Thrust Prediction of Hybrid Rocket Motor Based on Physical Feature Embedding and Residual Learning

Hybrid Rocket Motors (HRMs) have gained considerable attention as a significant rocket propulsion system due to their adaptive thrust modulation and multiple-start capability. The precise prediction of HRM thrust is crucial for various engineering applications, including real-time thrust control and failure recognition, particularly in the face of complex fuel regression laws and nozzle erosion. Conventional mechanism-based prediction method relies on zero-dimensional internal ballistic simulation, which deviates from the actual physical process and struggles with model parameter calibration under different conditions. On the other hand, emerging machine learning-based method leverages multi-source data characteristics but is hindered by limited datasets and inappropriate sample distribution. To address the limitations of both mechanism-driven and data-driven approaches, the paper proposes a deep-learning framework based on physical feature embedding and residual learning to enhance thrust prediction for long-duration working HRMs. The method involves establishing a medium-fidelity HRM internal ballistic model that accounts for nozzle erosion and non-ideal factors. Subsequently, a deep neural network (DNN) extracts essential physical features of the ballistic model based on input combustion chamber pressure and oxidizer mass flow rate. These features, encompassing parameters related to fuel regression rate and nozzle erosion, are incorporated into the mechanism model to evaluate baseline performance. A data-driven residual term is then added as a model discrepancy function to refine the thrust prediction. The proposed framework’s performance is validated using a series of hot-firing test data from a long-duration working 98%Hydrogen Peroxide/HTPB HRM, employing a cross-validation strategy to maximize the utility of limited datasets. Results indicate that the framework enhances the generalization of data-driven method under a small dataset, achieving 55.41% and 9.20% improvement compared with mechanism-driven and data-driven method alone respectively. Remarkably, the DNN-based prediction significantly increases efficiency, consuming only 0.33% of the time required by the mechanism-driven method. Further analysis confirms that the physical-guided term forms the main part of the prediction results, and the residual term has the ability to adaptively make up for the model discrepancy. In conclusion, the integration of mechanism-driven and data-driven in the proposed framework strikes a balance between generalization and accuracy compensation, positioning it as a promising approach for HRM performance prediction.