International Journal of Neural Networks and Advanced Applications
E-ISSN: 2313-0563
Volume 5, 2018
Notice: As of 2014 and for the forthcoming years, the publication frequency/periodicity of NAUN Journals is adapted to the 'continuously updated' model. What this means is that instead of being separated into issues, new papers will be added on a continuous basis, allowing a more regular flow and shorter publication times. The papers will appear in reverse order, therefore the most recent one will be on top.
Submit a paper | Submission terms | Paper format
Volume 5, 2018
Title of the Paper: DTC-ANN Control with Two Level Inverter Associated by Different Observers (MRAS and KUBOTA) for Induction Machine Drives
Authors: A. Dris, M. Bendjebbar, Aek.Belaidi, H. Benbouhenni
Pages: 55-63
Abstract: In this article, we present the construction of observers of rotor flux and mechanical speed needed for robust control of the asynchronous machine. Two observers will be developed for comparison. The first is based on the techniques MRAS and the second is observer of KUBOTA, with the enhanced DTC, "sensorless DTC". The validity of the proposed methods is confirmed by the simulation results. This work is devoted to the construction of rotor flux observers and the mechanical speed necessary for the robust control of the asynchronous machine. Two speed observers will be developed for comparison. The first is based on the MRAS technique and the second is based on KUBOTA, with the improved DTC-ANN command "sensorless DTC control". The THD (Total Harmonic Distortion) of stator current, torque ripple and stator flux ripple are determined and compared with conventional DTC control scheme using Matlab/Simulink environment.
Title of the Paper: Stock Price Prediction on Daily Stock Data using Deep Neural Networks
Authors: Sneh Jain, Roopam Gupta, Asmita A. Moghe
Pages: 45-54
Abstract: The price of a stock is volatile and complex in nature which makes its prediction a difficult task. This paper plans to predict the prices of Tata Consultancy Services (TCS) and Madras Rubber Factory Limited (MRF) stocks on a short-term basis. In this paper, a comparative analysis of various Deep Neural Network techniques applied for a stock price prediction application is done. The networks used are pertinent to the problem include Convolutional Neural Networks, Long Short-Term Memory Networks and Conv1D-LSTM. The different neural network models are trained on daily stock price data which includes Open, High, Low, and Close price values. These are used to predict the next day closing price. From the last 5 days of data, the prediction is made. Results of different models are compared with each other. In this paper, a deep neural network Conv1D-LSTM is proposed which is based on the combining of layers of two different techniques – CNN and LSTM to predict the price of a stock. The performance of the models is evaluated using RMSE, MAE and MAPE. These errors in Conv1D-LSTM model are found to be very low compared to CNN & LSTM. For stock price prediction, Conv1D-LSTM network is found to be effective, depending on the nature of stock hyper-parameters may require some variations.
Title of the Paper: Sequence Association in Two-Layered Feedforward Neural Network
Authors: Maja Sarevska
Pages: 40-44
Abstract: Here we study the properties of robustness of two-layer feedforward network that stores association between two sequences in the two layers. Our work shows that robustness comes from the overlapping feedforward projections from the first to the second layer in the network. Recurrent connections in the second layer further improve robustness, while noise is decreasing it. Incorporating biological aspects of neural network in more detail in network dynamic may improve neural networks in engineering applications.
Title of the Paper: Five-Level DTC Based on ANN of IM Drives Using 13-Level Hysteresis Control to Reduce Torque Ripple Comparing with Conventional Control
Authors: Habib Benbouhenni
Pages: 33-39
Abstract: In this paper, the author proposes a new switching table of direct torque control (DTC) of induction machine (IM) fed by five-level Neutral Point Clamped (NPC) inverter. Using the Artificial Neural Network (ANN) applied in switching select voltage. We used the torque hysteresis by using the 13-level hysteresis controller. The proposed DTC control in this paper can reduce the torque ripple, stator flux ripple and the THD (Total Harmonic Distortion) value of stator current. The validity of the proposed DTC control scheme is verified by simulation tests of an IM drive.
Title of the Paper: Bandwidth Oriented Image Compression Using Neural Network with ASAF
Authors: Prakash Jadhav, Siddesh G. K.
Pages: 25-32
Abstract: We have applied the adaptive slope of activation function in the supervised learning algorithm for multilayer feedforward neural networks to get the bandwidth based demand of image compression in multimedia based applications. Developed solutions have applied to compress and decompress the gray scale as well as color images also. The Proposed solution has also contained a smoothing approach, which applied after decompression process to avoid the distortion at the edge of pixels blocks and it has observed that with the higher value of compression, usefulness of smoothing process is increased. The algorithm is applied successfully to classic linearly nonseparable benchmark, XOR problem, to understand the fundamental benefit of adaptive slope in activation function. The algorithm has successfully tested in the presence of different requirement of image transmission bandwidth, and results have shown that proposed solution has better compression quality along with faster convergence.
Title of the Paper: Automatic Point-of-Interest Image Cropping via Ensembled Convolutionalization
Authors: Andrea Asperti, Pietro Battilana
Pages: 17-24
Abstract: Convolutionalization of discriminative neural networks, introduced also for segmentation purposes, is a simple technique allowing to generate heat-maps relative to the location of a given object in a larger image. In this article, we apply this technique to automatically crop images at their actual point of interest, fine tuning them with the final aim to improve the quality of a dataset. The use of an ensemble of fully convolutional nets sensibly reduce the risk of overfitting, resulting in reasonably accurate croppings. The methodology has been tested on a well known dataset, particularly renowned for containing badly centered and noisy images: the Food-101 dataset, composed of 101K images spread over 101 food categories. The quality of croppings can be testified by a sensible and uniform improvement (3-5%) in the classification accuracy of classifiers, even external to the ensemble.
Title of the Paper: Training Feed-Forward Neural Networks Using Asexual Reproduction Optimization (ARO) Algorithm
Authors: Mehdi Zekriyapanah Gashti, Rouhollah Habibey
Pages: 13-16
Abstract: Artificial neural networks have been increasingly used in many problems of data classification because of their learning capacity, robustness and extendibility. Training in the neural networks accomplished by identifying the weight of neurons which is one of the main issues addressed in this field. The process of network learning by back-propagation algorithm which is based on gradient, commonly fall into a local optimum. Due to the importance of weights and neural network structure, evolutionary neural networks have been emerged to obtain suitable weight set. This paper will concentrate on training a feed-forward networks by a modified evolutionary algorithm based on asexual reproduction optimization (ARO) in order to data classification problems. The idea is to use real representation (rather the binary) for adjusting weights of the network. Experimental results show a better result in terms of speed and accuracy compared with other evolutionary algorithms including genetic algorithms, simulated annealing and particle swarm optimization.
Title of the Paper: Data Mining Methods for Traffic Accident Severity Prediction
Authors: Qasem A. Al-Radaideh, Esraa J. Daoud
Pages: 1-12
Abstract: The growth of the population volume and the number of vehicles on the road cause congestion (jam) in cities that is one of the main transportation issues. Congestion can lead to negative effects such as increasing accident risks due to the expansion in transportation systems. The smart city concept provides opportunities to handle urban problems, and also to improve the citizens’ living environment. In recent years, road traffic accidents (RTAs) have become one of the largest national health issues in the world. Many factors (driver, environment, car, etc.) are related to traffic accidents, some of those factors are more important in determining the accident severity than others. The analytical data mining solutions can significantly be employed to determine and predict such influential factors among human, vehicle and environmental factors and thus to explain RTAs severity. In this research, three classification techniques were applied: Decision trees (Random Forest, Random Tree, J48/C4.5, and CART), ANN (back-propagation), and SVM (polynomial kernel) to detect the influential environmental features of RTAs that can be used to build the prediction model. These techniques were tested using a real dataset obtained from the Department for Transport of the United Kingdom. The experimental results showed that the highest accuracy value was 80.6% using Random Forest followed by 61.4% using ANN then by 54.8% using SVM. A decision system has been build using the model generated by the Random Forest technique that will help decision makers to enhance the decision making process by predicting the severity of the accident.
