×
The submission system is temporarily under maintenance. Please send your manuscripts to
Go to Editorial ManagerThe paper deals with neural networks identification of ultimate moment capacity of steel-concrete composite beams on base of experimental results. Basic information on artificial neural networks and its parameters suitable for analysis of experimental results are given. Two types of neural network algorithms are used. Results of identification are reported. The results show that artificial neural networks are highly suitable for assessing the ultimate moment capacity of composite section. The proposed neural network was also used to explore the effect of the various parameters on the behaviour of composite beams.
Artificial Neural Networks (ANN) have been applied to structural engineering in recent years. Most of the researches are based on backpropagation neural networks due to its well-studied theory. A backpropagation neural network has been used to predict the ultimate torsional strength of reinforced concrete rectangular beams. The effects of the parameters, such as the number of nodes in the input, output and hidden layers and the pre-process of the training patterns, on the behaviour of the neural network have been investigated. The algorithm called 'resilient propagation algorithm' has been used to the performance of the neural network. After training, the generalization of the neural network was tested by the patterns not included in the training patterns. Once the neural network has been trained, the ultimate torsional strength of reinforced concrete is obtained very easily and efficiently. Based on the ANN results, a parametric analysis was carried out to study the influence of parameters affecting the ultimate torsional strength of reinforced concrete beams and these results are compared with the equations of ACI-code.
This paper explores the potential of using artificial neural networks to predict the ultimate moment capacity of steel-concrete composite beams with metal deck slabs. Basic information on artificial neural networks and parameters suitable for the analysis of experimental results are given. A multilayer backpropagation neural network is used for training and testing the experimental data. A comparison study between the experimental values and two models (neural network and AJSC models) is also carried out. It was found that 1he neural network model provides better results. The proposed neural network is also used to explore the effect of the various parameters on the behavior of beams.
The effect of thermal aging on the tensile properties of cast stainless steel during service in light water reactors has been evaluated and recorded by the Argonne National Laboratory. Tensile data for several experimental and commercial heats of cast stainless steel (CF-8M) are presented for predicting the change in tensile flow and yield stresses and engineering stress-strain curve as a function of time and temperature of service in the light water reactors using Neural Networks. Thermal aging increases the tensile strength of this type of steel. The result and correlation described by this work may be used for assessing thermal embitterment of cast stainless steel components.
The present study deals with the analysis of short reinforced concrete columns subjected to axial load. One of the efficient techniques is applied, known as artificial neural networks. The descent gradient backpropagation algorithm is employed for analysis. The optimum topology (which gives the least mean square error for both training and testing with a fewer number of epochs) is presented. The effects of the number of nodes in input and hidden layer(s), and selecting of leaming rate and momentum coefficient, on the behavior of the neural network, have been investigated. Due to the slow convergence of results when using descent gradient backpropagation, the faster algorithm called "resilient backpropagation algorithm" has been used to improve the performance of the neural network and the results have been compared with those obtained using the descent gradient backpropagation algorithm.
A spandrel beam is a structural member lies at the edge of a frame and is connected by a joint to the floor beam extending into the slab. The spandrel beams are primarily responsible for transferring forces from a slab to the supporting edge columns. This work investigates the possibility of using the artificial neural networks to model the complicated nonlinear relationship between the various input parameters associated with reinforced concrete spandrel beams and the actual ultimate strength of them. The descent gradient backpropagation algorithm was employed for predicting the ultimate strength of the reinforced concrete spandrel beams. The optimum topology (which gives least mean square error for both training and testing with fewer number of epochs) is presented. Effects of parameters such as, number of hidden layer(s), number of nodes in the input layer, output layer and hidden layer(s), initialization weight factors and selection of the learning rate and momentum coefficient on the behaviour of the neural network have been investigated. Because of the slow convergence of results when using descent gradient backpropagation, another algorithm which is faster called "resilient backpropagation algorithm" has been used. The neural network trained with the resilient backpropagation RPROP algorithm gives better results than that trained with the steepest descent algorithm with momentum GDM algorithm.
Castor Oil is a natural raw material, used to prepare Brominated Castor Oil (BCO) and quaternary ammonium salt based castor oil (TEt-CO). The two products were tested as demulsifiers and compared with a commercial demulsifier (Chimec2439) by using bottle test method. BCO showed a high ability on water separation efficiency 90% with a dose of 150µl at 120min time settling while TEt- CO showed a low water separation efficiency reached to 10%. The effect of the demulsifier BCO was tested by varying different variables which have an obvious effect on water separation efficiency such as: dose, temperature, time of mixing emulsion, pH and salinity of aqueous phase of emulsion, and water ratio. The effect of some additives (i.e. methanol, ethanol, xylene and toluene) on the efficiency of the BCO was tested for the purpose of enhancing its effectiveness to break the crude oil emulsion. The experimental data obtained by using BCO were formulated as a model using the Artificial Neural Networks (ANNs) to evaluating the water separation efficiency. Multi-layer perceptron artificial neural network was developed based on the collected data of this study. The results showed that the training algorithm of back propagation (BP) is sufficient enough in predicting BCO efficiency under different operation conditions. It was found that the correlation coefficient values are 0.9995 and 0.9999 for the testing and training data, respectively and the mean square error (MSE) was 6.18*10^-5 at 200 epochs.
Tn this paper, a hierarchical phoneme recognition system is proposed. The hierarchical approach is applied here to recursively partition the recognition problem into smaller and smaller sub-problems those are independently handled at the di~tinct nodes of the hierarchy. The nodes are individually set to chara~erize different properties of the input phoneme, or more precisely to make separate d~isioos on its pertinence to the different reference subgroups of phonemes. The full characteri:zation of the input phoneme is achieved by traversing some root-to-leaf path through the hierarchy. The relationships between the different features of phonemes and their pertinence to the different reference subgroups. are to be objectively characteriz.ed ttnd optimized here. This involves specifying the decisive subset of features for each pertinence decision and neglecting the remaining features those are irrelevam to (or probably have negative effect on) that decision, at each node of the hierarchy. The optimization applied through the feature election process here, is not aimed at reducing the amount of feamres to be used in the recognition process, for the purpose of decreasing the time-complexity of the systcn1, but, is interested in enhancing the decision making accuracy of the system by avoiding the misleading features.
The land surface erosion is controlled by multifarious of different parameters, such as slope, soil physical properties (texture, structure, permeability, etc.), rainfall, runoff, and crop cover. However, it is impossible to develop precise simplest mathematical model that can predict the values of land surface soil erosion due to the behavior of controlled parameters. This paper presents the Neural Networks Model for assessing land surface soil erosion as amass per unit area per unit of time. The model derives from the analysis data obtained from available literature and was formulated as linear regression model and back propagation algorithm neural model. Both models were built by correlating firstly five watersheds variables with land surface erosion and secondly ten watershed variables with land surface erosion. The coefficients for independent variables were highly significant for both models. The case of correlating 10- watershed variables with land surface erosion gives R=0.978 & 0.976 for both models which is higher than that for 5- watershed variables. The mean absolute relative error (MARE%) is another procedure that used in order to evaluate the accuracy of the model and The average error % is 0.025 for (5) variables and 0.0064 for (10) variables. Both the supporting practices (P) and the slope length and slope steepness (LS) coefficients have a marked effect on the amount of land surface erosion in the case of 5- watershed variables. The amount of land surface erosion show a high level of sensitivity to the content of fine sand% in soil (FS) watershed variables on The amount of land surface soil erosion.
The dieless drawing process is an innovative method emanated and appeared in coincidence with development of the concept of metal superplasticity. It is utilized from the local heating of a wire or tube to a specified temperature and followed by a local cooling, so an additional deformation is inhibited. In this study, a special dieless drawing machine was designed to carry out an experimental program on SUS304-stainless steel wire having diameter of (1.6-2) mm to investigate the main process parameters such as speeds, heat quantity, heating coil width and heating-cooling separation distance. Also, a numerical model based on thermo-mechanical analysis was developed and validated with experimental program. Furthermore, an artificial neural network ANN model based on current experimental data was prepared to predict the dieless drawing behavior. A maximum area reduction of 40.7% was obtained in single pass. A 3.12mm/s feeding velocity and 4.97mm/s drawing velocity were realized through the experimental tests. The results showed that both drawing force and wire profile were effected by increasing of feeding speed, heating coil width and separation distance. Also, it is confirmed that strain rate was reduced by increasing the heating coil width and the reduction ratio was promoted. A maximum error of 21% was recorded between ANN model and experimental results. The results showed a good agreement among experimental, numerical and ANN models.
This paper considers the neural network based PID controller. The learning and generalization properties of neural network are utilized in improving the performance of a conventional PID controller. Two different schemes are introduced. Both schemes are studied and their performances are comparatively evaluated on an example for uncertain system.
This paper introduces a radial distribution feeder protection scheme based on certain features extraction from current signals measurement at the substation. The features are captured using the discrete wavelet transform (DWT). Two digital signals processing methods are used to introduce those features to the 1) fault detection 2) identification and 3) localization schemes; the first one is the energy method and the second one is the root mean square method. For the purpose of fault type identification, two systems are tested and compared, a Fuzzy Inference System (FIS) and Artificial Neural Network (ANN). Fault location scheme is then built based on ANNs. An effort is made to reduce the computational burden and the speed of detection provided by the fault detection and identification schemes. Since the short circuit faults are the most likely types of faults that can occur in power systems, the ten types of these faults taking into account different fault resistances are simulated in MATLAB environment and the protection scheme is built based on the idea of over current. The power quality disturbances such as switching transient events on the feeder is also taken into account in order to build a reliable and secure protection scheme.
Interest in neural networks as an alternative to the conventional algorithmic techniques has grown rapidly in recent years. Noise removal or noise suppression is an important task in image processing. In general, the results of the noise removal have a strong influence on the quality of the following image processing techniques. In this paper, two feed forward NN schemes have been presented for impulsive noise removal. The computation is reduced by using an artificial image in training. Results of NN schemes show high performance especially when the ratio of impulsive noise in testing are the same or greater than that of training image. The presented schemes are used for grayscale and also for truecolor.
Safwan-Zubair area is regarded as one of the important agricultural areas in Basrah province, South of Iraq. The aim of this study is to predict groundwater level in this area using ANNs model. The data required for building the ANNs model are generated using MODFLOW model (V.5.3). MODFLOW model was calibrated based on field measurements of groundwater level in 13 monitoring wells during a period of one year (Nov./2013 to Oct/2014). The neural network toolbox available in MATLAB version 7.1 (2010B) was used to develop the ANN models. Three layers feed-forward network with Log- sigmoid transfer function was used. The networks were trained using Levenberg-Marquradt back-propagation algorithm. The ANN modes are divided into two groups, each of four models. The input data of the first group include hydraulic heads, while, the input data of the second group include hydraulic heads and recharge rates. Based on results of this study it was found that; the best ANN model for predicting groundwater levels in the study area is obtained when the input data includes hydraulic heads and recharge rates of two successive months preceding the target month, the best structure of ANN model is of three layers feed-forward network type composes of two hidden layers, each of ten nodes, and the including of recharge rates as input data, beside the hydraulic heads has improved slightly the results.