MLP Based Feedback System for Gas Valve Control in a Madison Symmetric Torus

1
MLP Based Feedback System for Gas Valve Control
in a Madison Symmetric Torus

• Andrew Seltzman

Dec 14, 2010
2
Background and Project Description

• MST is a plasma physics experiment currently
  running in the physics department. MST operates
  in a pulsed mode, taking plasma shots that last
  approximately 40-60ms. Before the shot, gas input
  at 15 different points in time during the shot
  and plasma current is set. Density values drift
  from one shot to the next, requiring operator
  input to adjust the gas input profile.
• Currently a human operator is required to
  manually set gas input valves to stabilize
  density.
• Density fluctuates around desired level due to
  human error
• No automatic control system exists due to the
  complexity and non-linearity of the system in
  question
• A MLP will be designed to emulate and eventually
  replace the skilled human operator.

3
ANN Control System Setup

• Control system given system parameters and
  requested output
• Error computed from actual output
• BP learning
• Eventual convergence ?

Requested density
Human input
Error
Density
Feature space compression
MLP
MST (system to control)
Current
Gas input
4
MLP Design

• Simplest MLP required to adequately classify
  output data
• 40 element feature space
• lt 40 elements in first hidden layer
• lt 2 hidden layers
• 20 element output layer
• Pre-processing of raw date to compress feature
  space prior to MLP input
• BP training algorithm
• Initially train MLP from human operator responses
• Eventual self learning when integrated into
  control system
• Calculation of error (requested density actual
  density)
• BP learning to allow real time adaptability to
  varying system conditions
• Optimization of initial training parameters and
  network design in progress
• Number of layers, Number of elements
• Training parameters learning rate, momentum,
  etc

5
Feature Space Data
Initial Feature Space
Compressed Feature Space

• High speed digitizer captures experiment data
• 30000 data points for Ip and Ne
• 8000 actual relevant data points
• 21 data points for Gas input
• ANN complexity with 60000 input neurons would be
  wasteful of computing resourced
• Real time operation required (1 update every 2
  min)
• Feature space compressed by averaging points in a
  given data set
• 20 data points for Ip and Ne
• 2 data points for Gas input
• Relevant data extracted without loss accuracy

6
Expected Results / Initial Testing

• MLP given requested plasma density and plasma
  current
• System successfully generates gas waveform
• Gas waveform very similar to training set
• Not integrated with actual system
• Gas output appears similar to human operator
• Accurate model?
• Further testing / training required with multiple
  data sets

7
Discussion / Problems / Future Modifications

• Control method
• Output types
• Output gas values based on requested density
• Larger training data set
• Data available from every shot
• Known density result
• Output change in gas valued to error in density
• Smaller training data set
• Data only when operator changes gas settings
• Requested density value not known
• Over specification of fitted data
• Training sometimes results of MLP copying human
  gas output regardless of density