INTERNATIONAL JOURNAL OF ENGINEERING SCIENCE (ENGLISH)
Year:2004 | Volume:15 | Issue:3|Papers Count:18

Backpropagation algorithm performs gradient descent only in the weight space of a network with fixed topology. A very small network cannot learn the problem well, and a very larger network will lead to overfitting and poor generalization performance. Algorithms that can find an appropriate network architecture automatically are thus highly desirable. The algorithms that are introduced by researchers can be classified into five major groups. Pruning algorithms, constructive algorithms, hybrid algorithms, evolutionary algorithms, and learning automata based algorithms. Meybodi and Beigy introduced the first learning automata based algorithms, called survival algorithm. This algorithm produces networks with low complexity and high generalization. Survival algorithm by turning off and on the weights, tries to find the most important weights. At the beginning, all weights of the network are on and contribute to learning. The on weights, whose absolute values are less than a threshold value, are penalized and those, whose absolute value are larger than another threshold value, are rewarded. The on weights, whose absolute values lie between these two threshold values, neither rewarded, nor penalized. The values of these two thresholds are determinative and have considerable effect on the performance of the survival algorithm. Determination of the values of these thresholds is not an easy task and usually is determined by trial and error or using past experience. In this paper, we propose a method for adaptation of these two threshold values. The proposed method have been tested on number of problems and shown through simulations that the network generated by the survival algorithm when threshold values are adapted has lesser number of weights and neurons, comparing to the network generated by the first version of the algorithm reported earlier. Experimentation shows that the adaptive survival algorithm has nearly the same degree of generalization as the non-adaptive version.

The use of array antennas in mobile telecommunication systems has proven to increase the spectrum efficiency of the system. The use of beamforming techniques poses high linearity demands on both the RF/IF up and down conversion chain. In this paper, an overview of the various nonlinear effects is first presented. By the way amplitude distortion produced in the transmitter up conversion chain and amplification that can affect the antenna beam specification is simulated. Results show that the IMD products increase the level of spurious radiation and ultimately reduce the performance of the beamformer. In the receiving mode, AM-AM conversion due to gain compression of the Low Noise Amplifiers (LNAs) is modeled and the beam pattern degradation is simulated. Using behavioral modeling, a new relation for AM-PM conversion is established as a gain compression. Results show that the amplitude and phase distortion of the LNA vary for different gain setting and over signal bandwidth and indicate the need for regular calibration in using beamforming techniques. Also, results show that the produced radiation pattern is more affected by AM-PM distortion rather than AM-AM distortion.

The method of least squares is applied to the solution of the junction of cylindrical waveguides and its formulation is simplified. An error function is constructed from the boundary conditions (B.C.) at the junction which is a quadratic function of the modal amplitudes of the modal expansions inside the input and output guides. The unique minimum point of the error function gives the optimum conditions for the satisfaction of the junction B. C.'S. for the specified number of modes. A weighting factor is multiplied by the term due to the magnetic field B.C.'S. at the junction, which considerably affects the excited modal amplitudes and the field pattern at the junction. The conservation of reaction is used to compute the equivalent junction susceptance and together with the conservation of complex power determine the optimum range of values of the weighting factor.The special problem of the step-up and step-down junction of coaxial circular waveguides containing a circular metallic iris with TE11 mode excitation is analyzed in detail and several computer programs are. written in the C language.The typical problems solved by MLS show very good satisfaction of junction B.C.'s. and the equivalent junction susceptances compare very well with experimental and numerical results available in the literature.

Following the image processing steps on the structured light image, we have a binary image consisting of single pixel width pattern which shows where the projected lines are. The contours of subject surface mean that certain parts of the projected grid can not be seen on the captured image (dead shadow areas) and breaks occur in the projected grid itself around the sharp edges or holes. In order to produce a 3-D data set, all of the lines must be perfectly formed, of single-pixel thickness and be continuous. In this paper, the grid lines are interpreted and decisions must be made as to whether any given pixel belongs to which line. The main stage involved in this procedure is tracing. A novel approach of loop tracing introduced to overcome the problems encountered. As a pre-processing required before matching, the pixels detected during the feature extraction process are grouped into continuous lines using the tracing algorithms. These algorithms exploit the fact that the projected pattern was continuous, and allow us to recover the structured light image.

The number of satellite imageries, provide by remote sensing satellite sensors, is growing fast in the civilian market. At the present time, the problem with these images for many users is the lack of sensor calibration information and precise ephemeris data [1]. Therefore, there is a need for an alternative approach to extract 2D1 and 3D2 spatial information from these images. 2D geometric correction of satellite imageries one of the fastest and easiest methods of the registration of satellite imagery to a known coordinate system [2]. In this respect, global polynomial based models are widely used. However, finding the best terms of the polynomials that best first to the imagery is very important in order to reach best possible accuracy. This paper describes development and implementation of number of hybrid algorithms that each employs a genetic algorithm in addition to employing other existing algorithms, such as globalpolynomial and multi- quadric, for two- dimensional geometric correction of IKONOS3 imagery. The method at first uses a genetic Algorithm technique to find the best term for global polynomial, which provides the best fitness model for imagery to the ground space. Then, a number of effective ground control points "pointwise polynomial functions" utilizing genetic algorithm, are selected. Finally, in order to evaluate this new technique, the result of this method that applied to IKONOS imagery, is compared with global polynomial and multi-quadric mathematical models.

In this paper the effects of viscous dissipation on thermal entrance heat transfer in a circular tube filled with a saturated porous medium, is investigated analytically on the basis of a Darcy model. The case of isothermal boundary is treated. The local and the bulk temperature distribution along with the Nusselt number in the thermal entrance region were found. The fully developed Nusselt number -independent of the Brinkman number- is found to be 8. It was observed that neglecting the effects of viscous dissipation would lead to the well-known case of internal flows, with Nu equal to 5.78. Meanwhile the effect of viscous dissipation on the developing temperature distribution is presented and discussed in detail. Another solution is obtained through application of a finite difference scheme. A comparison between the results of these two methods (analytical and numerical) showed splendid confirmations.

This paper describes the application of eigenvalue analysis for surge prediction in multi-stage axial flow compressors. In this approach, the compression system is modeled using a stage-by-stage nonlinear modeling based on the conservation equations of mass, momentum and energy. By line arising the compressor flow model at each steady state point, the small perturbation system matrices have been obtained. By analyzing the eigenvalues of the linearised model, the stability of the compressor flow has then been investigated, and surge points for both low and high compressor rotational speeds have been predicted. Finally, the results for a seven stage axial flow compressor have been compared with experimental data in order to show the ability of the method.

Based on previous studies, prediction of flow in the draft tube, as a sample of complex flow phenomena, has been shown to be of vital importance in modeling hydroturbine flows.In this research, numerical solutions are presented for the turbulent flow through the draft tube of Masjed-e- Soliman HEPP., a new project constructed during recent years in the south of Iran [5]. The special features of the subject flow, such as rapid curvature, separation, swirl, and vortices require exact physical model investigation and also great challenges in numeric and turbulence modeling aspects.Unfortunately, as the physical measurements provided by the supplier are incomplete, no physical data are available. Thus, using known software (with light background) and grid independence study are two other ways for verification of results. However, this paper demonstrates the capability of CFX-tascflow, a commercial CFD software Widely used for complex flow prediction, and NS3D, a new CFD code developed by LHM. Reynolds-Averaged Navier-Stockes (RANS) equations are solved in both of them and standard k-ε and standard k- w turbulence models are examined In the end, the flow structures are discussed using various graphical post processing results.

A probabilistic analysis of the fatigue crack growth for reliability growth calculation on the mechanical component is presented on the basis of fracture mechanics and theory of random process. The loading is postulated to be stationary, narrow-band random Gaussian process and consequently, randomized Paris-Erdogan law is applicable. As a specific problem, a thin plate having a central crack is analyzed by two analytical methods "stochastic averaging (SA)" and "linear fatigue damage accumulation (LFDA)". The aforementioned plate is being analyzed by probabilistic finite element method (PFEM) on the basis of Monte-Carlo simulation. It is shown that SA method presents the most conservative results for brittle materials comparing with FEM and LFDA.

Polycrystalline diamond thin films were grown by the hot filament chemical vapor deposition (HFCVD) method. Trimethyl phosphite diluted in acetone was used as a source of in-situ doping. Thickness of the diamond films varied between 6 to 8 micrometers. These films were characterized by such methods as Raman spectroscopy, X-ray diffraction, and the scanning electron microscopy upon which the formation of the diamond phase was verified.Various metallic contacts were formed on the diamond films by either dc sputtering or thermal evaporation. Of these contacts, In and NiCr showed a rectifying behavior. The NiCr-diamond film-Si-AI structure showed sensitivity to light. Temperature measurement was then performed on this heterostructure in order to better understand the conduction mechanism. These measurements indicated that the Schottky emission and the Frenkel-Poole emission conduction mechanisms are the dominating ones. At higher temperatures, the space-charge-limited conduction mechanism dominates.

Clustering is widely used in many scientific areas. In an ESM system clustering the input radar pluses is a main part of operation, which needs considerable time. The main goal of different clustering algorithms is the reduction of computation time to reach to a real-time system, which is a vital feature of ESM systems.In this paper a matrix-based method, MMC, is implemented for clustering pulse train. MMC is a multi-step algorithm, which in each step the sub-clusters of previous step are splitted. This division continues until no dividable sub-cluster will remain. To reach a real-time operation MMC must perform in parallel manner. So in this paper MMC is implemented by systolic array, which is the best parallel architecture for matrix operation. To evaluate MMC, a radar environment is simulated. The simulation results show MMC can act as an effective clustering algorithm in an ESM system.

The unified power quality conditioner (UPQC) consists of a series active power filter (for conditioning voltage quality) and a shunt active power filter (for compensating current disturbances).This paper presents an improved scheme for controlling UPQC based on the instantaneous power theory. A model distribution network is simulated and analyzed that includes the different sources of temporary and sustained power quality disturbances. The main contribution of the paper are modeling of electric arc furnaces (as flicker producing loads) and simulation of single-phase short circuit faults (as source of voltage sags and swells), as well as the investigation of UPQC impact on damping of harmonic resonance. Based on the simulated results, the advantages and limitations of the proposed control scheme for UPQC are presented.

This paper proposes two Neural Networks (NNs) based approaches for on-line power system Transient Stability Evaluation (TSE). The first approach represents the application of feedforward neural networks in estimating the critical clearing time (t cr) for TSE. The second one presents a similar effort using the Kohonen neural network. Knowing that for a particular fault scenario, tcr is a function of only pre-fault system operating point, the main objective of the first approach is to investigate the way that one may develop a feedforward neural network based method to approximate this function by considering the smallest set of directly monitorable variables which characterize the pre-fault operating point adequately. In the other approach, it is shown that the small amount of changes in the pre-fault conditions leads to the small amount of changes in the tcr . So, a Kohonen neural network is used for the classification of the pre-fault conditions and estimation of the tcr. The proposed approaches have been successfully applied to a three-machine test system. The difficulties of using the Kohonen netbased approach in large power systems are also explored.

Periodic fully-developed fluid flow and heat transfer characteristics are obtained numerically for turbulent flow over a three-dimensional array of heated blocks deployed on a wall of parallel plate duct. The blocks have rectangular cross section and they are arranged in staggered array. This configuration simulates forced convection cooling of electronic equipment.The Lam-Brernhorst low Reynolds number form of the k-ε model is used for the computations. The computations are performed for the conditions of constants wall temperature and uniform heat generation per module. The Reynolds number ranged from 103 to 104, with the laminar prandtle number equal to 0.7 and the turbulent Prandtle number equal to 0.9. The results are compared with relevant experimental data from the literature and also with the numerical values obtained by other researchers for inline array.

This paper presents the results of mold filling and cures analyses of a Resin Transfer Molding (RTM) process. Three dimensional mold filling and cure analyses have been performed to model resin transfer molding of a hollow composite cylinder. Fiberglass and carbon fiber mats with EPON826 epoxy resin composites are considered. Numerical models are developed in commercial finite element software to simulate resin flow into a mold of cylindrical shape. The cylindrical part has inner and outer diameters of 9 and 10 centimeters, respectively while its length is 100 centimeters. Resin flow through fiber mats is modeled as flow through porous media with Darcy's law. The effects of reinforcement type on mold filling times are investigated. In modeling the composite cure cycle, three-dimensional energy and species equations are solved in the mold domain using a finite difference method. Thermal and cure histories of the composite part are determined in the cure cycle. Based on the results of this investigation the injection cycle of the process is the most time consuming part in the manufacturing of these composite parts. This information can help in the design of the mold, selection of injection pressure, and the cure cycle parameters prior to part manufacture.