Catálogo de publicaciones - libros

A neural oscillator model is applied to swing/giant-swing control of an under-actuated double pendulum. The oscillator receives the angle signal of the upper link and provides a relative angle of the lower link. The oscillator tunes itself to the natural dynamics of the pendulum system so as to increase the swing amplitude, and finally the pendulum system enters the phase of giant swing motion. A most remarkable result is that transition from simple swing to giant swing is attained without changing the values of the parameters of the neural oscillator at all.

Based on omnipotent approximation principle, a new neural network indirect adaptive sliding mode controller is designed for a class of nonlinear interconnected systems with uncertain dynamics. Different neural networks are adopted to approximate the affection of the uncertain terms in the subsystems and the interconnected terms to the whole system. It uses the mode transformation function to realize the changing between the NN indirect adaptive controller and fuzzy sliding mode controller, which keeps the state of the system changing in a close bounded set. By using Lyapunov method, it is proved that the close-loop system is stable and the tracking errors convergence to a neighborhood of zero. The result of the emulation proofs the validation of the designed controllers.

We introduce the support vector machine adaptive control by Lyapunov function derivative estimation. The support vector machine is trained by Kalman filter. Support vector machine is used to estimate the Lyapunov function derivative for affine nonlinear system, whose nonlinearities are assumed to be unknown. In order to demonstrate the availability of this new method of Lyapunov function derivative estimation, a simple example is given in the form of affine nonlinear system. The result of simulation demonstrates that the sequential training algorithm of support vector machine is effective and support vector machine control can achieve a satisfactory performance.

A model following adaptive control based on neural network for the electro-hydraulic servo system (EHSS) subjected to varied load is proposed. This proposed control utilizes multiple neural networks including a neural controller, a neural emulator and a neural tuner. The neural controller with specialized learning architecture utilizes a linear combination of error and the error’s derivative to approximate the back propagation error for weights update. The neural tuner is designed to adjust the parameters of the linear combination. The neural emulator is used to approximate the Jacobian of plant. The control of the hydraulic servo actuator is investigated by simulation and experiment, and a favorable model-following characteristic is achieved.

We use principal component analysis (PCA) to identify exons of a gene and further analyze their internal structures. The PCA is conducted on the short-time Fourier transform (STFT) based on the 64 codon sequences and the 4 nucleotide sequences. By comparing to independent component analysis (ICA), we can differentiate between the exon and intron regions, and how they are correlated in terms of the square magnitudes of STFTs. The experiment is done on the gene F56F11.4 in the chromosome III of C. elegans. For this data, the nucleotide based PCA identifies the exon and intron regions clearly. The codon based PCA reveals a weak internal structure in some exon regions, but not the others. The result of ICA shows that the nucleotides thymine (T) and guanine (G) have almost all the information of the exon and intron regions for this data. We hypothesize the existence of complex exon structures that deserve more detailed analysis.

As the linear method is difficult to recover the sources from the nonlinear mixture signals, in this paper a new nonlinear adaptive blind signal separation algorithm based kernel space is proposed for general invertible nonlinearities. The received mixture signals are mapped from low dimensional space to high dimensional kernel feature space. In the feature space, the received signals form a smaller submanifold, and an orthonormal basis of the submanifold is constructed in this space, as the same time, the mixture signals are parameterized by the basis in the high dimensional kernel space. In the noiseless or noisy situation, the sources are rebuilt online processing by M-EASI and subspace tracking. The results of computer simulations are also presented.

In this paper, a novel hybrid intelligent forecasting model based on empirical mode decomposition (EMD), support vector regression (SVR) and adaptive linear neural network (ALNN) is proposed, where these intrinsic mode components (IMCs) are adaptively extracted via EMD from a nonstationary time series according to the intrinsic characteristic time scales. Tendencies of these IMCs are forecasted with SVR respectively, in which kernel functions are appropriately chosen with these different fluctuations of IMCs. These forecasting results of IMCs are combinated with ALNN to output the forecasting result of the original time series. The proposed model is applied to the tendency forecasting of the Mackey-Glass benchmark time series and a vibration signal from a machine set. Testing results show that the forecasting performance of this proposed model outperforms that of the single SVR method under single-step ahead forecasting or multi-step ahead forecasting.

In this paper, a neural network is applied to process the color gamut mapping in real time. Firstly, a neural network is trained to learn the highly nonlinear input and output relationship of the color gamut mapping. And then, the trained neural network is simplified with a look-up table and an address decoder for a fast computation. The proposed method can be easily implemented in high speed and low cost hardware. Simulation result shows the soundness of the proposed method.

A novel method for identifying a chaotic system with time-varying bifurcation parameters via an observation signal which has been contaminated by additive white Gaussian noise (AWGN) is developed. This method is based on an adaptive algorithm which takes advantage of the good approximation capability of the Radial Basis Function (RBF) neural network and the ability of the Extended Kalman Filter (EKF) for tracking a time-varying dynamical system. It is demonstrated that, provided the bifurcation parameter varies slowly in a time window, a chaotic dynamical system can be tracked and identified continuously, and the time-varying bifurcation parameter can also be retrieved in a sub-window of time via a simple least-square-fit method.

Traffic behavior in a large-scale network can be viewed as a complicated non-linear system, so it is very difficult to describe the long-term network traffic behavior in a large-scale network. In this paper, according to the non-linear character of network traffic, the time series of network traffic is decomposed into trend component, period component, mutation component and random component by different mathematical tools. So the complicated traffic can be modeled with these four simpler sub-series tools. In order to check the decomposed model, the long-term traffic behavior of the CERNET backbone network is analyzed by means of the decomposed network traffic. The results are compared with ARIMA model. According to autocorrelation function and predicting error function, compounded model can get higher error precision to describe the long-term traffic behavior.