Fengzhi Dai

Application of a neuro-fuzzy system to control a mobile vehicle

We describe a fuzzy control based on a neural network, which is obtained by merging the advantages of a neural network, a competitive algorithm, and fuzzy control. This adaptive fuzzy control system can deal with data sampled by a neural network. From such training data, it can produce more reasonable fuzzy rules by a competitive (clustering) algorithm, and finally control the object by the optimized fuzzy rules. This is not a simple combination of the three methods, but a merger into one control system. Some experiments and future considerations are also given.

A Survey of Image Segmentation by the Classical Method and Resonance Algorithm

Computer vision and recognition plays more important role on intelligent control (Chen & Hwang, 1998). For an intelligent system, it is necessary to acquire the information of the external world by sensors, to recognize its position and the surrounding situation. Camera is one of the most important sensors for computer vision. That is to say, the intelligent system endeavours to find out what is in an image taken by the camera: traffic signs, obstacles or guidelines. For image analysis, image segmentation is needed, which means to partition an image into several regions that have homogeneous texture feature. This process is usually realized by the region-based, boundarybased or edge-based method (Castleman, 1998). And from the viewpoint of clustering, it is divided into supervised and unsupervised texture segmentation. Since before segmentation, the intelligent control system seldom knows the feature of the image, e.g. which type and how many types of textures exist in an image, thus the unsupervised segmentation algorithm is always needed, although it is more difficult than the supervised method (Dai, Zhao & Zhao, 2007). In this chapter, the classical method (Agui & Nagao, 2000) and the resonance theory (Heins & Tauritz, 1995; He & Chen, 2000) are proposed respectively for image segmentation. The classical method is simple but practicable, which will be introduced in section 2. But for some situations, it is not suitable for complex image segmentation (e.g., the gradient variations of intensity in an image). We know that human vision can recognize the same texture that has gradient variations of intensity. And many image segmentation methods are proposed based on the change of intensity (Nakamura & Ogasawara, 1999; Deguchi & Takahashi, 1999). But they always fail to handle the wide-ranged gradations in intensity (Jahne, 1995). It is usually difficult to give a suitable threshold for pixel-based image processing methods to deal with this gradation. Resonance algorithm is an unsupervised method to generate the region (or feature space) from similar pixels (or feature vectors) in an image. It tolerates gradual changes of texture to some extent for image segmentation. The purpose of section 3 is to propose the resonancetheory-based method for image segmentation, which means that the same texture in an image will be resonated into one region by seed pixels. This method assumes that the

Image segmentation by resonance algorithm

Computer vision and recognition plays more important role on intelligent control. In this paper, the image segmentation is done by the resonance theory. Resonance algorithm is an unsupervised method to generate the region (feature space) from similar pixels (feature vectors) in an image. It tolerates gradual changes of texture to some extent for image segmentation. The purpose of the paper is to propose a practical method for image segmentation, which is always the first step to control a real intelligent control system.

Pattern recognition and control by adaptive methods for an intelligent mobile vehicle

This paper describes adaptive methods for both pattern recognition and control in an experimental mobile vehicle (MV). An adaptive resonance theory (ART) neural network is used as the character recognizer. It can self-organize and self-stabilize in response to complex binary input vectors. New input patterns can be saved in such a fashion that the stored patterns are not forgotten or destroyed. By merging the advantages of the feed-forward neural network, adaptive algorithm, and fuzzy control, a neuro-fuzzy system also is proposed. This can deal with a large amount of training data in the neural network, from these data produce more reasonable fuzzy rules with the adaptive algorithm, and then control the object by fuzzy control. This is not a simple combination of the three methods, but a merger into one intelligent control system. Finally, the experimental results and some conclusions are given.

Fundamental research on polymer material as artificial muscle

Presently, breakthrough industrial technology and scientific technology produce various instruments of downscale and weight saving. But the metal parts are widely used for magnetic motors. It is difficult to lighten the weight and to miniaturize the motor. Therefore, we need small, light actuators that can be built into motor. We focus attention on the polymer material artificial muscle that responds to electrical stimulation with a significant shape or size change. In this research, the artificial muscle was produced by using the conductive grease and the polymer material. We used several kinds of polymer material films in our experiments to compare the performance of them. The displacement when the voltage is applied to electrode was measured. At last the result and conclusion are given.

Fuzzy fault query approach in the fault diagnosis of a power plant system

Based on the theory of fuzzy recognition, this method only depends on experience and statistical data to set up the fuzzy fault query relationship between the outside phenomena (fault characters) and the fault sources (fault states), so that it reaches the goal of quick diagnosis.