The partial discharges (PD) measurement approach in power equipments is a very important inspection technique for insulation deterioration assessment. The PD based approach possesses the greatest potential for further development. This study proposes a noncontact type acoustic measurement system. We first investigate an acoustic measurement method in the laboratory. To prove the accuracy of the acoustic measurements, we proceed with, in the laboratory, signal-pattern comparison between the acoustic measurement method and the pulse current method. This study creates polar-coordinate and discharge type identification patterns. We propose the use of the q-£p-t patterns, the polar-coordinate patterns and discharge type identification patterns, with mutual cross-reference, to identify the discharge type. Then this study applies the wavelet transform to suppress noises; a wavelet mother function most similar to the acoustic PD signals is chosen and then set the filtering threshold value for the wavelet transform. The signals' features will be extracted after the noises are eliminated. The experimental results indicate that the application of wavelet transform can effectively eliminate the field noises. Next, the features will be used to build the training database for the back-propagation neural network (BNN) to construct the discharge patterns' recognition and identification system. Finally, we apply the finished neural networks to field signal-pattern identification. The proposed acoustic measurement system is applied on line to epoxy-resin transformers, power distributors, and the like. The superior measurement results we obtained shall be able to correctly identify power equipment's PD fault types.

In this research, a single PAN-TILT image servo system has been developed with real-time face tracing technology. First, the target face is detected, and then the target template is kept at the image center with the integration of optical flow algorithm and control theory. In motion control, back-propagation neural network is taken to predict and estimate the target position. Experiments are made to analyze the performance of the video servo control system.

This paper presents a model transformation between System Dynamics Model (SDM) and Artificial Neural Network (ANN) to aid model construction and policy design. We first point out a similarity between a System Dynamics Model (SDM) and an artificial neural network, in which both store knowledge majorly in the structure (or linkages) of a model. Then, we design a method that can map a SDM to a special design Partial Recurrent Network (PRN), and prove in mathematics that they two operate under the same numerical propagation constraints. With the established foundation, we then showed that the SDM-PRN transformation could aid SDM construction in the following way: (1) start from an initial skeleton of a PRN model (mapping from an initial SDM), (2) incarnate its structure by learning and (3) convert it back to a corresponding SDM. This approach integrates the capability of neural network learning with a traditional process, which thus makes model construction more systematic and much easier for common people. In the same philosophy, the SDM-PRN transformation could also aid SD policy design. Since any PRN can learn some structures from a historical time series pattern, it can also learn a better structure from a better pattern set by designer. We have investigated the effectiveness and usefulness of two application of the SDM-PRN transformation described above and the results are satisfactory.

ABSTRACT In this thesis, some aspects of the non-fixed neural network for power load forecasting are discussed. Unlike traditional fixed neural network technique, the structure of neural network is non-fixed during its training and testing phases. Based on the characteristic of the desired forecasting day, the number of input node utilized is changeable. The modified learning algorithms, including fuzzy back-propagation learning algorithm and stochastic back-propagation learning algorithm, will be used in the load forecasters we developed. For precise input selection of the neural network model, the analysis of mutual relationship between load and temperature and gray relational analysis between desired forecasting load and the related previous load are studied. Two types of load forecasting, i.e., peak load forecasting and hourly load forecasting, are investigated. Short term (one-to-several-day-ahead) load forecasting is considered in this research. Hourly loads and relevant temperature data from 1992 to 1998 provided by Taipower Utility and the Central Weather Bureau is implemented for this research. For demonstrating the feasibility and superiority of the forecasters we develop, several forecasting models, including fixed neural network with constant learning rate and momentum, recursive time series model, and artificial neural network short term load forecaster (ANNSTLF) proposed by [Kho.2], are also performed for a comparison. From the results of the simulation, better performances could be obtained by the methods we proposed. Not only the over-training phenomenon is obviously reduced, the forecasting accuracy and the learning speed of the neural model are also effectively improved.

This objective of thesis is to derive the adaptive load shedding by artificial neural network (ANN) so that the amount of load shedding can be minimized. An actual industrial customer and Taipower system are selected for computer simulation to fit the ANN model. The mathematical models of generation, exciters, governors and loads are used in the simulator program. The back propagation neural method is considered for the neural network training of load shedding.To create the training data set for ANN models, the transient stability analysis is performed to fit the load shedding under different operation and fault condition. The back propagation method and L-M learning process are then used to fit the minimum load shedding without causing system stability problem. To verify the effectiveness of the proposed methodology for adaptive load shedding, three fault contingencies for both the industrial cogeneration system and Taipower system have been simulated. By compare to the conventional load shedding, it is found that the amount of load shedding can be minimized and adjusted according to the real time operation conditions of power systems.

The study of system dynamics starts from model construction and simulation to understand and solve dynamical complicated problems. Traditionally approaches of modeling process depend on an expert¡¦s experiences and the trial & error procedure. Chen¡¦s research proposes a transformation method that could map a System Dynamics Model (SDM) to a specially designed Partial Recurrent Network (PRN). Thus he could use the neural network training algorithm to assist model construction and policy design. In this paper, we will introduce a Genetic Algorithm (GA) in the model building process, which encodes a PRN into a string and uses an evolution process to select a best solution. The algorithm not only improves the PRN training, but also generates more candidate models for consideration. Thus, it enhances the SDM-PRN transformation¡¦s usability.

Texture boundary detection (or segmentation) is an important capability of the human visual system. Usually, texture segmentation is viewed as a 2D problem, as the definition of the problem itself assumes a 2D substrate. However, an interesting hypothesis emerges when we ask a question regarding the nature of textures: What are textures, and why did the ability to discriminate texture evolve or develop? A possible answer to this question is that textures naturally define physically distinct surfaces or objects, thus, we can hypothesize that 2D texture segmentation may be an outgrowth of the ability to discriminate surfaces in 3D. In this thesis, I investigated the relative difficulty of learning to segment textures in 2D vs. 3D configurations. It turns out that learning is faster and more accurate in 3D, very much in line with what was expected. Furthermore, I have shown that the learned ability to segment texture in 3D transfers well into 2D texture segmentation, but not the other way around, bolstering the initial hypothesis, and providing an alternative approach to the texture segmentation problem.

Traffic congestion has become a menace to civilized society. It degrades air quality, jeopardizes safety and causes delay. Traffic congestion can be alleviated by providing an effective traffic control signal system. Closed-loop traffic control systems are an example of such a system. Closed-loop traffic control systems can be operated primarily in either of two modes: Time of Day Mode (TOD) or Traffic Responsive Plan Selection Mode (TRPS). TRPS mode, if properly configured, can easily handle time independent variation in traffic volumes. It can also reduce the effect of timing plan aging. Despite these advantages, TRPS mode is not used as frequently as TOD mode. The reason being a lack of methodologies and formal guidelines for predicting the factors and thresholds associated with TRPS mode. In this research, a new methodology is developed for determining the thresholds and factors associated with the TRPS mode. This methodology, when tested on a closed-loop system in Odem, Texas, produced a classification accuracy of 94%. The classification accuracy can be increased to 98% with a proposed TRPS architecture.

Neural networks (NNs) are widely used to investigate the relationship among variables in complex multivariate problems. In cases of limited data, the network behavior strongly depends on factors such as the choice of network activation function and network initial weights. In this study, I investigated the use of neural networks for multivariate analysis in the case of limited data. The analysis shows that special attention should be paid when building and using NNs in cases of limited data. The linear activation function at the output nodes outperforms the sigmoidal and Gaussian functions. I found that combining network predictions gives less biased predictions and allows for the assessment of the prediction variability. The NN results, along with conventional statistical analysis, were used to examine the effects of folding, bed thickness, structural position, and lithology on the fracture properties distributions in the Lisburne Formation, folded and exposed in the northeastern Brooks Range of Alaska. Fracture data from five folds, representing different degrees of folding, were analyzed. In addition, I modeled the fracture system using the discrete fracture network approach and investigated the effects of fracture properties on the flow conductance of the system. For the Lisburne data, two major fracture sets striking north/south and east/west were studied. Results of the NNs analysis suggest that fracture spacing in both sets is similar and weakly affected by folding and that stratigraphic position and lithology have a strong effect on fracture spacing. Folding, however, has a significant effect on fracture length. In open folds, fracture lengths in both sets have similar averages and variances. As the folds tighten, both the east/west and north/south fracture lengths increase by a factor of 2 or 3 and become more variable. In tight folds, fracture length in the north/south direction is significantly larger than in the east/west direction. The difference in length between the two fracture sets creates a strong anisotropy in the reservoir. Given the same fracture density in both sets, the set with the greater length plays an important role for fluid flow, not only for flow along its principal direction but also in the orthogonal direction.

Several computer-aided diagnostic (CAD) methods for solitary pulmonary nodules (SPNs) have been proposed, which can be divided into two major categories: (1) the morphometric CT method, and (2) the perfusion CT method. The first goal of this work is to introduce a neural network-based CAD method of lung nodule diagnosis by combining morphometry and perfusion characteristics by perfusion CT. The proposed approach has the following distinctive features. Firstly, this work develops a very efficient semi-automatic procedure to segment entire nodules. Secondly, reliable nodule classification can be achieved by using only two time-point perfusion CT feature measures (precontrast and 90 s). This greatly reduces the amount of radiation exposure to patients and the data processing time. As demonstrated in previous work, classification tuberculomas from malignancies has been considered to be a challenging task. However the diagnosis accuracy for tuberculomas reaches 92.9% by applying the proposed CAD method. Another goal of this work is, by investigating the relative merits of 2D and 3D methods, to develop a two-stage approach that combines the simplicity of 2D and the accuracy of 3D methods. Experimental results show statistically significant differences between the diagnostic accuracy of 2D and 3D methods. The results also show that with a very minor drop in diagnostic performance the two-stage approach can significantly reduce the number of nodules needed to be processed by the 3D method and thus alleviates the computational demand.