using both numerical and symbolic models to create

USING BOTH NUMERICAL AND SYMBOLIC MODELS TO CREATE ECONOMIC VALUE : THE SACHEM SYSTEM EXAMPLE

1)

2)

Marc Le Goc , Claude Thirion ,

USINOR, Scientific and Advanced Computing Division LB1 SOLLAC-FOS 13776 Fos sur Mer Cedex, France 1)

Phone: +33 (0)4 42 47 29 82 Fax : +33 (0)4 42 47 22 32 E.mail: [email protected]

Nowadays, the Sachem system is known as one of the major achievements in Europe in the field of industrial knowledge based systems. The interviewing of thirteen experts in blast furnace supervision resulted, for example, in six blast furnaces in three plants being equipped with the Sachem software, and in over four hundred thousand program lines. This paper first describes the Sachem project and the industrial results. Theses results are then explained thought a discussion of the advantages we gained by merging numerical and symbolic models. Some aspects of the future of this work are surveying to conclude this paper.

2. WHAT SYSTEM?

IS

THE

SACHEM

2.1. Objectives and opportunities The main operational objectives of Sachem system are:

2)

Phone: +33 (0)4 42 47 33 82 Fax : +33 (0)4 42 47 22 32 E.mail: [email protected]

SUMMARY Since the past twenty years, the technology of knowledge engineering and Artificial Intelligence has shown how to transform an “empirical knowledge” in an operational model so that a knowledge based system can be designed to help the controllers in their work. These technologies propose to substitute the numbers, the paradigm usually used to represent the knowledge, with symbols. This substitution leads to the concept of knowledge based system, a new generation of experts systems able to manage the behavioural models used by the Operators to control a process. The Sachem system belongs to this new generation of systems. Sollac, the French steel manufacturing company, has developed it. In this paper, we propose to illustrate thanks to the Sachem system example how merging numerical and symbolic models leads to create economic value.

§

to assure the continuity of operations despite the change of operators (eight hours shift) and the limited capacity or performance for the human observation. This problem is getting more high and more complex, day after day, in relation with the computerisation development ;

§

to improve the regularity of operation by the way of integration of experts’ knowledge, and especially in the events early perception,

§

to prolong the life of the blast Furnace, resulting in a better and smoother operation,

§

to capitalise knowledge in order to better study past operations and to better train operators. SACHEM Numerica Data

Detection Detection diagnosis diagnosis System System BF State

Symbolic Data (BF State)

Recommendatio Recommendatio Action Action System System Preconise Actions

Actions

Fig. 1. The «closed» loop

1. INTRODUCTION The Sachem system is very large-scale knowledge-based system designed to help the Operators in the control room to control a blast furnace. It is one of the results of the Sachem project started in 1990 by Sollac, the French steel manufacturing company of flat products of the Usinor Group.

Consequently, the SACHEM system is intended to assist the return to optimum operating conditions when slight deviations or minor operating incidents occur. In the first case, it acts as an aid in the optimisation of the settings, in the second, as an aid for the diagnosis and resolution of minor incidents (Action Recommendation System, refer to Figure 1). The diagnosis and resolution of major incidents are out of its scope. In these cases, which are fortunately very rare,

Proceedings of the 27th McMaster Symposium on Iron and Steelmaking, pp. 206-218, Hamilton, Ontario, Canada, May 25-27, 1999.

all specialists are requested to resolve the problems and each situation is different from another. In order to achieve this, the system devised uses all the IRSID scientific knowledge, combined with the technical knowledge and experience of the specialist engineers in the Usinor Group. Finally, the entire field of blast furnace operation technology will be covered by the expertise integrated into the knowledge databases, excepted the hearth wearing. Another important point is the aim to design and to develop a generic system, which would be independent of the place of introducing. The target was to equip with acceptable economical conditions all the furnaces in the Usinor Group, in a short time, and to supply an evolving system in the future. The intended benefit is around $1 (around 1 Euro) per tonne of metal, by implementation of an improvement in the analytic dispersion of smelting, associated amongst other things with a 50% reduction in operating incidents detected late, and an improvement in the operating life of blast furnaces. Alongside this control system as such, a computerised system for monitoring the operation of the blast furnace, the ENQUET software environment, connected into the previous system, has been jointly developed at the attention of the technical sections teams. Finally, these two systems are supported by a softwareengineering department, responsible for providing management between versions, updating of the knowledge databases, configuration of the system to another blast furnace, etc. This environment is currently in use for the development phase of the first systems to be introduced on the Sollac sites in Dunkirk and Fos sur Mer, and in Lorraine

In co-operation with the operators, by the producing of relevant information helpful in carrying out the global mission, and by providing a concrete and efficient summary of information needed by the operations teammate. The amount of effort devoted to this subject represents approximately one third of the total project resources. The sustained presence of one, or even two ergonomics specialists within the team contributed to the management of the relationship between the project team and the future operators : the degree of understanding increased, and despite the length of the project, the future operators did keep in contact. A reliable service. When the information environment deteriorates, the system continues his reasoning, providing that the lack of information is not critical, and, through appropriate reasoning, gives reliable conclusions. In the same way, the system, along the “pseudo-pipe” is processing different types of mechanisms to check continuously the quality of the data. If the data are wrong, ore suspect, the system executes the appropriate treatments to give finally the best service to the operators.

2.3. Architectural principles To produce the above mentioned services, the Sachem system is integrated in the control room environment, connected directly to the control command system of the Blast Furnace. The figure 2 shows the pipeline architecture principle, developed in order to satisfy the functional requirements and to describe the current status of the process, for the operators in the control room. It presents a summary, in abstract form, of the successive data conversions, from elementary status to synthetic status. (Processing of 10 000 data items per minute). 1,000 data/minute

3,350 variables

2.2. Technical capabilities

70 msg / day Monitoring and operating

Yet today, the economic performance of the Blast Furnace is depending on improving the technical results. Uniformity is one of the key factors that contribute to improve product quality and the life of all the related equipment.

Acquiring data

Continuously. The system is indifferent to the shift changes, and analyses problems in the process, with the same seriousness ant the same strictness. It turns 24 hours per day as the BF itself on (Availability ratio = 99,7%).

Data processing + mathematical models processing

Systematic synchronous assistance Signals analysis

Detecting the Phenomena, managing the context

Warnings

Action Data

Actions

processing recommendation

Knowledge

Bases

Asynchronous assistance on request

Justifications

Data visualization

The Sachem system is designed to satisfy this major need, and have processing and analysis capabilities allowing it to act: In real time. The system uses all available expertise to detect earlier any anomaly that occurs. The covered domain is the furnace process, and its state is determined after reasoning on the data and on the time. Like an expert, the system is self-adaptive to the changing situations. In each part of the entire domain, the system describes all present anomalies, and potential anomalies or incidents that could occur in the short term. The consistency of the deductions is evaluated, as the magnitude of the problems.

Alarms

Level 3 computer

DB

Fig. 2. Architecture principle Following this model, the system contains the main following functions §

Data acquisition, synchronisation and verification

§

Data ordering, processing and validating (using physical and chemical models developed by IRSID),

§

Signal analysis, including neural networks [3], detection of phenomena occurring during operation and alarm generation,

§

Interpretation of the current situation , by reasoning using the knowledge bases,


§

Recommendation of correctives actions for thermal control of the BF,

same time, it is possible to see the relative importance of the metallurgical part in each component.

§

User friendly and context dependent Man Machine Interface.

The items of knowledge incorporated so far represent more than 150 000 lines of coding distributed between 21 Expert Systems working together simultaneously.

The dynamic architecture is intended to be general and simple and also testable. The main reason is that the development process is iterative and incremental, because of the incompleteness of the knowledge specification at the beginning of the development. The main features of the architecture are:

2.5. A generic system.

-

massive reasoning instead of focusing,

In relation with the main target, to produce the same system for all the blast furnace in the Group, it was necessary to intend to develop a generic system, which is independent of the blast furnace technologies.

-

adaptive reasoning. The perception part (signal analysis) is dynamically driven by the knowledge based system part.

One of the faculties which characterises an expert is the ability to work with a different type of equipment which he is used to work with.

-

saturation of the reasoning.

This therefore assumes that there is such a thing as «generic expertise». On this basis, the Sachem project undertook to produce a generic expression for coded expertise in the knowledge databases. The current 100% rate of success for the generic reasoning confirms the validity of this approach.

The system installation is quite easy, and requests only two main factors, as shown at the figure 3. Fos sur Mer BF plant Sachem system implantation

to the plant

2.6. Uncertainty, but reliability.

Sachem Sachem Network Network

others

Level 3 BF studies station

Qualification station

Ironmaking Ironmaking Dpt Dpt Building

SP2

Computer

Control room

BF control command

Fig. 3. Connecting the Sachem system. A data collecting system, to collect all necessary data at the appropriate frequency of scanning. This task can be successful by the process computer, at the third level, a specific connection between the level three computer and the Sachem system network. General General architecture architecture % of total code Situation report

3%

Numeric / Symbolic Data Base

Monitor

4%

The factors of uncertainty are examined from the point of view of the operationally. The expert knowledge places the operational decision at a higher level than the measurement precision. The expert choice for action is depending the risk and its associated cost. The result of the implementation is the fact that the system is producing «sure diagnosis». This property is also very important to perform the adaptation of the Sachem system on a «new blast furnace». To achieve easily the adaptation, and the corresponding robustness, the Knowledge Base includes the reasoning with degraded information.

2.7. The Man Machine Interface. From the customer’s point of view, the main objectives were achieved:

10%

- SACHEM speaks the language of the operator, Data Data Acquisition Acquisition

Data Data processing processing &&Modelling Modelling

5%

15%

IRSID models

Detecting Detecting the thephenomena phenomena Managing Managingthe thecontext context

Action Action recommandation recommandation

32%

Dk + Fos generic BF operating knowledge

Human Human Computer Computer Interface Interface

9%

21%

- Some problems are now followed up which were not before.

Dk + Fos generic collegiate BF operation

Knowledge Knowledgeimplementation implementation : :33 33%%of of403 403000 000of of written writtenlines lines of ofcode code

Fig. 4. BF knowledge distribution

2.4. Code characteristics. The total number of coded lines is around 500 000 lines. The figure 4 gives an idea of the distribution in the different components of the Sachem architecture. In the


The Usinor Group is running now 6 BF in France, in three different industrial sites. At the present time, the N°1 and N° 2 BF in Fos sur Mer, the large N°4 BF and the smaller N°2 in Dunkirk, the P3 and P6 in Lorraine are equipped with the Sachem system.

Fig. 5. The first page. Days and night, along years, Sachem always observes the behaviour of the blast furnace. So that now, Sachem is considered as an “artificial colleague” whose mission is to prevent the operators when something goes wrong. To resume the quality of the interaction, we indicate that only a few hours are necessary to train the operator to use the Sachem Human Interface.

Fig. 7. the Sachem systems implantation Now, the Sachem service monitoring is installed on a cumulative yearly production of 10,5 millions tonnes. The figure 7 shows the geographic distribution of the six Sachem instances. It is to note that nowadays, it is possible to install a version of Sachem in all the plants the same day. This result demonstrates the reliability of the software, in spite of its large size.

3.2. Efficiency. After one year in operation, more 8000 thousand phenomenon have been analysed by the BF Experts. 105 % 100 95

Fig. 6. An example of a phenomenon justification Of course, the training duration to obtain a total benefit in the Man Machine co-operation is rather near one year. It’s the mostly necessary time period to have a good maturity of the using, but also aid mainly an improvement of the operator level in the BF operation understanding and monitoring.

Target : 92 % 90 85

Pertinently detected Well dated Well qualified

80 75 1996

1997

1998

Fig. 8. Weekly operational results

3. INDUSTRIAL RESULTS 3.1. The Sachem instances.

The conclusion of this analyse is that Sachem is a better operator than the Sollac best operator. This result is over our target (see figure 8).

The first operational version of Sachem was connected in October 1996, on the BF1 in Fos sur Mer.


3.3. The BF operation quality. The figure 9 compares the operational results of the two BF of the Fos sur Mer plant. The first one is equipped with the Sachem system. The second is not. BF shut down time equivalent (hours per month)

Fig. 11. Improvement of the standard deviation for hot metal silicon content. The Standard deviation decreases by 0,04 when the system is connected.

3.5. ROI. With Sachem

8 6 4

BF1

2

After analysis by the BF people, it seems that the advantages of the system are following by different axes, such as : -

the incidents diminution,

-

the equipment availability increasing,

-

the regularity of the hot metal quality,

-

the protection of the equipment, due to less solicitation.

BF2

0 Jan. to Oct. 96

Oct. 96 to Apr. 97

Fig. 9. BF operation quality This comparison indicates a very good evolution for the BF1, with Sachem system, compared with the BF2 without Sachem. The figure 10 shows the favourable incident number decreasing with the new system: the number of incidents is divided by a factor 3; the tapping problems due to too low temperature are strongly minimised.

Mean shutdown time equivalent, per month

Low Hot Metal T°

70 60

Thermal losses

50

BF1 Incidents frequency Low Blast > 1 h

The total of the measured return of investment is around 1,7 Euros per tonne of hot metal. The table 1 gives a recapitulation of the results determined in the case of Fos sur Mer Blast Furnace Plant.

40

Burden descent

30

TOTAL / Month

20 10

0

0 1995

8 months 96

Without Sachem System

19 hours

With the Sachem System

14 hours

Yearly Production capacity saving

Gain on the Gain on the Gain on the Miscellaneous PCI Iron Quality BF life (Coke rate, duration indirect returns, etc.)

∆ Sigma [Si] = 0,03

30 000 THM, or 2,5 days

Direct valorisation, in Euros

0,15

Future target

40 000 THM

0,8

0,11

0,30

0,30

Sigma [Si] = 0,12

8 months 96 / 97

Tab. 1: ROI in Fos BF plant.

120 100

With Sachem

80

BF1 Incidents frequency Low Blast > 1/2 h

60 40 20

0

0 1995

Fig. 10.

8 months 96

8 months 96 / 97

BF operation quality

3.4. Thermal control.

§

Empirical knowledge (know how). This knowledge comes from the best Experts in blast furnace control of the Sollac Company. This kind of knowledge is elaborated along times during the human learning process. This knowledge is the core of the Sachem system.

§

Scientific knowledge (“know why”). This kind of knowledge comes from the Physics theories. Usually, such knowledge is represented as numerical models that are compiled in computer algorithm. The scientific knowledge produces an explanation and a justification of the empirical knowledge. It is then the theoretical basis for the Sachem system.

§

Ergonomic knowledge (“know how to be”). Precisely, this kind of knowledge is concerned with the culture of the Steelmakers people, their language, their

Before SACHEM After SACHEM

50 00 52 50 55 00 57 50 60 00 62 50 65 00 67 50 70 00

Silicon Standard Deviation *100

4.1. Classes of knowledge. Sachem merges 3 classes of knowledge:

The figure 11 shows the consequences of the utilisation of the Sachem system on the evolution of the silicon content in the hot metal. 18 16 14 12 10 8 6 4 2 0

4. MERGING NUMERICAL AND SYMBOLIC MODELS.

Daily production in THM


custom and so on. Such knowledge is fundamental to construct a Man-Machine Interface well suited for the operators that control the blast furnace. The ergonomic knowledge is the key factor to facilitate the interpretation of the outputs of the Sachem system: the users are not disturbed with the “mechanism” of the Sachem system but with the process problems which are the core of their job.

It is to note that the perception function includes what is usually called the “diagnostic function”. The perception function is in fact a more general cognitive function since the ability to “see” a particular object on a picture or a scene requires understanding the whole picture (or scene). In other word, the perception function is an intelligent activity which mixes both sub-symbolic (unconscious reasoning) and symbolic (conscious reasoning) knowledge.

These tree classes of knowledge are merging in a tool with the help of technologies as artificial intelligence (including artificial neural networks), signals processing and pattern recognition techniques, software development and knowledge acquisition and representation.

The input of the perception function is a set of around 500 variables whose values are computed by a set of scientific models like the MMHF one. Theses models are implemented in the second software process of Sachem (see figure 2).

To this end, a general model of the knowledge has been built to constitute a specification of an artificial agent, Sachem, as a component of the production team. The mission of this artificial agent is to analyse, at every time and always, the data flow coming from the thousand of sensors around a blast furnace with the aim to alarm the Operators when some thing goes wrong.

4.2. A closed loop approach. Functionally, this artificial agent is designed in a closedloop approach. Like all the automatic controllers, Sachem performs two basic functions (fig. 12): Desired behaviour Yd(t)

Process State X(t)

Perception

Fig. 12. 1.

2.

Action U(t) Control

Output Y(t) Process

The two basic functions of a control loop.

PERCEPTION. The aim of this function is to establish the state of the blast furnace. This state is the set of all the phenomena working inside the blast furnace. The presence of these phenomena is revealed with the evolution of the spatial and temporal signals constructed from the sensor data’s. When necessary, that is to say when a phenomenon is worrying in the current exploitation context of the blast furnace, the Operators are informed with an alarm of each detected phenomenon. So the state of a blast furnace is a set of short sentences, synthetically formulated in a specialisation of the professional language usually used by the operators: the phenomenological language of Sachem. Each sentences said what, where and when a phenomenon occurs (see figure 6 and 7). CONTROL. When a set of detected phenomena will leads to a problem with the production program (the quantity and/or the quality of the ping iron), this second basic function estimate the nature and the associated amount of correctives actions. Such an action is evaluated in a closed loop approach. So, if the operators engage the recommended action, the blast furnace would not have problems. At the present times, this function is working for the control of the temperature of the hot metal.

4.3. Symbolic models. The symbolic state is then elaborate through a recognition process. This process aim to map between observed evolutions of the variables of the process and characteristics evolutions of particular phenomenon. These latter evolutions are specified by control process Experts. Such specification constitutes symbolic representations, or models, of the evolution to be recognised by the perception function. The knowledge acquisition process concerned with the perception function aim to capture what an expert see on a set of signals when a phenomenon occurs. The formalisation, that is the construction of a symbolic model, of this knowledge is a necessity not only to specify the perception function but also to validate and to share this knowledge between Experts, Operators and knowledge engineers. Actually, Sachem is able to recognise around 150 class of phenomena, each class being specified with a symbolic model. This set of model covers the whole BF, excepted the hearth wearing. The KBS paradigm permits to use such symbolic models to design operational systems that implement human like functions. Within this approach, the numerical models aims to produce the values of indicators that put on the light the effects on the signals of the presence of phenomena. The decisions process that lead to the recognition of a phenomenon is based on the ability to give a semantic interpretation to specific evolutions of such indicators, temporal and/or spatial one.

4.4. Creating economic value. The interest is then a strong data reduction or compression: only few words suffer to formulate what is happening. And knowing what is happening facilitate knowing what to do. So, symbolic model permits to automate the interpretation of the results produced with numerical models in terms of physical phenomena. The advantage comes from the production of a synthetic and neutral description of the state of a process.


Such a description creates economic value because the decision making process is no more based on the question “what happen?” but shift to the question “what is to do?”. Because answering to this latter question is simpler than answering to the precedent, and especially when the process is as complex as a BF, the probability of making a mistake is dramatically decreasing. This property explains the financial results gained at the Sollac Company with the Sachem system.

5. CONCLUSION. 5.1. A technological lead. Of course, we were all the time proceeding benchmarking. Some conclusions of this activity show that the project has enabled significant progress to be achieved in the following directions: • Collection, Formalisation and Modelisation of expert knowledge with the KADS1 method [1], [2]. The project team in relation with the real needs, and with the largeness of the scope has adapted this original method. • Multiple Expert and Multiple Site Management approach, yielding important results in the fields of genericity (specifications and corresponding software) and therefore portability and adaptability of the application to other industrial sites, and to other hardware environments operating on the UNIX system. • Development of real times perception functions, applying Signal Processing, Pattern Analysis, artificial neural networks technologies [3] and Expert Adaptation relative to the blast furnace operating environment. • Development of new functions in the field of expert system generation (KOOL-94), in conjunction with databases handling large quantities of object related concepts.

• ENQUET for the on Line data examination. It is interesting in this case to note the original function of geographical access to data. • ENQUET to parameterise and to adjust the perception algorithms. This tool authorises to develop, for a low cost, a detection system using mainly continuous data, like curves.

5.2. Others applications. To re-use the methods and the tools, to recover the maximum of the R & D effort spent in the Sachem project, some other projects are engaged within the Usinor Group. The field of them concerns different ways like: -

other metallurgical processes (continuous casting, galvanisation, …)

-

ship loading planning,

-

distributed knowledge bases,

-

intelligent office for design.

Because it is not the job of the metallurgists to edit such a solution, we have taken some contacts with partners, specialists of the domain, to establish agreement to promote such technologies, and to prolong our Knowledge Engineering and our experiences.

5.3. Economic value. One of the fundamental reasons of the difficulty to control a blast furnace comes from the fact that it is not possible to “see” inside a blast furnace. Only the effects of the physical phenomena transforming the ore in pig iron can be perceived through the thousand of sensors equipping the periphery of the blast furnace. The controllers are then constrained to induce the phenomena, the causes, from theirs effects observed on the signals the sensors are delivering.

• Ergonomic approach, from the analysis of the requirements up to final testing, in the development of the Human Machine Interfaces (intermediate model, control room simulation...).

A collection of such causal relations between effects and causes constitutes a behavioural model of a process. It has not been yet possible to construct a numerical behavioural model of a blast furnace because the phenomenological of the blast furnace process is so complex that there is no scientific theory able to predict the blast furnace behaviour.

• ENQUET for the operation data studies. It is the tool for the technical people they are studying a lot of data and curves to understand the process. It offers the possibility to edit symbolic stamps of qualification to facilitate the interpretation of data. Signal processing functions are available for the phenomena detection and interpretation.

The technology of knowledge engineering and Artificial Intelligence propose to merge numerical models and symbolic one to construct knowledge based systems. Such KBS can be designed to help the controllers of complex systems in their work. This leads to a new generation of experts systems able to manage the behavioural models used by the Operators to control a process. The Sachem system belongs to this new generation of systems.

•

1

Explicit treatment of the uncertainty.

European studies conducted in the context of an ESPRIT project.

This paper aim to show that this approach truly creates value. The example of the Sachem system is used to this end where a real improvement of the industrial technical results has been achieved, especially in the field of the availability and the reliability of the Blast Furnace.


But it is also to note that indirect spin-offs have been observed: even before any technical achievements, the attitude of the Operators has evolved the expertise of each individual having been enhanced by being exposed to the expertise of others. Since May 1998, the Sachem project team is working as a permanent department. The main object of this new department is, of course to support the existing Sachem systems in place, but also to implement similar technology on other process lines, inside and outside of the Usinor Group.

F

[6] JM Libralesso, JM Steiler, JL Lebonvallet, CC Thirion, “The Blast Furnace Operation under High Supervision” Proceedings of the AIME Congress, in Toronto (Canada), in March 1998.

F

[7] C. Thirion, J.-M. Steiler, J.-L. Lebonvallet, B. Metz, D. Lao and J.-C. Lachat, “The Sachem Experience with BF Control”, METEC Congress 99, Proceedings of the International Conference on New Developments in Metallurgical Process Technology, Düsseldorf, June 13-15, 1999, pp 91-97.

Acknowledgements The Sachem project has been carried out with a financial support of the ECSC.

References F

[1] B.J. Wielinga A. Th. Schreiber and J.A. Breuker, "KADS: a modelling approach to knowledge engineering", Knowledge Acquisition, vol. 4, p5-53, 1992.

F

[2] C. Frydamn, L. Torres, N. Giambiasi and M. Le Goc, "Introduction of processing mechanisms in KnowledgeBased Conceptual Models", FLAIRS-96, Key West Florida; USA, 1996.

F

[3] M. Le Goc, C. Touzet, C.-C. Thirion «The Sachem Experience on ANN Application» Invited Paper at Neurap’98, Fourth International Conference on Neural Networks and their Applications, Marseille, France, 11-13 Mai 1998.

F

[4] N. Dolenc, A. Gobrecht, M. Helleisen, M. Lallier, D. Lemuet, F.M. Lesaffre, J.M. Libralesso, J.M. Steiler and C. Thirion, « The Sachem Project : Computer Assisted Blast Furnace Control System - Prospects and Opportunities, Development and Initial Results », Proceedings of the 3rd European Ironmaking Congress, Gent, Belgium, September 16-18, 1996, pp 120-127

F

[5] N. Dolenc, A. Gobrecht, J.M. Libralesso, M. Helleisen, M. Lallier, J.M. Steiler, D. Lemuet, F.M. Lesaffre and C. Thirion, « The Sachem Project : a Computer Assisted Control System - Prospects and Opportunities, Development and Initial Results », Proceedings of the 6th International Conference on Man-Machine Interaction and Intelligent Systems in Business - Montpellier, France, May 28-30, 1997, pp 255-259.
