Füze dinamik modelinin veri tabanlı yöntemler ile kestirimi ve kontrolü

Dynamical systems in nature are generally nonlinear and usually contains many hidden dynamics. Therefore, the need for data-driven model discovery and control methods continues. The most popular of these methods is neural networks-based methods nowadays. However, excessive data requirements, long training times and most importantly lack of interpretability of results are the main problems in neural networks-based system identification methods. Among the many other methods used for model discovery, SINDY (Sparce Identification of Non-linear Dynamical Systems) has recently attracted great attention with its simple and effective nature. SINDY, which has many extensions, also has various open problems. The proposed extension in this study is called SINDY-SAIC and combines the methods from Stepwise Sparse Regression (SSR) and Akaike Information Criteria (AIC) model selection algorithm. The need for tuning threshold parameter in SINDY is relaxed using SSR and the robustness to noisy measurements is increased with a newly used state derivative calculation method in sparse regression. In addition, presence of model selection with AIC enables sparse solution by penalizing the number of terms and prevents the algorithm to converge collinear basis. Studied dynamical systems are controlled by Model Predictive Control using discovered models. MPC is a control method that uses prediction models mostly discovered from data and try to minimize a given cost function subjected to the constraints. Both linear and nonlinear prediction models are generated using SINDY-SAIC and used in MPC as prediction models. The traditional state feedback (SF) controller is also presented for comparison. The proposed SINDY-SAIC algorithm and the controllers (MPC and SF) are tested for linear and highly non-linear longitudinal missile dynamics under moderate and high level of noise conditions.