Terrell Touchstone, Derek E. Blackwell, Grady E. Carter, Jolene K. Kramer
Chevron Corp.
Richmond, Calif.
A prototype, on-line, expert system has been developed at Chevron Corp.'s Richmond, Calif., refinery to advise and train process operators.
The system models the experience and decision making capability of the operator, rather than automating the decision process by modeling process equipment and operations.
The system, called OPAS for Operator Advisor System, coaches operators to make correct decisions so that consistent shift-to-shift control strategy can be maintained, thus minimizing variability in production, product qualities, and operating costs. It can also be used to train inexperienced operators.
The system can operate on-line, meaning it has access to live plant data through a serial link to the plant's monitoring computer.
The system monitors key, high-level information about the plant, and using the time-tested knowledge of experienced operators, it makes recommendations. It can also back up its recommendations with sound reasoning.
Key features of the system are its event-driven user interface and its development and curator tools. The first gives an operator a highly interactive and responsive tool for decision making. The second means the system can be both built and maintained with ease.
The system's architecture also allows easy access to other personal computer applications that are commonly used in plant control rooms, such as process and equipment troubleshooting, statistical process control, and process engineering calculations. Although currently in the prototype stage, a recent installation of a working prototype on the gas separation section of the refinery's lube plant is beginning to demonstrate the ability of the system.
PROCESS EXPERTS
Process operations are prime candidates to benefit from expert system technology.
Operating a process plant or unit is a skilled job requiring a good deal of experience.
Gaining experience takes time. And because length of service varies within operator crews, measurable differences in performance are typical among operators, and between operator crews. Experience is also very personal and invites stylistic expression-so much so that shift-to-shift changes in control strategy, even philosophy, are common.
Such differences lead to inconsistent operation, which, in turn, increases variability in production, qualities, and costs.
Consequently, the experience factor can be a major cause of variance in plant performance.
This is where expert systems can be advantageous to help capture and pass on the experience of veteran operators in a consistent manner.
Refinery and chemical processes typically exhibit a surprisingly complex problem space.
The inexperienced operator is routinely challenged to handle difficult problems in the face of rapidly changing circumstances. The job requires a continual cycle of assessment, response planning, decision, and action.
Over time, the experienced operator builds up two bodies of useful knowledge: a tree of all possible situations (the problem space) the process might assume, and a repertoire of control moves for correcting adverse situations.
The expert system can map out the problem space and supply it and a list of proven control strategies to inexperienced operators. It acts as a playbook similar to that used by football teams.
A team's playbook (the expert system) is both a learning tool and communication tool. Because plant operation is a team effort, much depends on everyone knowing what everyone else is thinking.
The overall idea is to capture and automate the factual and rule-based knowledge of expert operators and process engineers, and then to make that information available to less experienced operators at the precise time that it is needed.
Once operating strategies are crafted and recorded, they become indispensable for training.
It soon becomes second nature to think of calling specific plays in response to operating situations. Furthermore, plays can help to fine tune operating procedures because they will inevitably serve as discussion points in Monday morning quarterbacking sessions.
DEVELOPMENT ROUTE
At Chevron, the decision-making portion of an operator's job is described by an idealized decision pyramid (Fig. 1). The pyramid is composed of data and its associated forms: information and knowledge.
The flow of data to the top is similar to a biological food chain.
At the base are vast quantities of data. The data feed information, information feeds knowledge, and knowledge feeds a few decisions.
We define decision as a commitment to a plan of action (a sequence of plant moves). The total collection of data, information, and knowledge is defined as the knowledge base (KBase at Chevron).
Table 1 illustrates the various data forms, using an example situation drawn from a fluid catalytic cracking unit (FCCU).
Table 2 depicts the value-adding processes that transform one data form to the next. This idealized model defines the operator's job as the process of making knowledge-based decisions: application of understanding and wisdom to plant knowledge derived from the upward flow of data.
The ideal job depicted intentionally excludes the transformation of information into knowledge. Unfortunately, operators have, for a long time, been given limited knowledge, so that decisions have had to be made on information alone.
Today, with less-experienced operators on one hand, and information overload on the other, there is a wide knowledge gap. Because expert systems bear directly on the knowledge gap, Chevron, and others, are actively investigating this promising technology.
Out of this, we see two emerging routes to the development of expert systems: Real-time expert systems that close the knowledge gap from the bottom up, and online coaching that closes the knowledge gap from the top down.
REAL-TIME SYSTEMS
One way of coping with information overload is to automate routine tasks. This allows the operator to spend more time on important matters.
Accordingly, this approach has concentrated on automating plant sentry chores, and has mostly been concerned with fault detection, diagnosis, and alarming.
While this route has merit, its implementation has been difficult. Although the sentry tasks are simple, the sheer number of them, and the burden of maintaining the truth of their real-time aspects, has proven otherwise.
The mechanics of this approach require modeling the fine details of the plant's processes together with complex transition rules describing the physics and chemistry behind changing events. All of this necessitates using a highend, real-time expert system, and a cadre of talented people.
The result is the need for high-powered computers and manpower that lead to large, time-consuming, and expensive projects.
ON-LINE COACHING
Alternatively, the development route can narrow the knowledge gap from the top down. This is called on-line coaching at Chevron.
On-line coaching solves a different problem-process improvement by increasing operator knowledge, rather than automating plant sentry operations. It is a separate, noncompeting alternative to the real-time development route.
Basically, the approach taken is to start with top-level decisions that need to be made, and then gather and organize the data and information to create the necessary knowledge to support those decisions. Operators tend to work that way.
On-line coaching models the expert operator, and works to improve the decision making capability of the operator. In contrast, realtime expert systems model the process equipment and work to improve the decision-making capability of the process equipment.
Because it is easier to model how an expert operator works, the expert system for on-line coaching is simpler.
SYSTEM DETAILS
The expert system, OPAS, is an on-line expert system that provides operators with situation-appropriate advice for improving process performance. It runs on a personal computer that has access to real-time data through a serial data link to the plant's monitoring computer (the computer that records all sensor data, such as temperature, pressure, level, etc.).
Because it is on-line, the system can exploit on-the-job opportunities to teach inexperienced operators the skills routinely performed by experienced operators.
The system stores the know-how of the best operators for all known plant situations, from which it extracts advice appropriate to the current situation.
The system is designed so that the stored knowledge can be frequently changed or updated.
The best operator can mean just that, or it can be composite knowledge drawn from the experience of several persons, such veteran operators, experienced process engineers, and refinery planners.
In addition to offering advice on a real-time basis, the system can be used as a simulator for off-line training. Because the same system is used (except run in a simulation mode), the system retains the look and feel of being on-line.
Under training conditions, the trainee or instructor is free to experiment with dangerous or rarely seen situations, with the trainee's responses to the situations compared to the advice given by the system.
Knowledge is stored in a natural, organized way in the computer (KBase). Using an object-oriented programming language, the KBase building blocks (chunks of data, information, and knowledge) are networked to more closely mimic the way humans think.
This greatly facilitates the building, use, and upkeep of the knowledge base. For instance, advice is entered by adding operating experience, as cases, to a case-history file using fill-in-the-blanks templates.
The knowledge network is readily accessible to both the builder and the user by the use of windows and links. Nodes in the knowledge base correspond to windows on the screen on a one-to-one basis.
Nodes are examined through a windowing system that supports multiple windows and screen operations like opening, positioning, sizing, and closing of windows. The windows also have icons that link to other nodes in the knowledge base.
Links, which appear as buttons on the screen (activated by a mouse), enable the user to immediately jump to other nodes in the knowledge base or to run other conventional computer applications, such as simulations, calculations, or reports.
The system also allows browsing, not readily available today in conventional process plant work stations. Browsing allows random access to the knowledge base in an unconstrained manner, and collection of items for further study.
Browsing is described as hierarchical zooming. It gives a user quick access to information, and the ability to pick and choose items of immediate interest.
Once the items are collected, screen utilities allow the user to freely associate them as concepts. Information can be hidden, arranged, and sized, similar to using a blackboard as a visual aid in the thinking process.
The system allows easy editing of the knowledge base. Operators and others can make changes to reflect new discoveries in the operation of a process.
For instance, if upon using the system's advice, an operator discovers an adverse side effect, that side effect can be added to the knowledge base. That new information will then be included with that same advice on subsequent consultations.
HOW IT WORKS
The expert system provides the essential functions for computer-assisted decision making: data acquisition, analysis, and display. The system works by first taking a snapshot of the data it needs to infer the current performance state of the plant.
Each of the data are converted to a one-word qualitative description. Descriptions from related data are combined to form a word pattern.
The word pattern represents the current performance situation. The system then searches the KBase to find advice matching the situation.
The advice is displayed on the operator's screen in a pop-up window. Decision and action are then up to the operator. However, on request, the system will explain the reasoning behind the advice given.
The system doesn't process plant data in real time. That job is left to the plant's monitoring computer which has the computer resources to handle such a taxing chore.
What the system does is to take the information provided by the real-time computer and to begin symbolic processing. Using a data link to the plant's computer, the system gets statistics on performance variables to begin its analysis.
Performance variables are carefully selected indicators that tell how well the plant is running. Typically, they tie directly to production, quality, or operating costs.
Analysis is a two-step process. First, a word pattern representing the plant's current performance situation is formed. Then a search is conducted to find advice matching the situation.
The word pattern is formed using a value-adding process (fuzzy logic) that converts numbers from the real-time computer into a word list. The numerical value of a variable is classified according to the portion of the continuous range that value has at any particular sampling.
Typically, the verbal classifications are "High," "Ok," or "Low." Related performance variables are next grouped together so that if each variable in the group has been classified, a word pattern will result.
The word pattern is called a situation. For example, consider the four variables related to burning carbon off catalyst in an FCC regenerator.
The performance variables and the word descriptions they might assume are shown in Table 3. Taking percent oxygen, percent car bon dioxide, carbon on regenerated catalyst (CORC), and the regenerator temperature in order, one of the possible situations is: "Lo, Ok, Ok, Hi." There are, in all, 60 different word descriptions for this example.
Once the situation has been determined, the knowledge base is searched to find advice matching it. Advice needs to include a synopsis of the situation, the specific instructions for the situation, and reasons to support the advice.
It is worth noting that the knowledge base does not have to be complete. Without a complete set of recommendations, the system can still function, but if it sees a situation that hasn't been covered, it just says so.
At such a time, the user can take the opportunity to add advice for the new situation. This allows the knowledge base to be continually expanded with new knowledge by every user.
Table 4 shows how the situation and advice for the FCC example are displayed on the screen. The situationdefining information is given in table form with columns for the numerical value of the measurement, its qualitative value, the limiting condition, and the trend.
This matrix format seems to be the way most of us choose to describe plant situations, and its compact form promotes pattern recognition.
The trend column is an attempt to restore feel-of-the-plant information that disappeared when control room readout devices were changed from analog to digital.
The trend is a word description of the direction and rate at which the variable's numerical value is changing.
Because it is believed that advice is best given only when asked for, the system's advice is displayed only if the operator requests it. Table 5 shows the kind of advice the expert system provides. Note the conversational style of the information.
Also note that the computer limits the information to advice. The actual decision is left up to the operator, because, as suggested by Table 2, operators, not data, run the plant.
PROCESS UNIT SYSTEM
Currently, installation of the expert system on the gas recovery section of the lube oil plant at the Richmond refinery is nearing completion. The gas recovery section is a train of distillation columns that separate light refinery gases.
The gas recovery section was chosen because the process is well understood, and operating expertise is readily available. Although the expert system has not been formally commissioned, a working system has been built and checked out.
Because much of the basis for the design of OPAS came from the architecture of the Mackintosh personal computer and HyperCard, an authoring tool and information organizer for the beginning computer user, these were the obvious choices for the prototype development. (Mackintosh and HyperCard are both products of Apple Computer Corp.)
The expert system runs on the Mackintosh personal computer that has access to realtime data through a serial link to the lube oil plant's monitoring computer. The plant monitoring computer does all of the calculations and data analysis, while the personal computer provides the man/machine interfacing and symbolic processing.
The OPAS system is implemented as a HyperCard stack where the inference engine has been programmed using HyperTalk (another Apple product).
The prototype's main objective is to teach the fundamentals of gas recovery, not necessarily to handle every imaginable exception. Once trained, the operator is expected to know how to handle these common situations without assistance. Knowing why helps the operator adjust basic advice to handle exceptions.
The HyperCard version of OPAS has some limitations. HyperCard is similar to Basic in many ways.
It is an interpreted language, so it cannot deliver user friendliness without sacrificing performance. Even though portions of the code have been compiled, the system is too slow to support background processing or a periodic examination of the entire plant.
It can only attend to the portion of the plant the operator has selected to view on the screen and, therefore, runs only on demand. Even with those limitations, HyperCard has proven to be an excellent environment in which to shape the OPAS concept.
POTENTIAL BENEFITS
One of the major benefits of installing an expert system, such as OPAS, is that it provides management of process expertise, an important company resource. The installation of a system forces builders of the system to collect and record detailed information about the plant.
Building the system forces operators and engineers to think systematically about how they run the plant. This fosters a shared understanding about how the plant is, or should be, run.
For instance, engineers have the opportunity to frame the problem by defining performance measures, and by categorizing operational situations by giving symbolic meaning to important plant circumstances. Within the framework, it becomes clear to veteran operators how their experience fits into the entire plant picture.
The cooperative, team-building results that come from this shared information process are very similar to results achieved by quality-improvement teams.
Once all of the plant knowledge is acquired, it is more useful to have it in electronic form than in the form of hardbound instruction manuals. Electronic information is, many times, easier to access, and easier to frequently update.
Rather than collecting dust, as is often the case with printed manuals, electronic systems foster the addition of new knowledge.
So, once the expertise is electronically on hand, it is easy to teach, apply, and improve it.
ACKNOWLEDGMENT
The authors thank Ashok S. Krishna of Chevron Resrearch Co. and Steve J. Stadnicki of Chevron's engineering technology department for providing the FCC example used to describe the OPAS system.
Copyright 1990 Oil & Gas Journal. All Rights Reserved.
