Smart projectile parameter estimation using meta-optimization

Parameter estimation;Numerical optimization;Smart projectiles;Reinforcement learning

Gross, Matthew ; Costello, Mark Aerospace Engineering Johnson, Eric German, Brian Kennedy, Greame Ferri, Aldo ; Costello, Mark

University:Georgia Institute of Technology

Department:Aerospace Engineering

关键词: Parameter estimation;    Numerical optimization;    Smart projectiles;    Reinforcement learning;   

Others  :  https://smartech.gatech.edu/bitstream/1853/59190/1/GROSS-DISSERTATION-2017.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

System identification and parameter estimation are valuable tools in the analysis and design of smart projectile systems. Given the complexity of these systems, it is convenient to work with mathematical models in place of the actual system. Parameter estimation uses time history data of the system to determine a model that accurately matches the data. Many techniques have been developed to perform parameter estimation, including regression methods, maximum likelihood estimators, and Kalman filters.Maximum likelihood methods, in particular the output error method (OEM), pose the estimation problem in terms of an optimization problem. OEM has seen extensive use on projectile systems, utilizing a numerical optimizer such as a Newton style algorithm to solve for unknown parameters. These algorithms are prone to converging on local minima present in the projectile dynamics, requiring reasonable initial guesses of the parameters to ensure convergence. However, for new smart projectile systems, prior estimates of the control parameters may not be available.Thus, there is a need for reliable and robust parameter estimation methods that are not dependent a priori knowledge of the parameters.This thesis proposes a new method for smart projectile parameter estimation based on OEM. To achieve robust and reliable parameter estimates, a new underlying optimization algorithm is formed, dubbed meta-optimization. Meta-optimization employs a diverse set of individual optimization algorithms with both local and global search capabilities. The meta-optimizer operates by iteratively selecting a single algorithm to deploy in a stochastic manner, giving preference to algorithms which have performed well on the problem. This approach allows synergies to develop between the individual optimizers, boosting performance beyond what each optimizer is capable of individually.A suite of benchmark functions commonly used in the optimization field are used to analyze the meta-optimization framework and compare it to other existing algorithms. These functions have the benefit of known structure and solutions and are less computationally intensive than the parameter estimation problem. A series of trade studies are performed to evaluate each component of meta-optimization and determine a robust configuration for use on general optimization problems. Meta-optimization is also compared to the individual algorithms it employs, showing superior performance and reliability over this benchmark suite. Finally, meta-optimization is compared to other state of the art algorithms, showing comparable performance.The new parameter estimation method is applied to an example smart projectile system equipped with a new microspoiler control mechanism. Both synthetic and experimental trajectory data is used to evaluate the effectiveness of the proposed method. From the synthetic data, the parameter estimation algorithm accurately estimates the aerodynamic coefficients of a standard projectile as well as parameters for a smart projectile executing a maneuver. This synthetic data is also used to conduct trade studies investigating how the data itself impacts the accuracy of the parameter estimates. Lastly, the method is applied to flight test data collected at the U.S. Army Research Laboratory spark range with good results in the presence of large measurement errors.