Toolbox Integration for Instability Prediction at Redcar Blast Furnace,

Teesside Cast Products, Corus UK

www.chem-dss.org

REDCAR (1 Big Blast Furnace)

SCUNTHORPE (3 Medium Blast Furnaces)PORT TALBOT

(2 Medium Blast Furnaces)

Each site produces3.2 to 3.6 million tonnes of liquid iron per year.All sites use G2.

Location of Corus UK Blast Furnaces.

Sinter,oreand coke

Hot blast1100oC4 bar

Top gas2 bar/120oC

21%CO21%CO25% H2

Iron and Slag

1500oC

Steel shellWater cooled

pressure vessel

Carbonbricks

The Blast Furnace Process

Melting Zone

2000'C

14 Rows ofCooling Staves48 – 60 stavesPer row

= Passage of Reducing gas

Melting Zone

The Objective

Predict aerodynamic instability in order to enable the controller to reduce the blast volume in time to reduce the effect.

Effect usually seen as Sudden slip of the material in the furnace,

which can lead to a Surge of gas at higher than normal pressure through the furnace stack, hence lifting the pressure relief valves.

Channelling of gas through the burden which can lead to

high local heat load onto the furnace wall cooling plates. Poorer gas distribution in the furnace hence reduction in process

efficiency.

Blast Furnace process:Combinations of

Toolboxes

iMSPC alone iMSPC with Qualtrend and G2 rules to analyse sequences of

episodes iMSPC with SALSA Qualtrend with SALSA

(iMSPC is a multivariate SPC toolbox by Computas, Norway. G2 Qualtrend is a qualitative trend analysis toolbox by University of

Girona. G2 Generation of objects, known as episodes, from a univariate

signal Salsa is a pattern recognition toolbox by University of Toulouse.

Labviews

Toolboxes communicate using XML Blaster, freeware, using A G2 module as the ‘data store’

Salsa

Classification

Qualtrend

Contribution analysisOn PC1, PC2. Alarm

Salsa

iMSPC

PC1, PC2

PC1, PC2, PC3, SPE

Raw data

QualtrendEpisodes

G2

iMSPC

G2PC1, PC2

G2 rules

Analyse sequences of episodes

Episodes

Classification Data updatedEvery minute

iMSPC iMSPC

Raw data PrincipalComponents

Contribution Analysis

iMSPC Alone

Principal Component Analysis

Data compression without loss of informationSmaller number of new variables generated called

‘Principal Components’i.e., reduce dimensionality of the data

Each principal component is a linear combination of the original normalised variables

Variables Selected for PCA Stability Index

NW Row 6 to Row 9 differential pressure (Quadrant 1)

NE Row 6 to Row 9 differential pressure (Quadrant 2)

SW Row 6 to Row 9 differential pressure (Quadrant 3)

SE Row 6 to Row 9 differential pressure (Quadrant 4)

CO utilisation [100 * CO2/(CO + CO2) in off gas]

Sum of CO + CO2 in off gas

Permeability

These 7 selected after much testing with many other variables

Top gasComposition,

pressure

Wall pressuretappings

Blast pressuretemperature

volume

Row 6 to 9DP over

4 quadrants

Blast Furnace Signals used for PCA Models

Permeability =f(blast pressure,top pressure,blast volume)

Calculation of principal component scores

PC2 factorsPC1 = 0.26 * CO Utilisation -0.33

+ 0.40 * Permeability Resistance -0.26 + 0.063 * (CO + CO2) 0.81

+ 0.47 * Row 6 to 9 DP Quadrant 1 0.06 + 0.45 * Row 6 to 9 DP Quadrant 2 0.13 + 0.45 * Row 6 to 9 DP Quadrant 3 -0.16 + 0.38 * Row 6 to 9 DP Quadrant 4 0.31

Variables must be normalised:

Normalised value = (actual value - mean) / standard deviation

Mean and standard deviation derived from stable period of operation

We use an adaptive mean

Inputs updated every minute

Calculate5 minute moving average

Inputs to model

Link to model

Outputs from model

Outputs to G2 object (to Qualtrend)

iMSPC Model Configuration in G2

iMSPC iMSPC

Raw data PrincipalComponents

Contribution Analysis

iMSPC Alone

iMSPC Contribution Analysis

Contribution Analysis monitors the bi-variate trend of PC1 v PC2 (These 2 PC’s represent 70% of the variability in the data) Identifies which variables have contributed

the most to the change in principal component.

Alarm if 6/7 points outside action limit and significant change in at least 1 quadrant for 6-9 Differential Pressure.

Blast Furnace Wall Pressure trends

0

0.2

0.4

0.6

0.8

1

09:00 09:30 10:00 10:30 11:00 11:30 12:00 12:30

Q. 1

Q. 2

Q. 3

Q. 4

Row 6 - 9 DP17 aug 03 09:00

Warning message

Bi-variate score plot when in control

Yellow region outside warning limitPink outside action limit

Contribution Analysis: 6/7 points outside Action limit

12:50 Blast volume reduced for poor permeability13:30 1.5m slip14:10 2m slip

iMSPC Qualtrend

Raw data PC1, PC2

Sequence of episodes analysed in G2 procedure

Episodes

iMPSC with Qualtrend and G2 rules

Data entry (PC1) Range check

Filter

Configureattributes to be stored in episodesand hold currentvalues

Calculate 1st derivative

Signal block (level = normal/low)

Attributes of current episode. List of past episodes

Filtered signal

First derivative Limits

Episode Types:Type Level First derivative7 Normal Normal8 Normal Low16 Low Low31 Low Normal7 6 16

31

Qualtrend: development of rules

22 * 24 hour periods of 1 minute data supplied to UDG from Jan 2002 to Oct 2003.

PC1 and PC2 Episodes generated in Qualtrend.Sequences of episodes analysed.Possible rules tested in Matlab.

Successful rules programmed into G2 and run on line at Redcar since October 2003.Within the same G2 as iMSPC. (The live plant G2).

G2 Rules

Rule 1 looks for a sequence of episode types from PC1. Criteria set for minimum rate of change (slope)

and degree of change (amplitude).

Another rule looks for a similar sequence of episodes from PC2, and generates an alarm if the most recent episode from PC1 satisfies certain conditions. Effectively, this detects a sequence of events in

the process. To prevent false alarms, an ‘enabler’ has been

added based on the recent trend in heat flux.

2m Slip at 09:40. 40 minutes warning.

Current episode = 31 and Max-min of previous episode > 2.2And min slope < -0.0015

Episode Types:Type Level First derivative6 Normal Low31 Low Normal

Filtered PC1

First derivative

2m Slip at 09:40. 35 minutes warning. Confirms previous message

PC2 Current episode = 31 and previous episode = 6Min slope of last episode of PC1 < -0.0015And finished within 10 minutes

Filtered PC2

First derivative

Episode Types:Type Level First derivative6 Normal Low31 Low Normal

Summary Statistics

Event

Type

Number of

events

Predicted by

iMPSC alone

Predicted by iMSPC/Qualtrend

PC1

Predicted by iMSPC/

Qualtrend PC1/PC2

Notpredict

ed

Major

19 8 13 6 0

Minor

10 1 1 2 7

Events detected over 22 days Jan 2002 – Oct 2003.Classed as predicted if more than 10 minutes warning.Major event: Slip >=1m and/or excessive heat flux.Minor event: Smaller slip and/or significant rise in heat flux.Sometimes alarms also generated during event (high heat flux).

Conclusion

All of the major events were predicted (19/19)

Only 3/10 of the minor events were predicted. However, it is unlikely that action would

have been taken for minor events.

iMSPC SALSA

Raw data

PC1, PC2, PC3, SPE Classification to

Normal,Pre-slip orSlip

G2 Windows

iMSPC with Salsa

iMSPC and SALSA

Same data as used in for iMSPC/Qualtrend/G2 rules (PC1 – PC4, SPE and T2 for 22 * 24 hour

periods)

Best classification gained with PC1, PC2, PC3 and SPE

However, too many false alarms

Qualtrend Salsa

Raw data(4 * differentialPressures)

EpisodesClassification toNormal,Pre-slip or Slip

G2 Windows

Raw data with Qualtrend and Salsa

Raw data with Qualtrend and SALSA

Classification based on data from early 2002. Classification based on

Quantitative values (values at end of previous episode) Qualitative values (current episode types) So 8 inputs (4 differential pressure signals: 4 sets of

episodes)Can give more advanced warning than other methods

described. e.g., 4 Jan 2002. 30 mins before iMSPC/Qualtrend.

IssuesSALSA on-line reliability – stalls after a day.Need to write classifications back from SALSA to DTM.

Row 6-9 Differential Pressure. 3-4 Jan 2002

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00 23:00 0:00 1:00 2:00 3:00 4:00 5:00 6:00 7:00 8:00 9:00 10:00 11:00

QualtrendAlarms

= 2 metre Slips

SALSA alarm

Did not exceed action limits for 6 minutes, so no alarm

iMSPC alone on 4th Jan 2002.

Qualtrend

Salsa

Episodes

Classification

G2

Qualtrend

iMSPC

G2PC1, PC2

G2 rules

Analyse sequences of episodes

Contribution analysisOn PC1, PC2. Alarm

Salsa

iMSPC

PC1, PC2

PC1, PC2, PC3, SPE

Raw data

Episodes

Classification

Blast Furnace processSummary of Results

1. iMSPC alone All alarms generated by action limits are valid Many events are missed

2. iMSPC with Qualtrend and G2 rules to analyse sequences of episodes

Predicted remaining major events and very few false alarms once heat flux trend ‘enabler’ added

1 and 2 predicted all the major events.

3. iMSPC with SALSA Many false alarms

4. Qualtrend with SALSA Predicts certain types of faults with good warning Salsa not robust enough for continuous on line Salsa needs to send classifications back to DTM