Virtual Sensor TechnologyVirtual Sensor Technologyfor Process Optimizationfor Process Optimization

Edward WilsonNeural Applications Corporation

ewilson@neural.com

Virtual Sensor (VS)

l Also known as “soft sensor,” “smart sensor,” “estimator,” etc.

l Used in place of real sensor (RS)l Takes readings from RSs and control

variables, calculates values of (unsensed) process variables.

Reheat Furnace Virtual Sensor

l Real Sensor -> process model -> VS outputs

Virtual TemperatureSensors

Virtual Sensor - Example Output

l Reheat furnace example

0 10 20 30 40 500

200

400

600

800

1000

1200

1400

1015 1127 1208

Virtual Sensor Utilization

## ## ##

Process Data

Process Data

wall temp. history

beam cycle, freq.

Billet size, grade, loading pattern

mill conditions, delays

Control Decision

Goals of Talk (Outline)l Introduce technology (background

information for following talk)– Virtual sensor– Neural network

l Example applicationsl Discuss how to identify potential

applications of this technology

Types of VS Applications

l Replace a temporarily installed sensorl Provide continuous output based on periodic

RS measurements (e.g., lab analyzers)l Predict ahead for systems with built-in delay

- allows predictive controll Provide robustness - substitute VS when RS

fails or is down for maintenancel In all cases, model is needed

Basic Virtual Sensor Technology

l VS - gives measurement in place of a RSl Requires system model to process data from RSs

RealSystem

VSProcessing

controlinputs

outputs measuredwith real sensors

VSoutputs

usesdynamical model of system

or "transfer function"

Gear Vibration VSl Gear research by Joel Limmer at

Mechanical Diagnostics Laboratory at RPIl Uses temporarily installed rot. vib. sensor

Transfer Function Example

model ofTranfer

Function

error

Gearsrotational vibration

translational vibration

adaptation

model ofTranfer

Function

translationalvibration

(from RS)

rotationalvibration

(VS output)

Dynamical Model Example

l Example from Reheat Furnacel Develop dynamical model that can be run

forward in time to predict future outputsl RS used to develop model (adapt parameters)

errorFurnace/

billetmodel

adaptation

TkTk+1

Real Sensor(pyrometer)

furnace variables

Dynamical Modeling

l Can be modeled, even with intermittent RS datal If accurate model, can predict ahead, optimize

control inputs

Furnace/billet

modeladaptation

Tk+2Tk+3

Real Sensor(pyrometer)Furnace/

billetmodel

Tk+1Furnace/

billetmodel

Tk

Furnace/billet

model

Tk+2Tk+3

Furnace/billet

model

Tk+1Furnace/

billetmodel

Tk

Model Accuracy is Critical

l VS output depends on model accuracyl RS accuracy important

– used to build model– used as inputs to VS processing (GIGO)

l VS measurement must be “observable” from RS data

l Often, this is where the real challenge is

Modern Control - Estimator

l Primarily for Linear Systems. Also Kalman Filter, EKF.

l State-feedback control

l If model unknown, must be “identified.”

Virtual Sensor Modeling

l Industrial control systems generally don’t use state feedback control -> not full estimator, just certain VSs for (e.g., PID) control loops.

l Often nonlinear / poorly understood / time-varying processes

l Use Neural Networks (NN) for modeling. (e.g., hybrid NN with linear model)

Outline

l Introduce Technology– Virtual sensor– Neural network

l Example applicationsl Discuss how to identify potential

applications of this technology

Properties of Neural Networks

l Neural Networks (NNs) are known to have valuable capabilities such as:– Nonlinear ==> deal with real-world – Adaptive ==> “trained” with data to solve problems,

adapt to changing systems– Parallel architecture ==> fast in hardware– Generic functional element ==> can model anything

l However …– costs must be weighed vs. these benefits

Biological Motivation,Engineering Application

l Human brain:– Massively parallel network of simple

processors with great capabilities 15 billion neurons

– 10,000 inputs per neuron– 1-2 ms neuron response time

l ANNs studied in a variety of fields– Engineering– Psychology

Model of Single Neuron

y = sigmoid(w1x + w2x + ... + wnx + bias)

sigmoid(wx) = 1 - e-wx

1 + e-wx

W1x1

W2x2

Wnxn

W0

x0 (= 1)

(bias)

w x sigmoid( ) y

l Implement in software or hardwarel Loosely modeled from biology, but chosen for processing and trainingl This type most common for engineering applications

Model of Neural Network

l output = W2*sigmoid(W1*inputs)l proven to be a generic nonlinear functional elementl Functionality defined by architecture, weights, (training)

OutputsInputs

Neurons(nonlinear

processing units)

Weights(determine

connection strengthbetween neurons)

W1 W2

NN Background - Summaryl Generic nonlinear functional element - can implement any

MIMO mapping function to arbitrary accuracy (universal approximator)

l “Trained” with data

l Solid mathematical foundation - BP gets derivatives, then standard gradient-based optimization problem

l Parallel architecture, but usually implemented in software on serial computer

l Black box - difficult to understand inner workings

Training NN Model

l NN trained to emulate a physical processl Parameter ID issues :sample rate, sufficient

data, sufficient dof in model, etc.

NN

input

NNoutput

output

error cost0.5 ( )2

BP training

physicalprocess

Data Pre-processing, Structure

NN

input

NNoutput

output

error cost0.5 ( )2

BP training

physicalprocess

remove outliersfilter data

select/combine inputs and delayed inputspreprocess with known functions

linear/nonlinear

model

linear/nonlinear

model

Technology Summaryl Virtual sensors

– make virtual measurements by processing control inputs and measurements from real sensors

– depend on accurate system model– for nonlinear, complex systems, NN model used

l Neural networks– generic nonlinear processing element– functionality set by “training” with data– can be used in “hybrid” modeling structures

Outline

l Introduce Technology– Virtual sensor– Neural network

l Example applicationsl Discuss how to identify potential

applications of this technology

Neural-Network Dryer Example

Rotary DryerBurnerFurnace

Combustion Air/Fuel

Tempering Air

DischargeHousing

Discharge Screw

FeedScrewSpeed

FanSpeed

DriveSpeedAmps

ManualMoistureSample

Feed

Temp.DEGF

Drum Drive

Oulet TeperatureInlet Teperature

Temp.DEGF

ScrewSpeed

Air/Fuel

Recycled Air%

valveopen

REAL-TIMEOUTPUT-MOISTURE

PREDICTION

lDelayed sensing of product water content

1 5

2 0

2 5

3 0

3 5

4 0

4 5

1 1 0 1 2 0 1 3 0 1 4 0 1 5 0 1 6 0 1 7 0 1 8 0 1 9 0 1

T i m e ( m i n )

D e s i r e d

A c tu a l

Chemical Reaction TankReduce Process Variation

Standard Deviation BEFORE = 1.38Standard Deviation BEFORE = 1.38

Chemical Reaction TankReduce Process Variation

34

39

44

1 101 201 301 401 501 601 701 801 901

Time

Pla

nt O

utpu

t

Desired

Actual

Standard Deviation AFTER = 0.85Standard Deviation AFTER = 0.85

ICC

PlantPlant Output

Σ

ICPError

ΣDesiredOutput

ProcessVariables

DCS

ProcessSet Points

ProcessVariables

Neural

Fuzzy

ProcessSet Points

ProcessVariables

ProcessSet Points

Predicted Plant Output

ICCP Block Diagram

+ -

+-

Green Sand Process Example

Binder

Water

Hopper

Muller

MullerMotor

Temp Moisture

Amps

Compact-ibility

GreenStrength

Flow Flow

lDelayed, intermittent sensing of compactibility, green strength

Rolling Mill Application

Neural IRMC

Entry Gauge

Exit Gauge

Force

Rolling Mill Gauge Predictor/Controller

Model Error

Plant

ΣRegulationCorrection

DesiredExitThickness

Exit Gaugeat Time N

NeuralModel

U(N+1)Control

(PID) Σ

Σ Modeled Plant Output

PredictedExit Gauge

at TimeN+10P

LANT

Gauge Predicted vs. Actual Exit GaugeT

rain

ed o

n m

any

coils

, T

este

d on

coi

l #2

Tra

ined

on

coil

#1,

Tes

ted

on c

oil #

2Plotted vs. Time Scatter Plot

Hybrid Sensing/Control SolutionHybrid Sensing/Control Solution

ll No Universal Solution to Control ProblemsNo Universal Solution to Control Problemsll VS with PID VS with PID vsvs. Intelligent . Intelligent controllercontrollerll Best solution may draw upon various technologiesBest solution may draw upon various technologies

–– Neural NetworksNeural Networks–– Fuzzy LogicFuzzy Logic–– StatisticsStatistics–– Classical MethodsClassical Methods

ll Always requires some level of process knowledgeAlways requires some level of process knowledge

Model Error

Plant

Σ

RemoteSetpoints

(Fuzzy, Expert, ...)

Desired StatesS(N+1)

S(N+1)

NeuralModel

U(N+1)Control

(PID)

Modeled Plant Output

Neural/Fuzzy Remote Set-point GeneratorNeural/Fuzzy Remote Set-point Generator

Model Error

Plant

ΣRegulationCorrection

DesiredPlantStates

S(N+1)

S(N+1)

NeuralModel

U(N+1)Control

(PID) Σ

Σ Modeled Plant Output

Neural Network Predictive/Corrective Controller

Neural Network Predictive/Corrective Controller

Model Error

Plant

Σ

State Valuesfor Time

N, N-1, N-2, ...

ΣDesiredPlantStates

S(N+1)

S(N+1)

Regulation Error

NeuralRegulator

NeuralModel

U(N+1)

Modeled Plant State

Neural Network Based Predictive ControllerNeural Network Based Predictive Controller

Application Summary

l NN may serve key role, but is part of systeml NN modeling is parameter optimization

– choose structure of function to be adapted with minimal but sufficient dof

– need sufficient data– use known structure to extent possible

» e.g., linear + polynomial + NN» allows input of pre-calculated solutions» gradient-based optimization

l Preprocessing important

Outline

l Introduce Technology– Virtual sensor– Neural network

l Example applicationsl Discuss how to identify potential

applications of this technology

Types of VS Applications

l Replace a temporarily installed sensorl Provide continuous output based on periodic

RS measurements (e.g., lab analyzers)l Predict ahead for systems with built-in delay

- allows predictive controll Provide robustness - substitute VS when RS

fails or is down for maintenancel In all cases, model is needed

Where to use VS / NN

“This is a very hard problem and I don’t know to solve it, so I’ll see how a NN does at it.”

l Important decisionl Benefits vs. costsl Evaluate other solution methods

– cost of nonlinear optimizationl Significant effort analyzing physical system and

developing data pre-processing, system architecture, NN architecture, etc.

l System-level analysis

Where to Use NNs

one liner: “Use NN when data availability outweighs process understanding”

l Benefits - Nonlinear, Adaptive, Generic, Scalable processing, Parallel hardware

l Costs - Nonlinear optimization, Requires data, Black box

l Evaluation of “conventional” methodsl Use NN where these fall shortl Structure total solution to use NN in conjunction

with these

Summary

l Virtual sensing technology can provide:

– improved control by providing virtual measurements

– predictive capability

– continuous output from periodic real measurements

– robustness to RS failure

l VS output limited by accuracy of model and RSs

l Model structure important - process understanding needed

l Neural-network technology useful for modeling data-rich/theory-poor processes