Вы находитесь на странице: 1из 11

FINITE ELEMENTS WITH ARTIFICIAL INTELLIGENCE

Stafanos Drakos1, Hyu-Soung Shin2 and Gyan N. Pande3 1 Director, DG Consulting Engineers, Department of Geotechnical Engineering, Rhodes, Greece 2 Senior Researcher, Geotechnical Engineering Research Department, Korea Institute of Construction Technology, ilsan-Gu, Kyunggi-Do, South Korea 3 Professor Emeritus, Centre for Civil & Computational Engineering, University of Wales Swansea, Singleton Park, Swansea, SA2 8PP, UK E-mail: g.n.pande@swan.ac.uk ABSTRACT
This paper deals with two aspects of Neural Network based Constitutive Models. Firstly, it demonstrates how a complex stress-strain model, such as the Hardening Soil Model of PLAXIS can be cast in the form of a neural network. The main advantage of this strategy is computational efficiency. Secondly, it shows how an intelligent finite element code can determine various elastic and strength parameters of a material from the analysis of a single structure having a nonuniform state of stress and strains. It is conjectured that such a strategy may be superior to conducting a number of different types of tests to determine specific parameters/properties. KEYWORDS Artificial Intelligence, FE analysis, Constitutive Modelling, Material Parameter Identification. 1. INTRODUCTION Numerical methods such as the finite element method play an important role as a tool for analysing a variety of problems in engineering. One of the most crucial components of the finite element analyses is the constitutive model used for representing the mechanical response of the material(s) involved. Phenomenological models of complex materials such as soils, concrete, composites etc. are formulated within an assumed mathematical framework and involve determination of a multitude of material parameters. It is generally admitted that in spite of considerable complexities of constitutive theories proposed, it has not been possible to capture the material response universally along all complex stress paths under a wide range of confining pressures. Furthermore, the complexities of constitutive models, in many cases, have inhibited their incorporation in general purpose finite element codes; thus restricting their usefulness in engineering practice. Motivated by the above considerations, a radically different framework for developing constitutive models is proposed in this paper. In recent years a number of applications of Artificial Neural Networks (ANNs) leading to Neural Network based Constitutive Models (NNCMs) have been proposed by a number of researchers [1, 2, 6]. There are three different ways in which NNCMs can be used by the engineers and analysts with advantage. These are:

(1) Firstly, NNCMs can be developed for any material from the raw test data without invoking any constitutive theory [3, 6]. This approach has many advantages, the most important being that one does not necessarily have to identify material parameters of the model. However, if an engineer does need to identify them to get a feel, they can be identified by carrying out what is known as virtual tests. (2) Secondly, since many constitutive models required to represent real material behaviour are very complex, incorporating them in a finite element code and using them for solving real life problems may not be a trivial task. Here, using synthetic data generated from systematic exploration of strain space and corresponding stress response, a NNCM can be trained and plugged in a Finite Element (FE) code. This leads to computational efficiency since the computation of stress increment for a strain increment from a trained NNCM is instantaneous. (3) Thirdly, NNCMs can be trained from load displacement data of structures. By the term structures, we mean solids of arbitrary shape subjected to monotonically increasing loads having a non-uniform states of strain and stress. Thus, a cylindrical specimen having glued platens, subjected to uniaxial load is a structure. This methodology leads to intelligent finite elements [2] and is advantageous for condition monitoring of real structures. It can also be used to identify material parameters for complex materials such as masonry [7]. This paper deals with the second and third categories of applications as described above. Firstly, it is demonstrated how synthetic data from any constitutive model can be used to successfully train a NNCM. We choose the example of the Hardening Soil Model (HSM) of the well-known commercial code PLAXISTM for illustration. Secondly, we present a methodology through which a NNCM embedded in the FE code can be trained from the monitored data of load and displacement of a structure made of the corresponding material. This methodology leads to a third generation (3G) intelligent self-learning FE codes in which the disparity between predicted and monitored response of a structure can be minimised. The main advantage of this methodology is that as further monitored data become available and re-analyses are performed, NNCM learns and is able to detect the pattern of aging or degradation. The layout of the paper is as follows: Section 2 gives brief details of NNCM adopted for training and methodology for the generation of synthetic data from a known constitutive law. Results of the response of the trained NNCM are given in section 3 for various stress paths experienced at certain points beneath a foundation. Section 4 gives the outline of a self-learning finite element code in which embedded NNCM is recursively trained to match the observed response of a structure. Examples of successful training and capabilities of NNCMs are given in Section 5 followed by conclusions in Section 6. 2. ARTIFICIAL NEURAL NETWORKS Artificial Neural Networks are pattern recognition algorithms using which relationship between a set of causes and effects can be captured. Any set of numeric data can be used to discover the pattern in it. There are a large number of applications of this methodology in almost all disciplines of science. Here we restrict ourselves to the application of ANN in constitutive modelling. 2.1 Neural network based constitutive model - methodology and architecture In the field of constitutive modelling, strains and stresses or their increments are used as causes and effects respectively. In many geotechnical problems, geomaterials are subjected to cyclic and transient loads. For example, the foundations of an offshore structure are subjected cyclic loads. Moreover, even in quasi-static problems many elements are subjected to unloading and reloading. For constitutive models to be valid, in such situations, it is proposed to adopt intrinsic time [10]

or length of strain trajectory, , as an independent input parameter. Mathematically, , is defined as follows: (1) L = d L Where, d is an increment of deviatoric strain defined as

d L = d ij ij kk

(2)

Intrinsic time, , is a monotonically increasing positive parameter. The above definition can be changed to include volumetric strain as well. The NNCM adopted here [9] has strain increments, stresses and intrinsic time as the input variables. Increments of stress are the output variables. For two- dimensional analyses, the optimal architecture of the NN is presented in Figure 1. It is constituted by 9 input nodes two hidden layer of 18 and 8 nodes respectively and 4 output nodes (Fig. 1). In order to train the neural network the resilient back propagation (RPROP) algorithm is used [5, 8]. It is a local adaptive learning scheme based on the standard back propagation framework.
xn yn n z n

xn-1 yn-1 n-1 zn-1 dx dy d dz L

Fig. 1. Architecture of NNCM [9-18-8-4] for two-dimensional analysis 2.2 Data enrichment The strain-stress pairs from the triaxial tests are in terms of principal stresses and strains since no shear stresses/strains are involved. If such data were used for training, NNCM will have to extrapolate wherever shear stress/strain components are involved. This will lead to large inaccuracies in stress-strain response of the NNCM. To overcome this limitation, a data enrichment strategy was proposed by Shin & Pande [6]. They recommend transforming the strainstress data into different local axes, which gives rise to shear strain and shear stress components which are then included in the data set for training. 2.3 Data generation In order to train a NNCM a number of data sets are needed. In place of using real experimental data for any material, we have chosen to generate data (these are termed as synthetic data) from the well known constitutive model, the Hardening Soil Model (HSM) available in the commercially available software, PLAXISTM. For this purpose typical parameters for sand have been arbitrarily chosen and are shown in Table 1. The triaxial test paths corresponding to conditions of Loading in Compression (LC), Loading in Extension (LE), Unloading in Compression (UC) and Unloading in Extension (UE) have been simulated under stress controlled drained conditions using PLAXIS software with HSM. Three different confining pressures of 50, 100 and 150 kPa were used for each of the above stress paths. In engineering practice, if such test data were available for a soil, one would perhaps term them as extensive. These stress-strain data obtained from these simulations were then used for training the NNCM.

Plots of q versus axial strain, yy, as predicted by the trained NNCM and the original HSM data used for training, for various stress paths are shown in Fig. 2. An excellent match is seen confirming that NNCM has been adequately trained Table1. Arbitrary sand parameters E50ref (for pref=100): 20000, Eurref (for pref=100): 60000, Eoedref (for pref=100): 20000 Cohesion c=0, Friction angle =30, Dilatancy angle =0, Poisson ratio v=0.2 Power m 0.5 Konc 0.5 Tensile strength 0 Failure ratio 0.9

250

80 70

200 LE 150 LC
60 50 UE UC

q
100 HSM NNCM

q
40 HSM 30 20 10 NNCM

50

0 -0.06

-0.04

-0.02

0.00

0.02

0.04

0.06

0.08

0 -4.0E-02

-3.0E-02

-2.0E-02

-1.0E-02

0.0E+00

1.0E-02

2.0E-02

3.0E-02

4.0E-02

yy

yy

Fig. 2: Graph of q versus yy under LC and LE & UC and UE conditions, respectively, for a confining pressure of 100 kPa 3. STRAIN PATHS AT VARIOUS POINTS UNDER A FOUNDATION The excellent match between NNCM prediction and the actual data is not surprising, since predictions are being made for the same set of darta as were used in the training of the ANN. In this section we will compare the response in a problem of practical importance. In Fig. 3 the geometry of a typical foundation problem is presented using the sand data given in Table 1. Six noded triangular isoparametric elements have been used for the analysis. A uniform load of 150 kPa has been incrementally applied and the resulting strain stress curves at two monitoring points, A & B, marked on the Fig. 3, for NNCM and HSM are shown in Fig 4.

Figure 3: Geometry and monitoring points of the foundation problem

80 70 60 xx, yy 50 40 30 20 10 0 -6.E-03 -4.E-03 -2.E-03 0.E+00 2.E-03 4.E-03 6.E-03 8.E-03 1.E-02 1.E-02 xx, yy HSM NNCM
xx

160 140
yy

120 ,yy 100 80 60 40 20 0 -0.0040 -0.0020 0.0000 0.0020 xx, yy

yy

HSM NNCM

0.0040

0.0060

0.0080

0 -5 -10 txy -15 -20 -25 -30 -3.0E-02 -2.5E-02 -2.0E-02 -1.5E-02 -1.0E-02 -5.0E-03 0.0E+00 xy
35 30 25 q 20 15 10 5 0 0 10 20 30 p 40 50 60 HSM NNCM
q

0 -5 -10

HSM NNCM

txy

-15 -20 -25 -30 -35 -8.0E-03 -6.0E-03 -4.0E-03 xy -2.0E-03 0.0E+00

HSM NNCM

80.0 70.0 60.0 50.0 40.0 30.0 20.0 10.0 0.0 0 20 40 p 60 80 100 HSM NNCM

Fig. 4: Various stress paths at point A & B respectively 4. ENHANCED FE CODES WITH ARTIFICIAL INTELLIGENCE An intelligent FE code can be essentially developed by incorporation of a NNCM in a standard FE code. The nonlinear incremental analyses and learning are undertaken using a two-step staggered scheme in which in the first step, the boundary value problem is solved using a pre-primed NNCM (an arbitrary linear tangential stiffness matrix (TSM), for example). This gives the n computed displacement field from which a vector of displacements for the monitored points ( c ) can be assembled. In the following step, a second boundary value problem (conjugate problem) is solved in which displacements corresponding to the delinquent displacement vector ( n ) defined d as:
n n = c n d m

(3)

are applied. Here, n is the vector of monitored displacement at the nth monitoring point. The m stresses and strains at each Gauss point and at each increment constitute a set of data for training of the NNCM. The intelligent FE code can be used for structural analysis with pre-trained NNCM, or alternatively can be used for training the existing NNCM from structural tests. In order to adjust the current stress-strain sets into proper direction for target constitutive relationships, there are two schemes investigated in this study, i.e. a stress correction scheme and a strain correction scheme. Basically the two schemes differ in the manner that the stress or strain components are corrected to

correspond to the stress-strain law for the current load increment and are equivalent to the initial stress or initial strain corrections in the well-known methodology of nonlinear FE analysis. Based on either of two schemes, the currently recognised stress-strain training sets become stationary and consistent with the response of the structure at monitoring points within limits of tolerance based on a error measure, specified below:

1 lk

(
l k n =1 i =1

im - ic

im

(4)

where l and k are the number of total load increments and monitoring points, respectively. The is used to judge whether the error between the computed and monitored displacements has reached an acceptable level, otherwise, retraining of the NNCM has to be undertaken. It is noted that the error, as defined by Eqn. (4) is also dependent on the number of monitoring points. Thus, even if the error is small after some cycles of self-learning, it is admitted that the currently trained NNCM may not be consistent with the actual constitutive behaviour of the material, specially in three-dimensional problems. Selection of the number and location of monitoring points is therefore of considerable importance in identifying a valid NNCM for any material from the structural tests. The details for the self-learning mode of the intelligent FE code can be found in the previous publication [3] which also provides some illustrative examples to verify the computational strategy. 5. NUMERICAL EXAMPLES Here we give two examples of application of intelligent finite element simulations, viz. a cylindrical specimen with rigid glued platens under uniaxial compression and a panel subjected to a uniform load. 5.1 Cylindrical rock specimen with glued platens It is well known that this arrangement leads to severe end-effects giving rise to a nonuniform state of stress, which is an issue in material testing requiring elaborate arrangements to minimise friction. In this case, when loaded in triaxial compression, the failure takes place at a stress level lower than implied by real F Axi-symmetric FE model strength of the material. On the model Stiff plate shown in Fig. 5, in general, a number of monitoring points are chosen on the surface of the structure in certain directions. In this specific case, displacemens are measured at a monitoring point located on the stiff plate in the vertical (loading) direction and at three monitoring points on vertical surface in the Specimen radial direction as indicated in Fig. 5. Using 50 mm the radial displacements measured at the : Monitoring point Stiff plate three monitoring points, a rational Fig. 5: A sample with glued end platens polynomial interpolation
10 10 Inhomogeneous displacements 45

Glued

is used to approximate radial displacements at other positions of the surface. It is obvious that a larger number of monitoring points would guarantee more consistent stress-strain relationships extracted from the self-learning process. The authors did not wish to divert their attention to an experimental programme to obtain monitored displacement data. It was therefore decided to carry out a conventional finite element analysis of the model (geometry as shown in Fig. 5) using material parameters indicated as given in Table 2. A quarter symmetric model is adopted with axisymmetric conditions applied. Mohr-Coulomb yield criteria with assosiated flow rule and no

45

(mm)

strain hardening is assumed. The failure load from the analysis was observed as 141 kN (failure stress = 18 MPa) after 18 load increments. It is noted here that the failure stress observed should be compared with the uniaxial compressive strength (=19.2 MPa) which is simply calculated from a cohesion and a internal friction angle given in Table 2. TABLE 2 MATERIAL PROPERTIES ASSUMED FOR OBTAINING MONITORED DISPLACEMENT DATA Parameter Young's modulus (E), Mpa Poisson's ratio () Cohesion (c), Mpa Internal friction angle (), degrees Assumed value 20000 0.35 5.0 35.0

Displacements at various monitoring points at different load levels, resulting from the FE analysis are assumed as physically measured data from the structural test. These are the only data needed for training the NNCM embedded in the intelligent FE code. Some arbitrary data are, however, used for priming the neural network. This is similar to the situation in initial starting of an iterative process. The values of E and arbitrarily adopted for priming were 10,000 MPa and 0.4 respectively. The self-learning process begins with the code predicting displacement field based on the trained NNCM in the previous cycle and comparing it with the monitored displacement field at the corresponding load level. The self-learning procedure is continued until the monitored load-displacment responses at monitoring points are captured within a specified tolerance given by Eqn. (4):
160
Failure point 2.89% 4.39% 18.8% =49.7%
Initial stage After 1st SL After 2nd SL after 3rd SL after 4th SL Target

120

80

Target
e1 1 5.0E-04 4.6E-04 4.3E-04 3.9E-04 3.6E-04 3.2E-04 2.9E-04 2.5E-04 2.1E-04 1.8E-04 1.4E-04 1.1E-04 7.1E-05 3.6E-05 0.0E+00

After 4th SL

40

0 0.05

After 1th SL

Initial state

-0.05

-0.1

-0.15

-0.2

Displacement (mm) at the vertical monitoring point

Fig. 6: Self-learning processes: the contour plots is for major principal strain

The vertical load-vertical displacement relationship at monitoring point on the stiff plate is presented in Fig. 6. It can be seen that at the 4th cycle of self-learning process structural responses was captured with an error of ( =2.89 %). Experience shows that this order of tolerance is acceptable and NNCM assembled at the 4th cycle of selflearning process is accurate enough to represent constitutive relationships of the real material. The insets in the Fig. 4 show contours of major principal strains for each cycle of self-learning process. These contours also indicate that after the 4th cycle of self-learning strain distribution in the structure is close to one obtained when the structure was analysed with the conventional finite element code with assigned material parameters.

Applied load (kN)

5.2 Identification of the uniaxial compressive and tensile strengths The constitutive models embedded in FE codes are usually in the strain controlled format. This implies that stress increments can be readily computed from the given strain increments. In the intelligent FE code also, the NNCM embedded has strain components as input or cause and stress components as output or effect. To identify the compressive or tensile strength of the material, an obvious way is to run the intelligent FE code with the trained NNCM embedded in it in purely compressive or tensile mode. Here the finite element model has just one element and smooth

boundaries to induce a uniform state of stress. Thus, uniaxial stress-strain curves in tension and compression are straightforwardly obtained from virtual tests from which compressive and tensile strengths can be identified.
20 Target value

0.40
2.5% 4%

Initial value 26.7%

E (GPa)

18

16 14 12 10 Initial value
1

15% Error = 26.5%

0.35

20% 3.3% 0%

0.30
2 3 4

Target value
1 2 3 4

Cycles of Self-learning process

Cycles of Self-learning process

Fig. 7: Evolution of elastic constants during run of the self-learning mode of an intelligent FE code. It is seen from Figure 8 that the uniaxial compressive and tensile strengths obtained are 19.2 MPa and 1.2 MPa respectively. The compressive strength obtained is remarkably close to the strength implied by c and parameters given in Table 2. However, the tensile strength obtained in this case is considerably below (approximately 77%) the actual value (5.2 MPa) implied by c and parameters.
-20
Actural compressive strength (= 19.2 MPa)

6 5
Computed tesile strength (= 5.2 MPa)

-15

Observed compressive strengh (= 18.0 MPa)

Stress (MPa)

-10

Stress (MPa)

4 3 2

extrapolated from NNCM

-5
abserved on specimen with 'end-effect' predicted from NNCM

Failure stress (= 0.9 MPa)

1 0 0.00

0 0.0 -0.5 -1.0


-1

0.02

0.04

0.06
-1

0.08

Strain (10 , %)

Strain (10 , %)

Figure 8: Determination of a uniaxial compressive strength (left) and tensile strength (right). There is therefore some evidence to indicate that once an NNCM embedded in a FE code has been trained on the basis of load displacement data of a structure, the code can be used to predict strengths in various modes of failure. However, the accuracy of the prediction is primarily dependent on the stress paths followed by various monitoring points in the structural tests. Thus, for example, the points at the middle of the specimen are subjected to a state of compressiontension-tension whilst the region close to the rigid platens is subjected to triaxial compression. Since the mode of failure of the structure is in compression, constitutive information in tension is not rich or comprehensive and neural network has to perform an extrapolation, which admittedly is an inaccurate process. This readily explains the considerable error in the prediction of tensile strength. 5.3 A plane stress panel with a circular cavity subjected to uniform load For heuristic purposes, a simple structure viz. a plane stress masonry panel with a circular hole at its contre is chosen. The panel which is assumed to be made of an orthotropic material with known principal material axes. Fig. 9 shows the geometry, boundary and loading conditions. It is well known that severe stress concentrations take place around the cavilty. A number of monitoring points at convenient locations have to be chosen. It is obvious that larger the number of monitoring points more accurate stress-strain relation can be extracted from the NNCM.

Although, collecting data for a large number of monitoring points may lead to higher costs, it should be noted that modern instrumentation techniques with automatic data-loggers make such measurements relatively easy to conduct. Here again, the authors used synthetic computed load displacement data instead of monitored displacement data. A 3-D FE analysis of the panel configuration using assumed values of the nine independent orthotropic elastic constants as given in Table 3, was carried out. A threedimensional analysis, instead of a two dimensional plane stress analysis, was intentionally carried out so as to prove that the methodology is generic and embedded NNCM can be trained from load displacement data. Here elastic parameters were also identified once the NNCM had been fully trained. A quarter symmetric model was discretised with 578 nodes and 70 isoparametric 20 noded brick elements1. Displacements of 66 nodes on the surface of the model at 5 load levels, resulting from the FE analysis were assumed as physically measured data from the field test, although a much smaller number of monitoring points would be adequate as shown [2]. These are the only data needed for training the NNCM embedded in the self-learning FE code. It is noted that although the material is linear with proportional load-displacement relationship, the NNCM cannot recognise this unless incremental data are given.

Table 3. Arbitrarily chosen orthotropic elastic constants. Elastic constants Ex (MPa) Ey (MPa) Ez (MPa) Gxy (MPa) Gyz (MPa) Assumed values 650.4 967 1735.4 582.2 203.6 694.5

2.5m

Y
0.75m

Y
0.375m 2.125m

Gxz (MPa)

xy yz xz

0.269 0.139 0.13

: Displacement monitoring point in a certain direction

Fig. 9. A panel with circular cavity at its centre.

As mentioned in section 5.1, some initial data are required for priming the NNCM. In this example, the values of isotropic elastic constants, E = 1.0 GPa and = 0.2 were arbitrarily adopted for priming. At the 3rd cycle of self-learning process, the overall discrepency between predicted and monitored displacement field reduced from 37.21% to 2.76%. 5.4 Identification of orthotropic elastic constants The constitutive matrix (DNNik) can be assembled from the first derivative of the ith effect of the NN with respect to kth cause. strain and stress vectors defined as: i = {1 , 2 , 3 , 12 23 31}, k = { 1 , 2 , 3 , 12 23 31} (5) are introduced as input nodes oi and output nodes ok in the following Eqn. respectively: ok SLi ; DNN ik (oi , ok ) = oi SLk

It is noted that the accuracy of the FE model adopted is not an issue here

SLi =

1 , ( Max (oi ) Min (oi ))

SLk =

1 ( Max (tk ) Min (tk ))

(6) where DNNik is a matrix of size (I K) composed of partial derivatives of the trained BPNN. SLi and SLk are linear scale factors for the output variables oi and target variables tk, respectively. Here, I and K are the maximum values of causesand effects, respectively. Then the derivative, DNNik of the NNCM is the TSM for a material at a certain given strain state, i. It can be decribed as follows: k DNN = DNNik ( , ) = (7) i At the beginning of each iteration of the self-learning FE analysis, constitutive matrices at all Gauss points are computed using Eq. (7). These are compared to the form of the conventional orthotropic elasticity matrix as the material assumed for this example is an orthotropic material with material axis coinciding with the axis of the model. The elastic constants are simply determined by forming a set of equations with elastic parameters as unknowns. It is noted that NN based constitutive matrices do not invoke any symmetry and as many as 36 elastic constants can be evaluated. The elastic constants are evaluated at the middle of each load increments of each cycle of the self-learning procedure. Table 4 shows the evolution of nine orthotropic elastic constants with the number of cycles of re-training of the NNCM embedded in the FE code. At the 3rd cycle of selflearning process all the values of the orthotropic constants were less than 3.8% in error against the target values. It is found from numerical experiments that in practical problems tested so far no more than five self-learning cycles are required. Table 4. Prediction of orthotropic elastic constants after each self-learning process.
Ratios Ex,p / Ex,a Ey,p / Ey,a Ez,p / Ez,a Gxy,p / Gxy,a Gyz,p / Gyz,a Gxz,p / Gxz,a xy,p / xy,a yz,p /yz,a xz,p /xz,a (%) Prediction after 1st selflearning 103.4% 75.67% 65.39% 42.82% 108.4% 64.78% 72.86% 25.9% 111.54% 37.21% Prediction after 2nd self-learning 100.06% 98.55% 74.63% 69.36% 106.1% 67.86% 99.63% 83.45% 113.85% 13.93% Prediction after 3rd self-learning 100.24% 99.02% 103.46% 101.53% 99.95% 102.95% 101.49% 99.28% 96.15% 2.76%

Denominators, p and a indicate predicted and assumed values, respectively.

6. SUMMARY AND CONCLUSIONS Artificial intelligence is likely to play a crucial role in engineering analysis and design in the future. Constitutive modelling of materials which has engaged the attention of a large number of researchers in geotechnical engineering for many decades is reaching a stage when very complex models are being proposed and used by engineers. Artificial Neural Networks provide a viable alternative to modelling within various mathematical frameworks. It has been shown in this study that many material parameters can be obtained from a single structural test. After the 3rd cycle of re-training of the NNCM, the nine elastic parameters of the orthotropic material were computed with the maximum error in any parameter being less than 4%.

Many critical structures such as dams and nuclear structures etc. are monitored frequently. Increase or decrease of displacements at the monitored points is frequently observed. From these results, it is difficult to judge the health or condition of the structure. The methodology of parameter identification for anisotropic elastic constants proposed here is valid for fully anisotropic materials although the example of an orthotropic material has been provided in this paper. Although, a rather simple test configuration of plane stress panel with a central cavity was used for illustration, the methodology is generic and elastic constants can be identified from instrumented load-displacement data of three-dimensional bodies. The discussed methodology of determining, material parameters can be used to provide an indication of the health of the structure. In view of the existing technology, it is believed that testing of small scale structures of arbitrary shape is more convenient to determine properties of materials from a single test as compared to numerous conventional material tests in which the state of stress is homogenous. More research, however, is needed to confirm this. REFERENCES [1] Ghaboussi, J., Garrett, J. H. & Wu, X. (1991). Knowledge-based modelling of material behavior with neural networks, Journal of Engineering Mechanics, ASCE 117:1, 132-153. [2] Shin, H. S. and Pande, G. N. (2000). On self-learning finite element code based on monitored response of structures. Computers and Geotechnics 27, 161-178. [3] H.S. Shin & G.N. Pande, (2001a), Intelligent finite elements, Proceeding of Asian-Pacific Conference for Computational Mechanics (APCOM'01), edited by S. Valliappan and N. Khalili, Sydney, Australia, pp. 1301-1310. [4] H.S. Shin & G.N. Pande, (2001b), Intelligent finite elements in masonry research, Proceeding of International Symposium on Computer Methods in Structural Masonry (STRUMAS V), edited by T.G. Hughes & G.N. Pande, Rome, Italy, pp. 221-230. [5] Shin, H.S. (2001c) Neural Network Based Constitutive Models for Finite Element Analysis, Ph.D. thesis, Department of Civil Engineering, University of Wales Swansea. [6] Shin, H. S. and Pande, G. N. (2002). Enhancement of data for training neural network based constitutive models for geomaterials, Proceeding of The Eighth International Symposium on Numerical Models in Geomechanics (NUMOG VIII), edited by G.N. Pande & S. Pietruszczak, Rome, Italy, pp. 141-146. [7] H.S. Shin, G.N. Pande, Identification of elastic constants for orthotropic materials from a structural test, Computers and Geotechnics (International Journal-SCIE), Vol 30, No. 7, pp. 571577, 2003.10. [8] Riedmiller, M. and Braun, H. (1993), A direct adaptive method for faster backpropagation learning: the RPROP algorithm, Proceedings of the IEEE International Conference on Neural Networks, San Francisco, CA, March 28-April 1. [9] S. I. Drakos, G. N. Pande, A Neural Network Equivalent of Hardening Soil Model of PLAXIS, NUMGE06, Graz, Austria, pp [10] Valanis, K.C. & Read, H. E., A new endochronic plasticity model for soils, Chapter 14, Soil Mechanics Transient &Cyclic Loads, Edited by G. N.Pande & O. C. Zienkiewicz, J. Wiley & Sons, 1982.

Вам также может понравиться