Application Of Backpropagation Neural Networks Research Articles

Safety Risk Assessment Analysis of Bridge Construction Using Backpropagation Neural Network

The evaluation of construction safety risks has become a crucial task with the increasing development of bridge construction. This paper aims to provide an overview of the application of backpropagation neural networks in assessing safety risks during bridge construction. It introduces the situation, principles, methods, and advantages, as well as the current status and future development directions of backpropagation-related research.

Pattern recognition analysis of marine oil spills in airborne passive infrared multispectral remote sensing images.

Pattern recognition methodology was developed for the automated detection of marine oil spills in passive infrared multispectral remote sensing images. The images employed in this work were collected from the Deepwater Horizon oil spill accident in 2010. The imaging instrument for data collection was a downward-looking infrared line scanner equipped with eight optical bandpass filters in the spectral range of 8-12 μm on a fixed-wing aircraft. Oil slicks may show either positive or negative thermal contrast against the surrounding sea water, depending on the sun glint conditions or the oil thickness. Classifiers were developed separately to detect oil with different contrasts by the application of backpropagation neural networks to the preprocessed radiances. Preprocessing strategies included: (1) assembly of training data through k-means clustering analysis; (2) elimination of variation in radiance magnitudes by a customized temperature correction method; (3) removal of sun glint artifacts in images by polynomial correction; and (4) extraction of the most representative features as inputs for the neural networks by a subset selection approach. The classifiers designed to detect oil with positive and negative thermal contrast relative to water achieved overall classification accuracies of 88.7 and 92.2%, respectively. Composite classification images were generated by integrating classification scores produced by the two classifiers. The prediction performance of the classification system was demonstrated through its application to images not involved during the training of the networks.

Application of back-propagation neural networks in the distribution of oil sand reservoirs\u2014a case study from the western slope of the Songliao Basin, China

High-quality hydrocarbon source rocks are present in the upper Cretaceous layer in the western slope of the Songliao Basin. Oil and gas have accumulated in these rocks at the shallow edge of the basin, which has led to the formation of oil sand resources. This study uses the back-propagation (BP) neural network method to predict the distribution of oil sand reservoirs and is the first study of its kind in China. First, based on the basic data collected by core sample, well log and geochemical analyses, and the reasonable selection of samples, the cores are divided into mudstone, siltstone, fine sandstone, medium sandstone, and sand, according to lithology. Second, a three-layer BP neural network model is constructed with two hidden S-type layers and one linear output layer. Third, through a comparison of the effect of different numbers of training sessions of the sample data, we demonstrate that the accuracy of the model can be increased to 90% after training the network 100,000 times. Then, the log-derived data of rocks with unknown lithologies are input into the neural network to predict whether they contain oil sands. We show that the BP neural network method can predict the distribution of oil sand reservoirs in the target horizon of the study area, and the results are consistent with research results on the corresponding sand reservoirs and sedimentary facies. Thus, we conclude that it is feasible to use the BP neural network method to predict the distribution of oil sand reservoirs.

Mechanical behavior prediction of CF/PEEK-titanium hybrid laminates considering temperature effect by artificial neural network

This paper focuses on the mechanical behaviors of carbon fiber-reinforced PEEK-titanium hybrid laminates (TiGr) under different temperatures. The tensile strength, flexural strength and inter-laminar shear strength (ILSS) of TiGr laminates with different stacking structures were investigated at room temperature, 80, 150 and 220 °C, respectively. The results show that the tensile strength, flexural strength and ILSS gradually decrease as the temperature increases, and reduced by 12.3%, 45.4% and 29.1% for Type I, 12.9%, 48.8% and 30.1% for Type II, 16.1%, 45.1% and 39.4% for Type III at 220 °C. Moreover, a three-layers back-propagation (BP) neural network was developed to predict the corresponding mechanical behaviors of TiGr laminates, then the genetic algorithm (GA) was introduced to optimize the established BP neural network and the accuracies of the two prediction models were compared. It is found that both the trained BP and GA-BP neural networks could predict the mechanical performances of TiGr laminates well while the GA optimized BP neural network has better prediction results, with average absolute errors of 1.78%, 1.92% and 2.41% for tensile, flexural and inter-laminar shear strengths, respectively. Furthermore, the predictability of the trained models was validated with data from published literature, providing a new concept for the application of BP neural networks to fiber metal laminates.

Application of back-propagation neural networks to defect characterization using eddy current testing

Eddy current testing is widely used for the automatic detection of defects in conductive materials. However, this method is strongly affected by probe scanning conditions and requires signal analysis to be carried out by experienced inspectors. In this study, back-propagation neural networks were used to predict the depth and length of unknown slits by analyzing eddy current signals in the presence of noise caused by probe lift-off and tilting. The constructed neural networks were shown to predict the depth and length of defects with relative errors of 4.6% and 6.2%, respectively.

Application of BP Neural Networks on the Thickness Prediction of Sherardizing Coating

Sherardizing is a surface protection method for obtaining Fe–Zn coating on the steel surface by thermal diffusion. The thickness of coating is critical for its performance. Therefore, it is necessary to propose a model to predict coating thickness based on different process parameters. The model for the relationship between sherardizing process parameters and the total thickness and each layer phase thickness of the coating is built using a three-layer backpropagation (BP) artificial neural network based on Levenberg–Marquardt algorithm. The results show that neural network model has a good predictive effect on the total thickness and each layer phase thickness of Fe–Zn coating. The prediction error of the model on the thickness of the coating is within 4.71%, and the coefficient of determination is 0.99974. It presents a new method for prediction of the thickness of sherardizing coating.

Applying theoretical spectra to artificial neural networks for real-time estimation of thin film thickness

This paper presents the application of backpropagation neural networks (BPNNs) for estimating the thickness of deposited silver (Ag) films on polyethylene terephthalate substrate via a roll-to-roll magnetron sputtering system. The thickness of thin films affects the optical properties of films, while the transmittance implies the thickness. Nevertheless, thin films are unlike their bulk counterparts whose absorptions of light are proportional to their thicknesses. Moreover, the interference is considerable. Thus, BPNNs are applied for estimating thickness of Ag films. BPNNs were trained via theoretical transmittance spectra because they can be quickly generated and reduce actual experiments. The BPNNs were applied to estimate thickness via actual spectra. Different levels of noise were also added to the theoretical spectra to improve the performance of BPNNs. The results show that the estimation of BPNNs is more accurate when adding slight noise to the theoretical spectra. The average error is ∼0.027 when 3% noise is added to the training spectra, while the error of spectra without adding noise is greater than 0.12.

Improving the forecasting accuracy of air passenger and air cargo demand: the application of back-propagation neural networks

This study employs back-propagation neural networks (BPN) to improve the forecasting accuracy of air passenger and air cargo demand from Japan to Taiwan. The factors which influence air passenger and air cargo demand are identified, evaluated and analysed in detail. The results reveal that some factors influence both passenger and cargo demand, and the others only one of them. The forecasting accuracy of air passenger and air cargo demand has been improved efficiently by the proposed procedure to evaluate input variables. The established model improves dramatically the forecasting accuracy of air passenger demand with an extremely low mean absolute percentage error (MAPE) of 0.34% and 7.74% for air cargo demand.

Artificial Neural Network Simulation of Combined Permeable Pavement and Earth Energy Systems Treating Storm Water

Artificial intelligence techniques, such as neural networks, are modeling tools that can be applied to analyze urban runoff quality issues. Artificial neural networks are frequently used to model various highly variable and nonlinear physical phenomena in the water and environmental engineering fields. The application of neural networks for analyzing the performance of combined permeable pavement and earth energy systems is timely and novel. This paper presents the application of back-propagation neural networks and the testing of the Levenberg-Marquardt, Quasi-Newton, and Bayesian Regularization algorithms. The neural networks were statistically assessed for their goodness of prediction with respect to the biochemical oxygen demand (BOD), ammonia-nitrogen, nitrate-nitrogen, and ortho-phosphate-phosphorus by numerical computation of the mean absolute error, root-mean-square error, mean absolute relative error, and the coefficient of correlation for the prediction compared with the corresponding measured data. The three neural network models were assessed for their efficiency in accurately simulating the effluent water quality parameters from various experimental pavement systems. The models predicted all key parameters with high correlation coefficients and low minimum statistical errors. The back-propagation and feed-forward neural network models performed optimally as pollutant removal predictors with regard to these two sustainable technologies.

Application of BP neural networks in non-linearity correction of optical tweezers

The back-propagation (BP) neural network is proposed to correct nonlinearity and optimize the force measurement and calibration of an optical tweezer system. Considering the low convergence rate of the BP algorithm, the Levenberg-Marquardt (LM) algorithm is used to improve the BP network. The proposed method is experimentally studied for force calibration in a typical optical tweezer system using hydromechanics. The result shows that with the nonlinear correction using BP networks, the range of force measurement of an optical tweezer system is enlarged by 30% and the precision is also improved compared with the polynomial fitting method. It is demonstrated that nonlinear correction by the neural network method effectively improves the performance of optical tweezers without adding or changing the measuring system.

Application of back-propagation neural network modeling for free residual chlorine, total trihalomethanes and trihalomethanes speciation

The application of back-propagation neural networks (BPNNs) was assessed for modeling residual chlorine decay, total THM concentrations (TTHMs), and three individual THM species (CHCl3, CHBrCl2, and CHBr2Cl) in water that was chlorinated under laboratory-scale conditions. Data for modeling chlorine decay and TTHM were generated in chlorination experiments carried out with water collected in water utilities of the Quebec City region, whereas data for THM species were provided by the US Geological Survey for the Mississippi River and its tributaries. The BPNN models were compared with conventional models developed with exactly the same data. Results showed that the ability of BPNN to model residual chlorine decay is, in general terms, comparable to the ability of kinetic first- and second-order models. For TTHMs and THM speciation, however, the ability of BPNN was clearly higher in comparison with multivariate regression models, in particular when brominated disinfection by-products (DBPs) (CHBrCl2 and CHBr2Cl) were modeled. The successful application of BPNN presented in this study opens the door to other potential applications of BPNN for field-scale data concerning THMs as well as for other relevant disinfection by-products. Key words: back-propagation neural networks, drinking water, chlorination, residual chlorine decay, trihalomethanes, prediction.

Neural simulation of dynamic response of prefabricated buildings subjected to paraseismic excitations

On the basis of measurements on a group of 5-storey flat buildings experimental data were collected. Back-propagation neural networks (BPNNs) are formulated as replicators for data compression. The compressed values of the discretized displacement response spectrum are associated with paraseismic excitations caused by explosions in nearby quarries. The compressed excitations are completed by buildings parameters and used in the master BPNN for simulation of compressed outputs which correspond to the building displacement response on the 4th floor of buildings. The BPNN replicator is also formulated for the compression of the building displacement records in time domain into the target vectors of the master BPNN. After the training the compressed outputs of the master BPNN are decompressed by means of the BPNN decompressor into the building displacement records. In the paper it was proved that the discussed application of BPNNs gives satisfactory neural simulation of displacement records in time domain for vibrations with large amplitudes without analysis of the building motion equations.

Application of neural networks in natural periods identification and simulation of prefabricated buildings response

The paper deals with an application of neural networks for detection of natural periods of vibrations of prefabricated, medium height buildings. The neural network technique is also used to simulate the dynamic response at selected floor of one of the analysed buildings subject to seismic loading induced by explosives in a nearby quarry. Both the training and testing patterns were formulated on the basis of measurements performed on actual structures. The results of neural network identification of natural periods of the considered buildings obtained with different soil, geometrical and stiffness parameters are compared with the results of experiments. The application of back-propagation neural networks enables us to identify the natural periods of the buildings with accuracy quite satisfactory for engineering practice. The experimental and generated data of vibration displacements are compared and much clearer comparison is given on the phase plane: displacements versus velocities. It was stated that a good generalization takes place both with respect to displacements and velocities.

Importance of input data normalization for the application of neural networks to complex industrial problems

Recent advances in artificial intelligence have allowed the application of such technologies in real industrial problems. We have studied the application of backpropagation neural networks to several problems of estimation and identification in nuclear power plants. These problems often have been reported to be very time-consuming in the training phase. Among the different approaches suggested to ease the backpropagation training process, input data pretreatment has been pointed out, although no specific procedure has been proposed. We have found that input data normalization with certain criteria, prior to a training process, is crucial to obtain good results as well as to fasten significantly the calculations. This paper shows how data normalization affects the performance error of parameter estimators trained to predict the value of several variables of a PWR nuclear power plant. The criteria needed to accomplish such data normalization are also described.

Application of back-propagation neural networks to identification of seismic arrival types

A back-propagation neural network (BPNN) approach is developed to identify P- and S-arrivals from three-component recordings of local earthquake data. The BPNN is trained by selecting trace segments of P- and S-waves and noise bursts converted into an attribute space based on the degree of polarization (DOP). After training, the network can automatically identify the type of arrival on earthquake recordings. Compared with manual analysis, a BPNN trained with nine groups of DOP segments can correctly identify 82.3% of the P-arrivals and 62.6% of the S-arrivals from one seismic station, and when trained with five groups from a training dataset selected from another seismic station, it can correctly identify 76.6% of the P-arrivals and 60.5% of S-arrivals. This approach is adaptive and needs only the onset time of arrivals as input, although its performance cannot be improved by simply adding more training datasets due to the complexity of DOP patterns. Our experience suggests that other information or another network may be necessary to improve its performance.

Application of backpropagation neural networks to diagnosis of breast and ovarian cancer

Neural network programs have been developed in an attempt to improve the diagnosis of breast and ovarian cancer using a group of laboratory tests and the age of the patient. The laboratory tests employed in this study include albumin, cholesterol, HDL-cholesterol, triglyceride, apolipoproteins Al and B, NMR linewidth (the Fossel Index) and a tumor marker (i.e., CA 15-3 or CA 125). The breast cancer study involved 104 patients (45 malignant and 59 benign subjects). The ovarian cancer study involved 98 individuals (35 malignant, 36 benign and 27 control subjects). Methods are outlined for identification of the most influential input parameters and optimization of network structure and training. Network characteristics were contrasted with the test results of the appropriate serum tumor marker assay. For the breast cancer study, the best neural network program, using six input parameters, had a sensitivity of only 55.6% and a specificity of 72.9%. The tumor marker CA 15-3 alone gave results of 61.3% and 64.4%, respectively. For the ovarian cancer study, the best neural network program, using six input parameters, had a sensitivity of 80.6% and a specificity of 85.5%. The tumor marker CA 125 alone gave results of 77.8% and 82.3%, respectively. These methods provide an objective approach to neural network optimization and parameter selection applicable to other data bases of clinical and laboratory data.

Solving the interconnection problem with a linked assembly of neural networks

To reduce the complexity of training a single, large neural network, partitioning of the problem is introduced to facilitate the identification of smaller and, where possible, replicable networks which are more readily trained. A system for the linked assembly of these neural networks (ASLANN) has been developed and is used to generate the final neural system. To demonstrate this approach we discuss the application of backpropagation neural networks to the routing of an integrated circuit.