accumulated losses per shift. The ISS also provides mechanisms to allow the user to confirm or correct on-line some of the findings and results obtained by the ISS. Should the system begin to diagnose a fault that has not really occurred, the user can override the system and prevent an erroneous prediction.
The ISS generates reports at the end of each shift with a summary of supervision results. A file with a record of all relevant events is also generated by the ISS. This file is accessed by the plant metallurgist to assist in daily or weekly analysis using spreadsheet data-processing. Eventually, the system will be interfaced with the spreadsheet program to generate the analysis report directly.
The ISS has been successfully installed and is currently operating at HVC. The diagnostic results obtained from the ISS were stored on files to be analyzed at the end of an evaluation period of four weeks. Operators did not have access to the system nor to the results during this period.
The performance of the ISS is very encouraging. The ISS appropriately detected the occurrence of each of the restrictions covered by its current set of heuristics. Table 1 shows the results reported by the ISS during the evaluation period, as well as the results reported by a metallurgist obtained by following the heuristic approach proposed by Hans Raabe, our Expert. From this table we can see that losses estimated by the ISS are within the range of those reported by a metallurgist. It is anticipated that with future system tuning the ISS system will match the performance of the metallurgical staff on average providing consistent analysis of future tonnage restrictions.
The difference between the results shown in Table 1, is due to two factors. First, the sensitivity of the diagnosis analysis. As mentioned earlier, the ISS performs a more rigorous diagnosis; it reports even those restrictions that only cause minor losses. Metallurgists, on the other hand, tend to consider losses caused by major restrictions only. This accounts for the difference in both number and duration of restrictions shown in Table 1. Second, the set of heuristics incorporated into the ISS prototype is incomplete: it does not cover all the possible circumstances in which a restriction may occur. One such case occurs when C-line is fed with material transferred from other grinding circuits and the control system cuts fresh feed in order to maintain mill power. The ISS detects this tonnage drop but due to the current lack of information and appropriate heuristic knowledge it does not report this occurrence as a tonnage restriction.
On a restriction-by-restriction basis, the diagnosis results provided by the ISS were very close to those reported by metallurgists. Over a period of time, however, these results may differ from each other due to the factors described above. As the evaluation period is extended, some factors may cancel out each other so the results are expected to be acceptable.
These two difficencies in the diagnotic procedure will be addressed in updated versions of the ISS. The sensitivity of the ISS, for example, can be adjusted to meet the requirements of any specific application. The ISS prototype evaluated at HVC was operating at its highest point of sensitivity, i.e. it would detect and report all restrictions that caused tonnage losses. The knowledge base of the ISS needs to be updated with new heuristics, some of which has already been identified as a result of the evaluation process The updating of the ISS is considered to be a natural procedure during the lifetime of such systems.
Based upon the results obtained from both simulation studies and the HVC application, we conclude the following:
The application of the ISS to the diagnosis of tonnage restrictions in a grinding line at HVC demonstrates the validity of the proposed pseudo-qualitative modeling approach. Feasibility of incorporating this technique into a commercial realtime SCADA system widely used in industrial applications has been proven.
An ISS, with the structure and elements proposed in this research, can be applied effectively to a realtime full-scale industrial operation. In the HVC application, the ISS dealt with aspects that are difficult to tackle with tools provided by current knowledge-based systems. The ISS could mimic human expertise and heuristic knowledge to handle all the subjective aspects involved in this application.
In addition, this application has highlighted a major advantage of such systems to capture human expertise and deal with problems that involve qualitative information and highly subjective heuristic procedures at a speed unattainable by a human.
The authors wish to thank Colin Lindsay of HVC for permission to conduct this research in the mill. Software and technical help from Comdale Technologies is greatly appreciated. Financial support received from MITEC (Mining Industry Technology Council of Canada) is also acknowledged.
1. Hans Raabe. "Mill tonnage restrictions". Internal communications. HVC, Logan Lake, BC, Canada, 1994.
2. E. I. Cifuentes. Incorporating AI elements into the design of an intelligent supervisory control system. Ph.D. thesis. The University of British Columbia, Vancouver, BC, Canada, 1995.
3. ________ ProcessVision V 5.2: User's Manual and Reference Guide. Comdale Technologies, Toronto, ON, Canada, 1993
Return to the publications index