Continuous and discrete zeroing neural network for a class of multilayer dynamic system
Introduction
Neural networks (NNs) have been regarded as the key to solving various difficult problems in science and engineering fields with the heuristics of the biological nervous system [1], [2], [3] and references therein. In the past decade NNs have been investigated comprehensively and developed rapidly in a wide range of areas like image processing [4] and industrial automation [5]. Zeroing neural network (ZNN), as a class of NNs, can be applied for solving time-varying problems. In 1993, [6] took the initiative in putting forward a recurrent neural network(RNN). Zhang et al. proposed a novel class of RNN, termed zeroing neural network (ZNN) (or Zhang neural network after its inventors) to analyze and solve time-varying problems [7]. Researchers proposed various kinds of ZNN models for solving problems with different highlights such as matrix equation solving [8], [9], [10], time-varying matrix pseudoinversion [11], [12], robot manipulator [13], [14], [15], root-finding [16] and so on.
Especially, ZNN method can be applied to solve time-varying multilayer dynamic system. Multilayer dynamic system has more than two layers different from common system with signal layer. It plays an important role in industrial engineering. A typical application is manipulator operation: the manipulator tracking control with additional constraints can be transformed into solving multilayer dynamic system problems. Researchers have made progress along this direction and proposed various continuous and discrete ZNN models [13], [17], [18]. Different-Level Dynamic Linear System(DLDLS) was first formulated in [13]. Hybrid Multilayered Time-Varying Linear System(HMTVLS) was studied in [17]. And Future Different-Layer Nonlinear and Linear Equation System(DLNLES) was investigated in [18]. Continuous and discrete ZNN models were proposed in [13], [17], [18] to solve the DLDLS, HMTVLS and DLNLES respectively.
Note that, these models were all adopted the linear ZNN design method in the above mentioned papers [13], [17], [18]. Actually, at the beginning of the study, conventional ZNN models usually adopt the original linear ZNN design method. And with further research, researchers have tried to find new ZNN design method to achieve better convergence speed of ZNN models. Thereupon, activation functions have been applied in ZNN models [19], [20], [21], [22], [23]. In previous studies, researchers presented various ZNN models with different activation functions for different time-varying problems. [19] selected power-sum activation function to construct ZNN models for time-varying QP problem solving. [20] used hyperbolic sine activation function in the ZNN model for time-varying matrix square roots finding. [21] presented a new activation function with two tunable parameters. [22] adopted Sign-Bi-power (SBP) activation function for time-varying Sylvester equation. It has been shown that the selection of nonlinear functions can speed up the convergence of ZNN models [23]. Moreover, some nonlinear activation functions will bring finite convergence. Both theoretical analysis and simulation result in the above literature were provided to verify the outstanding performance of these nonlinear activation functions [19], [20], [21], [22], [23].
Besides, conventional ZNN models usually assume free of noise in the solving process [24]. However, noise actually exists in real time. [25] firstly proposed an integration-enhanced Zhang neural network(IEZNN) based on the novel integration-enhanced design formula and proved its noise resistance. Since the first empirical research, the noise-tolerant design formula has been widely used to construct ZNN models for solving time-varying problems with noise-suppression [26], [27], [28], [29], [30], [31]. In this paper, we both adopt linear and nonlinear ZNN design method to propose ZNN models for solving a class of time-varying continuous and discrete multilayer dynamic system(MLDS). Further, we take the noise pollution into account and propose noise-tolerant ZNN model.
The rest of this paper is organized as follows. Section 2 presents the continuous and discrete problem formulation of MLDS respectively. In Section 3, three kinds of ZNN models for continuous MLDS are discussed through theoretical analysis. In Section 4, ZNN models for discrete MLDS with m-step ZeaD formula are proposed. Section 5 is the numerical simulation and Section 6 is the conclusion of this paper. The major contributions to this paper are as follows.
- 1.
Different from common systems, MLDS with two levels is formulated and investigated.
- 2.
Linear, nonlinear and noise-tolerant ZNN models for continuous MLDS are presented and the convergence of solutions is analyzed.
- 3.
The existence of m-step ZeaD formula is proved. Linear, nonlinear and noise-tolerant ZNN models for discrete MLDS with m-step ZeaD formula are proposed and related algorithms are obtained for easier hardware realization.
Section snippets
Problem Formulation
Notation: For a matrix denotes its transpose matrix. denotes the n-dimensional Euclidean space. For a vector represents its 2-norm. denotes time duration . denotes a set of integers . and denote the sampling gap and the final time respectively. defines the time derivative of function f. Symbol denotes the factorial operator. denote the second-order, and the m-th-order derivatives of with respect to t at time
Continuous Zeroing Neural Network for MLDS
In this section, continuous ZNN models for continuous MLDS (1)–(2) are proposed. In subSection 3.1, we assume that the model solving process is free of noise. We propose linear ZNN model (MLDS-linear-ZNN) and nonlinear ZNN model (MLDS-nonlinear-ZNN) by selecting different activation functions. In subSection 3.2, we take noise pollution into account and propose noise-tolerant ZNN model(MLDS-noise-tolerant-ZNN).
Discrete Zeroing Neural Network for MLDS
In this section, ZNN models for discrete MLDS (3) and (4) are proposed based on ZeaD formula, a new Zhang et al. discretization formula [7]. There are some ZeaD formulas deduced in papers: Euler formula, Taylor formula, four order formula and so on. In fact, by Taylor expansion method we can prove that can always be expressed like Theorem 5. Theorem 5 For any positive integer m, if is ()-order differentiable there always exists constants such that can be obtained by
Numerical Simulation
In this section, matlab programming is used for numerical simulation. Let and , we take the following MLDS [17] as an example for numerical simulation:where , , , and satisfy:
Conclusion
In this paper, zeroing neural network(ZNN) models for a class of multilayer dynamic system(MLDS) have been formulated and investigated. In the case of continuous time, we have obtained ZNN models for MLDS(MLDS-linear-ZNN and MLDS-nonlinear-ZNN) with different activation functions. MLDS-noise-tolerant-ZNN has been further proposed when solving process influenced by noise. Meanwhile, for hardware realization we considered the discrete time case. We have proved the existence of m-step ZeaD formula
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
The authors express their gratitude to the editor and the reviewers for constructive comments and helpful suggestions. This work is supported by the National Natural Science Foundation of China under Grant 61673296.
Yuting Xue was born in Anhui, China, in 1998. She received her Bachelor’s degree in School of Mathematical Sciences, Tongji University, Shanghai, China, in 2021. She is currently pursuing her Ph.D. degree with School of Mathematical Sciences, Tongji University, Shanghai, China. Her current research interests include game and hybrid systems.
References (31)
A recurrent neural network for real-time matrix inversion
Appl. Math. Comput.
(1993)- et al.
Zeroing neural network methods for solving the Yang-Baxter-like matrix equation
Neurocomputing
(2020) - et al.
Zhang neural network versus gradient-based neural network for time-varying linear matrix equation solving
Neurocomputing
(2011) - et al.
From different ZFs to different ZNN models accelerated via Li activation functions to finite-time convergence for time-varying matrix pseudoinversion
Neurocomputing
(2014) - et al.
Robust model-free control for redundant robotic manipulators based on zeroing neural networks activated by nonlinear functions
Neurocomputing
(2021) - et al.
Finite time dual neural networks with a tunable activation function for solving quadratic programming problems and its application
Neurocomputing
(2014) - et al.
Integration enhanced and noise tolerant ZNN for computing various expressions involving outer inverses
Neurocomputing
(2019) - K. Shi, J. Wang, S.Zhong, et al., Non-fragile memory filtering of T-S fuzzy delayed neural networks based on switched...
- K. Shi, J. Wang, Y.Tang, et al., Reliable asynchronous sampled-data filtering of T-S fuzzy uncertain delayed neural...
- S. Dong, H.Zhu, S.Zhong, et al., Hybrid control strategy of delayed neural networks and its application to sampled-data...
Zhang neural networks and neural-dynamic method
A noise-suppression ZNN model with new variable parameter for dynamic Sylvester equation
IEEE Trans. Ind. Informat.
Inverse-free discrete ZNN models solving for future matrix pseudoinverse via combination of extrapolation and ZeaD formulas
IEEE Trans. Neural Netw. Learn. Syst.
Cited by (1)
A zeroing neurodynamics-based location method for nodes in underwater acoustic sensor network
2023, CAAI Transactions on Intelligence Technology
Yuting Xue was born in Anhui, China, in 1998. She received her Bachelor’s degree in School of Mathematical Sciences, Tongji University, Shanghai, China, in 2021. She is currently pursuing her Ph.D. degree with School of Mathematical Sciences, Tongji University, Shanghai, China. Her current research interests include game and hybrid systems.
Jitao Sun received the B.Sc. degree in Mathematics from Nanjing University, Nanjing, China, in 1983, and the Ph.D. degree in control theory and control engineering from the South China University of Technology, Guangzhou, China, in 2002.
He was with the Anhui University of Technology, Ma’anshan, China, from 1983 to 1997. From 1997 to 2000, he was with Shanghai Tiedao University. In 2000, he joined the School of Mathematical Sciences, Tongji University, Shanghai, China. He was a Professor with the Anhui University of Technology and Shanghai Tiedao University from 1995 to 2000. He is currently a Professor with Tongji University. He has authored or coauthored more than 210 journals papers. His recent research interests include impulsive control, time delay systems, hybrid systems, and systems biology.
Dr. Sun is a Member of Technical Committee on Nonlinear Circuits and Systems, IEEE Circuits and Systems Society, and a Reviewer of Mathematical Reviews of American Mathematical Society.
Ying Qian is an undergraduate student of the Class of 2023 studying electrical engineering and applied math at the School of Engineering, Vanderbilt University.