Feedforward Multilayer Perceptron Research Articles

Effect of Geometrical Factors on Magnetic Induction Distribution of Toroidal Cores Using Numerical Methods

Toroidal cores wound from grain-oriented 3 % SiFe 0.27-mm thick (M4) electrical steel are widely used in many applications. Apart from the processing factors, the geometry of the toroid has an effect on its magnetic properties. In magnetic path length of a toroid from its inside to its outside circumference, the magnetic field varies at different layers. Artificial neural network have been successfully used for the prediction of magnetic performance in electromagnetic devices. Experimental data obtained from the previous measurements have been used as a training data to a feed forward multilayer perceptron neural network for the prediction of magnetic field and flux density distribution. The input parameters were outer and inner diameters and strip thickness while the output parameter was the magnetic field and flux density. When the network was tested by untrained sample data, the average correlation of the models was found to be 99 % and the overall prediction error was in the range of 2.31 and 0.02 in Ampere per meter and Tesla. An analytical equation as depending on theoretical data and prediction results has been determined by using MATLABⓇ Curve Fitting Toolbox TM for magnetic field and flux density distribution in the toroid.

Two-stage neural network regression of eye location in face images

Automatic eye localization is a crucial part of many computer vision algorithms for processing face images. Some of the existing algorithms can be very accurate, albeit at the cost of computational complexity. In this paper, a new solution to the problem of automatic eye localization is proposed. Eye localization is posed as a nonlinear regression problem solved by two feed-forward multilayer perceptrons (MLP) working in a cascade. The input feature vector of the first network is constructed from coefficients of a two dimensional discrete cosine transform(DCT) of a face image. The second network generates corrections based on small image patches. Feature extraction and neural network prediction have known and efficient implementations, thus the entire procedure can be very fast. The paper hints at the neural network structure and the procedure for generating artificial training samples from a low number of face images. In terms of accuracy, the method is comparable to state-of-the-art techniques; however it is based on numerical procedures that could be highly optimized (fast Fourier transform and matrix multiplication).

Speaker identification based on artificial neural networks. Case study: the Polish vowel (pilot study)

Abstract Current statistical methods and technologies used for speaker identification via dynamic formant frequency often involve classic multivariate analyses that must meet a number of criteria in order to be considered trustworthy. The authors propose more advanced classification techniques, including artificial neural networks. Owing to iterative learning algorithms, neural networks can be trained to detect highly complex, nonlinear relations hidden in input data. This study specifically considers feed-forward multilayer perceptron and radial basic function network models. The investigation involves an analysis of the Polish vowel (stressed or unstressed) in selected contexts described by the four lowest formant frequencies. Results indicate high accuracy of neural networks as a speaker identification tool reaching up to 100%. In addition, the authors have determined that the accuracy of classification is similar when based on a single context to when input data are aggregated over several different contexts.

Prediction of solubility of CO2 in ethanol–[EMIM][Tf2N] ionic liquid mixtures using artificial neural networks based on genetic algorithm

This paper proposes artificial neural networks (ANNs) technique as a new approach to predict solubility of CO2 in ethanol–[EMIM][Tf2N] ionic liquid mixtures. The solubility equilibrium data of CO2 were measured at 0, 20, 50 and 80 and 100wt.% mixture of [EMIM][Tf2N] ionic liquid, temperatures of 313.2 and 333.2°C, and pressure range of 0–7MPa. A feed-forward multilayer perceptron (MLP) neural network with Levenberg-Marquardt learning algorithm was developed for prediction task. The ANN model was trained, validated and tested using 70%, 15% and 15% of all solubility data, respectively. An optimization procedure was performed based on genetic algorithm to select the best ANN architecture. Therefore, a three layer feed-forward neural network with Levenberg-Marquardt back-propagation training algorithm was developed and designed with four variables as inputs and one variable as output, 15 neurons in the hidden layer, log-sigmoid transfer function in the hidden layer and linear transfer function in the output layer. Based upon statistical analysis, results obtained demonstrated that there is a very little difference between predicted and experimental data of CO2 capture rate giving very low value of average absolute deviation (AAD) and high value of least square (R2) very close to one, indicating high accuracy of this model to predict output variable. The results also proved that the developed ANN model outperforms the Peng-Robinson model.

Intelligent Controller Design for a Blowdown Supersonic Wind Tunnel

In this paper, first, the nonlinear mathematical model of special Blowdown Supersonic Wind Tunnel (BSWT) consisting of a set of ordinary differential and algebraic equations is developed in Matlab/Simulink software environment. At the second step, an Artificial Neural Network (ANN) is used for finding the optimum membership functions of the Fuzzy Logic Controller (FLC) system. This method can help for reasonable system recognition. In this step, by designing and training a feed-forward multilayer perceptron neural network according to the available database which is generated from mathematical model; a number of different reasonable functions for Valve Opening Angle (VOA) in various test conditions are determined. These functions are used to define the desired VOA fuzzy Membership Functions (MFs). Next, a Proportional-Derivative FLC (PD-FLC) system is developed in the Simulink toolbox to control a relationship between the stagnation pressure and the temperature in the plenum chamber, which presents the Reynolds number in the test section. A synthetic algorithm combined from FLC and ANN is used to design a controller for a BSWT with the aim of achieving the accurate and acceptable desired results. Performance of the BSWT using optimized fuzzy controller by ANN is found to be satisfactory, as confirmed by the results.

Application of wavelet energy and Shannon entropy for feature extraction in gearbox fault detection under varying speed conditions

The fault detection and feature extraction of varying speed machinery with multi-component signals are full of difficulties caused by non-stationary machine dynamics and vibrations. In monitoring the vibrations of varying speed machinery, mainly formal signal processing methods based on digital sampling accomplished in equal time intervals become unsuitable. On the other hand, energy and Shannon entropy distribution of gear vibration signals measured in time–frequency plane would be different from the distribution under the normal state, when faults occur in the gear. Therefore, it is possible to detect a fault by comparing the energy and Shannon entropy distribution of gear vibration signals with and without fault conditions. In this paper, for fault diagnosis of gearbox in the run-up condition, primarily the obtained vibration signals from an acceleration sensor of automotive gearbox test setup are sampled at constant time increment by an acquisition card. To process the non-stationary vibration signals, the re-sampling technique at constant angle increment is combined with the continuous wavelet transform (CWT) and the wavelet coefficients of the signals are obtained. The Morlet wavelet is used; because impulses in many mechanical dynamic signals are always the indication of faults and the Morlet wavelet is exceedingly comparable to an impulse component. Then, statistical parameters of the wavelet coefficients are extracted that constitute the feature vectors. As a new method, the optimal range of wavelet scales is selected based on the maximum energy to Shannon entropy ratio criteria and consequently feature vectors are reduced. In addition, energy and Shannon entropy of the wavelet coefficients are used as two new features along other statistical parameters as input of the classifier. Finally, a feed-forward multilayer perceptron (MLP) neural network uses the extracted features for classification. The experimental results show that the presented method can diagnose the faults of the gear chip and wear efficiently.

Use of Artificial Neural Networks for the Development of an Inverse Kinematic Solution and Visual Identification of Singularity Zone(s)

This paper presents a non-conventional technique for solving the inverse kinematics problem using artificial neural networks. A feed forward multi-layer perceptron with backpropagation neural network is selected for this research. An inverse kinematic solution for a PUMA 560 robot is developed by training the neural network with the robot's end-effector Cartesian co-ordinates and its corresponding joint configurations. Once the network is well trained (90th percentile) and confident predictions can be achieved, a test input set (singularity conditions) is introduced to the trained network to simulate results. This technique proves promising since it requires little computation time over other traditional methods.

Investigation on Fluidized Bed Bioreactor Treating Ice Cream Wastewater Using Response Surface Methodology and Artificial Neural Network

Abstract In this study, a comparative approach was developed between response surface methodology (RSM) and artificial neural network (ANN) in the predictive capabilities for the removal of chemical oxygen demand (COD) from ice cream industry wastewater using fluidized bed bioreactor. The effects of process variables such as pH, temperature, flow rate and agitation speed investigated using a four-factor three-level Box–Behnken experimental design (BBD). Same design was utilized to train a feed-forward multilayered perceptron (MLP) ANN with back-propagation algorithm. The predictive capabilities of the two methodologies were compared in terms of statistical parameters including coefficient of determination (R2). The results showed that properly trained ANN model is more accurate in prediction as compared to RSM model. Under the optimum conditions (pH of 7, temperature of 40°C, flow rate of 20 ml/min and agitation speed of 175 rpm), 91% of COD was removed.

Neuro-controllers for Synchronous Generators

Automatic voltage regulators (AVRs) are controllers used to maintain constant voltage at the generator terminals. Artificial Neural Networks (ANNs) are nonlinear maps that have the potential to make the realisation of practical nonlinear controllers possible. This paper is concerned with the development of a Feedforward Multilayer Perceptron (MLP) Neural Networks and its use as an Automatic Voltage Regulator (AVR) with Power System Stabiliser (PSS). The performance of the MLP-AVR is compared with that of a conventional AVR with PSS. The MLP-AVR shows good performance compared to that of conventional AVR with PSS.

Prediction of Physicochemical Properties of Raspberry Dried by Microwave-Assisted Fluidized Bed Dryer Using Artificial Neural Network

The main target of this research is to dry raspberries in a microwave-assisted fluidized bed dryer. Artificial neural network (ANN) modeling was used in order to evaluate and predict the physicochemical properties of this fruit. In this research, the effects of five variables—microwave power (0, 300, and 600 W), temperature (55, 70, and 85°C), air flow rate (15, 20, and 25 m/s), starting time of microwave input (from the moment when the moisture content decreased until 334, 400, and 466 g water/g dry matter), and amount of loaded material (50, 100, and 150 g)—on nine outputs—drying time, rehydration capacity, density, porosity, hardness, water activity, phenolic compounds content, anthocyanins content, and the antioxidant activity of dried raspberries—were studied. A feed-forward multilayered perceptron trained by back-propagation algorithms for five independent variables was developed to predict these nine outputs. The optimal configuration of the neural network model had a hidden layer with nine neurons. The predictive ability of the ANN was compared using a separate data set of 52 unseen experiments based on root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2) for each output parameter. The optimum model was able to predict the nine output parameters with a coefficient of determination higher than 0.92. The results indicated that the experimental and ANN-predicted data sets were in good agreement, so it is feasible to use an ANN to predict the physicochemical properties of dried black raspberry.

Forensic voice comparison by means of artificial neural networks

Abstract This article examines the effectiveness of artificial neural networks (ANNs) as forensic voice comparison techniques. This study specifically considers feed-forward multilayer perceptron (MLP) and radial basic function (RBF) network models. Formant frequencies of Polish vowel e (stressed or unstressed) in selected contexts were used as predictors. This has already been confirmed in an earlier investigation that determined that dynamic formant frequencies of vowels are powerful elements in distinguishing the voice. It has been concluded that neural networks might assist in distinguishing speakers from the others with very good accuracy, reaching 100%. MLP models should be given preference. The results of the investigation have shown the influence of vowel e triads on the effectiveness of correct classification rates. In addition, the authors have determined that the accuracy of classification is greater when based on a single context than for similar input data aggregated over several different contexts.

Modeling and optimization of a trench layer location around a pipeline using artificial neural networks and particle swarm optimization algorithm

The main objective of the present work is to utilize particle swarm optimization algorithm (PSOA) integrated with feed-forward multi-layer perceptron (MLP) type of artificial neural networks (ANN) to find the optimum positions of a trench layer around a pipeline in order to obtain the minimum liquefaction potential. The mesh free local radial basis function differential quadrature method (LRBF-DQ) was used to solve the governing equations of seismic accumulative excess pore pressure containing pore pressure source term. This data was used to train the ANN using back propagation weight update rule. Then the trained ANN predicts the liquefaction potential and PSOA was used to find the best location of the trench layer. The results obtained by the MATLAB codes of LRBF-DQ, ANN and PSOA are showed that there was a linear relation between the location of the pipeline and the optimum location of the trench layer. Moreover the minimum liquefaction potential has been occurred when the trench layer placed beneath of the pipeline.

Enhancing the Performance of Jet Engine Fuel Controller Using Neural Networks

This paper proposes a fuzzy logic controller for a specific turbojet engine. The turbine engines require control systems to achieve the appropriate performance. The control systems typically featured loops to prevent engine flame out, over speeds, compressor surge, and check turbine inlet temperature limit, either by scheduling the fuel flow during accelerations and decelerations or by controlling the acceleration and deceleration rates of engine spool. This paper presents a successful approach in designing a Fuzzy Logic Controller for a specific Jet Engine. At first a suitable mathematical model for the jet engine is presented by the aid of SIMULINK simulation software. Then by applying different reasonable fuel flow functions via the engine model, some important engine continuous time operation parameters (such as: thrust, compressor surge margin, turbine inlet temperature and engine spool speed...) are obtained. These parameters provide a precious database which can be used by a neural network. At the second step, by designing and training a feedforward multilayer perceptron neural network according to this available database; a number of different reasonable fuel flow functions for various engine acceleration operations are determined. These functions are used to define the desired fuzzy fuel functions. Indeed, the neural networks are used as an effective method to define the optimum fuzzy fuel functions. At the next step we design a fuzzy logic controller by using the engine simulation model and the neural network results. The proposed control scheme is proved by computer simulation using the designed engine model. The simulation results of engine model with fuzzy controller in comparison with the engine testing operation illustrate that the proposed controller achieves the desired performance.

Artificial Neural Network–Based System Identification for a Single-Shaft Gas Turbine

During recent decades, artificial intelligence has been employed as a powerful tool for identification of complex industrial systems with nonlinear dynamics, such as gas turbines (GT). In this study, a methodology based on artificial neural network (ANN) techniques was developed for offline system identification of a low-power gas turbine. The processed data was obtained from a SIMULINK model of a gas turbine in matlab environment. A comprehensive computer program code was generated and run in matlab for creating and training different ANN models with feed-forward multilayer perceptron (MLP) structure. The code consisted of various training functions, different number of neurons as well as a variety of transfer (activation) functions for hidden and output layers of the network. It was shown that the optimal model for a two-layer network with MLP structure consisted of 20 neurons in its hidden layer and used trainlm as its training function, as well as tansig and logsid as its transfer functions for the hidden and output layers. It was also observed that trainlm has a superior performance in terms of minimum mean squared error (MSE) compared with each of the other training functions. The resulting model could predict performance of the system with high accuracy. The methodology provides a comprehensive view of the performance of over 18,720 ANN models for system identification of the single-shaft gas turbine. One can use the optimal ANN model from this study when training from real data obtained from this type of GT. This is particularly useful when real data is only available over a limited operational range.

Artificial neural network and response surface methodology modeling in mass transfer parameters predictions during osmotic dehydration of Carica papaya L.

In this study, a comparative approach was made between artificial neural network (ANN) and response surface methodology (RSM) to predict the mass transfer parameters of osmotic dehydration of papaya. The effects of process variables such as temperature, osmotic solution concentration and agitation speed on water loss, weight reduction, and solid gain during osmotic dehydration were investigated using a three-level three-factor Box-Behnken experimental design. Same design was utilized to train a feed-forward multilayered perceptron (MLP) ANN with back-propagation algorithm. The predictive capabilities of the two methodologies were compared in terms of root mean square error (RMSE), mean absolute error (MAE), standard error of prediction (SEP), model predictive error (MPE), chi square statistic (χ2), and coefficient of determination (R2) based on the validation data set. The results showed that properly trained ANN model is found to be more accurate in prediction as compared to RSM model.

Use of artificial neural networks for prognosis of charcoal prices in Minas Gerais

Energy is an important factor of economic growth and is critical to the stability of a nation. Charcoal is a renewable energy resource and is a fundamental input to the development of the Brazilian forest-based industry. The objective of this study is to provide a prognosis of the charcoal price series for the year 2007 by using Artificial Neural Networks. A feedforward multilayer perceptron ANN was used, the results of which are close to reality. The main findings are that: real prices of charcoal dropped between 1975 and 2000 and rose from the early 21st century; the ANN with two hidden layers was the architecture making the best prediction; the most effective learning rate was 0.99 and 600 cycles, representing the most satisfactory and accurate ANN training. Prediction using ANN was found to be more accurate when compared by the mean squared error to other studies modeling charcoal price series in Minas Gerais state.

Artificial Neural Network Modeling of Plastic Viscosity, Yield Point, and Apparent Viscosity for Water-Based Drilling Fluids

A feed-forward multilayer perceptron neural network model with Levenberg-Marquardt learning algorithm has been developed to predict rheological properties of water-based drilling fluids. The developed model was based on a three-layer network containing input, hidden, and output layer. The sigmoid function was applied as the transfer function in the hidden layer and linear transfer function in the output layer. Two statistical parameters, mean square errors and correlation coefficient, were used as a criterion for evaluating artificial neural network modeling performance. Accordingly, a comparison between experimental values and those predicted by the artificial neural network showed a good coincidence proving its high accuracy in estimating target value.

DIMENSIONALITY REDUCTION FOR SENSORY DATASETS BASED ON MASTER–SLAVE SYNCHRONIZATION OF LORENZ SYSTEM

In this study, we propose a novel method for reducing the attributes of sensory datasets using Master–Slave Synchronization of chaotic Lorenz Systems (DPSMS). As part of the performance testing, three benchmark datasets and one Electronic Nose (EN) sensory dataset with 3 to 13 attributes were presented to our algorithm to be projected into two attributes. The DPSMS-processed datasets were then used as input vector to four artificial intelligence classifiers, namely Feed-Forward Artificial Neural Networks (FFANN), Multilayer Perceptron (MLP), Decision Tree (DT) and K-Nearest Neighbor (KNN). The performance of the classifiers was then evaluated using the original and reduced datasets. Classification rate of 94.5%, 89%, 94.5% and 82% were achieved when reduced Fishers iris, crab gender, breast cancer and electronic nose test datasets were presented to the above classifiers.

A Prediction Model for Population Dynamics of Cotton Pest (Thrips tabaci Linde) using Multilayer-Perceptron Neural Network

agricultural sector faces a series of problems when it comes to increasing crop productivity. Despite the efforts of researchers to discover productive agricultural practices, crop yield has not been the most pleasing, and one global reason stated for this poor crop yield is the insect pests. Predicting in advance the occurrence of peak activities of a given pest could enable the development of a suitable pest control mechanism that would initiate better production. Researchers have attempted to comprehend the pest population dynamics by applying analytical and other techniques on pest surveillance data sets. In this paper, An intelligent system for effectual prediction of pest population dynamics of Thrips Tabaci Linde (Thrips) on cotton (Gossypium Arboreum) crop is presented. The raw data used in the proposed system was obtained from the College of Agriculture, Raichur, India. Initially, the raw (pest surveillance) data is prepared by 1) Data preprocessing 2) Normalization and 3) Data transformation. The feed forward Multi-Layer Perceptron (MLP) Neural Network with backpropagation training algorithm is employed in the design of the intelligent system. The neural network is trained and tested with the data prepared. The experimental results portray the effectiveness of the proposed system in predicting pest population dynamics of Thrips on cotton crop. Moreover, a comparative analysis is performed between the proposed system and two of the existing works. The results showed that the proposed system based on feed forward neural networks was best suited for effective pest prediction.

Artificial Neural Network Modeling of Deposition Thickness Through the Pipeline

Modeling and prediction of deposition thickness is a key to developing crude oil transportation process. Usually, the heat transfer method has been used to calculate deposition thickness through the pipeline, but this method requires vast information and cumbersome calculations. To eliminate this problem, a feed-forward multilayer perceptron neural network (MLPNN) algorithm has been applied to predict the deposition thickness through the pipeline. We present an optimized one-layer feed-forward neural network using properties of the oil pipeline such as inlet and outlet oil temperatures, environmental (mixture of water and ethylene glycol) temperature, and oil Reynolds’ numbers. Different networks are considered and trained using 15313 data sets; the accuracy of the proposed network is validated by 5104 testing data sets. Using validating data set, the network that is having the lowest mean square error (MSE) and average absolute relative deviation percent (AARD%) and the highest regression coefficient (R2) is selected as an optimal configuration. Statistical analysis show that the artificial neural network (ANN) predictions have an excellent agreement (MSE = 9.4949e-006, AARD% = 1.0109 and R2 = 0.9999) with the heat transfer method. In the other words, accuracy of the proposed ANN model is approximately equal to the heat transfer method without required too much information and cumbersome calculations.