Virtual STFI Sensor

1
Virtual STFI Sensor
Willard Reed, Process Control Supervisor
Weyerhaeuser Company
2
VIRTUAL STFI SENSOR

• Outline
• Objective Predict Strength
• Background
• STFI -- What is it?
• No.1 Machine process
• Experience with On-Line Strength Measurement
• Neural Net model
• Results of study
• Next Steps

3
VIRTUAL STFI SENSOR

• Objective Predict Strength
• Key bets
• Knowing the strength can improve operators
  ability to reduce cull and optimize speed.
• If it can be predicted, it can be automatically
  controlled.

4
VIRTUAL STFI SENSOR

• Background
• What is STFI ?
• It is the principle strength property.
• It is typically measured on reel turn-up by the
  Paper Test Lab.
• Operators must wait 15-minutes before test
  results are available.
• Operators use Jet/Wire and refining to maintain.

5
VIRTUAL STFI SENSOR

• Background
• No.1 Machine Process
• Produce Linerboard 35-69 lb./MSF
• Rule 41 grades (mullen)
• Alternate Rule grades (STFI)
• 3-Ply sheet with 2 headboxes
• 6 stages of refining
• 5-30 recycled fiber (OCC)

6
VIRTUAL STFI SENSOR
7
VIRTUAL STFI SENSOR

• Background
• Experience with On-Line Measurement
• Worked on early generation gauge.
• Sold by the gauge for nearly 2-years
• Value was the information gained by the operator,
  i.e. faster testing.
• Use of the virtual sensor was intended to help
  reconcile calibration issues with the gauge.

8
VIRTUAL STFI SENSOR

• Process Characteristics
• Non-Linear
• example --- impact of mid-hole refiner on STFI
  may be different at high HPd/T versus low HPd/T.
• Interactions amongst many variables
• example --- impact of mid-tickler refiner may
  vary at different levels of OCC and broke
• The interactions made it nearly impossible to
  gain information from traditional bump tests.

9
VIRTUAL STFI SENSOR

• Pavilion Predicter
• Predicted STFI is calculated every 5-minutes for
  display to operator.
• Model is run in a separate computer and result
  (STFI) is written into PI database.
• Monitor model performance with X-Y and SQC tools
  in PI.

10
VIRTUAL STFI SENSOR

• Neural Net Model
• Using 17 process measurements, including
• total head
• refiner specific energy (HPd/T)
• couch vacuum
• headbox flow
• OCC broke flows
• The process variables have remained unchanged,
  but their relative impact varies from month to
  month.

11
VIRTUAL STFI SENSOR

• Results of Study
• Quickly achieved same accuracy as the on-line
  gauge.
• Model stability has been good. Re-trained once in
  8-months.
• Model step-outs still occur. Reasons unknown.

12
VIRTUAL STFI SENSOR
Trends of lab test model.
13
VIRTUAL STFI SENSOR
Overall difference between lab test model.
14
VIRTUAL STFI SENSOR
Model fit to lab test for all grades.
On the trained data, R² ?0.95
15
VIRTUAL STFI SENSOR
Stability of model prediction.
16
VIRTUAL STFI SENSOR
Model prediction compared to TAPPI test variation.
17
VIRTUAL STFI SENSOR
Forecasting STFI 30-minutes ahead.
18
VIRTUAL STFI SENSOR

• Lessons Learned
• Understand the statistical properties of the
  modeled variable, such as sample and test errors.
• Look for system changes to facilitate change. For
  example, use of SPC/SQC tools.
• Change, even for the better, is hard for
  operators to accept.

19
VIRTUAL STFI SENSOR

• Next Steps
• Work with the operators to make model more
  useful.
• Improve performance of the 30-minute prediction.
• Re-train model after changes to refiner plates.
• Assess opportunity for closed loop control.

20
VIRTUAL STFI SENSOR
Questions ?
21
Study of CONCORA
VIRTUAL STFI SENSOR
Collected 85 variables from PI system for 3
months.
22
Neural Net Model
VIRTUAL STFI SENSOR
Formula using
CONCORA
These variables are used to calculate
CONCORA, not for control.
23
Prediction Results for 26-lb.
VIRTUAL STFI SENSOR
3-month study
24
Residual Analysis Results
VIRTUAL STFI SENSOR
Predicted is within 2.5 of test 95 of the
time.
25
Results of Feasibility Study
VIRTUAL STFI SENSOR

• A neural net model has been built using process
  data from 3 months operation.
• CONCORA was predicted within 2.5 approximately
  95 of the time.