The design of neural networks for the performance estimation of satellite transponders

It has become increasingly important to improve upon the performance of satellite transponders, whose function is to receive and transmit signals automatically when triggered by an interrogator. The transponder of INTELSAT V whose performance parameters include: noise figure, group delay, gain, frequency translation, and power output is considered for investigation. The focus of this project is to design an Artificial Neural Network (ANN) as a new analytical model and computational technique to assess the intricate interactions between the transponder performance parameters and the environment in order to improve future satellite transponder performance and design.

The rationale for the use of ANN as a means of estimating the performance of a transponder lies in their parallel computation, learning ability, optimization capability, distributed data presentation, and ability to handle various tasks that are difficult for traditional computer techniques. Computer analysis tools are used to generate an optimal ANN model that meets the design specifications.

Finally, several candidate ANN models are investigated and the proposed models are selected based upon the result that minimizes the mean square error. The analysis will address the design of ANN using hypothetical training and validation data which incorporates a comparative assessment of the ANN estimation accuracy relative to the number of training patterns, iterations, hidden units, and learning parameters. Consequently, the impact of dynamic threshold values to interpret the ANN's response relative to transponder performance specification by the postprocessor is discussed.