Online, Artificial Intelligence-Based Turbine Generator Diagnostics

Transcription

1 AI Magazine Voume 7 Number 4 (1986) ( AAAI) Robert L. Osborne, Ph. D Onine, Artificia Inteigence-Based Turbine Generator Diagnostics introduction The need for onine diagnostics in the eectric powergeneration industry is driven by a number of significant factors. Due to the ow number of new power pants being buit by eectric utiities, the average age of existing power pant equipment in the United States and its susceptibiity to faiure is increasing rapidy. Figure 1 shows the percentage of power-generation equipment over 20 years od as a function of year. Note the rapid increase of average age after 1980 and the fact that by the year 2000 fuy 50 percent of a generation equipment in the United States wi be over 20, the odest average age of power pant equipment ever experienced by U.S. utiities. Thus, there is a need to know what the actua operating condition of the equipment is at a times, so that outages can be avoided by taking corrective actions at the eariest possibe time and by prepanning for outages if they become necessary in order to to minimize their ength. In order to provide increased information on the actua operating condition of the equipment, the utiity industry has instaed additiona monitoring capabiity utiizing power pant computer systems that measure system variabes and present these variabes effectivey. Using data highways, these power pant computers aow the operator to dispay monitored variabes on coor CRTs in a variety of ways. Dispays incude showing a variabes above an aarm eve, sensors that are out of service, variabes superimposed on diagrams of the equipment, and variabes potted as a function of time. The vaue of variabes can be printed periodicay or when required. Unfortunatey, regardess of how sophisticated the pant computer system is, knowing the vaue of variabes such as temperature, pressure, and vibration eve, that are reported by the computer does not aways aow you to know what Dr Robert L. Osborne is manager of Diagnostics, Monitor and AI Deveopment at Westinghouse Eectric Corporation, Service Technoogy Division in Orando, Forida, is wrong with the equipment; that is, the system does not te the operator a bearing is faiing or a conductor is broken. Today, it takes a skied person to interpret the vaue of measured variabes to determine what is actuay wrong with the equipment. In order to aid the equipment operator in making better operating decisions, the need was recogniized in the mid- 1970s to pace in a computer the capabiities of diagnosticians who know the reationships between measured variabes and the condition of the equipment. At this time, an approach utiizing probabiities was pursued; a sma microprocessor-based system was buit and demonstrated to the utiity industry at a symposium in The resuts showed that there was utiity interest in this kind of product. It was aso recognized, however, that a commercia-size system using hundreds of variabes and identifying hundreds of conditions coud not be obtained through this approach. Pacing the knowedge in the computer was awkward and time consuming, and ony one mafunction at a time coud accuratey be identified. Thus, we were eft knowing what the product was that we wanted but not an acceptabe method of impementing that product. Abstract The deveopment of an onine turbine generator diagnostic system is described from conception to initia fied verification. The system is composed of a data center ocated in the power pant that coects data from onine measurement devices and communicates these data to a centraized diagnostic faciity in Orando, Forida, where the actua diagnosis is done. The resuting diagnosis and recommended actions are transmitted to the power pant where they are dispayed to the operator by the data center. The marketpace need, initia approaches to the product, system specification generation, rue base deveopment, and initia system fied verification are described. The artificia inteigence (AI) diagnostic program has been diagnosing seven arge utiity generators since Juy 1984 and has correcty diagnosed a significant number of generator and instrumentation probems. Issues such as a centraized approach, rue base quaity contro, and the range of resources needed for a successfu product are discussed. FALL

2 Expert System Approach A decision to pursue expert systems as the basic too in the deveopment of diagnostic systems was made in 1981; this decision was dictated by a number of factors. The deveopment effort is essened consideraby with an AI approach when compared with conventiona programming in a anguage such as Fortran. Athough expert knowedge can be written into conventiona software, the program input process requires not ony experts but aso programmers. Using AI software, nonprogrammer knowedge engineers can create expert systems through interactive input sessions with experts. Using AI software aso aows easier modification to an existing expert system. Conventiona faut tree anaysis does not indicate the vaidity of the answer, but an expert system, just ike a human diagnostician, can give the confidence associated with each of the diagnoses presented to the user. This confidence is extremey important because few power pant equipment diagnoses are 100 percent certain, and the user can be basing mutimiion doar decisions on the information presented by the system. An expert system can dispay the method it used to reach its concusion. This carifies the diagnosis and provides an exceent vehice for training new peope in the fied. Conventiona programming discourages this method of training. Thus, when appied in a cost-competitive environment, AI becomes very effective compared with standard programming approaches. The first step in the deveopment of the diagnostic system product was to initiate the deveopment of the basic AI too. The resut of this deveopment was the Process Diagnostic System (PDS). PDS is a forward-chaining, rue-based system in which sensors, hypotheses (see figure 2), mafunctions, rues, and turbine structure are represented as schemata in SRL (Fox 1979; Wright and Fox 1982). The initia impementation utiized the MYCIN (Shortiffe 1976) approach to the representation and propagation of certainty. The impementation was modified, though, because of the existence of erroneous sensor data due to sensor degradation or spurious readings. In particuar, the foowing modifications were made: The fuzzy minimum for conjunctive evidence was repaced with a weighted average that refected the degree to which the evidence shoud be considered usefu in the decision. These weights can dynamicay be atered according to sensor heath. This removes the probem of MYCIN s underestimation of beief identified by Wise (1986). A rue s certainty factor was extended to incude the specification of the necessity and sufficiency of the evidence. Earier sensor readings and hypotheses were archived to aow time-series anaysis as part of the reasoning. Logica sensors which are the composite of mutipe physica sensors, were used to reduce the impact of erroneous readings. 50. s w kj I 1 Figure I. Od. 11 pds node U.S. CAPACITY 20 YEARS AND OLDER 850 GIGA WATTS YEAR Percentage of Eectric Power Generation over 20 Years MB: "eve of beief in the node being true" MD: "eve of disbeief in the node being true" CF: "eve of certainty = mb - md" SUPPORTING-RULES: "rues for which this node is hypothesis" SUPPORTED-RULES: "rues for which this node is evidence" SIGNAL: "contains signa schema name(s)" DESCRIPTION: "Engish description of the node" HAS-IS-A: (or sensor hypothesis mafunction) }] Figure 2. Generic Node in PDS. Metarues were used to ater diagnostic rues, weights, and certainty factors when sensor degradation was identified. Once the forward-chaining deduction cyce is competed for the current sensor readings, the operator can enter into a mixed-initiative interaction in which the system can eicit information not avaiabe through sensors. PDS was deveoped using SRL and Franz Lisp. The production version is written in C. For a compete description of PDS, the reader is referred to Fox, Keinosky, and Lowenfed (1983). In the initia stage of product deveopment, one of the critica decisions was whether to pace the expert system in a computer ocated in each power pant or to utiize a centraized computer. The centraized approach was chosen for a number of reasons. First, the knowedge of the reationships between variabes that can be measured and the condition of the turbine generator resided primariy with design and service engineers who had worked with the equipment for a number of years. A centraized approach aowed a diagnos- 98 AI MAGAZINE

3 tic rue base to be written that coud improve over the years because of its growing knowedge of actua component probems and its abiity to recognize mafunctions that had not occurred to date. Because of the reative infrequency of equipment faiures on any one turbine generator, a basic question was how the knowedge base coud be improved in the future, utiizing both the designers knowedge and the users experience, without a centraized faciity. By utiizing a centra diagnostic center approach, any knowedge gained in one power pant to improve the rue base coud immediatey be made avaiabe to a users. This is in contrast to a decentraized system where every individua computer in each power pant woud have to be reprogrammed each time a rue base improvement was made. In addition, by manning the center 24 hours a day, if a new mafunction appeared in any unit that was not in the rue base, a human diagnostician woud have many resources immediatey avaiabe to aid the operator and to update the rue base if necessary. These resources incude databases that contain the outage histories of a units being diagnosed, detaied design information, and the equipment design engineers themseves. After deveoping the basic too, PDS, the next step was to convince management a commercia-sized system, that is, one with hundreds of variabe inputs and hundreds of equipment conditions diagnosed, coud be made a commercia reaity. It was decided that a steam chemistry diagnostic rue base woud be generated. Knowing the state of steam chemistry in a power pant is important in preventing damage to those parts of the power pant where the steam fows. Steam chemistry was chosen for the first diagnostic rue base primariy because more steam chemistry upsets occur per month than actua equipment probems; more upsets meant that this rue base woud be exercised consideraby more frequenty than one for a turbine or a generator, aowing much more evauation of the AI approach in a shorter period of time. Aso, much more steam chemistry fied data were stored that coud be used offine to evauate the diagnostic rue base. In eary 1982, a sma 25rue chemistry diagnostic rue base system was written using PDS. This system demonstrated that the AI approach was practica and coud be expanded to arge systems. Additiona work on the steam chemistry rue base (Beows 1984) resuted in a demonstration, primariy for management, of a much arger rue base. This abiity to do arger rue bases convinced division management to significanty expand the entire program, which ed to estabishing the diagnostic center at the power generation headquarters in Orando, Forida. Diagnostic Center The diagnostic center was designed to centraize our entire diagnostic program in one geographic ocation. Within the Figure 3. Diagnostic Center Information Area.. power generation headquarters buiding, the center was ocated next to the service engineering department because that department woud have the most interaction with the diagnostic center on a day-to-day basis. The diagnostic center has five distinct areas: the information area, the diagnostic conference room, the operations center, a customer aboratory, and two AI aboratories. Each of these areas has its own mission as part of the diagnostic product ine. Initiating a new technoogy in a 70-year-od division presented an educationa chaenge to the empoyees in the division as we as to our customers. The information area, shown in figure 3, is one of severa effective techniques used to inform peope about the diagnostic product and the AI technoogy it is based on. So as not to interrupt the engineers working on the product, an IBM-XT is programmed to aow anyone to wak up and request, through the use of a mouse, subjects such as AI, the diagnostic center, and expert systems. For each subject requested, severa paragraphs of expanation are presented to the onooker. When the person has finished reading these paragraphs they can return to the menu and choose another subject. Another portion of this information area contains an infrared detector that senses the presence of anyone in the area, triggering a four-minute recording keyed into a seectivey ighted diagram which expains the diagnostic center and how it is integrated into the service engineering department. Athough this method was effective, especiay for groups, the frequent repaying of the recording distracted the engineers in the area, so the recording is now ony used when groups are visiting the center. The diagnostic conference room, shown in figure 4, is an important part of the diagnostic center. This futuristicooking area provides a pace where personne can interact in probem-soving sessions; demonstrations of the products can be made; and direct communication between division FALL

4 Figure 4. Diagnostic Center Conference Room. Figure 5. Arti$cia Inteigence Laboratory. personne and other ocations, such as our research and deveopment center and the fied, can instanty take pace. The room itsef was designed around the concept of a paperess conference room. Ideay, the ony paper that needs to be brought into the room is for persona note taking. A other information is contained in databases or comes in ive from the fied and can be accessed by personne in the room. To faciitate ive viewing, there are two singe-gun video projectors that can be seen in the upper portion of figure 4. In addition, there are two video monitors and a video tape deck to the right of the main screens at the front of the room. These machines faciitate reviewing of videotapes taken of mafunctioned equipment in the fied and sent to the diagnostic center. The operation of the entire room is controed from a consoe at the back. From this consoe, information in various company computers can be accessed. For exampe, engineering information on the design of a particuar turbine generator can be obtained, thus aowing personne to concentrate on probem soving rather than searching through voumes of papers and notebooks for information. A data and diagnoses of equipment in the fied are avaiabe instantaneousy to everyone in the room. The diagnostic conference room is used as an education area for both customers and company personne. For those who have had no exposure to the product or to AI, severa demonstrations are avaiabe, One demonstration is composed of an introductory portion that gives the overa goas of the project, a medica exampe to expain expert systems and how they are used in diagnoses, and an expanation of the diagnostic system being demonstrated. Next, a demonstration is presented of the diagnostic system in operation when a generator is experiencing an abnorma condition. Fied data stored in the computer are fed into the diagnostic program, and the resuts are presented on the screen. As the condition of the generator continues to deteriorate, each diagnosis is presented to the audience as if the diagnostic center were actuay diagnosing the unit using ive data. The audience is drawn into the action and begins to understand what the product is and how it works in a way that is hard to obtain using other methods. Over a thousand peope have now seen this demonstration. The customer aboratory is another area of the diagnostic center, used to test software and hardware systems before shipment to the fied. It aso dispays how the data center and various fied monitors are integrated into the diagnostic center to form the diagnostic system product. The operations center is dedicated to 24-hour-per-day, seven day-a-week support of our utiity customers. The center is staffed by a diagnostician who reviews a onine diagnoses being done on customer equipment. A description of how the operations center works with the data centers in the fied is given in The Data Center. Two AI aboratories are aso in the diagnostic center. The interior of one of the aboratories is shown in figure 5. The knowedge engineer sits in the area in the center of the picture, in front of two CRT screens. The experts sit at the tabe in the foreground and have CRTs that contain the same information as those in front of the knowedge engineer. The aboratory is designed to faciitate the transfer of knowedge from the experts to the computer. The knowedge engineer is responsibe for rue base generation. The engineer s responsibiities incude interviewing the experts in each equipment area, extracting their knowedge, and pacing that knowedge in the computer. The knowedge engineer is the ony person who has the authority and the responsibiity for entering the rues into the computer. When experts disagree, the knowedge engineer must determine what actuay is entered into the rue base. Whenever possibe, two or three experts are interviewed at a time. This provides a synergy where the 100 AI MAGAZINE

5 experts reinforce and add to each other s knowedge. It aso provides checks and baances on the knowedge that goes into the rue base because severa experts are ikey to identify a questionabe rues or confidence factors, preventing them from going into the rue base. Interviewing sessions are generay imited to a haf day because the process is very tiring. Probems of noncooperation of experts has been virtuay nonexistent, which is a tribute to those invoved. Aso, a demonstration of the operation of the diagnostic system is given to the experts so that they have a cear understanding of the product they are contributing to before the interview begins. The interview process invoves the experts and the knowedge engineer concentrating on a particuar equipment area. When a the rues have been generated in a session, the experts eave. The knowedge engineer subsequenty draws a diagram of the rue base and enters the rues into the computer. Once the knowedge engineer is satisfied with the operation of the new portion of the rue base, the experts are reconvened. The engineer then exercises the rue base by inputting data. The experts are asked if they agree with the resuting diagnosis. If they agree, no changes are made. If they disagree however, corrections are agreed to, the rue base is modified, and the resuting rue base is again exercised with appropriate data. The process is repeated unti a satisfactory rue base is created. Another method used to ensure a high-quaity rue basis is the design review. Design reviews are conducted by a pane of engineers knowedgeabe of the equipment being diagnosed. Most of these pane members are not invoved in the generation of the rue base. A ist of action items is generated by the pane, and the knowedge engineer is responsibe for modifying the rue base to satisfy these action items. The design review is a very effective too in rue base deveopment. In order to be effective, the knowedge engineer must have a good knowedge of the equipment the diagnostic rues are for. Without such knowedge, the engineer is severey handicapped in deaing with the experts and in making judgments of what finay goes into the rue base. Basicay, the experts shoud fee that the knowedge engineer understands the detais of the knowedge being transmitted. However, the knowedge engineer does not need to know Lisp. Generay, it is very difficut to find a knowedge engineer who knows the equipment and who aso has a Lisp background. Engineers both in Orando and at the research and deveopment center in Pittsburgh who know Lisp and can have the PDS too modified when necessary, are avaiabe to the knowedge engineer. Virtuay a of the many improvements made in the diagnostic center programs are to satisfy the needs of the knowedge engineers. We have found that there is a continua need to improve and extend the programs to suit the particuar requirements of the diagnostic products being deveoped. PLANT A PLANT B CENTRALIZED (A DIAGNOSTIC SOFTWARE) A MODEM PLANT Figure 6. Diagnostic System Schematic. The Data Center Other extremey important parts of the diagnostic system product are the data center (Osborne, Kemper, Emery, Trosky, Logan, and Rodriquez 1985) and the monitors that are ocated in the power pant. Figure 6 is a diagram of the entire diagnostic system, showing the data centers feeding data to the diagnostic center and receiving the resuting diagnosis for dispay to the operator. Figure 7 shows a data center. The consoe is designed to hande one or two generators. The operator utiizes a touch screen, coor CRT for primary communication with the system as shown on the right side of the picture. On the eft is the input/output cabinet. A keyboard is avaiabe for the few times that data are entered or for the eectronic mai feature that is used to communicate with the diagnostic center. The data center was designed to be both an advanced, inpant data monitoring system and an extension of the diagnostic center that can dispay the diagnosis to the operator and give direct recommendations on what actions shoud be taken. The diagnosis is presented in the form of a ist of C FALL

6 Figure 7 Data Center Consoe. variabes, there might be an action the operator can take, such as change the oad or read a pressure gage and enter the vaue into the data center. The diagnostic center then sends a new diagnosis based on both the onine variabes and any vaues manuay input into the data center. The reason such procedures are given is that if a variabe is used infrequenty and primariy to increase the confidence factor, instaing an onine sensor can t be justified economicay. The procedure, in essence, substitutes for the onine sensor during the few times that data is needed. An eectronic mai capabiity is buit into the data center to aow the power pant operator to send messages to the human diagnostician at the diagnostic center and vice versa. Because the diagnostic center is manned 24 hours a day, assistance can be obtained beyond the onine diagnosis. For exampe, the diagnostician can access databases on equipment history and detaied engineering design information. One of the key functions the diagnostician performs is recognizing those few instances when a potentia probem occurs that is not in the rue base. The diagnostician assists the operator with the immediate diagnosis and then has the rue base reviewed and updated with new rues that reate measured variabes to the mafunction. Without this centraized approach, it woud be difficut to improve the rue base. Figure 8. Typica Diagnostic Dispay. candidate conditions. A typica ist is shown in figure 8. The condition (mafunction) with the highest confidence factor is aways at the top of the ist. Confidence factors have vaues from negative one to positive one. A confidence factor of positive one means the diagnostic system is absoutey sure that the condition actuay exists and a confidence factor of negative one indicates absoute surety the condition does not exist. Confidence factors near zero indicate uncertainty. This condition typicay occurs when measured variabes associated with the condition deviate very itte from norma. This is consistent with what a human diagnostician woud te the operator under simiar circumstances. The system is designed to give the operator the abiity to ask the diagnostic center computer through the in-pant data center for a procedure that can be impemented to increase the quaity of the diagnosis. For exampe, if two different equipment conditions have simiar effects on the monitored Customer invovement An important part of the deveopment of the diagnostic system has been customer invovement. A project to verify the diagnostic system was jointy undertaken in 1984 with the Texas Utiities Generating Company (TUGCO) (Osborne, Gonzaez, Weeks, and Martin 1986). The objective was to monitor and diagnose seven of their generators in east Texas. Phase 1 of the project incuded modifying the existing generator monitoring systems at each pant to permit them to transmit reports of data every hour as we as when an aarm condition took pace. These data were then transated into a diagnosis by the generator diagnostic system and dispayed at the diagnostic center in Orando. If a deveoping abnorma condition was diagnosed, diagnostic center personne contacted the pant operators with the diagnosis and recommendations. The phase 1 system was imited because the modified monitoring system did not have the capabiity of communicating between the Orando diagnostic center and the pant. Any communication in this direction took pace by a separate eectronic mai system or by teephone ine. However, the data center now has the capabiity of fu, two-way communications, so that the diagnostic output is accessibe to the pant personne. The data center data-acquisition system monitors approximatey 110 different sensors in a gas-cooed generator and its auxiiaries. This is compared to the phase 1 system tota of 22 sensors, which imited the number of identifiabe conditions in the phase 1 system to 94. The phase 2 system is 102 AI MAGAZINE

7 presenty abe to identify approximatey 350 conditions, using over 8,500 rues. Another major difference between the phase 1 and phase 2 systems as it appies to diagnostics is the frequency of data transmissions. The data center (phase 2) can transmit continuousy through a dedicated data ink, thus aowing a high degree of trending. Trending was ess feasibe with the phase 1 system, which caed once an hour and ony transmitted one set of data at a time. Operationa Experience The phase 1 system started operation in Juy Since that time, a number of incidents have been correcty diagnosed. The incidents range in importance from sensor faiures to broken conductors in the stator winding of the generator. The vaue of the diagnostic system is best iustrated by an incident on January 25, 1985, when broken phase coi conductors in one of the monitored units caused the unit to be brought offine. The diagnostic expert system correcty diagnosed the situation 2-1 I2 hours before the conditions reached the aarm eve. The unit was removed from service before serious damage occurred and was on turning gear ready for synchronization four days ater. The utiity avoided a potentiay costy event by recognizing the condition eary and taking the appropriate action. The capabiity of this system to discriminate fase aarms from rea emergencies was demonstrated during phase 1. In February 1985, correction factor inaccuracies in a temperature-normaizing agorithm caused temperature deviations to exist. These deviations coud have been interpreted as broken conductors on the generator stator winding. However, the diagnostic system correcty diagnosed the situation as bad correction factors. Thus, an unnecessary unit shutdown was avoided. In September 1984, the expert system made use of a temperature comparison scheme to identify a sensor mafunction that otherwise coud not be identified through conventiona range checks. In a second incident in September 1984, the diagnostic system diagnosed a conductor discontinuity at a particuar ocation. Because the severity was not high, a recommendation was made to continue running for the remaining three weeks before a panned outage but to be ready to repair the probem at that time. Inspection of the generator three weeks ater proved the diagnosis to be accurate. This incident iustrates the vaue of the diagnostic system as a predictive maintenance too. The phase 1 expert system, due to the sensor input imitations, was ony abe to monitor and diagnose conditions in the generator stator winding system. The phase 2 system, however, wi not ony be abe to do what the phase 1 system did but wi aso be abe to diagnose probems in the remainder of the generator, which incudes the hydrogen auxiiary system, the sea oi system, the generator rotor, and the excitation system and its contro. Presenty, the data centers have been instaed in each pant, and initia testing has been done by TUGCO instrument and contro engineers. As a resut of both this testing and earier discussions with TUGCO personne, a number of changes to the data centers have been made. These changes have improved the functionaity of the system and increased its user friendiness to the operators. Other improvements incude the abiity to utiize time histories of variabes and the use of procedures dispayed to pant personne to obtain additiona data. Concusions Based on our fied experience to date, severa concusions can be made. The use of a software too such as PDS is very important in the deveopment of diagnostic systems because it aows personne who are experts in equipment diagnosis to be utiized as knowedge engineers, even though they do not know a speciaized computer anguage such as Lisp. The use of an AI approach to onine equipment diagnostics is practica. To create a commercia onine diagnostic system takes a arge number of highy trained, experienced engineers in a number of discipines dedicated soey to the system s deveopment. References Beows,.I. C 1984 An Artificia Inteigence Chemistry Diagnostic System In Proceedings of the Fortyfifth Internationa Water Conference, Pittsburgh, Pennsyvania Fox, M S On Inheritance in Knowedge Representation. In Proceedings of the Sixth Internationa Joint Conference on Artificia Inteigence, Tokyo, Japan Fox, M S ; Keinosky, P ; and Lowenfed, S Techniques for; Sensor-Based Diagnosis In Proceedings of the Eighth Internationa Joint Conference on Artificia Inteigence, Meno Park, Caif.: American Association for Artificia Inteigence Osborne, R. L ; Gonzaez, A J ; Weeks, C A ; and Martin, J First Year s Experience with Onine Generator Diagnostics Paper presented at 1986 American Power Conference, forthcoming. Chicago, IL. Osborne, R. L ; Kemper, C T ; Emery, F. T.; Trosky, W T ; Logan J R ; and Rodtiquez, R Turbine Generator Data Center for Artificia Inteigence-Based Diagnostics Paper presented at 1985 Joint Power Generation Conference, no pubished proceedings Miwaukee, WI Shortiffe, E. H 1976 Computer-Based Medica Consutations MYCIN. New York: American Esevier. Wise, B. P An Experimenta Comparison of Uncertain Inference Systems Ph D diss, Dept of Engineering and Pubic Poicy, Carnegie-Meon Univ Wright, J M., and Fox, M S 1982 SRL/.S User Manua, The Robotics Institute, Carnegie-Meon Univ FALL