The Development of Neural Network Based System Identification ...

186 CHAPTER 7 FLIGHT CONTROL SYSTEM DESIGN: RESULTS AND DISCUSSION
• The mathematical calculations that can be done without iterations or involve a
predetermined matrix size are placed outside processing loops as much as possible
to reduce the computation overhead.
• The debugging option is disabled in the program to reduce overhead.
• Input constraints are only imposed on the first sample of the control horizon to
reduce computational efforts. There is no significant performance improvement
obtained from imposing constraints on all control sequence as confirmed in Wang
[2009a].
7.3 EXPERIMENTAL RESULTS
This section presents the control experimental results for the unmanned helicopter
system. The first part of the experiments is conducted to identify and demonstrate
the effect of tuning parameters on the NNAPC flight controller performance. The
tuning parameters are identified only in the pitch, roll and yaw channels. The altitude
controller uses the same tuning parameters obtained from the pitch, roll and yaw
controller experiment. The tuning parameters comparison is done for control penalty
factor r w and prediction horizon length N p . The control penalty factor r w is determined
first in the experiment and the best r w value is then used for the N p effect experiment.
Table 7.2 lists the specification of the HMLP networks used in the implementation of the
NNAPC controllers. The HMLP network is selected to model the helicopter’s dynamics
due to the network capability to model the dynamics with fewer weight connections
compared with the standard MLP network.
Table 7.2
The HMLP network parameters used in the NNAPC controller implementation.
Network
Dynamic Channels
Parameters Coupled Roll-Pitch Yaw Altitude
Output(s) φ, θ (rad) r (rad) w (m/s)
Regression Vector, ϕ φ, θ, δ lat , δ lon r, δ ped w, δ col
Number of Past Outputs 2 2 2
Number of Past Inputs 1 1 1
Number of Hidden Neurons 2 2 2