
Citation: Liyun Liu, Zichi Wang, Zhenxing Qian, Xinpeng Zhang, Guorui Feng. Steganography in beautified images[J]. Mathematical Biosciences and Engineering, 2019, 16(4): 2322-2333. doi: 10.3934/mbe.2019116
[1] | Er-yong Cong, Li Zhu, Xian Zhang . Global exponential synchronization of discrete-time high-order BAM neural networks with multiple time-varying delays. AIMS Mathematics, 2024, 9(12): 33632-33648. doi: 10.3934/math.20241605 |
[2] | Nina Huo, Bing Li, Yongkun Li . Global exponential stability and existence of almost periodic solutions in distribution for Clifford-valued stochastic high-order Hopfield neural networks with time-varying delays. AIMS Mathematics, 2022, 7(3): 3653-3679. doi: 10.3934/math.2022202 |
[3] | Ailing Li, Mengting Lv, Yifang Yan . Asymptotic stability for quaternion-valued BAM neural networks via a contradictory method and two Lyapunov functionals. AIMS Mathematics, 2022, 7(5): 8206-8223. doi: 10.3934/math.2022457 |
[4] | Xiaofang Meng, Yongkun Li . Pseudo almost periodic solutions for quaternion-valued high-order Hopfield neural networks with time-varying delays and leakage delays on time scales. AIMS Mathematics, 2021, 6(9): 10070-10091. doi: 10.3934/math.2021585 |
[5] | R. Sriraman, P. Vignesh, V. C. Amritha, G. Rachakit, Prasanalakshmi Balaji . Direct quaternion method-based stability criteria for quaternion-valued Takagi-Sugeno fuzzy BAM delayed neural networks using quaternion-valued Wirtinger-based integral inequality. AIMS Mathematics, 2023, 8(5): 10486-10512. doi: 10.3934/math.2023532 |
[6] | R. Sriraman, R. Samidurai, V. C. Amritha, G. Rachakit, Prasanalakshmi Balaji . System decomposition-based stability criteria for Takagi-Sugeno fuzzy uncertain stochastic delayed neural networks in quaternion field. AIMS Mathematics, 2023, 8(5): 11589-11616. doi: 10.3934/math.2023587 |
[7] | Wenxiang Fang, Tao Xie, Biwen Li . Robustness analysis of fuzzy BAM cellular neural network with time-varying delays and stochastic disturbances. AIMS Mathematics, 2023, 8(4): 9365-9384. doi: 10.3934/math.2023471 |
[8] | Saim Ahmed, Ahmad Taher Azar, Ibraheem Kasim Ibraheem . Nonlinear system controlled using novel adaptive fixed-time SMC. AIMS Mathematics, 2024, 9(4): 7895-7916. doi: 10.3934/math.2024384 |
[9] | Yongkun Li, Xiaoli Huang, Xiaohui Wang . Weyl almost periodic solutions for quaternion-valued shunting inhibitory cellular neural networks with time-varying delays. AIMS Mathematics, 2022, 7(4): 4861-4886. doi: 10.3934/math.2022271 |
[10] | Houssem Jerbi, Izzat Al-Darraji, Saleh Albadran, Sondess Ben Aoun, Theodore E. Simos, Spyridon D. Mourtas, Vasilios N. Katsikis . Solving quaternion nonsymmetric algebraic Riccati equations through zeroing neural networks. AIMS Mathematics, 2024, 9(3): 5794-5809. doi: 10.3934/math.2024281 |
Quaternion is a hypercomplex extended from real and complex numbers, which was first introduced by W. R. Hamilton in 1843 [1]. A famous characteristic of quaternion numbers is that the commutativity law no longer holds for its multiplication. In the past decades, quaternion-valued neural networks (QVNNs) have become a hot topic in research due to its strong ability to cope with high-dimensional data [2]. QVNNs are extended from real-valued NNs (RVNNs) and complex-valued NNs (CVNNs). The state value, connection weights, and activations of QVNNs are all taken values in the quaternion area. Compared with CVNNs and RVNNs, QVNNs improve the computation speed substantially and have great potential in high-dimensional data processing, attitude control [3], computer graphics [4,5], image compression [6], and optimization [7]. Recently, some results on the dynamical property of QVNNs have been reported [8,9,10,11,12,13,14,15]. Q. Song and X. Chen [11] investigate the multi-stability of delayed QVNNs with a decomposition of the state space. In [14], sufficient conditions for the global μ-stability were derived by using the decomposition technique and quaternion linear matrix inequality (LMI). The state estimation issue of QVNNs was considered in [13], in which some criteria are achieved via quaternion LMI. Till now, the investigation on stabilization of QVNNs are still very few.
The BAMNNs is a famous network model, which was first introduced by Kosko in 1987 [16,17]. By iterations of backward and forward information flows between two layers, this network owns the ability of information association and information memory. BAMNNs have broad applications in various areas including associative memory, pattern recognition and automatic control. Moreover, the research of BAMNNs has received broad interests, in particular, various kinds of dynamical behaviors of this type of network have been investigated extensively, such as exponential stability, synchronization, Lagrange stability, dissipativity, and etc. [18,19,20,21,22]. However, to our best knowledge, the study on the dynamics of BAMNNs has been mainly concentrated on the real and complex field, corresponding results on the quaternion field has not appeared yet. Due to the abundant dynamical behavior of QVNNs and BAMNNs, our research is novel and has a promising application prospect. On the other hand, most of the previous literatures on BAMNNs and QVNNs only consider constant connection weights. In fact, due to the existence of external disturbances and uncertainty, the time-varying parameters are more reasonable than constant coefficients and can better adapt to the real-world systems.
Stability and stabilization are important topics in the study of network systems, which have been widely applied to various fields, such as associative memory, optimization, and pattern recognition [23,24,25,26,27,28,29,30,31]. So far, the traditional stabilization control has been focused on the case that the convergence time tends to infinity. However, in many practical applications, the state trajectories are required to converge to zero in finite time. To meet with this need, the concept of finite-time stability (FTS) was proposed. The finite-time control does not only reduce the convergence time effectively but also improve the robustness of the system [25,26,27]. Unfortunately, the settling time of finite-time control relies on the initial state of a system, which may be unavailable in many engineering processes. To overcome this limitation, the fixed-time control was introduced. As a special case of FTS, the settling time of fixed-time stability can be estimated even without the initial information. This is a quite meaningful advantage and due to this reason, the fixed-time control has become a hot topic [28,29,30,31]. Furthermore, to reduce the high control strength and save energy, the adaptive control strategy is proposed in this paper. By designing proper adaptive laws, the control gains are increasing according to the adaptive laws [32,33,34,35,36]. Therefore, adaptive control can be applied even when the accurate information of the system parameters is unavailable. To our best knowledge, the adaptive fixed-time control method has not been applied to quaternion systems yet.
The purpose of this paper is to explore the fixed-time stabilization of quaternion-valued BAMNNs with time-varying coefficients. The main novelty of this paper is as follows.
1) It is the first time that the quaternion and time-varying coefficients are introduced into BAMNNs. The new network model is more general than previous suggested ones and can arouse a more complex dynamical behavior.
2) A novel fixed-time convergence method is adopted to deal with the stabilization problem, the traditional fixed-time technique is a particular case of our method.
3) To overcome the noncommutativity of quaternion, two different methods are proposed, i.e., a decomposition and a non-decomposition method.
4) For the first time, the adaptive control strategy is applied to the fixed-time stabilization of QVNNs, which can reduce the control strength effectively and avoid the computation of control gains.
The contents of our paper is as follows. In chapter 2, the BAM OVNNs model is introduced. The main theorem are given in chapter 3. In chapter 4, simulations are given to demonstrate the correctness of our results. Lastly, conclusion is derived in chapter 5.
Notations. In this work, R,C and Q denote the real field, complex field, and quaternion field, respectively. For any vector χ=(χ1,⋯,χn)T∈Rn, it is noted that |χ|=(|χ1|,⋯,|χn|)T. The 1-vector norm of χ is defined as ‖χ‖1=n∑q=1|χq|. For any vector x,y∈Rn, x≤y means that xi≤yi for i=1,⋯,n. For any vector x∈Q, |x| denote the modulus of x. ∀a,b∈R, a⋀b denotes the minimum of a and b, a⋁b represents the maximum of a and b.
Quaternions are a kind of hypercomplex, which are an extension of complex numbers. A quaternion y∈Q can be described as
y=yR+yIi+yJj+yKk, |
where yR,yI,yJ,yK∈R, the imaginary parts i,j,k obey the Hamilton rule:
i2=j2=k2=−1,ij=−ji=k,jk=−kj=i,ki=−ik=j. |
For any quaternion q=qR+qIi+qJj+qKk, the conjugate of q is denoted by q∗=qR−qIi−qJj−qKk. The modulus of q is defined as
|q|=√q∗q=√(qR)2+(qI)2+(qJ)2+(qK)2. |
For any two quaternions y=yR+yIi+yJj+yKk and z=zR+zIi+zJj+zKk, addition is defined as
y+z=yR+zR+(yI+zI)i+(yJ+zJ)j+(yK+zK)k. |
Based on the Hamilton rule, the product of any two quaternion numbers is defined as
yz=(yRzR−yIzI−yJzJ−yKzK)+(yRzI+yIzR+yJzK−yKzJ)i+(yRzJ+yJzR+yKzI−yIzK)j+(yRzK+yKzR+yIzJ−yJzI)k. |
With the introduction of time-varying coefficients and quaternion into traditional BAMNNs, the model of time-varying quaternion-valued BAMNNs is introduced as follows:
dxp(t)dt=−cp(t)xp(t)+m∑q=1apq(t)fq(yq(t))+m∑q=1bpq(t)fq(yq(t−τ1(t)))+ωp(t)+up(t)dyq(t)dt=−dq(t)yq(t)+n∑p=1˜aqp(t)gp(xp(t))+n∑p=1˜bqp(t)gp(xp(t−τ2(t)))+ξq(t)+vq(t) | (2.1) |
where p=1,2,⋯,n,q=1,2,⋯,m; xp(t),yq(t)∈Q represent the state value of the pth neuron in the FX-field and the qth neuron in the FY-field, respectively. cp(t),dp(t)∈R are the time-varying self-feedback coefficients; fq(yq(⋅)),gp(xp(⋅)):Q→Q denote the activation functions of the qth neuron from the FY-field and the pth neuron from the FX-field, respectively. apq(t), bpq(t),˜aqp(t),˜bqp(t)∈Q represent the time-varying connection weights. τ1(t),τ2(t) are time-varying delays. ωp(t),ξq(t)∈Q are bounded external disturbances satisfying that |ωp(t)|≤ˆωp,|ξq(t)|≤ˆξq. up(t),vq(t) are the control input vectors.
dx(t)dt=−C(t)x(t)+A(t)f(y(t))+B(t)g(y(t−τ1(t)))+ω(t)+u(t)dy(t)dt=−D(t)y(t)+˜A(t)f(x(t))+˜B(t)g(x(t−τ2(t)))+ξ(t)+v(t) | (2.2) |
where A(t)=(apq(t))n×m,B(t)=(bpq(t))n×m,˜A(t)=(˜aqp(t))m×n,˜B(t)=(˜bqp(t))m×n.
Assumption 1. Suppose that the quaternion-valued activation function f(x(t))∈Qm,g(x(t))∈Qn can be expressed as
f(x(t))=fR(xR(t))+fI(xI(t))i+fJ(xJ(t))j+fK(xK(t))kg(x(t))=gR(xR(t))+gI(xI(t))i+gJ(xJ(t))j+gK(xK(t))k |
where fπ(xπ(t))∈Rm,gπ(xπ(t))∈Rn,π=R,I,J,K.
Remark 1. In fact, Assumption 1 is quite restrictive because not every quaternion activation function can be explicitly expressed as real and imaginary parts. There are still a large number of quaternion functions that do not satisfy this condition. Moreover, there is a more general assumption as follows:
f(x)=fR(xR,xI,xJ,xK)+fI(xR,xI,xJ,xK)i+fJ(xR,xI,xJ,xK)j+fK(xR,xI,xJ,xK)kg(x)=gR(xR,xI,xJ,xK)+gI(xR,xI,xJ,xK)i+gJ(xR,xI,xJ,xK)j+gK(xR,xI,xJ,xK)k |
However, this assumption may cause complex computation in our analysis process. Furthermore, it can be extended by Assumption 1 without difficulty. Thus, we take Assumption 1 in our later discussion.
Based on the above assumptions, now we separate the QVNNs (2.2) into RVNNs. Let
x(t)=xR(t)+ixI(t)+jxJ(t)+kxK(t),y(t)=yR(t)+iyI(t)+jyJ(t)+kyK(t). |
According to the quaternion multiplication rule, the quaternion system (2.2) can be decomposed into four real-valued NNs
dxR(t)dt=−CR(t)xR(t)+AR(t)fR(yR(t))−AI(t)fI(yI(t))−AJ(t)fJ(yJ(t))−AK(t)fK(yK(t))+BR(t)fR(yR(t−τ(t)))−BI(t)fI(yI(t−τ(t)))−BJ(t)fJ(xJ(t−τ(t)))−BK(t)fK(yK(t−τ(t)))+ωR(t)+uR(t),dxI(t)dt=−CI(t)xI(t)+AI(t)fR(yR(t))+AR(t)fI(yI(t))−AK(t)fJ(yJ(t))+AJ(t)fK(yK(t))+BI(t)fR(yR(t−τ(t)))+BR(t)fI(yI(t−τ(t)))−BK(t)fJ(xJ(t−τ(t)))+BJ(t)fK(yK(t−τ(t)))+ωI(t)+uI(t),dxJ(t)dt=−CJ(t)xJ(t)+AJ(t)fR(yR(t))+AK(t)fI(yI(t))+AR(t)fJ(yJ(t))−AI(t)fK(yK(t))+BJ(t)fR(yR(t−τ(t)))+BK(t)fI(yI(t−τ(t)))+BR(t)fJ(xJ(t−τ(t)))−BI(t)fK(yK(t−τ(t)))+ωJ(t)+uJ(t),dxK(t)dt=−CK(t)xK(t)+AK(t)fR(yR(t))−AJ(t)fI(yI(t))+AI(t)fJ(yJ(t))+AR(t)fK(yK(t))+BK(t)fR(yR(t−τ(t)))−BJ(t)fI(yI(t−τ(t)))+BI(t)fJ(xJ(t−τ(t)))+BR(t)fK(yK(t−τ(t)))+ωK(t)+uK(t),dyR(t)dt=−DR(t)yR(t)+˜AR(t)gR(xR(t))−˜AI(t)gI(xI(t))−˜AJ(t)gJ(xJ(t))−˜AK(t)gK(xK(t))+˜BR(t)gR(xR(t−τ(t)))−˜BI(t)gI(xI(t−τ(t)))−˜BJ(t)gJ(xJ(t−τ(t)))−˜BK(t)gK(xK(t−τ(t)))+ξR(t)+vR(t)dyI(t)dt=−DI(t)yI(t)+˜AI(t)gR(xR(t))+˜AR(t)gI(xI(t))−˜AK(t)gJ(xJ(t))+˜AJ(t))gK(xK(t))+˜BI(t)gR(xR(t−τ(t)))+˜BR(t)gI(xI(t−τ(t)))−˜BK(t)gJ(xJ(t−τ(t)))+˜BJ(t)gK(xK(t−τ(t)))+ξI(t)+vI(t)dyJ(t)dt=−DJ(t)yJ(t)+˜AJ(t)gR(xR(t))+˜AK(t)gI(xI(t))+˜AR(t)gJ(xJ(t))−˜AI(t)gK(xK(t))+˜BJ(t)gR(xR(t−τ(t)))+˜BK(t)gI(xI(t−τ(t)))+˜BR(t)gJ(xJ(t−τ(t)))−˜BI(t)gK(xK(t−τ(t)))+ξJ(t)+vJ(t)dyK(t)dt=−DK(t)yK(t)+˜AK(t)gR(xR(t))−˜AJ(t)gI(xI(t))+˜AI(t)gJ(xJ(t))+˜AR(t)gK(xK(t))+˜BK(t)gR(xR(t−τ(t)))−˜BJ(t)gI(xI(t−τ(t)))+˜BI(t)gJ(xJ(t−τ(t)))+˜BR(t)gK(xK(t−τ(t)))+ξK(t)+vK(t) | (2.3) |
Let
X(t)=(xR(t)T,xI(t)T,xJ(t)T,xK(t)T)T∈R4n,Y(t)=(yR(t)T,yI(t)T,yJ(t)T,yK(t)T)T∈R4m, |
then the compact form of system (2.3) is derived
{dX(t)dt=−C∗(t)X(t)+A∗(t)F(Y(t))+B∗(t)F(Y(t−τ1(t)))+ω∗(t)+U(t),dY(t)dt=−D∗(t)Y(t)+˜A∗(t)G(X(t))+˜B∗(t)G(X(t−τ2(t)))+ξ∗(t)+V(t), | (2.4) |
where X(t)=(xR(t)T,xI(t)T,xJ(t)T,xK(t)T)T∈R4n, Y(t)=(yR(t)T,yI(t)T,yJ(t)T,yK(t)T)T∈R4m, C∗(t)=diag(CR(t),CI(t),CJ(t),CK(t))∈R4n×4n, D∗(t)=diag(DR(t),DI(t),DJ(t),DK(t))∈R4m×4m, F(X(t))=(fR(xR(t))T,fI(xI(t))T,fJ(xJ(t))T,fK(xK(t))T)T∈R4m, G(X(t−τ(t)))=(gR(xR(t−τ(t)))T,gI(xI(t−τ(t)))T,gJ(xJ(t−τ(t)))T,gK(xK(t−τ(t)))T)T∈R4n, U(t)=diag{uR(t),uI(t),uJ(t),uK(t)}∈R4n, V(t)=diag{vR(t),vI(t),vJ(t),vK(t)}∈R4m, ω∗(t)=(ωR(t),ωI(t),ωJ(t),ωK(t))T∈R4n, ξ∗(t)=(ξR(t),ξI(t),ξJ(t),ξK(t))T∈R4m.
A∗(t)=(AR(t)−AI(t)−AJ(t)−AK(t)AI(t)AR(t)−AK(t)AJ(t)AJ(t)AK(t)AR(t)−AI(t)AK(t)−AJ(t)AI(t)AR(t)), |
B∗(t)=(BR(t)−BI(t)−BJ(t)−BK(t)BI(t)BR(t)−BK(t)BJ(t)BJ(t)BK(t)BR(t)−BI(t)BK(t)−BJ(t)BI(t)BR(t)), |
˜A∗(t)=(˜AR(t)−˜AI(t)−˜AJ(t)−˜AK(t)˜AI(t)˜AR(t)−˜AK(t)˜AJ(t)˜AJ(t)˜AK(t)˜AR(t)−˜AI(t)˜AK(t)−˜AJ(t)˜AI(t)˜AR(t)), |
˜B∗(t)=(˜BR(t)−˜BI(t)−˜BJ(t)−˜BK(t)˜BI(t)˜BR(t)−˜BK(t)˜BJ(t)˜BJ(t)˜BK(t)˜BR(t)−˜BI(t)˜BK(t)−˜BJ(t)˜BI(t)˜BR(t)), |
Assumption 2. Suppose that there exist positive diagonal matrices L1∈R4m×4m,L2∈R4n×4n such that the real-valued functions F(⋅),G(⋅) satisfy the following conditions
|F(Y)−F(Y′)|≤L1|Y−Y′|,|G(X)−G(X′)|≤L2|X−X′|, |
where X,X′∈R4n,Y,Y′∈R4m and
|X−X′|=(|X1−X′1|,⋯,|X4n−X′4n|)T,|Y−Y′|=(|Y1−Y′1|,⋯,|Y4n−Y′4m|)T,|F(Y)−F(Y′)|=(|F1(Y1)−F1(Y′1)|,|F2(Y2)−F2(Y′2)|,⋯,|F4m(Y4m)−F4m(Y′4m)|)T,|G(X)−G(X′)|=(|G1(X1)−G1(X′1)|,|G2(X2)−G2(X′2)|,⋯,|G4n(X4n)−G4n(X′4n)|)T. |
Moreover, F(0)=G(0)=0.
Remark 2. For later discussion, we make the following notations: L1=diag(L11,L12,⋯,L1,4m), L2=diag(L21,L22,⋯,L2,4n). Note that, according to the boundedness of disturbances ωp(t),ξq(t)∈Q, ω∗p(t),ξ∗q(t)∈R are also bounded.
Definition 1. For the initial value x(t0),y(t0) of system (2.2), if there exists a constant T(x(t0),y(t0))>0 such that
limt→T(x(t0),y(t0))|x(t)|+|x(t)|=0|x(t)|+|y(t)|=0,t≥T(x(t0),y(t0)). | (2.5) |
Then, the system (2.2) is said to be finite-time stabilized to the origin, T(x(t0),y(t0)) is called the settling time.
Definition 2. System (2.2) can reach fixed-time stabilization if the conditions in Definition 1 hold and the settling time T(x(t0),y(t0)) has an upper-bound Tmax. The number Tmax is not dependent on initial condition of the system. That is, T(x(t0),y(t0))≤Tmax,∀x(t0),y(t0)∈Qn.
Lemma 1. [30] Assume that the function V(⋅):Rn→R+⋃{0} is continuous radically bounded and if there exists an indefinite function q1(t) and a nonpositive function q2(t) such that
˙V(t)≤q1(t)Vα(t)+q2(t)Vβ(t)fort∈[t0,+∞) | (2.6) |
for some 0<α<1,β>1 and q1(t),q2(t) satisfy that
∫tt0q+1(s)ds≤N,∫tt0q−1(s)ds≤−γ1(t−t0)+M1 | (2.7) |
and
∫tt0q2(s)ds≤−γ2(t−t0)+M2 | (2.8) |
for all t>t0, where q+1(s)=q1(s)⋁0, q−1(s)=q1(s)⋀0 and N,Mi,γi(i=1,2) are positive constants. Then, V(t)=0 for all t≥Tmax, and the settling time is
Tmax=t0+(1−α)(N+M1)+1γ1(1−α)+(β−1)(N+M2)+1γ2(β−1) |
Proof. The proof can be referred to [30].
Lemma 2. For constants x1,⋯,xn≥0,0<p<1,q>1, the following condition holds.
n∑i=1xpi≥(n∑i=1xi)p,n∑i=1xqi≥n1−q(n∑i=1xi)q. | (2.10) |
Lemma 3. For any a,b∈Q, the following inequalities hold.
a∗⋅b+b∗⋅a≤a∗⋅a+b∗⋅b=|a|2+|b|2,a∗⋅b+b∗⋅a≤2|a||b|. |
In this section, we aim to derive the criteria for fixed-time stabilization of quaternion-valued BAMNNs (2.1). Due to the equivalence between QVNNs (2.1) and RVNNs (2.4), we now focus on the fixed-time stabilization of RVNNs (2.4) instead of QVNNs (2.1). The state-feedback controller for RVNNs (2.4) is chosen as follows
U(t)=−K(t)X(t)−Γ(t)sgn(X(t))−Λ(t)SGN(X(t))|X(t−τ2(t))|+ρ(t)SGN(X(t))|X(t)|α+ϱ(t)SGN(X(t))|X(t)|βV(t)=−Θ(t)Y(t)−Υ(t)sgn(Y(t))−Ξ(t)SGN(Y(t))|Y(t−τ1(t))|+ρ(t)SGN(Y(t))|Y(t)|α+ϱ(t)SGN(Y(t))|Y(t)|β, | (3.1) |
where 0<α<1,β>1, sgn(X(t))=(sgn(X1(t)),⋯,sgn(X4n(t)))T∈R4n, sgn(Y(t))=(sgn(Y1(t)),⋯,sgn(Y4m(t)))T∈R4m, SGN(X(t))=diag(sgn(X1(t)),sgn(X2(t)),⋯,sgn(X4n(t)))∈R4n×4n, SGN(Y(t))=diag(sgn(Y1(t)),⋯,sgn(Y4m(t)))∈R4m×4m. |Y(t)|α=(|Y1(t)|α,⋯,|Y4m(t)|α)T∈R4m, |X(t)|β=(|X1(t)|β,|X2(t)|β,⋯,|X4n(t)|β)T∈R4n.
The matrices K(t)=diag(K1(t),K2(t),⋯,K4n(t))∈R4n×4n, Γ(t)=diag(Γ1(t),Γ2(t),⋯,Γ4n(t))∈R4n×4n, Λ(t)=diag(Λ1(t),Λ2(t),⋯,Λ4n(t))∈R4n×4n, Θ(t)=diag(Θ1(t),Θ2(t),⋯,Θ4m(t))∈R4m×4m, Υ(t)=diag(Υ1(t),⋯,Υ4m(t))∈R4m×4m, Ξ(t)=diag(Ξ1(t),⋯,Ξ4m(t))∈R4m×4m are control gain matrices to be determined. ρ(t),ϱ(t)∈R are control gains to be decided.
Theorem 1. Under the Assumption 1 and 2, if the control gains Kp(t),Γp(t),Λp(t),Θq(t),Ξq(t),Υq(t) satisfy the following conditions
Kp(t)≥−C∗p(t)+4m∑q=1|˜a∗qp(t)|L2p,Θq(t)≥−D∗q(t)+4n∑p=1|a∗pq(t)|L1q,Λp(t)≥4m∑q=1|˜b∗qp(t)|L2p,Ξq(t)≥4n∑p=1|b∗pq(t)|L1q,Υq(t)≥|ξ∗q(t)|,Γp(t)≥|ω∗p(t)| | (3.2) |
where p=1,⋯,4n,q=1,⋯,4m, and there exist positive constants N,M1,M2,γ1,γ2 such that
∫tt0ρ+(s)ds≤N,∫tt0ρ−(s)ds≤−γ1(t−t0)+M1,∫tt0ϱ(s)ds≤−γ2(t−t0)+M2. | (3.3) |
Then, system (2.4) can be stabilized in fixed-time under controller (3.1). Equivalently, the quaternion-valued BAMNNs (2.1) can be stabilized in fixed-time.
Proof. Choosing the Lyapunov functional as below
V(t)=‖X(t)‖1+‖Y(t)‖1. |
Computing its derivative along trajectories (2.4), we have
dV(t)dt=sgnT(X(t)){−C∗(t)X(t)+A∗(t)F(Y(t))+B∗(t)F(Y(t−τ1(t)))+ω∗(t)+U(t)}+sgnT(Y(t)){−D∗(t)Y(t)+˜A∗(t)G(X(t))+˜B∗(t)G(X(t−τ2(t)))+ξ∗(t)+V(t)}≤−4n∑p=1C∗p(t)|Xp(t)|+4n∑p=14m∑q=1|a∗pq(t)|L1q|Yq(t)|+4n∑p=14m∑q=1|b∗pq(t)|L1q|Yq(t−τ1(t))|+4n∑p=1|ω∗p(t)|−4m∑q=1D∗q(t)|Yq(t)|+4m∑q=14n∑p=1|˜a∗qp(t)|L2p|Xp(t)|+4m∑q=14n∑p=1|˜b∗qp(t)|L2p|Xp(t−τ2(t))|+4m∑q=1|ξ∗q(t)|−4n∑p=1Kp(t)|Xp(t)|−4n∑p=1Γp(t)−4n∑p=1Λp(t)|Xp(t−τ2(t))|−4m∑q=1Θq(t)|Yq(t)|−4m∑q=1Υq(t)−4m∑q=1Ξq(t)|Yq(t−τ1(t))|+ρ(t)4n∑p=1|Xp(t)|α+ρ(t)4m∑q=1|Yq(t)|α+ϱ(t)4n∑p=1|Xp(t)|β+ϱ(t)4m∑q=1|Yq(t)|β |
Thus, we get
dV(t)dt≤4n∑p=1(−C∗p(t)+4m∑q=1|˜a∗qp(t)|L2p−Kp(t))|Xp(t)|+4m∑q=1(−D∗q(t)+4n∑p=1|a∗pq(t)|L1q−Θq(t))|Yq(t)|+4m∑q=1{4n∑p=1|b∗pq(t)|L1q−Ξq(t)}|Yq(t−τ1(t))|+4n∑p=1{4m∑q=1|˜b∗qp(t)|L2p−Λp(t)}|Xp(t−τ2(t))|+4m∑q=1|ξ∗q(t)|−4m∑q=1Υq(t)+4n∑p=1|ω∗p(t)|−4n∑p=1Γp(t)+(ρ+(t)+ρ−(t)){4n∑p=1|Xp(t)|α+4m∑q=1|Yq(t)|α}+ϱ(t){4n∑p=1|Xp(t)|β+4m∑q=1|Yq(t)|β} | (3.4) |
According to Lemma 2 and condition (3.2), we have
dV(t)dt≤(ρ+(t)+ρ−(t)){4n∑p=1|Xp(t)|α+4m∑q=1|Yq(t)|α}+ϱ(t){4n∑p=1|Xp(t)|β+4m∑q=1|Yq(t)|β}≤ρ+(t)(4n+4m)1−α{4n∑p=1|Xp(t)|+4m∑q=1|Yq(t)|}α+ρ−(t){4n∑p=1|Xp(t)|+4m∑q=1|Yq(t)|}α+ϱ(t)(4n+4m)1−β{4n∑p=1|Xp(t)|+4m∑q=1|Yq(t)|}β={ρ+(t)(4n+4m)1−α+ρ−(t)}Vα(t)+ϱ(t)(4n+4m)1−βVβ(t) | (3.5) |
According to Lemma 1, the fixed-time stabilization of system (2.4) can be reached via controller (3.1). Furthermore, the settling time is estimated as
Tmax=t0+(1−α){(4n+4m)1−αN+M1}+1γ1(1−α)+(β−1){(4n+4m)1−αN+(4n+4m)1−βM2}+1(4n+4m)1−βγ2(β−1) |
The proof is completed.
Remark 3. Recently, most of the existing studies on QVNNs [21,22,23,24,25,26] are mainly focused on a model with constant parameters. However, since the external disturbances widely exist, the time-varying coefficients considered in this paper are more reasonable, which makes the QVNNs model more general and it can be better applied to real world. Furthermore, the traditional fixed-time convergence approach requires the derivative of the Lyapunov functional to be indefinite, which is a significant limitation. Thus, the novel fixed-time convergence method is required to deal with this issue.
Remark 4. Note that Theorem 1 only focuses on the situation that the activation functions and connection weights can be decomposed into real numbers. In this way, the difficulty caused by noncommutativity of quaternion numbers are overcome. Unfortunately, in practical engineering systems, there are large number of quaternion-valued activations that can not be expressed explicitly by real and imaginary parts, which can cause the invalidity of Theorem 1. Thus, it gives us the motivation to develop a non-decomposition method for quaternion-valued BAMNNs.
Next, we will derive non-decomposition fixed-time stabilization criteria for system (2.1).
Assumption 3. Suppose that there exist positive constants l1q,l2p such that the activation functions fq(⋅),gp(⋅)∈Q satisfy the following conditions
|fq(x)−fq(y)|≤l1q|x−y|,q=1,⋯,m|gp(x)−gp(y)|≤l2p|x−y|,p=1,⋯,n. |
Furthermore, fq(0)=gp(0)=0, where x,y∈Q.
The state-feedback controller for system (2.1) is designed as follows
up(t)={−12kp(t)xp(t)|xp(t)|−12θp(t)xp(t)−12λp(t)|xp(t−τ2(t))|2|xp(t)|2xp(t)+ρ(t)2xp(t)|xp(t)|1−α+ϱ(t)2xp(t)|xp(t)|1−β,if|xp(t)|≠00,if|xp(t)|=0,vq(t)={−12γq(t)yq(t)|yq(t)|−12δq(t)yq(t)−12μq(t)|yq(t−τ1(t))|2|yq(t)|2yq(t)+ρ(t)2yq(t)|yq(t)|1−α+ϱ(t)2yq(t)|yq(t)|1−β,if|yq(t)|≠00,if|yq(t)|=0, | (3.6) |
where 0<α<1,β>1, and kp(t), θp(t), λp(t), γq(t), δq(t), μq(t)∈R are control gains to be determined.
Theorem 2. Under Assumption 3, if the control gains kp(t), θp(t), λp(t), γq(t), δq(t), μq(t) satisfy the following conditions
δq(t)≥−2dq(t)+n∑p=1l2p(|˜aqp(t)|+|˜bqp(t)|)+n∑p=1l1q|apq(t)|,θp(t)≥−2cp(t)+m∑q=1l1q(|apq(t)|+|bpq(t)|)+m∑q=1l2p|˜aqp(t)|,kp(t)≥2ˆωp,γq(t)≥2ˆξq,λp(t)≥m∑q=1l2p|˜bqp(t)|,μq(t)≥n∑p=1l1q|bpq(t)| | (3.7) |
and there exist positive constants N,M1,M2,γ1,γ2 such that
∫tt0ρ+(s)ds≤N,∫tt0ρ−(s)ds≤−γ1(t−t0)+M1∫tt0ϱ(s)ds≤−γ2(t−t0)+M2. | (3.8) |
Then, the fixed-time stabilization of system (2.1) can be achieved under controller (3.6).
Proof. Choosing the Lyapunov functional as below
V(t)=n∑p=1x∗p(t)xp(t)+m∑q=1y∗q(t)yq(t) |
Computing the derivative along trajectories (2.1), we have
dV(t)dt=n∑p=1x∗p(t)˙xp(t)+˙x∗p(t)xp(t)+m∑q=1y∗q(t)˙yq(t)+˙y∗q(t)yq(t)≤n∑p=1x∗p(t){−cp(t)xp(t)+m∑q=1apq(t)fq(yq(t))+m∑q=1bpq(t)fq(yq(t−τ1(t)))+ωp(t)+up(t)}+n∑p=1{−cp(t)xp(t)+m∑q=1apq(t)fq(yq(t))+m∑q=1bpq(t)fq(yq(t−τ1(t)))+ωp(t)+up(t)}∗xp(t)+n∑q=1y∗q(t){−dq(t)yq(t)+n∑p=1˜aqp(t)gp(xp(t))+n∑p=1˜bqp(t)gp(xp(t−τ2(t)))+ξq(t)+vq(t)}+n∑q=1{−dq(t)yq(t)+n∑p=1˜aqp(t)gp(xp(t))+n∑p=1˜bqp(t)gp(xp(t−τ2(t)))+ξq(t)+vq(t)}∗yq(t). | (3.9) |
From the feedback controller (3.6), we yield
x∗p(t)up(t)+u∗p(t)xp(t)=2x∗p(t){−12kp(t)xp(t)|xp(t)|−12θp(t)xp(t)−12λp(t)|xp(t−τ2(t))|2|xp(t)|2xp(t)+12ρ(t)xp(t)|xp(t)|1−α+12ϱ(t)xp(t)|xp(t)|1−β}=−kp(t)|xp(t)|−θp(t)|xp(t)|2−λp(t)|xp(t−τ2(t))|2+ρ(t)|xp(t)|α+1+ϱ(t)|xp(t)|β+1 | (3.10) |
y∗q(t)vq(t)+v∗q(t)yq(t)=2y∗q(t){−12γq(t)yq(t)|yq(t)|−12δq(t)yq(t)−12μq(t)|yq(t−τ1(t))|2|yq(t)|2yq(t)+12ρ(t)yq(t)|yq(t)|1−α+12ϱ(t)yq(t)|yq(t)|1−β}=−γq(t)|yq(t)|−δq(t)|yq(t)|2−μq(t)|yq(t−τ1(t))|2+ρ(t)|yq(t)|α+1+ϱ(t)|yq(t)|β+1 | (3.11) |
According to Lemma 3, we get
n∑p=1{x∗p(t)ωp(t)+ω∗p(t)xp(t)}≤2n∑p=1|xp(t)||ωp(t)|≤2n∑p=1|xp(t)|ˆωp, | (3.12) |
m∑q=1{y∗q(t)ξq(t)+ξ∗q(t)yq(t)}≤2m∑q=1|yq(t)||ξq(t)|≤2m∑q=1|yq(t)|ˆξq, | (3.13) |
n∑p=1m∑q=1{x∗p(t)apq(t)fq(yq(t))+f∗q(yq(t))a∗pq(t)xp(t)}≤n∑p=1m∑q=12|apq(t)||xp(t)|l1q|yq(t)|≤n∑p=1m∑q=1l1q|apq(t)|(|yq(t)|2+|xp(t)|2), | (3.14) |
n∑p=1m∑q=1x∗p(t)bpq(t)fq(yq(t−τ1(t)))+n∑p=1m∑q=1f∗q(yq(t−τ1(t)))b∗pq(t)xp(t)≤n∑p=1m∑q=1l1q|bpq(t)|(|yq(t−τ1(t))|2+|xp(t)|2) | (3.15) |
y∗q(t)˜aqp(t)gp(xp(t))+g∗p(xp(t))˜a∗qp(t)yq(t)≤l2p|˜aqp(t)|(|yq(t)|2+|xp(t)|2), | (3.16) |
y∗q(t)˜bqp(t)gp(xp(t−τ2(t)))+g∗p(xp(t−τ2(t)))˜b∗qp(t)yq(t)≤l2p|˜bqp(t)|(|yq(t)|2+|xp(t−τ2(t))|2) | (3.17) |
Combining the above inequalities, we have
dV(t)dt≤n∑p=1{−2cp(t)+m∑q=1l1q(|apq(t)|+|bpq(t)|)+m∑q=1l2p|˜aqp(t)|−θp(t)}|xp(t)|2+m∑q=1{−2dq(t)+n∑p=1l2p(|˜aqp(t)|+|˜bqp(t)|)+n∑p=1l1q|apq(t)|−δq(t)}|yq(t)|2+n∑p=1{2ˆωp−kp(t)}|xp(t)|+m∑q=1{2ˆξq−γq(t)}|yq(t)|+m∑q=1{n∑p=1l1q|bpq(t)|−μq(t)}|yq(t−τ1(t))|2+n∑p=1{m∑q=1l2p|˜bqp(t)|−λp(t)}|xp(t−τ2(t))|2+ρ(t){n∑p=1|xp(t)|α+1+m∑q=1|yq(t)|α+1}+ϱ(t){n∑p=1|xp(t)|β+1+m∑q=1|yq(t)|β+1}. | (3.18) |
According to Lemma 2 and the condition in (3.7), we get
dV(t)dt≤(ρ+(t)+ρ−(t)){n∑p=1|xp(t)|α+1+m∑q=1|yq(t)|α+1}+ϱ(t){n∑p=1|xp(t)|β+1+m∑q=1|yq(t)|β+1}≤ρ+(t)(n+m)1−α2{n∑p=1|xp(t)|2+m∑q=1|yq(t)|2}α+12+ρ−(t){n∑p=1|xp(t)|2+m∑q=1|yq(t)|2}α+12+ϱ(t)(n+m)1−β2{n∑p=1|xp(t)|2+m∑q=1|yq(t)|2}β+12={ρ+(t)(n+m)1−α2+ρ−(t)}V(t)α+12+(n+m)1−β2ϱ(t)V(t)β+12. | (3.19) |
Thus, the fixed-time stabilization of system (2.1) can be achieved under feedback controller (3.6). Furthermore, the settling time is estimated as
Tmax=t0+(1−α)((n+m)1−α2N+M1)+1γ1(1−α)+(β−1)((n+m)1−α2N+(n+m)1−β2M2)+1(n+m)β−12γ2(β−1). | (3.20) |
The proof is completed.
Remark 5. In most of the existing literature on the stability of NNs [22,23,24,25], the derivative of the Lyapunov functions are required to be negative definite, which is a strong limitation in practical use. Compared with that, since ρ(t) can be chosen as positive value, the derivative of the Lyapunov function in this work may not be negative definite. Thus, our result can be applied to a broader engineering area since the limitation of traditional Lyapunov functions are relaxed here.
Remark 6. The model proposed in this work combines both the characteristics of QVNNs and BAMNNs. As we know, the BAMNNs have a great information association and information memory ability. On the other hand, the imaginary parts of the quaternion neuron can better describe the human visual neuron cells and deal with multidimensional data more efficiently. Based on these strong engineering background, the BAM QVNNs model has a good application prospect and deserves further investigation.
It is well known that an adaptive control strategy can greatly reduce the control strength. In the following, we further investigate the fixed-time stabilization via an adaptive method. The update law for the adaptive parameters in controller (3.6) is
˙kp(t)=|xp(t)|+(12)α+12ρ(t)|kp(t)−kp|αsgn(kp(t)−kp)+(12)β+12ϱ(t)|kp(t)−kp|βsgn(kp(t)−kp),˙θp(t)=|xp(t)|2+(12)α+12ρ(t)|θp(t)−θp|αsgn(θp(t)−θp)+(12)β+12ϱ(t)|θp(t)−θp|βsgn(θp(t)−θp),˙λp(t)=|xp(t−τ2(t))|2+(12)α+12ρ(t)|λp(t)−λp|αsgn(λp(t)−λp)+(12)β+12ϱ(t)|λp(t)−λp|βsgn(λp(t)−λp),˙γq(t)=|yq(t)|+(12)α+12ρ(t)|γq(t)−γq|αsgn(γq(t)−γq)+(12)β+12ϱ(t)|γq(t)−γq|βsgn(γq(t)−γq),˙δq(t)=|yq(t)|2+(12)α+12ρ(t)|δq(t)−δq|αsgn(δq(t)−δq)+(12)β+12ϱ(t)|δq(t)−δq|βsgn(δq(t)−δq),˙μq(t)=|yq(t−τ1(t))|2+(12)α+12ρ(t)|μq(t)−μq|αsgn(μq(t)−μq)+(12)β+12ϱ(t)|μq(t)−μq|βsgn(μq(t)−μq) | (3.21) |
where p=1,⋯,n,q=1,⋯,m, 0<α<1,β>1. kp,θp,λp,γq,δq,μq are positive constants to be determined.
Assumption 4. Suppose that the time-varying coefficients apq(t),bpq(t),˜apq(t),˜bpq(t) are bounded, cp(t),dq(t) are lower bounded. i.e., there exist positive constants aMpq, bMpq, ˜aMpq,˜bMpq, cmp,dmq such that |apq(t)|≤aMpq, |bpq(t)|≤bMpq, |˜apq(t)|≤˜aMpq, |˜bMpq(t)|≤˜bMpq, cp(t)≥cmp, and dq(t)≥dmq.
Theorem 3. Under Assumption 3 and 4, if the condition (3.8) in Theorem 2 holds. Then, the fixed-time stabilization of system (2.1) can be achieved under controller (3.6) and adaptive law (3.21).
Proof. Considering the Lyapunov functional as below
V(t)=V1(t)+V2(t) |
and
V1(t)=n∑p=1x∗p(t)xp(t)+m∑q=1y∗q(t)yq(t)V2(t)=12n∑p=1(kp(t)−kp)2+12n∑p=1(θp(t)−θp)2+12n∑p=1(λp(t)−λp)2+12m∑q=1(γq(t)−γq)2+12m∑q=1(δq(t)−δq)2+12m∑q=1(μq(t)−μq)2 | (3.22) |
where the parameters are
δq=−2dmq+n∑p=1l2p(˜aMqp+˜bMqp)+n∑p=1l1qaMpq,θp=−2cmp+m∑q=1l1q(aMpq+bMpq)+m∑q=1l2p˜aMqp,kp=2ˆωp,γq=2ˆξq,μq=n∑p=1l1qbMpq,λp=m∑q=1l2p˜bMqp | (3.23) |
Calculating the derivative of V2(t), we get
dV2(t)dt=n∑p=1(kp(t)−kp)˙kp(t)+n∑p=1(θp(t)−θp)˙θp(t)+n∑p=1(λp(t)−λp)˙λp(t)+m∑q=1(γq(t)−γq)˙γq(t)+m∑q=1(δq(t)−δq)˙δq(t)+m∑q=1(μq(t)−μq)˙μq(t)=n∑p=1(kp(t)−kp)|xp(t)|+n∑p=1(θp(t)−θp)|xp(t)|2+n∑p=1(λp(t)−λp)|xp(tτ)|2+m∑q=1(γq(t)−γq)|yq(t)|+m∑q=1(δq(t)−δq)|yq(t)|2+m∑q=1(μq(t)−μq)|yq(tτ)|2+ρ(t)n∑p=1(12)α+12{|kp(t)−kp|α+1+|θp(t)−θp|α+1+|λp(t)−λp|α+1}+ρ(t)m∑q=1(12)α+12{|γq(t)−γq|α+1+|δq(t)−δq|α+1+|μq(t)−μq|α+1}+ϱ(t)n∑p=1(12)β+12{|kp(t)−kp|β+1+|θp(t)−θp|β+1+|λp(t)−λp|β+1}+ϱ(t)m∑q=1(12)β+12{|γq(t)−γq|β+1+|δq(t)−δq|β+1+|μq(t)−μq|β+1} | (3.24) |
Combining (3.24) with (3.19) in Theorem 2, and applying Assumption 3, we have
\begin{align} \frac{dV(t)}{dt}\leq &\rho(t)\sum\limits_{p = 1}^{n}\{|x_{p}(t)|^{\alpha+1}+(\frac{|k_{p}(t)-k_{p}|^{2}}{2})^{\frac{\alpha+1}{2}}+(\frac{|\theta_{p}(t)-\theta_{p}|^{2}}{2})^{\frac{\alpha+1}{2}}+(\frac{|\lambda_{p}(t)-\lambda_{p}|^{2}}{2})^{\alpha+1}\}\\ &+\rho(t)\sum\limits_{q = 1}^{m}\{|y_{q}(t)|^{\alpha+1}+(\frac{|\gamma_{q}(t)-\gamma_{q}|^{2}}{2})^{\frac{\alpha+1}{2}}+(\frac{|\delta_{q}(t)-\delta_{q}|^{2}}{2})^{\frac{\alpha+1}{2}}+(\frac{|\mu_{q}(t)-\mu_{q}|^{2}}{2})^{\frac{\alpha+1}{2}}\}\\ &+\varrho(t)\sum\limits_{p = 1}^{n}\{|x_{p}(t)|^{\beta+1}+(\frac{|k_{p}(t)-k_{p}|^{2}}{2})^{\frac{\beta+1}{2}}+(\frac{|\theta_{p}(t)-\theta_{p}|^{2}}{2})^{\frac{\beta+1}{2}}+(\frac{|\lambda_{p}(t)-\lambda_{p}|^{2}}{2})^{\frac{\beta+1}{2}}\}\\ &+\varrho(t)\sum\limits_{q = 1}^{m}\{|y_{q}(t)|^{\beta+1}+(\frac{|\gamma_{q}(t)-\gamma_{q}|^{2}}{2})^{\frac{\beta+1}{2}}+(\frac{|\delta_{q}(t)-\delta_{q}|^{2}}{2})^{\frac{\beta+1}{2}}+(\frac{|\mu_{q}(t)-\mu_{q}|^{2}}{2})^{\frac{\beta+1}{2}}\} \end{align} | (3.25) |
According to Lemma 2, we yield
\begin{align} \frac{dV(t)}{dt} \leq&\{(4n+4m)^{\frac{1-\alpha}{2}}\rho^{+}(t)+\rho^{-}(t)\}V_{1}^{\frac{\alpha+1}{2}}(t)+(4n+4m)^{\frac{1-\beta}{2}}\varrho(t)V^{\frac{\beta+1}{2}}_{2}(t) \end{align} | (3.26) |
Based on above discussion, system (2.1) can be stabilized with the fixed-time controller (3.6) and the adaptive law (3.21). According to Lemma 1, the settling time can be estimated as
\begin{align} T_{\max} = &t_{0}+\frac{(1-\alpha)((4n+4m)^{\frac{1-\alpha}{2}}N+M_{1})+1}{\gamma_{1}(1-\alpha)}\\ &+\frac{(\beta-1)\{(4n+4m)^{\frac{1-\alpha}{2}}N+(4n+4m)^{\frac{1-\beta}{2}}M_{2}\}+1}{(4n+4m)^{\frac{\beta-1}{2}}\gamma_{2}(\beta-1)} \end{align} | (3.27) |
Remark 7. In many previous studies on NNs [21,22,23,24,25], the time-varying delays are assumed to be bounded or derivative bounded. Compared with that, the time-varying delays \tau_{1}(t), \tau_{2}(t) considered in our paper do not have such restrictions. The basic assumption \tau(t)\leq\tau, \dot{\tau}(t)\leq1 is removed. Thus, the network model proposed in this work is more general and our research is very meaningful.
Remark 8. In Theorem 2 and 3, the QVNNs is coped with its quaternion form without a decomposition. Therefore, it can be used to deal with the activations or connection weights that cannot be explicitly expressed as real and imaginary parts, which make our results more general and promising for real applications. However, a disadvantage of this approach is that the Hamilton rule has not been fully taken into consideration, thus the effect of the sign of each real and imaginary parts are neglected. Hence, both decomposition and non-decomposition method are proposed in our work.
Remark 9. In [11,12,13,14,15], the dynamical behaviors of NNs are investigated by linear feedback control. The drawback of traditional feedback control is that the control gains must be maximal. Compared with that, by applying the adaptive method in this work, the control strength can be effectively decreased. Another advantage is that we even do not need to calculate the exact value of control gains, which is quite meaningful in real applications.
To verify the effectiveness of our theoretical results, some simulation examples are presented. Firstly, we consider the simulation of Theorem 1.
Example 1. Consider the following time-varying BAM QVNNs
\begin{align} \frac{dx_{p}(t)}{dt} = &-c_{p}(t)x_{p}(t)+\sum\limits_{q = 1}^{2}a_{pq}(t)f_{q}(y_{q}(t))+\sum\limits_{q = 1}^{2}b_{pq}(t)f_{q}(y_{q}(t-\tau_{1}(t)))+\omega_{p}(t)+u_{p}(t)\\ \frac{dy_{q}(t)}{dt} = &-d_{q}(t)y_{q}(t)+\sum\limits_{p = 1}^{2}\tilde{a}_{qp}(t)g_{p}(x_{p}(t))+\sum\limits_{p = 1}^{2}\tilde{b}_{qp}(t)g_{p}(x_{p}(t-\tau_{2}(t)))+\xi_{q}(t)+v_{q}(t) \end{align} | (4.1) |
where d_{1}(t) = d_{2}(t) = 2+|\cos t|, c_{1}(t) = c_{2}(t) = 1+|\sin t| , \omega_{p}(t) = \sin(t), \xi_{q}(t) = -\cos(t) . The transmission delay is \tau_{1}(t) = 0.5\sin(t)+0.5, \tau_{2}(t) = 0.5\cos(t)+0.5 . The time-varying connection weights are given as below.
\begin{align} &a_{pq}(t) = \frac{1}{1+t^{2}}+\frac{1}{1+t}i+\frac{1}{1+\sqrt{t}}j+\frac{1}{1+t}k, \; \; b_{pq}(t) = \frac{\sin t}{1+t}+\cos t\cdot i+\sin t\cdot j+\frac{\sin t}{1+\sqrt{t}}k\\ &\tilde{a}_{qp}(t) = \cos t+\frac{1}{1+t}i+\frac{1}{1+\sqrt{t}}j+\frac{1}{1+t}k, \; \; \tilde{b}_{qp}(t) = \sin t+\frac{\cos t}{1+t}\cdot i+\sin t\cdot j+\frac{\sin t}{1+\sqrt{t}}k \end{align} |
The activation functions are chosen as
\begin{align} f_{q}(x_{q}(t)) = &-0.8\sin(x^{R}_{q}(t))+0.6\cos(x^{I}_{q}(t))i-0.8\cos(x^{J}_{q}(t))j+0.6\sin(x^{K}_{q}(t))\}k, q = 1, 2. \end{align} |
\begin{align} g_{p}(x_{p}(t)) = &-0.5\sin(x^{R}_{p}(t))+0.5\cos(x^{I}_{p}(t))i-0.5\cos(x^{J}_{p}(t))j+0.5\sin(x^{K}_{p}(t))\}k, p = 1, 2. \end{align} |
Based on the above conditions, it is not difficult to check that L_{1} = diag(0.8, 0.8, 0.6, 0.6, 0.8, 0.8, 0.6, 0.6) , L_{2} = diag(0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5) . Thus, the conditions in Assumption 1, 2 are satisfied. Applying the quaternion decomposition, the real-valued compact system is achieved as
\begin{align} \left\{ \begin{aligned} \frac{dX(t)}{dt} = &-C^{*}(t)X(t)+A^{*}(t)F(Y(t))+B^{*}(t)F(Y(t-\tau_{1}(t)))+\omega^{*}(t)+U(t), \\ \frac{dY(t)}{dt} = &-D^{*}(t)Y(t)+\tilde{A}^{*}(t)G(X(t))+\tilde{B}^{*}(t)G(X(t-\tau_{2}(t)))+\xi^{*}(t)+V(t). \end{aligned} \right. \end{align} | (4.2) |
The feedback controller is in the following form
\begin{align} U(t) = &-K(t)X(t)-\Gamma(t)sgn(X(t))-\Lambda(t)SGN(X(t))|X(t-\tau_{2}(t))|\\ &+\rho(t)SGN(X(t))|X(t)|^{\alpha}+\varrho(t)SGN(X(t))|X(t)|^{\beta}, \\ V(t) = &-\Theta(t)Y(t)-\Upsilon(t)sgn(Y(t))-\Xi(t)SGN(Y(t))|Y(t-\tau_{1}(t))|\\ &+\rho(t)SGN(Y(t))|Y(t)|^{\alpha}+\varrho(t)SGN(Y(t))|Y(t)|^{\beta} \end{align} | (4.3) |
where K_{p}(t) = 7-|\sin t| , \Gamma_{p}(t) = 1 , \Lambda_{p}(t) = 8 , \Theta_{q}(t) = 10.8-|\cos t| , \Upsilon_{q}(t) = 1 , \Xi_{q}(t) = 12.8 , \rho(t) = \frac{1}{1+t^{2}}-\frac{t}{2}|\cos t|, \varrho(t) = -t|\sin t|\notag . It can be checked that
\begin{align} &K_{p}(t)\geq-C_{p}^{*}(t)+\sum\limits_{q = 1}^{4m}|\tilde{a}^{*}_{qp}(t)|L_{2p}, \; \; \Theta_{q}(t)\geq-D_{q}^{*}(t)+\sum\limits_{p = 1}^{4n}|a^{*}_{pq}(t)|L_{1q}, \\ &\Lambda_{p}(t)\geq \sum\limits_{q = 1}^{4m}|\tilde{b}^{*}_{qp}(t)|L_{2p}, \; \; \Xi_{q}(t)\geq \sum\limits_{p = 1}^{4n}|b^{*}_{pq}(t)|L_{1q}, \; \; \Gamma_{p}(t)\geq|\omega_{p}^{*}(t)|, \; \; \Upsilon_{q}(t)\geq|\xi_{q}^{*}(t)|, \\ &\int_{t_{0}}^{t}\rho^{+}(s)ds\leq N, \; \; \int_{t_{0}}^{t}\rho^{-}(s)ds\leq -\gamma_{1}(t-t_{0})+M_{1}, \; \; \int_{t_{0}}^{t}\varrho(s)ds\leq -\gamma_{2}(t-t_{0})+M_{2} \end{align} | (4.4) |
Therefore, K_{p}(t) , \Gamma_{p}(t) , \Lambda_{p}(t) , \Theta_{q}(t) , \Xi_{q}(t) , \Upsilon_{q}(t) satisfy condition (3.2), and \rho(t), \varrho(t) satisfy condition (3.8) with N = \frac{\pi}{2}, \gamma_{1} = \frac{2}{3}\pi, M_{1} = 1, \gamma_{2} = \frac{4}{3}\pi, M_{2} = 2 . Thus, the conditions in Theorem 3.1 are satisfied so that system (4.2) can be stabilized with fixed-time controller (4.3). Choosing 10 initial values in the interval [-0.5, 0.5] , it is depicted in Figures 1–4 that each imaginary part of system (4.1) will converge to zero with fixed-time controller (4.3). Thus, the correctness of Theorem 3.1 is verified.
Next, we consider the Theorem 3.
Example 2. Consider system (4.1) under the adaptive controller (3.6) and the adaptive law (3.21) in Theorem 3. According to the condition in Example 1, the time-varying coefficients are satisfied that |a_{pq}(t)|\leq a^{M}_{pq} = 2 , |b_{pq}(t)|\leq b^{M}_{pq} = 2 , |\tilde{a}_{qp}(t)|\leq \tilde{a}^{M}_{qp} = 2 , |\tilde{b}_{qp}(t)|\leq\tilde{b}^{M}_{qp} = 2 , c_{p}(t)\geq c^{m}_{p} = 1, d_{q}(t)\geq d^{m}_{q} = 2 , l_{1q} = 0.8 , l_{2p} = 0.5 . The parameters are chosen as
\begin{align} &\delta_{q} = -2d^{m}_{q}+\sum\limits_{p = 1}^{n}l_{2p}(\tilde{a}^{M}_{qp}+\tilde{b}^{M}_{qp})+\sum\limits_{p = 1}^{n}l_{1q}a^{M}_{pq} = 3.2, \\ &\theta_{p} = -2c^{m}_{p}+\sum\limits_{q = 1}^{m}l_{1q}(a^{M}_{pq}+b^{M}_{pq})+\sum\limits_{q = 1}^{m}l_{2p}\tilde{a}^{M}_{qp} = 6.4, \\ &k_{p} = 2\hat{\omega}_{p} = 2, \gamma_{q} = 2\hat{\xi}_{q} = 2, \\ &\mu_{q} = \sum\limits_{p = 1}^{n}l_{1q}b^{M}_{pq} = 3.2, \lambda_{p} = \sum\limits_{q = 1}^{m}l_{2p}\tilde{b}^{M}_{qp} = 2 \end{align} | (4.5) |
Thus, the conditions in Theorem 3 are satisfied. Take 10 initial random values in the interval [-0.8, 0.8] , Figures 5–8 describe the state trajectories of x^{R}_{1}(t) , x^{I}_{1}(t) , x^{J}_{1}(t) , x^{K}_{1}(t) , y^{R}_{2}(t) , y^{I}_{2}(t) , y^{J}_{2}(t) , y^{K}_{2}(t) of system (4.1) under the adaptive controller (3.6), respectively. According to our numerical results, the fixed-time stabilization of system (4.1) can be achieved with the adaptive controller (3.6) and the adaptive law (3.21). Thus, the effectiveness of Theorem 3 is verified.
In this paper, the traditional BAMNNs has been extended to quaternion area. A newly fixed-time convergence theory is employed to solve the fixed-time stabilization of time-varying BAM QVNNs, which has a great advantage compared with traditional fixed-time control methods. Due to the non-commutativity feature of quaternions, two different control approaches are proposed, a decomposition and a non-decomposition method. Moreover, with the designing of an adaptive controller and an update law, an adaptive control strategy is proposed to reduce the control cost and improve the efficiency. Lastly, numerical examples are given to verify our results.
Our future research will focus on these directions: 1) The dynamical behavior of stochastic QVNNs, 2) The basin stability of QVNNs.
This research was supported by Scientific Research Foundation of Graduate School of Southeast University under Grant Nos. YBPY1919 and Innovation project for College Graduates of Jiangsu Province of China under Grant No. KYCX18-0054.
All authors declare no conflicts of interest in this paper.
[1] | B. Li, S. Tan and M. Wang, et al., Investigation on cost assignment in spatial image steganography, IEEE. TIFS., 9 (2014), 1264–1277. |
[2] | Y. Ma, X. Luo and X. Li, et al., Selection of rich model steganalysis features based on decision rough set α-Positive region reduction, IEEE. TCSVT., (2018), In press. |
[3] | Y. Zhang, C. Qin and W. Zhang, et al., On the fault-tolerant performance for a class of robust image steganography, J. Sigpro., 146 (2018), 99–111. |
[4] | S. Li and X. Zhang, Towards construction based data hiding: from secrets to fingerprint images, IEEE. TIP., 28 (2019), 1482–1497. |
[5] | J.Frdrich and D. Soukal, Matrix embedding for large payloads, IEEE. TIFS., 1(2006), 390–395. |
[6] | X. Zhang and S. Wang, Efficient steganographic embedding by exploiting modification direction, IEEE. LCOMM., 10 (2016), 781–783. |
[7] | W. Zhang, X. Zhang and S. Wang, Maximizing steganographic embedding efficiency by combining hamming codes and wet paper codes, Proc. IH2008, Santa Barbara, CA, USA, (2008), 60–71. |
[8] | T. Filler, J.Judas and J. Fridrich, Minimizing additive distortion in steganography using syndrome-trellis codes, IEEE. TIFS., 6 (2011), 920–935. |
[9] | X. Luo, X. Song and X. Li, et al., Steganalysis of HUGO steganography based on parameter recognition of syndrome-trellis-codes, MTAP., 75 (2016), 13557–13583. |
[10] | T. Pevný, T. Filler and P. Bas, Using high-dimensional image models to perform highly undetectable steganography, Proc. IH2008, Calgary, Canada, (2010), 161–177. |
[11] | V. Holub and J. Fridrich, Designing steganographic distortion using directional filters, Proc. IEEE. WIFS., Binghamton, NY, USA, (2012), 234–239. |
[12] | B Li, M. Wang, and J. Huang, et al., A new cost function for spatial image steganography, Proc. IEEE. ICIP., Paris, France, (2014), 4206–4210. |
[13] | V. Sedighi, J. Fridrich and R. Cogranne, Content-adaptive pentary steganography using the multivariate generalized gaussian cover model, Proc. SOP., San Francisco, California, USA, (2015), 94090H–94090H–13. |
[14] | V. Holub and J. Fridrich, Digital image steganography using universal distortion, Proc. ACM IH&MMS., New York, NY, USA, (2013), 59–68. |
[15] | L. J. Guo, J. Q. Ni and Y. Q. Shi, uniform embedding for efficient JPEG steganography, IEEE. TIFS., 9 (2014), 814–825. |
[16] | L. J. Guo, J. Q. Ni and W. K. Su, et al., Using statistical image model for JPEG steganography: Uniform Embedding Revisited, IEEE. TIFS., 10 (2015), 2669–2680. |
[17] | Z. Wang, X. Zhang and Z. Yin, Hybrid distortion function for JPEG steganography, JEI., 25 (2016), 050501. |
[18] | Z. Wang, Z. Qian and X. Zhang, et al., On improving distortion functions for JPEG steganography, IEEE Access, 6 (2018), 74917–74930. |
[19] | C. Qin, C. Chang and Y. Chiu, A novel joint data-hiding and compression scheme based on SMVQ and image inpainting, IEEE. TIP., 23 (2014), 969–978. |
[20] | Z. Qian, H. Zhou and X. Zhang, et al., Separable reversible data hiding in encrypted JPEG bitstreams, IEEE. TDSC., 15 (2018), 1055–1067. |
[21] | Z. Qian and X. Zhang, Reversible data hiding in encrypted image with distributed source encoding, IEEE. TCSVT., 26 (2016), 636–646. |
[22] | Z. Qian, X. Zhang and S. Wang, Reversible data hiding in encrypted JPEG bitstream, IEEE. TMM., 16 (2014), 1486–1491. |
[23] | C. Qin, W. Zhang and F. Cao, et al., Separable reversible data hiding in encrypted images via adaptive embedding strategy with block selection, J. Sigpro., 153 (2018), 09–122. |
[24] | J. Deng, W. Dong and R. Socher, et al., ImageNet: A large-scale hierarchical image satabase, IEEE CVPR., (2009), 248–255. |
[25] | Beautification softwares "BeautyCam", Meitu Company, Fujian, China, 2018, Available from: https://mt.meipai.com/. |
[26] | Beautification softwares "CAMERA360", Chengdu Pinguo technology co. LTD, Chengdu, China, 2015, Available from: http://www.camera360.com/. |
[27] | Beautification softwares "TianTianPiTu", Tencent Company, Shenzhen, China, 2014, Available from: https://tu.qq.com/. |
[28] | P. Joshi, D. M. Esciriva and V. Godoy, Opencv By example, Penerbit Packt Publishing Ltd, (2016), 87–92. |
[29] | Z. Wang, J. Lv and Q. Wei, et al., Distortion function for spatial image steganography based on the polarity of embedding change, Proc. IWDW., Beijing, China, (2016), 487–493. |
[30] | V. Holub and J. Fridrich, Low complexity features for JPEG steganalysis using undecimated DCT, IEEE. TIFS.,10 (2014), 219–228. |
[31] | J. Kodovsky, J. Fridrich and V. Holub, Ensemble classifiers for steganalysis of digital media, IEEE. TIFS., 7 (2012), 432–444. |
[32] | J. Kodovsky and J. Fridrich, Steganalysis of JPEG images using rich models, Proc. ISOP., San Francisco, CA, USA, (2012), 83030A–83030A–13. |
1. | Tengda Wei, Xiang Xie, Xiaodi Li, Input-to-state stability of delayed reaction-diffusion neural networks with multiple impulses, 2021, 6, 2473-6988, 5786, 10.3934/math.2021342 | |
2. | Adnène Arbi, Najeh Tahri, J. Zhu, Almost anti-periodic solution of inertial neural networks model on time scales, 2022, 355, 2261-236X, 02006, 10.1051/matecconf/202235502006 |