Four-layer Feedforward Neural Network Research Articles

Neural network-based method for determining vessel position by seabed relief

A seabed relief-based vessel position fixing model on the basis of a four-layered feedforward neural network is proposed. Hidden neurons have hyperbolic tangent activation functions. The model is constructed for 1-D case that can be considered as vessel motion throw a narrow channel or alongside fairway axis. A sequence of spot soundings is given for the network input. The linear coordinate registered for the last sounding forms the network output. The training set is formed by means of the intentional pseudorandom alteration of input samples in accordance with suspected limits of sea level variations and the constant error of its measurements. The validation set is not used. The Adamax algorithm is implemented for the neural network training. The maximum of absolute value of the prediction error is used as a performance criterion of the net. Modeling has been conducted with the Python programming language. The Tensorflow library is used for the creation, training and testing of the neural network. The depth is modelled as a piecewise polynomial function of the coordinate. The results of neural network testing with the use of noised input samples let to state that the neural net can determine a ship position by means of soundings with acceptable accuracy. Different combinations of the sea level error and the number of hidden neurons have been considered. For each of such combinations the network accuracy indicators have been calculated. The best results are obtained for the network with 100 hidden neurons per each layer.

Artificial neural network based photovoltaic module diagnosis by current–voltage curve classification

In this paper a model-based procedure for fault detection and diagnosis of photovoltaic modules is presented. A four-layered feedforward artificial neural network learns the correlation between the features of the current vs. voltage curve and the environmental variables, which are the irradiance and the temperature. This correlation describes the behavior of the module at normal conditions. Moreover, the effect of anomalous variation of some parameters is learnt and correlated to the shape of the same curve, thus associated to a specific failure mechanism and to some assigned ranges quantifying the fault severity. The neural network is trained by using synthetic curves simulated by employing the single diode model and some well assessed and validated translation formulae. The obtained results over the simulated set of curves with different failures allow to achieve a classification error lower than 1.5%. The proposed approach has been also validated for detecting anomalous increases of the series resistance in a large experimental set of curves; in this case, a classification error of 2.7% has been achieved.

A statistical bias correction technique for global climate model predicted near-surface temperature in India using the generalized regression neural network

Abstract Global climate models (GCMs) developed by the numerical simulation of physical processes in the atmosphere, ocean, and land are useful tools for climate prediction studies. However, these models involve parameterizations and assumptions for the simulation of complex phenomena, which lead to random and structural errors called biases. So, the GCM outputs need to be bias-corrected with respect to observed data before applying these model outputs for future climate prediction. This study develops a statistical bias correction approach using a four-layer feedforward radial basis neural network – a generalized regression neural network (GRNN) to reduce the biases of the near-surface temperature data in the Indian mainland. The input to the network is the CNRM-CM5 model output gridded data of near-surface temperature for the period 1951–2005, and the target to the model used for bias correcting the input data is the gridded near-surface temperature developed by the Indian Meteorological Department for the same period. Results show that the trained GRNN model can improve the inherent biases of the GCM output with significant accuracy, and a good correlation is seen between the test statistics of observed and bias-corrected data for both the training and testing period. The trained GRNN model developed is then used for bias correction of CNRM-CM5 modelled projected near-surface temperature for 2006–2100 corresponding to the RCP4.5 and RCP8.5 emission scenarios. It is observed that the model can adapt well to the nature of unseen future temperature data and correct the biases of future data, assuming quasi-stationarity of future temperature data for both emission scenarios. The model captures the seasonal variation in near-surface temperature over the Indian mainland, having diverse topography appreciably, and this is evident from the bias-corrected output.

Deep Neural Network for Compressive Sensing and Application to Massive MIMO Channel Estimation

In this paper, we consider the problem of sparse signal recovery using a learned dictionary in multiple measurement vectors (MMVs) case. Employing deep neural networks, we provide two new greedy algorithms to solve sparse MMV problems. In the first algorithm, we create a stacked vector of measurement matrix columns and a new measurement matrix, which can be assumed as the Kronecker product of the primary compressive sampling matrix and a unitary matrix. In order to reconstruct sparse vector corresponding to this new set of equations, a four-layer feed-forward neural network is applied. In the second algorithm, joint sparse structure of the sparse vectors is considered. Recurrent neural networks are employed to extract the joint sparsity structure. In addition, we utilize an over-complete dictionary obtained from an unsupervised learning procedure. Simulation results illustrate the benefit of using the proposed methods. Finally, the proposed algorithms are applied for pilot-based channel estimation in massive multiple-input multiple-output systems to improve the channel state information recovery performance.

Estimation of PM10 levels using feed forward neural networks in Igdir, Turkey

In areas with high air pollution, it is important to estimate the PM10 value causing this. In this way, it can be determined which input variables the PM10 value depends on and steps can be taken. Preventive measures can be taken by the public authority for the factors causing air pollution. The aim of this paper is to introduce the air pollution that has reached the threatening dimensions of human health today and to determine the effect of these elements by specifying the environmental conditions affecting it. In this study, we showed the effect of NOx and NOx concentrations on PM10 together with SO2. Also here, the effects of 9 different environmental factors (Nitrogen oxides (NOx), Sulfur dioxide (SO2), Nitrogen monoxide (NO), Nitrogen dioxide (NO2), Ozone (O3), Air temperature, Air humidity, Air pressure, and Wind speed) on PM10 value were investigated. For this reason, PM10 value was estimated by using these factors as input in the feed forward neural network. For this purpose, four-layered feed forward neural network architectures are used. 8 different experiments were done according to different layer sizes. In addition, experiments were carried out by reducing the size of the features with Principal Component Analysis (PCA) to see the results of a small number of input parameters. Five months of data obtained daily from public sources were used as a data set to conduct the experiments. Mean absolute error (MAE), correlation coefficient (r), root mean square error (RMSE) and R2 were used as performance metrics. MAE, r, RMSE and R2 were found to be 27.09, 0.95, 27.89 and 0.881, respectively. These results are quite satisfactory despite low data.

Estimating chlorophyll-A concentration in the Caspian Sea from MODIS images using artificial neural networks

Nowadays, due to various pollution sources, it is essential for environmental scientists to monitor water quality. Phytoplanktons form the end of the food chain in water bodies and are one of the most important biological indicators in water pollution studies. Chlorophyll-A, a green pigment, is found in all phytoplankton. Chlorophyll-A concentration indicates phytoplankton biomass directly. Therefore, Chlorophyll-A is an indirect indicator of pollutants, including phosphorus and nitrogen, and their refinement and control are important. The present study, Moderate Resolution Imaging Spectroradiometer (MODIS) satellite images were used to estimate the chlorophyll-A concentration in southern coastal waters in the Caspian Sea. For this purpose, Multi-layer perceptron neural networks (NNs) were applied which contained three and four feed-forward layers. The best three-layer NN has 15 neurons in its hidden layer and the best four-layer one has 5 in each. The three- and four- layer networks both resulted in similar root mean square errors (RMSE), 0.1(<math xmlns="http://www.w3.org/1998/Math/MathML"><mfrac><mrow><mi>μ</mi><mi>g</mi></mrow><mi>l</mi></mfrac></math>), however, the four-layer NNs proved superior in terms of R<sup>2</sup> and also required less training data. Accordingly, a four-layer feed-forward NN with 5 neurons in each hidden layer, is the best network structure for estimating Chlorophyll-A concentration in the southern coastal waters of the Caspian Sea.

Stability prediction of gate roadways in longwall mining using artificial neural networks

Roadways stability in longwall coal mining is critical to mine productivity and safety of the personnel. In this regard, a typical challenge in longwall mining is to predict roadways stability equipped with a reliable support system in order to ensure their serviceability during mining life. Artificial neural networks (ANNs) were employed to predict the stability conditions of longwall roadways based on roof displacements. In this respect, datasets of the roof displacements monitored in different sections of a 1.2-km-long roadway in Tabas coal mine, Iran, were set up to develop an ANN model. On the other hand, geomechanical parameters obtained through site investigations and laboratory tests were introduced to the ANN model as independent variables. In order to predict the roadway stability, these data were introduced to a multilayer perceptron (MLP) network to estimate the unknown nonlinear relationship between the rock parameters and roof displacements in the gate roadways. A four-layer feed-forward backpropagation neural network with topology 9-7-6-1 was found to be optimum. As a result, the MLP proposed model predicted values close enough to the measured ones with an acceptable range of correlation. A high conformity (R 2 = 0.911) was observed between predicted and measured roof displacement values. Concluding remark is the proposed model appears to be a suitable tool for prediction of gate roadways stability in longwall mining.

Misbehavior in a Neural Network Model

This paper describes a neural network account of misbehavior with an extant neural network model of conditioning. The model makes no distinction between learning (weight-change mechanisms) in operant and Pavlovian conditioning, but preserves the standard behavioral distinctions between types of stimuli, responses, and contingencies, with connectionist interpretations of some possible neuroanatomical substrates. Misbehavior has been traditionally conceived as a species-specific response R* that is unnecessary for a biologically significant reward S* but interferes with another response R that is necessary for S*. Misbehavior thus conceived has been explained as interfering Pavlovian conditioned responding. Three four-layer feedforward neural networks were designed to differ only in their output layers, as a connectionist interpretation of three hypothetical operant-Pavlovian relations in misbehavior, namely, interference (Pavlovian output to operant output lateral inhibitory connection), compatibility (Pavlovian output to operant output lateral excitatory connection), and independence (no lateral connection in the output layer). These relations are proposed as neural-network interpretations of neuroanatomical substrates of conditioning with three biologically significant stimuli, namely, food, water, and sexual mate, respectively. Each network first received pairings of contextual cues with its respective S*, to simulate pretraining with such stimuli. Then, networks received operant contingencies where S* was paired with the same contextual cues, as well as cues from a token dependently on R responding, defined as a minimal R activation of 0.5. Networks showed substantial misbehavior (qua conditioned R* responding) that interfered with R to different extents, food causing the most, sexual mate the least interference. Limitations, future directions, and implications for biological constraints and the generality of learning are discussed.

Advantages of Artificial Neural Network in Neutron Spectra Unfolding

Advantages of using the artificial neural network method in neutron spectra unfolding are investigated in comparison with the maximum entropy unfolding method. By introducing the information entropy theory, we find that for the spectrum with the information entropy over 3.5, the four-layer feed-forward neural network (11-35-55-60) and the maximum entropy method generally demonstrate the same unfolding performance, while the spectrum with the information entropy lower than 3.5, the artificial neural network unfolding model is recommend due to the fact that the artificial neural network method has a stronger negative correlation between the entropy of the spectra and the mean squares error of the spectra than the maximum entropy method.

Defects detection on TFT lines of flat panels using a feed forward neural network

This paper proposes a novel approach for inspection of open/short defects on thin film transistor (TFT) lines of flat panel displays (FPD). The inspection is performed on digitized waveform data of voltage signals that are captured by a capacitor based non-contact sensor by scanning over TFT lines on the surface of mother glass of FPD. The sudden deep falls (open circuits) or sharp rises (short circuits) on the captured noisy waveform are classified and detected by employing a four-layer feed forward neural network with two hidden layers. The topology of the network consist of an input layer with two units, two hidden layers with two and three units respectively and an output layer with one unit and a standard sigmoid function as the activation function of each unit. The network is trained with a fast adaptive back-propagation algorithm to find an optimal set of associated weights of neurons by feeding a known set of input data. This method is an alternative to the existing thresholding based non-contact method which has always its own limitations and drawbacks due to some non-avoidable features of input data such as non-stationary patterns and varying magnitude levels at defect points. Experimental results show that this method can adapt fast for new input patterns and avoids the ambiguity of threshold definitions and therefore it is more feasible than the existing thresholding method.

A Neural Network Approach for Non-contact Defect Inspection of Flat Panel Displays

This paper proposes a neural network-based approach for the inspection of electrical defects on thin film transistor lines of flat panel displays. The inspection is performed on digitized waveform data of voltage signals that are captured by a capacitor-based non-contact sensor by scanning over thin film transistor lines on the surface of the mother glass of flat panels. The sudden deep falls (open circuits) or sharp rises (short circuits) on the captured noisy waveform are classified and detected by employing a four-layer feed-forward neural network with two hidden layers. The topology of the network comprises an input layer with two units, two hidden layers with two and three units, and an output layer with one unit; a standard sigmoid function as the activation function for each unit. The network is trained with a fast adaptive back-propagation algorithm to find an optimal set of associated weights of neurons by feeding a known set of input data. The ambiguity of the threshold definition does not arise in this method because it uses only local features of waveform data at and around selected candidate points as inputs to the network, unlike the existing thresholding-based method, which is inherently prone to missed detections and false detections.

Prediction of tunnel convergence using Artificial Neural Networks

This research intends to develop a method based on Artificial Neural Network (ANN) for prediction of convergence in tunnels. In this respect, data sets of the convergence monitored in different section of a tunnel and geomechanical and geological parameters obtained through site investigations and laboratory tests are introduced to an ANN model. This data is used to estimate the unknown non-linear relationship between the rock parameters and convergence. Tunnel convergence model is developed, calibrated and tested using the above data from the perspective of Ghomroud water conveyance tunnel in Iran. The dominating rock masses in this case are metamorphic and sedimentary and are considered to be of weak to fair quality. In this tunnel there were some problems due to the convergence and instability of the tunnel. The tunnel boring machine has had several stoppages including a few major delays related to being trapped in squeezing ground and also delays due to face collapses. In order to predict the tunnel convergence a Multi-Layer Perceptron (MLP) analysis is used. A four-layer feed-forward back propagation neural network with topology 9-35-28-1 was found to be optimum. Simultaneously, the methods Radial Basis Function (RBF) analysis as another approach of ANN and Multi-Variable Regression (MVR) as a linear regression using statistical approach are used to analyze the problem and the results are compared. As a result, the MLP proposed model predicted values closer to the measured ones with an acceptable range of correlation. After the calibration and assessment of the ANN model, a parametric study is also carried out to estimate the intensity of the impact of the geological and rock mechanics parameters on tunnel convergence. It is observed that C, Φ, E and UCS parameters are the most effective factors and σt is the least effective one. Concluding remark is the proposed model appears to be a suitable tool for the prediction of convergence in the unexcavated zones of the tunnel as well as in new tunnels to be excavated in the similar geological environment. The results show that an appropriately trained neural network can reliably predict the convergence in tunnels.

Analysis and modeling of multivariate chaotic time series based on neural network

A new nonlinear multivariate technique is proposed for modeling and predicting chaotic time series with a view to improve estimates and predictions. With analysis of the relations among different state spaces by the proposed method, which introduces the reverse-predictability and time spans to discover the underlying relationship, the connections among multivariate time series are discussed before prediction. Then we predict the time series by multivariate prediction. Though multivariate time series can bring more information about the complex system, which can enhance the accuracy of prediction, they also bring a too large number of input variables which may result in overfitting and poor generalization abilities. To overcome the shortcomings, principal component analysis (PCA) based on singular value decomposition (SVD) is used to extract main features of multivariate time series and reducing the dimension of the model inputs. Then based on Takens’ delay time theory, the multivariate time series are reconstructed. Subsequently, a four-layer feedforward neural network is trained as the multivariate predictive model. Three simulation examples, that are coupled Henon equation and two set of real world time series, are used to demonstrate the validity of the proposed method.

Automated Identification of Diabetic Type 2 Subjects with and without Neuropathy Using Wavelet Transform on Pedobarograph

Diabetes is a disorder of metabolism-the way our bodies use digested food for growth and energy. The most common form of diabetes is Type 2 diabetes. Abnormal plantar pressures are considered to play a major role in the pathologies of neuropathic ulcers in the diabetic foot. The purpose of this study was to examine the plantar pressure distribution in normal, diabetic Type 2 with and without neuropathy subjects. Foot scans were obtained using the F-scan (Tekscan USA) pressure measurement system. Various discrete wavelet coefficients were evaluated from the foot images. These extracted parameters were extracted using the discrete wavelet transform (DWT) and presented to the Gaussian mixture model (GMM) and a four-layer feed forward neural network for classification. We demonstrated a sensitivity of 100% and a specificity of more than 85% for the classifiers.

Line flow contingency selection and ranking using cascade neural network

Line flow or real-power contingency selection and ranking is performed to choose the contingencies that cause the worst overloading problems. In this paper, a cascade neural network-based approach is proposed for fast line flow contingency selection and ranking. The developed cascade neural network is a combination of a filter module and a ranking module. All the contingency cases are applied to the filter module, which is trained to classify them either in critical contingency class or in non-critical contingency class using a modified BP algorithm. The screened critical contingencies are passed to the ranking module (four-layered feed-forward artificial neural network (ANN)) for their further ranking. Effectiveness of the proposed ANN-based method is demonstrated by applying it for contingency screening and ranking at different loading conditions for IEEE 14-bus system. Once trained, the cascade neural network gives fast and accurate screening and ranking for unknown patterns and is found to be suitable for on-line applications at energy management centre.

CONTROLLING MANIPULATOR OFF-LINE USING ELECTROMYOGRAPHY SIGNALS

The objective is to control manipulator ofF-line by using electromyography signals. Electromyography signals are collected from the biceps and triceps muscles of normal subjects when they move their elbow flexion-extension with time-varying loads. The raw electromyography signals are processed and the new defined characteristic is picked up. A four-layer feed-forward neural network model with the characteristic as its input is developed. The weighted values of the model are optimized with the adjusted back-propagation algorithm. By training the model the transformation can be mapped: From the processed eletromyography signals to the elbow joint angles. The predicted angles are used to control the manipulator by inverse-control method. The angles of the manipulator are compared with those of the elbow joint. The experimental results show that the root mean square error between the joint angle of the manipulator and the actual joint angle measured by the goniometer is less than 1°.

Neural Network Model for Asphalt Concrete Permeability

In this study, a four-layer feed-forward neural network is constructed and applied to determine a mapping associating mix design and testing factors of asphalt concrete samples with their performance in conductance to flow or permeability. To generate data for the neural network model, a total of 100 field cores from 50 different mixes (two replicate cores per mix) are tested in the laboratory for permeability and mix volumetric properties. The significant factors that affect asphalt permeability are identified using simple and multiple regression analysis. The analyses results show that permeability of an asphalt concrete is affected mainly by five factors: (1) air void ( Va ) ; (2) the grain size through which 10% materials pass ( d10 ) ; (3) the grain size through which 30% materials pass ( d30 ) ; (4) saturation, or the CoreLok Infiltration Coefficient (CIC); and (5) effective asphalt to dust ratio ( Pbe ∕ P0.075 ) . The significant factors are then used to define the domain of a neural network. Regar...

A Hybrid Neuro-Fuzzy Model for Mineral Potential Mapping

A GIS-based hybrid neuro-fuzzy approach to mineral potential mapping implements a Takagi–Sugeno type fuzzy inference system in a four-layered feed-forward adaptive neural network. In this approach, each unique combination of predictor patterns is considered a feature vector whose components are derived by knowledge-based ordinal encoding of the constituent predictor patterns. A subset of feature vectors with a known output target vector (i.e., unique conditions known to be associated with either a mineralized or a barren location), extracted from a set of all feature vectors, is used for the training of an adaptive neuro-fuzzy inference system. Training involves iterative adjustment of parameters of the adaptive neuro-fuzzy inference system using a hybrid learning procedure for mapping each training vector to its output target vector with minimum sum of squared error. The trained adaptive neuro-fuzzy inference system is used to process all feature vectors. The output for each feature vector is a value that indicates the extent to which a feature vector belongs to the mineralized class or the barren class. These values are used to generate a favorability map. The procedure is applied to regional-scale base metal potential mapping in a study area located in the Aravalli metallogenic province (western India). The adaptive neuro-fuzzy inference system demarcates high favorability zones occupying 9.75% of the study area and identifies 96% of the known base metal deposits. This result is significant both in terms of reduction in search area and the percentage of deposits identified.

Thunderstorm identification from AMSU‐B data using an artificial neural network

AbstractArtificial Neural Networks (ANNs) that were originally proposed to model brain functions mathematically are now being used in many branches of research, including the atmospheric sciences. The main contribution of this paper is the development of ANN to identify the presence of thunderstorms (and their location) based on microwave measurements from all five AMSU‐B frequencies and on surface observations over Iran (45–65°E,25–45°N). From the surface observations, the position of each of the different meteorological events (thunderstorm, heavy rain, rain, light rain, snowfall and clear sky) is located in AMSU‐B images. An area around this position with roughly the same brightness (visually) is selected. Then the brightness temperatures (TBs) of a few randomly selected pixels from this area are noted. In all, there were 560 pixel‐points, over 20 passes of AMSU‐B, in these selected areas. It is found that the thunderstorms are related to image features that are bright in all of the five‐frequency images (i.e. have low TB), as expected.The architecture of the neural network that was finally selected for identifying thunderstorm events (as against other reported events), after a series of trial and error for the best accuracy, was achieved using the four‐layered Feed‐Forward Artificial Neural Network (FFANN) with five nodes in input, two hidden layers with four nodes each, and one node in the output layer. In the ANN output, the thunderstorms clustered clearly away from other phenomena. ANN gave more correct thunderstorm recognition (96.7%), whereas the discriminant analysis gave 95.3% correct recognition without clear separation between the two classes. Copyright © 2003 Royal Meteorological Society.

Universal approximations of continuous fuzzy-valued functions by multi-layer regular fuzzy neural networks

The fact that four-layer feedforward regular fuzzy neural networks with sigmoid function in the first hidden layer are capable of approximately representing continuous fuzzy valued functions on any compact set of R is shown. At first, Bernstein polynomials associated with fuzzy valued functions are employed to approximate continuous fuzzy valued function defined on a compact set. Secondly, by the conclusions related to standard feedforward networks, universal approximations of continuous fuzzy valued functions by regular fuzzy neural networks are obtained.