Professional Documents
Culture Documents
Comparisons of State Estimation Methods For A Quadrotor Attitude Model
Comparisons of State Estimation Methods For A Quadrotor Attitude Model
Model
Terefe Bayisa1, Yohannes Bekuma2, Rituraj Jain3
Department of Electrical and Computer Engineering, Wollega University
1.2.3
1
terefebayisa2018@gmail.com, 2yohbek@gmail.com, 3jainrituraj@yahoo.com
Abstract— Quadrotors have a wide range of applications in concluded from the estimation-accuracy comparison that the
all walks of life in recent decades, so their attitude control resulting state estimation from the proposed one is
accuracy is of high demand. Due to their good applicability comparable to those from the Kalman filtering and
to nonlinear multivariable complex control objectives, smoothing.
numerous advanced state-feedback controller design
methods have been applied in the quadrotor attitude control. Keywords—Quadrotor, Attitude model, Subspace
However, the traditional state observer becomes obsolete identification-based state estimation, Kalman filtering,
due to the presence of disturbing noises in practice. Without Kalman smoothing.
an accurate model, the Kalman filter and smoother are even
inapplicable. This thesis will propose a subspace 1. Introduction
identification-based state estimation method suited to the State estimation is one of the fundamental problems in control
state-variable estimation for the attitude model in a systems in which the true states of the system are estimated
quadrotor attitude control system in the absence of models based on the measured data and knowledge about the system.
but with disturbing noises. Subspace identification-based state estimation method has
In terms of the existing research results on the quadrotor attracted much attention during the past few years because it is
attitude model, this thesis will take the yaw-axis attitude simple in parametrization, the compact model is minimal
control loop as an example to carry out the following realization, no linear optimization technique is required and it
research work with MATLAB. is Kalman filter framework [1]. This interest is raised to
Above all, an observer-based discrete state feedback control provide accurate state-space models for multivariable linear
system is designed starting from attitude control systems directly from input-output data. The method is a
performances given by users. In order to obtain high- realization-based technique. In this research, using the 4SID
accuracy quadrotor attitude model and state estimation, a method is very important to estimate the state parameters
sufficiently excited attitude-angle reference input is using subspace algorithms [2,3].
designed as a pseudo-random binary series (PRBS) using A quadrotor is a highly linear, multi-variable, strongly
the settling time and the maximum operation frequency coupled, and underactuated system since it has three degrees
figured out from closed-loop pre-experiments. Then, the of freedom (pitch, roll, and yaw) and only three control inputs
control quantity and the real attitude angle being input and (pitching, rolling, and yaw moments) [4,5]. It is a type of
an output for the quadrotor attitude model, respectively, can unmanned aerial vehicle which is lifted by a set of four rotors
be collected as outputs. These experimental data pave the [6,7]. It has two pairs of identical fixed propellers when two
way for the quadrotor attitude model identification and its of them rotate clockwise and the other two propellers rotate
associated state estimation. counterclockwise [8,9,].
Next, this project proposes a subspace identification-based The first quadrotor attitude model is estimated and in
state estimation method to deal with the yaw-axis for consideration of input and output noise in the true
quadrotor attitude model state estimation problem under the environment, the quadrotor closed-loop error in variable (EIV)
conditions that the measurements for the control quantity system is introduced. Then using pseudo random binary
and the output attitude are corrupted with disturbing noises sequence (PRBS) as the reference input to collect input and
and that there are no available models. Firstly, a high-order output data and identify a linear model of the quadrotor three
coprime factor (CF) model for the quadrotor attitude model attitude angles [10,11]. But, in this thesis attitude model state
is identified by using 4S identification method. An estimation of a quadrotor using subspace identification-based
adaptable CF model for the quadrotor attitude model can state estimation is very interesting due to the ability to provide
then be obtained by applying the balanced stochastic model accurate state-space models for multivariable linear systems
truncation (BSMT) reduction method according to the directly from input-output data.
mechanical quadrotor attitude model structure. At last, the The main contribution of this paper is to extract the state using
estimation states for the quadrotor attitude model can be a subspace identification based state estimation method to
given by similarity transform and the relations between the estimate the state of the quadrotor in consideration of input
quadrotor attitude model and its CF model. and output noise in the true environment and compare the
Finally, on the basis of the identified CF model for the estimated states with the estimated states of Kalman filtering
quadrotor attitude model from the proposed state estimation and Kalman smoothing.
method, the Kalman filter and soother are applicable and
adopted to estimate the quadrotor attitude model states. It is 2. State-space quadrotor generalized attitude model
A quadrotor state-space representation is a mathematical The amount of change in angle during maneuvering , and
model of a physical system as a set of input, output, and state
variables related to first-order differential equations. State- is always very small and it is nearly zero. Finally, the
space refers to space whose axes are the state variables. The three quadrotor attitude angle models can be obtained as [12].
state of the system can be represented as a vector within that
space. This research, start by selecting the system states with M2
the three degrees of freedom. =
Let's consider the inputs that can be applied to the quadrotor J xx
attitude model to control the behavior of the quadrotor. There
M 3
are four propellers and three degrees of freedom. Commonly, (3)
J yy
the control inputs that will be detailed here are considered for
the vertical thrust across yaw and of the quadrotor [12]. M
4
J zz
3.1 Zero-order hold discretization Figure 3: A closed-loop EIV system with an observer-based state
feedback
Discretization is the process of transferring continuous-time
state-space models into discrete-time state-space models and it
is very important in this research because the quadrotor state- where r is the reference input angle for the quadrotor, and
space model identified is in continuous state-space form. ( k ) is real attitude angle of the quadrotor to be estimated,
Hence, transferring the continuous state-space form of the
quadrotor attitude model into discrete counterparts is crucial. u4 (k ) is the control quantity for yaw axis, xˆ (k ) is denoting the
This process is usually carried out as a first step toward estimated state, and the disturbance noises in the input and
making them suitable for observer design stated in the
output can then be given as m (k ) (k ) n (k ) and x(k 1) A1 x( k ) B1 r (k ) e( k )
u (k ) u4 ( k ) nu (k ) in which nu (k ) and n (k ) are
u (k ) nu (k ) (13)
(k ) C x ( k ) D ( k )
the measurement noises disturbing on the output u ( k ) and 1 1 r
n (k )
m
m (k ) respectively, ( A, B, C , D)
is a continuous state-space
model for the quadrotor GAM and ( F , G, C , D) is the discrete T
where the covariance for e( k ) and nu n are W and
state-space model of the quadrotor based on which one can
design a feedback gain K . R. The matrices C1 and D1 are partitioned as
The main objective of this paper is to estimate state xˆ (k ) ,
using the subspace identification-based state estimation
method in the available measurement noises in the inputs and C11
C1
outputs. C12
Subspace identification-based state estimation, Kalman
filtering, and Kalman smoothing will be applied. D
Coprime factorization is a technique that makes a control D1 11
system ( F , G, C , D) and factorizes it into a product of two D12
matrices. Coprime factorization is used to describe systems
and applied in controller synthesis for a given dynamic system 3.4 Derivation of estimated state
with uncertainties [13]. A state variable representation of a system is not unique.
Then the CF description for quadrotor attitude model can be There are infinitely many representations. Methods for
given by the designed state feedback law transforming one set of state variables to another is discussed
below. The way to transfer from one internal representation to
another is called a similarity transformation [13]. To perform
u4 (k ) r (k ) Kx(k ) (10) the transformation P , two different sets of state-space model
Based upon the figure 3, we can have a reference input angle can be given as
and
(11) z (k 1) Az (k ) Bu
(k )
(k ) Du
(k )
(15)
y (k ) Cz
Hence, the real quadrotor generalized state-space model can
be given as Let’s have
z ( k ) Px(k )
A0 F GK z ( k 1) Px ((k 1)
(k ) Bu
(k )
B0 G z (k 1) P ( Ax
C DK 1 z (k ) Bu
(k )
z (k 1) P( AP
C0 (12)
K 1 z (k ) PBu
z (k 1) PAP (k )
D
D0 (k ) Du
y (k ) Cx (k )
Iq 1 z (k ) Du (k )
y (k ) CP
3.3 A quadrotor attitude model subspace identification
Coprime factor model can be identified from the observer-
The transformed matrices ( A, B , C , D ) of the new
based feedback control system in figure 3 by experimental
measurements contaminating with output noises as well as transformed state-space model can be given as
state noise
1
A PAP A1 Pc Pc A1T B1 B1T 0
B PB Bw Pc C1T B1 D1T
(16)
1
C CP QA1 A1T Q (QBw C1T )( D1 D1T )(QBw C1T )T 0
D D 2. The state-space model in step 1 and transform and
partition of the state space model transformation to a
These equations represent the same dynamical system for any balanced realization { Aˆ1 , Bˆ1 , Cˆ1 , Dˆ1 }if and only if it is
nonsingular matrix P , and hence, can regard these asymptotically stable and minimal.
representations as equivalent. It is easy to see that the
input/output transfer matrix is not changed under the 3. Obtain subsystem realization ( Aˆ1 , Bˆ1 , Cˆ1 , D
ˆ ) via BSMT.
1
coordinate transformation. Then the realization ( A1 , B1 , C1 , D1 ) is balanced and
It be can simply possible to find the linear transformation stable.
matrix P as 4. Obtain the reduced-order models using
C1 P 1 C0
P C0 1C1 (17)
x2(k+1)
x2(k)
measurements of the system input and outputs. The observer x1(k+1) x1(k)
x 1(degree)
the characteristic equation can be given as 0
-1 xr
* ( z ) ( z 0.7408)( z 0.2231) -2
xe
z 2 0.9639 z 0.1653
Time(sec.)
0.02
x 2(degree/sec)
0.01
z 0 1 T T / 3 3
-0.01 xr
0 z 0 1 T
1.5 2
Time(sec.)
2.5 3
Thus, equating the state feedback gain matrix for this case can Figure 8: Subspace identification based state estimation result .
be obtained as K [201.3505 93.5376]
Based on the simulation result obtained in figure 8, the
On the basis of the discrete state estimation from it thus is
possible to realize the quadrotor attitude model state feedback estimated state xe1 , xe 2 is very nearest to the true state,
control system using MATLAB Simulink shown in figure 6 x1 , x2 . Hence subspace identification based state estimation
below. In the controller synthesis, the designed control gains
are connected in closed-loop feedback by giving step input as method is feasible and satisfactory.
shown in the following. After, the gain K is designed for the
quadrotor attitude model, the state feedback control system 7 Application of the Kalman smoothing
model can be designed as shown in the following figure. Generally, to do state estimation for the quadrotor attitude
model using Kalman smoothing, the model is important.
Hence, I have used the model estimated using the subspace
nu_psi
n_psi identification method and apply Kalman smoothing to
u_4 Psi 1/4.9665*1e3
estimate quadrotor attitude model states. To estimate the
K_n
quadrotor attitude model state using Kaman smoothing, the
model estimated using the subspace identification method is
u_4
Psi
xh2
1
x1(degree)
0.02
x 2(degree/sec)
0.01
-0.01 xr
xe s
-0.02
1.5 2 2.5 3
Time(sec.)
1
4.28668 10 7
x1(degree)
0
0.000548678 0.034946433
-1 xr xef 2 1.21346 10 6
xe f
-2
0.000130151 1.499632937
xes1 10 6
1.5 2 2.5 3
Time(sec.)
1.10935
0.02
0.000564674
xes 2 1.64224 10 6 0.035965206
x2(degree/sec)
0.01
-0.01 xr
-0.02
xe f
In terms of the computational load as can be seen from the
simulation results obtained in Table 1, subspace identification-
1.5 2 2.5 3
Time(sec.)
Figure 10: Subspace identification based state estimation result. based state estimation method is better than Kalman filtering,
In figure 10 are the true states, are and Kalman smoothing. Because the result J1 shows that
estimated states for Kalman filtering. The results shows that subspace identification-based state estimation is more
Kalman filtering estimated state and the true state are very preferable. But it is not obvious for the norms J2 and J3 and
nearest to each other. Hence, Kalman filtering is feasible and subspace identification-based state estimation method has less
satisfactory to estimate the quadrotor attitude model state. computational load. Hence, it can be concluded that, subspace
9. State estimation comparison identification-based state estimation method is more
Lastly, using different norms I compare the estimated state preferable than Kalman filtering and Kalman smoothing.
and the true state. In this research, I have used three different
norms to compare the true state with the estimated state.
9 Conclusion
The true state is very nearest to the estimated state using the
N 2
subspace identification-based state estimation method.
J1 x(k ) xˆ(k )
k 1
(30) Comparing the true state with the estimated states for
subspace identification-based state estimation method and
J 2 max x(k ) xˆ (k ) (31) Kalman filtering Kalman smoothing estimation method, the
k
estimated states are very nearest to the true states subspace
x(k ) xˆ ( k )
identification-based state estimation method shows feasible
J 3 max (32) result. Hence, subspace identification-based state estimation
k
xˆ ( k )
method is the most satisfactory and feasible method than
Kalman filtering and Kalman smoothing
From the estimated coprime factor, quadrotor state-space
where x( k ) is the true state, xˆ( k ) is estimated states and, model, the estimation states for quadrotor attitude model can
J1 , J 2 , J 3 are norms. be derived. The newly proposed subspace identification
Using the above norms, the comparison of estimated states method is very influential in state estimation. As I make a
with true states after long computation is listed in the table 1. rough analysis for a computational load for subspace
From the state estimation comparison of subspace identification method, Kalman filtering, and Kalman
identification-based state estimation method, Kalman filtering, smoothing, subspace identification-based state estimation
and Kalman smoothing estimation result is clearly shown in method is the most preferable. As we have seen the
the table above, subspace identification based estimated states, comparative assessment of state estimation using different
Kalman filtered estimated states and Kalman smoothed norms in Tab.1 subspace identification-based state estimation
estimated states have different norms. Based on the results method is better than Kalman filtering and Kalman
obtained, the subspace identification-based state estimation smoothing.
method and Kalman filtering show better estimation results. The main contribution of this research is to extract the state
Subspace identification-based state estimation method shows based on subspace identification methods. Because the state
a better estimation accuracy than Kalman filtering and Kaman estimation methods using Kalman filtering and Kalman
smoothing. smoothing will fail without a model. Their identical models
are both borrowed from the subspace identification-based
state estimation method to apply Kalman filtering and Kalman [12] Jin-Cang Liu, Research on Closed Loop Identification
smoothing. Because Kalman filtering and Kalman smoothing Method for a Generalized Attitude Model. Dissertation,
are inapplicable without a model. Master Degree Thesis, Tianjin University of Technology and
The estimation comparison in Table 1 indicates that the Education,2020.
subspace identification-based state estimation method is [13] Kemin Zhou, John C. Doyle, and Keith Glover. 1996.
fitting with the true states. Hence, the subspace identification- Robust and optimal control. Prentice-Hall, Inc., USA.
based state estimation method is better than Kalman filtering [14] Viberg, M. Subspace-based state-space system
and Kalman smoothing. identification. Circuits Systems and Signal Process 21, 23–37
(2002). https://doi.org/10.1007/BF01211649
Reference
[1] Yolanda Bolea, Nicolas Chefdor, Antoni Grau, "MIMO
LPV State-Space Identification of Open-Flow Irrigation Canal
Systems", Mathematical Problems in Engineering, vol. 2012,
Article ID 948936, 16 pages, 2012.
https://doi.org/10.1155/2012/948936
[2] Peter Van Overschee and Bart De Moor. 1994. N4SID:
subspace algorithms for the identification of combined
deterministic-stochastic systems. Automatica 30, 1 (Jan.
1994), 75–93. DOI:https://doi.org/10.1016/0005-
1098(94)90230-5
[3] Tanelli, Mara & Ardagna, Danilo & Lovera, Marco.
(2011). Identification of LPV State Space Models for
Autonomic Web Service Systems. Control Systems
Technology, IEEE Transactions on. 19. 93 - 103.
10.1109/TCST.2010.2063250.
[4] S. Jackson, J. Tisdale, M. Kamgarpour, B. Basso and J. K.
Hedrick, "Tracking controllers for small UAVs with wind
disturbances: Theory and flight results," 2008 47th IEEE
Conference on Decision and Control, 2008, pp. 564-569, doi:
10.1109/CDC.2008.4739415.
[5] T. Bresciani, Modelling, Identification, and Control of a
Quadrotor Helicopter, Master‟s thesis, Lund University,
Sweden, 2008.
[6] M. Bergamasco and M. Lovera, "Identification of Linear
Models for the Dynamics of a Hovering Quadrotor," in IEEE
Transactions on Control Systems Technology, vol. 22, no. 5,
pp. 1696-1707, Sept. 2014, doi:
10.1109/TCST.2014.2299555.
[7] Liang, Wenjing, "Attitude Estimation of Quadcopter
through Extended Kalman Filter" (2017). Theses and
Dissertations. 2685.
http://preserve.lehigh.edu/etd/26
[8] Dong, Wei & Gu, Guoying & Zhu, Xiangyang & Ding,
Han. (2014). High-Performance Trajectory Tracking Control
of a Quadrotor with Disturbance Observer. Sensors and
Actuators A: Physical. 211. 10.1016/j.sna.2014.03.011.
[9] S. Bouabdallah and R. Siegwart, "Full control of a
quadrotor," 2007 IEEE/RSJ International Conference on
Intelligent Robots and Systems, 2007, pp. 153-158, doi:
10.1109/IROS.2007.4399042.
[10 Kemin Zhou. 1993. Frequency-weighted model reduction
with L∞ error bounds. Syst. Control Lett. 21, 2 (Aug.
1993), 115–125. DOI:https://doi.org/10.1016/0167-
6911(93)90114-L
[11] Evensen, G. The Ensemble Kalman Filter: theoretical
formulation and practical implementation. Ocean Dynamics
53, 343–367 (2003). https://doi.org/10.1007/s10236-003-
0036-9