Multilayer Feedforward Network Research Articles

Use of artificial neural network to evaluate the vibration mitigation performance of geofoam-filled trenches

Development activities in a city often generate ground vibration that can cause discomfort to the occupants in nearby buildings, disturbances to the activities undertaken in the buildings and possible damage to nearby structures. This ground vibration is caused by construction activities such as pile driving, ground compaction etc., and road and rail traffic. The use of trenches has been an effective way to mitigate the adverse effects of such ground vibration. The effectiveness of the trench depends on many parameters including the properties of the vibration source, soil medium and trench in-fill material, trench dimensions and the requirements of the receiver. The process of selecting an effective trench for vibration mitigation can therefore become complex due to the influence of a number of parameters and their wide range of values. This paper investigates the use of artificial neural network (ANN) as a smart and efficient tool to predict the effectiveness of geofoam-filled trenches to mitigate ground vibration. Towards this end, a database is developed from an extensive study on the effects of the controlling parameters through numerical simulations with a validated finite element (FE) model. At a certain distance from the vibration source, a geofoam-filled trench is introduced to evaluate the efficiency of vibration mitigation with changes in key parameters such as excitation frequency, amplitude of load, trench configuration (i.e. depth and width), soil shear wave velocity, soil density and damping ratio. These were selected as the input parameters for the ANN while amplitude reduction ratio and peak particle velocity (PPV) were considered as outputs. A multilayer feed forward network was used and trained with the Levenberg-Marquardt algorithm. Neural networks with different configurations were evaluated by comparing coefficient of determination (R2) and mean square error (MSE). The optimum architecture was then used to predict previous results, which revealed the accuracy and the effectiveness of the ANN approach. The findings of this study will provide useful information for vibration mitigation using geofoam-filed trenches.

Growing Co-doped TiO2 nanosheets on reduced graphene oxide for efficient photocatalytic removal of tetracycline antibiotic from aqueous solution and modeling the process by artificial neural network

A one-pot hydrothermal synthesis was applied for growth of cobalt-doped TiO2 nanosheets (Co-TNs) having different quantities of cobalt on the reduced graphene oxide surfaces (Co-TNs/rGO (x)). The synthesized nanocomposites were characterized by a range of analyses including XRD, UV–Vis DRS, FESEM/EDX, elemental mapping, TEM, HRTEM and Raman spectroscopy. The visible light degradation of Tetracycline antibiotic (TC) by synthesized samples was investigated and the degradation percentage of TC was 60% by Co-TNs/rGO (0.152) (the optimal sample). Reusing the optimal photocatalyst after five successive cycles showed ∼7% decline in its activity for degrading of TC. Active species trapping experiments showed that OH radicals and h+ are the main active species in the degradation process. An artificial neural network (ANN) model was used to predict the photocatalytic removal of tetracycline antibiotic. The multilayered feed forward networks were trained by using a backpropagation algorithm; a three-layer network with 14 neurons in the hidden layer gave the optimal results. The relative importance of different parameters on the photoactivity of the as-obtained photocatalysts were evaluated.

Forecasting coking coal prices by means of ARIMA models and neural networks, considering the transgenic time series theory

Price forecasting is a vital matter for mining investment decisions, as it represents the credibility of any financial outcome claimed by the feasibility studies presented to investors in financial markets. Most of these financial studies use forecasts from well-known providers of price assessments and market data that, ultimately, constitute a black box for the investors. This is why, to achieve credibility, user-friendly forecasting techniques through which the future price instability can be bounded are needed.This paper examines the performance in forecasting coking coal prices of traditional time series models, ROBUST models, ARIMA models, generalized regression neural networks (GRNNs) and multi-layer feedforward networks (MLFNs). Two cases are analysed: (1) considering the full time series, and (2) considering the transgenic time series. The latter is based on a theory that addresses the existence of raw material price cycles and the presence of anomalous phenomena due to periods with spiked prices, and it eliminates these periods from the time series to improve the accuracy of the forecasts.At this point, the question of which forecasting performance measure provides a better projection of the errors becomes critical. Keeping in mind that we are focusing on commodity prices and that these results are going to be used for estimating incomes, percentual deviations represent a greater compromise for the forecasts. Among the RMSE, MAE and MAPE, the mean absolute percentage error (MAPE) is the most representative since it is the only one that considers percentual deviations. On the other hand, since the error does not follow a normal distribution, the MAE can be considered to be the second-most representative performance measure, and the RMSE is third.When the full time series is used for the forecast, the GRNN model gives more accurate forecasts than the MLFN model, the ARIMA model, the traditional time series models and the ROBUST model. On the other hand, when the transgenic time series is used for the forecast, the performance of the GRNN model is improved, but, in this case, the ARIMA model achieves a more accurate forecast.

Laminar burning velocity of hydrogen and carbon-monoxide enriched natural gas (HyCONG): An experimental and artificial neural network study

The laminar burning velocity of H2 and CO enriched natural gas (HyCONG) mixtures measured through spherically expanding flames method under three different fuel-air equivalence ratios (0.6, 0.8, and 1.0) at room temperature and atmospheric pressure. Experiments have been conducted into constant volume combustion chamber by using high-speed schlieren photography technique. The five different CO fraction (0%, 20%, 40%, 60%, and 80%) and three HCNG (0%, 20%, and 40%) blends combinations are selected for the experiments. The laminar burning velocity and Markstein length were obtained for 45 operational conditions. The results reveal that the laminar burning velocity increases drastically with an increase of hydrogen and carbon monoxide in HyCONG blends and increasing fuel-air equivalence ratio. The maximum laminar burning velocities are determined as 24.42 cm/s, 45.7 cm/s, and 62.78 cm/s for 20–80 HyCONG at ϕ = 0.6, 40–60 HyCONG at ϕ = 0.8, and 40–60 HyCONG at ϕ = 1.0 respectively. Additionally, the artificial neural network model (ANN) of combustion chamber has been constructed to predict the laminar burning velocity of HyCONG blends. The ANN’s well known backpropagation algorithm is applied in multilayered feedforward networks. The forecasting of laminar burning velocity has been evaluated using four-input and one-output network structure with two different learning algorithm namely: Levenberg-Marquardt (LM) algorithm and Scale Conjugate Gradient (SCG) algorithm. The four-input parameters considered as; fuel-air equivalence ratio, hydrogen fraction, methane fraction, and carbon monoxide fraction while output parameter is laminar burning velocity. The model training has been accomplished using hyperbolic tangent transfer activation function (tansig) and various number of neurons.The maximum and minimum value of correlation coefficient (R) and mean square error (MSE) for validation were obtained for laminar burning velocity. The values of R (max-min) and MSE (max-min) are 0.9987–0.9944 & 2.1305–1.0195 for Levenberg-Marquardt (LM) algorithm and 0.9990–0.9952 & 2.7965–0.9501 for Scale Conjugate Gradient (SCG) algorithm. Presented article not only provides the experimental study of LBV as well as explores the possibility of artificial neural network application in combustion research area.

Deep neural networks for human pose estimation from a very low resolution depth image

The work presented in the paper is dedicated to determining and evaluating the most efficient neural network architecture applied as a multiple regression network localizing human body joints in 3D space based on a single low resolution depth image. The main challenge was to deal with a noisy and coarse representation of the human body, as observed by a depth sensor from a large distance, and to achieve high localization precision. The regression network was expected to reason about relations of body parts based on depth image, and to extract locations of joints, and provide coordinates defining the body pose. The method involved creation of a dataset with 200,000 realistic depth images of a 3D body model, then training and testing numerous architectures including feedforward multilayer perceptron network and deep convolutional neural networks. The results of training and evaluation are included and discussed. The most accurate DNN network was further trained and evaluated on an augmented depth images dataset. The achieved accuracy was similar to a reference Kinect algorithm results, with a great benefit of fast processing speed and significantly lower requirements on sensor resolution, as it used 100 times less pixels than Kinect depth sensor. The method was robust against sensor noise, allowing imprecision of depth measurements. Finally, our results were compared with VGG, MobileNet, and ResNet architectures.

Generalized space time autoregressive (gstar)-artificial neural network (ann) model with multilayer feedforward networks architecture

The GSTAR model is one linear space time model used to model time series data with inter-location linkages. One of the weaknesses of the GSTAR model is that the model has not been able to capture any nonlinear patterns that may arise. The ANN is one model of nonlinear artificial intelligence that has a flexible functional form and is a supervised engine learning that provides a good framework to represent a relationship in time series data. Due to the advantages of such ANN, it can be combined with GSTAR model. In this research conducted study of GSTAR-ANN model and its network architecture. The model is constructed in 3 layers namely input, hidden and output. The GSTAR model is used as input in the training process. Each input will receive an input signal and forward the signal to the hidden layer, then to the output layer. The GSTAR-ANN model has network architecture with one neuron unit at the output layer, p input neuron, q neuron in hidden layer (multilayer feedforward networks).

Hippocampal Neurogenesis Reduces the Dimensionality of Sparsely Coded Representations to Enhance Memory Encoding.

Adult neurogenesis in the hippocampal dentate gyrus (DG) of mammals is known to contribute to memory encoding in many tasks. The DG also exhibits exceptionally sparse activity compared to other systems, however, whether sparseness and neurogenesis interact during memory encoding remains elusive. We implement a novel learning rule consistent with experimental findings of competition among adult-born neurons in a supervised multilayer feedforward network trained to discriminate between contexts. From this rule, the DG population partitions into neuronal ensembles each of which is biased to represent one of the contexts. This corresponds to a low dimensional representation of the contexts, whereby the fastest dimensionality reduction is achieved in sparse models. We then modify the rule, showing that equivalent representations and performance are achieved when neurons compete for synaptic stability rather than neuronal survival. Our results suggest that competition for stability in sparse models is well-suited to developing ensembles of what may be called memory engram cells.

A top-down approach for semantic segmentation of big remote sensing images

The increasing amount of remote sensing data has opened the door to new challenging research topics. Nowadays, significant efforts are devoted to pixel and object based classification in case of massive data. This paper addresses the problem of semantic segmentation of big remote sensing images. To do this, we proposed a top-down approach based on two main steps. The first step aims to compute features at the object-level. These features constitute the input of a multi-layer feed-forward network to generate a structure for classifying remote sensing objects. The goal of the second step is to use this structure to label every pixel in new images. Several experiments are conducted based on real datasets and results show good classification accuracy of the proposed approach. In addition, the comparison with existing classification techniques proves the effectiveness of the proposed approach especially for big remote sensing data.

Comparison of Neural Network–Based Adaptive Controllers Using Hypercomplex Numbers for Controlling Robot Manipulator

This study investigates an adaptive controller by applying a neural network, in which all the network parameters, states, signals and functions are expressed using hypercomplex numbers and algebras; its application to dynamics control of a robot manipulator. To design hypercomplex–valued neural networks where each neural network is a multilayer feedforward network with a split–type activation function of neurons using a tapped–delay–line input, we consider the following four types of hypercomplex numbers: complex, hyperbolic, bicomplex and quaternion numbers. In the control system, we utilise a feedback error–learning scheme to conduct the training of the network through a back–propagation algorithm. In the computational experiments, we explore a hypercomplex–valued neural network–based controller as a trajectory control problem of a three–link robot manipulator, in which the position of the end–effector follows to the desired trajectory in a 3–dimensional space. The simulation results validate the feasibility and effectiveness of the quaternion neural network–based controller for this task.

Multilayer Feed Forward Models in Groundwater Level Forecasting Using Meteorological Data in Public Management

Managing the groundwater resources is very vital for human life. This research proposes a methodology for predicting the groundwater levels which can be very valuable in water resources management. This study investigates the application of multilayer feed forward network models for forecasting the groundwater values in the region of Montgomery country in Pennsylvania. Multiple training algorithms and network structures were investigated to develop the best model in order to forecast the groundwater levels. Several multilayer feed forward models were created in order to be tested for their performance by changing the network topology parameters so as to find the optimal prediction model. The forecasting models were developed by applying different structures regarding the number of the neurons in every hidden layer and the number of the hidden network layers. The final results have shown a very good forecasting accuracy of the predicted groundwater levels. This research can be very valuable in water resources and environmental management.

Artificial neural network approach for air brake pushrod stroke prediction in heavy commercial road vehicles

In heavy commercial road vehicles, the air brake system is a critical vehicle safety system whose performance degradation increases the risk of accidents and hence requires periodic inspection and maintenance. The wear of brake pad lining and brake drum during operation leads to increase in the stroke of a component called pushrod whose ‘out-of-adjustment’ creates severe brake performance degradation. The fact that the driver does not receive a corresponding tactile feedback till it is too severe adds to the complexity of manual detection. Motivated by the increase in onboard sensing, electronics, and computation capabilities, this study proposes an artificial neural network–based approach to predict pushrod stroke based on measurement of brake chamber pressure. Here, a back propagation algorithm was used to train the multilayer feed-forward network. The effect of excessive pushrod stroke on vehicle braking response was first studied using a Hardware-in-Loop system that consists of brake system hardware and a commercial vehicle dynamics simulation software (IPG TruckMaker®). Experimental data collected from this system with manual slack adjuster and automatic slack adjuster have then been used to train and test the artificial neural network for pushrod stroke prediction. The performance of the prediction scheme has been tested over the entire range of brake operating conditions. The prediction error corresponding to manual slack adjuster was found to be within ±15% in 322 out of the entire test set of 328 instances (98.17%) and automatic slack adjuster within ±8% in all 57 test sets (100%). Statistical analysis based on confidence interval revealed a prediction error between −1.62% and −3.05% for manual slack adjuster and 0.43% and −1.62% for automatic slack adjuster for 99% confidence interval, which demonstrated the efficacy of the proposed prediction scheme.

Experimental and artificial neural network (ANN) study of hydrogen enriched compressed natural gas (HCNG) engine under various ignition timings and excess air ratios

In today's era, the computational capabilities of artificial neural network has endorsed to be a bedrock and inferences in many fields, including internal-combustion engines. The presented research in ANN has been germinated to anticipate the performance and emission characteristics of a turbocharged SI engine fueled with various HCNG mixtures. The experiments were accomplished at various excess air ratios (λ), ignition timings (θi) at MAP of 105 kPa and 140 kPa, while engine speed was kept constant at 1600 rpm to obtain data for testing and training ANN model. The test results show that with the increase in values of λ, MAP, and hydrogen addition, the torque output effectively decreases while BSFC first decreases and after attaining minimum value it further increases. The NOx, CO, THC, and CH4 emissions all declined with the hike of ignition advance angle, and inclined with increase of the load. ANN’s popular backpropagation algorithm is adopted in multilayered feedforward networks. In order to predict the performance and emission characteristics of HCNG engine, the four-input and one-output network structure are used. HCNG0, HCNG20 and HCNG40 blends has been studied in the presented ANN model in which the excess air ratio (λ), engine load, ignition timing, and HCNG blends has been taken as four-input parameters. The bestowed model has been trained using hyperbolic tangent transfer activation function (tansig) and Levenberg-Marquardt learning algorithm (LM) along with numerous neurons.The lowest and highest value of correlation coefficient (R) and mean square errors (MSE) for validation were dig-out for various parameters, which were BSFC, torque, NOx, CO, THC, CH4. The values of R (min–max) and MSE (min–max) were BSFC; 09519–0.9963 and 5.72–64.29, torque; 0.9948–0.9997 and 8.3935–95.8350, NOx; 0.9938–0.9992 and 0.5271–3.4340, CO; 0.8595–0.9960 and 0.0188–1.0154, and THC; 0.9644–0.9972 and 0.0714–1.1695, CH4; 0.7965–0.9973, 0.0351–1.1789 respectively for various number of neurons. This research work provides the gist of ANN as an attractive choice to researchers for classical modelling techniques. In this regard, the ANN modeling can be utilized to minutely predict the performance and emission characteristics of hydrogen added CNG engines.

Hydraulic flow units' estimation from seismic data using artificial intelligence systems, an example from a gas reservoir in the Persian Gulf

In recent years, considering the reservoir pressure drop in productive wells, designing the optimal well trajectory for production and injection in enhanced oil recovery (EOR) plans requires to determine hydraulic flow unit (HFU) in the reservoir. HFUs can also be used for petrophysical zonation of reservoirs as well as permeability predictions in uncored intervals or zones with low quality core data of wells. In the present study we have tried to integrate 3D seismic data with well data in order to find a quantitative relationship between flow zone index (FZI) and seismic attributes through employing linear regression and artificial intelligence models. Using this approach, FZI can be predicted using the information which is propagated in the whole field and achievable in the early stage of field development and consequently a suitable HFUs model may be represented. To this end, the suitable attributes for FZI estimation were selected by stepwise linear regression from extracted acoustic impedance (AI) and sample based attributes. Afterward, three optimal intelligent systems including probabilistic neural network (PNN), multi-layer feed forward network (MLFN), and radial basis function networks (RBFN) were employed. The obtained results reveal that PNN is the most accurate estimator compared to MLFN, RBFN, and multi-attribute regression methods. In the final stage, PNN was applied to develop 3D hydraulic flow unit model for the reservoir section of the investigated carbonate gas field located in Persian Gulf.

Desain Kontrol Sistem Eksitasi Steam Turbine Generator (Stg) Pada Combine Cycle Power Plant 120 Mw Pt.Krakatau Daya Istrik Menggunakan Jaringan Syaraf Tiruan Backpropagation

Excitation System is a vital part of generating electricity in the generator, the function of the excitation system in the generator is to control the output of the generator to remain stable at the changing system load. The excitation control system of the synchronous generator is one of the steps to improve the stability of the power system and to ensure the quality of the electricity generated by the network. In this research used one of the implementation of artificial intelligence technology is Artificial Neural Network designed as controller excitation system generator. The network is designed using a backpropagation learning algorithm, and the Lavenberg-Marquardt training algorithm with a feedforward multilayer network architecture. The result of simulated test and training using Matlab application obtained with presentation of system success equal to 96,387%.

A Surrogate Modeling for Storm Surge Prediction Using an Artificial Neural Network

Kim, S.-W.; Lee, A., and Mun, J., 2018. A Surrogate Modeling for Storm Surge Prediction Using an Artificial Neural Network. In: Shim, J.-S.; Chun, I., and Lim, H.S. (eds.), Proceedings from the International Coastal Symposium (ICS) 2018 (Busan, Republic of Korea). Journal of Coastal Research, Special Issue No. 85, pp. 866–870. Coconut Creek (Florida), ISSN 0749-0208.A surrogate model for storm surge prediction is developed using on an artificial neural network with the measured tidal level in Korea peninsula. The 59 historical storms during 1978 to 2014 years are used in this modelling. Tidal data recorded for 15 years was applied. The neural network between seven input parameters (i.e., latitude, longitude, moving speed, heading direction, central pressure, radius of strong wind speed, maximum wind speed) and the storm surge is trained by Levenberg-Marquardt backpropagation algorithm. The type of network is a multilayer feedforward network. The data is divided by 70% for training, 15 % for validation and 15 % for test. The six save points in southern Korea are analysed by the surrogate model. The performance of the storm surge surrogate model is expressed as the correlation coefficient and mean square error at the six save points. The minimum and maximum correlation coefficients are respectively 0.861 and 0.979. The developed surrogate model satisfies high-accuracy and high-speed for predicting he storm surge based on an artificial intelligence method and a grid-free system.

Sycamore failure hazard classification model (SFHCM): an environmental decision support system (EDSS) in urban green spaces

Although the pruning of trees is known as the one of the principal domains of green space management, it is includes shortcomings in practical models or methodologies to classify or prioritize hazardous trees. This paper aims to formulate physical and biological and associated tree factors with sycamore failure probability which results to classify tree failure hazard using artificial neural network (ANN) modeling, as an environmental decision support system (EDSS). Considering the outputs of modeling process, multi-layer feed-forward network (MLFN), which is used to train sycamore (Platanus orientalis) failure hazard classification model (SFHCM), indicates that the performance of SFHCM is statistically accurate. Changes in physical and biological tree factors with the maximum priority in sensitivity analysis assist decision makers to increase SFHCM class and decrease tree failure hazard. Finally a soundly applicable EDSS with new tool is designed for green space managers to prioritize criteria in sycamore failure hazard in order to plan for the most appropriate mitigation plan.

Completeness Problem of the Deep Neural Networks

Hornik, Stinchcombe & White have shown that the multilayer feed forward networks with enough hidden layers are universal approximators. Roux & Bengio have proved that adding hidden units yield a strictly improved modeling power, and Restricted Boltzmann Machines (RBM) are universal approximators of discrete distributions. In this paper, we provide yet another proof. The advantage of this new proof is that it will lead to several new learning algorithms. We prove that the Deep Neural Networks implement an expansion and the expansion is complete. First, we briefly review the basic Boltzmann Machine and that the invariant distributions of the Boltzmann Machine generate Markov chains. We then review the θ-transformation and its completeness, i.e. any function can be expanded by θ-transformation. We further review ABM (Attrasoft Boltzmann Machine). The invariant distribution of the ABM is a θ-transformation; therefore, an ABM can simulate any distribution. We discuss how to convert an ABM into a Deep Neural Network. Finally, by establishing the equivalence between an ABM and the Deep Neural Network, we prove that the Deep Neural Network is complete.

Resilient Back-Propagation Algorithm in the Prediction of Mother to Child Transmission of HIV

Prediction of a child HIV status poses real challenges in medical research. Even though there are different statistical techniques and machine learning algorithms that have been used to predict models like HIV for the clinical data with binary outcome variables, yet neural network techniques are major participants for prediction purposes. HIV is the primary cause of mortality among women of reproductive age globally and is a key contributor to maternal, infant and child morbidity and mortality. In this paper, resilient back propagation algorithm is used for training the Neural Network and Multilayer Feed forward network to predict the mother to child transmission of HIV status.

Outdoor Path Loss Predictions Based on Extreme Learning Machine

In a typical outdoor environment, the propagation of radio waves is usually random in nature, to the extent that the characterization of the wireless channel often becomes very difficult. Several models have been developed to predict the average Received Signal Strength (RSS) for specified distance ranges. However, the use of deterministic models requires high computational efficiency while the prediction results of empirical models may not be as accurate as required. On machine learning approach, the performances of multi-layered feed-forward network models are limited by slow convergence and local minimum, such that a global optimal solution is not guaranteed. In this paper, Extreme Learning Machine (ELM) algorithm is considered in the development of an optimal path loss prediction model for outdoor propagation scenario. Single Hidden Layer Feed-forward Neural Networks (SHLFNNs) are trained and tested with the path loss data that were computed based on the RSS data of a commercial 1800 MHz base station located along Lagos-Badagry highway in Nigeria. The training speed, learning effectiveness, and the generalization ability of Artificial Neural Network Back-Propagation (ANN-BP) and ELM algorithms are analysed. Experimental results show that ELM models are 140 times faster to train than the ANN-BP models. On prediction accuracy, the outputs of ELM, ANN-BP, Okumura–Hata, and COST-231 models have Root Mean Squared Error (RMSE) values of 2.896, 2.449, 7.456, and 6.116 dB respectively; and regression coefficient (R) values of 0.959, 0.973, 0.935, and 0.935 respectively, when compared to the target variable of the training dataset. When the models were tested with new input data that were excluded from the training process, RMSE values of 4.250, 6.622, 8.732, and 7.087 respectively; and R values of 0.893, 0.876, 0.904, and 0.904 respectively are obtained. In conclusion, the findings of this study confirm that ELM algorithm guarantees an optimal path loss model with fast training convergence, high prediction accuracy, and good generalization ability for radio network planning and optimization in outdoor environments.

Prediction of Cooling Load for a Standing Wave Thermoacoustic Refrigerator through Artificial Neural Network Technique

This study involves the application of artificial neural network (ANN) as a new approach to predict cooling load of thermoacoustic refrigerator under some operating conditions. One ANN model for a standing wave thermoacoustic refrigerator had been developed based on the experimental data from other literature. Cooling load was chosen as a response to input parameters, mean pressure and frequency in ANN model. A multi-layer feed-forward network with a back propagation algorithm had been proposed for predicting cooling load of the thermoacoustic refrigerator. This proposed ANN model has three layers with the configuration of 2-10-1, namely, input layer with two neurons representing the two operating parameters, one hidden layer with an optimal 10 hidden neurons, and output layer with one neuron representing the cooling load. The ANN model had been proven to be desirable in accuracy for predicting the cooling load by comparing model results with experimental results in literature work. This research work would provide a new modelling approach based on ANN technique for solving complex thermoacoustic problems with linear or nonlinear nature.