![[logo] US EPA](http://www.epa.gov/epafiles/images/logo_epaseal.gif){kind=logo}
Artificial Intelligence Abstracts
Neural Networks for Development of Apparent Zero Lag Time Sensors for Process Control
C.A. Miller, P.M. Lemiuex, K.J. Fritsky, P.J. Chappell U.S. Environmental Protection Agency Air and Energy Engineering Research Laboratory Combustion Research Branch (MD-65) Research Triangle Park, NC 27711 and R.L. Capone Ronald L. Capone Associates Arlington, VA
Abstract
Almost every process control uses some form of sensor as a means of
measuring the state of the controlled process. The time between measuring
the state of the process and the output of that measurement for use
in a control system can often be longer than the length of a transient
process change. Unfortunately, there is often no alternative to the
slow response and, although faster responding sensors may be available,
their cost is high. By predicting the state of the process a short
time ahead, artificial neural networks (ANNs) have the capability
to create "virtual sensors" which have apparent zero lag
time. ANNs have been successfully used to predict the behavior of
time-series data such as daily stock market prices after being adequately
trained using historical system behavior. In this paper, back-propagation
ANNs have been trained using extended delta-bar-delta paradigms to
predict the concentration levels of oxygen, carbon monoxide, and total
hydrocarbons from the transient incineration of surrogate liquid hazardous
wastes. Given a sensor lag time of t, the ANNs use the measurements
of the gas concentrations at the current time t and previous times
t-dt and t-2dt to predict the gas concentration values at time t+dt.
Since the current measurement output reflects the process state which
occurred in the system at time t-dt, the ANN output value for time
t+dt will be a prediction of the actual system state at time t. Time-series
results are plotted for several test runs, and the performance measurement
of the ANNs is given for all test runs completed. Criteria are presented
for evaluating the success or failure of network training, and for
identifying the need for additional network training and sensor failure.
"Development of an
Artificial-Intelligence-Based System to Control Transient Emissions
from Secondary Combustion Chambers of Hazardous Waste Incinerators,"
P.M. Lemieux C.A. Miller, K.J. Fritsky, P.J. Chappell, Paper presented
at the 1995 International Incineration Conference, Seattle, WA, May
8-12, 1995.
Development of an Artificial-Intelligence-Based System to Control Transient Emissions from Secondary Combustion Chambers of Hazardous Waste Incinerators
P.M. Lemieux, C.A. Miller, K.J. Fritsky, P.J. Chappell United States Environmental Protection Agency Air and Energy Engineering Research Laboratory Research Triangle Park, NC 27711
Abstract
Experiments have been performed on a 73 kW (250,000 Btu/hr) rotary kiln
incinerator simulator equipped with a 73 kW secondary combustion chamber
including an afterburner to develop and evaluate the performance of a
dynamic system to control transient emissions from the rotary kiln by
augmenting the afterburner flame with pure oxygen, injected at a rate
determined by a fuzzy logic control algorithm. Liquid surrogate waste
consisting of reagent grade toluene was injected into the rotating section
of the natural-gas-fired kiln. The toluene was injected in a controlled
transient manner, programmed to simulate the sudden release of liquid
waste from a container that was batch fed into an incinerator. The investigators
developed a fuzzy rule set based on system response during previous experiments
with the same facility, and developed the custom process control software.
The control software used signals, measured at both the kiln exit and
the afterburner exit, to determine an appropriate injection rate of oxygen
into the afterburner. System performance was measured by using a combined
performance indicator consisting of integrated weighted mass emissions
of carbon monoxide, total hydrocarbons, and soot.