DrillEdge is a software system that provides real-time deci-sion support when drilling oil wells. Decisions are sup-ported through analyzing real-time data streams of

parameters measured both on the surface and downhole whendrilling. The real-time analysis identifies symptoms of problems,which are combined to provide best practices for how to handlethe current situation. Verdande Technology has developedDrillEdge to reduce the cost and decrease the probability of fail-ures in oil well drilling. Currently, DrillEdge continuously mon-itors around 30 oil well drilling operations in parallel for sever-al customers and has been deployed commercially for two years.Verdande Technology’s customers include Baker Hughes, Petro-leum Development Oman (PDO), and Shell, among others. InMarch 2011, Verdande Technology was awarded the MeritousAward for Engineering Excellence for its DrillEdge software plat-form by E&P Magazine.

More and more oil well drilling operators monitor a large partof their drilling operations in real time from off-site real-timeoperation centers (RTOCs), which are located close to theirexperienced subject matter experts of various kinds. In this waythe operators seek to provide remote support and ensure thatknowledge can be transferred between shifts, operations, andregions. Typically, less experienced personnel are located at thedrilling rig and more experienced personnel are stationed in theRTOCs. The experiences gained from handling a problem at one

Articles

SPRING 2013 21Copyright © 2013, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602

A Real-TimeDecision Support System

for High Cost Oil WellDrilling Operations

Odd Erik Gundersen, Frode Sørmo, Agnar Aamodt, Pal Skalle

n In this paper we present DrillEdge — a com-mercial and award-winning software systemthat monitors oil well drilling operations inorder to reduce nonproductive time (NPT).DrillEdge utilizes case-based reasoning withtemporal representations on streaming real-time data, pattern matching, and agent systemsto predict problems and give advice on how tomitigate the problems. We document the meth-ods utilized, the architecture, the GUI. and thedevelopment cost in addition to two case stud-ies.

rig can be utilized by the personnel in the RTOCwhen handling another problem later on at a dif-ferent rig probably located in another region.Another advantage of RTOCs is cost reduction assubject matter experts are centralized. Booth(2011) has documented the advantages and histo-ry of RTOCs.

Monitoring a situation is to become aware of it.This is the main concern for decision makers. End-sley (1995) investigated situation awareness indepth and identified three levels of situationawareness. First the elements of the situation areperceived, then the situation must be compre-hended, and finally the future state of the situationcan be projected. It is only after the decision mak-er is aware of the situation that a decision can bemade regarding how to handle it. These three lev-els apply to oil well drilling operations. Experiencewith similar situations is valuable when makingdecisions. By recalling similar problems, differentoptions for actions and risk assessment may beidentified (Crichton, Lauche, and Flin 2005).

Reusing past experience, that is, being remindedof similar situations and making use of decisionsteps made earlier, has turned out to be an efficientway for human beings to handle new situations.Case-based reasoning (CBR) (Kolodner 1992) is acomputational method (Watson 1999) that isbased on this principle. One of its roots stems fromRoger C. Schank’s work on representation of mean-ing for generation of natural language (Schank1980). While working with natural language pro-cessing, Schank started considering inference asbeing the most important part and proposedscripts as a method to connect events together bythe remembrance that they have been connectedbefore, which again led to the study of dynamicmemory (Schank 1983) for situation understand-ing, problem solving, and learning.

The leap from memory structures and retrieval algo-rithms that allow understanding to happen to theuse of cases in other sorts of reasoning is a shortone. [...] Further observations showed that, indeed,cases were useful in problem solving (Kolodner[1993], page 102).

In CBR, new problems are solved by reusing thesolutions of the most similar past problems storedin a case base. Furthermore, newly solved problemsare stored in the case base and thus incrementaland sustained learning is supported intrinsically bythe reasoning method. The reasoning process canbe viewed as a cyclic, four-step process oftenreferred to as the CBR cycle (Aamodt and Plaza1994).

First, the current problem is compared to thepast cases in the case base, and the most similarcases are retrieved. Second, the solutions of themost similar cases are reused to solve the currentproblem. Third, the suggested solution is revised

and possibly updated, and finally the current prob-lem with its revised solution is retained in the casebase. A case in DrillEdge is a concrete drilling situ-ation that comprises a collection of symptoms thatpreviously lead to a problem and a recommonda-tion for how to handle a similar situation. Casesare captured from actual historic drilling data andcontain the operators’ best practices for how tohandle the situations.

DrillEdge is not alone in being a commerciallydeployed CBR system, as CBR systems have a longtradition in this regard (Watson 1997, Cheethamand Watson 2005). General Electric has deployedseveral CBR systems internally, and one of them isa software tool for determining color formulas thatmatch requested colors (Cheetham 2004). CBReven helped users with specially compiled TVguides that suit their particular viewing preferences(Cotter and Smyth 2000).

In the next two sections we explain some coreterms of oil well drilling and explain the type ofproblem addressed by the DrillEdge system. This isfollowed by three sections in which the rationalefor our method choices are described, also describ-ing the two main method types — pattern recog-nition and CBR — in more detail. After a sectionon related work, the DrillEdge system architectureis described. Information about the system devel-opment process and costs are provided, and adescription of the system’s deployment status isgiven. Finally, we present two case studies fromactual deployment.

The ProblemThe main task of drilling engineers situatedremotely in a RTOC is to monitor and understandthe situation on the oil well drilling rigs in order tobe able to support the rig crews if a problematic sit-uation occurs. Problems that might occur that canbe detected by remote observation of drillingparameters are when the drilling pipe gets stuck inthe ground because of gravel pack around it, thedrill string twists, off and drilling mud is lost in theformation. There are several problems related tothe task of monitoring the drilling situationremotely.

Perceiving elements of the drilling situations isdone through monitoring the real-time parametergraphs. For some drilling engineers, this meansstaring at graphs for 12 hours straight. This is ofcourse a tiring and boring task, which might resultin important symptoms and trends being over-looked. Therefore all elements of the situationmight not be perceived.

To comprehend the current situation, all rele-vant information must be available. Even if theperception of elements is performed perfectly, astwo or more shifts perform the monitoring task,

Articles

22 AI MAGAZINE

important information might be lost because ofrotating the shifts. Symptoms of problems thatmight have been deemed unimportant in the pastmight be crucial in understanding what is hap-pening currently. If these seemingly unimportantsymptoms were observed by another shift but werenot communicated, making the right decision inthe current situation is hard.

Another problem is the huge amount of dataprovided by different sources. Understanding aproblem is to build a mental model of it based onthe relevant information. Often the problem for adrilling engineer is to combine all the informationor to find the right clue even when the data isavailable (Booth 2011). Some of the information isprovided in real time, such as some physical mod-els that are updated accordingly, while other infor-mation needs to be computed manually. Otherinformation is stored in the drilling plan or dailydrilling reports. Combining all the relevant infor-mation to get the full view constitutes a high work-load and is not practically possible.

Being able to project the future status of the sit-uations requires not only theoretical knowledgebut also experience. One of the reasons for movingexpert knowledge from the rigs to RTOCs was tocolocate the experience. However, there is stillunsatisfactory experience transfer in the oil welldrilling industry, which easily leads to a lack ofexperience when it would be needed. The twomain reasons for this are the age gap and the con-nection between personnel demands and the oilprice. Because the oil price governs the personneldemands, a significant number of personnel arelaid off when the oil price is low. The age gapdescribes the current distribution of personnelexperience. Recently, the oil price has been highand the demand for personnel is high. Also, manyare close to retirement age, and there are few peo-ple in between. Therefore the industry expects agap in the experience needed.

Lack of experience can be mitigated by lookingup best practices or incident reports that are rele-vant to the current situation. However, as search-ing for the right information in a data base orreport often results in too many or too few relevantdocuments, this is rarely done.

Motivation for Using AI Technology

In the past decade, the focus in the oil and gasindustry has been on collecting drilling data andmaking it available remotely in real time. Thisfocus has seen the development of standards fortransferring data as well as web-service-based stan-dards for accessing databases. As a result, there is

now a common platform to plug into data streams.However, applications that plug into this infra-structure to do useful things with the data havebeen limited and initially focused on visualizationtools to support manual interpretation of data incentralized real-time operation centers. The ideabehind RTOCs is that if a problem develops, deepexpertise from across the organization can bebrought in to assist, rather than having to dependsolely on people at the rig site. However, in orderto bring expertise in on a well, the rig or the RTOCmust first recognize that a problem is developing,and manually monitoring sensor data from a rig isa highly skilled and work-intensive task. It is diffi-cult for even an experienced person to monitormore than a few operations at a time manually.This means that unless such monitoring can bepartially or fully automated, the value of RTOCscannot be fully realized.

The most obvious way to automate rig monitor-ing is to set more or less complex alarms that willnotify human operators if some value or combina-tion of values crosses a threshold. This type oftechnology exists, but is of very limited usebecause most symptoms are much more complexthan what can be captured by such an alarm. Tra-ditionally, the oil & gas industry has relied onphysical models to generate expected data formany sensor values, such as the pressure down inthe hole with the drill bit. These physical modelsare essential tools in planning a well, and muchresearch in the industry has been directed towardmaking real-time versions of these models that arecontinuously updated and tuned with data fromthe well. This is a challenge because it involvessuch issues as inverse modeling of fluids, takingsingle point measurements of pressure and flowand working out the fluid dynamics of the wholesystem. Further complications arise as the physicalproperties of the formation are not completelyunderstood in many cases, and differences inequipment cause differences in sensor responses.Still, drilling specialists are able to interpret thedata to understand what is going on in the opera-tion. In this situation, we are interested in investi-gating a data-driven AI approach to modeling theheuristics used by domain experts, rather than theuse of full physical models. This approach hasproved highly successful.

Our initial approach to the problem was to useinstance-based reasoning and other machine-learning approaches to predict problems directly.This very quickly proved difficult. First, there arerelatively few examples of serious incidents. Anincident such as stuck pipe, where the drill stringgets stuck and part of the well has to be redrilled,can happen a few times a year. It is therefore nec-essary to have a system that is able to learn fromvery few examples, ideally a single case. The data

Articles

SPRING 2013 23

dimensionality and dynamics increase the difficul-ty of the task. Although there may not be morethan 10 to 20 parameters to monitor, it is not pos-sible to detect symptoms of these problems with-out taking the dynamics into account, often look-ing at trends and the frequency of occurrence ofevents over the last 12 to 24 hours. Learning suchdynamics directly from sensor data would beextremely difficult even with an abundance of cas-es, but to do so with only a few examples is clearlynot feasible. The approach we have taken is to usea two-stage approach. First, pattern matchingagents are used to identify symptoms in the databy forming abstractions over the sensor data, thenCBR is used to identify whether this set of symp-toms has caused problems in similar situations inthe past.

The pattern-matching agents making up the firststep of this process are designed so that each agentseeks to identify one type of symptom that thedomain experts have identified as important indiagnosing a particular problem. For instance, onetype of symptom that can be predictive of stuckpipe is the pack off (plugging of the wellborearound the drill string) symptom. The process ofcreating these agents is a classical knowledge-acquisition process seen in expert systems in gen-eral. The domain experts are typically able to pro-vide some pointers on what to look for and caneasily identify examples, but the expert is typical-ly not conscious of all the heuristics he or she usesin identifying the pattern. Thus, the creation ofagents that reliably identify symptoms is an itera-tive process where the agent is tested on as muchdata as possible, and the results are studied by theknowledge-acquisition engineer and the domainexpert together in order to improve it for the nextiteration. The abstraction provided by the pattern-matching agents provides an abstraction wheremachine learning can be performed even with rel-atively few instances. Our main reason for choos-ing case-based reasoning during this stage is thatDrillEdge is a decision support system used bydomain experts, and we need to be able to providenot only an answer but also explanatory supportfor that answer. In DrillEdge we have even chosennot to make the prediction of the system explicit.DrillEdge simply displays all the most similar cases(if any is sufficiently similar) and sorts the cases bythe prediction they imply. This allows us to presentDrillEdge as an automated search tool that willbring relevant cases to the attention of the userwhenever a situation develops, rather than a clas-sical style expert system that provides a predictionor a diagnosis. This has helped in presentingDrillEdge as a tool rather than a threat to theexpertise of the user.

Symptom RecognitionSeveral methods of symptom recognition are beingand have been tested. So far we have settled on twostable and reliable methods: rudimentary mathe-matical modeling and deviation from expectedtrend.

Rudimentary mathematical modeling isable to describe the following type of symp-toms: formation hardness, soft formation, andhard stringers. Advanced models provide goodunderstanding of the underlying physical process-es, but they can be computationally very expensiveand highly dependent on data quality and on theassumptions. Therefore, in many occasions, it isbetter to use simple models by just focusing onselected effects. A simple or a rudimentary model,as opposed to an advanced model, is characterizedby selecting only the most obvious influencingparameters and by ignoring unimportant effects.Such models do not explain or model every aspectof the process but are limited to those parametersthat are important for the specific deviation fromthe normal background level. A simple model offormation hardness is exemplified below:

ROP = C1 · DR · WOBC2 · RPMC3 (1)

Here, ROP is rate of penetration, DR is drillingresistance, WOB is weight on bit, and RPM is rota-tions per minute. Testing this equation for roller-cone bits has shown that best results were foundwhen C2 and C3 were equal to 1.5 and 1 respec-tively. C1 is taking into account all effects notexplicitly included in the previous equation, likebit type, bit characteristics, hydraulics, pore pres-sure, and so on. These effects are assumed constantafter drilling has started.

FH = C1 · DR = ROP/(WOB1.5 · RPM) (2)

If formation hardness, FH, is above a certainthreshold value recorded over a drilled distance ofless than 2 meters, then hard stringers and soft for-mation are triggered and marked in the real-timedrilling data.

We have selected the problem of interpretingrepeatedly mechanical resistance during the trip-ping operation to introduce deviation from expect-ed trend, illustrated by the two symptoms overpulland took weight. Tripping out is to remove the drillpipe from the hole to replace a dull drill bit, forexample, while tripping in is to run the drill bitback to the bottom of the hole. Overpull is whenthe weight on the hook carrying the whole drillstring suddenly gets higher when tripping out, anda took weight is when the weight of the whole drillstring suddenly gets lower when tripping in. To beable to detect the event overpull when tripping outon the basis of this method we need first to estab-lish a normal HKL (hook load) value.

Articles

24 AI MAGAZINE

Case-Based ReasoningWhile the heuristic mathematical models identifythe symptoms observed while drilling, the CBRengine compares the current drilling situationwith past drilling situations stored in the case basein order to find out whether the current situationdevelops into a problematic one. Thus the CBRengine is diagnosing the current situation and thisencompasses finding out which type of problem isbeing observed. The past drilling situations storedin the case base are selected because they representproblems experienced in the past that the operatorwants to avoid. However, the role of a case is notonly to classify the current situation, but also toprovide experience and best practices for how tohandle the situation. Also, the outcome of the pastsituation is known, and it indicates what mighthappen in the current situation if the problemdevelops and corrective action is not taken.

Symptoms of drilling problems typically happenaround the drill bit, and the drill bit can be placedanywhere in the well. As the drill bit is not at thebottom of the well at all times but often is pulledup from the bottom when cleaning the hole orcompletely out of the hole when a change of toolsis required, the position where a symptom isobserved is essential. Some sections of the well(that is, vertical parts) might be particularly prob-lematic, and in such sections the symptoms typi-cally cluster together. Also, the exact time a symp-tom has been observed is important, as it isessential for the operator to know whether the rigis experiencing problems right now. So, both thelocation in the well and the time when a symptomwas observed are important.

Some problematic situations are characterizedby symptoms occurring in the same section whereprevious symptoms were observed, while other sit-uations are characterized by symptoms occurringrepeatedly over the last couple of hours. In somesituations the operators should be warned thatthey are pulling through a section in which prob-lems were experienced earlier.

However, symptoms of problems are not theonly features that characterize a problematic oilwell drilling situation. Some problematic situa-tions are characterized by the design of the bottomhole assembly that is used, such as which drill bit,the number of and the placement of stabilizers,and whether a mud motor is used or not. Others,again, are related to a specific formation or its com-position. Also the type of drilling fluid that is usedand its composition might be relevant, as somedrilling fluids react with the formation, which cancause swelling of the well making it thinner. Theproblems experienced in vertical wells might bequite different from the ones experienced in onewith a horizontal trajectory. Considering the factthat paths of wells are getting more and more com-

plex, some wells have trajectories that curve morethan 90 degrees from the vertical (going upwardrather than downward), the trajectory of the wellis important too. Hence, the context in which thesymptoms occur is an important factor when dif-ferentiating between problematic situations, espe-cially when trying to identify possible root causes.

Cases have two parts: The case description andthe case solution. The case description is used bythe computer to compare the two cases while thecase solution is an experience transfer from onehuman to another. All the above-mentioned fea-tures are parts of the case description. The problemsolution part is a textual description intended forhumans, and it has four subparts. A problemdescription part that describes the operation theywere performing in the specific situation, a symp-toms part that describes the symptoms they wereexperiencing, a response action that describes theactual actions they took in this situation to rectifythe problem, and a recommended action, based ona postanalysis, that describes what should havebeen done. The problem description can easilycontain pointers and links to best practices or inci-dent reports that are stored in different softwaresystems. Cases are captured and written by drillingexperts that are trained in recognizing problemat-ic situations with symptoms that can be identifiedby DrillEdge.

Cases are represented in tree structures andstored as XML files. The root node of the case con-tains two main sections, the problem descriptionand the problem solution. Sections might containother sections or leaf nodes. For example, the prob-lem description contains the formation, drillingfluid, bottom hole assembly, well geometry, andsymptoms. The formation section contains the for-mation name and the formation composition(lithology), while the drilling fluid section con-tains properties describing the drilling fluid that isused, such as mud weight and whether it is oilbased or water based. The bottom hole assemblydescribes the design of the bottom hole assembly.Important features are how long it is, which drillbit is used, the number of stabilizers, and theirpositions. The section well geometry representstarget depth of the current section and the depthof where the section started. The sequence sectionis a special kind of section that contains events,and events are representations used for symptoms.Events have different types corresponding to thetype of symptoms they represent. The symptomssection contains two sequence sections. One rep-resents the distribution of events over a givendepth around the drill bit, while the other repre-sents the events distributed over a limited timeperiod. The end of the time sequence is the timethe drill bit was located at a given depth, and thestart of the time sequence represents the time the

Articles

SPRING 2013 25

current situation started. The ranges above andbelow the drill bit in the depth sequence indicatethe section of the well that is relevant for the situ-ation in the opinion of the expert that built thecase.

The degree of similarity between two cases isfound by comparing the root nodes in the casetrees. The similarities of root nodes are aggregatedinto section similarities, and section similarities arecombined recursively until the similarity of theroot nodes is found. The similarity of the rootnodes is the resulting similarity of the case com-parison. Root nodes can be of different types, likeintegers, doubles, enumerations, and sequences.For each type of node a set of similarity measurescan be configured for comparison. For example,not all numeric features are compared using thesame type of similarity measure. Both standard

similarity measures (Richter 2008) and custom-made similarity measures, which are domain spe-cific, are used to compare features.

The CBR process continuously compares thecurrent situation with cases stored in the case base.Figure 1 illustrates the reasoning process. For eachtime step, the real-time data parameters are inter-preted, and if symptoms of problems are identified,events are fired. The current situation is represent-ed by both important events and contextual infor-mation. Events are stored in the case as depth andtime sequences sorted on the distance from thedrill bit and distance from the current time, respec-tively. The CBR system searches for and retrievescases from the case base and compares them to thecurrent case. The comparison result is sorted onsimilarity. All past cases with a degree of similarityabove a given threshold are visualized on a GUI

Articles

26 AI MAGAZINE

Real-TimeData (1)

Search &Retrieve

Alert &

Advice

New Case

Learn

Case Radar (5)

InterpretFormCase

SituationalDescription

Advice &Experience

Retained Case (6)

Took Weight

ImportantEvents (2)

Pack Off

Case Base

Retrieved Case (4)

SituationalDescription

Advice &Experience

Input Case (3)

SituationalDescription

Figure 1. The DrillEdge CBR Cycle.

element, the radar, to alert and advise the user ofpast historic cases that are similar to the current sit-uation. By investigating the similar situations theuser is advised on what others did and what shouldhave been done in similar situations in the past. Anew case is created by the drilling engineer if thecurrent situation is not covered by any cases storedin the case base or if other advice applies to this sit-uation. New cases are typically quality assuredthrough peer review by a group of domain experts.The system learns when new cases are added to thecase base.

Related WorkSeveral researchers and companies have tried outcase-based reasoning methods for oil drilling assis-tance. Very few systems have reached deployment,however, and no other system than DrillEdge, toour knowledge, links online data streams to pastcases for real-time decision support in an ongoingdrilling process. Partly based on a previous study(Shokouhi, Aamodt, and Skalle 2010), we discussrelated work where CBR methods are applied to anongoing oil drilling operation as well as applica-tions in related drilling domains, such as well plan-ning, reservoir engineering, and petroleum geolo-gy.

An application of CBR in planning of a drillingoperation was described by the Australian researchorganization CSIRO (Kravis and Irrgang 2005). Thesystem, Genesis, can use multiple cases at varyinglevels of generalization. Mendes, Guilherme, andMorooka (2001) developed an application of CBRin offshore well design. A genetic algorithm (GA)was used to determine the proper trajectory of thewell, starting out from a set of solutions that formthe initial population. The cases retrieved throughCBR constitute the initial population. Perry et al.(2004) developed a case-based system for drillingperformance optimization. Project documents,well summary documents, and technical lessondocuments are three levels of documents in theknowledge base hierarchy.

Popa et al. (2008) applied CBR to determine theoptimum cleaning technique for filter failures. Asmall subset of historical cases was taken from thedatabase to evaluate the proposed solution withthe actual results. According to the similarityassessment, 80 percent of the cases were correctlyassigned to the successful cleaning method.Bhushan and Hopkinson (2002) developed a CBRsystem to search for reservoir analogs in the plan-ning of new fields. A knowledge-sharing tool,called the Smart Reservoir Prospector (SRP), wasdeveloped.

A CBR framework was developed by Schlum-berger to assess the applicability of seven lift meth-ods for land, platform, and subsea wells (Sinha,

Yan, and Jalali 2003). Abel, Silvaa, Campbell, andDe Rosc developed a CBR system to support theinterpretation and classification of new rock sam-ples (Abel, Reategui, and Castilho 1996). To pro-vide petrographic analyses, the system achieves itsreasoning power through the set of previous casescombined with other domain knowledge. Later thesystem was introduced into a real corporate envi-ronment (Abel et al. 2005).

Research related to the DrillEdge system, but notpart of it, has been done as joint work between Ver-dande Technology, Statoil, and our universitygroups. One line of research has been to study theeffects of including an explicit model of generaldomain knowledge with the cases, as done in theearlier Creek system (Aamodt 2004). Shokouhi etal. (2009) utilized a newly developed version ofCreek to integrate case-based and model-based rea-soning (MBR) for oil drilling. Abdollahi et al. stud-ied the problem of uncontrolled release of forma-tion fluids into the well through the lifecycle of awell. It was shown that predefined rules could suc-cessfully be integrated with CBR to obtain causes ofwell leakages (Abdollahi et al. 2008).

DrillEdgeDrillEdge has a client-server architecture in whichthe server is a distributed system running on largecomputing clusters like the Amazon Elastic Com-puting Cloud (ECC) while the clients run on regu-lar desktop machines.

ArchitectureAs depicted in the leftmost part of figure 2, theprocess view of the architecture can be illustratedas a layered architecture with four layers. The low-est level is data acquisition in which data areacquired from the data sources. Data sources canbe both manually updated when starting to moni-tor an operation and updated automatically in realtime. The second lowest level is data interpretationwhere patterns are recognized and symptoms areflagged by software agents. The symptoms are fedinto the case-based reasoning engine, which rec-ognizes broader patterns, not focusing on a small-er set of parameters, but the complete situation.Finally, at the topmost level the data and findingsare visualized for the users.

The middle part of figure 2 depicts a client-serv-er view of the architecture. Each machine in thecluster is indicated by a gray, squared box that runsservices, which are drawn as white boxes with softcorners. Different kinds of services are provided bythe server cluster. The main service is the operationservice, and it is responsible for the three lowestlevels of the processes view, which are related toone real-time drilling operation. In the illustrationtwo machines run a total of four operations desig-

Articles

SPRING 2013 27

nated with an O and a subscript for its ID. Otherservices are the license service, which ensures thatthe operator is not running more operations thanit has licenses for and is designated L, and the man-agement service, designated M, which manages thesetup and configuration of operations. Clients, onthe right, communicate with services through afront-end server, whose only task is to ensure thatthe services get the right messages. All communi-cation between the server cluster and the clients isdone using HTTP.

The rightmost part of figure 2 shows the archi-tecture of the operation service. The operationservice includes a WITSML (Wellsite InformationTransfer Standard Markup Language) client thatrequests data from a WITSML server and is respon-sible for storing the received data in the appropri-ate data structure. Some data are stored in a depthtable while others are stored in a time table. Theagents fetch data from the data structures to per-form their calculations. Some of the agents com-pute new values each time step, for example, trendagents, while others store data only when they findnew symptoms. The CBR engine is also imple-mented as an agent. It is however dependent onthe calculations from all the other agents, so it willnot execute until all other agents have done theircalculations for a given time step.

ClientThe intention behind the client is that it should beeasy to use and self-explanatory. Users can choosebetween several different tabs that visualize data indifferent manners, like parameters plotted against

time or depth, the static properties that the opera-tion is configured with, or an overview. The twomain tabs are the time tab and the overview tab.Figure 3 shows screen captures of the two tabs withthe overview tab on the left and the time tab onthe right. The screen shots are included in order toillustrate the structure of the screens and not theirdetailed contents.

The overview provides three main types of infor-mation: the depth view, the radar, and the casesolutions. The depth view visualizes the drill bit inthe hole, and it shows the current depth and theformation layers and their lithology. Both rotationof the drill bit and mud flow are animated to makeit easier for the users to understand the currentactivity. Also, events are plotted according to depthso if the drill bit is pulled through a section of thewell that was problematic to drill, this can be seenfrom the depth view. The most important GUI ele-ment on the overview screen is the radar, and thisis emphasized by the amount of screen real estateit occupies. Past cases are plotted on the radaraccording to how similar they are to the currentsituation. The more similar a case is to the currentsituation the closer to the center of the radar thecase is located; if a case is 100 percent similar, itwill be in the center. All cases that are less similarthan a threshold will not be plotted on the radar.A threshold of 50 percent has shown to ensure thatonly relevant cases are brought to the attention ofthe user. The radar is divided into different sectorsaccording to the problem area the cases in the casebase are representing, such as lost circulation andmechanically stuck pipe. When clicking on a case on

Articles

28 AI MAGAZINE

O3

O4

O1

O2

M1

L1

FE

Amazon ECC

Https

Client

Client

Client

Case-basedreasoning

Visualization

Data interpretation

Data acquisition

Agent system

A3

A4

A2

A1Timedata

Depthdata

WITSMLclient

Operation service

Figure 2. The DrillEdge Architecture Explained

Process view to the left, the client-server architecture in the middle, and the operation service to the right.

the radar, the case solution will appear on the rightdescribing the past situation, the problem, and rec-ommended actions.

The time tab provides a list of available parame-ters that can be dragged over to the parametergraph columns so that the values will be drawn asgraphs with time flowing downwards. Events aredrawn in the event column so that users can seeexactly when symptoms appeared. As event namesuse common terminology for the users, they caneasily look at the corresponding parameters to seewhether they agree with the symptoms. Casegraphs can be drawn for each case in the case base,and they show the similarity degree of the cases atevery time step. Thus the case graphs convey thehistory of how the current situation has developedseen through the lens of the past cases stored inthe case base.

Development and CostThe development of DrillEdge was based onresearch performed at the Norwegian University ofScience and Technology (NTNU), more specificallywithin the AI and Drilling Engineering researchgroups (Skalle, Sveen, and Aamodt 2000; Aamodt2004). In 2006, the Norwegian Research Council’sPetromaks program in conjunction with Statoil,the largest Norwegian oil company, provided fund-ing for a two-year project to develop a pilot system

for studying whether case-based reasoning andrelated technologies could be used to detect andpredict drilling problems. This project funded thetwo first years of commercial development, pro-viding approximately $3 million to fund 14 man-years of development over these two years. At theend of this period, in 2008, a prototype ofDrillEdge was finished, although another 6 man-years of development were required before the firstcustomer version was released in 2009, placing thetotal development cost (excluding marketing, butincluding all overhead) at around $4.5 million.Since then, development has continued both tobroaden the types of drilling problems that can bedetected, but also to develop features such asadministrator tools, email alerts, support for clouddeployments, and so on. Today, the DrillEdgedevelopment team consists of 12 full-time devel-opers and testers.

Commercial DeploymentDrilling for oil and gas is very expensive, especial-ly in offshore locations. In the North Sea, a single6000 meter well can cost $100 million to drill. Theindustry also recognizes that there are significantsavings possible here — depending on the wellcomplexity, the average nonproductive time (NPT)can be as much as 30 percent. Some of this NPT isunavoidable, for example, waiting out especially

Articles

SPRING 2013 29

Figure 3. Two Different Screens from the DrillEdge Client.

The overview screen is on the left side and the time view is shown to the right.

bad weather, but DrillEdge is currently able toaddress problems causing about half of this down-time. At the time of writing, there are 1190 off-shore rigs in the world, about 70 percent of whichare operational at any time.1 These rigs have anaverage day rate (cost for the operator company torent the rig) of $144,000 per rig. Provided that theNPT can be reduced by 5 percent across the off-shore rig fleet, this translates to a cost saving ofover $2.1 billion per year. In addition, there aremore than 3000 rotary land rigs in operation in theworld, a number that is increasing rapidly due tothe drilling for shale gas in the United States (thereare about 2000 rotary land rigs in the United Statesalone, an increase of about 300 over the last year)(Baker Hughes, inc. 2012). These rigs have lowerNPT and day rates than offshore rigs, but we areinterested in real-time drilling data to improve effi-ciency and consistency of drilling.

DrillEdge is in use by Shell in its Real-Time Oper-ating Centers in the United States, where it hasbeen used on many of the most challenging wellsdrilled by Shell across the world in the last sixmonths. It is also installed as a core part of the real-time monitoring infrastructure in the national oilcompany of Oman, PDO. At the time of this writ-ing, DrillEdge monitors more than 20 rigs in Omanalone. In September 2011, Verdande Technologysigned a partnership with Baker Hughes, a majoroil service company, which means Baker Hugheswill use DrillEdge as a key component in its real-time monitoring services as sold to operators, start-ing in 2012. In addition, Verdande Technologyalso executed successful pilot installations for morethan 15 oil operators during 2011.

Case StudiesDrillEdge has been used successfully for analyzingdata for several customers both in pilot tests andcommercial deployment. Two of the case studiesthat have been made are summarized below. Ineach of them DrillEdge analyzed several sets of his-torical data. The tests were performed blindly; Ver-dande Technology was not informed beforehandof which or when the operator had experiencedproblems. Nor was any information about the out-come given.

Case Study: Stuck PipeA major service company wanted to minimize therisk of stuck pipe events for their clients in LatinAmerica and around the world. DrillEdge wasemployed in a postwell historical analysis to deter-mine if the losses and ballooning events couldhave been recognized in advance.

A project was launched to evaluate the potentialof DrillEdge technology to recognize the stuck pipeproblems. Time and depth-based data from several

prior land drilling operations in Latin Americawere employed in a rigorous testing routine. Over-pull events, erratic torque, and pack offs werethought to be key contributors to the stuck pipescenarios. Cases were built in DrillEdge and thesecases were associated with other data, such as BHA,trajectory, mud properties, and the formation. Welldrilling data was then streamed while DrillEdgetechnology monitored it for precursor events. Testswere highly successful with the DrillEdge technol-ogy consistently predicting stuck pipe incidents 6to 8 hours prior to their occurrence. While drilling,this window was deemed sufficient to proactivelyaddress a stuck pipe risk. The success of this proj-ect set the stage for further testing in live field tri-als from RTOCs.

Case Study: Lost CirculationA major operator encountered costly NPT due tomud losses and subsequent hole ballooning.DrillEdge was employed in a postwell historicalanalysis to determine whether the losses and bal-looning events could have been recognized inadvance.

Using a combination of information includingpit volumes, well and drill string geometry, ROPand lag times, the DrillEdge system was able to rec-ognize changes in volume that were not account-ed for by the drilling process. In a historical analy-sis of a well with known problems related to mudlosses and hole ballooning, DrillEdge first saw indi-cations of losses three days before the customerexperienced total losses. During this time, a seriesof loss events was detected, combined with holeballooning at connections. Had DrillEdge beenemployed on this well, the risk associated withthese problems could have potentially been betterassessed, leading to remedial actions based oncompany best practices that could have reducedcostly NPT.

ChallengesWe face mainly three challenges in our work. Theseare revising the symptom recognition agents, thetime used to find and capture a proper case, andthe real-time demands on the similarity compari-son.The symptom-recognition agents recognize pat-tern signatures for the different symptoms by ana-lyzing the change of several parameters in theWITSML data streams. Several factors affect thesymptom signatures, such as the depth, the holediameter, and the tools used. Hence, when enter-ing new markets the conditions can change sig-nificantly, and thus the symptom-recognitionagents need to be revised. Our goal is to avoidregional customization of agents, which makes theagent development considerably more difficult.

Articles

30 AI MAGAZINE

Building good cases that capture significant sit-uations and provide proper advices require morework than first anticipated. Finding the data con-taining problematic situations and quality assur-ance of the data in regard to data frequency andmissing parameters are very time consuming, ascompanies have not really used the stored databefore. After finding and assuring the quality ofthe data, experts analyze it and assure that it con-tains symptoms supporting the claimed problems.Then, the symptom analysis is performed beforeexperts can start identifying where the casesshould be captured. Captured cases need to bequality assured to check how early and whetherthey actually would create relevant alarms for theproblems. Building one case requires around 40man-hours on average by a drilling expert trainedin identifying proper situations.

Providing decision support in real time puts con-straints on the case comparison. We have foundthat sequence matching is the bottleneck, and weare currently studying methods for speeding upthe sequence comparison (Gundersen 2012).

Conclusion and Future WorkIn this article, we have presented DrillEdge, a soft-ware system for supporting decisions in real timewhen performing oil well drilling operations.DrillEdge has been developed since late 2006 andhas been commercially deployed for more thantwo years on more than 200 wells. We show howAI methods such as case-based reasoning, patternmatching, and agent systems have helped large oilwell drilling operators to prevent doing the samemistakes over and over again.

Verdande Technology is currently expandinginto new industries and offers solutions in financeand health care in addition to oil well drilling. Incontrast to the oil well drilling domain, in whichfive minutes reaction time is sufficient, financialservices are very focused on low latency, and thisrequires further improvement of the case-basedreasoning engine.

Note1. See the RIGZONE 2012 offshore rig day rates (www.rig-zone.com).

ReferencesAamodt, A. 2004. Knowledge-Intensive Case-Based Rea-soning in Creek. In Advances in Case-Based Reasoning: The7th European Conference, volume 3155, Lecture Notes inComputer Science, ed. P. Funk and P. González-Calero, 1–15. Berlin: Springer.

Aamodt, A., and Plaza, P. 1994. Case-Based Reason-ing: Foundational Issues, Methodological Variations,and System Approaches. AI Communications 7(1): 39–59.

Abdollahi, J.; Carlsen; I. M.; Randhol, P.; Tenold, E.; Haga,

H.; and Jakobsen, T. 2008. A Case-Based Approach toUnderstand the Complexity of Causal ConnectionsRelated to Well Integrity Problems. SPE 111129-MS. Paperpresented at the IADC/SPE Drilling Conference, Orlando,FL, 4–6 March.

Abel, M.; Silvaa, L.; Campbell, J.; and De Rosc, L.2005. Knowledge Acquisition and InterpretationProblem-Solving Methods for Visual Expertise: Studyof Petroleum-Reservoir Evaluation. Journal of PetroleumScience and Engineering 47(1–2): 51–69.

Abel, M.; Reategui, E. B.; and Castilho, J. M. 1996. UsingCase-Based Reasoning in a System That Supports Petro-graphic Analysis. In Artificial Intelligence in the PetroleumIndustry: Symbolic and Computational Applications II, chap-ter 7, 159–172. Paris: Technip.

Baker Hughes, Inc. 2012. Baker Hughes Investor Rela-tions: Overview and FAQ. Houston, TX: Baker HughesIncorporated.

Bhushan, V., and Hopkinson, S. 2002. A Novel Approachto Identify Reservoir Analogues. Paper presented at theEuropean Petroleum Conference, Aberdeen, Scotland,UK, 29–31 October.

Booth, J. 2011. Real-Time Drilling Operations Centers: AHistory of Functionality and Organizational Purpose —The Second Generation. SPE Drilling & Completion 26(2):295–302.

Cheetham, W. 2004. Tenth Anniversary of the PlasticsColor Formulation Tool. In Proceedings of the 16th Confer-ence on Innovative Applications of Artifical Intelligence, 770–776. Menlo Park, CA: AAAI Press.

Cheetham, W., and Watson, I. D. 2005. Fielded Applica-tions of Case-Based Reasoning. Knowledge EngineeringReview 20(3): 321–323.

Cotter, P., and Smyth, B. 2000. PTV: Intelligent Person-alised TV Guides. In Proceedings of the Seventeenth Nation-al Conference on Artificial Intelligence, 957–964. MenloPark, CA: AAAI Press.

Crichton, M. T.; Lauche, K.; and Flin, R. 2005. IncidentCommand Skills in the Management of an Oil IndustryDrilling Incident: A Case Study. Journal of Contingenciesand Crisis Management 13(3): 116–128.

Endsley, M. R. 1995. Toward a Theory of SituationAwareness in Dynamic Systems. Human Factors: The Jour-nal of the Human Factors and Ergonomics Society 37: 32–64(33).

Gundersen, O. E. 2012. Toward Measuring the Similarityof Complex Event Sequences in Real-Time. Paper pre-sented at the 20th International Conference on Case-Based Reasoning, Lyon, France, 3–6 September.

Kolodner, J. L. 1992. An Introduction to Case-Based Rea-soning. Artificial Intelligence Review 6(1): 3–34.

Kolodner, J. L. 1993. Case-Based Reasoning. San Francis-co: Morgan Kaufmann Publishers.

Kravis, S., and Irrgang, R. 2005. A Case Based Systemfor Oil and Gas Well Design with Risk Assessment.Applied Intelligence 23(1): 39–53.

Mendes, J.; Guilherme, I.; and Morooka, C. 2001. Case-Based System: Indexing and Retrieval with Fuzzy Hyper-cube. In Proceedings of the Joint 9th IFSA World Congressand 20th NAFIPS International Conference. Piscataway, NJ:Institute of Electrical and Electronics Engineers.

Perry, P.; Curry, D.; Kerridge, J.; Lawton, J.; Bowden, D.;

Articles

SPRING 2013 31

and Flett, A. 2004. A Case Based Knowledge Repositoryfor Drilling Optimization. Paper presented at the NinthBiennial International Association of Drilling Contrac-tors (IADC) and the Society of Petroleum Engineers (SPE)Asia Pacific Drilling Technology Conference, SPE 87994-MS. Tianjin, China, 9–11 July.

Popa, A.; Popa, C.; Malamma, M.; and Hicks, J. 2008.Case-Based Reasoning Approach for Well Failure Diag-nostics and Planning. Paper presented at the SPE WesternRegional and Pacific Section AAPG Joint Meeting SPE114229-MS, Edmonton, AB, Canada, 29 March–2 April.

Richter, M. 2008. Similarity. In Case-Based Reasoning onImages and Signals, volume 73, Studies in ComputationalIntelligence, ed. P. Perner, 25–90. Berlin: Springer.

Schank, R. C. 1983. Dynamic Memory: A Theory of Remind-ing and Learning in Computers and People. New York: Cam-bridge University Press.

Schank, R. C. 1980. Language and Memory. Cognitive Sci-ence 4(3): 243 – 284.

Schank, R. C. 1983. Dynamic Memory: A Theory of Remind-ing and Learning in Computers and People. New York: Cam-bridge University Press.

Shokouhi, S.; Aamodt, A.; and Skalle, P. 2010. Applica-tions of CBR in Oil Well Drilling. In Intelligent Information

Processing V, Proceedings of the 6th International Conferenceon Intelligent Information Processing, 102–111. Berlin:Springer.

Shokouhi, S.; Aamodt, A.; Skalle, P.; and Sørmo, F. 2009.Comparing Two Types of Knowledge-Intensive CBR forOptimized Oil Well Drilling. Paper presented at the 4thIndian International Conference on Artificial Intelli-gence, Tumkur, India, 16–18 December.

Sinha, S.; Yan, M.; and Jalali, Y. 2003. Completion Archi-tecture: A Methodology for Integrated Well Planning.Paper presented at the SPE/IADC Middle East DrillingTechnology Conference and Exhibition, SPE 85315-MS.Abu Dhabi, United Arab Emirates, 20–22 October.

Skalle, P.; Sveen, J.; and Aamodt, A. 2000. Improved Effi-ciency of Oil Well Drilling Through Case-Based Reason-ing. In Topics in Artificial Intelligence, 6th Pacific Rim Inter-national Conference on Artificial Intelligence, Lecture Notesin Computer Science 1886, 712–722. Berlin: Springer.

Watson, I. 1999. Case-Based Reasoning Is a MethodologyNot a Technology. Knowledge-Based Systems 12(56): 303–308.

Watson, I. D. 1997. Applying Case-Based Reasoning —Techniques for the Enterprise Systems. San Francisco: Mor-gan Kaufmann.

Odd Erik Gundersen is the research and developmentdirector at Verdande Technology and a research sci-entist at the Norwegian University of Science andTechnology (NTNU). He has published papers oncase-based reasoning, intelligent agents, and physi-cally based animation of fire.

Frode Sørmo is the chief technology officer at and acofounder of Verdande Technology. He received hisPh.D. in 2007 in computer science and artificial intel-ligence from NTNU.

Agnar Aamodt is a professor of computer science andartificial intelligence at the Norwegian University ofScience and Technology and a cofounder of Ver-dande Technology. He received his Ph.D. in 1991 incomputer science and artificial intelligence from theNorwegian University of Science and Technology.His main research interest is in knowledge-baseddecision support, combing case-based and other rea-soning methods. He has spent longer periods in theAI labs at the University of Texas, Austin (1987–88),Free University of Brussels, VUB (1991–92), and at theInstitut d’Investigació en Intel.ligència Artificial(IIIA), Barcelona (2001–02). He is currently a visitingprofessor at the University of Auckland.

Pal Skalle is an associate professor at the NorwegianUniversity of Science and Technology and acofounder of Verdande Technology. He received hisPh.D. in 1983 in drilling engineering from NTNU. Hehas served both as technical and associate editor forthe journal SPED&C, and was in 2005 honored by SPEas an outstanding technical editor and in 2008 as apeer apart. Skalle’s present research activities andfields of interest cover topics like knowledge engi-neering, drilling fluid engineering, evaluation ofreal-time drilling data, and pressure control duringdrilling.

Articles

32 AI MAGAZINE

IAAI-13Bellevue, Washington USA

July 16–18 2013