Real-time identification of an unmanned quadcopter flight dynamics using fully tuned radial basis function network

• Main Author: Pairan, Mohammad Fahmi
• Format: Thesis
• Language: English; English; English
• Published: 2018
• Subjects: TL500-777 Aeronautics. Aeronautical engineering
• Online Access: http://eprints.uthm.edu.my/216/1/24p%20MOHAMMAD%20FAHMI%20BIN%20PAIRAN.pdf; http://eprints.uthm.edu.my/216/2/MOHAMMAD%20FAHMI%20BIN%20PAIRAN%20COPYRIGHT%20DECLARATION.pdf; http://eprints.uthm.edu.my/216/3/MOHAMMAD%20FAHMI%20BIN%20PAIRAN%20WATERMARK.pdf; http://eprints.uthm.edu.my/216/

A quadcopter is a four-rotor unmanned aerial vehicle (UAV) with nonlinear and strongly coupled dynamics system. A precise dynamics model is important for developing a robust controller for a quadcopter. NN model capable to obtain the accurate dynamics model from actual data without having any govemmg mathematical model or priori assumptions. Recursive system identification based on neural network (NN) offers an alternative method for quadcopter dynamics modelling. Recursive learning algorithms, such as Constant Trace (CT) can be implemented to solve insufficient training data and over-fitting problems by developing a new model from real-time flight data in each time step. The modelling results from the NN model could be inaccurate due to inappropriate model structure selection, excessive number of hidden neurons and insufficient training data. Typically, the model structures and hidden neuron are determined by using trial and error approach to obtain the best network configuration. This study utilised a fully tuned radial basis function (RBF) neural network to obtain a minimal structure and avoid pre-determining the number of hidden neurons by introducing the adding and pruning neuron strategy. The prediction performance of the proposed fully tuned RBF was compared with Multilayer Perceptron (MLP), Hybrid Multilayer Perceptron (HMLP) and RBF networks trained with CT algorithm. The findings indicated that the fully tuned RBF with minimal resource allocating networks (MRAN) automatically selected seven neurons with 9.5177 % prediction accuracy and 5.89ms mean training time. The results also showed that the proposed extended minimal resource allocating networks (EMRAN) algorithm is capable to adapt with dynamics changes and infer quadcopter model with an even shorter training time (4.16ms) than MRAN and suitable for real-time system identification.