Article Preview
Top1. Introduction
In recent years, the concept of Internet of things put forward on the basis of the Internet has been widely concerned, and the applications of intelligent Internet of things are increasing, such as smart community, smart home, smart car, intelligent green house, etc. (Xia et al., 2012). These applications mainly receive and analyze data in real time through sensors and other devices and feedback the results to the control system, so as to achieve the intelligent effect, and the actual system modeling is an important part. System identification theory for linear systems has been well-established and many applications of system identification have been reported. On the other hand, system identification theory for nonlinear systems has not been established so systematically, because of its extremely wide scope (Adachi et al., 2004). Neural network provides a new way for the modeling of unknown nonlinear dynamic systems. As long as the input and output of a dynamic system in the Internet of things are measurable, it can be identified online by neural network. In recent years, there has been many research on modeling, prediction, and control of the Internet of things using neural networks. Patra. J. C. proposed a neural network-based interface framework to automatically compensate for the nonlinear influence of the environmental temperature and the nonlinear-response characteristics of a capacitive pressure sensor to provide correct read out (2005). In the study of Manonmani, a neural network was used to model and control sufficient growth conditions of a greenhouse system (GHS) resulting in high cross yield, advanced production period, better quality, and less use of protective chemicals (2016). In the study of Hamid Taghavifar, the potential of a supervised artificial neural network (ANN) approach was assessed to diagnose the energy consumption and environmental indexes of application production in the learning location (2015).
The research of neural network in nonlinear dynamic system identification is as follows. In the past decade, due to the ability of learning on the basis of appropriate error function optimization and the good performance of approximation of nonlinear function (Antsaklis, 1992), ANN based on different examples (MLP, RBF, etc.) have been widely used as powerful learning tools for complex system identification and control tasks. In 1990, Narendra and Artasarathy proposed effective identification and control of nonlinear dynamic systems using MLP (1990). Because of its simple structure, RBF (Haykin, 2008; Pislaru & Shebani, 2014) is considered as an alternative to MLP. At the same time, in order to obtain better performance, some algorithms (such as BP) are used to train the role of network-based neural networks in complex power system/plant modeling and control (Narendra & Parthasarathy, 1990; Park & Sandberg, 1991; Parlos et al., 1994). Although studies have shown that these networks with corresponding algorithms can effectively identify and control complex process dynamics, they can achieve better performance (Zhao & Zhang, 2009) on the premise of increasing computational complexity. In general, the training of MLP and RBF is usually based on the back propagation (BP) or gradient descent (GD) training algorithm with fixed learning rate, but it is difficult to find the optimal learning rate in BP or GD algorithm. Basically, if the selected learning rate is very small, then the convergence speed of the network is very slow, and it takes a long time to converge. On the other hand, the high learning rate will lead to unstable learning process and network dispersion (Cao & Lin, 2008; Yoo et al., 2006). There are also other novel designs of FNN, such as self-organizing fuzzy neural networks (SOFNN) (Han et al., 2017) and self-organizing deep belief networks (SODBN) (Qiao et al., 2018), which are capable of constructing a simple fuzzy network without expert knowledge. A new growing and pruning algorithm was proposed in (Han et al., 2010) which is named as self-organizing radial basis function (SORBF) for RBFNN.