Presentation is loading. Please wait.

Presentation is loading. Please wait.

© Marc LE GOC - LSIS - DLS03 Dec 3-4 - Tucson1 Extensive Large Scale Real Time Knowledge Based Systems Design: the Sachem Example Marc LE GOC LSIS, Laboratory.

Similar presentations


Presentation on theme: "© Marc LE GOC - LSIS - DLS03 Dec 3-4 - Tucson1 Extensive Large Scale Real Time Knowledge Based Systems Design: the Sachem Example Marc LE GOC LSIS, Laboratory."— Presentation transcript:

1 © Marc LE GOC - LSIS - DLS03 Dec Tucson1 Extensive Large Scale Real Time Knowledge Based Systems Design: the Sachem Example Marc LE GOC LSIS, Laboratory of Sciences of Information and System Marseille, France

2 © Marc LE GOC - LSIS - DLS03 Dec Tucson2 Content Sachem and the Blast Furnace Sachem Design Current Works

3 © Marc LE GOC - LSIS - DLS03 Dec Tucson3 Arcelor is the world s first steel producer 45 millions tons per year, > employees, $33 billions of turnover. Estimates 43,5 28,6 20,2 19,3 19,1 17,7 16, ,3 27,1 Arcelor (Luxembourg) Posco (South Korea) Nippon Steel (Japan) NKK (Japan) LMN (Ispat London) Shanghai Baosteel (China) Corus (United Kingdom) Thyssen Krupp Steel (Germany) Riva (Italy) Kawasaki (Japan)

4 © Marc LE GOC - LSIS - DLS03 Dec Tucson MW for6,000 tons of hot metal / day ~1200 sensors Extreme conditions (high T°, dust, acidity, etc) No mathematical model of the dynamic Long period for learning (5-10 years)

5 © Marc LE GOC - LSIS - DLS03 Dec Tucson5 Sachem goal: Optimize the BF operation in order to save up 1/ton of hot metal (i.e. 1% of the cost price) MW for6,000 tons of hot metal / day ~1200 sensors Extreme conditions (high T°, dust, acidity, etc) No mathematical model of the dynamic Long period for learning (5-10 years)

6 © Marc LE GOC - LSIS - DLS03 Dec Tucson6 Sachem optimises the BF by reducing the deviation of its parameters Instrumentation computer ~5500 Variables State Correction Event DB Data Acquisition & Model Processing State Perception & Diagnosis Operator Communication Monitoring 1,100 data/mn Process DB Data Base Management ~150 Phenomenon / day ~10 Actions / day ~70 Msg / day Alarms/Action Justifications

7 © Marc LE GOC - LSIS - DLS03 Dec Tucson7 Sachem represents ~ lines of « C » equivalent code State Correction 10% Data Acquisition & Model Processing 20% State Perception & Diagnosis 33% Operator Communication 20% Monitoring 7% Data Base Management 10% 21 Knowledge Bases 1060 Object Classes 1100 First Order Rules 140 Chronicles Development: ~30 engineers from 1991 to man.year (~30 m) Knowledge acquisition & modeling: 12 Experts, ~6 knowledge engineers 14 man.year over 3 years

8 © Marc LE GOC - LSIS - DLS03 Dec Tucson8 With 6 Sachem in operation, the Arcelor Group earns ~18,5 m per year Frequency of Incidents Reference Level Expected level Without Sachem With Sachem Low Hot Metal T° Thermal Losses Burden Descent TOTAL TargetResult Sachem save up ~1,7 per ton of hot metal

9 © Marc LE GOC - LSIS - DLS03 Dec Tucson9 Extensive Large Scale Real Time Knowledge Based Systems Design: the Sachem Example Sachem and the Blast Furnace Sachem Design Current Works

10 © Marc LE GOC - LSIS - DLS03 Dec Tucson10 Sachem is designed a as recursive abstraction process of discrete events Process State Signals Quantization Signal Events Discrete trajectories Interpretation Signal Phenomena Detection Process Phenomena Logical Sensors State Perception Process Behavior Analysis Process Problems & Causes Recommend Recommended actions & warnings Process State Logical Actuators Diagnose Correct Process Phenomena Recognition

11 © Marc LE GOC - LSIS - DLS03 Dec Tucson11 Requirements definition Software Specification Technical Design Implementation CommonKads Model Transformation Process, within a spiral development Sachem Requirement Specification Model Design Model Data Model Function Model Behavior Model Knowledge Analysis & Modeling Knowledge Model Concepts Model Inferences Model Tasks Models SACHEM 25,000 objects: 3200 concepts, 2000 relations 75 inference structures 33 goals, 27 tasks

12 © Marc LE GOC - LSIS - DLS03 Dec Tucson12 Requirements definition Software Specification Technical Design Implementation Sachem Template Knowledge Model for Monitoring and Diagnosing Sachem Requirement Specification Model Design Model Data Model Function Model Behavior Model Knowledge Analysis & Modeling Conceptual Generic Sachem Knowledge Model Concepts Model Inferences Model Tasks Models Validated in collaboration with the LSIS by simulation (using Petri Nets formalism) SACHEM

13 © Marc LE GOC - LSIS - DLS03 Dec Tucson13 Sachem Framework for an Adaptative Generic Monitoring Cognitive Agent Generic Sachem Monitoring and Diagnosing Method Kernel Compilation Implemented Sachem Template Knowledge Model Generic Blast Furnace Generic Process Knowledge Knowledge Bases Configuration … Generic Process#2 Fos#1 Specific Process Instrumentation Structure Functions Parameterization 150 man.year MT#1 ~5 man.year Fos#2 Duk#4 ~1 man.year Pat#6 … … ~0.5 man.year

14 © Marc LE GOC - LSIS - DLS03 Dec Tucson14 Extensive Large Scale Real Time Knowledge Based Systems Design: the Sachem Example Sachem and the Blast Furnace Sachem Design Current Works

15 © Marc LE GOC - LSIS - DLS03 Dec Tucson15 One of the advantages of discrete event representation is compactness octets of the Sachem Process Data Base are required to produce one octet in the Sachem Events Data Base One year of blast furnace behavior is ~30,000 events Events Data Base = 4.7Mb Sachem Digital Data base = 235,000 Mb 15 years of blast furnace (450,000 events) is only ~70Mb ! To be compared with 15*235,000Mb = 3,525Gb ! Sachem Process Data Base(~5500 variables) 500 < Factor < 1000 Events Log 50 < Factor < 100 Compactness Factor = ~ Process Data (~1100 variables)

16 © Marc LE GOC - LSIS - DLS03 Dec Tucson16 The « Elp Laboratory » aims at learning from an event flow Events Expert Knowledge Engineer Signatures ELP Lab SACHEM A signature show the way a type of event is generated New Knowledge to improve the perception of the process state

17 © Marc LE GOC - LSIS - DLS03 Dec Tucson17 Events Data base Signatures ELP Models Editor Expert ELP Models (DEVS formalism) Model Recognition (DEVS Simulator) Event Flow Analysis (Markov Theory) ELP Lab Sequences of Events A signature show the way an event type is produced by a couple (Process, MCA) Discrete Event Models


Download ppt "© Marc LE GOC - LSIS - DLS03 Dec 3-4 - Tucson1 Extensive Large Scale Real Time Knowledge Based Systems Design: the Sachem Example Marc LE GOC LSIS, Laboratory."

Similar presentations


Ads by Google