In modern local area networks, the communication signals sent from one computer to another across the lines of transmission are degraded because of reflection at the receiver. This reflection can be characterized through the impedances of the transmitter and the receiver, and is defined by the Institute of Electrical and Electronic Engineers (IEEE) as the S11 return loss. The specifications for S11 return loss in Gigabit Ethernet are given in terms of magnitude only in the IEEE 802.3 guidelines. This does not fully take into account, however, the effects of frequency dependant impedances within the bandwidth of interest. With a range of 30% error in the category 5, or CAT5, transmission line impedance used in this specification and no further requirements for individual components within the Gigabit Ethernet port, such as the RJ45 magjack or the physical layer, the system can easily be out of tolerance for return loss error. A simple impedance matching circuit could match the CAT5 cable to the physical layer such that the return loss is minimized and the S21 transmission is maximized. The first part of the project was commissioned by Dell Computer to characterize the return loss of all of its platforms. This thesis goes further in the creation of a system that can balance these two impedances so that the IEEE specification failure rate is reduced with the lowest implementation cost, size, power and complexity. The return loss data were used in the second phase of the project as the basis for component ranges needed to balance the impedance seen at the front of the physical layer to the CAT5 transmission line. Using the ladder network theory, an impedance matching circuit was created that significantly reduced the S11 return loss in the passband of the equivalent ladder network. To manage this iterative process, a control loop was also designed. While this system does not produce the accuracy that a programmable finite impulse response (FIR) filter could, it does improve performance with relatively minimal cost, power, area and complexity.

This thesis proposes innovative methods to extract information embedded in the frequency and time domain response of the transformer winding, and utilizes them to suggest solutions to a few tasks that have until now been thought difficult, if not impossible, to resolve. Pursuing this philosophy originated from the basic under- standing that the response of any physical system (behaving largely as a linear time invariant system) has embedded information that characterizes it completely. So, the prerequisite is to evolve ways to extract this information from measured responses. Once that is done, a variety of interesting applications can be envisaged. The two applications considered in this thesis are- •Investigate indirect measurement of the series capacitance of a transformer winding using the measured frequency or time domain response •Explore the possibility of increasing the physical resolution of the ladder network used to model a fully interleaved-disk winding In the former application, since direct measurement of series capacitance is impossible, alternatives based on indirect measurement were also not attempted. Similarly, in the latter application, the upper limit is known to be fixed by the number of distinctly observable peaks in the magnitude frequency response, so the question of increasing this limit was also never explored. Solutions to these tasks are proposed after a systematic analysis of frequency/time domain responses of a winding, initially modeled as a lumped parameter ladder network, to extract correlations that exist between them and winding parameters, and finally examine how these relations can be exploited together with the measured responses. Each of the five chapters is dedicated to describe the solution to one task. In each chapter, analytical formulation is presented first, followed by experimental results. Good agreement with the predicted results demonstrates its practicability. In final summary, indirect measurement of the series capacitance of a winding and en- hancing physical resolution of a ladder network model to represent a fully interleaved- disk winding was successfully demonstrated and they are the main contributions of this thesis.

Large capacity high voltage power transformers are one of the most expensive items of equipment in an electrical power network. Power utilities can ill-afford breakdown of transformers, especially, in a deregulated scenario. The consequences of such a failure are well known. Under these circumstances, utilities have figured-out that condition-based monitoring and diagnosis is worth pursuing, in spite of increased expenditure. Thus, monitoring and diagnosis is an integral part of operation and maintenance. Mechanical forces generated during short-circuits is the main cause leading to displacement/deformation of windings. Frequency response measurements have attained worldwide acceptance as a highly sensitive monitoring tool for detecting occurrence of such events. This is evident from the fact that customized commercial equipment are available (popularly called FRA or SFRA instruments), and with recent introduction of an IEEE draft trial-use guide for application and interpretation of frequency response analysis. Once a damage is detected, the next task is to identify its location along the winding and, if possible, determine its extent of severity. Understandably, these two tasks are best achieved, without disassembling the transformer and should ideally be based on off-line and on-site terminal measurements. In this regard, literature analysis reveals that recent research efforts have successfully demonstrated possibilities of using frequency response data for localization of discrete change in windings. This is indeed noteworthy, in spite of one major drawback. This pertains to excessive computing time needed to synthesize large-sized ladder-network, which automatically limits its practical use. Keeping these issues in mind, a research was initiated to find alternatives. The primary objective of this thesis is to examine the use of- • Complexnetwork-function-lociforlocalizationofadiscretechangeinasingle,isolatedtransformerwinding,basedonterminalmeasurements It goes without saying that the proposed method should be non-invasive, simple, time-efficient and overcome drawbacks in the earlier approach. A brief summary of the proposed method follows- This thesis presents a different approach to tackle the problem of localization of winding deformation in a transformer. Within the context of this thesis, winding deformation means, a discrete and specific change imposed at a particular position on the winding. The proposed method is based on the principle of pre-computing and plotting the complex network-function-loci (e.g. driving-point-impedance) at a selected frequency, for a meaningful range of values for each element (increasing and decreasing) of the ladder network. This loci diagram is called the nomogram. After introducing a discrete change (to simulate a deformation), the driving-point-impedance (amplitude and phase) is measured again .By plotting this single measurement on the nomogram, it is straightforward to estimate the location and identify the extent of change. In contrast to the earlier approach (wherein the entire ladder-network had to be synthesized for every new measurement), the proposed method overcomes the drawbacks, is non-iterative and yields reasonably accurate localization. Experimental results on a model coil and two actual transformer windings (continuous-disc and interleaved-disc) were encouraging and demonstrate its potential. Further details are presented in the thesis.

No description available.

京都大学 / 新制・課程博士 / 博士(工学) / 甲第23511号 / 工博第4923号 / 新制||工||1769(附属図書館) / 京都大学大学院工学研究科電気工学専攻 / (主査)教授 松尾 哲司, 教授 雨宮 尚之, 准教授 久門 尚史 / 学位規則第4条第1項該当 / Doctor of Philosophy (Engineering) / Kyoto University / DFAM

Weight-sharing property in convolutional neural network (CNN) is useful in reducing number of parameters in the network and also introduces regularization effect which helps to gain high performance. Non-weight-shared convolutional neural networks also known as Locally connected networks (LCNs) has potential to learn more<br>in each layer due to large number of parameters without increasing number of inference computations as compared to CNNs. This work explores the idea of where Locally connected layers can be used to gain performance benefits in terms of accuracy and computations, what are the challenges in training the locally connected networks and what are the techniques that should be introduced in order to train this network and achieve high performance. Partially-local connected network (P-LCN) VGG-16 which is hybrid of convolutional layers and Locally connected layers achieves on average 2.0% accuracy gain over VGG-16 full convolutional network on CIFAR100 and 0.32% on CIFAR10. Modified implementation of batch normalization for Full LCNs (all layers in network are locally connected layers) gives improvement of 50% in training accuracy as compared to using CNN batch normalization layer in full LCN. Since L1, L2 and Dropout regularization does not help improve accuracy of LCNs, regularization methods which focuses on kernels rather than individual weight for regularizing the network were explored. Ladder networks with semi supervised learning achieves this goal. Training methodology of ladder networks was modified to achieve ∼2% accuracy improvement on Pavia-University hyper-spectral image dataset with 5 labels per class.