Configurability in a Diagnostic Expert System for Paper Machine Dryer Sections

J.R. Amyot,1 J.D. Gowing,2 R.H. Wylie,1 R. Henzell3

1 National Research Council of Canada Ottawa, Ontario, Canada K1A 0R63 KanEng Industries, Inc. 16847 Hymus Blvd.

Kirkland, Quebec, Canada H9H 3L4

2 EDS of Canada Ltd. 1615 Dundas Street East

Whitby, Ontario, Canada L1N 7S6

Abstract

A \"Paper Drying Expert System\" (PDES) prototype is being developed as a diagnostic consultant

for troubleshooting dryer sections of paper machines. It is currently undergoing validation in two pulp andpaper mills, scheduled to be completed in September 1994. A requirement that the PDES be configurablefor many possible dryer configurations is a major consideration. This paper describes the project with afocus on how this requirement is satisfied in the design.

Keywords: diagnosis, troubleshooting, expert system, paper making.

Introduction

A \"Paper Drying Expert System\" (PDES) prototype is being developed as a diagnostic consultant fortroubleshooting dryer sections of paper machines. A requirement that it be configurable for many possibledryer configurations is a major consideration. The purpose of this paper is to describe the project with afocus on how this requirement is satisfied in the design.

The project has five partners, i.e., three pulp and paper companies, one manufacturer, and one softwaredeveloper who operate under the Software for Integrated Manufacturing Consortium (SIMCON) umbrella.The development is led by the National Research Council of Canada. The project's duration isapproximately thirteen months. The scheduled completion is October 1994, after a three-month validation.The prototype is PC-based and initially off-line. Diagnosis proceeds on the basis of departures frombaseline operation regarding such parameters as total steam consumption. Given that the energy required toproduce steam is costly, significant economic benefits are anticipated.

The PDES has two major modes of operation: a configuration mode and an operational mode. Theconfiguration mode provides the user with the ability to configure the PDES for the paper machine andpaper grade of interest. This involves specifying the physical characteristics of the dryer section and a setof baseline operational values (including paper speed, basis weight, trim, dryer pressures, and moisturecontent of the paper web into and out of the dryer section). A diagnostic session is initiated by accessing theappropriate configuration and baseline data files. The diagnostic portion of the system is programmed inCLIPS, an object-oriented expert system shell from NASA, whereas the remainder is in C++. Thisselection of software was made on the basis of cost, functionality and ability to be integrated.______________________NRC 37145

In the operational mode, the user sees a graphical representation of the dryer section. He is presented withsome baseline parameter values and asked to enter current values. From the baseline and current values,calculations are performed to determine if the paper machine's dryer section is operating acceptablyaccording to two energy-related performance measures. If the performance is not acceptable, the PDESinitiates a dialog with the user to selectively obtain more information with the objective of diagnosing theproblem and suggesting remedies. The PDES uses the configuration data to specialize the computationsand the diagnostic dialog to the specific dryer section and paper grade under investigation.

There has been much work on the integration of model based and heuristic (or fault based) diagnosisrecently. Typical approaches to this problem include switching from one strategy to the other when a dead-end is encountered, using one strategy to confirm, explain or refine results obtained with the other (Lee andKim, 1993). While practical and appropriate in many cases, such approaches do not relieve the inherentinflexibility and brittleness of the heuristic knowledge. Another approach involves coercing the heuristicknowledge into a model oriented representation, (typically called a functional model), so that model basedreasoning techniques may be applied (Guida et al., 1993). While elegant, this approach disregards the factthat domain experts think in terms of diagnostic procedures.

The PDES represents a different approach to the integration of model and heuristic based reasoning indiagnosis. The approach is motivated mostly by the requirement that the PDES be configurable fordiagnosis of the steam section of any typical paper machine. Succinctly, all information about the state ofthe paper machine is stored in a model, and all computations and tests on system state information areperformed by the model. As a result, the information remaining in the diagnostic tree is independent of anyparticular machine configuration.

Paper Drying Process Description

The paper machine drying process consists of passing the moisture-laden paper web, coming from the presssection of the paper machine, over steam-filled rotating cylinders, called dryers (or cans, or drums). Dryertemperature is controlled by steam pressure. Temperature increases with pressure. Dryers are controlled insteam groups to maintain the same pressure (temperature) within a group. There is usually an increase insteam group temperature from the wet end to the dry end. The number of dryers in a group and the numberof steam groups in a paper machine dryer section vary between machines. A felt is used with many gradesof paper to hold the paper web tightly against the surface of the dryers, to improve heat transfer, and toprevent or reduce wrinkling of the paper as it dries. The dryers are usually driven by electrical motors ingroups that correspond to the felts. These drive groups do not necessarily correspond to the steam groupsor to the differential pressure groups described below.

The steam that is condensed in a dryer (in the process of giving heat to the paper) is evacuated from thedryer and collected as water in various possible configurations of separator tanks. The evacuation ofcondensate from dryers, via syphons, requires that a differential pressure (DP) be maintained, and itusually results in bi-phase flow of steam and condensate (S&C) involving approximately 20% steam.Therefore, in addition to being controlled in steam groups, dryers are also controlled in DP groups thatusually correspond to one or more steam groups. The uncondensed steam may be reused in the dryers of alower-pressure steam group, or it may be reused in the same steam group after having its pressure boostedby means of a thermocompresser. Otherwise, the excess steam is condensed in a condenser, contributing todryer inefficiency in the form of energy losses. Furthermore, some of the dryers at the wet end are oftenoperated at vacuum pressures (i.e., below atmospheric pressure), with resultant low surface temperatures,to avoid sheet problems. These low pressures are maintained by sending the bi-phase flow directly to acondenser (rather than to a separator tank) in a closed system, with the condensing steam creating a

vacuum or negative back pressure. The condensate from the condenser goes to a vacuum receiver beforebeing pumped away, usually back to the boilers. For PDES purposes, a condensate drainage section (orsimply condensate section) can be one of several pre-defined combinations of separator tanks, condensers,and vacuum receiver.

The various configurations of dryers, tanks, condensers and control systems that are encountered in papermachine dryer sections define the configurability requirements of the PDES.

Functionality

The PDES consists of distinct functional entities that either handle or store information.

(a) User Interface

The user interface is a windows based graphical user interface (GUI) that conforms to the look and feel ofthe Microsoft Windows operating system. It provides typical GUI entities such as pull-down menus, pop-up dialog windows, and push-buttons. It can display a simple schematic of the S&C system and will havethe ability to display graphs such as the TAPPI (Technical Association of the Pulp and Paper Industry)graphs of drying rate versus steam temperature. It provides help facilities in the form of Windows helpfiles, that explain the various user interface elements and terminology. The user interface supports bothBritish and metric units on input and output, and is available in either English or French.

(b) Configuration

The PDES can manage descriptions of various S&C systems. A user can create a new description or selectand modify an existing one. The system guides the user through the specification process, providingdefaults where appropriate. The configuration of the S&C system requires detailed information about theindividual components and their groupings.

(c) Calculations

Numerical calculations are performed by the PDES to provide a quantitative description of the currentoperating state of the S&C system. The data required for the calculations consists of previously enteredconfiguration data as well as current operating data provided by the user.

(d) Interpretation

Interpretation is required to convert the numerical description of the S&C state into a symbolic descriptionthat is meaningful for diagnostic purposes (e.g., total steam condensed = 60 000 kg/h; therefore, total steamcondensed is \"too high\"). Interpretation involves comparisons of calculated data or current user-suppliedoperating data, with predetermined baseline values. This baseline is usually predetermined by a completefield study of the S&C system. Other sources, such as the historical data base of the system, paper machinedryer literature, and the user, may also be useful in establishing a baseline.

The baseline describes an acceptable state of the S&C system. All comparisons are made to it via user-specified thresholds. The accuracy of determination of baseline values depends on instrumentationaccuracy and normal process fluctuations. The determination of threshold values is related to the baselineaccuracies, and amounts to tuning the sensitivity of the PDES so that meaningful comparisons can be madebetween the current and baseline operating states. If it is determined that the S&C system is currentlyoperating more efficiently than the baseline state (i.e., the ratio weight of water evaporated from the paperto weight of steam condensed is greater than the baseline value), then the user can reset the baseline to thecurrent operating state. However, If the S&C system is operating below the baseline efficiency, then theresult of the interpretation is: (1) total steam condensed is too high, and/or (2) U-factor is too low. Eitherone of these initiating symptoms is sufficient to trigger the diagnostic process.

(e) Diagnosis

The diagnosis is based on a fault tree approach. The knowledge base is composed of symptom and faultnodes connected by edges corresponding to the different outcomes of diagnostic tests associated with thenodes.

The diagnostic system examines the initiating symptoms and hypothesizes on possible faults or problemsin the dryer section. It then begins a diagnostic session consisting of a series of system questions and userresponses regarding the S&C operating conditions. Some of the observable symptoms that the user may bequestioned about are: (i) the paper web is bagging around one or more dryers, (ii) the electrical current in adrive motor is above normal, (iii) a blow down valve is open, (iv) a condensate tank is flooded, and (v) adifferential pressure is out of range.

The responses are used to confirm or refute the possible hypotheses. The session terminates with a list ofpossible faults and corresponding recommendations for their elimination. Some of the possible faults are:(1) steam supply problem, (2) wet end water removal problem, (3) control loop problems such as steampressure, differential pressure, water level, or moisture control loops, (4) incorrect pressure or DP setpoints, (5) defective components such as thermocompressers, vacuum pumps, condensate pumps, andsyphons, and (6) felt tension problems.

Information Handling System Design

The graphical user interface (GUI) consists of two basic components: (1) a model of the physical dryersection appearing as a schematic diagram on the screen, and (2) the visual objects, such as pull-downmenus, pop-up dialog windows, and push-buttons, that allow the user to create or modify the physicalmodel in conformity with a given dryer section.

The physical model is a collection of C++ classes that represent physical components found in dryersections of paper machines. The main class is the Dryer class. One instance of the Dryer class exists foreach paper machine and contains all the other sub-components in the dryer section. Since the sub-components are variable in number, they are included as lists in the Dryer object.

The dryer section of paper machines can typically be characterized by four functional components, andthere can be a variable number of each per machine. These components, or building blocks, make itpossible to configure the PDES for various paper machines. They also allow for grouping the atomicphysical components so that analysis and diagnosis can be performed on each group, independently ofexactly how the components within the group are connected. Baseline, actual and threshold values ofcomponent variables are stored as attribute values in the component objects.

(1) The first component, a pressure control group or steam group, is a set of dryer cans whose steampressure is independently controlled. This component is used because it is necessary to: (i) know the steampressure in each dryer can to calculate the dryer efficiency, and (ii) perform diagnosis on each pressurecontrol loop.

(2) The next component, a DP group, is a set of steam groups whose DP is controlled by the same DPcontroller. This component is used because it is necessary to perform diagnosis on each DP control loop.Several steam groups may be connected to the same DP control valve(s). In such cases, only one of thesteam groups acts as the master of the DP control loop. This is the steam group whose DP is actuallybeing measured and controlled. The other slave steam groups have a DP equal to that of the master steamgroup plus the difference in steam pressure between the slave and the master.

(3) The next component, a condensate section, is a set of steam groups whose condensates flow to thesame location and whose total condensate flow rate can be determined. The distinction is made between aDP group and a condensate section because the condensate from several steam groups which have differentDP controllers may flow to the same location. A condensate section may contain a separator tank and/orcondensers, but it has only one condensate flow rate. The total amount of condensate being condensed by apaper machine's dryer section is needed to determine the efficiency of the dryer section.

(4) The last component, a drive group, is a set of dryer cans that are rotated by the same drive mechanism.The visual objects are sub-classes of the Microsoft Foundation Class Library (MFC). This is because theinterface is designed for the Microsoft Windows environment. The MFC uses a document / viewarchitecture. The physical model of the paper machine is stored in a document class object, which providesfile access functionality. The visual objects are view and dialog class objects which allow the user to viewand modify different areas inside the document. For example, a visual object exists for each component ofthe Dryer object to allow the user to modify the configuration attributes of each component. Other visualobjects exist to allow the user to view and modify the current operating parameters or the baseline values ofthe paper machine.

The separation of the physical model from the visual objects allows the physical model to be portable todifferent target platforms (windowing environments) as only the visual objects would need to be convertedto the new windowing system.

Diagnostic System Design

Diagnostic reasoning in the PDES can be viewed as a guided tour of the state of the paper machine. At anypoint during the diagnosis, there is a current node and a current component. Each diagnostic inferencecycle involves a change in one or both of these. The effect is to shift the diagnostic focus to the next mostinteresting aspect of the machine's behaviour until enough knowledge about the machine state has beenaccumulated to isolate a specific fault. Only when the diagnostic process cannot complete some test (due tomissing information) does the system interact with the user.

A clear distinction is made between model based knowledge and diagnostic knowledge.

(a) Model based knowledge. Two types of objects, components and variables, are used to model thestructure and behaviour of the physical system. The principal role of the model in the PDES is to allowgeneral knowledge about paper machines (in the form of a standard set of components) to be assembledinto valid descriptions of specific machines. Variables are contained in components, and they encapsulateknowledge about the state of the system. Each variable can manage a number of distinct values. Diagnostictests make comparisons between these values.

(b) Diagnostic knowledge. Diagnostic heuristics are represented as nodes in a diagnostic tree. Nodescontain no knowledge about the state of the system being diagnosed. The dynamic knowledge that can beassociated with a node consists of the diagnostic test results, and whether or not it or its children have beenvisited. There are four types of node: (1) model traversal node, (2) variable test node, (3) informationdisplaying node, and (4) node for moving about the diagnostic tree based on user input information, with noreference to the model.

The first two types of diagnostic node are of special interest here because they interact with the model.

(1) Model traversal node. Model traversal nodes provide a mechanism for identifying components of aparticular type in a particular position. For each component found, the same diagnostic test is performed.Model traversal nodes are essential to the configurability of the PDES. They provide a way of specifyingmodel independent algorithms for shifting the diagnostic focus of attention.

Example 1: Creating an instance of a model traversal node.

(make-instance (gensym m4-) of model-traversal-node (print-string \"check for flooded tank\")

(query-string \"please select a tank for flooding check\") (child-node-spec \"check tank level'')

(traversal-cmd-list supercomp supercomp subcomp) (traversal-class-list SteamSect CondensateSect Tank))

This node, named \"check for flooded tank'', (a) identifies all the tanks in the same condensate section as thecurrent component, and (b) visits each one of these components, invoking the \"check tank level'' node foreach one. Its role is to compile lists of candidate components for further diagnosis. The two lists named\"traversal-command-list'' and \"traversal-class-list'' specify a simple algorithm for moving about the model.The first step (specified by the first element of each list: \"supercomp'' and \"SteamSect'') says \"get all thesteam sections (i.e., steam groups) which contain the current component''. The second command says \"getall the condensate sections which contain (in the sense that they receive condensate from) these steamsections'' and the last command says \"get all the tanks in these condensate sections''. For each one of thetanks which are identified by this algorithm, the \"check tank level'' node is visited.

(2) Variable test node. Variable test nodes provide the basic ability to perform comparisons of values ofsome variable (in a component of the model), and to use the result of the test to decide which node shouldbe visited next.

Example 2: Creating an instance of a variable test node.

(make-instance (gensym m1-) of var-test-node (print-string \"check tank level\") (test-var-name Level)

(children ``low tank level: probably OK''_ ``normal tank level: probably OK'' ``high tank level: check controller'') (reference-behaviour actual-value)

(child-behaviours threshold-value threshold-value threshold-value) (child-expected-results LT EQ GT))

In this example of a variable test node, the actual value of the level variable in the current component(which must be a tank) is compared to its threshold value. The result of this test determines which childnode is visited next.

These examples illustrate several features of the diagnostic mechanism relevant to configurability.(i) The model traversal node provides a general model navigation facility that makes the diagnosticknowledge independent of any particular machine configuration. It is only necessary to build a model ofanother paper machine dryer section (using the same basic set of components) to reconfigure the PDES tothe new machine.

(ii) No system state information is stored in the diagnostic tree. The diagnostic nodes know nothing of thestate of the machine being diagnosed. They make references to particular variables (in the currentcomponent), specify what tests are to be performed, and what to do (i.e., which node to visit next) uponreceipt of the test results. By making sure that the diagnostic tree does not store any system stateinformation, a diagnostic node may be used repeatedly in the course of a single diagnosis (i.e., it may bevisited once for each component of the appropriate type). Furthermore, the risk of redundancy andinconsistency in the knowledge base is reduced significantly. Because the model has a well definedstructure (isomorphic with the system being diagnosed), it is unlikely that components will be inadvertentlyduplicated. Because the tests are all performed by reference to the model, the likelihood of redundant (andhence possibly inconsistent) state information creeping into the system is reduced.

(iii) All numerical information is represented internally as intervals, and all arithmetic is performed onintervals. This provides a simple, uniform mechanism for representing both allowable ranges and pointvalues for variables (e.g., the \"threshold-value'' referred to in example 2 is a range whereas the \"actual-value'' is typically a point value). A test compares two intervals and determines which of \"less than'' \"equal''or \"greater than'' is most likely to be true (treating the intervals as representing uniform random variables).Intervals could also be used to represent error or uncertainty, although it has not been done in the currentsystem. Finally, the PDES has been built so that the value representation, test and arithmetic computationmechanisms may all be easily replaced (and in fact the replacement of intervals with fuzzy sets is beingconsidered).

(iv) Tests are localized. All tests defined on the model must compare two values for the same variable. Inexample 2, the \"level'' variable in each tank component is tested by comparing the actual-value to thethreshold-value. The localization of tests has two effects. First, it eliminates certain types of dependencybetween the diagnostic nodes and the structure of the model. Second, it enforces a certain discipline on theknowledge engineer in the sense that it is impossible to compare quantities with different physicalmeanings; the knowledge engineer is forced to introduce a new variable into the model for any valuesufficiently important as to warrant a diagnostic test.

Besides configurability, the strict partitioning of physical system knowledge from diagnostic reasoningknowledge has other advantages. The model can be viewed as rather unspecialized knowledge about thecurrent and desired states of the dryer section of a paper machine. As such it could be used for a number ofother reasoning tasks. For example it would be a simple matter to replace (or augment) the diagnosticreasoning mechanism (and diagnostic knowledge) with other reasoning mechanisms for such tasks as alarmmanagement or process tuning. These modules could employ completely different mechanisms, sharingonly their common use of the model for storage, retrieval, computation and comparison of physicalquantities.

Integration of the Configuration, Analysis and Diagnosis Components

The Configuration and Analysis components have been built in C++. The Diagnosis component has beenbuilt using CLIPS 6.0. Integration consists of a set of CLIPS function calls for diagnostic modelconstruction and to initiate or continue diagnosis. User interaction during diagnosis is managed by the C++portion of the system. Messages from CLIPS to C++ pass through a message buffer.

Conclusion

The \"Paper Drying Expert System\" (PDES) project is a thirteen-month (elapsed time) and three person-year effort to produce a working prototype of a diagnostic consultant expert system for troubleshootingdryer sections of paper machines. We described the physical process, the special requirement that the

PDES be adaptable to various dryer section configurations by paper mill personnel (who are not knowledgeengineers), and the design methods used to satisfy this requirement.

The original design evolved significantly during the course of the project. Functionally, it changed in scopeand focus. The representation of knowledge and manner of reasoning evolved significantly duringimplementation. Having explored several alternatives, we are satisfied that the resulting design istheoretically sound and practical. In retrospect, the decision to use two development environments (C++and CLIPS) was somewhat costly in terms of integration effort.

The strict partitioning of physical system knowledge from diagnostic reasoning knowledge is applicable tomany other physical processes. Furthermore, the design of the diagnostic reasoning mechanism is such thatit could be readily augmented with other reasoning mechanisms for such tasks as alarm management orprocess tuning.

References

Chittaro, L., Guida, G., Tasso, C. and E. Toppano; Functional and Teleological Knowledge in theMultimodelling Approach for Reasoning about Physical Systems: A case study in Diagnosis; IEEETransactions on Systems, Man and Cybernetics; Vol 23, No 6 (Nov/Dec)

Lee, J.M. (1993); An Integration of Heuristic and Model-based Reasoning in Fault Diagnosis; EngineeringApplications of Artificial Intelligence; Vol 6, No. 4 (pp. 345-356)

Acknowledgement

Since the focus of this paper is on software design considerations to meet the configurability requirementsof the PDES, the authors are all significantly involved in this aspect of the development. The first threeauthors are members of the software development team; the last is the main domain expert who providedsubstantial input on interface design. However, we would like to express our gratitude to the people whowere involved in defining the configurability requirements: Jim Futcher (consultant) and Paul Henzell(KanEng Industries, Ltd.) for their detailed technical contributions as dryer section domain experts, RohanJayatilaka (Stone Consolidated) and John Reinsborough (Abitibi-Price) for their general technical advice aspaper makers, Oliver Vadas (Pulp and Paper Research Institute of Canada) for sharing his experience withthe development and marketing of the \"Pitch Expert\" (an expert system for diagnosing pitch problems inKraft mills), and David Cook (EDS of Canada) for promoting the project and providing advice towardsfuture development and marketability.

pa270594.doc