Virtual Sensor Technology for Process Optimization Edward Wilson Neural Applications Corporation ewilson@neural.com
Virtual Sensor (VS) Also known as soft sensor, smart sensor, estimator, etc. Used in place of real sensor (RS) Takes readings from RSs and control variables, calculates values of (unsensed) process variables.
Reheat Furnace Virtual Sensor Virtual Temperature Sensors Real Sensor -> process model -> VS outputs
Virtual Sensor - Example Output Reheat furnace example 1400 1200 1000 800 600 400 1015 1127 1208 200 0 0 10 20 30 40 50
Virtual Sensor Utilization wall temp. history Process Data beam cycle, freq. Billet size, grade, loading pattern mill conditions, delays Control Decision Process Data ## ## ##
Goals of Talk (Outline) Introduce technology (background information for following talk) Virtual sensor Neural network Example applications Discuss how to identify potential applications of this technology
Types of VS Applications Replace a temporarily installed sensor Provide continuous output based on periodic RS measurements (e.g., lab analyzers) Predict ahead for systems with built-in delay - allows predictive control Provide robustness - substitute VS when RS fails or is down for maintenance In all cases, model is needed
Basic Virtual Sensor Technology VS - gives measurement in place of a RS Requires system model to process data from RSs control inputs Real System outputs measured with real sensors VS VS outputs Processing uses dynamical model of system or "transfer function"
Gear Vibration VS Gear research by Joel Limmer at Mechanical Diagnostics Laboratory at RPI Uses temporarily installed rot. vib. sensor
Transfer Function Example Gears rotational vibration translational vibration model of Tranfer Function error adaptation translational vibration (from RS) model of Tranfer Function rotational vibration (VS output)
Dynamical Model Example Example from Reheat Furnace Develop dynamical model that can be run forward in time to predict future outputs RS used to develop model (adapt parameters) furnace variables Real Sensor T Furnace/ k T k+1 (pyrometer) billet error model adaptation
Dynamical Modeling Can be modeled, even with intermittent RS data If accurate model, can predict ahead, optimize control inputs T k Furnace/ billet model T k+1 Furnace/ billet model Real Sensor T Furnace/ k+2 T k+3 (pyrometer) billet model adaptation T k Furnace/ billet model T k+1 Furnace/ billet model T k+2 Furnace/ billet T k+3 model
Model Accuracy is Critical VS output depends on model accuracy RS accuracy important used to build model used as inputs to VS processing (GIGO) VS measurement must be observable from RS data Often, this is where the real challenge is
Modern Control - Estimator Primarily for Linear Systems. Also Kalman Filter, EKF. State-feedback control If model unknown, must be identified.
Virtual Sensor Modeling Industrial control systems generally don t use state feedback control -> not full estimator, just certain VSs for (e.g., PID) control loops. Often nonlinear / poorly understood / timevarying processes Use Neural Networks (NN) for modeling. (e.g., hybrid NN with linear model)
Outline Introduce Technology Virtual sensor Neural network Example applications Discuss how to identify potential applications of this technology
Properties of Neural Networks 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 However costs must be weighed vs. these benefits
Biological Motivation, Engineering Application 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 ANNs studied in a variety of fields Engineering Psychology
Model of Single Neuron x 0 (= 1) x 1 W 1 W 0 (bias) x 2 W 2 w x sigmoid( ) y x n W n sigmoid(wx) = 1 - e-wx 1 + e -wx y = sigmoid(w 1 x + w 2 x +... + w n x + bias) Implement in software or hardware Loosely modeled from biology, but chosen for processing and training This type most common for engineering applications
Model of Neural Network Weights (determine connection strength between neurons) Inputs W 1 W 2 Outputs output = W 2 *sigmoid(w 1 *inputs) Neurons (nonlinear processing units) proven to be a generic nonlinear functional element Functionality defined by architecture, weights, (training)
NN Background - Summary Generic nonlinear functional element - can implement any MIMO mapping function to arbitrary accuracy (universal approximator) Trained with data Solid mathematical foundation - BP gets derivatives, then standard gradient-based optimization problem Parallel architecture, but usually implemented in software on serial computer Black box - difficult to understand inner workings
Training NN Model NN trained to emulate a physical process Parameter ID issues :sample rate, sufficient data, sufficient dof in model, etc. input physical process output NN NN output error 0.5 ( ) 2 cost BP training
Data Pre-processing, Structure input physical process output remove outliers filter data select/combine inputs and delayed inputs preprocess with known functions linear/ nonlinear model NN NN output error 0.5 ( ) 2 cost linear/ nonlinear model BP training
Virtual sensors Technology Summary make virtual measurements by processing control inputs and measurements from real sensors depend on accurate system model for nonlinear, complex systems, NN model used Neural networks generic nonlinear processing element functionality set by training with data can be used in hybrid modeling structures
Outline Introduce Technology Virtual sensor Neural network Example applications Discuss how to identify potential applications of this technology
Neural-Network Dryer Example Feed Recycled Air % valve open REAL-TIME OUTPUT -MOISTURE PREDICTION Feed Screw Speed Tempering Air Burner Furnace Combustion Air/Fuel Fan Speed Inlet Teperature Temp. DEGF Air/ Fuel Rotary Dryer Drive Speed Amps Drum Drive Discharge Housing Discharge Screw Oulet Teperature Temp. DEGF Manual Moisture Sample Screw Speed Delayed sensing of product water content
Chemical Reaction Tank Reduce Process Variation Standard Deviation BEFORE = 1.38 1.38 45 40 35 30 D e s ire d Actu al 25 20 15 1 101 201 301 401 501 601 701 801 901 Tim e (m in)
Chemical Reaction Tank Reduce Process Variation Standard Deviation AFTER = 0.85 0.85 Plant Output 44 39 34 Desired Actual 1 101 201 301 401 501 601 701 801 901 Time
ICCP Block Diagram ICC Fuzzy Process Set Points Process Variables DCS Process Set Points Process Variables Plant Plant Output Error Process Set Points Process Variables ICP Neural - Σ + Desired Output + - Σ Predicted Plant Output
Green Sand Process Example Hopper Temp Moisture Flow Flow Water Binder Delayed, intermittent sensing of compactibility, green strength Amps Muller Muller Motor Compactibility Green Strength
Rolling Mill Application Entry Gauge Force Exit Gauge Neural IRMC
Rolling Mill Gauge Predictor/Controller Desired Exit Thickness Control (PID) S U(N+1) Plant Exit Gauge at Time N Regulation Correction S P L A N T Neural Model Model Error Modeled Plant Output Predicted Exit Gauge at Time N+10 S
Gauge Predicted vs. Actual Exit Gauge Plotted vs. Time Scatter Plot Trained on many coils, Tested on coil #2 Trained on coil #1, Tested on coil #2
Hybrid Sensing/Control Solution No Universal Solution to Control Problems VS with PID vs.. Intelligent controller Best solution may draw upon various technologies Neural Networks Fuzzy Logic Statistics Classical Methods Always requires some level of process knowledge
Neural/Fuzzy Remote Set-point Generator Desired States Remote Setpoints (Fuzzy, Expert,...) Control (PID) U(N+1) Plant S(N+1) Neural Model S(N+1) S Modeled Plant Output Model Error
Neural Network Predictive/Corrective Controller Desired Plant States Control (PID) S U(N+1) Plant S(N+1) Neural Model Regulation Correction S(N+1) S S Model Error Modeled Plant Output
Neural Network Based Predictive Controller State Values for Time N, N-1, N-2,... U(N+1) Plant S(N+1) Neural Regulator Neural Model S(N+1) S Desired Plant States S Regulation Error Modeled Plant State Model Error
Application Summary NN may serve key role, but is part of system 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 Preprocessing important
Outline Introduce Technology Virtual sensor Neural network Example applications Discuss how to identify potential applications of this technology
Types of VS Applications Replace a temporarily installed sensor Provide continuous output based on periodic RS measurements (e.g., lab analyzers) Predict ahead for systems with built-in delay - allows predictive control Provide robustness - substitute VS when RS fails or is down for maintenance 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. Important decision Benefits vs. costs Evaluate other solution methods cost of nonlinear optimization Significant effort analyzing physical system and developing data pre-processing, system architecture, NN architecture, etc. System-level analysis
Where to Use NNs one liner: Use NN when data availability outweighs process understanding Benefits - Nonlinear, Adaptive, Generic, Scalable processing, Parallel hardware Costs - Nonlinear optimization, Requires data, Black box Evaluation of conventional methods Use NN where these fall short Structure total solution to use NN in conjunction with these
Summary Virtual sensing technology can provide: improved control by providing virtual measurements predictive capability continuous output from periodic real measurements robustness to RS failure VS output limited by accuracy of model and RSs Model structure important - process understanding needed Neural-network technology useful for modeling datarich/theory-poor processes