Deep Learning-Based Regression Prediction of Relative Separation Distance of Rockets

The assessment for the rocket separation process still stays in qualitative observation, and the research is based on the deep learning method to realize the relative separation distance regression prediction in the rocket separation process, and realize the shift from qualitative observation to quantitative assessment for the rocket separation process. Based on the data acquired from the rocket separation video and the separation stroke sensor, a deep regression neural network relying on the feature extraction capability of ResNet50 is constructed to learn the complex mapping relationship between the separation image features and the relative separation distance labels, in order to realize the accurate prediction of the relative separation distance of unfamiliar separation images. It is experimentally verified that the predicted separation distances are in good agreement with the measured values of the sensors (the average absolute error is 0.0404, the Pearson correlation coefficient is 0.9986, and the coefficient of determination is 0.9961), and the prediction results of the unfamiliar separation images show an average relative error of 11.97%, and the maximum relative error is 17.42%. The evaluation results show that the proposed method has good practicality for realizing the quantitative evaluation of the rocket separation process.