A New Separated Fault Estimator and Fault-Tolerant Control

Design Strategy for Uncertain Nonlinear Systems Using T-S Fuzzy

Modeling
Mohammad Cheraghiyan a, *
a
Department of Automation & Instrumentation Engineering, Petroleum University of Technology, Ahwaz, 6199171183, Iran
*
Corresponding author at: Department of Automation & Instrumentation Engineering, Petroleum University of Technology, Ahwaz,
6199171183, Iran
E-mail address: [email protected].

Abstract- In this paper, a new separated design of adaptive observer-based fault estimator (FE) and fault-
tolerant control (FTC) is introduced for non-linear systems along with sensor and actuator faults in the
presence of external disturbance and parametric modeling uncertainty by T-S fuzzy modeling, which can
easily challenge integrated design strategies. Firstly, a new fuzzy adaptive FE estimates actuator/sensor
faults and the system states simultaneously and also can automatically estimate and compensates bi-
directional uncertainties between FE function and control system. Secondly, a fuzzy state feedback FTC
controller is designed based on these estimations. Then the proposed separated FE/FTC design presented
which is solved by linear matrix inequality via 𝐻∞ optimization. Finally, a tutorial example of an inverted
pendulum on a cart demonstrates the validity and effectiveness of the proposed design.
Keywords: Separated FE/FTC; T-S fuzzy modeling; fuzzy adaptive fault estimation;

1. Introduction
Fault-Tolerant Control (FTC) is one of the research areas that has attracted a great deal of attention
in the two last decades. In the event of part and/or instrument failures, the FTC allows the actual output to
be retained close to the desired level [1]. A control system's capacity for accommodation depends on several
factors, such as the magnitude of the fault, nominal system robustness, and mechanisms that incorporate
redundancy in sensors and/or actuators. A bibliographic review of the FTC method is published in [2].
The literature defines two main classes of approaches from the viewpoint of FTC strategies: passive
and active. Passive FTC strategy is a control law that takes fault appearance into account as a system
disturbance [3]. The active FTC technology, however, is to adapt the control law using fault detection and
isolation (FDI) block information. using this information, some automatic changes are made trying to
achieve control objectives [4].
Many FDI method are developed in the last decades. FDI approaches can be classified into three
different categories, such as model-based, data-based, and hybrid methods [5],[6]. The model-based FDI
strategies are presented in [6],[7],[8] and can be categorized into basic approaches, such unknown input
observer (UIO), Kalman filter, algorithms that focused on optimization. Model-based observers are most
popular FDI structure. It not only allows the identification and isolation of different faults but also provides
information about the magnitude of the faults and their occurrence time and location.
The Takagi- Sugeno (T–S) fuzzy method has become increasingly interested because it is a
powerful solution to bridge the gap between linear control systems and nonlinear ones. T-S fuzzy modeling
has a significant advantage due to its universal approximation of any smooth nonlinear functions by mixing
of several local linear sub-models, which significantly makes it easy for the complex nonlinear system to
analyze and synthesize. There have been several important analytical and synthesis findings for the T–S
fuzzy modeling (see [9],[10]).
In real processes, in addition to the faults, there is also modeling uncertainty in the system and the
presence of uncertainty affects both the FE function and the control system, so it may lead to an FE-based
FTC with decreased performance when designed separately. The existence of uncertainty in the FE function
along with the uncertainty associated with the control system leads to the presence of bi-directional
robustness interactions between the FE and FTC. Such interactions lead to the need for a joint design of the
FE and FTC for robust overall system performance [11]. So, these so-called bi-directional robustness
interactions between FE and FTC have naturally lead to the need for integrated FE/FTC strategy [12].
The integrated FE/FTC design has much computation and even has its own limitation for more
complicated systems. So a need for a much simpler and more applicable design of FE and FTC in the
presence of uncertainty been felt. The main contribution of this paper is to propose a fuzzy adaptive
observer-based FE which capable of estimation and compensation these so-called bi-directional robustness
interactions, and by using it in a separated FTC and FE design, it can be a substitute for integrated FE/FTC
design.
The remainder of this paper is as follows, In Section 2 the system statement and preliminaries are
presented. Sections 3 and 4, respectively, introduce the fuzzy AAO and state feedback FTC designs. The
single-step integrated design of robust FE based FTC is provided in section 5. A simulation example of an
inverted nonlinear pendulum on a cart is presented in Section 6 to demonstrate the applicability of the
proposed design methods, and finally, the conclusion is given in Section 7.
The rest of this paper is as follows, In Section 2 the problem formulation is presented. Sections 3
and 4, respectively, introduce the fuzzy adaptive fault estimator and fault-tolerant control designs. then
separated design of FE and FTC is provided in section 5. In section 6 a comparison experiment is conducted
between the proposed separated and an integrated FE/FTC designs to illustrate the applicability of the
proposed design methods. In the end, a conclusion is also provided in section 7.

2. Problem Formulation
A nonlinear system influenced by unknown disturbance, actuator and sensor faults, and faced to
modeling uncertainty is written as:
𝑥̇ (𝑡) = ℎ𝑥 (𝑥(𝑡), 𝑢(𝑡), 𝑓𝑎 (𝑡), 𝑑(𝑡), 𝛿(𝑡)),

𝑦(𝑡) = ℎ𝑦 (𝑥(𝑡), 𝑓𝑠 (𝑡)), (1)

where 𝑥(𝑡) ∈ 𝑅𝑛 ,𝑢(𝑡) ∈ 𝑅𝑚 ,and 𝑦(𝑡) ∈ 𝑅𝑝 are the state of the system, control input, and output,
respectively. 𝑓𝑎 (𝑡) ∈ 𝑅𝑞 , 𝑓𝑠 (𝑡) ∈ 𝑅𝑞1 and 𝑑(𝑡) ∈ 𝑅𝑙 denote the actuator fault, sensor fault and unknown
external disturbance, respectively. 𝛿(𝑡) ∈ 𝑅𝑛 is the modeling uncertainty, ℎ𝑥 (. ) And ℎ𝑦 (. ) are continuous
nonlinear functions.
Consider a Takagi-Sugeno (T-S) fuzzy system, consisting of a set of If-Then rules derived from
system (1). The 𝑖𝑡ℎ rule of the fuzzy model is as follows:

Plant Rule 𝑖 (𝑖 = 1, … , ℎ)
 IF 𝜃1 is 𝜇𝑖1 and … and 𝜃𝑠 is 𝜇𝑖𝑠 , then

𝑥̇ (𝑡) = (𝐴𝑖 + 𝛥𝐴𝑖 )𝑥(𝑡) + 𝐵𝑖 𝑢(𝑡) + 𝐹𝑎𝑖 𝑓𝑎 (𝑡) + 𝐷𝑖 𝑑(𝑡)

{ (2)
𝑦(𝑡) = 𝐶𝑥(𝑡) + 𝐹𝑠 𝑓𝑠 (𝑡)

where 𝐴𝑖 , 𝐵𝑖 , 𝐹𝑎𝑖 , 𝐷𝑖 , 𝐶, 𝐹𝑠 are known real constant matrices with appropriate dimensions. Δ𝐴𝑖 ∈ 𝑅𝑛×𝑛
denotes for the bounded uncertainty matrix satisfying ‖Δ𝐴𝑖 ‖ ≤ 𝜀𝑖 . In general, 𝜃𝑗 (𝑗 = 1,2, … , 𝑠) denote
premise variables which can be some measurable variables of the system states or the function of states.
Moreover, 𝜇𝑖𝑗 are the fuzzy sets that are characterized by membership function, ℎ is the number of sub-
models and 𝑠 is the number of premise variables. The overall fuzzy model, being achieved from the plant
rules (2), can be described as:

𝑥̇ (𝑡) = ∑ℎ𝑖=1 𝜌𝑖 (𝜃(𝑡))[(𝐴𝑖 + 𝛥𝐴𝑖 )𝑥(𝑡) + 𝐵𝑖 𝑢(𝑡) + 𝐹𝑎𝑖 𝑓𝑎 (𝑡) + 𝐷𝑖 𝑑(𝑡)]

𝑦(𝑡) = 𝐶𝑥(𝑡) + 𝐹𝑠 𝑓𝑠 (𝑡), (3)

in which,
𝓃𝑖 (𝜃)
𝜌𝑖 (𝜃) = ∑ℎ , 𝓃𝑖 (𝜃) = ∏𝑠𝑗=1 𝓃𝑖𝑗 (𝜃𝑗 ),
𝑤=1 𝓃𝑤 (𝜃)

where 𝜌𝑖 (𝜃(𝑡)) are membership functions depending on the premise variable vector 𝜃(𝑡) = [𝜃1 , … , 𝜃𝑠 ] ,
𝓃𝑖𝑗 (𝜃𝑗 ) are grades of the membership of 𝜃𝑗 in the fuzzy sets 𝓃𝑖𝑗 , 0 ≤ 𝜌𝑖 (𝜃) ≤ 1 and ∑ℎ𝑖=1 𝜌𝑖 (𝜃) = 1.

Assumption 1. All sub-models (𝐴𝑖 , 𝐵𝑖 ) and (𝐴𝑖 , 𝐶) are controllable and observable, respectively and
actuator and sensor faults are observable, which means that the following rank conditions are satisfied:
𝐴 𝐹𝑎𝑖 𝐴 0
𝑟𝑎𝑛𝑘([ 𝑖 ]) = 𝑛 + 𝑞, 𝑟𝑎𝑛𝑘([ 𝑖 ]) = 𝑛 + 𝑞1 .
𝐶 0 𝐶 𝐹𝑠
Assumption 2 [12]. The actuator fault 𝑓𝑎 is matched i.e. 𝑟𝑎𝑛𝑘(𝐵𝑖 , 𝐹𝑎𝑖 ) = 𝑟𝑎𝑛𝑘(𝐵𝑖 ), 𝑖 = 1,2, … , ℎ.
Assumption 3. The external disturbance 𝑑(𝑡), actuator fault 𝑓𝑎 (𝑡) and sensor fault 𝑓𝑠 (𝑡) fulfill ‖𝑑(𝑡)‖ ≤ 𝑑0 ,
‖𝑓𝑎 (𝑡)‖ ≤ 𝜎𝑎 and ‖𝑓𝑠 (𝑡)‖ ≤ 𝜎𝑠 , respectively, and 𝜎𝑎 , 𝜎𝑠 and 𝑑0 are some unknown positive scalars.

Remark 1. Assumption 1 ensures that the 𝑖th sub-model are controllable and observable, and actuator fault
and sensor fault are observable also, and the adaptive observer method proposed in the paper can be used
to estimate them. Assumption 2 implies that the actuator fault can be compensated through control action.
Assumptions 3 guarantees that actuator/sensor faults and disturbance are bounded.

Lemma 1 [13]: Let 𝑥 and 𝑦 denote two constant vectors, then 𝑥 𝑇 𝑦 = 𝑡𝑟(𝑥𝑦 𝑇 ), where 𝑡𝑟(𝐺) is trace of the
constant matrix 𝐺.

3. Fuzzy Adaptive Fault Estimator Design

𝑇
Defining 𝑥̅ = [𝑥 𝑇 𝑓𝑎𝑇 𝑓𝑠𝑇 ]𝑇 and 𝑑̅ = [𝑑 𝑇 𝑓𝑎𝑇̇ 𝑓𝑠𝑇̇ ] ; The system (3) can be augmented into:

̅𝑖 𝑑̅ (𝑡)]
𝑥̅̇ (𝑡) = ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[(𝐴̅𝑖 + ∆𝐴̅𝑖 )𝑥̅ (𝑡) + 𝐵̅𝑖 𝑢(𝑡) + 𝐷

𝑦(𝑡) = 𝐶̅ 𝑥̅ (𝑡), (4)

where:
 𝐴𝑖 𝐹𝑎𝑖 0 ∆𝐴𝑖 0 0 𝐵𝑖 𝑑 0 0
𝐴̅𝑖 = [ 0 0 0] , ∆𝐴̅𝑖 = [ 0 0 0], 𝐵𝑖 = [ 0 ] , 𝐷𝑖 = [0
̅ ̅ 𝐼𝑞 0 ] , 𝐶̅ = [𝐶 0 𝐹𝑠 ].
0 0 0 0 0 0 0 0 0 𝐼𝑞1

The estimation of the augmented state vector of (4) can be attained by the fuzzy adaptive observer as
follows:

𝑥̅̂̇ (𝑡) = ∑ℎ𝑖=1 𝜌𝑖 (𝜃) [𝐴̅𝑖 𝑥̅̂ (𝑡) + ∆𝐴̂̅𝑖 𝑥̅̂ (𝑡) + 𝐵̅𝑖 𝑢(𝑡) + 𝐺𝑖 (𝑦(𝑡) − 𝑦̂(𝑡))]

𝑦̂(𝑡) = 𝐶̅ 𝑥̂̅ (𝑡), (5)

where 𝑥̅̂ (𝑡) ∈ 𝑅𝑛+𝑞+𝑞1 and 𝑦̂(𝑡) ∈ 𝑅𝑝 are the estimate of 𝑥̅ (𝑡) and observer output, respectively. ∆𝐴̂̅𝑖 (𝑡) is
the estimation of ∆𝐴̅𝑖 , and 𝐺𝑖 represents observer gain matrix with appropriate dimensions to be solved,
𝑖 = 1, … , ℎ.
State and output estimation errors are defined as follow:

𝑒̅ (𝑡) = 𝑥̅ (𝑡) − 𝑥̅̂ (𝑡) (6)

𝑒𝑦 (𝑡) = 𝑦(𝑡) − 𝑦̂(𝑡) = 𝐶̅ 𝑒̅(𝑡), (7)

By using (4) to (6), the error dynamics can be written as:

𝑒̅̇ (𝑡) = ∑ℎ𝑖=1 𝜌𝑖 (𝜃) [(𝐴̅𝑖 − 𝐺𝑖 𝐶̅ )𝑒̅ (𝑡) + ∆𝐴̅𝑖 𝑥̅ (𝑡) − ∆𝐴̂̅𝑖 (𝑡)𝑥̅̂ (𝑡) + 𝐷
̅𝑖 𝑑̅ (𝑡)]. (8)

Now assumed that ∆𝐴̅𝑖 and ∆𝐴̂̅𝑖 (𝑡) can be defined as:

∆𝐴̅𝑖 = 𝑁𝜑̅𝑖 (9a)

and

∆𝐴̂̅𝑖 (𝑡) = 𝑁𝜑̅̂𝑖 (𝑡) (9b)

where 𝑁 ∈ 𝑅(𝑛+𝑞+𝑞1)×𝑝 is a constant matrix, and 𝜑̅𝑖 ∈ 𝑅𝑝×(𝑛+𝑞+𝑞1) denotes an uncertain constant matrix,
satisfying ‖𝜑̅𝑖 ‖ ≤ 𝛽𝑖 and 𝜑̅̂𝑖 is the estimate of 𝜑̅𝑖 . Furthermore, Since ‖Δ𝐴𝑖 ‖ ≤ 𝜀𝑖 so it can deduce that
‖Δ𝐴̅𝑖 ‖ ≤ 𝜀𝑖 , with known positive scalar 𝜀𝑖 .

By substituting (9a) and (9b) into (8), the error dynamics can be rewritten as:
̅𝑖 𝑑̅ (𝑡)]
𝑒̅̇ (𝑡) = ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[(𝐴̅𝑖 − 𝐺𝑖 𝐶̅ )𝑒̅ (𝑡) + 𝑁𝜑̅𝑖 𝑥̅ (𝑡) − 𝑁𝜑̅̂𝑖 (𝑡)𝑥̅̂ (𝑡) + 𝐷

= ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝐴1𝑖 ̅𝑖 𝑑̅ (𝑡)],

̅ 𝑒̅ (𝑡) + 𝑁𝜑̅𝑖 𝑒̅ (𝑡) + 𝑁𝜑̅𝑒𝑖 (𝑡)𝑥̅̂ (𝑡) + 𝐷 (10)

̅ = (𝐴̅𝑖 − 𝐺𝑖 𝐶̅ ).
where 𝜑̅𝑒𝑖 (𝑡) = 𝜑̅𝑖 − 𝜑̅̂𝑖 (𝑡) and 𝐴1𝑖

4. Fault-Tolerant Control Design

An FTC is designed for the system (3) as:

𝑢 = ∑ℎ𝑖=1 𝜌𝑖 (𝜃)𝐾𝑖 𝑥̅̂ (11)

where 𝐾𝑖 = [𝐾𝑥𝑖 𝐾𝑓𝑖 0𝑚×𝑞1 ] with 𝐾𝑥𝑖 and 𝐾𝑓𝑖 the state-feedback control gains and the actuator fault
compensation gains, respectively. By putting (11) into (3) closed-loop system can be written:

𝑥̇ (𝑡) = ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝑗(𝜃)[(𝐴𝑖 + ∆𝐴𝑖 )𝑥(𝑡) − 𝐵𝑖 𝐾𝑗 𝑒̅ + 𝐵𝑖 𝐾𝑥𝑗 𝑥 + 𝐵𝑖 𝐾𝑓𝑖 𝑓𝑎 + 𝐹𝑎𝑖 𝑓𝑎 + 𝐷𝑖 𝑑]. (12)

Under Assumption 2, 𝐾𝑓𝑖 can be chosen as 𝐾𝑓𝑖 = −𝐵𝑖† 𝐹𝑖 and 𝐵𝑖 = 𝐹𝑎𝑖 , where 𝐵𝑖† defined as left
pseudo-inverse of 𝐵𝑖 . By putting 𝐾𝑓𝑖 into (12) the closed-loop system rewritten as:

𝑥̇ (𝑡) = ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃) 𝜌𝑗 (𝜃)[(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )𝑥(𝑡) + 𝐸𝑖𝑗 𝑒(𝑡) + ∆𝐴𝑖 𝑥(𝑡) + 𝐷𝑖 𝑑(𝑡)], (13)

where 𝐸𝑖𝑗 = −𝐵𝑖 𝐾𝑗 = [−𝐵𝑖 𝐾𝑥𝑗 𝐹𝑖 0].

∆𝐴𝑖 can be expressed as:

𝛥𝐴𝑖 = 𝑌1 𝛥𝐴̅𝑖 𝑌2 (14)

where 𝑌1 = [𝐼𝑛 0𝑛×(𝑞+𝑞1 ) ] and 𝑌2 = 𝑌1𝑇 . Substituting (14) and (9a) into (13), gives:

𝑥̇ (𝑡) = ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃) 𝜌𝑗 (𝜃)[(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )𝑥(𝑡) + 𝐸𝑖𝑗 𝑒(𝑡) + 𝑌1 𝑁𝜑̅𝑖 𝑌2 𝑥(𝑡) + 𝐷𝑖 𝑑(𝑡)]. (15)

5. Separated Design of FE/FTC

Uncertain +
Actuator Sensor
System -

Step 2: Step 1:

Controller Observer

Separated FE/FTC design

Fig. 1. The separated FE/FTC design framework

The separated design method is by first designing the fault estimator and then the FTC controller
(see Fig. 1). This separate FE/FTC design concept is accomplished based on the satisfactory of the
Separation Principle as a result of the proposed fuzzy adaptive observer-based FE, which reduces the bi-
directional robustness interactions. In this respect the error dynamics are:

𝑒̅̇ (𝑡) = ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝐴1𝑖 ̅𝑖 𝑑̅ (𝑡)]

̅ 𝑒̅ (𝑡) + 𝑁𝜑̅𝑖 𝑒̅ (𝑡) + 𝑁𝜑̅𝑒𝑖 (𝑡)𝑥̅̂ (𝑡) + 𝐷

𝑧𝑒 = 𝐶𝑒 𝑒̅ , (16)

where 𝑧𝑒 ∈ 𝑅𝑟2 is the measured output and 𝐶𝑒 is a known constant matrix of appropriate dimension. It is
assumed that the observer is already stable, i.e., 𝑒(𝑡) = 0, then the closed-loop system (15) becomes into:
 𝑥̇ (𝑡) = ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃) 𝜌𝑗 (𝜃)[(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )𝑥(𝑡) + 𝑌1 𝑁𝜑̅𝑖 𝑌2 𝑥(𝑡) + 𝐷𝑖 𝑑(𝑡)]

𝑧𝑥 = 𝐶𝑥 𝑥

𝑦𝑐 = 𝑦 − 𝐹𝑠 𝑓̂𝑠 (17)

where 𝑧𝑥 ∈ 𝑅𝑟3 , 𝑦𝑐 , 𝑓̂𝑠 and 𝐶𝑥 are the measured output, corrected system output, the sensor fault estimation,
and a known constant of appropriate dimension, respectively.
For separate determination of the observer and controller gain matrices, theorems 1 and 2 are
sufficient prerequisites.

Theorem 1: Given positive scalars 𝛾1 and 𝜀𝑖 , suppose Δ𝐴̅𝑖 satisfies (9) with ‖Δ𝐴̅𝑖 ‖ ≤ 𝜀𝑖 and ‖𝜑̅𝑖 ‖ ≤ 𝛽𝑖 ,
and the error dynamics (15) are stable with 𝐻∞ optimization ‖𝐺𝑧𝑒 𝑑̅ ‖ < 𝛾1 , if there exists a symmetric
∞
positive definite matrix 𝑋 ∈ 𝑅𝑛1 ×𝑛1 , any symmetric positive definite Γ𝑖 ∈ 𝑅𝑝×𝑝 and 𝑊1𝑖 ∈ 𝑅𝑛1 ×𝑝 , such
that:
1
𝛺𝑖 ̅𝑖
𝑋𝐷
[ 𝛾1 ]<0 (18)
∗ −𝐼(𝑛2 )

𝜑̅̂̇𝑖 (𝑡) = 𝜌𝑖 (𝜃)𝛤𝑖 𝐶̅ 𝑒̅ 𝑥̅̂ 𝑇 (𝑡) (19)

𝐶̅ 𝑇 = 𝑋𝑁 (20)

where Ω𝑖 = 𝑠𝑦𝑚(𝑋𝐴̅𝑖 ) − 𝑠𝑦𝑚(𝑊1𝑖 𝐶̅ ) + 2𝜀𝑖 𝑋 + 𝐶𝑒𝑇 𝐶𝑒 , 𝑛2 = 𝑙 + 𝑞 + 𝑞1 . Then observer gains are given
by: 𝐺𝑖 = 𝑋 −1 ∗ 𝑊1𝑖 , 𝑁 = 𝑋 −1 𝐶̅ 𝑇 , 𝑖 = 1,2, … , ℎ.

Proof: Consider a Lyapunov function 𝑉𝑒 = 𝑒̅ 𝑇 (𝑡)𝑋𝑒̅ (𝑡) + ∑ℎ𝑖=1 𝑡𝑟(𝜑̅𝑒𝑖

𝑇 (𝑡)Γ −1
𝑖 𝜑 ̅ 𝑒𝑖 (𝑡)). The time derivative
of 𝑉𝑒 is:

𝑉𝑒̇ = 𝑒̅̇ 𝑇 (𝑡)𝑋𝑒̅ (𝑡) + 𝑒̅ 𝑇 (𝑡)𝑋𝑒̅̇ (𝑡) + ∑ℎ𝑖=1 𝑡𝑟(𝜑̅̇𝑒𝑖

𝑇 (𝑡)𝛤 −1
𝑖 𝜑 ̅ 𝑒𝑖 (𝑡)) + ∑ℎ𝑖=1 𝑡𝑟(𝜑̅𝑒𝑖
𝑇 (𝑡)𝛤 −1 ̇ (𝑡))
𝑖 𝜑̅ 𝑒𝑖 (21)

As 𝜑̅̇𝑒𝑖 (𝑡) = 𝜑̅̇𝑖 − 𝜑̅̂̇𝑖 (𝑡) = −𝜑̅̂̇𝑖 (𝑡) then

𝑉𝑒̇ = ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝑒̅ 𝑇 (𝑡)(𝐴1𝑖

𝑇 ̅ )𝑒̅ (𝑡)] + 2 ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑒̅ 𝑇 (𝑡)𝑋𝑁𝜑̅𝑖 𝑒̅ (𝑡)) +
̅ 𝑋 + 𝑋𝐴1𝑖

𝑇 (𝑡)𝛤 −1 ̂̇ (𝑡))
̅𝑖 𝑑̅ (𝑡))) + 2 ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑒̅ 𝑇 (𝑡)𝑋𝑁𝜑̅𝑒𝑖 (𝑡)𝑥̅̂ (𝑡)) − 2 ∑ℎ𝑖=1 𝑡𝑟(𝜑̅𝑒𝑖
∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑠𝑦𝑚(𝑒̅ 𝑇 (𝑡)𝑋𝐷 𝑖 𝜑 ̅𝑖

(22)

If (19) and (20) hold, by Lemma 1, the sum of the last two terms of (22) will be zero. Thus, (22) is
reduced to:

𝑉𝑒̇ = ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝑒̅ 𝑇 (𝑡)(𝐴1𝑖

̅𝑇 𝑋 + 𝑋𝐴1𝑖 ̅𝑖 𝑑̅ (𝑡))) + 𝑌1𝑖
̅ )𝑒̅ (𝑡)] + ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑠𝑦𝑚(𝑒̅ 𝑇 (𝑡)𝑋𝐷 (23)

where:

𝑌1𝑖 = 2 ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑒̅ 𝑇 (𝑡)𝑋𝑁𝜑̅𝑖 𝑒̅ (𝑡)), (24)

Since ‖Δ𝐴̅𝑖 ‖ ≤ 𝜀𝑖 and (9a), the following can be written:

𝑌1𝑖 ≤ 2 ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑒̅ 𝑇 (𝑡)𝑋‖𝑁𝜑̅𝑖 ‖𝑒̅ (𝑡)) = 2 ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑒̅ 𝑇 (𝑡)𝑋‖𝛥𝐴̅𝑖 ‖𝑒̅ (𝑡)) ≤

2 ∑ℎ𝑖=1 𝜌𝑖 (𝜃)(𝑒̅ 𝑇 (𝑡)𝑋𝜀𝑖 𝑒̅ (𝑡)), (25)

By using (23) and (25):

𝑉𝑒̇ ≤ ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝑒̅ 𝑇 (𝑠𝑦𝑚(𝑋𝐴1𝑖 ̅𝑖 𝑑̅ )] − 𝑧𝑒𝑇 (𝑡)𝑧𝑒 (𝑡).

̅ ) + 2𝜀𝑖 𝑋 + 𝐶𝑒𝑇 𝐶𝑒 )𝑒̅ + 𝑠𝑦𝑚(𝑒̅ 𝑇 𝑋𝐷 (26)

The 𝐻∞ performance ‖𝐺𝑧𝑒 𝑑̅ ‖ ≤ 𝛾1 is represented by

∞
∞
𝐽 = ∫0 (𝑧𝑒𝑇 𝑧𝑒 − 𝛾12 𝑑̅ 𝑇 𝑑̅ )𝑑𝑡 < 0 (27)

Under zero initial conditions,

∞ ∞
𝐽 = ∫0 (𝑧𝑒𝑇 𝑧𝑒 − 𝛾12 𝑑̅ 𝑇 𝑑̅ + 𝑉𝑒̇ )𝑑𝑡 − (𝑉𝑒 (∞) − 𝑉𝑒 (0)) ≤ ∫0 (𝑧𝑒𝑇 𝑧𝑒 − 𝛾12 𝑑̅ 𝑇 𝑑̅ + 𝑉𝑒̇ )𝑑𝑡 < 0

Then, a sufficient condition for (27) is:

𝐽1 = 𝑧𝑒𝑇 𝑧𝑒 − 𝛾12 𝑑̅ 𝑇 𝑑̅ + 𝑉𝑒̇ < 0 (28)

By using (28):

̅ 1𝑖 ) + 2𝜀𝑖 𝑋 + 𝐶𝑇𝑒 𝐶𝑒 ) 𝑒̅ + 𝑠𝑦𝑚(𝑒̅ 𝑇 𝑋𝐷

𝐽1 ≤ ∑ℎ𝑖=1 𝜌𝑖 (𝜃) [𝑒̅ 𝑇 (𝑠𝑦𝑚(𝑋𝐴 ̅𝑇 ̅
̅ )] − 𝛾2 𝑑
̅ 𝑖𝑑
1 𝑑=

𝑇
̅ 1𝑖 ) + 2𝜀𝑖 𝑋 + 𝐶𝑇𝑒 𝐶𝑒 ) 𝑒̅ ] + ∑ℎ 𝜌 (𝜃) [− (𝛾 𝑑 1 ̅𝑇 ̅ 𝑇 𝑋𝑒̅ ) +
̅− 1 𝐷
∑ℎ𝑖=1 𝜌𝑖 (𝜃) [𝑒̅ 𝑇 (𝑠𝑦𝑚(𝑋𝐴 ̅
1 − 𝐷 𝑖 𝑋𝑒
̅ ) (𝛾1 𝑑
𝑖=1 𝑖 𝛾1 𝛾1 𝑖

1 𝑇 ̅ ̅𝑇
𝛾21
𝑒̅ 𝑋𝐷𝑖 𝐷𝑖 𝑋𝑒̅ ] <0 (29)

Then:

∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝑒̅ 𝑇 (𝑠𝑦𝑚(𝑋𝐴1𝑖

̅ ) + 2𝜀𝑖 𝑋 + 𝐶𝑒𝑇 𝐶𝑒 )𝑒̅ ] + ∑ℎ𝑖=1 𝜌𝑖 (𝜃)[𝑒̅ 𝑇 𝑋𝐷 ̅𝑖𝑇 𝑋𝑒̅ ] < 0
̅𝑖 𝐷 (30)

By applying the Schur complement to (30) and denoting 𝑊1𝑖 = 𝑋𝐺𝑖 , (18) will be achieved. ∎

Theorem 2 : Given positive scalars 𝛾2 and 𝜀𝑖 , suppose Δ𝐴̅𝑖 satisfies (8) with ‖Δ𝐴̅𝑖 ‖ ≤ 𝜀𝑖 and ‖𝜑̅𝑖 ‖ ≤ 𝛽𝑖 ,
and the closed-loop system (16) is stable with 𝐻∞ optimization ‖𝐺𝑧𝑥 𝑑 ‖ < 𝛾2 , if there exists a symmetric
∞
positive definite matrix 𝑄 ∈ 𝑅𝑛×𝑛 , and 𝑊2𝑗 ∈ 𝑅𝑚×𝑛 , such that:

𝛺𝑖𝑗 𝑄𝐶𝑥𝑇
[ ]<0 (31)
∗ −𝐼(𝑛)

1
where Ω𝑖𝑗 = 𝑠𝑦𝑚(𝐴𝑖 𝑄) + 𝑠𝑦𝑚(𝐵𝑖 𝑊2𝑗 ) + 𝑠𝑦𝑚(𝑌1 𝜀𝑖 𝑌2 𝑄) + (𝐷𝑖 𝐷𝑖𝑇 ). Then controller gains are given
𝛾22
by: 𝐾𝑥𝑗 = 𝑊2𝑗 ∗ 𝑄 −1 and 𝑖, 𝑗 = 1,2, … , ℎ.

Proof: Consider a Lyapunov function 𝑉𝑥 = 𝑥 𝑇 𝑃𝑥. The time derivative of 𝑉𝑥 is:

 𝑉𝑥̇ = 𝑥̇ 𝑇 𝑃𝑥 + 𝑥 𝑇 𝑃𝑥̇ (32)

By substituting (16) into (32):

𝑉𝑥̇ = ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃)[𝑥 𝑇 𝑠𝑦𝑚(𝑃(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 ))𝑥 + 𝑠𝑦𝑚(𝑥 𝑇 𝑃𝐷𝑖 𝑑) + 𝑌2𝑖 ] (33)

where:

𝑌2𝑖 = 𝑥 𝑇 𝑠𝑦𝑚(𝑃𝑌1 𝑁𝜑̅𝑖 𝑌2 )𝑥, (34)

Similarly, since ‖Δ𝐴̅𝑖 ‖ ≤ 𝜀𝑖 and (9a), the following can be written:

𝑌2𝑖 ≤ ∑ℎ𝑖=1 𝜌𝑖 (𝜃) 𝑥 𝑇 𝑠𝑦𝑚(𝑃𝑌1 ‖𝑁𝜑̅𝑖 ‖𝑌2 )𝑥 = ∑ℎ𝑖=1 𝜌𝑖 (𝜃) 𝑥 𝑇 𝑠𝑦𝑚(𝑃𝑌1 ‖𝛥𝐴̅𝑖 ‖𝑌2 )𝑥 ≤

∑ℎ𝑖=1 𝜌𝑖 (𝜃) 𝑥 𝑇 𝑠𝑦𝑚(𝑃𝑌1 𝜀𝑖 𝑌2 )𝑥, (35)

By using (33) and (35):

𝑉𝑥̇ ≤ ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃) [𝑥 𝑇 (𝑠𝑦𝑚 (𝑃(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )) + 𝑠𝑦𝑚(𝑃𝑌1 𝜀𝑖 𝑌2 ) + 𝐶𝑥𝑇 𝐶𝑥 ) 𝑥 +

𝑠𝑦𝑚(𝑥 𝑇 𝑃𝐷𝑖 𝑑)] − 𝑧𝑥𝑇 (𝑡)𝑧𝑥 (𝑡). (36)

The 𝐻∞ performance ‖𝐺𝑧𝑥 𝑑 ‖ ≤ 𝛾2 is represented by

∞
∞
𝐽= ∫0 (𝑧𝑥𝑇 𝑧𝑥 − 𝛾22 𝑑 𝑇 𝑑)𝑑𝑡 < 0. (37)

Under zero initial conditions,

∞ ∞
𝐽 = ∫0 (𝑧𝑥𝑇 𝑧𝑥 − 𝛾22 𝑑 𝑇 𝑑 + 𝑉𝑥̇ )𝑑𝑡 − (𝑉𝑥 (∞) − 𝑉𝑥 (0)) ≤ ∫0 (𝑧𝑥𝑇 𝑧𝑥 − 𝛾22 𝑑 𝑇 𝑑 + 𝑉𝑥̇ )𝑑𝑡 < 0

Then, a sufficient condition for (37) is

𝐽2 = 𝑧𝑥𝑇 𝑧𝑥 − 𝛾22 𝑑 𝑇 𝑑 + 𝑉𝑥̇ < 0, (38)

By using (38):

𝐽2 ≤ ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃) [𝑥 𝑇 (𝑠𝑦𝑚 (𝑃(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )) + 𝑠𝑦𝑚(𝑃𝑌1 𝜀𝑖 𝑌2 ) + 𝐶𝑥𝑇 𝐶𝑥 ) 𝑥 +

𝑠𝑦𝑚(𝑥 𝑇 𝑃𝐷𝑖 𝑑)] − 𝛾22 𝑑 𝑇 𝑑 = ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃) [𝑥 𝑇 (𝑠𝑦𝑚 (𝑃(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )) + 𝑠𝑦𝑚(𝑃𝑌1 𝜀𝑖 𝑌2 ) +

1 𝑇 1 1
𝐶𝑥𝑇 𝐶𝑥 ) 𝑥] + ∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃) [− (𝛾2 𝑑 − 𝐷𝑖𝑇 𝑃𝑥) (𝛾2 𝑑 − 𝐷𝑖𝑇 𝑃𝑥) + 𝑥 𝑇 𝑃𝐷𝑖 𝐷𝑖𝑇 𝑃𝑥] < 0
𝛾2 𝛾2 𝛾22

(39)

Then:
1
∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃)𝑥 𝑇 [𝑠𝑦𝑚 (𝑃(𝐴𝑖 + 𝐵𝑖 𝐾𝑥𝑗 )) + 𝑠𝑦𝑚(𝑃𝑌1 𝜀𝑖 𝑌2 ) + 𝐶𝑥𝑇 𝐶𝑥 + 𝑃𝐷𝑖 𝐷𝑖𝑇 𝑃] 𝑥 < 0 (40)
𝛾22
A congruence transformation with 𝑃−1 and denoting 𝑄 = 𝑃−1 and 𝑊2𝑗 = 𝐾𝑥𝑗 𝑄, turn (40) into:
1
∑ℎ𝑖=1 ∑ℎ𝑗=1 𝜌𝑖 (𝜃)𝜌𝑗 (𝜃)𝑥 𝑇 [𝑠𝑦𝑚(𝐴𝑖 𝑄 + 𝐵𝑖 𝑊2𝑗 ) + 𝑠𝑦𝑚(𝑌1 𝜀𝑖 𝑌2 𝑄) + 𝑄𝐶𝑥𝑇 𝐶𝑥 𝑄 + 𝐷𝑖 𝐷𝑖𝑇 ] 𝑥 < 0. (41)
𝛾22

By applying the Schur complement to (41), (31) will be achieved. ∎

6. Numerical Example
In this section, a tutorial example of an inverted pendulum on a cart borrowed from [14], has been
given in the nonlinear model as follows:
𝑥̇1 = 𝑥2

𝑔 𝑠𝑖𝑛(𝑥1 )−𝑎𝑚𝑙𝑥22 𝑠𝑖𝑛(2𝑥1 )⁄2−𝑎 𝑐𝑜𝑠(𝑥1 )𝑢

𝑥̇ 2 = 4𝑙⁄ −𝑎𝑚𝑙 𝑐𝑜𝑠 2 (𝑥 )
3 1

𝑦 = [𝑥1 𝑥2 ]𝑇 (41)

where 𝑥1 and 𝑥2 denote the pendulum angle (rad) from the vertical and the angular velocity (rad/s),
respectively. 𝑔, 𝑚, 𝑀 and 2𝑙 stand for gravity constant, pendulum mass, cart mass, and pendulum length,
respectively, 𝑢 denotes the applied force to the cart, and 𝑎 = 1⁄(𝑚 + 𝑀). In this study, parameters of the
model are chosen as follows, 𝑚 = 2.0 𝑘𝑔, 𝑀 = 8.0 𝑘𝑔, and 2𝑙 = 1.0 𝑚. The T-S fuzzy model derived
from (41), assumed to have sensor/actuator faults, external disturbance and modeling uncertainty. The
nonlinear pendulum system can be modeled by the following two-rule fuzzy system, which according to
[14], in the controllable region 𝑥1 ∈ (−𝜋/2, 𝜋/2) is valid.
2

𝑥̇ = ∑ 𝜌𝑖 (𝑥1 )[(𝐴𝑖 + ∆𝐴𝑖 )𝑥 + 𝐵𝑖 (𝑢 + 𝑓𝑎 ) + 𝐷𝑖 𝑑 ]

𝑖=1

𝑦 = 𝐶𝑥 + 𝐹𝑠 𝑓𝑠 , (42)
2 2
where 𝜌1 (𝑥1 ) = 1 − |𝑥1 |, 𝜌2 (𝑥1 ) = |𝑥1 |,
𝜋 𝜋

0 1 0 0 1 0
𝐴1 = [ 0], 𝐵1 = [− ], 𝐴2 = [ 0], 𝐵2 = [− ], 𝐶 = 𝐼2 ,
𝑔 𝑎 2𝑔 𝑎𝛽
4𝑙 ⁄3−𝑎𝑚𝑙 4𝑙 ⁄3−𝑎𝑚𝑙 𝜋(4𝑙 ⁄3−𝑎𝑚𝑙𝛽2 ) 4𝑙 ⁄3−𝑎𝑚𝑙𝛽2

0 0.1
𝐷1 = 𝐷2 = [ ], 𝐹 = [ ] and 𝛽 = 𝑐𝑜𝑠(88° ).
0.01 𝑠 0.3

The faults and disturbance are:

1, 0𝑠 ≤ 𝑡 ≤ 5𝑠 0.1, 0𝑠 ≤ 𝑡 ≤ 14𝑠
𝑓𝑎 = {𝑠𝑖𝑛(𝑡) , 5𝑠 < 𝑡 ≤ 20𝑠 , 𝑓𝑠 = {0.2, 14𝑠 < 𝑡 ≤ 23𝑠 , 𝑑 = 0.01 𝑠𝑖𝑛(10𝑡).
1, 20𝑠 < 𝑡 ≤ 30𝑠 0.1, 23𝑠 < 𝑡 ≤ 30𝑠
 To demonstration the effectiveness of the proposed separated FE/FTC design strategy, a
comparison with same initial angles is conducted between the integrated FE/FTC design mentioned in [12]
and the proposed separated design in this paper.
The proposed separated FE based FTC design for the system (42) is solved with the following
parameters: 𝜀𝑖 = 0.1, 𝛾1 = 1, 𝛾2 = 1, 𝐶𝑥 = 𝐼2 , 𝐶𝑒 = 0.1𝐼4 , Γ𝑖 = 𝐼2 and 𝑖 = 1,2. Solving Theorem 1 and 2
with the chosen parameters and by using the lmilab solver in Yalmip [15] of MATLAB, gives the following
observer and control gains:
𝐾𝑥1 = [4500.79 1533.8], 𝐾𝑥2 = [4493.86 1528.9],

−711.94 314.145 14.57 51.11 0.00293 0.01168

𝐺1 = [−2331.69 1020.92 ], 𝐺 = [ 50.237 157.94 ], 𝑁 = [ 0.00862 0.0366 ],
15532.47 −5311.51 2 539.5 10.576 0.0575 0.0258
6738.14 −2862.93 −109.64 −385.97 −0.0206 −0.09252

and the matrix gains for the proposed integrated Augmented State Unknown Input Observer (ASUIO)-
based FE/FTC areas the same mentioned in [12].
The initial conditions for the integrated ASUIO-based FE/FTC design are chosen 𝑍(0) =
0.1𝑜𝑛𝑒𝑠(7,1) and for the proposed separated FE/FTC design are chosen 𝑥̅̂ (0) = 0.1 ∗ 𝑜𝑛𝑒𝑠(4,1) , 𝜑̅̂𝑖 (0) =
0.1 ∗ 𝑜𝑛𝑒𝑠(2,4) (𝑖 = 1,2). It should be noted that the uncertainty is defined as parametric uncertainty on
the parameters 𝑚, 𝑀 and 𝑙 as follows:
̅ + 𝛿𝑀, and 𝑙 = 𝑙 ̅ + 𝛿𝑙 ,
̅ + 𝛿𝑚, 𝑀 = 𝑀
𝑚=𝑚

where 𝑚̅, 𝑀̅ and 𝑙 ̅ denote for nominal values of the pendulum mass, the cart mass and half of the pendulum
length, respectively, and 𝛿𝑚, 𝛿𝑀 𝑎𝑛𝑑 𝛿𝑙 denote for their corresponding variations, which are 𝛿𝑚 =
0.01cos(𝑡), 𝛿𝑀 = 0.01sin(𝑡) and 𝛿𝑙 = 0.01cos(𝑡).

A comparison has been made with the same initial state conditions (𝑥(0) = [18° , 0]). From Figs.
2-5, it has been observed that the proposed separated FE based FTC design has better performance and
covers a wider range of initial angles than the one proposed in [12].
 Fig. 2. Comparison of states between separated and integrated FE/FTC designs

Fig. 3. Comparison of actuator and sensor faults estimation errors between separated and integrated
FE/FTC designs
 Fig. 4. Comparison of states estimation errors between separated and integrated FE/FTC designs

Fig. 5. Comparison of control signals between separated and integrated FE/FTC designs

In the face of both faults, unknown external disturbance, and with the same uncertainties, the
proposed separated FE based FTC design able to balance the pendulum angle in range |𝑥1 (0)| ≤ 39.3 deg
(𝑥2 (0) = 0). Whilst the proposed Integrated ASUIO-based FE/FTC in [12] only able to balance the
pendulum angle in range |𝑥1 (0)| ≤ 28.8 deg (𝑥2 (0) = 0).

7. Conclusion
Observer and control systems influence each other in the face of uncertainty, and cannot be
designed independently. The presence of uncertainty also restricts applications of the Separation Principle.
The key reason for using the Integrated Fault Estimator (FE) and Fault-Tolerant Control (FTC) framework
is these effects, which are called bi-directional robustness interactions. In this paper, a new separated design
of adaptive observer-based FE and FTC is introduced for non-linear systems along with sensor and actuator
faults in the presence of external disturbance and parametric modeling uncertainty by T-S fuzzy modeling,
which can easily challenge integrated design strategies.
The proposed fuzzy adaptive FE estimates actuator/sensor faults and the system states
simultaneously and also can automatically estimate and compensates the bi-directional uncertainties
between FE function and control system. Hence, by successfully reducing the bi-directional robustness
interactions it resulting in a wider control range and much better performance. Then a fuzzy state feedback
FTC controller is designed based on these estimations. The proposed separated design has less complexity
than the integrated ones and shows a better performance in the face of faults which makes it more feasible
for practical applications.
 This method in the future can be extended to large-scale interconnected systems, multi-agent
systems, networked systems, and even more complicated systems.

References
[1] M. Blanke, R. Izadi-Zamanabadi, S. Bogh, and C. P. J. C. E. P. Lunau, "Fault-tolerant control
systems-a holistic view," vol. 5, no. 5, pp. 693-708, 1997.
[2] Y. Zhang and J. Jiang, "Bibliographical review on reconfigurable fault-tolerant control systems,"
Annual reviews in control, vol. 32, no. 2, pp. 229-252, 2008.
[3] Z. Qu, C. M. Ihlefeld, Y. Jin, and A. J. A. Saengdeejing, "Robust fault-tolerant self-recovering
control of nonlinear uncertain systems," vol. 39, no. 10, pp. 1763-1771, 2003.
[4] M. Blanke, M. Kinnaert, J. Lunze, M. Staroswiecki, and J. Schröder, Diagnosis and fault-tolerant
control. Springer, 2006.
[5] X. Zhang, L. Tang, and J. J. I. T. o. C. S. T. Decastro, "Robust fault diagnosis of aircraft engines:
A nonlinear adaptive estimation-based approach," vol. 21, no. 3, pp. 861-868, 2012.
[6] B. Pourbabaee, N. Meskin, and K. J. I. T. o. C. S. T. Khorasani, "Sensor fault detection, isolation,
and identification using multiple-model-based hybrid Kalman filter for gas turbine engines," vol.
24, no. 4, pp. 1184-1200, 2015.
[7] V. Venkatasubramanian, R. Rengaswamy, K. Yin, S. N. J. C. Kavuri, and c. engineering, "A review
of process fault detection and diagnosis: Part I: Quantitative model-based methods," vol. 27, no. 3,
pp. 293-311, 2003.
[8] W.-M. Lin, T.-C. J. I. g. Ou, transmission, and distribution, "Unbalanced distribution network fault
analysis with hybrid compensation," vol. 5, no. 1, pp. 92-100, 2011.
[9] H. R. Patel and V. A. J. E. Shah, "Stable Fault Tolerant Controller Design for Takagi–Sugeno Fuzzy
Model-Based Control Systems via Linear Matrix Inequalities: Three Conical Tank Case Study,"
vol. 12, no. 11, p. 2221, 2019.
[10] W. Ji, J. Qiu, and H. R. J. I. T. o. F. S. Karimi, "Fuzzy-model-based output feedback sliding mode
control for discrete-time uncertain nonlinear systems," 2019.
[11] J. Lan and R. J. Patton, "A new strategy for integration of fault estimation within fault-tolerant
control," Automatica, vol. 69, pp. 48-59, 2016.
[12] J. Lan and R. J. Patton, "Integrated Design of Fault-Tolerant Control for Nonlinear Systems Based
on Fault Estimation and T–S Fuzzy Modeling," IEEE Transactions on Fuzzy Systems, vol. 25, no.
5, pp. 1141-1154, 2017.
[13] T. V. Dang, W. J. Wang, L. Luoh, and C. H. Sun, "Adaptive observer design for the uncertain
Takagi–Sugeno fuzzy system with output disturbance," IET control theory & applications, vol. 6,
no. 10, pp. 1351-1366, 2012.
[14] H. O. Wang, K. Tanaka, and M. F. Griffin, "An approach to fuzzy control of nonlinear systems:
Stability and design issues," IEEE transactions on fuzzy systems, vol. 4, no. 1, pp. 14-23, 1996.
[15] J. Lofberg, "YALMIP : a toolbox for modeling and optimization in MATLAB," pp. 284-289.