Analysis and Prediction Research for Bipropellant Thruster Mixture Ratio Based on BP-RNN Chain Method

Abstract

In order to improve the mixture ratio accuracy of the bipropellant thruster and prolong the service life of a satellite on orbit, a variety of different liquid-flow tests and the final firing test are critical in the production testing process. However, after comparison with a large number of tests data, it was found that the liquid-flow test results were far away from matching the firing-test data without clear regularity, resulting in a qualified rate of mixture ratio of less than 40%. This study developed a BP-RNN (back propagation—recurrent neural network) chain method based on machine learning, which uses multi-dimensional nonlinear parameters to construct the specific dataset after data enhancement. Then, the mapping characteristic of the neural network was used to fit the historical data for the weight analysis and mixture ratio prediction, and effectively improved the qualified rate of the mixture ratio. The back propagation neural network was used to learn the association rules of the 10-dimension characteristic data and the firing test results generated in the historical process of thruster production. Then, the features with high influencing weight were extracted and sorted, so the “many-to-one” mixture ratio prediction was conducted through the subsequent recurrent neural network. The accepted prediction accuracy could reach around 75% after the test data verification. By using this method, most of bipropellant thrusters could directly reach the qualified mixture ratio in the firing test after adjusting the throttle orifice size in the liquid-flow tests. This chain method first bridges the data between the liquid-flow test and the firing test.