Resilient H State Estimation for Discrete-Time Stochastic Delayed Memristive Neural Networks a Dynamic Event-Triggered Mechanism
Resilient H State Estimation for Discrete-Time Stochastic Delayed Memristive Neural Networks a Dynamic Event-Triggered Mechanism
Abstract—In this article, a resilient H∞ approach is put forward to prospective applications in various domains, such as signal pro-
deal with the state estimation problem for a type of discrete-time delayed cessing, combinatorial optimization, image processing, and pattern
memristive neural networks (MNNs) subject to stochastic disturbances
recognition [1], [6], [9], [12], [20], [26], [33], [39], [47]. According to
(SDs) and dynamic event-triggered mechanism (ETM). The dynamic
ETM is utilized to mitigate unnecessary resource consumption occur- the connection manner of the neurons and the transmission direction
ring in the sensor-to-estimator communication channel. To guarantee of the signals, the ANNs can be roughly divided into two categories,
resilience against possible realization errors, the estimator gain is per- namely, feedforward neural networks (FNNs) and the recurrent neu-
mitted to undergo some norm-bounded parameter drifts. For the delayed ral networks (RNNs). Different from the FNNs that mainly receive
MNNs, our aim is to devise an event-based resilient H∞ estimator that
not only resists gain variations and SDs but also ensures the exponen- inputs with certain structures, the RNNs allow the states to be trans-
tial mean-square stability of the resulting estimation error system with a ferred among the networks and are therefore able to accept inputs
guaranteed disturbance attenuation level. By resorting to the stochastic with extensive structures [14], [15], [23], [24], [36], [46]–[48], [50].
analysis technique, sufficient conditions are acquired for the expected In practice, the RNN can be realized by integration circuits, where
estimator and, subsequently, estimator gains are obtained via figuring
out a convex optimization problem. The validity of the H∞ estimator is
the synapses of the neurons are executed via resistors and each
finally shown via a numerical example. neuron is linked with multiple resistors. Nonetheless, the resistor
itself has inherent weakness regardless of its usefulness in realiz-
Index Terms—Dynamic event-triggered mechanism (ETM), H∞ ing some specific functions. Taking the neural circuit as an example,
performance, memristive neural networks (MNNs), resilient state the volatility of the resistor renders the disappearance of the stored
estimation, time delays. information when the voltage is cut off and, moreover, the huge
amount of resistors leads inevitably to a substantial reduction of
the integration degree of the neural circuits. As such, there is a
I. I NTRODUCTION practically urgent demand to develop alternative devices that could
Over the last few decades, a persistent research interest has not only preserve the stored information in the absence of volt-
been paid to the artificial neural networks (ANNs) because of their age but also help to reduce the number of components in the
circuit.
Manuscript received April 18, 2020; revised August 1, 2020; accepted Due to the enormous success made by the Hewlett-Packard Lab
August 31, 2020. Date of publication October 1, 2020; date of current ver-
sion May 19, 2022. This work was supported in part by the National Natural on experimental prototyping, the memristors have been introduced
Science Foundation of China under Grant 61933007 and Grant 61873148; to take place of the resistors in implementing RNNs for the purpose
in part by the Heilongjiang Province Postdoctoral Science Foundation of of achieving higher integration of the circuits and better process-
China under Grant LBH-Z19048; in part by the Natural Science Foundation ing of the information [3]. In particular, the so-called memristive
of Universities in Anhui Province of China under Grant gxyqZD2019053
and Grant KJ2019A0160; in part by the AHPU Youth Top-Notch Talent
neural networks (MNNs) have attracted much attention in various
Support Program; in part by the Key Laboratory of Advanced Perception domains, such as combinatorial optimization, knowledge acquisition,
and Intelligent Control of High-End Equipment of Ministry of Education and brain emulation [27], [34]. A key feature of the MNNs is their
in Anhui Polytechnic University of China under Grant GDSC202014 and strong dependence of the connection weights (CWs) on the neuron
Grant GDSC202016; in part by the Royal Society of the U.K.; and in
states (NSs), which results in complicated dynamic behaviors of the
part by the Alexander von Humboldt Foundation of Germany. This arti-
cle was recommended by Associate Editor S. Tong. (Corresponding author: MNNs compared to the conventional RNNs. Accordingly, the anal-
Zidong Wang.) ysis problems (e.g., stability, convergence, and synchronization) of
Hongjian Liu is with the Key Laboratory of Advanced Perception and many kinds of dynamics of the MNNs have aroused much research
Intelligent Control of High-End Equipment, Ministry of Education, Anhui attention, and some pioneering work has been available in [7], [21],
Polytechnic University, Wuhu 241000, China, and also with the Artificial
Intelligence Energy Research Institute, Northeast Petroleum University, [24], [41]–[43], and [46].
Daqing 163318, China (e-mail: [email protected]). A prerequisite for many engineering applications of MNNs is the
Zidong Wang is with the College of Electrical Engineering and Automation, exact acknowledgment of the NSs. In fact, for many MNN appli-
Shandong University of Science and Technology, Qingdao 266590, China, and cations, it is often unrealistic to acquire the precise information
also with the Department of Computer Science, Brunel University London,
of the NSs that are necessary for fulfilling the requirements of
Uxbridge UB8 3PH, U.K. (e-mail: [email protected]).
Weiyin Fei is with the Key Laboratory of Advanced Perception and monitoring, approximation, and optimization [49], [52]. Due mainly
Intelligent Control of High-End Equipment, Ministry of Education, Anhui to resource constraints and technical limits, the NSs are usu-
Polytechnic University, Wuhu 241000, China, and also with the School of ally unavailable to end users but only the measurement outputs
Mathematics and Physics, Anhui Polytechnic University, Wuhu 241000, China can be accessible [5], [11], [25], [28]–[31], [44], [45]. On the
(e-mail: [email protected]).
Jiahui Li is with the Artificial Intelligence Energy Research Institute, other hand, the implementation of the MNNs in engineering prac-
Northeast Petroleum University, Daqing 163318, China, and also with tice is easily susceptible to stochastic disturbances (SDs) induced
the Heilongjiang Provincial Key Laboratory of Networking and Intelligent by random fluctuations of the environment or the network itself,
Control, Northeast Petroleum University, Daqing 163318, China (e-mail: for example, the release of neurotransmitters [28]. These prob-
[email protected]).
Color versions of one or more figures in this article are available at
lems, if not adequately addressed, could significantly degrade the
https://doi.org/10.1109/TCYB.2020.3021556. performance and reliability of the MNNs. As such, much research
Digital Object Identifier 10.1109/TCYB.2020.3021556 attention has recently been drawn toward state estimation problems
2168-2267
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
3334 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022
(SEPs) for MNNs with SDs (see [2], [22], [28], and the references
therein).
When realizing filter/controller algorithms in traditional system
design, it has been explicitly assumed that the filter/controller gains
can always be exactly implemented. This assumption goes, unfor-
tunately, against the engineering practice as the actual execution of
the filter/controller gains might undergo parameter variations/drifts Fig. 1. Remote state estimation under the dynamic ETM.
originating from many factors, including the analog–digital conver-
sion, finite resolution of instruments, and rounding errors [10]. Such II. P ROBLEM F ORMULATION
parameter variations/drifts, though possibly small, would undoubtedly Consider a class of stochastic delayed MNNs in the following form:
impair the performance of the expected filter/controller [13], [32]. ⎧
Accordingly, the design of resilient (also called nonfragile) estima- ⎪
⎪ z(s + 1) = D(z(s))z(s) + A(z(s))f (z(s))
⎪
⎪
tors that are insensitive to gain fluctuations has become a research ⎨ + B(z(s))f (z(s − τ (s))) + L(s)v(s)
hotspot. For example, in [13], a resilient finite-time l2 − l∞ SEP has + σ (s, z(s), z(s − τ (s)))w(s) (1)
⎪
⎪
been addressed for Markov jumping neural networks with unreliable ⎪
⎪ y(s) = Cz(s) + Mv(s)
⎩
communication. Nevertheless, to the best of our knowledge, very few z̄(s) = N(s)z(s)
results have been acquired so far on the resilient SEPs for MNNs, where
not to mention the case where the time delays and SDs are both T
involved. z(s) z1 (s) z2 (s) ··· zn (s)
The circuit implementation of large-scale MNNs is often faced T
y(s) y1 (s) y2 (s) ··· ym (s)
with the consumption of a huge amount of various resources (e.g.,
processing, storage, and communication) and, thus, the resource- D(z(s)) diag{d1 (z1 (s)), d2 (z2 (s)), . . . , dn (zn (s))}
saving issue has emerged as a hot topic when addressing the SEPs are the NS vector, the ideal measurement output, and the self-
for MNNs [4], [51]. To reduce unnecessary resource consumption, feedback matrix, respectively. z̄(s) ∈ Rp is the NS to be estimated.
the event-triggered mechanism (ETM) serves as an efficient means di (zi (s)) > 0. A(z(s)) = (aij (zi (s)))n×n and B(z(s)) = (bij (zi (s)))n×n
when implementing the sensor-to-estimator communication, where are CWs with no delays and discrete delays, respectively
the current data are sent only if the prescribed triggering conditions T
(TCs) are satisfied [16]. It is worth mentioning that most available f (z(s)) = f1 (z1 (s)) f2 (z2 (s)) · · · fn (zn (s))
ETMs are static, namely, the thresholds of the TCs are fixed (rather
than adaptive or dynamic) [22], [35]. To further conserve resource is the nonlinear neuron activation function (NAF). v(s) ∈ l2 [0, +∞)
expenses, the so-called dynamic ETMs have been established where is the disturbance vector, and w(s) is a scale Wiener process
the threshold parameters are dynamically adjusted at each checking (Brownian motion) defined on (, F , P) where
time [8], [17], [18], [37], [38]. For example, in [17], the synchro-
E{w(s)} = 0, E w2 (s) = 1, E{w(s)w(t)} = 0, (s = t) (2)
nization control problem has been investigated for a class of discrete
time-delay complex dynamical networks. For the efficiency of energy and σ : R × Rn × Rn → Rn is the noise intensity function satisfying
utilization, a dynamic event-triggering strategy has been introduced
into the design of synchronization controllers. In comparison with σ T (s, a, b)σ (s, a, b) ≤ ρ1 aT a + ρ2 bT b ∀a, b ∈ Rn (3)
its static counterpart, the dynamic ETM is capable of decreasing the
frequency of event releasing, thereby getting around needless data where ρ1 and ρ2 are known positive constants. The positive integer
transmissions and achieving conservative solutions. As such, it is τ (s) is the time delay satisfying
of practical significance to explore how the dynamic ETM can be τm ≤ τ (s) ≤ τM , s ∈ N+ (4)
employed to coordinate the considerable data transmission between
the MNN and the remote estimator, and this constitutes one of our where τm and τM are known constants satisfying 0 ≤ τm ≤ τM . C,
motivations in the present investigation. L, M, and N are known matrices. The initial
T condition of (1) has the
In this article, we aim at developing a resilient H∞ estimator form φ(s) = φ1 (s) φ2 (s) · · · φn (s) for s ∈ [−τM , 0].
for delayed MNNs with SDs under the dynamic ETM. The primary Assumption 1: The NAF f (z(s)) is continuous that satisfies
contributions we deliver in this article are outlined in three-fold. T
1) The ETM is introduced to govern the data transmission in the f (x) − f (y) − 1 (x − y) f (x) − f (y) − 2 (x − y)
sensor-to-estimator communication, thereby relieving the trans- ≤ 0 ∀x, y ∈ Rn (x = y) (5)
mission frequency of the data and saving the consumption of
resources. where 1 and 2 are constant matrices and f (0) = 0.
2) A unified H∞ estimation scheme is built to cope with the math- By using the similar techniques employed in [22], the MNNs given
ematical complexities resulting from the time delays, SDs, and by (1) can be rewritten as the following form:
⎧
ETM. ⎪
⎪ z(s + 1) = D̄ + D(s) z(s) + Ā + A(s) f (z(s))
⎨
3) A delay-dependent measure is established to ensure the expo- + B̄ + B(s) f (z(s − τ (s))) + Lv(s)
(6)
nential mean-square stability (EMSS) of the resulting estima- ⎪
⎪ y(s) = Cz(s) + M(s)v(s)
⎩
tion error system with guaranteed H∞ performance constraint. z̄(s) = N(s)z(s)
Notation: Rn and Rn×m define, respectively, the n-dimensional
Euclidean space and the set of all n × m real matrices. N+ denotes where
the set of all non-negative integers. I is the identity matrix. λmin (A)
D̄ diag{d1 , d2 , . . . , dn }, Ā aij n×n , B̄ bij n×n
(λmax (A)) is the smallest (largest) eigenvalue of matrix A. diag{· · · }
means a block-diagonal matrix. E{·} stands for the expectation D(s) HF1 (s)E1 , A(s) HF2 (s)E2
operator. x describes the Euclidean norm of x. B(s) HF3 (s)E3 .
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CYBERNETICS 3335
Here, dj , aij , and bij are known positive scalars, H = H1 H2 · · · Hn Furthermore, we now have the augmented system
T T
T T (i = 1, 2, 3) are known matrices, and
and Ei = Ei1 Ei2 · · · Ein ⎧
the time-varying matrix Fi (s) (i = 1, 2, 3) satisfies ⎪
⎪ η(s + 1) = D(s)η(s) + A(s)f (η(s))
⎨
+ B(s)f (η(s − τ (s)))
FiT (s)Fi (s) ≤ I. (13)
(7) ⎪
⎪ + K(s)ζ (s) + L(s)v(s) + σ (s)w(s)
⎩ ˜ = N (s)η(s)
Remark 1: Compared to the traditional RNNs, a remarkable fea- z̄(s)
ture of the MNNs is the dependence of their memristive CWs on
where
the NSs, which makes it mathematically difficult to analyze the
dynamics of the MNNs. Thanks to the state-dependent switching T T
η(s) zT (s) eT (s) , ζ (s) 0 T (s)
(toward norm-bounded uncertainties) that relies on the characteris- T
tics of the memristors and the current–voltage, we are able to convert f (η(s)) f T (z(s)) f̃ T (s) , D(s) D̄(s) + D(s)
the MNN (1) with state-dependent uncertainties into (6) with norm- D(s) D1 (s) + D2 (s), A(s) Ā + A(s)
bounded uncertainties displayed in the terms D(s), A(s), and
Ā diag Ā, Ā , N (s) 0 N(s)
B(s).
In this article, in order to reduce resource consumption in the B(s) B̄ + B(s)
sensor-to-estimator channel, a dynamic ETM is taken to decide if B̄ diag B̄, B̄ , K(s) K̄(s) + K(s)
the current data should be sent or not (see Fig. 1). Denote the trig- D̄ 0
gering instant sequence as 0 ≤ s0 < s1 < · · · < s < · · · , where L(s) L̄(s) + L(s), D̄(s)
0 D̄ − K(s)C
s +1 is determined by
D(s) 0 0 0
ς (s) D1 (s) , D2 (s)
D(s) 0 0 K(s)C
s +1 = min s ∈ [0, N]|s > s , + yT (s)y(s)
ϑ A(s) 0 B(s) 0
A(s) , B(s)
− T (s)(s) ≤ 0 (8) A(s) 0 B(s) 0
0 0 0 0
K̄(s) , K(s)
where ϑ and are known positive scalars, (s) y(s)−y(s ), y(s ) is 0 K(s) 0 K(s)
the latest transmitted measurement, and the dynamical variable ς (s) L(s) 0
L̄(s) , L(s)
satisfies L(s) − K(s)M − K(s)M
ς (s + 1) = ς (s) + yT (s)y(s) − T (s)(s) σ (s, z(s), z(s − τ (s)))
ς (0) = ς0
(9) σ (s)
σ (s, z(s), z(s − τ (s)))
where 0 < < 1 is a known positive scalar. T
Remark 2: From the TC (8), it is straightforward to see that (8) and f (η(s − τ (s))) f T (z(s − τ (s))) f̃ T (s − τ (s)) .
will reduce to a traditional static ETM when ϑ approaches +∞. That Our main purpose is to design the estimator (10) such that:
is to say, the dynamic ETM (8) includes its static counterpart as a 1) system (13) with v(s) = 0 is exponentially stable in the mean
special case. square;
Based on the dynamic ETM, the desired estimator is built as 2) for nonzero v(s)
⎧ ⎧ ⎫
⎨ ẑ(s + 1) = D̄ẑ(s) + Āf ẑ(s) + B̄f ẑ(s − τ (s)) ⎨+∞ ⎬ +∞
+ (K(s) + K(s)) y(s ) − Cẑ(s) ˜ 2
(10) E z̄(s) ≤ γ2 v(s)2 (14)
⎩ ˆ = Nẑ(s) ⎩ ⎭
z̄(s) s=0 k=0
ˆ
where ẑ(s) and z̄(s) are the estimates of z(s) and z̄(s), respectively, where γ is a prescribed disturbance attenuation level.
K(s) is the parameter to be designed, and the gain perturbation K(s)
satisfies
III. M AIN R ESULT
K(s) HK FK (s)EK (11)
Lemma 1 [17]: Given the dynamic ETM (8), (9) with ς0 ≥ 0, if
where HK and EK are known matrices, and FK (s) satisfies parameters (0 < < 1) and ϑ (ϑ > 0) satisfy ϑ ≥ 1, then ς (s)
T (s)F (s) ≤ I.
FK K satisfies ς (s) ≥ 0 for all s ∈ [0, N].
Let Remark 3: Lemma 1 tells that the dynamic variables ς (s) are non-
e(s) z(s) − ẑ(s) negative for s ∈ [0, N]. Unlike the static case, the negative values
˜ z̄(s) − z̄(s)
ˆ of yT (s)y(s) − T (s)(s) are now acceptable. As a result, the trig-
z̄(s)
gering frequency is greatly reduced, thereby alleviating the resource
f̃ (s) f (z(s)) − f ẑ(s) consumption in the sensor-to-estimator substantially.
f̃ (s − τ (s)) f (z(s − τ (s))) − f ẑ(s − τ (s)) . Theorem 1: Let the estimator gain K and the attenuation level
γ > 0 be given. Then, the EMSS of system (13) with v(s) = 0
Following from (6) and (10), we obtain the following dynamics of
is guaranteed if there exist a positive-definite matrix P and scalars
the estimation error:
⎧ λ1 , λ2 , λ3 , and λ∗ such that
⎪ e(s + 1) = D̄ − K(s)C e(s) − K(s)Ce(s)
⎪
⎪ ⎡ ⎤
⎪
⎪ + D(s)z(s) + Āf̃ (s) + A(s)f (z(s)) ¯ 11 0 ¯ 13 14 15 0
⎪
⎪
⎪
⎨ + B̄f̃ (s − τ (s)) + B(s)f (z(s − τ (s))) ⎢ ∗ ¯ 22 0 ¯ 24 0 0 ⎥
⎢ ⎥
+ (L(s) − K(s)M)v(s) + K(s)(s) (12) ⎢ ∗ ∗ ¯
33 34 35 0 ⎥
⎪ ¯ =⎢
⎥<0 (15)
⎪
⎪ − K(s)Mv(s) + K(s)(s) ⎢ ∗ ∗ ∗ ¯ 44 45 0 ⎥
⎪
⎪ ⎢ ⎥
⎪
⎪ + σ (s, z(s), z(s − τ (s)))w(s) ⎣ ∗ ∗ ∗ ∗ ¯ 55
0 ⎦
⎪
⎩
˜ = Ne(s).
z̄(s) ∗ ∗ ∗ ∗ ∗ ¯ 66
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
3336 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022
Noting (3), we obtain Along the similar line of the proof in [22], the EMSS of
system (13) with v(k) = 0 can be assured from (15).
σ T (s)Pσ (s) ≤ λmax (P)σ T (s)σ (s)
Theorem 2: Let the estimator gain K and the attenuation level
≤ λ∗ ρ1 ηT (s)I2T I2 η(s) γ > 0 be given. Then, system (13) is EMSS and satisfies the H∞
+ λ∗ ρ2 ηT (s − τ (k))I2T I2 η(k − τ (s)). (19) constraint (14) if there exist a positive-definite matrix P and scalars
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CYBERNETICS 3337
× f (η(s − τ (s)))
T
T
˜
E z̄˜T (s)z̄(s) − E γ 2 vT (s)v(s)
+ ζ T (s)KT PKζ (s) + vT (s)LT PLv(s) s=0 s=0
+ σ (s)T Pσ (s)
T
≤ ˜ ξ̃ (s) − E{V(T + 1)}
E ξ̃ T (s)
+ 2ηT (s)DT (s)PA(s)f (η(s))
s=0
+ 2ηT (s)DT (s)PB (s)
T
× f (η(s − τ (s))) + 2ηT (s)DT (s)PKζ (s) ≤ ˜ ξ̃ (s) .
E ξ̃ T (s) (29)
s=0
+ 2η (s)D (s)PLv(s)
T T
Noticing (24) and making T → ∞ result in the satisfaction of the
+ 2f T (η(s))AT (s)PB (s)
H∞ constraint (14).
× f (η(s − τ (s))) + 2f T (η(s))AT (s)PKζ (s) In terms of Theorem 2, our desired estimator is given as follows.
+ 2f T (η(s))AT (s)PLv(s) Theorem 3: The resilient H∞ SEP for MNN (1) is solvable if there
exist matrix P diag{P1 , P2 } and scalars λ1 , λ2 , λ3 , λ∗ , and κ such
+ 2f T (η(s − τ (s))) × B T (s)PKζ (s)
that
+ 2f T (η(s − τ (s)))B T (s)PLv(s) ⎡ ⎤
ˆ H̆ κ Ĕ T
−1 T ˘ = ⎣ ∗ −κI
− ηT (s)Pη(s) + ς (s)ς (s) 0 ⎦<0 (30)
ϑ ∗ ∗ −κI
+ ηT (s)C˜T C˜η(s)
ϑ where
⎡ ⎤
+ vT (s)M T Mv(s) + 2 ηT (s)C˜T Mv(s) −P 0 ¯1
ϑ ϑ ˆ ⎣ ∗
1 T −I 2 ⎦
− ζ (s)I1 I1 ζ (s) .
T
(25) ∗ ∗ 3
ϑ
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
3338 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022
with
˘ 11 λ∗ ρ1 I T I2 + λ3 C˜T C˜ − λ1
˜ −P
2
˘ ¯ ˘ ˜T
13 λ1 , 17 λ3 C M, 33 −λ1 I, ˘ ˘ 44 −λ2 I
1 + λ3 T
˘ 55 −
I1 I1 , ˘ 77 λ3 MT M − γ 2 I
ϑ
¯ 1 PD̄ 0 PĀ PB̄ PK̄ 0 PL̄
N 0 0 0 0 0 0
2 √ √
/ϑ C˜ 0 0 0 0 0 /ϑM
⎡˘ ˘ 13 ˘ 17 ⎤
11 0 0 0 0
⎢ ∗ ¯
22 0 ¯
24 0 0 0 ⎥
⎢ ⎥
⎢ ∗ ∗ ˘ 33 0 0 0 0 ⎥
⎢ ⎥
3 ⎢ ⎢ ∗ ∗ ∗ ˘ 44 0 0 0 ⎥ ⎥
⎢ ∗ ∗ ∗ ∗ ˘ 55
0 ⎥
⎢ 0 ⎥
⎣ ∗ ∗ ∗ ∗ ∗ ¯ 66 0 ⎦ Fig. 2. State z1 (s) and its estimate ẑ1 (s).
∗ ∗ ∗ ∗ ∗ ∗ ˘ 77
⎡ ⎤T
H̃T PT 0 0 0 0 0 0 0 0
⎢ H̃D T ⎥ Then, it implies by K = P−1 Y and the S-procedure that inequal-
⎢ A PT 0 0 0 0 0 0 0 0⎥
⎢ T T ⎥ ity (32) holds if inequality (30) holds.
H̆ ⎢ H̃B P 0 0 0 0 0 0 0 0⎥
⎢ T T ⎥ Remark 4: In Theorem 3, the design of the discussed resilient H∞
⎣H̃K P 0 0 0 0 0 0 0 0⎦
T T state estimator is successfully converted into a convex optimization
H̃ P 0 0 0 0 0 0 0 0
⎡ L ⎤ problem via the semidefinite programming method. The computa-
0 0 ẼD 0 0 0 0 0 0 tional complexity of the proposed LMI-based algorithm depends
⎢0 0 0 ẼA 0⎥
⎢ 0 0 0 0 ⎥ polynomially on the dimensions of the system state and the measured
⎢ ⎥
Ĕ ⎢0 0 0 0 0 ẼB 0 0 0⎥ output vector, and the number of the IF - THEN rules.
⎢ ⎥
⎣0 0 0 0 0 0 ẼK 0 0⎦ Remark 5: Regarding the concerned resilient H∞ SEP for delayed
0 0 0 0 0 0 0 0 ẼL stochastic MNNs, there are four primary factors that increase the
design complexity of our estimator, that is, the dynamic ETM, param-
H 0 0 H 0
H̃D , H̃A H̃B = eter drifts/variations, time delays, and additive deterministic noises.
H 0 −HK H 0
⎡ ⎤ All these factors are well reflected in the derivation of the sufficient
E1 0
0 0 conditions that assure the existence of our estimator. Besides, the
H̃K H̃L = , ẼD ⎣ 0 0 ⎦
0 −HK solvability condition of the estimator gain is formulated by means
0 −EK C of the feasibility of certain linear matrix inequalities whose solutions
E 0 E 0 can be acquired by standard software packages.
ẼA 2 , ẼB 3
0 E2 0 E3 Remark 6: Among the rich body of existing literature, our main
EK 0 0 results make themselves unique from the following two aspects: 1)
ẼK , ẼL . the dynamic ETM is, for the first time, introduced to schedule the
0 EK EK M
data traffic in the sensor-to-estimator channel for the sake of mitigat-
Furthermore, if (30) is feasible, then K = P2 X. ing resource consumption and 2) a new yet unified H∞ estimation
Proof: Denote scheme is established to account for the SDs, dynamic ETM, time
1 PD(s) 0 PA(s) PB(s) PK 0 PL . delays, and parameter drifts/variations.
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CYBERNETICS 3339
Fig. 3. State z2 (s) and its estimate ẑ2 (s). Fig. 5. Estimation error of z̄(s).
Fig. 4. ˆ
State z̄(s) and its estimate z̄(s). Fig. 6. Event-based release instants and release interval: dynastic case.
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
3340 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 52, NO. 5, MAY 2022
More specifically, in order to verify the superiority of the dynamic [15] J. Li, Z. Wang, H. Dong, and F. Han, “Delay-distribution-dependent
ETM than the static one on reducing resource consumption, the cor- state estimation for neural networks under stochastic communication
protocol with uncertain transition probabilities,” Neural Netw., vol. 130,
responding release instants and intervals of the static one are shown
pp. 143–151, Oct. 2020.
in Fig. 7. We can come to a conclusion exactly that the dynamic [16] Q. Li, B. Shen, Z. Wang, and W. Sheng, “Recursive distributed filtering
ETM can reduce the resource consumption more effectively than the over sensor networks on Gilbert–Elliott channels: A dynamic event-
traditional static one. triggered approach,” Automatica, vol. 113, Mar. 2020, Art. no. 108681.
[17] Q. Li, B. Shen, Z. Wang, T. Huang, and J. Luo, “Synchronization con-
trol for a class of discrete time-delay complex dynamical networks: A
V. C ONCLUSION dynamic event-triggered approach,” IEEE Trans. Cybern., vol. 49, no. 5,
pp. 1979–1986, May 2019.
In this article, we have coped with the dynamic event-triggered [18] Q. Li, Z. Wang, N. Li, and W. Sheng, “A dynamic event-
resilient SEP for MNNs with time delays and SDs. Combining triggered approach to recursive filtering for complex networks
the Lyapunov–Krasovskii functional and stochastic analysis tech- with switching topologies subject to random sensor failures,” IEEE
nique, a dynamic event-triggered criterion has been presented to Trans. Neural Netw. Learn. Syst., early access, Dec. 11, 2019,
doi: 10.1109/TNNLS.2019.2951948.
guarantee both the EMSS of the error dynamics and the prescribed [19] Y. Liu, Z. Wang, and X. Liu, “Global exponential stability of generalized
H∞ constraint. Furthermore, the desired estimator insensitive to recurrent neural networks with discrete and distributed delays,” Neural
parameter variations/fluctuations has been devised via solving an Netw., vol. 19, pp. 667–675, Jun. 2006.
optimization problem. In the end, the validity of our estimator has [20] Y. Li, G. Wang, L. Nie, Q. Wang, and W. Tan, “Distance metric
optimization driven convolutional neural network for age invariant face
been verified via a quintessential example. Note that the developed
recognition,” Pattern Recognit., vol. 75, pp. 51–62, Mar. 2018.
event-based resilient strategy might be applied to systems with [21] H. Liu, Z. Wang, B. Shen, and H. Dong, “Delay-distribution-dependent
higher complexities, such as genetic regulatory networks and sensor H∞ state estimation for discrete-time memristive neural networks with
networks. mixed time-delays and fading measurements,” IEEE Trans. Cybern.,
vol. 50, no. 2, pp. 440–451, Sep. 2020.
[22] H. Liu, Z. Wang, B. Shen, and X. Liu, “Event-triggered H∞ state esti-
R EFERENCES mation for delayed stochastic memristive neural networks with missing
measurements: The discrete time case,” IEEE Trans. Neural Netw. Learn.
[1] S. Arik, “A modified Lyapunov functional with application to stability of Syst., vol. 29, no. 8, pp. 3726–3737, Aug. 2018.
neutral-type neural networks with time delays,” J. Frankl. Inst., vol. 356, [23] Y. Liu, Z. Wang, L. Ma, and F. E. Alsaadi, “A partial-nodes-
no. 1, pp. 276–291, 2019. based information fusion approach to state estimation for discrete-time
[2] H. Bao, J. Cao, J. Kurths, A. Alsaedi, and B. Ahmad, “H∞ state esti- delayed stochastic complex networks,” Inf. Fusion, vol. 49, pp. 240–248,
mation of stochastic memristor-based neural networks with time-varying Sep. 2019.
delays,” Neural Netw., vol. 99, pp. 79–91, Mar. 2018. [24] Y. Liu, B. Shen, and Q. Li, “State estimation for neural networks
[3] L. Chen, C. Li, T. Huang, Y. Chen, and X. Wang, “Memristor crossbar- with Markov-based nonuniform sampling: The partly unknown transition
based unsupervised image learning,” Neural Comput. Appl., vol. 25, probability case,” Neurocomputing, vol. 357, pp. 261–270, Sep. 2019.
no. 2, pp. 393–400, 2014. [25] Y. Liu, Z. Wang, Y. Yuan, and F. E. Alsaadi, “Partial-nodes-based state
[4] W. Chen, D. Ding, J. Mao, H. Liu, and N. Hou, “Dynamical performance estimation for complex networks with unbounded distributed delays,”
analysis of communication-embedded neural networks: A survey,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp. 3906–3912,
Neurocomputing, vol. 346, no. 21, pp. 3–11, 2019. Aug. 2018.
[5] Y. Chen, Z. Wang, J. Hu, and Q.-L. Han, “Synchronization control [26] F. Moretti, S. Pizzuti, S. Panzieri, and M. Annunziato, “Urban traffic
for discrete-time delayed dynamical networks with switching topology flow forecasting through statistical and neural network bagging ensemble
under actuator saturations,” IEEE Trans. Neural Netw. Learn. Syst., early hybrid modeling,” Neurocomputing, vol. 167, pp. 3–7, Nov. 2015.
access, Jun. 10, 2020, doi: 10.1109/TNNLS.2020.2996094. [27] Y. V. Pershin and M. Di Ventra, “Experimental demonstration of asso-
[6] J. Ding, B. Chen, H. Liu, and M. Huang, “Convolutional neural network ciative memory with memristive neural networks,” Neural Netw., vol. 23,
with data augmentation for SAR target recognition,” IEEE Geosci. no. 7, pp. 881–886, 2010.
Remote Sens. Lett., vol. 13, no. 3, pp. 364–368, Mar. 2016. [28] R. Sakthivel, R. Anbuvithya, K. Mathiyalagan, and P. Prakash,
[7] S. Ding, Z. Wang, and H. Zhang, “Quasi-synchronization of delayed “Combined H∞ and passivity state estimation of memristive neural
memristive neural networks via region-partitioning-dependent intermit- networks with random gain fluctuations,” Neurocomputing, vol. 168,
tent control,” IEEE Trans. Cybern., vol. 49, no. 12, pp. 4066–4077, pp. 1111–1120, Nov. 2015.
Dec. 2019. [29] B. Shen, Z. Wang, D. Wang, and Q. Li, “State-saturated recursive filter
[8] A. Girard, “Dynamic triggering mechanisms for event-triggered control,” design for stochastic time-varying nonlinear complex networks under
IEEE Trans. Autom. Control, vol. 60, no. 7, pp. 1992–1997, Jul. 2015. deception attacks,” IEEE Trans. Neural Netw. Learn. Syst., early access,
[9] F. G. Harmon, A. A. Frank, and S. S. Joshi, “The control of a parallel Nov. 14, 2016, doi: 10.1109/TNNLS.2019.2946290.
hybrid-electric propulsion system for a small unmanned aerial vehi- [30] Y. Shen, Z. Wang, B. Shen, F. E. Alsaadi, and A. M. Dobaie, “ 2 −
cle using a CMAC neural network,” Neural Netw., vol. 18, nos. 5–6, ∞ state estimation for delayed artificial neural networks under high-
pp. 772–780, 2005. rate communication channels with round-Robin protocol,” Neural Netw.,
[10] N. Hou, H. Dong, Z. Wang, W. Ren, and F. E. Alsaadi, “Non- vol. 124, pp. 170–179, Apr. 2020.
fragile state estimation for discrete Markovian jumping neural networks,” [31] Y. Shen, Z. Wang, B. Shen, and F. E. Alsaadi, “H∞ state estimation
Neurocomputing, vol. 179, pp. 238–245, Feb. 2016. for multi-rate artificial neural networks with integral measurements: A
[11] N. Hou, Z. Wang, D. W. C. Ho, and H. Dong, “Robust partial-nodes- switched system approach,” Inf. Sci., vol. 539, pp. 434–446, Oct. 2020.
based state estimation for complex networks under deception attacks,” [32] Y. Shen, Z. Wang, B. Shen, and F. E. Alsaadi, “Nonfragile H∞ filtering
IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 6, pp. 2793–2802, for discrete multirate time-delayed systems over sensor networks char-
Jun. 2020. acterized by Gilbert–Elliott models,” Int. J. Robust Nonlinear Control,
[12] L. Jin, S. Li, H. La, and X. Luo, “Manipulability optimization of redun- vol. 30, no. 8, pp. 3194–3214, May 2020.
dant manipulators using dynamic neural networks,” IEEE Trans. Ind. [33] B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural network
Electron., vol. 64, no. 6, pp. 4710–4720, Feb. 2017. for image-based sequence recognition and its application to scene text
[13] F. Li, H. Shen, M. Chen, and Q. Kong, “Non-fragile finite-time recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 11,
l2 − l∞ state estimation for discrete-time Markov jump neural networks pp. 2298–2304, Jul. 2016.
with unreliable communication links,” Appl. Math. Comput., vol. 271, [34] D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, “The
pp. 467–481, Nov. 2015. missing memristor found,” Nature, vol. 453, no. 7191, pp. 80–83, 2008.
[14] J. Li, Z. Wang, H. Dong, and G. Ghinea, “Outlier-resistant remote [35] E. Tian, Z. Wang, L. Zou, and D. Yue, “Probability-constrained filtering
state estimation for recurrent neural networks with mixed time-delays,” for a class of nonlinear systems with improved static event-triggered
IEEE Trans. Neural Netw. Learn. Syst., early access, May 18, 2020, communication,” Int. J. Robust Nonlinear Control, vol. 29, no. 5,
doi: 10.1109/TNNLS.2020.2991151. pp. 1484–1498, 2019.
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON CYBERNETICS 3341
[36] L. Tian, Y. Cheng, C. Yin, D. Ding, Y. Song, and L. Bai, “Design of the [45] F. Zeng and L. Sheng, “State estimation for neural networks with random
MOI method based on the artificial neural network for crack detection,” delays and stochastic communication protocol,” Syst. Sci. Control Eng.,
Neurocomputing, vol. 226, pp. 80–89, Feb. 2017. vol. 6, no. 3, pp. 54–63, 2018.
[37] S. Tong and Y. Li, “Robust adaptive fuzzy backstepping output feedback [46] H. B. Zeng, Y. He, M. Wu, and H. Q. Xiao, “Improved conditions for
tracking control for nonlinear system with dynamic uncertainties,” Sci. passivity of neural networks with a time-varying delay,” IEEE Trans.
China Inf. Sci., vol. 53, no. 2, pp. 307–324, 2010. Cybern., vol. 44, no. 1, pp. 785–792, Jan. 2014.
[38] S. Tong and Y. Li, “Observer-based adaptive fuzzy backstepping con- [47] C.-K. Zhang, Y. He, L. Jiang, and M. Wu, “Stability analysis for delayed
trol of uncertain nonlinear pure-feedback systems,” Sci. China Inf. Sci., neural networks considering both conservativeness and complexity,”
vol. 57, pp. 1–14, Mar. 2014. IEEE Trans. Neural Netw. Learn. Syst., vol. 27, no. 7, pp. 1486–1501,
[39] A. T. Vemuri, M. M. Polycarpou, and S. A. Diakourtis, “Neural network Jul. 2015.
based fault detection in robotic manipulators, IEEE Trans. Robot. [48] G. Zhang and Y. Shen, “Exponential synchronization of delayed
Autom., vol. 14, no. 2, pp. 34–348, Apr. 1998. memristor-based chaotic neural networks via periodically intermittent
[40] Z. Wang, Y. Liu, and X. Liu, “On global asymptotic stability of neural control,” Neural Netw., vol. 55, pp. 1–10, Jun. 2014.
networks with discrete and distributed delays,” Phys. Lett. A, vol. 345, [49] X.-M. Zhang, Q.-L. Han, Z. Wang, and B.-L. Zhang, “Neuronal state
nos. 4–6, pp. 299–308, 2005. estimation for neural networks with two additive time-varying delay
[41] S. Wen, T. Huang, Z. Zeng, Y. Chen, and P. Li, “Circuit design and components,” IEEE Trans. Cybern., vol. 47, no. 10, pp. 3184–3194,
exponential stabilization of memristive neural networks,” Neural Netw., Oct. 2017.
vol. 63, pp. 48–56, Mar. 2015. [50] X.-M. Zhang, Q.-L. Han, X. Ge, and D. Ding, “An overview of
[42] S. Wen, Z. Zeng, T. Huang, and Y. Zhang, “Exponential adaptive lag recent developments in Lyapunov–Krasovskii functionals and stabil-
synchronization of memristive neural networks via fuzzy method and ity criteria for recurrent neural networks with time-varying delays,”
applications in pseudorandom number generators,” IEEE Trans. Fuzzy Neurocomputing, vol. 313, pp. 392–401, Nov. 2018.
Syst., vol. 22, no. 6, pp. 1704–1713, Dec. 2014. [51] D. Zhao, Z. Wang, Y. Chen, and G. Wei, “Proportional-integral observer
[43] A. Wu and Z. Zeng, “Exponential stabilization of memristive neural design for multi-delayed sensor-saturated recurrent neural networks: A
networks with time delays,” IEEE Trans. Neural Netw. Learn. Syst., dynamic event-triggered protocol,” IEEE Trans. Cybern., early access,
vol. 23, no. 12, pp. 1919–1929, Oct. 2012. Feb. 17, 2020, doi: 10.1109/TCYB.2020.2969377.
[44] Y. Xu, C. Liu, J.-Y. Li, C.-Y. Su, and T. Huang, “Finite-horizon H∞ state [52] Z. Zhao, Z. Wang, L. Zou, and G. Guo, “Finite-time state estima-
estimation for time-varying neural networks with periodic inner coupling tion for delayed neural networks with redundant delayed channels,”
and measurements scheduling,” IEEE Trans. Syst., Man, Cybern., Syst., IEEE Trans. Syst., Man, Cybern., Syst., early access, Oct. 29, 2019,
vol. 50, no. 1, pp. 211–219, Jan. 2020. doi: 10.1109/TSMC.2018.2874508.
Authorized licensed use limited to: Anhui University of Technology. Downloaded on March 19,2025 at 02:00:32 UTC from IEEE Xplore. Restrictions apply.