Abstract
The last chapter has discussed adaptive control design methods for switched nonlinear systems with uncertainties.
Access provided by CONRICYT-eBooks. Download chapter PDF
Keywords
- Stochastic Nonlinear System
- Stochastic Linear System
- Adaptive Neural Control
- Unknown Nonlinear Function
- Euclidean Vector Norm
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.
5.1 Background and Motivation
The last chapter discussed adaptive control design methods for switched nonlinear systems with uncertainties. However, the system structures considered in the last chapter are somewhat simple, which greatly limits the applications of the results in practice.
It is well known that stochastic disturbance is inevitably encountered in practical systems. Therefore, control of stochastic systems with or without switching has become an active research field and received much attention recently, see, e.g., [1–5] and the references therein. The authors in [6] considered global stabilization for high-order stochastic nonlinear systems with stochastic integral input-to-state stability inverse dynamics. The moment stability and sample path stability of switched stochastic linear systems were investigated in [7]. In [8] dissipativity-based sliding mode control for switched stochastic linear systems was adopted. Stabilization problems for stochastic nonlinear systems with Markovian switching were studied in [9]. The \(p^{th}\) moment exponential stability and global asymptotic stability in probability for a class of switched stochastic nonlinear retarded systems with asynchronous switching were solved in [10].
Moreover, dead-zone characteristics are encountered in many physical components of control systems. They are particularly common in actuators, such as hydraulic servovalves and electric servomotors. They also appear in biomedical systems. The system model is more realistic and reliable when the dead-zone nonlinearities are taken into consideration.
On the other hand, since the input-to-state stability (ISS) property was proposed in [11], it has rapidly become an important tool to investigate the stability problem of nonlinear systems. In view of the crucial importance of ISS, it is natural to introduce this concept to switched nonlinear systems. In this chapter, we consider some control problems of switched high-order nonlinear systems. Some complex dynamics such as stochastic disturbances, uncertainties, dead-zone nonlinearities and input-to-state stability inverse dynamics are considered in the systems under investigations. The considered mathematical models can provide a good description of a large number of practical switched nonlinear systems.
Notation \(\mathbb {R}\) denotes the n-dimensional space, \(\mathbb {R}^{n}\) is the set of all nonnegative real numbers. \(\mathscr {C}^{i}\) stands for a set of functions with continuous \(i^{th}\) partial derivatives. For a given matrix A (or vector v), \(A^{T}\) (or \(v^{T}\)) denotes its transpose, and \(Tr\{A\}\) denotes its trace when A is a square. \(\mathscr {K}\) represents the set of functions: \(\mathbb {R}^{+}\rightarrow \mathbb {R}^{+}\), which are continuous, strictly increasing and vanishing at zero; \(\mathscr {K}_{\infty }\) denotes a set of functions that is of class \(\mathscr {K}\) and unbounded. In addition, \(\left\| \cdot \right\| \) refers to the Euclidean vector norm. \(\mathbb {R}\) denotes the n-dimensional space, \(\mathbb {R}^{+}\) denotes the set of all nonnegative real numbers, and \(\mathbb {R}^{*}=\{q\in \mathbb {R}^{+} :q\ge 1\) is an odd integer\(\}\). \(\mathscr {C}^{i}\) denotes a set of all functions with continuous \(i^{th}\) partial derivatives. For a given matrix A (or vector v), \(A^{T}\) (or \(v^{T}\)) denotes its transpose, and \(Tr\{A\}\) denotes its trace when A is a square. \(\mathscr {K}\) denotes the set of all functions: \(\mathbb {R}^{+}\rightarrow \mathbb {R}^{+}\), which are continuous, strictly increasing and vanishing at zero; \(\mathscr {K}_{\infty }\) denotes a set of functions that are of class \(\mathscr {K}\) and unbounded. In addition, \(\left\| \cdot \right\| \) refers to the Euclidean vector norm.
5.2 Adaptive Tracking Control for Switched Stochastic Nonlinear Systems with Unknown Actuator Dead-Zone
5.2.1 Problem Formulation and Preliminaries
Consider the following switched stochastic nonlinear system in nonstrict-feedback form.
where \(x=(x_{1},\,x_{2},\ldots ,x_{n})^{T}\in \mathbb {R}^{n}\ \) is the system state, w is an r-dimensional independent standard Brownian motion defined on the complete probability space \(\left( \varOmega , \mathscr {F} ,\left\{ \mathscr {F} _{t}\right\} _{t\ge 0},P\right) \) with \(\varOmega \) being a sample space, \( \mathscr {F} \) being a \(\sigma \)-field, \(\left\{ \mathscr {F} _{t}\right\} _{t\ge 0}\) being a filtration, and P being a probability measure, and y is the system output; \(\sigma (t):[0,\infty )\rightarrow M=\{1,2,\ldots ,m\}\) represents the switching signal; \(v_{\sigma (t)} ,\,u_{\sigma (t)}\in \mathbb {R}\) are the actuator output and input. For any \(i=1,2,\ldots ,n\) and \(k\in M,\,\,f_{i,k}(x):\) \(\mathbb {R}^{n}\rightarrow \mathbb {R},\,\psi _{i,k}:\mathbb {R}^{n}\rightarrow \mathbb {R}^{r}\ \) are locally Lipschitz unknown nonlinear functions and \(\,g_{i,k}\) are positive known constants.
The nonsymmetric dead-zone nonlinearity is considered in the chapter, which is defined as the form in [12]:
Here, \(m_{r_{k}}>0\) and \(m_{l_{k}}>0\) represent the right and the left slopes of the dead-zone characteristic. \(b_{r_{k}}>0\) and \(b_{l_{k}}>0\) stand for the breakpoints of the input nonlinearity.
It is assumed that the nonsymmetric dead-zone nonlinearity can be reformulated as:
where \(D_{k}^{^{\prime }}(u_{k})\) is a smooth function, \(\iota _{k}\,\)is the error between \(D_{k}(u_{k})\,\)and \(D_{k}^{^{\prime }}(u_{k})\) with \(\left| \iota _{k}\right| \le \bar{\iota }_{k}\).
Moreover, we have
where \(\eta _{k}^{^{\prime }}(u_{k})=D_{k}^{^{\prime }}(u_{k})-u_{k}\) is an unknown function.
The controller can be designed as
Then (5.4) can be rewritten as
where \(u_{\phi _{k}}\) is the compensator of dead-zone nonlinearity and \(u_{c_{k}}\) is a main controller of system (5.1).
Our control objective is to design a state-feedback controller such that the output of system (5.1) can track a given time-varying signal \(y_{d}(t)\), and the problem of the actuator dead-zone can be solved. The following assumptions are supposed to be true.
Assumption 5.1
The tracking target \(y_{d}(t)\) and its time derivatives up to \(n^{th}\) order \(y_{d}^{(n)}(t)\) are continuous and bounded; it is further assumed that \(|y_{d}(t)|\le d\).
Assumption 5.2
There exist strictly increasing smooth functions \(\phi _{i,k} (\cdot ),\,\rho _{i,k}(\cdot ):R^{+}\rightarrow R^{+}\) with \(\phi _{i,k} (0)=\,\rho _{i,k}(0)=0\) such that for \(i=1,2,\ldots ,n\) and \(k\in M,\)
Remark 5.1
The increasing properties of \(\phi _{i,k}(\cdot ),\,\rho _{i,k}(\cdot )\) imply that if \(a_{i},b_{i}\ge 0,\,\)for \(i=1,2,\ldots ,n,\) then \(\phi _{i,k}(\sum _{i=1}^{n}a_{i})\le \sum _{i=1}^{n}\phi _{i,k}(na_{i})\), \(\rho _{i,k}(\sum _{i=1}^{n}b_{i})\le \sum _{i=1}^{n}\rho _{i,k}(nb_{i}).\) Notice that \(\phi _{i,k}(s),\,\rho _{i,k}(s)\) are smooth functions, and \(\phi _{i,k}(0)=\rho _{i,k}(0)=0.\) Therefore, there exist smooth functions \(h_{i,k}(s),\,\eta _{i,k}(s)\) such that \(\phi _{i,k}(s)=sh_{i,k}(s),\,\rho _{i,k}(s)=s\eta _{i,k}(s)\) which results in
We use the radial basis function (RBF) neural networks to approximate any a real function f(Z) over a compact set \(\varOmega _{Z}\subset \mathbb {R}^{q}.\) For arbitrary \(\bar{\varepsilon }>0,\) there exists a neural network \(W^{T}S(Z)\) such that
where \(Z\in \varOmega _{Z}\subset R^{q},\) \(W=[w_{1},w_{2},\ldots ,w_{l}]^{T}\) is the ideal constant weight vector, and \(S(Z)=[s_{1}(Z)\), \(s_{2}(Z)\),\(\,\ldots \) ,\(\,s_{l}(Z)]^{T}\) is the basis function vector, with \(l>1\) being the number of the neural network nodes and \(s_{i}(Z)\) being chosen as Gaussian functions; i.e., for \(i=1,2,\ldots ,l\),
where \(\mu _{i}=[\mu _{i1},\mu _{i2},\ldots ,\mu _{iq}]^{T}\) is the center vector, and \(\zeta _{i}\) is the width of the Gaussian function.
Definition 5.1
For any given \(V(x_{i},t)\in \mathscr {C}^{2,1}\) associated with system (5.1), define the differential operator \(\mathscr {L}\) as follows;
where \(F_{i,k}=g_{i,k}x_{i+1}+f_{i,k}(x)\).
Definition 5.2
The trajectory \(\{x(t),t\ge 0\}\) of switched stochastic system (5.1) is said to be semi-globally uniformly ultimately bounded (SGUUB) in the \(p^{th}\) moment, if for some compact set \(\varOmega \in \mathbb {R}^{n}\) and any initial state \(x_{0}=x(t_{0})\), there exist a constant \(\varepsilon >0\), and a time constant \(T=T(\varepsilon ,x_{0})\) such that \(E(|x(t)|^{p})<\varepsilon \), for all \(t>t_{0}+T\). Especially, when \(p=2\), it is usually called SGUUB in mean square.
Lemma 5.1
([13]) Suppose that there exist a \(C^{2,1}\) function V(x, t) \(:\mathbb {R}^{n}\times \mathbb {R}^{+}\rightarrow \mathbb {R}^{+}\), two constants \(c_{1}>0\) and \(c_{2}>0\), class \(\mathscr {K}_{\infty }\) functions \(\bar{\alpha }_{1}\) and \(\bar{\alpha }_{2}\) such that
for all \(x\in \mathbb {R}^{n}\) and \(t>t_{0}\). Then, there is an unique strong solution of system (5.1) for each \(x_{0}\in \mathbb {R}^{n}\), that satisfies
Lemma 5.2
([14]) For any \(\xi \in \mathbb {R}\) and \(\varpi >0,\) the following inequality holds:
with \(\delta =0.2785.\)
5.2.2 Main Results
Based on the backstepping technique, a control design and stability analysis procedure is presented in this section. For \(i=1,2,\ldots ,n-1,\) define a common virtual control function \(\alpha _{i}\) as
where \(\lambda _{i}\), \(a_{i}>0\) are design parameters, \(g_{i,\min } =\min \{g_{i,k}:k\in M\}\), \(z_{i}\) is the new state variable after the coordinate transformation: \(z_{i}=x_{i}-\alpha _{i-1}\), \(\alpha _{0}=y_{d}\). \(\hat{\theta }\) is an unknown constant that is specified later. \(S_{i} =S_{i}(X_{i})\) is the basis function vector. \(X_{i}=[\bar{x}_{i}^{T},\bar{\hat{\theta }}_{i},\bar{y}_{d}^{(i)}]^{T}\) with \(\bar{x}_{i}=[x_{1} ,x_{2},\ldots ,x_{i}]^{T}\), \(\,\bar{\hat{\theta }}_{i}=[\hat{\theta }_{1} ,\hat{\theta }_{2},\ldots ,\hat{\theta }_{i}]^{T}\), \(\,\bar{y}_{d}^{(i)}=[y_{d} ,\dot{y}_{d},\ldots ,y_{d}^{(i)}]^{T}\). The z-system is obtained as
where the differential operator \(\mathscr {L}\) is defined in Definition 5.1; \(\mathscr {L}\alpha _{i-1}\) is given by:
Consider the following common stochastic Lyapunov function candidate
where \(r_{1},\,r_{2}>0\) are design parameters; \(\theta \) and \(\vartheta \) are specified later. \(\hat{\theta }\) and \(\hat{\vartheta }\) stand for the estimations of \(\theta \) and \(\vartheta \), respectively; \(\tilde{\theta } =\theta -\hat{\theta }\), \(\tilde{\vartheta }=\vartheta -\hat{\vartheta }\).
Lemma 5.3
From the coordinate transformations \(z_{i}=x_{i}-\alpha _{i-1}\), \(i=1,2,\ldots ,n,\) \(\alpha _{0}=y_{d}\), the following results hold,
where \(\varphi _{i}(z_{i},\hat{\theta })=\frac{1}{g_{i,\min }}[(\lambda _{i} +\frac{3}{4})+\frac{1}{2a_{i}^{2}}z_{i}^{2}\hat{\theta }S_{i}^{T}S_{i}]+1,\) for \(i=1,2,\ldots ,n-1,\) and \(\varphi _{n}=1\).
Proof
From Assumption 5.1 and (5.15), one can get that
The proof of Lemma 5.3 is completed here. \(\square \)
The \(\mathscr {L}V\) can be given by
By resorting to Assumption 5.2 and Lemma 5.3, one has that
where \(\bar{\phi }_{s,k}^{4}(z_{l}\), \(\hat{\theta })=\frac{1}{4}(n+1)^{4} \varphi _{l}^{4}(z_{l}\), \(\hat{\theta })h_{s,k}^{4}((n+1)|z_{l}|\varphi _{l}(z_{l}\), \(\hat{\theta }))\), \(\frac{\partial \alpha _{0}}{\partial x_{i}}=0\) and \(\frac{\partial \alpha _{i-1}}{\partial x_{i}}=-1.\)
Then, the following inequality can be obtained,
where \(l_{ij}\) is a positive constant, and \(\frac{\partial \alpha _{0}}{\partial x_{j}}=0\) because \(\alpha _{0}=y_{d}\), and
where \(\bar{\rho }_{s,k}^{4}(z_{l},\hat{\theta })=\frac{1}{2}(n+1)^{4} \varphi _{l}^{4}(z_{l},\hat{\theta })\eta _{s,k}^{4}((n+1)|z_{l}|\varphi _{l}(z_{l}\), \(\hat{\theta }))\), \(s=1,2,\ldots ,i-1.\)
Substituting (5.21), (5.22) and (5.23) into (5.20) gives that
Define \(U_{i,k}\) as
By using Lemma 5.2 one has
Note that
Therefore, one has
For any \(i=1,2,\ldots ,n\) and \(k\in M\), define \(\bar{f}_{i,k}\) as
with \(z_{n+1}=0.\)
Substituting (5.6) and (5.26)–(5.28) into (5.24) yields that
By exploring the neural networks’ approximation capability and Young’s inequality, one can get the following inequalities.
where \(\theta _{i,k}=\left\| W_{i,k}\right\| ^{2}\), \(\vartheta _{\eta ,k}=\left\| W_{\eta ,k}\right\| ^{2}\), \(\theta _{i}=\max \{\theta _{i,k}:k\in M\}\), \(\vartheta _{\eta }=\max \{\vartheta _{\eta ,k}:k\in M\}\), \(\left| \varepsilon _{i,k}\right| \le \bar{\varepsilon }_{i}\), \(\left| \varepsilon _{\eta ,k}+\iota _{k}\right| \le \bar{\varepsilon }_{\eta }\).
Substituting (5.30) and (5.31) into (5.29) gives
where \(\,\varpi _{i}:=\max \{\varpi _{i,k},\,k\in M\}.\)
Design the virtual control function as
where \(\hat{\theta }=\sum _{i=1}^{n}\hat{\theta }_{i}\) is the estimation of \(\theta \); \(\,\lambda _{i}>0\) is a design parameter.
The actual actuator input is given as
where
\(\lambda _{n}\), \(\lambda _{\eta }\), \(a_{n}\), \(a_{\eta }>0\) are design parameters, \(g_{n,\max }=\max \{g_{n,k}\), \(k\in M\}\), \(g_{n,\min } =\min \{g_{n,k}\), \(k\in M\}\), \(\hat{\vartheta }\) is the estimation of \(\vartheta \).
The adaptive laws can be designed as
Then, one can get from (5.32)–(5.38) that
It is clear that
Combining (5.39) with (5.40) and (5.41) yields that
where \(\lambda _{n}:=\lambda _{n}+\lambda _{\eta },\,p_{0}=\min \{4\lambda _{i},\beta _{1},\beta _{2}:1\le i\le n\},\,q_{0}=\sum _{i=1}^{n}(\frac{a_{i}^{2}}{2}+\frac{\bar{\varepsilon }_{i}^{4}}{4})+\sum _{i=1}^{n}\left( \delta \varpi _{i}+\sum _{j=1}^{i}l_{ij}^{2}\right) +\frac{\beta _{1}\theta ^{2} }{2r_{1}}+\frac{\beta _{2}\vartheta ^{2}}{2r_{2}}+g_{n,k}(\frac{a_{\eta }^{2}}{2}+\frac{\bar{\varepsilon }_{\eta }^{4}}{4}).\)
By using Lemma 5.1, we have
Then, the following inequality holds
where \(V(0)=\sum _{j=1}^{n}\frac{z_{j}^{2}(0)}{4}+\frac{1}{2r_{1}}\tilde{\theta }(0)^{2}+\frac{1}{2r_{2}}\tilde{\vartheta }(0)^{2}\). Equation (5.44) implies that all the signals in the closed-loop system are bounded in probability. In particular, we have
Now, we are ready to provide our main result in the following theorem.
Theorem 5.1
Consider the closed-loop system (5.1) with unknown nonsymmetric actuator dead-zone (5.2). Suppose that for \(1\le i\le n,\) \(k\in M,\) the packaged unknown functions \(\bar{f}_{i,k}\) can be approximated by neural networks in the sense that the approximation error \(\varepsilon _{i,k}\) are bounded. Under the state feedback controller (5.34) and the adaptive laws (5.37), (5.38), the following statements hold,
(i) All the signals of the closed-loop z-system (5.17) are SGUUB in the fourth moment and
(ii) The output y of the closed-loop system (5.1) can be almost surely regulated to a small neighborhood of the target signal.
Proof
It is not difficult to complete the proof by using the above developments. \(\square \)
5.2.3 Simulation Results
In this section, an example about the control of a ship manoeuvring system are used to illustrate the effectiveness of the obtained results.
The ship maneuvering system can be described by the following Norrbin nonlinear model [15].
where \(T_{\sigma (v_{s})}\) is the time constant, \(h=\dot{\psi }\) denotes the yaw rate, \(\psi \) stands for the heading angle, \(\alpha _{\sigma (v_{s})}\) is the Norrbin coefficient, \(K_{\sigma (v_{s})}\) represents the rudder gain, \(\delta \) is the rudder angle and w stands for an r-dimensional independent standard Brownian motion, \(\phi _{\sigma (v_{s})}(\psi ,h,\delta ):\mathbb {R}^{3}\rightarrow \mathbb {R}^{3\times r}\) is an unknown function, and \(\sigma (v_{s})\) is the switching signal that satisfies:
\(v_{L},\,v_{M},\,v_{T}\) represent the value of low speed, middle speed and top speed, respectively.
A simplified mathematical model of the rudder system can be described as follows,
where \(T_{E,\sigma (v_{s})}\) represents the rudder time constant, \(\delta \) stands for the actual rudder angle, \(K_{E,\sigma (v_{s})}\) denotes the rudder control gain and \(\delta _{E,\sigma (v_{s})}\) is the rudder order.
Let \(x_{1}\) \(=\psi \), \(x_{2}\) \(=h\), \(x_{3}\) \(=\delta \), \(v_{\sigma (v_{s})}=\delta _{E,\sigma (v_{s})}\); we can get the following switched nonlinear system model with actuator dead-zone to describe the dynamic behavior of the ship with low speed, middle speed and high speed, respectively.
where \(f_{\sigma (v_{s})}=-\frac{1}{T_{\sigma (v_{s})}}x_{2}-\frac{\tau _{\sigma (v_{s})}}{T_{\sigma (v_{s})}}x_{2}^{3},\) \(b_{\sigma (v_{s})} =\frac{K_{\sigma (v_{s})}}{T_{\sigma (v_{s})}}\).
The vessel data comes from a ship that has a length overall of 160.9 m. \(v_{L}=3.7\) m/s, \(K_{1}=32\) s\(^{\text {-1}}\), \(T_{1}=30\) s, \(\tau _{1}=40\) s\(^{\text {2}}\), \(T_{E,1}=4\) s, \(K_{E,1}=2\); \(v_{M}=7.5\) m/s, \(K_{2}=11.4\) s\(^{\text {-1}}\), \(T_{2}=63.69\) s, \(\tau _{2}=30\) s\(^{\text {2}}\), \(T_{E,2}=2.5\) s, \(K_{E,2}=1\); \(v_{T}=15.3\) m/s, \(K_{3}=5.1\) s\(^{\text {-1}}\), \(T_{3}=80.47\) s, \(\tau _{3}=25\) s\(^{\text {2}}\), \(T_{E,3}=1\) s, \(K_{E,3}=0.72\). The initial conditions are \(x_{1}(0)=2\),\(\,x_{2}(0)=-0.05\), \(x_{3}(0)=0.03\), \(\hat{\theta }(0)=10\), \(\hat{\vartheta }(0)=1\). We construct the basis function vectors \(S_{1}\), \(S_{2}\), \(S_{3}\) and \(S_{\eta }\) using 11, 15, 21 and 48 nodes, the centers \(\mu _{1}\), \(\mu _{2}\), \(\mu _{3}\), \(\mu _{\eta }\) evenly spaced on \([-1.5,\) \(4.5]\times [-3,\,4]\times [-10,\,8]\), \([-5,\,4]\times [-30,\,20]\times [-0.5,\,5.5]\), \([-5.5,\,8]\times [-12,\,25]\times [-0.1,\,2]\) and \([-10,\,2]\times [-60,\,2]\times [-0.2,\,10.5]\), and the widths \(\zeta _{1}=1.2\), \(\zeta _{2}=2.2\), \(\zeta _{3}=2\), \(\zeta _{\eta }=1.8\). The design parameters are \(a_{1}=a_{2}=a_{3}=a_{\eta }=10\), \(r_{1}=2,\) \(r_{2}=10\), \(\beta _{1}=0.5\), \(\beta _{2}=0.1\), \(\lambda _{1}=\lambda _{2}=\lambda _{3}=5\), and \(\lambda _{\eta }=3\). The desired trajectory is \(y_{d}=10\sin 0.08t\).
According to Theorem 5.1, the adaptive laws \(\dot{\hat{\theta }}\), \(\dot{\hat{\vartheta }}\) and the control laws \(u_{c_{k}},\,u_{\phi _{k}}\) are chosen, respectively, as
where \(u_{k}=u_{c_{k}}-u_{\phi _{k}},\,z_{1}=x_{1}-y_{d},\,z_{2}=x_{2} -\alpha _{1},\,z_{3}=x_{3}-\alpha _{2}\) and \(\alpha _{1},\) \(\alpha _{2}\) are given by
In order to give the simulation results, we assume that
and \(\phi _{1}=0.5x_{1}\sin x_{2}x_{3}\), \(\phi _{2}=0.25x_{1}^{2}x_{2}\cos x_{2}\), \(\phi _{3}=0.1x_{1}x_{3}\). The simulation results are shown in Figs. 5.1–5.4. Figure 5.1 depicts the responses of system output \(\psi \) and target signal \(y_{d}\). Figure 5.2 shows the trajectories of adaptive laws. Figure 5.3 demonstrates the responses of \(D(u_{c_{k}})\) (without dead-zone compensation controller) and \(D(u_{c_{k}}-u_{\phi _{k}})\) (with dead-zone compensation controller) and Fig. 5.4 illustrates the evolution of the switching signal. From Fig. 5.1, it can be seen that the output \(\psi \) can track the target signal \(y_{d}\) within a small bounded error. On the other hand, Fig. 5.3 proves that the dead-zone nonlinearity can be compensated by \(u_{\phi _{k}}\).
5.2.4 Conclusions
The tracking control problem for a class of stochastic switched nonlinear systems under arbitrary switchings has been investigated, where the unknown nonsymmetric actuator dead-zone is taken into account. A state feedback controller is designed for the systems under consideration. It is shown that the target signal can be almost surely tracked by the system output within a small bounded error, and the tracking error is SGUUB in \(4^{th}\) moment.
5.3 Adaptive Neural Control for Switched Stochastic High-Order Uncertain Nonlinear Systems with SISS Inverse Dynamic
5.3.1 Problem Formulation and Preliminaries
Here, we consider the following stochastic switched high-order nonlinear systems with SISS inverse dynamic,
where \(\zeta \in \mathbb {R}^{r}\) are immeasurable stochastic inverse dynamics; \(x=(x_{1},x_{2},\ldots ,x_{n})^{T}\in \mathbb {R}^{n}\) and \(y\in R\) are the system state and output, respectively; \(p_{i}\) is a positive odd integer and \(\omega \) is an m-dimensional standard Wiener process defined on the complete probability space \(\left( \varOmega , \mathscr {F} ,\left\{ \mathscr {F} _{t}\right\} _{t\ge 0},P\right) \) with \(\varOmega \) being a sample space, \( \mathscr {F} \) being a \(\sigma \)-field, \(\left\{ \mathscr {F} _{t}\right\} _{t\ge 0}\) being a filtration, and P being a probability measure; \(\sigma (t):[0,+\infty )\rightarrow M=\{1,2,\ldots ,m\}\) is the switching signal; \(u_{k}\in R\) is the control input of the k-th subsystem; \(f_{0,k}:\mathbb {R}^{r}\times R\rightarrow \mathbb {R}^{r}\), \(\psi _{0,k}:\mathbb {R}^{r}\times \mathbb {R}\rightarrow \mathbb {R}^{m\times r}\); For any \(i=1,2\ldots ,n\) and \(k=1,2,\ldots ,m\), \(f_{i,k}:\mathbb {R}^{r}\times \mathbb {R}^{n}\rightarrow \mathbb {R}\), \(\psi _{i,k}:\mathbb {R}^{r}\times \mathbb {R}^{n}\rightarrow \mathbb {R}^{m}\) are unknown nonlinear functions assumed to be locally Lipschitz with \(f_{i,k}(0)=0\), \(\psi _{i,k}(0)=0\), and \(g_{i,k}:R^{r}\times \mathbb {R}^{n}\rightarrow R\) is a strictly either positive or negative known function.
Remark 5.2
System (5.46) reduces to the well-known normal form when \(p_{i}=1\), \(\zeta =0\) and \(m=1\). In the case that \(p_{i}>1\), \(\zeta =0\) and \(m=1\), the Jacobian linearization of the system is neither controllable nor feedback linearizable. This makes the control design very challenging. To solve this problem, Lin and Qian [16] proposed a fruitful deterministic technique: adding a power integrator. Subsequently, many excellent results are proposed based on the adding a power integrator technique, see, e.g., [17–19] and the references therein.
Definition 5.3
For any given \(V(x_{i},t)\in \mathscr {C}^{2,1}\) associated with system (5.46), define the differential operator \(\mathscr {L}\) as follows,
where \(F_{i,k}=g_{i,\sigma (t)}(\zeta ,x)x_{i+1}^{p_{i}}+f_{i,\sigma (t)} (\zeta ,x)\).
Assumption 5.3
The sign and the upper bound of function \(g_{i,k}\) for \(1\le i\le n\) and \(k\in M,\) are known, and without loss of generality, it is assumed that
where \(\underline{d}_{i}\) and \(\bar{d}_{i}\) stand for the lower and upper bound values of \(g_{i,k}(\zeta ,x)\), respectively.
Assumption 5.4
For \(1\le i\le n\) and \(k\in M,\) there exists a \(\mathscr {C}^{2}\) function \(V_{0}(\zeta )\), which is positive definite and proper, such that \(\mathscr {L}V_{0}\le -\lambda _{0}\zeta ^{4}+\bar{\lambda } _{0}x_{1}^{p+3}\), where \(\lambda _{0}\) and \(\bar{\lambda }_{0}\) are positive constants.
Lemma 5.4
Let \(p\in \mathbb {R}^{*}\) and \(x,\,y\) be real-valued functions. There exists a constant \(c>0\) such that
Lemma 5.5
Suppose that there exists a \(\mathscr {C}^{2,1}\) function \(V(x,t):\mathbb {R}^{n}\times \mathbb {R}^{+}\rightarrow \mathbb {R}^{+}\), two constants \(c_{1}>0\), \(c_{2}>0\), and \(\mathscr {K}_{\infty }\) functions \(\bar{c}_{1}\), \(\bar{c}_{2}\) such that
for all \(x\in R^{n}\) and \(t>t_{0}\). Then, there is an unique strong solution for each \(x_{0}\in \mathbb {R}^{n}\) and it satisfies:
In the following control design procedure, radial basis function (RBF) neural networks are used to approximate a continuous real function f(X). For arbitrary \(\varepsilon >0,\) there exists a neural network \(W^{T}S(X)\) such that
where \(X\in \varOmega _{X}\subset \mathbb {R}^{q}\) is the input vector with q dimension, \(S(X)=[s_{1}(X),s_{2}(X),\ldots , s_{l}(X)]^{T}\) is the basis function vector, and \(W=[w_{1},w_{2},\ldots ,w_{l}]^{T}\) is the ideal constant weight vector with \(l>1\) being the number of the neural network nodes, and \(s_{i}(X)\) are chosen as Gaussian functions; i.e., for \(i=1,2,\ldots ,l\),
where \(\zeta _{i}\) is the width of the Gaussian function, and \(\mu _{i} =[\mu _{i1},\mu _{i2},\ldots ,\mu _{iq}]^{T}\) is the center vector.
Lemma 5.6
Consider the Gaussian RBF networks. Let \(\rho :=\frac{1}{2}\min _{i\ne j}||\mu _{i}-\mu _{j}||\); then an upper bound of ||S(X)|| is taken as
It has been proven in [20] that the constant D in Lemma 5.6 is a limited value and is independent of the variable X.
5.3.2 Main Results
In the following, the adaptive tracking control design is carried out by using a standard backstepping procedure. Firstly, define \(p=\max _{i=1,\ldots ,n}\{p_{i}\}\). The following lemma is also given.
Lemma 5.7
Suppose that the Lyapunov function
is positive-definite and proper, satisfying
Then, the following inequality holds
where
Proof
Let \(a=\phi ^{1/(p+3)}\) and \(b=\xi _{i}\). Then, by using Young’s inequality
which implies that
Substituting (5.49) into (5.48) yields that
The proof of Lemma 5.7 is completed here. \(\square \)
Step 1: Define the variable \(z_{1}=x_{1}\). Then, consider the following Lyapunov function candidate
It follows from (5.47) and Assumption 5.4 that
where \(f_{1,k}\) and \(\left\| \psi _{1,k}\right\| ^{2}\) are unknown. Then, two neural networks \(W_{1,k}S_{1}\) and \(\varPhi _{1,k}P_{1,k}\) are used to approximate the unknown function \(f_{1,k}\) and the norm \(\left\| \psi _{1,k}\right\| \) such that for any given \(\varepsilon _{1,k}>0\) and \(\tau _{1,k}>0\),
where \(X_{1}:=[\zeta ^{T},x^{T}]^{T}\in R^{r+n}\), \(\left| \delta _{1,k}(X)\right| \le \varepsilon _{1,k}\), \(\bar{\delta }_{1,k}(X_{1} )\le \tau _{1,k}\).
One can get from the Young’s inequality and Lemma 5.6 that
where \(l_{1},\,\eta _{1}>0\) are design parameters; \(\left\| S_{1}\right\| \le D_{1}\); \(\theta _{1}:=\max \{\left\| W_{1,k}\right\| ^{(p+3)/(p-p_{1} +3)}:k\in M\}\); \(b_{1}=l_{1}^{-(p+3)/p_{1}}+\eta _{1}^{-(p+3)/p_{1}} \varepsilon _{1}^{(p+3)/p_{1}}.\)
Moreover, one has that
where \(\xi _{1}\), \(\ m_{1}>0\) are design parameters; \(\left\| P_{1}\right\| \le Q_{1}\); \(\bar{b}_{1}=\xi _{1}^{-(p+3)/(p_{1}+1)}\) \(+\) \(m_{1}^{-(p+3)/(p_{1}+1)}\tau _{1}^{(p+3)/(p_{1}+1)}\); \(\varphi _{1} =\max \{\left\| \varPhi _{1,k}\right\| ^{(p+3)/(p-p_{1}+2)}:k\in M\}\).
Substituting (5.51) and (5.52) into (5.50), yields that
where \(\tilde{b}_{1}:=b_{1}+0.5(p-p_{1}+3)\bar{b}_{1}.\)
Then, the common virtual control function can be designed as
where \(\hat{\theta }_{1}\), \(\hat{\varphi }_{1}\) are the estimations of \(\theta _{1}\), \(\varphi _{1}\) respectively; \(\bar{\lambda }_{1}>1+\bar{\lambda }_{0}\) is a positive design parameter; \(\underline{d}_{1}\) is defined in Assumption 5.3.
It follows from (5.53) and (5.54) that
where \(\tilde{\theta }_{1}=\theta _{1}-\hat{\theta }_{1}\), \(\tilde{\varphi } _{1}=\varphi _{1}-\hat{\varphi }_{1}.\)
Step 2: Denote \(z_{2}=x_{2}-\alpha _{1}\), and define \(d\alpha _{1}\) as
where \(\dot{\hat{\theta }}_{1}\) and\(\ \dot{\hat{\varphi }}_{1}\) will be specified later, \(x_{n+1}:=u\) will given at final step.
Choose the Lyapunov function as
Then, \(\mathscr {L}V_{2}\) is given by
By using Lemma 5.4 and Young’s inequality, one can obtain that
where \(\bar{\beta }_{1}=1+\beta _{1}^{(p_{1}-1)(p+1)}\); \(c_{1}\) is chosen as \(1/2\bar{d}_{1}\); \(\bar{d}_{1}\) is defined in Assumption 5.3.
Substituting (5.58) into (5.57), gives that
where \(\bar{f}_{2,k}=f_{2,k}-\bar{\alpha }_{1}\), \(\bar{\psi }_{2,k}:=\) \(\left\| \psi _{2,k}-\frac{\partial \alpha _{1}}{\partial x_{1}}\psi _{1,k}\right\| ^{2}\). Then, neural networks \(W_{2,k}^{T}S_{2,k}(X_{2})\) and \(\varPhi _{2,k}^{T}P_{2,k}(X_{2})\) are used to approximate the unknown functions \(\bar{f}_{2,k}\) and \(\bar{\psi }_{2,k}\) such that for any given \(\varepsilon _{2,k}>0\) and \(\tau _{2,k}>0\),
where \(X_{2}:=[\zeta ^{T},x^{T},\hat{\theta }_{1},\hat{\varphi }_{1}]^{T}\in R^{r+n+2}\), \(\left| \delta _{2,k}(X_{2})\right| \le \varepsilon _{2,k}\), \(\bar{\delta }_{2,k}(X_{2})\le \tau _{2,k}\).
Similar to the procedure in (5.51), one can obtain that
where \(l_{2},\,\eta _{2}>0\) are design parameters, \(\theta _{2}:=\max \{\left\| W_{2,k}\right\| ^{(p+3)/(p-p_{2}+3)}:k\in M\}\), \(b_{2} =l_{2}^{-(p+3)/p_{2}}+\eta _{2}^{-(p+3)/p_{2}}\varepsilon _{2}^{(p+3)/p_{2}}.\)
Using a similar way to (5.52), one gets that
where \(\xi _{2}\), \(m_{2}>0\) are design parameters; \(\bar{b}_{2}=\xi _{2} ^{-\frac{p+3}{p_{2}+1}}+m_{2}^{-\frac{p+3}{p_{2}+1}}\tau _{2}^{\frac{p+3}{p_{2}+1}}\); \(\varphi _{2}=\max \{ \left\| \varPhi _{2,k}\right\| ^{\frac{p+3}{p-p_{2}+2}}:k\in M\}\); \(P_{2}(X_{2})\) and \(\tau _{2}(X_{2})\) represent the basis function vector and the estimation error of \(\varphi _{2}\).
Design the common virtual control function as
where \(\hat{\theta }_{2}\), \(\hat{\varphi }_{2}\) are the estimation of \(\theta _{2}\) and \(\varphi _{2}\) respectively; \(\lambda _{2}>1\) is a design parameter; \(\underline{d}_{2}\) is defined in Assumption 5.3.
By substituting (5.60)–(5.62) into (5.59), one has
where \(\tilde{b}_{j}:=b_{j}+0.5(p-p_{j}+3)\bar{b}_{j}\), \(\tilde{\theta } _{j}=\theta _{j}-\hat{\theta }_{j}\), \(\tilde{\varphi }_{j}=\varphi _{j} -\hat{\varphi }_{j}\), \(j=1,2\).
Step i: Suppose at step i \((3\le i\le n-1)\) that, there is a set of virtual control functions \(\alpha _{3},\ldots ,\alpha _{n-1}\), defined by
and assume that a set of unknown nonlinear functions \(\bar{f}_{i,k}\) and \(\bar{\psi }_{i,k}\) can be approximated by neural networks \(W_{i,k}^{T} S_{i,k}(X_{i})\) and \(\varPhi _{i,k}^{T}P_{i,k}(X_{i})\) for any given \(\varepsilon _{i,k}>0\), \(\tau _{i,k}>0.\)
where \(X_{i}:=[\zeta ^{T},x^{T},\hat{\theta }_{1},\ldots ,\hat{\theta }_{i} ,\hat{\varphi }_{1},\ldots ,\hat{\varphi }_{i}]^{T}\in R^{r+n+2i}.\)
A straightforward calculation gives that
where \(\tilde{\theta }_{j}=\theta _{j}-\hat{\theta }_{j}\), \(\tilde{\varphi } _{j}=\varphi _{j}-\hat{\varphi }_{j}\); \([\hat{\theta }_{j}\), \(\hat{\varphi }_{j} ]\) is the estimation of
and \(\bar{b}_{j}=\xi _{j}^{-(p+3)/(p_{j}+1)}+m_{j}^{-(p+3)/(p_{j}+1)}\tau _{j}^{(p+3)/(p_{j}+1)}\).
Step n: Let \(z_{n}=x_{n}-\alpha _{n-1},\) define \(d\alpha _{n-1}\) as
where \(x_{n+1}:=u\) is provided later.
We construct the Lyapunov function as
By using (5.64), \(\mathscr {L}V_{n}\) is given by
where \(\bar{f}_{n,k}=f_{n,k}-\bar{\alpha }_{n-1}\), \(\bar{\psi }_{n,k}=\) \(\left\| \psi _{n,k}-\sum _{i=1}^{n-1}\frac{\partial \alpha _{n-1}}{\partial x_{i}}\psi _{i,k}\right\| ^{2}\). Then, neural networks \(W_{n,k}^{T} S_{n,k}(X_{n})\) and \(\varPhi _{n,k}^{T}P_{n,k}(X_{n})\) are used to approximate unknown functions \(\bar{f}_{n,k}\) and \(\bar{\psi }_{n,k}\) such that for any given \(\varepsilon _{n,k}>0\) and \(\tau _{n,k}>0,\)
where \(X_{n}:=[\zeta ^{T},x^{T},\hat{\theta }_{1},\ldots ,\hat{\theta }_{n} ,\hat{\varphi }_{1},\ldots ,\hat{\varphi }_{n}]^{T}\in R^{r+3n}\), \(\left| \delta _{n,k}(X_{n})\right| \le \varepsilon _{n,k}\), \(\bar{\delta } _{n,k}(X_{n})\le \tau _{n,k}\).
Similar to (5.60) and (5.61), one has
where \(l_{n},\,\eta _{n},\,\xi _{n},\,m_{n}>0\) are design parameters; \(\theta _{n}:=\max \{\left\| W_{n,k}\right\| ^{(p+3)/(p-p_{n}+3)}:k\in M\}\); \(b_{n}=l_{n}^{-\frac{p+3}{p_{n}}}+\eta _{n}^{-\frac{p+3}{p_{n}} }\varepsilon _{n}^{\frac{p+3}{p_{n}}}\), \(\varphi _{n}:=\max \{\left\| \varPhi _{n,k}\right\| ^{(p+3)/(p-p_{n}+2)}:k\in M\}\); \(\bar{b}_{n}=\xi _{n}^{-(p+3)/(p_{n}+1)}+m_{n}^{-(p+3)/(p_{n}+1)}\tau _{n}^{(p+3)/(p_{n}+1)}\).
Furthermore, it is not hard to get that
where \(\bar{\beta }_{n-1}=1+\beta _{n-1}^{(p_{n-1}-1)(p+1)}\), \(c_{n-1}\) is chosen as \(1/2\bar{d}_{n-1}\), and \(\bar{d}_{n-1}\) is defined in Assumption 5.3.
Substituting (5.68) and (5.69) into (5.67), the following inequality can be obtained.
Design the controller u as
where \(\hat{\theta }_{n}\) is the estimation of \(\theta _{n};\) \(\lambda _{n}>1\) is a positive design parameter; \(d_{n}\) is defined in Assumption 5.3.
It follows from (5.71) and (5.72) that
where \(\lambda _{1}:=\bar{\lambda }_{1}-\bar{\lambda }_{0}\).
Last Step: Choose the final Lyapunov function as
where \(r_{j}\) is a positive design parameter.
\(\mathscr {L}V\) is given by
The adaptive laws are defined as the solutions to the following differential equations
where \(\,j=1,2,\ldots ,n,\) \(B_{j},\,\bar{B}_{j}>0\) are design parameters.
This, together with (5.77), means that
Notice that
By using (5.79), (5.80) and Lemma 5.7, one has
where \(q_{0}=\min \{(p-p_{i}+4)(\lambda _{j}-1)\phi ^{(p_{i}-1)/(p+3)}\), \(B_{j}\), \(\bar{B}_{j}\), \(2\lambda _{0}:1\le j\le n\}\), \(\phi =\sum _{j=1}^{n}(\frac{B_{j}}{2r_{j}}\theta _{j}^{2}+\frac{\bar{B}_{j} }{2\bar{r}_{j}}\varphi _{j}^{2}+\tilde{b}_{j})\), \(q_{1}=(\lambda _{j} -1)(n+1)\phi \).
According to Lemma 5.5, we have that
which indicates that all the signals in the closed-loop system are bounded. The design is completed here. Next, we address our main result.
Theorem 5.2
For \(1\le i\le n,k\in M\) assume that all the unknown nonlinear functions \(\bar{f}_{i,k}\) and \(\bar{\psi }_{i,k}\) can be approximated by neural networks in the sense that the approximation errors are bounded, and all the initial values of \(\hat{\theta }_{i}\) and \(\hat{\varphi }_{i}\) satisfy \(\hat{\theta }_{i}(0)\ge 0\) and \(\hat{\varphi }_{i}(0)\ge 0\), respectively. Then, under the state feedback controller (5.72) and the adaptive laws (5.78), the equilibrium at the origin of the closed-loop system is boundedly stable in probability and
Proof
It is not difficult to complete the proof by the above discussions, and thus we omit the proof here.
In the following, a corollary is given by using only two adaptive laws.
Corollary 5.1
For \(1\le i\le n,k\in M\) assume that all the unknown nonlinear functions \(\bar{f}_{i,k}\) and \(\bar{\psi }_{i,k}\) can be approximated by neural networks in the sense that the approximation errors are bounded, and all the initial values of \(\hat{\theta }_{i}\) and \(\hat{\varphi }_{i}\) satisfy \(\hat{\theta }_{i}(0)\ge 0\) and \(\hat{\varphi }_{i}(0)\ge 0\), respectively. Consider the following controller and adaptive laws:
where \(\,\lambda _{n}>1\), \(l_{j},\,\xi _{j},\,m_{n},\,\eta _{n},\,r\,,B\,,\bar{r}\,,\bar{B}>0\) are positive design parameters, \(\hat{\theta }=\sum _{j=1} ^{n}\hat{\theta }_{j}\), \(\hat{\varphi }=\sum _{j=1}^{n}\hat{\varphi }_{j}\). Then, the equilibrium at the origin of the closed-loop system is boundedly stable in probability and
Proof
It should be pointed out that \(\hat{\theta }\ge \) \(\hat{\theta }_{j}\ge 0\), \(\hat{\varphi }\ge \hat{\varphi }_{j}\ge 0\), \(\,j=1,\ldots ,n\). Therefore, we can use \(\hat{\theta }\) and \(\hat{\varphi }\) instead of \(\hat{\theta }_{j}\) and \(\hat{\varphi }_{j}\) in (5.54), (5.62), (5.63) and (5.72). In (5.75), the parameters \(\tilde{\theta }_{j}\) and \(\tilde{\varphi }_{j}\) in Lyapunov function V should be rewritten as \(\tilde{\theta }\) and \(\tilde{\varphi }.\) The detailed proof is omitted here because it is similar to the one of Theorem 5.2. \(\square \)
5.3.3 Simulation Results
An example with two controllers (multiple adaptive laws and two adaptive laws respectively) is presented in the following to demonstrate the effectiveness of our main results.
Consider the following switched stochastic high-order nonlinear systems with SISS inverse dynamics:
where \(g_{1,1}\), \(f_{1,1}\), \(\psi _{1,1}\), \(g_{2,1}\), \(f_{2,1}\), \(\psi _{2,1}\), \(g_{1,2}\), \(f_{1,2}\), \(\psi _{1,2}\), \(g_{2,2}\), \(f_{2,2}\), and \(\psi _{2,2}\) are all completely unknown functions; \(p_{1}=3\), \(p_{2}=5\). First, a controller under multiple adaptive laws is designed by Theorem 5.2. The initial conditions are \(\zeta (0)=1\) \(x_{1}(0)=0.5,\,x_{2}(0)=-0.5\) and \(\hat{\theta }_{1}(0)=2\), \(\hat{\theta }_{2}(0)=3.5\), \(\hat{\varphi }_{1}(0)=3\), \(\hat{\varphi }_{2}(0)=4\). The controller parameters are chosen as: \(\lambda _{1}=\lambda _{2}=5\), \(l_{2}=l_{2}=\eta _{1}=\eta _{2}=\xi _{1}=\xi _{2}=m_{1}=m_{2}=4\), \(r_{1} =r_{2}=\bar{r}_{1}=\bar{r}_{2}=1\), \(B_{1}=B_{2}=\bar{B}_{1}=\bar{B}_{2}=0.1\). We apply three nodes for each input dimension of \(W_{1}^{T}S_{1}\), \(W_{2} ^{T}S_{2}\), \(\varPhi _{1}^{T}P_{1}\) and \(\varPhi _{2}^{T}P_{2}\). Therefore, each of them contains 81 nodes with centers spaced evenly in the interval \([-0.5,\) \(0.5]\times [-0.5,\,0.5]\times [-0.5,\) \(0.5]\times [-0.5,\) 0.5], and the widths still being equal to 2.5. Second, a controller under two adaptive laws is designed by Corollary 5.1 with same conditions except \({\hat{\theta }}(0)=3\), \({\hat{\varphi }}(0)=4\), \(r=1\), \(B=0.1\).
In order to give the simulation results, it is assumed that \(f_{0,1} =-15\zeta +0.1x_{1}^{2}\), \(\psi _{0,1}=\left( \zeta ^{2}+0.3x_{1}^{4}\right) ^{1/2}\), \(g_{1,1}=\sin (x_{1}x_{2}+\zeta )+2\), \(f_{1,1}=x_{1}x_{2}+\zeta \), \(\psi _{1,1}=\sin (x_{1}x_{2}+\zeta )\), \(g_{2,1}=\cos (x_{1}+x_{2}^{2}+\zeta )+2\), \(f_{2,1}=x_{1}x_{2}^{2}+\zeta \sin \zeta \), \(\,\psi _{2,1}=x_{1}\cos x_{2} +\zeta ^{2}\); \(f_{0,2}=-13\zeta +0.3x_{1}^{2}\), \(\psi _{0,2}=\left( 0.17\zeta ^{2}+0.13x_{1}^{4}\right) ^{1/2}\), \(g_{1,2}=\sin (x_{1}^{2} +x_{2}+\zeta )+2\), \(f_{1,2}=x_{1}^{2}x_{2}+\zeta ^{2}\), \(\psi _{1,2}=\sin (x_{1}+x_{2})+\zeta ^{3}\), \(g_{2,2}=\cos (x_{1}x_{2}^{2}+\zeta )+2\), \(f_{2,2}=x_{1}x_{2}+\zeta \cos \zeta \), \(\,\psi _{2,2}=x_{1}\sin (x_{1}x_{2} )+\zeta ^{2}.\)
The simulation results based on Theorem 5.2 are shown in Figs. 5.5, 5.6 and 5.7, respectively. Figure 5.5 depicts the responses of system states. The trajectories of adaptive laws are shown in Figs. 5.6, and 5.7 describes the switching signal. From Fig. 5.5, it can be seen that all the system states eventually converge to a small neighborhood of the origin. The simulation results based on Corollary 5.1 are shown in Figs. 5.8, 5.9 and 5.10, respectively. It can be seen that all the system states eventually converge to a small neighborhood of the origin by using only two adaptive laws.
5.3.4 Conclusions
The adaptive neural control for a class of stochastic high-order switched nonlinear systems with SISS inverse dynamic is studied. An adaptive neural control algorithm is proposed. It can be shown that the equilibrium at the origin of the closed-loop system is BIBO stable in probability.
References
Hou M, Fu F, Duan G (2013) Global stabilization of switched stochastic nonlinear systems in strict-feedback form under arbitrary switchings. Automatica 49(8):2571–2575
Liu L, Xie X (2013) State feedback stabilization for stochastic feedforward nonlinear systems with time-varying delay. Automatica 49(4):936–942
Wu Z, Xia Y, Xie X (2012) Stochastic Barbalat’s lemma and its applications. IEEE Trans Autom Control 57(6):1537–1543
Xie X, Duan N (2010) Output tracking of high-order stochastic nonlinear systems with application to benchmark mechanical system. IEEE Trans Autom Control 55(5):1197–1202
Zhang H, Wu Z, Xia Y (2014) Exponential stability of stochastic systems with hysteresis switching. Automatica 50(2):599–606
Xie X, Duan N, Yu X (2011) State-feedback control of high-order stochastic nonlinear systems with SiISS inverse dynamics. IEEE Trans Autom Control 56(8):1921–1926
Feng W, Tian J, Zhao P (2011) Stability analysis of switched stochastic systems. Automatica 47(1):148–157
Wu L, Zheng WX, Gao H (2013) Dissipativity-based sliding mode control of switched stochastic systems. IEEE Trans Autom Control 58(3):785–791
Wu Z, Yang J, Shi P (2010) Adaptive tracking for stochastic nonlinear systems with Markovian switching. IEEE Trans Autom Control 55(9):2135–2141
Zhai D, Kang Y, Zhao P, Zhao Y (2012) Stability of a class of switched stochastic nonlinear systems under asynchronous switching. Int. J. Control, Autom. Syst. 10(6):1182–1192
Sontag E (1989) Smooth stabilization implies coprime factorization. IEEE Trans Autom Control 34(4):435–443
Tao G, Kokotović P (1994) Adaptive control of plants with unknown dead-zones. IEEE Trans Autom Control 39(1):59–68
Krstić M, Hua D (1998) Stabilization of nonlinear uncertain systems. Springer, London
Polycarpou M, Ioannou P (1996) A robust adaptive nonlinear control design. Automatica 32(3):423–427
Lim C, Forsythe W (1983) Autopilot for ship control, part 1: theoretical design. IEE Proc D Control Theory Appl 130(6):281–287
Lin W, Pongvuthithum R (2003) Nonsmooth adaptive stabilization of cascade systems with nonlinear parameterization via partial-state feedback. IEEE Trans Autom Control 48(10):1809–1816
Li W, Xie X (2009) Inverse optimal stabilization for stochastic nonlinear systems whose linearizations are not stabilizable. Automatica 45(2):498–503
liu L, Duan N (2010) State-feedback stabilization for stochastic high-order nonlinear systems with a ratio of odd integers power. Nonlinear Anal Model Control 15(1):39–53
Xie X, Tian J (2007) State-feedback stabilization for high-order stochastic nonlinear systems with stochastic inverse dynamics. Int. J. Robust Nonlinear Control 17(14):1343–1362
Wang C, Hill DJ, Ge SS, Chen G (2006) An iss-modular approach for adaptive neural control of pure-feedback systems. Automatica 42(5):723–731
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
Copyright information
© 2017 Springer International Publishing Switzerland
About this chapter
Cite this chapter
Zhao, X., Kao, Y., Niu, B., Wu, T. (2017). Adaptive Control of Switched Stochastic Nonlinear Systems. In: Control Synthesis of Switched Systems. Studies in Systems, Decision and Control, vol 80. Springer, Cham. https://doi.org/10.1007/978-3-319-44830-5_5
Download citation
DOI: https://doi.org/10.1007/978-3-319-44830-5_5
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-44828-2
Online ISBN: 978-3-319-44830-5
eBook Packages: EngineeringEngineering (R0)