EMPLOYING INTELLIGENT ADVISORY SYSTEMS AT ALCOA OF AUSTRALIA LIMITED
Alcoa of Australia Limited, Kwinana Refinery,
P.O. Box 161,Kwinana
W.A., Australia, 6167
Alcoa of Australia Limited, Wagerup Refinery,
P.O. Box 84, Waroona
W.A., Australia, 6167
85 The Esplanade, South Perth
W.A., Australia, 6151
In this paper, the measures taken to enhance the performance of the filtration operation by allowing the operator to make more and better informed decisions through the implementation of an Intelligent Advisory System in the Alcoa of Australia Ltd Wagerup Alumina refinery are discussed. The G2 expert system (Gensym Corp.) has been used to implement a knowledge-based operator guidance system that not only relies on past experience via sets of rules, but also enables several possible operating scenarios to be evaluated using a mathematical cost function. It then provides operators with the best solution to a particular process problem.
EMPLOYING INTELLIGENT ADVISORY SYSTEMS AT
ALCOA OF AUSTRALIA LIMITED
Ali Nooraii, Jim Mather and Murray Wood
In recent years with the evolution of computer technology, powerful computational platforms have provided Control Engineers with efficient tools, enabling them to apply more effective and advanced techniques. Meanwhile the evolution of manufacturing systems has moved towards information and software based manufacturing. There are many factors contributing to this evolution including the need for continuous improvement in a competitive market environment, the requirements stemming from environmental regulation, the need for process safety management and the international drive for improved product quality.
An Intelligent Advisory System is classified as one of the new tools that can play an important role in management of process operations. An Intelligent Advisory System or Operator Guidance System acts as a process supervisor and is used to enhance the performance of the plant by allowing the operators to make more frequent and better informed decisions. This intelligent decision making process may include tasks such as data acquisition, regulatory control, performance monitoring, fault diagnosis, supervisory control, scheduling and planning. These tasks should be included in an integrated framework that allows the operational decision-making to be more comprehensive and effective (Cochran T. and D. Rowan, Supervision and Decision Support of Process Operations).
Alcoa operates three alumina refineries in Western Australia. These refineries, in common with most other alumina refineries in the world, use the Bayer process to produce alumina from bauxite. Bauxite, the major raw material used in the process, is a mineral composed of hydrated aluminas in combination with other non-valuable components that can be present in a variety of physical and chemical forms. The product, which is alumina (aluminium oxide), must be of high purity for smelting to aluminium. This is due to the nature of the electrolytic smelting process, which reduces most alumina impurities simultaneously, leaving them alloyed in the product aluminium. Other impurities also remain in the smelter bath and adversely affect its operation. The basic principle of the Bayer alumina refinery process is that alumina hydrates are dissolved in a caustic soda solution at elevated temperature and then recrystallised from the same solution at a lower temperature. The other bauxite components are separated from the liquor prior to recrystallisation.
Figure 1 represents a typical Bayer process flowsheet. One of the unit operations before recrystallisation is the process of filtering out bauxite residue solids from the process liquor stream. Obviously the quality of the stream out of the filtration operation will directly affect product quality and act as a major constraint in the process operation. Being a batch process in the middle of a continuous operation, filters can also impose a significant constraint in terms of process flow, with adverse effect on plant efficiency and production rate. Due to these important effects on production rate and quality, it is well known in Alcoa of Australia that the filter building can have a great impact on overall process operation. Studies during the last few years have shown that employing an Intelligent Advisory System, together with modified filter control strategies, can have a significant impact on maximising plant performance.
Typical flowsheet of the Bayer Process
In the filtration operation at Wagerup, process liquor containing bauxite residues is fed into two surge tanks. These tanks are then used to feed two parallel units composed of several vertical and horizontal filter presses. The operational objective of the building is to meet the flow requirements of the plant while maintaining the quality of the filtrate within an acceptable limit. Although the objective looks simple, it can be extremely difficult to achieve, considering all the factors affecting the process:
As can be seen, there are a range of factors which affect the buildings operation. Effective scheduling of filter presses can have a significant impact on the performance of the building in general and can assist in keeping the level of the feed tanks in control, therefore employing an Intelligent Advisory system is crucial to:
In this paper, the implementation of such an Intelligent Support System in Alcoas Wagerup Alumina Refinery is discussed. The measures taken to enhance the performance of the filtration operation by allowing the operators to make more frequent and better-informed decisions are addressed.
2.0 EXPERT SYSTEM AND THE MODULE HIERARCHY
We can define an expert system to be a computer program that uses knowledge and procedures to solve complex and difficult problems at the level of professionally trained human(Klein R.M. and L.B. Methlie, Knowledge-based Decision Support Systems). When knowledge and inference procedures are modelled after human experts, we call such knowledgebased system an expert system. Based on our requirements, it was found that G2 expert system provides an efficient environment for developing such a complex intelligent system.
It is possible to characterise the components of an expert system as belonging to the knowledge-base subsystem. Most workers in expert systems share the belief that good design practice makes these subsystems as independent of another as possible. However complete independence is neither possible nor desirable. It is necessary to know the structural form of the knowledge base to properly make use of the knowledge base by the cognitive (or inference) engine. However, what is not desired, is that the operation of the cognitive engine subsystem be dependent upon the specific domain-dependent knowledge that is contained in the knowledge base. This independence allows efficient and effective updating of the expert system as new knowledge is acquired.
G2 knowledge base module hierarchy
The module hierarchy of the knowledge base developed for filter building application at Wagerup refinery is shown in Figure 2. The functionality of the expert system is divided into two main modules:
The Diagnosis Module is responsible for providing insight into properties in the building that are difficult to measure and also advice on the status of the equipment. The Scheduler Module is responsible for providing conclusions on status of the building and the changes required in the press arrangements to reach to specified objectives. These two modules are basically flowsheet independent and can perform their objectives with any flowsheet.
Other modules in the knowledge base are:
1. Building Views: this module provides all the information on the dynamic changes of the presses as well as the building flowsheet configuration for the specific flowsheet in the filter building at Wagerup refinery.
2. Data Validation: this module is used to validate the measurements and avoid using measurements with gross errors.
3. Bridge: this module provides means of communication between the Honeywell control system and the G2 platform
4. Watch Dog Timer: watch dog timer was designed to monitor the integrity of G2/Honeywell communication.
In the next few sections two main modules will be discussed further and results from employing such an Advisory system are presented.
2.1 Diagnostics Module
In this module the following main calculations are performed:
Cloth resistance for each press.
Mud resistance for each press.
Average mud and cloth resistances for each unit.
Feed lines flow resistance.
The theoretical equation for the behaviour of a filter press can be written as follows(Cheremisinoff N. P. and D. S. Azbel, Liquid Filtration,):
Since mud properties are changing with time, so equation 1 can be rewritten as:
The variable Z is readily calculated from the ratio of pressure and instantaneous flow. As it can be seen from equation 2, at t=0, R2=Z. This is a simple way to determine R2 from which cloth resistance is estimated. Because of extensive amount of noise in the measurement in practice, it was found more satisfactory to take the average values over first six minutes. For the case of R1, from which mud resistance is estimated, the calculation is not simple. R1 is essentially the instantaneous slope of the plot of Z vs. liquor volume passes through press. In the presence of noisy data, the noise on the derivative is two orders of magnitude greater than the value of the derivative. This demanded a more elegant approach in determining the true value of cake or mud resistance. A model of the process was built using equation 2 and the parameter R1 is then estimated adaptively to minimise the error between the model derived value of Z, and the actual value of Z. This was done using Kalman Filter(Kailath T., Linear Systems):
Where K is the Kalman gain and a is forgetting factor. The Kalman gain depends upon the statistics of the input signals and output signals to the real plant. As the gain changes, it is possible to get better estimation of the state, but poorer noise rejection. Determining the optimum gain required the solution of a steady state Riccatti equation, the result, being used in the module.
Figure 3 shows the convergence of the state estimator for one of the presses in G2. As it can be seen estimation
Comparison of Z and of one of presses
from the model, (blue), closely follows Z (red) which is calculated from the measurements (q and P). Figure 4 shows the average of mud resistance for unit 1 and unit 2. These calculations assist in targeting problems in the different units, and also in looking at caustic wash scheduling and filter aid control based on cloth and cake resistances.
Average mud resistance of unit 1 and unit
Another function of the module is to estimate scale in the feed lines based on pressure measurements, which will ultimately provide information for scheduling the line cleaning.
2.2 Scheduler Module
To elicit real problem-solving knowledge from human experts (engineers and operators), a performance modelling strategy was used in developing the Scheduler. The challenging part of knowledge modelling is the process of transferring and transforming knowledge about the problem from the knowledge source to a computer program. This process of knowledge acquisition or knowledge processing is defined in terms of the following stages(Klein R. M. and L. B. Methlie, Knowledge-based Decision Support Systems):
The ultimate goal in the Wagerup filter building is to filter the green liquor being presented to the building while maintaining the maximum surge capacity in the feed tanks. In mathematical form we can write a flow objective as:
The Flow Objective = Flow to the tanks + (Current level Target level)´ (Area of the Tank)/(the time in which the target level is achieved)
In reaching this objective, several constraints should be met. In order to integrate all the constraints into one ultimate figure, a performance index is introduced which makes the comparison between different scenarios possible. Consequently, the final outcome from the scheduler is a set of performance indices for all the possible scenarios. In cases where hard constraints do not allow a scenario to become feasible, the performance index is still calculated but will be indicated as infeasible. In cases where the infeasible scenario is the most attractive one, the operator may be able to remove the constraint by proper action.
In the scheduler, the four worst performing presses in terms of flow are determined and their remaining on-line times are estimated. This is based on the cloth and cake resistance calculated in the diagnostics module. If the diagnosis module is not available then this section is bypassed and a performance index is calculated for four scenarios (The number of scenarios is limited by physical requirements involved in changing the press configuration):
- Leave the current configuration as it is.
- Take one press off (the program will advise which press to be taken off).
- Bring one press on (the program will advise which press to be brought on).
- Take one-press-off and bring one-press-on (the program will advice which presses).
For scenarios 3 and 4, in cases where the press can be fed from either unit, the program will also advise the unit from which the press must fed. Figure 5 is a screen in G2 that shows the recommendations. It must be noted that only developers have access to displays in G2. All the necessary results are sent to the Honeywell LCN control system via SCAN3000 and the operators monitor the advice from the program directly on their LCN Universal station.
Performance Index coming from advisory system
Figure 6 Performance Index coming from advisory system
Average Feed Tank level
Figures 6 and 7 are showing recommendations from the advisory system and the average tank levels in a period of 24 hours, respectively. From 01:00 till 10:00 the program is recommending no changes as it can be seen that the tank levels are around 70% vs. target level of 65%. Around 00:30, the levels start dropping and the program is advising the operator to take one press off. Following the implementation of this recommendation, the levels start to increase again. This situation is repeated several times during which the level has been maintained around 65%. Tuning parameters in the G2 knowledge base will allow us to decide how close we want to keep the level to the target. Obviously optimum tuning is achieved when the tank levels are kept within a range to avoid either a flow cut or the need for more frequent press changes than necessary.
An Intelligent Advisory System is used to seek judgmental and problem-solving advice and has opened the doors towards better decision making for our plant operators. The preliminary study showed potential benefits of the order of several hundreds of thousands dollars per year. Now that the program is well into commissioning stages, the results are very encouraging and we expect to be able to get full benefit of the program. We have clearly noticed potential improvements in the efficiency of the filter press building. Interestingly we have also noticed that the user can be a tutor to the system as soon as they get more familiar with the system and its potential, such that the expert system increases or improves its knowledge and its ability to perform. Consequently, the process of knowledge acquisition cannot be totally separated from that of knowledge representation and utilisation. In order to determine consistency and completeness of the knowledge base, it is necessary for acquired knowledge to be subjected to appropriate tests. In our on-going experience with developing the expert system, we are continuously facing new and unforeseen aspects of the process as we trying to improve the knowledge base. Some influences that were considered to be minor now are showing to be important in certain circumstances and must be taken into consideration. Thus, identification of existing deficiencies in the knowledge base (and our knowledge of process) drives the process of knowledge acquisition. In so doing, the process of knowledge acquisition should become more efficient and effective.
A = Total press area
dp = Pressure drop across the press
q = Instantaneous flow
r 0 = Cake resistance
r f = Cloth resistance
V = Bayer liquor volume passed through press
x0 = Mud load
m = Viscosity
Cochran T. and D. Rowan, Supervision and Decision Support of Process Operations, Proceedings of the First International Conference on Intelligent Systems in Process Engineering, ISPE-95, Snowmass, Colorado (1995).
Klein R.M. and L.B. Methlie, Knowledge-based Decision Support Systems, John Wiley and Sons, (1995).
Cheremisinoff N.P. and D.S. Azbel, Liquid Filtration, Ann Arbor Science, (1983).
Kailath T., Linear Systems, Prentice-Hall, (1980).