Review of the state of the art of intelligent control for large stationary engines
- Colorado State Univ., Fort Collins, CO (United States). Dept. of Mechanical Engineering
A project was undertaken to assess the state-of-the-art artificial intelligence in light of application to the monitoring and control of large bore natural gas engines. Several artificial intelligence/intelligent control techniques were examined: heuristic search techniques, adaptive control, expert systems, fuzzy logic, neural networks, and genetic algorithms. Of these, it was concluded that neural networks have the most potential for use on large bore engines due to their capability to recognize patterns in incomplete, noisy data. Neural networks are systems which consist of a large number of very simple interconnected processing units which are capable of `learning` through manipulation of connection strengths between elements. An examination of the information available from combustion cylinder pressure-time histories confirms the potential for predicting emissions from these curves. An examination of current techniques used for emissions predictions from combustion pressure indicates significant potential for improvement by exploiting the pattern recognition capabilities of neural networks. To test the potential of neural networks to predict emissions from pressure-time histories, a two-part experimental program was conducted. In the first part of the program, neural networks were used for a very simple pattern recognition task: determination of ignition timing from combustion pressure histories measured in an automotive engine. A series of combustion pressure histories was discretized into a series of pressures corresponding to particular crank angles. Four different neural networks were constructed which utilized 37, 3, 2, and 1 pressure values as inputs, respectively. The best network with 37 input points showed the ability to predict ignition timing within a specified tolerance level 98.8% of the time, when tested on an averaged testing set. The best networks with one, two, or three input points all showed approximately 90% recognition.
- Research Organization:
- American Gas Association, Inc., Arlington, VA (United States). Pipeline Research Committee
- Sponsoring Organization:
- American Gas Association, Inc., Arlington, VA (United States)
- OSTI ID:
- 576048
- Report Number(s):
- AGA-98002888; PR-179-9131; TRN: 98:001464
- Resource Relation:
- Other Information: PBD: Sep 1996
- Country of Publication:
- United States
- Language:
- English
Similar Records
Hybrid system for fault diagnosis using scanned input: A tutorial
Adaptive Remote-Sensing Techniques Implementing Swarms of Mobile Agents