On-line Performance Monitoring of a Chemical Process

1
On-line Performance Monitoring of a Chemical
Process
BP Chemicals/CPACT/MDC
2
Summary

• Will talk about application of multivariate SPC.
• A data visualisation system for overview of plant
  operation.
• Tried on Hull site plant.
• Will aid operators control of plant.

3
Introduction

• Plants have data overload.
• MSPC gives overview of plant operation on just a
  few graphs.
• PCA is used to compress correlated plant
  variables to just a few PCs.
• Technique was applied to the BP Hull A4 CO plant.
• Plant manufactures CO by steam reforming of nat.
  gas. CO is feedstock for acetic acid production.

4
Plant Schematic
MEA
Steam
Reformer Heat
CO2 Removal
Natural Gas (desulphurised)
H2 (NH3 plant)
Water Removal
Cold Box Separation
CO (acetic acid plant)
5
MSPC

• Data point on PC scores plots represent plant
  status at that time.

• Data points due to plant problem appear outside a
  confidence ellipse.
• Problem points also show up using statistical
  measures (e.g. SPE and T2 statistic) - distances
  from model.
• Problem points interrogated using contribution
  plots for causal variables.

6
First Model

• An MSPC model was built of normal operation for
  the A4 CO plant.
• The model used 27 main plant variables, including
  temperatures, pressures, flows and analyser
  results.
• Model training data was collected at 4 minute
  snapshots over a 1 week period of stable
  operation.
• 6 PCs explains 70 of variance.
• This is effectively then used as a basis to
  compare future operation.

7
Off-line Analysis

• Using CPACT MultiDAT and PreScreen2 Software

Off-line Analysis of Operator Error
8
On-line with MDC

• PC scores plots, statistics vs time, etc
• Zoom in
• Click on point to select contribution plot
• Plus off-line tools for model building
• Plus PLS and adaptive models

On-line Model, Feedstock Upset (N2)
9
On-line with MDC

• Normalised
• Greatest first
• Scrollable
• Click for time trend

Process Variable Contribution Plots
Time Trend of a Process Variable
10
Problems

• But plant operates at different rates.
• What data to use for model?
• What variables to use?
• Dynamic data influenced by the past.
• Serially correlated (invalid control limits).
• Result hard to find balance between alerts and
  false alarms.
• So concentrated on smaller section of plant
  tried new techniques.

11
Plant Section - MEA

• Model for MEA (10 tags)
• Has oscillation upset at high rates

The Oscillation
Upset in Column Level

• Clusters 1, 2 and 3 represent different modes
  of operation

3
2
1
Score Statistics Plots
12
Live Demo

• Live demo of MDC MSPC with the previous data for
  MEA.

13
Multi Rate Model

• Work by Ewan Mercer et al (CPACT Newcastle
  University).
• Model for MEA.
• Need models for different plant rates.
• Modes of operation seen as clusters.
• But better to collapse clusters together

PC1 vs PC2 Scores Plot
14
New Technique

• Also by Ewan Mercer et al.
• Based on plant model mismatch (PMM).
• Build state space model for MEA (1m data, I/Ps
  and O/Ps).
• Build PCA model on differences between predicted
  and actual plant data (residuals).
• Will effectively collapse clusters.
• Use standard MSPC graphs to monitor plant.

15
Plant/Model Mismatch
Schematic of Technique
16
PCA on Residuals

• Model for multiple plant rates.
• Near normal distribution with low serial
  correlation.
• Picks up upset with fewer false alarms.

17
Parallel Coords

• Another potential technique
• Light cluster is normal MEA operation
• Other darker data is upset
• Can also use to visualise many PCs

• Each y-axis is a plant variable
• Each path is state of plant at one point in time

18
Conclusion/Next Steps

• Model built for overview of the A4 CO plant.
• Tested on-line using MDCs MSPC software.
• Overview of plant operation with drill down.
• Picks up process problems and helps diagnose
  cause.
• Will improve running of plants.
• Gain site acceptance for deployment.
• Initially to see plant changes.
• Use alerting later with the new techniques (i.e.
  PMM).

19
Acknowledgements

• BP Hull Site
• Steve Batty, Zaid Rawi et al.
• CPACT/Newcastle University
• Ewan Mercer, Julian Morris, Elaine Martin et al.
• MDC
• Chris Hawkins, Paul Booth et al.