Predicting the Crushing Behavior of Axially Loaded

download Predicting the Crushing Behavior of Axially Loaded

of 13

Transcript of Predicting the Crushing Behavior of Axially Loaded

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    1/13

    Predicting the Crushing Behavior of Axially Loaded

    Elliptical Composite Tubes Using Artificial NeuralNetworks

    Hany El Kadi

    Received: 11 September 2008 / Accepted: 30 October 2008 /

    Published online: 13 November 2008

    # Springer Science + Business Media B.V. 2008

    Abstract In this research work, the artificial neural networks (ANN) technique is used in

    predicting the crushing behavior and energy absorption characteristics of axially-loaded

    glass fiber/epoxy composite elliptical tubes. Predictions are compared to actual experimental

    results obtained from the literature and are shown to be in good agreement. Effects of

    parameters such as network architecture, number of hidden layers and number of neurons per

    hidden layer are also considered. The study shows that ANN techniques can effectively be

    used to predict the crushing response and the energy absorption characteristics of elliptical

    composite tubes with various ellipticity ratios subjected to axial loading.

    Keywords Composite tubes . Artificial neural networks . Crushing behavior. Ellipticity ratio

    1 Introduction

    Impact energy absorbers are used to protect automobile passengers and pedestrians from the

    effects of sudden impact caused by collisions. This is accomplished by converting the

    impact energy into many different types of deformation energy keeping the peak forceexerted on the individual below the level causing damage [1, 2]. The absorbers must also

    provide a long deformation path to slow the deceleration of the protected person. These

    impact energy absorbers will mostly rely on a crushable energy absorber to cushion the

    passenger compartment during impact.

    The use of advanced materials in the design of energy absorber devices has been

    hindered by a lack of experimental and numerical simulation work that would guide

    designers to the optimum energy absorber device. Extensive utilization of advanced

    composites in energy absorber design will mainly depend on finding accurate techniques to

    predict their response to different loading conditions. To that goal, many studies have lately

    investigated the quasi-static crushing behavior of composite tubes both experimentally and

    Appl Compos Mater (2008) 15:273285

    DOI 10.1007/s10443-008-9074-2

    H. El Kadi (*)

    Mechanical Engineering Department, American University of Sharjah, Sharjah,

    United Arab Emirates

    e-mail: [email protected]

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    2/13

    numerically using finite elements analysis [112]. In these cases, the load was either

    applied to the tube in a transverse or in an axial direction. In these studies, both the load-

    carrying capacity and the energy-absorption capability of composite tubes were

    investigated.

    The behavior of axially-loaded elliptic composite tubes under compression loading hasrecently been investigated both experimentally and numerically [11]. In this study, the

    effect of the ellipticity ratio (a/b; a=inner major radius, b=inner minor radius) on the load-

    carrying capacity of the tubes as well as the energy absorption until failure were also

    investigated. Finite elements analysis was used for the numerical investigation of elliptic

    tubes with ellipticity ratios varying from 1 to 2 (1 signifying a circular tube). Although

    good agreement was obtained from the finite element analysis compared to the

    experimental results, the authors emphasized that typical imperfections existing in the

    manufacturing process of the tubes could not possibly be accounted for by the analysis.

    They suggested using a non-linear finite element analysis to physically include these

    imperfections in the model.

    One way of including specimen irregularities and material inhomogeneities in modeling

    is to use the results of the available experiments to predict the behavior of composite tubes

    subjected to different loading. Artificial neural-networks (ANN) is a technique that uses

    existing experimental data to predict the behavior under a variety of testing conditions.

    Using this method, details regarding bonding properties between fiber & matrix, strength

    variation of fibers and any manufacturing-induced imperfections are implicitly incorporated

    within the input parameters fed to the neural network.

    Caliskan [13] has one of the few published studies dealing with the use of these

    networks in predicting the crushing energy absorption of carbon fiber-reinforced circulartubes under axial loading. A simple neural network with back-error-propagation algorithm

    was trained using 84 data sets of crush energy absorption of circular tubes obtained from

    the literature. Ten input parameters were fed into the network; these included material and

    geometric properties of the tubes. The network was then used to predict the specific

    energy absorption of a single tube. Comparing the average experimental results obtained

    from six tests to the ANN predictions resulted in an error of 14%. The author suggested

    that neural networks could more accurately predict the crushing behavior of these tubes if

    a more complex network was used or if the properties of all input parameters were

    experimentally measured rather than calculated using micromechanics and laminate plate

    theory.In a recent work, Mahdi and El Kadi [14] evaluated the prediction of both load-carrying

    capacities and energy absorption of elliptical composite tubes using artificial neural

    networks (ANN). In this study, the experimental behavior and corresponding ANN

    predictions of circular and elliptical tubes subjected to lateral compressive loads were

    presented and discussed. The ANN was shown to successfully predict the crushing

    behavior of tubes for a wide range of ellipticity ratios. The predicted results obtained from

    the neural network were compared with actual experimental data in terms of load-carrying

    capacity and energy absorption capability, showing an excellent agreement. It was

    concluded that ANN techniques could effectively be used to predict the response ofcollapsible composite energy absorber device subjected to different loading conditions.

    In the current work, the prediction of both load-carrying capacities and energy

    absorption for axially loaded elliptic composite tubes is evaluated using artificial neural

    networks (ANN). To test the validity of using ANN in determining the crushing behavior of

    these tubes, the study will compare the predictions obtained using ANN to the experimental

    results obtained from the literature [11] for various ellipticity ratios.

    274 Appl Compos Mater (2008) 15:273285

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    3/13

    2 Experimental Investigation

    The current work will make use of the comprehensive experimental program developed by

    Alkoles et al [11] which deals with the crushing behavior and energy absorption

    characteristics of axially loaded glass fiber/epoxy composite elliptical tubes. Axial quasi-static crushing tests were carried out on thin-walled glass/epoxy composite elliptical tubes

    with various ellipticity ratios by compressing them between parallel rigid plates. The load

    and displacements were recorded by an automatic data acquisition system and, as the quasi-

    static crushing tests were carried-out, instant photographs were taken throughout the test.

    Figures 1, 2, 3, 4, 5 show the typical deformation history and corresponding load-end

    shortening path in elliptical composite tubes for the various ellipticity ratios considered. For

    more specific details about the load deformation relation of each of the tubes, one could

    refer to the work of Alkolose et al [11]. The experimental results are shown here for the

    sake of completeness and for proper comparison with the predictions obtained by the

    artificial neural networks introduced later in this work. Since the purpose of the current

    work is to gauge the effectiveness of ANN in predicting the crushing behavior of axially

    loaded composite tubes rather than compare the experimental results obtained in [11] to

    other published experimental result, only the experiments published in [11] will be used

    Fig. 1 Typical load-end shortening path histories of axially loaded elliptical composite tube with a/b ratio of

    1.00 [11] (With kind permission of Springer Science and Business Media)

    Appl Compos Mater (2008) 15:273285 275

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    4/13

    here. Using ANN to predict other experimental results published in the literature will

    necessitate training the networks using the corresponding experimental data.

    3 Artificial Neural Networks (ANN)

    Artificial Neural Networks have emerged as one of the useful artificial intelligence concepts

    successfully used over the past decade in modeling the mechanical behavior of fiber-

    reinforced composite materials (see for example [15, 16].)

    In general, ANN consist of a layer of input neurons, a layer of output neurons and one or

    more layers of hidden neurons [1719]. Neurons in each layer are interconnected to

    preceding and subsequent layer neurons with each interconnection having an associated

    connection strength (or weight).

    A training algorithm is commonly used to iteratively minimize the following cost

    function with respect to the interconnection weights and neuron thresholds:

    E 1

    2

    XM

    1

    XN

    i1

    di zi 2

    where M is the number of training patterns and N is the number of output nodes. di and ziare the desired and actual responses for output node i, respectively.

    Fig. 2 Typical load-end shortening path histories of axially loaded elliptical composite tube with a/b ratio of

    1.25 [11] (With kind permission of Springer Science and Business Media)

    276 Appl Compos Mater (2008) 15:273285

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    5/13

    The training process is terminated either when the mean-square-error (MSE) between

    the observed data and the ANN outcomes for all elements in the training set has reached

    a pre-specified threshold or after the completion of a pre-specified number of learning

    epochs.

    Although all neural network models share common operational features, their underlyingstructures, input requirements and modeling and generalization abilities are different.

    Consequently, each paradigm would have advantages and disadvantages depending on the

    particular application. Hence, selecting the appropriate network class with suitable parameters

    is vital to ensure a successful application. The following neural network architectures will be

    considered here in predicting the crushing behavior in composite tubes:

    Feedforward neural networks (FNN) This is the most known and commonly used class of

    neural networks. Although the main success of neural networks has been in the application

    of the multilayer FNN with back-propagation training, they suffer from some drawbacks

    such as local convergence and the need for large training cases in order to make adequate

    modeling generalization [19].

    Recurrent Neural Networks (RNN) RNN distinguish themselves from FNN in that the

    outputs from some neurons are fed back to the same neurons or to the neurons in preceding

    layers. Thus signal can flow in both forward and backward directions [ 18]. The Elman

    Fig. 3 Typical load-end shortening path histories of axially loaded elliptical composite tube with a/b ratio of

    1.50 [11] (With kind permission of Springer Science and Business Media)

    Appl Compos Mater (2008) 15:273285 277

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    6/13

    neural networks (ENN) are multi-layer back-propagation partially recurrent networks, with

    the addition of a feedback connection from the output of the hidden layer to its input. In

    partially recurrent networks, the main network structure is feedforward. These feedforwardconnections are trainable. The feedback connections are formed through a set ofcontext

    units and are not trainable. The context units memorize some past states of the hidden units,

    and so outputs of the network depend on an aggregate of the previous states and the current

    input [20].

    Modular neural networks (MNN) The central idea behind such networks is task

    decomposition, where in this case the concept of using a combined (or averaged) estimator

    may be able to exceed the limitation of a single estimator [18, 19]. Using a modular

    network, the task of predicting the crushing behavior of composite tubes is split up amongseveral local neural networks (sub-networks) not communicating with each other.

    Combination weights that determine the degree by which each sub-network should

    contribute to the final composite material properties are estimated with an integrating unit.

    Furthermore, this integrating unit decides which module should learn which training

    pattern.

    Fig. 4 Typical load-end shortening path histories of axially loaded elliptical composite tube with a/b ratio of

    1.75 [11] (With kind permission of Springer Science and Business Media)

    278 Appl Compos Mater (2008) 15:273285

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    7/13

    4 Prediction of Axial Load-Carrying Capacity

    The three neural network architectures introduced were used to predict the axial load

    carrying capacity of the elliptic composite tubes. The Neurosolution-5 software [21] wasused to construct, train and test the networks. In all cases, the input experimental data to the

    network was the ellipticity ratio and the specimen deformation. The output was the tube

    load-carrying capacity. In each case, the network was trained using all but one of the

    ellipticity ratios obtained experimentally. The network was then required to predict the

    behavior of the composite tube for the ellipticity ratio it was not trained for. The predictions

    obtained were then compared to the experimental results for this ellipticity ratio. Once

    assured that the predictions obtained are reliable, the network could be used in the future to

    predict the behavior of a tube with any ellipticity ratio for which experimental results do not

    exist. Ellipticity ratios of 1.0, 1.25, 1.5, 1.75 and 2.0 were used in this study. Since the

    ANN cannot be accurately used to predict behavior outside the area of training, predicting

    the behavior of the elliptic tubes with 1.0 and 2.0 ellipticity ratios was not attempted.

    In the work done for transverse loading [14], the effect of the number of hidden layers

    and the number of neurons per layer was not considered. This is because the main goal of

    that study was to establish the feasibility of using ANN to predict the crushing behavior of

    Fig. 5 Typical load-end shortening path histories of axially loaded elliptical composite tube with a/b ratio of

    2.00 [11] (With kind permission of Springer Science and Business Media)

    Appl Compos Mater (2008) 15:273285 279

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    8/13

    composite tubes. In the current study, the effect of varying the number of hidden layers as

    well as the number of neurons per layer was also examined. Prediction was attempted using

    the three above-mentioned ANN configurations with one and two hidden layers. The

    number of training epochs was held constant at 5,000 epochs.

    Tables 1, 2, 3 show the mean square error obtained from the predicted values of the load-displacement relation of the tubes compared to the experimental data for each of the three

    eccentricity ratios (1.25, 1.5, and 1.75). The tables show the effect of varying the neural

    network architecture, the number of hidden layers and the number of neurons per hidden

    layer. Although no single neural network configuration consistently resulted in the best

    load-displacement predictions, predictions using ANN with two hidden layers consistently

    lead to lower MSE values when compared to experimental data. Figures 6, 7, 8 graphically

    compare the load-displacement relationship obtained experimentally to typical predictions

    from the three neural networks for each of the three eccentricity ratios investigated. In these

    figures, FFN1 and FFN2 relate to predictions obtained using feedforward neural networks

    with one and two hidden layers respectively, ENN1 and ENN2 relate to predictions

    obtained using Elman neural networks with one and two hidden layers respectively, and

    MNN1 and MNN2 relate to predictions obtained using modular neural networks with one

    and two hidden layers respectively. The predictions obtained using the Elman neural

    network with one and two hidden layers were the only type of ANN able to match the

    experimental behavior at low deformations; all other ANN structures predicted an initial

    load carrying capacity at the start of the test. From these typical results, it can be concluded

    that artificial neural networks can, in general, be used to predict the load-displacement

    relationship for the composite elliptical tubes subjected to axial loading.

    5 Crushing Behavior

    In order to study the effects ofa/b ratio on the crashworthiness of elliptical composite tubes,

    the instantaneous load is normalized with respect to the cross section area of the tube. Crush

    stress were chosen to eliminate the influence of different cross-section area so that the effect

    of ellipticity ratio remains. Accordingly, Fig. 9 describes the variation of the instantaneous

    crush stress with ellipticity ratio. Experimental results show that the load carrying capacity

    at pre-crush failure stage is independent of the ellipticity ratio. On the other hand, the load

    carrying capacity at post crush failure stage is strongly sensitive to the ellipticity ratio. Thesame figure also shows the predicted crush stress using the ANN. The recurrent neural

    network (ENN) resulted in the best predictions and was therefore used here. The figure

    shows that these networks accurately predict the crushing stress behavior. They also

    accurately show the independency of the crush behavior of the elliptic tubes at the pre-crush

    Table 1 MSE for eccentricity ratio=1.25

    One Hidden Layer Two Hidden Layers

    Number of neurons per layer Number of neurons per layer

    6 8 10 12 16 6 8 10 12 16

    FFNN 0.0298 0.0243 0.0279 0.0242 0.0237 0.0181 0.0242 0.0146 0.0192 0.0173

    ENN 0.0218 0.0219 0.0194 0.0233 0.0294 0.0192 0.0174 0.0172 0.0194 0.0163

    MNN 0.0266 0.0238 0.0169 0.0266 0.0150 0.0258 0.0259 0.0165 0.0171 0.0142

    280 Appl Compos Mater (2008) 15:273285

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    9/13

    failure stage as well as the dependency on the ellipticity ratio during the post-crush failure

    stage; the same behavior identified by the experimental data presented in [ 11]. In spite of

    the fact that the ANN accurately predict the overall crush stress-strain behavior at various

    ellipticity ratios, one can notice a slight deviation in the predicted trend for the ellipticityratio of 1.25; this is especially clear for the crush strain range of 0.1 to 0.45. The deviation

    between the experimental behavior and the ANN prediction can be related to the dissimilar

    trend of the experimental data for the ellipticity ratio of 1.25 compared to all other

    experimental results. Since the ANN depends on the training data to predict the behavior at

    a certain ellipticity ratio, having this particular behavior different from all data used in the

    training set would understandably result in a deviation in the prediction.

    6 Prediction of Energy Absorption Capability

    Energy absorption capability during the structural crash is a requirement for the complete

    spectrum of passenger transport vehicles. The total work done during the axial crushing of

    the tubes is equal to the area under the load-displacement curve. Instantaneous specific

    energy absorption capability of the elliptical composite tube defined as the energy absorbed

    per unit mass was computed. Figure 10 shows the experimental relation between specific

    energy absorption and the deformation for the elliptic tubes with different eccentricity ratios

    [11]. The figure also shows typical predicted results obtained using artificial neural

    networks. Since, as mentioned before, load-deformation predictions using ENN were more

    accurate than those obtained with other types of ANN structures; they also resulted in the

    best energy absorption predictions and were therefore used here for comparison purposes.

    The average error in predicting the energy absorption capability for the ellipticity ratios of

    1.25, 1.5 and 1.75 were calculated to be 14.3%, 9.5% and 31.5% respectively. The

    percentage error obtained for the ellipticity ratio of 1.75 is greatly exaggerated by the high

    error obtained at the very low values of deformations. As the deformation increases, the

    Table 2 MSE for eccentricity ratio=1.50

    One Hidden Layer Two Hidden Layers

    Number of neurons per layer Number of neurons per layer

    6 8 10 12 16 6 8 10 12 16

    FFNN 0.0316 0.0237 0.0274 0.0254 0.0311 0.0150 0.0096 0.0192 0.0073 0.0337

    ENN 0.0256 0.0293 0.0299 0.0244 0.0188 0.0230 0.0180 0.0167 0.0160 0.0074

    MNN 0.0323 0.0311 0.0311 0.0263 0.0300 0.0260 0.0210 0.0107 0.0146 0.0151

    Table 3 MSE for eccentricity ratio=1.75

    One Hidden Layer Two Hidden Layers

    Number of neurons per layer Number of neurons per layer

    6 8 10 12 16 6 8 10 12 16

    FFNN 0.0170 0.0180 0.0110 0.0189 0.0114 0.0107 0.0156 0.0107 0.0065 0.0083

    ENN 0.0125 0.0180 0.0133 0.0155 0.0166 0.0163 0.0125 0.0188 0.0129 0.0198

    MNN 0.0249 0.0133 0.0166 0.0147 0.0211 0.0155 0.0156 0.0113 0.0103 0.0096

    Appl Compos Mater (2008) 15:273285 281

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    10/13

    Fig. 6 Typical experimental vs. predicted load-deformation behavior of axially-loaded composite tubes with

    an eccentricity ratio of 1.25

    Fig. 7 Typical experimental vs. predicted load-deformation behavior of axially-loaded composite tubes with

    an eccentricity ratio of 1.50

    282 Appl Compos Mater (2008) 15:273285

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    11/13

    Fig. 9 Crush stress-strain behavior of elliptical composite tubes: experimental results vs. ANN predictions

    Fig. 8 Typical experimental vs. predicted load-deformation behavior of axially-loaded composite tubes with

    an eccentricity ratio of 1.75

    Appl Compos Mater (2008) 15:273285 283

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    12/13

    error decreases significantly. Also, as expected, the ellipticity ratio of 1.5 which falls

    midway through the range of tested data gave the most accurate predictions. Figure 10

    shows that, in general, ANN can suitably predict the energy absorption characteristics of

    axially-loaded elliptic composite tubes.

    7 Conclusion

    The experimental behavior and corresponding ANN predictions of elliptical composite

    tubes subjected to axial compressive load were presented and discussed. The ANN has been

    shown to successfully predict the crushing behavior of a wide range of elliptic tubes. The

    predicted results obtained from the ANN were compared to actual experimental data in

    terms of load-carrying capacity, energy absorption capability and crushing load prediction,

    showing a very good agreement. In particular, the Elman Neural Network was shown to

    consistently lead to the best predictions of the experimental data. Additional work might be

    needed to determine whether ANN can also be used to accurately predict the crushing

    behavior of non-elliptical composite tubes.

    From the current work, it could be concluded that ANN techniques can be used toeffectively predict the response of composite energy absorber devices with elliptical cross-

    sections subjected to axial loading conditions.

    Acknowledgment The author would like to thank Dr. El-Sadig Mahdi, Associate Professor in the Kulliyyah

    of Engineering at the International Islamic University in Malaysia for providing the experimental data used in

    this work.

    Fig. 10 Specific energy absorption-deformation curves of elliptical composite tubes: experimental results vs.

    ANN predictions

    284 Appl Compos Mater (2008) 15:273285

  • 8/3/2019 Predicting the Crushing Behavior of Axially Loaded

    13/13

    References

    1. Mahdi, E., Sahari, B.B., Hamouda, A.M.S., Khalid, Y.A.: Effect of hybridisation on crushing behaviour

    of carbon/glass fibre/epoxy circular cylindrical shells. J. Mater. Process. Technol. 132, 4957 (2003)

    2. Bisagni, C., Di Pietro, G., Fraschini, L., Terletti, D.: Progressive crushing of fiber-reinforced composite

    structural components of a Formula One racing car. Compos. Struct. 68, 491503 (2005)

    3. Mahdi, E., Sahari, B.B., Hamouda, A.M.S., Khalid, Y.A.: On the axial collapse of cotton/epoxy tubes.

    Appl. Compos.Mater. 10, 6784 (2003)

    4. Mahdi, E., Mokhtar, A.S., Asari, N.A., Elfaki, F., Abdullah, E.J.: Nonlinear finite element analysis of

    axially crushed cotton fibre composite corrugated tubes. Compos. Struct. 75, 3948 (2006)

    5. Mamalis, A.G., Manolakos, D.E., Ioannidis, M.B., Papapostolou, D.P.: The static and dynamic axial

    collapse of CFRP square tubes: Finite element modeling. Compos. Struct. 74, 213225 (2006)

    6. Abosbaia, A.S., Mahdi, E., Hamouda, A.M.S., Sahari, B.B., Mokhtar, A.S.: Energy absorption capability

    of laterally loaded segmented composite tubes. Compos. Struct. 70, 356373 (2005)

    7. Mamalis, A.G., Manolakos, D.E., Ioannidis, M.B., Papapostolou, D.P.: On the response of thin-walled

    CFRP composite tubular components subjected to static and dynamic axial compressive loading:

    experimental. Compos. Struct. 69, 407420 (2005)

    8. Mahdi, E., Hamouda, A.S.M., Mokhtar, A.S., Majid, D.L.: Many aspects to improve damage tolerance

    of collapsible composite energy absorber devices. Compos. Struct. 67, 175187 (2005)

    9. Elgalai, A.M., Mahdi, E., Hamouda, A.M.S., Sahari, B.S.: Crushing response of composite corrugated

    tubes to quasi-static axial loading. Compos. Struct. 66, 665671 (2004)

    10. Mamalis, A.G., Manolakos, D.E., Ioannidis, M.B., Kostazos, P.K.: Crushing of hybrid square sandwich

    composite vehicle hollow bodyshells with reinforced core subjected to axial loading: numerical

    simulation. Compos. Struct. 61, 175186 (2003)

    11. Alkolose, O., Mahdi, E., Hamouda, A.M.S., Sahari, B.B.: Ellipticity ratio effects in the energy absorption

    of axially crushed composite tubes. Appl. Compos. Mater. 10, 339363 (2003)

    12. Mahdi, E., Alkolose, O., Hamouda, A.M.S., Shari, B.B.: Ellipticity ratio effects in the energy absorption

    of laterally crushed composite tubes. Adv. Comp. Mater. 15, 95113 (2006)

    13. Caliskan, A.G.: Prediction of the behavior of composite materials and structures using neural networks.Collection of Technical Papers - AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and

    Materials Conference 4, 29382946 (2001)

    14. Mahdi, E., El Kadi, H.: Crushing behavior of laterally compressed composite elliptical tubes:

    Experiments and predictions using artificial neural networks. Compos. Struct. 83, 399412 (2008)

    15. El Kadi, H.: Modeling the Mechanical Behavior of Fiber-Reinforced Polymeric Composite Materials

    Using Artificial Neural Networks A Review. Compos. Struct. 73, 123 (2006)

    16. Zhang, Z., Friedrich, K.: Artificial neural networks applied to polymer composites: a review. Compos.

    Sc. Tech. 63, 20292044 (2003)

    17. Schalkoff, R.J.: Artificial neural networks. McGraw-Hill (1997)

    18. Haykin, S.S.: Neural networks - a comprehensive foundation, 2nd edition. Prentice Hall, New Jersey

    (1999)

    19. Skapura, D.: Building neural networks. Addison-Wesley, New York (1996)20. Fausett, L.: Fundamentals of Neural Networks. Prentice Hall, New Jersey (1994)

    21. Neurosolutions 5 software: http://www.nd.com (2005)

    Appl Compos Mater (2008) 15:273285 285

    http://www.nd.com/http://www.nd.com/