An important secondary topic addressed in the research and development effort described in the preceding article is the use of artificial neural networks to improve the monitoring and thus the control and safety of multiple free-piston Stirling engines. Information collected by monitoring subsystems constitutes essential feedback for use by control and safety subsystems. This information includes such externally measurable quantities as heater-head temperatures, motions of engine housings, and output currents and voltages.

An Artificial Neural Network could be part of the feedback control system of a Stirling engine or an assembly of such engines. The neural network would estimate internal operating parameters (e.g., pressure or piston motions) from externally measurable quantities (e.g., output current, output voltage, and current/voltage phase angle).

Prior to this effort, typical approaches to the monitoring and control of Stirling engines involved the use of extensive data-acquisition systems that collected, in addition to externally measurable quantities, such critical internal operating parameters as pressures and motions of components as measured by invasive pressure and position probes, respectively. Unfortunately, such probes are expensive, are potential sources of failure, and compromise design options.

The neural-network approach offers an inexpensive, simple, and highly reliable alternative to the use of invasive probes. An artificial neural network can be particularly useful for modeling and monitoring a complex mechanical/electrical system like a Stirling engine, for which the mathematical relationships among input and output variables are either unknown or too complex to be represented by an analytical model. The basic idea is to train an artificial neural network to infer information about internal operating conditions from measurements by minimal (only external) instrumentation in order to detect actual or incipient failure or deterioration.

An autonomous control system could process the output of such a neural network (see figure) into control commands to perform corrective actions to maintain reliable engine operation. For example, in the case of two or more Stirling engines operating in synchronization as described in the preceding article, a neural network could compensate for deterioration of one of the engines by triggering a command to ramp down the output of that engine and ramp up the output(s) of the other engines. The feasibility of this neural-network concept was demonstrated in an experiment on two coupled Stirling engines: While one engine was operated under nominal conditions, the other was operated at a series of reduced pressures to simulate the effects of a slow leak. An artificial neural network, fed only data on root-mean-square currents, output power, and current/voltage phase angles from the two engines was found to infer the decrease in pressure in the affected engine with high accuracy.

This work was done by Laurence B. Penswick of Stirling Technology Co. for Glenn Research Center.

Inquiries concerning rights for the commercial use of this invention should be addressed to

NASA Glenn Research Center
Commercial Technology Office
Attn: Steve Fedor
Mail Stop 4—8
21000 Brookpark Road
Ohio 44135.

Refer to LEW-16822.