Catálogo de publicaciones - libros

This paper presents a neural network based implementation of space vector modulation (SVM) of a voltage-fed inverter that only includes the under-modulation region. SVM has recently grown as a very popular pulse width modulation (PWM) method for voltage-fed converter ac drives because of its superior harmonic quality and extended linear range of operation. However, a difficulty of SVM is that it requires complex on-line computation that usually limits its operation up to several kHz of switching frequency. Switching frequency can be extended by using high-speed digital signal processing (DSP) card and simplifying computations by using look-up tables. In our work a low cost microcontroller was used instead of DSP card. The performances of the drive with artificial neural network (ANN) based sector selected SVM are faster than conventional SVM. The performances of the drive with ANN based SVM are excellent.

This paper presents a novel approach for automatic detection of white blood cells in bone marrow microscopic images. Far more different from traditional color imaging analysis methods, a multispectral imaging techniques for image analysis is introduced. Multispectral image can not only show the spatial features of a cell, but also reveal the unique spectral information of each pixel. The supported vector machine (SVM) classifier is employed to train the spectrum vector of a pixel, and the output of the classifier can indicate the class type of the pixel: nucleus, erythrocytes, cytoplasm and background. Experimental results show that, compared with any other method previously reported, our method is more robust, precise and insensitive to smear staining and illumination condition.

According to the problems in the on-line measurement and automatic control of component content in rare-earth countercurrent extraction process, soft sensor strategies based on the mechanism modeling of the extraction process and neural network technology are proposed. On this basis, the intelligent optimal control strategy is provided by combining the technologies based on soft sensor and CBR (case-based reasoning) for the extraction process. The application of this system to a HAB yttrium extraction production process is successful and the optimal control, optimal operation and remarkable benefits are realized.

User-friendly Human-Computer interaction becomes more important accordance with rapid development of various information systems. In this paper we describe a three-dimensional gesture recognition algorithm and a system that adopts the algorithm for non-contact human-computer interaction. From sequence of stereo images, five feature regions are extracted with simple color segmentation algorithm and then those are used for three dimensional locus calculation processing. However, the result is not so stable, noisy, that we introduce principal component analysis method to get more robust gesture recognition results. This method can overcome the weakness of conventional algorithms since it directly uses three-dimensional information for human gesture recognition.

This paper presents a direct adaptive control scheme based on multi-layer neural networks for a class of single-input-single-output (SISO) uncertain nonlinear systems. The on-line updating rules of the neural networks parameters are obtained by Lyapunov stability theory. All signals in the closed-loop system are bounded and the output tracking error converges to a small neighborhood of zero. In this sense the stability of the closed-loop system is guaranteed. The effectiveness of the control scheme is verified by a simulation of inverted pendulum.

Most control applications closed over a shared network are suffering from the time-varying characteristics of flexible network workload. This gives rise to non-deterministic availability of communication resources and may significantly impact the control performance. In the context of integrating control and scheduling, a novel feedback scheduler based on neural networks is suggested. With a modular architecture, the proposed feedback scheduler mainly consists of a monitor, a predictor, a regulator and an actuator. An online learning Elman neural network is employed to predict the network conditions, and then the control period is dynamically adjusted in response to estimated available network utilization. A fast algorithm for period regulation is employed. Preliminary simulation results show that the proposed feedback scheduler is effective in managing workload variations and can provide runtime flexibility to networked control applications.

A humanoid walking gait synthesizing approach, which is able to generate gaits in both sagittal and frontal planes, is presented in this paper. To further improve the humanoid walking gait in consideration of both ZMP (Zero Moment Point) and energy consumption constraints, a two-stage optimization method is proposed. At the first stage, real-coded GAs (genetic algorithms) are used to generate a set of near-optimal walking gaits. At the second stage, the near-optimal walking gaits are used as training samples for a GA-based NN (neural network) to further improve the humanoid walking gait. By making use of the global optimization capability of GAs, the GA-based NN can solve the local minima problem. The proposed approach is able to generate near-optimal walking gait at any speed in feasible range. Experiments are conducted to verify the effectiveness of the proposed method.

An adaptive neural networks internal model controller is designed for a tilt rotor aircraft platform. The behavior of the research platform, in certain aspects, resembles that of a tilt rotor aircraft. The proposed control architecture can compensate external disturbances and dynamic inversion error. The controller includes an on-line learning neural network of inverse model and an off-line trained neural network of forward model. Lyapunov stability analysis guarantees tracking errors and network parameters are bounded. The performance of the controller is demonstrated using the tilt rotor aircraft platform, including nacelle tilting flight.

This paper presents a novel learning feed-forward controller design approach for accurate robotics trajectory tracking. Based on the joint nonlinear dynamics characteristics, a model-free learning algorithm based on Support Vector Machine (SVM) is implemented for friction model identification. The experimental results verified that SVM based learning feed-forward controller is a good approach for high performance industrial robot trajectory tracking. It can achieve low tracking error comparing with traditional trajectory tracking control method.

This note presents a robust adaptive neural network (NN) control scheme for multi-fingered robot hand manipulation system in the constrained environment to achieve arbitrarily small motion and force tracking errors. The controllers consist of the model-based controller, the NN controller and the robust controller. The model-based controller deals with the nominal dynamics of the manipulation system. The NN handles the unstructured dynamics and external disturbances. The NN weights are tuned online, without the offline learning phase. The robust controller is introduced to compensate for the effects of residual uncertainties. An adaptive law is developed so that no priori knowledge of the bounds for residual uncertainties is required. Most importantly, the exponential convergence properties for motion and force tracking are achieved.