Using Deep Learning to Predict Aircraft and Missile Trajectories in Simulated Scenarios – Development and Evaluation of a CNN-LSTM and an iTransformer Model Trained on MATLAB/Simulink Simulation Data

Sammanfattning

This thesis proposes and evaluates two deep learning models developed to support and extend the MATLAB/Simulink models currently used at the Swedish Defence Research Agency (FOI). These models, which are used for conducting simulations of different combat scenarios, such as between an aircraft and a missile, are computationally expensive and limit the number of scenarios that can be efficiently explored. To enable faster generation of additional simulated data, a CNN-LSTM hybrid model and a Transformer-based model (iTransformer) are trained on nearly 45 000 previously simulated scenarios. The dataset includes static input parameters and multivariate time series data representing the 3D positions of the aircraft and the missile. The models are trained to predict the continuation of trajectories based on initial conditions and are evaluated using standard regression metrics and inference time. Results show that both models capture most trajectory patterns well with low positional error and significantly reduced simulation time. Although direct comparison is not the primary aim of this thesis, differences in input sequence lengths and downsampling prevent the proposed models from being directly compared with each other or with the MATLAB/Simulink model. This thesis presents the development of the models, along with a discussion of their limitations. While further validation is needed, the results demonstrate the potential of deep learning models to support and accelerate scenario generation alongside traditional simulation methods.