Inverse Procedure to Initial Parameter Estimation for Air-Dropped Packages Using Neural Networks

This study presents an inverse approach utilizing neural networks to support the initial parameter estimation for air-dropped packages. The research is motivated by the growing need for accurate and autonomous systems in aerial delivery missions, particularly in environments where satellite navigation signals (e.g., GPS) are unavailable or degraded. In the first step, a forward neural network is trained to predict key flight parameters—range, flight time, and impact velocity—based on predefined initial conditions such as drop velocity, angle, and height. In the second step, an inverse neural network is developed to estimate these initial conditions based solely on the desired outcome parameters. A hybrid computing architecture combining simulation and neural modeling is proposed to generate sufficient training data and reduce computational cost. The results demonstrate that the proposed inverse model can accurately reconstruct the initial parameters under both deterministic and random test scenarios, provided that the input values lie within the well-represented domain of the training set. The approach offers a robust, low-cost alternative to traditional methods, potentially enabling real-time mission planning and decision support for humanitarian, scientific, or military applications.