0 évaluation0% ont trouvé ce document utile (0 vote)
10 vues10 pages
Genetic algorithms have been the subject of considerable interest in recent years, since they appear to provide a
robust search procedure for solving dicult problems. Due to the way the genetic algorithm explores the region of
interest it avoids getting stuck at a particular local minimum and locates the global optimum. The genetic algorithm
is slow in execution and is best applied to dicult problems. This paper applies a genetic algorithm to the problem
of damage detection using vibration data. The objective is to identify the position of one or more damage sites in a
structure, and to estimate the extent of the damage at these sites. The genetic algorithm is used to optimize the
discrete damage location variables. For a given damage location site or sites, a standard eigensensitivity method is
used to optimize the damage extent. This two-level approach incorporates the advantages of both the genetic
algorithm and the eigensensitivity methods. The method is demonstrated on a simulated beam example and an
experimental plate example
Titre original
A Combined Genetic and Eigensensitivity Algorithm for The
Genetic algorithms have been the subject of considerable interest in recent years, since they appear to provide a
robust search procedure for solving dicult problems. Due to the way the genetic algorithm explores the region of
interest it avoids getting stuck at a particular local minimum and locates the global optimum. The genetic algorithm
is slow in execution and is best applied to dicult problems. This paper applies a genetic algorithm to the problem
of damage detection using vibration data. The objective is to identify the position of one or more damage sites in a
structure, and to estimate the extent of the damage at these sites. The genetic algorithm is used to optimize the
discrete damage location variables. For a given damage location site or sites, a standard eigensensitivity method is
used to optimize the damage extent. This two-level approach incorporates the advantages of both the genetic
algorithm and the eigensensitivity methods. The method is demonstrated on a simulated beam example and an
experimental plate example
Genetic algorithms have been the subject of considerable interest in recent years, since they appear to provide a
robust search procedure for solving dicult problems. Due to the way the genetic algorithm explores the region of
interest it avoids getting stuck at a particular local minimum and locates the global optimum. The genetic algorithm
is slow in execution and is best applied to dicult problems. This paper applies a genetic algorithm to the problem
of damage detection using vibration data. The objective is to identify the position of one or more damage sites in a
structure, and to estimate the extent of the damage at these sites. The genetic algorithm is used to optimize the
discrete damage location variables. For a given damage location site or sites, a standard eigensensitivity method is
used to optimize the damage extent. This two-level approach incorporates the advantages of both the genetic
algorithm and the eigensensitivity methods. The method is demonstrated on a simulated beam example and an
experimental plate example
A combined genetic and eigensensitivity algorithm for the
location of damage in structures
M.I. Friswell a, *, J.E.T. Penny b , S.D. Garvey b a Department of Mechanical Engineering, University of Wales Swansea, Swansea SA2 8PP, U.K. b Department of Mechanical and Electrical Engineering, Aston University, Birmingham B4 7ET, U.K. Received 20 January 1997; accepted 28 April 1998 Abstract Genetic algorithms have been the subject of considerable interest in recent years, since they appear to provide a robust search procedure for solving dicult problems. Due to the way the genetic algorithm explores the region of interest it avoids getting stuck at a particular local minimum and locates the global optimum. The genetic algorithm is slow in execution and is best applied to dicult problems. This paper applies a genetic algorithm to the problem of damage detection using vibration data. The objective is to identify the position of one or more damage sites in a structure, and to estimate the extent of the damage at these sites. The genetic algorithm is used to optimize the discrete damage location variables. For a given damage location site or sites, a standard eigensensitivity method is used to optimize the damage extent. This two-level approach incorporates the advantages of both the genetic algorithm and the eigensensitivity methods. The method is demonstrated on a simulated beam example and an experimental plate example. # 1998 Elsevier Science Ltd. All rights reserved. Keywords: Damage location; Genetic algorithm; Eigensensitivity 1. Introduction Genetic algorithms are now frequently applied to problems of maximising or minimizing a given objec- tive function, often subject to some constraints. Genetic algorithms have been applied to a wide range of optimization problems in engineering which have this form, such as the design of structures [14], pipe- line optimization [5], nite element model updating [6 8] and the problem of choosing the number and lo- cation of actuators for the control of large exible space structures [9]. This paper approaches the pro- blem of damage location by formulating an objective function composed of measured and analytical vi- bration data and applying a genetic algorithm to opti- mize it. Other evolutionary strategies [10] and algorithms such as simulated annealing [7, 11] have been used for design optimization and model updating. In recent years, there has been a considerable demand for more accurate techniques to detect and locate damage, particularly in large structures. Damage will cause the stiness distribution in the structure to change which may be detected by measur- ing its dynamic response. Doebling et al. [12] gave an overview of damage detection and location. Three dis- tinct philosophies have been considered to locate the damage in a structure using measured vibration data. The rst group of methods uses techniques of nite el- ement model updating and error localization to obtain a corrected set of physical parameters that reproduce the measured data. Mottershead and Friswell [13] gave a survey of model updating methods, many of which can also be used for damage location. The direct model updating approach, where the whole stiness matrix is updated, does not require parameterization of the nite element model [14]. Otherwise, the model Computers and Structures 69 (1998) 547556 0045-7949/98/$ - see front matter # 1998 Elsevier Science Ltd. All rights reserved. PII: S0045- 7949( 98) 00125- 4 PERGAMON * Corresponding author. E-Mail: M.I.Friswell@swansea. ac.uk. is parameterized and all the parameters updated [15, 16]. Generally, the identication will be underdetermined and some form of regularization should be employed. Genetic algorithms have also been used to identify damage in structures by optimiz- ing the force residual [17] or modal residuals [18]. A second group of methods, called forward methods, assume a candidate set of possible damage scenarios, which can include both damage mechanism and location. The change in dynamic response due to the damage is predicted, usually in terms of the natural frequencies. The predicted change in response of the structure for all the damage scenarios is then compared to the measured change and the closest damage case is chosen. Statistical methods are often used to identify the most likely location of the damage. The dierence between these methods and those from standard model updating is that only a subset of parameters are assumed to be in error. Cawley and Adams [1921] used this type of method and found that damage in specimens fabricated from composite materials could be detected. The method of Cawley and Adams has the advantage that damage location is identied separ- ately from damage extent, although this only works for a single damage site. Friswell et al. [22] proposed a similar method but used a more rened statistical analysis. A similar approach is to directly nd the best subset of parameters that models the damage [23]. Neural networks are able to treat damage mechan- isms implicitly, so that it is not necessary to model the structure in so much detail [2430]. The method can easily deal with non-linear damage mechanisms. Models are still required to provide the training cases for the networks, and the algorithms should be robust against systematic errors between the model used for training and the actual structure. For success, neural networks require that the essential features in the damaged structure are represented in the training data. 2. Overview of genetic algorithms The striking feature of genetic algorithms is that they are based on ideas from the science of genetics and the process of natural selection. This cross fertili- zation from one eld of science to another has led to stimulating and fruitful applications in many elds and particularly in computer science. The genetic algorithm works with an initial popu- lation which may, for example, correspond to numeri- cal values of a particular variable. The size of this population may vary and is generally related to the problem under consideration. The members of this population are usually strings of zeros and ones i.e. binary strings. For example a small initial or rst gen- eration population may take the form: 1000010 1110000 1010101 1111001 1000001: In practice, the population may be far larger than this and the strings longer. The strings themselves may be the encoded values of a variable or variables that we are examining. This initial population is generated ran- domly and we can use the terminology of genetics to characterize it. Each string in the population corre- sponds to a chromosome and each binary element of the string to a gene. A new population must now develop from this initial population and to do this we implement the analogue of specic fundamental genetic processes. These are: . reproduction based on tness; . crossover; . mutation. A set of chromosomes is selected at the reproduction stage based on natural selection. Thus, members of the population are chosen for reproduction on the basis of their tness dened according to some specied cri- teria. The ttest are given a greater probability of reproducing in proportion to the value of their tness. The actual process of mating is implemented by using the simple idea of crossover. This means that two members of the population exchange genes. There are many ways of implementing this crossover, for example having a single crossover point or many cross- over points. These crossover points are selected ran- domly. A simple crossover is illustrated below for two chromosomes selected according to tness. Here we have randomly selected a crossover point after the fourth digit. 1110 + 000; 1010 + 101: After crossover this gives 1110 + 101; 1010 + 000: Applying this procedure to our original population we produce a new generation. The nal process is mutation. Here, we randomly change a particular gene in a particular chromosome. Thus, 0 may be changed to a 1 or vice versa. The pro- cess of mutation in a genetic algorithm occurs very rarely and hence this probability of a change in a string is kept very low. Mutation ensures genetic diver- M. Friswell et al. / Computers and Structures 69 (1998) 547556 548 sity is maintained by generating new, random chromo- somes. The application of genetic algorithms to optimiz- ation problems is relatively clear since the test of t- ness can be based on the function we are trying to maximize, thus the members of the population which give the largest values of the objective function are taken as the ttest. The chromosomes are viewed as a concatenation of binary strings representing the values individual variables in the optimization problem. Continuous variables must be discretized, as described later. Constraints are usually incorporated into the objective function using a penalty function which ensures that during the optimization process the sol- ution will satisfy the constraints. This penalty must be carefully chosen to ensure the constraints are satised and so the correct optimum is reached. The reason why a genetic algorithm diers from a simple direct search procedure is that it involves two special features: crossover and mutation. Thus, starting from an initial population the algorithm develops new generations which rapidly explore the region of inter- est. This is useful for dicult optimization problems and in particular for those where we wish to nd the global maximum or minimum of function which has many local maxima and minima. In this case, standard optimization methods, such as the conjugate gradient method, can locate only the local optimum. A genetic algorithm, however, may locate the global optimum although this is not guaranteed. This is due to the way it explores the region of interest avoiding getting stuck at a particular local minimum. Since we are dealing with a stochastic process, the solution should be taken as the best of a number of runs. Genetic algorithms are a developing area of research and many amendments could be made to the functions which we have used to implement a genetic algorithm. For example, the weighted roulette wheel selection can be implemented in many dierent ways, crossover can be changed to multi-point crossover or other alterna- tives. Elitism, where the best solution is always passed on to the next generation, is a particular feature for which good results have been reported. An alternative coding for the variables of the problem such as Gray coding have also been used by workers in the eld. It will often be noticed that a genetic algorithm is slow in execution, but it is usually applied to problems where standard optimization techniques perform poorly; for example those which have multiple optima and where the global optimum is required. Problems which have both real valued and integer variables pose particular problems for standard algorithms. Since standard al- gorithms often fail in these cases, the extra time taken by the genetic algorithm is worthwhile. Having described the basic principles of a genetic algorithm we will now illustrate how it may be applied by consider- ing an engineering application. 3. Penalty functions for damage location The location of damage in a structure using vi- bration data is very dicult. We will make the assumption that damage causes a change in the sti- ness of the structure. In many cases this is a gross assumption and the distinction between strength and stiness should be made. For example, in a concrete highway bridge the concrete is held in compression by steel tendons and the stiness of the bridge comes pre- dominantly from the concrete. The strength of the bridge comes from the steel tendons, since concrete has little strength or stiness in tension. Thus, as a bridge is damaged its dynamic response will change very little until just before failure. The simplied problem to be addressed by this paper is to identify the location and extent of a limited number of stiness reductions that best reproduce the measured dynamics. The stiness reductions will be modelled by a reduction in the Young's modulus for a given element, and its location is specied by the el- ement number. Localization of the damage is depen- dent on a suciently rened mesh. A very ne mesh could be used with a tness penalty included to ensure that adjacent elements are chosen. The examples that follow only include beam and plate examples, although any structure with its associated nite element model may be used. A maximum of two damage sites will be allowed, although the extension to more sites is straightforward. The measured data will be the natural frequencies and mode shapes of the structure. We next consider the objective function which is based on the measured data. This will consist of three terms: a term relating to the error in natural frequen- cies, J o , a term relating to the error in mode shapes, J f , and a term to weight against two damage sites. These terms will be weighted to give a total error of J = W o J o W f J f W ns d ns (1) where W o , W f , W ns are weighting factors and d ns = 0 if one site is damaged, 1 if more than one site is damaged.
The term added to weight against too many damage
sites is very important. Suppose damage only occurs at one site. Due to measurement noise and model inade- quacies, the modal model will not be reproduced exactly, even if the damage is correctly identied. Introducing a second damage site may make a minor improvement in the correlation between test and analy- sis, although in reality one site is damaged. The ten- M. Friswell et al. / Computers and Structures 69 (1998) 547556 549 dency will always be to nd damage at every site. Thus, a penalty is introduced to weight against an increased number of damage sites. 3.1. The frequency error The obvious objective function for frequency errors is just a weighted sum of squares of the dierences in the natural frequencies, provided the modes are paired correctly. Any natural frequencies corresponding to analytical and experimental modes that do not pair with sucient condence are simply omitted from the objective function. Thus, one possibility for J o is, J o = X r j=1 W oj o mj o aj o mj ! 2 (2) where o mj and o aj are the jth measured and analytical natural frequency of the damaged structure, and r is the number of measured modes. If required, the indi- vidual frequencies may be given dierent weights, denoted by W oj , although often these weights are taken as unity. The analytical natural frequencies, o aj , could be used in place of the measured frequencies in the denominator in Eq. (2), although the dierences in these frequencies should be small and using the ana- lytical frequencies makes the sensitivity calculations much more dicult. The problem with Eq. (2) is that the model for the undamaged structure should correlate well with the measured dynamics if the structure was undamaged. If periodic measurements of a structure are undertaken, so that measurements of the undamaged structure exist, then the model of the undamaged structure may be updated to correspond closely to the undamaged structure. If this updating is not performed, then the algorithm to identify the location and extent of the damage will try to compensate for both the original modelling error and the change due to the damage. Obviously, the identication of damage in this case is very dicult. A dierent method to reduce the eect of modelling errors is to consider changes in frequency from the undamaged to the damaged structure. Of course, this assumes that the natural frequencies of the undamaged structure were measured. In this case, J o is given by J o = X r j=1 W oj do mj do aj o mj ! 2 (3) where do mj and do aj are the changes, due to damage, in the jth measured and analytical natural frequencies. Eq. (3) gives the penalty function that will be used in subsequent sections. One possible disadvantage with Eqs. (2) and (3) is that the range of the penalty function is not known. Using the penalty function J o = X r j=1 W oj o 2 mj o 2 aj o 2 mj o 2 aj ! 2 (4) in place of Eq. (2) overcomes this problem since the in- dividual natural frequency error terms are between 0 and 1. This may not be a great advantage, since Eqs. (2) and (4) are approximately the same when the dierences between the measured and analytic frequen- cies are small. Calculating the sensitivities for Eq. (4) is also dicult. The problem of the relative weighting between the frequencies and mode shapes will be addressed later. Expressions similar to Eq. (4), but using frequency changes, may also be dened, and instead of squaring the terms the absolute value may be taken. 3.2. The mode shape error The error in the mode shapes may be incorporated into a penalty function as J f = X r j=1 W fj (f mj f aj ) T (f mj f aj ) (5) where f mj and f aj are the jth measured and analytical mode shapes. Only those degrees of freedom that are measured are picked out of the analytical mode shapes. To apply Eq. (5) correctly, both sets of mode shapes should be normalized consistently, and typically mass normalization is used. Alternatively, the modes may be normalized using the modal scale factor (MSF) [31]. Using the MSF also makes sure that no modes are 1808 out of phase. Of course the modes must also be real if the nite element model does not include damp- ing. The terms in Eq. (5) can also be weighted using the analytical mass or stiness matrix, reduced to the measured degrees of freedom if necessary. An alternative method that does not require the mode shape scaling, is to use the modal assurance cri- terion (MAC) [31]. Thus, the new penalty function is J f = X r j=1 W fj [1 MAC(f mj ; f aj )] (6) where MAC(f mj , f aj ) is the MAC value for the jth paired mode. This penalty function has the advantage that each term is between 0 and 1, since the MAC values are between 0 and 1. It is dicult to use penalty functions based on the MAC with eigensensitivity methods, since the derivative of the penalty function is dicult and costly to obtain. The [1 MAC(f mj , f aj )] terms may also be square rooted. M. Friswell et al. / Computers and Structures 69 (1998) 547556 550 3.3. The relative weighting of frequencies and mode shapes How much weight should be given to the errors in the natural frequencies, compared to the errors in the mode shapes? Although the mode shapes can provide valuable information, one has to be careful. The mode shapes are measured with less accuracy than the natu- ral frequencies, and generally the mode shapes are less sensitive to damage in the structure. As an example, consider a uniform cantilever beam with 15 elements. Let the stiness of the fourth element from the clamped end be reduced to simulate damage. Figs. 1 and 2 give the change in the natural frequency and mode shape (through the MAC value) for the rst ve modes as the element stiness is reduced. Obviously, the stiness reduction has a relatively large eect on the natural frequencies, as compared with the MAC values. Because of the insensitivity of the modes to the damage, only the natural frequency penalty function will be used in the following examples. 4. The application of the genetic algorithm Let us rst consider the parameters to be optimized by the genetic algorithm. One advantage of the genetic algorithm is the ability to optimize a combination of discrete and continuous variables. In this application we have both discrete variables, the location of the damage given by the element number, and continuous variables, the extent of the damage given as a percen- tage reduction in stiness. These continuous variables must be expressed in terms of a xed number of binary digits. Thus, in rea- lity, genetic algorithms optimize discrete variables, and continuous variables are approximated to the required accuracy. For example, the damage extent variables are in the range 0100% and if 8 bits are allocated to represent these variables, the maximum resolution is 100/102310.1%. The location variables, or element numbers, are already discrete variables. Zero will be included in this variable set to denote no damage. Thus, if there are 15 elements, or possible damage lo- cations, in the nite element model then the location variable will range from 0 to 15, requiring a 4 bit binary number. If the model contains less elements, then some numbers would have to be assigned a zero tness. For example, if there were only 12 elements, then the objective function would be zero for the vari- able values 13, 14 and 15. The objective is to minimize the function given in Eq. (1), whereas genetic algorithms try to nd the t- test members of the population, or maximize an objec- tive function. This is easily rectied by subtracting J, dened in Eq. (1) from a suitable number. Negative t- ness values will be taken to be zero. The choice of this number is critical as it gives the range of tness values for the algorithm, which is used to generate the prob- ability that a particular member of a generation will be selected for the next generation. In general, the most satisfactory results are obtained when the population has a good spread of tness values, from zero to the ttest. The problems that occur are easily shown by an example. Suppose the objective function we wish to minimize takes values between 0 and 1. If we subtract these values from 100 we have tness values between 99 and 100. The tness values are approximately equal, and all members of a generation will have ap- proximately equal probability of survival to the next generation. This is not desirable, as we want only the ttest to survive. If, instead, we subtract the value of the objective function from 1 we obtain tness values between 0 and 1, and a more satisfactory performance of the algorithm. The direct application of genetic al- gorithms to damage detection was discussed by the authors in an earlier paper [32]. Fig. 1. The eect of damage on the rst ve natural frequen- cies of the cantilever beam (mode 1 w; 2 r; 3 q; 4 X; 5 +). Fig. 2. The eect of damage on the MAC for the rst ve modes of the cantilever beam (mode 1 w; 2 r; 3 q; 4 X; 5 +). M. Friswell et al. / Computers and Structures 69 (1998) 547556 551 5. Combining eigensensitivity and genetic algorithms Genetic algorithms may be used to directly minimize the penalty function given in Eq. (1), by forcing the continuous variables to take on a nite number of dis- crete values. Genetic algorithms are best used to opti- mize dicult problems. In damage location and extent estimation most of the diculty arises from trying to locate the damage. Given one or two damage lo- cations, then the optimum values for the extent of the damage are easily calculated from sensitivity type methods, based on a Taylor series expansion of the natural frequencies and mode shapes in terms of the unknown extent variables (see for example, Ref. [33]). In this way the length of the binary string representing each member of the generation is considerably reduced. In practice the penalty functions given by Eq. (3) for frequency and Eq. (5) for mode shapes, are most appropriate for the eigensensitivity method. 6. A simulated cantilever beam example The location of damage using the combined eigen- sensitivity and genetic algorithm method was demon- strated using a simulated steel, cantilever beam of length 1 m and cross section 2550 mm. Bending in the plane of the thinner dimension only, was con- sidered, that is in the most exible of the two bending planes. The changes in the rst 5 natural frequencies were used to locate the damage, and thus the penalty function is given by Eq. (3). Four cases were tested: case 1 demonstrated the per- formance of the method to locate a single damage site; in case 2 the damage was to an element near the free end of the cantilever and hence dicult to locate; in case 3 the damage was located at two sites; case 4 demonstrated the performance of the method to sys- tems with systematic errors, namely the addition of extra mass into the ``measured'' system. In all cases the population had 10 members, 60% of which were mated at every generation. Each binary bit had a 0.5% chance of mutation at every generation. In all cases, the nite element model used to locate the damage contained 15 elements, or 30 degrees of free- dom. These elements were numbered from the xed end. Only one run of the genetic algorithm was used for each optimization. Mating was accomplished using a single point crossover, with a randomly generated crossover point. The variables used for the genetic al- gorithm optimization were the location of two damage sites. For a given set of damage sites the extent of the damage was optimized using the eigensensitivity approach. The genetic algorithm has been modied so that the most t member of a population survives into the next generation, without mating or mutation. This is called elitism and ensures that the ``best'' solution is maintained from generation to generation. The average tness is also shown in the plots, although care must be exercised in interpreting these plots. The optimum was always selected as the best member of a gener- ation, and the convergence of the average tness should not be taken as any sort of convergence to the optimum. The increasing average tness values do indi- cate a reduction in the genetic diversity of the popu- lation. 6.1. Case 1 The simulated data was obtained from a nite el- ement model with 15 elements. The model used to locate the damage was the same as that used to simu- late the measurements. Damage was simulated as a 30% reduction in the stiness of element 4. Table 1 shows the damaged and undamaged natural frequen- cies. The weights given to the terms in the objective function were W oj = 100; W ns = 0:25 and the objective function, Eq. (1), was subtracted from 1, giving tness values between 0 and 1. Remember, negative values of tness are taken as zero. Fig. 3 shows the tness of the most t member of the population, and also the average tness of the population. The damage was correctly located at el- ement 4 at the sixth generation. The average tness increased with the generation number. Table 1 The damaged and undamaged natural frequencies for cases 1 and 4 Natural frequencies (Hz) Without added mass (case 1) With added mass (case 4) Mode number Undamaged Damaged Undamaged Damaged 1 21.0 20.4 20.5 20.0 2 131.3 131.2 131.3 131.2 3 367.7 362.2 365.5 360.1 4 720.7 705.7 709.2 694.5 5 1191.7 1181.7 1176.0 1165.5 M. Friswell et al. / Computers and Structures 69 (1998) 547556 552 6.2. Case 2 Damage was simulated by a 30% reduction in sti- ness at element 14. The cantilever beam model has only 15 elements, and the damage was very close to the free end of the cantilever beam. Changes in sti- ness in this region have very little inuence on the lower modes of the beam. Fig. 4 shows the results in this case. The genetic algorithm correctly located the damage, although the high average tness even at the rst generation shows that many estimated damage lo- cations tted the measured data reasonably well. 6.3. Case 3 Damage was simulated at two locations, by 30% stiness reductions at elements 4 and 12, and the results are shown in Fig. 5. The algorithm has less suc- cess here but still located the damage at elements 5 and 12, which was very close to the actual damage lo- cations. Note that the maximum tness expected in this case was 0.75, since damage was located at 2 sites. 6.4. Case 4 This case demonstrated the eect of a systematic error in the ``measurements''. Substantial discretization errors in the nite element model are often found in practice, although in our example these were relatively small because the original nite element model was suciently rened. The systematic error introduced into the 15 element cantilever beam model was a mass of 0.2 kg added at node 12. Table 1 shows the error in the rst 5 natural frequencies of the damaged and undamaged model, and shows the level of systematic error introduced. The evolution of the tness values is shown in Fig. 6. The damage was correctly located in element 4, despite the systematic error in the ``measured'' data. The per- formance in many respects was better than that in case 1, without the systematic error. This is probably because only a single run of the algorithm was per- formed. Fig. 3. Fitness values for the simulated beam problem, case 1(most t member ; average tness - - - - -). Fig. 4. Fitness values for the simulated beam problem, case 2(most t member ; average tness - - - - -). Fig. 5. Fitness values for the simulated beam problem, case 3(most t member ; average tness - - - - -). Fig. 6. Fitness values for the simulated beam problem, case 4(most t member ; average tness - - - - -). M. Friswell et al. / Computers and Structures 69 (1998) 547556 553 7. An experimental example The second example consists of a 3 mm steel plate which was clamped along one edge. The exposed plate measured 305 mm 357 mm and was clamped along the shorter side. Fig. 7 shows the nite element mesh used to model the undamaged plate. This mesh resulted in a 48 degree of freedom model and produced the estimated natural frequencies given in Table 2. Table 2 also gives the natural frequencies of the plate measured experimentally using impact excitation. The 12th mode was not measured because the acceler- ometer was placed at a node of this mode. The analyti- cal and experiment modes were easily paired in this case using the displayed mode shapes. Clearly, the nite element mesh was very coarse and there were considerable modelling and discretization errors pre- sent. Any damage location method should be robust to such errors, and, to increase the robustness, the dier- ence in natural frequencies between the undamaged and damaged plate was used to locate the damage. The plate was damaged by 4 saw cuts each of length 40 mm in the shape of a cross centred on element 4 of the 48 degree of freedom nite element model, as shown in Fig. 7. The population had 10 members, 60% of which were mated at every generation. Each binary bit had a 0.5% chance of mutation at every generation. The weights given to the terms in the objective function were W oj = 1000; W ns = 0:25 and the objective function, Eq. (1), was subtracted from 1, giving tness values between 0 and 1. Negative values of tness were taken as zero. A maximum of 2 damage locations were assumed, and each location variable was 4 bits long; zero was taken as no damage, integers 112 as damage in the corresponding element and integers 13, 14 and 15 were given a zero tness value. Fig. 8 shows the tness of the best member of Fig. 7. The experimental cantilever plate example, showing the nite elements and the position of the saw cut. Table 2 Analytical and experimental natural frequencies and frequency changes due to damage for the experimental cantilever plate example Mode number Analytical undamaged natural frequency (Hz) (48 degrees of freedom model) Experimental undamaged natural frequency (Hz) Experimental natural frequency change due to damage (Hz) 1 19.1 20.0 0.4 2 55.7 56.7 1.0 3 119.7 124.6 3.8 4 192.8 198.1 4.7 5 214.6 212.6 3.0 6 341.6 353.9 10.9 7 367.6 380.8 3.0 8 405.6 427.5 12.0 9 544.7 530.1 9.0 10 602.3 639.9 17.5 11 663.6 690.7 24.5 12 665.8 Not measured Not measured 13 705.7 774.5 25.2 Fig. 8. Fitness values for the experimental cantilever plate example(most t member ; average tness - - - - -). M. Friswell et al. / Computers and Structures 69 (1998) 547556 554 the population in each generation. The correct location of the damage (element 4) was found in the 8th gener- ation, showing the approach is robust to modelling and measurement errors. 8. Conclusions A combined genetic algorithm and eigensensitivity method has been used to identify the location and magnitude of damage from measured vibration data. Essentially, the genetic algorithm is used to identify the damage located and the eigensensitivity is used to identify the damage extent. Damage at one and two sites have been successfully located in the simulated example of a cantilever beam. Damage was also suc- cessfully location in an experimental cantilever plate. The algorithm is robust to systematic errors in the measured data, demonstrated in simulation by the ad- dition of a discrete mass and experimentally by using a crude model for the plate. One feature of the numeri- cal example given is that genetic diversity was reduced as the generation number increased. Alternatively, this could be viewed as the convergence to a uniform ``op- timum'' population, although it would not be guaran- teed that this population would be the global optimum. This may be a feature of the small number of genes associated with each member of the popu- lation. Acknowledgements Dr Friswell gratefully acknowledges the support of the Engineering and Physical Sciences Research Council through the award of an advanced fellowship. References [1] Rajeev S, Krishnamoorthy CS. Discrete optimization on structures using genetic algorithms. Journal of Structural Engineering 1992;118(5):123349. [2] Jenkins WM. Plane frame optimum design environment based on genetic algorithm. ASCE Journal of Structural Engineering 1992;118(11):310312. [3] Jenkins WM. Towards structural optimization via the genetic algorithm. Computers and Structures 1991;40(5):13217. [4] Keane AJ. Passive vibration control via unusual geome- tries: the application of genetic algorithm optimization to structural design. Journal of Sound and Vibration 1995;185(3):44153. [5] Goldberg DE, Kuo CH. Genetic algorithms in pipeline optimization. Journal of Computing in Civil Engineering 1987;1(2):12841. [6] Larson CB, Zimmerman DC. Structural model rene- ment using a genetic algorithm approach. 11th International Modal Analysis Conference, Kissimmee, Florida, 1095101, 1993. [7] Levin RI, Lieven NAJ. Dynamic nite element model updating using simulated annealing and genetic algor- ithms. Mechanical Systems and Signal Processing 1998;12(1):91120. [8] Dunn SA. Modied genetic algorithm for the identi- cation of aircraft structures. Journal of Aircraft 1997;34(2):2513. [9] Zimmerman DC. A Darwinian approach to the actuator number and placement problem with non-negligible actuator mass. Mechanical Systems and Signal Processing 1993;7(4):36374. [10] Cai JB, Thierauf G. Evolution strategies for solving dis- crete optimization problems. Advances in Engineering Software 1996;25(23):17783. [11] Topping BHV, Khan AI, Leite JPD. Topological design of truss structures using simulated annealing. Structural Engineering Review 1996;8(23):30114. [12] Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage identication and health monitoring of struc- tural and mechanical systems from changes in their vi- bration characteristics: a literature review. Los Alamos National Laboratory report LA-13070-MS. Electronic version available on WWW: http://lib-www.lanl.gov/la- pubs/00285981.pdf, 1996. [13] Mottershead JE, Friswell MI. Model updating in struc- tural dynamics: a survey. Journal of Sound and Vibration 1993;167(2):34775. [14] Zimmerman DC, Kaouk M. Structural damage detection using a minimum rank update theory. ASME Journal of Vibration and Acoustics 1994;116:22231. [15] Hemez FM, Farhat C. Structural damage detection via a nite element model updating methodology. Modal Analysis: The International Journal of Analytical and Experimental Modal Analysis 1995;10(3):15266. [16] Topole KG, Stubbs N. Nondestructive damage evalu- ation in complex structures from a minimum of modal parameters. Modal Analysis: The International Journal of Analytical and Experimental Modal Analysis 1995;10(2):95103. [17] Mares C, Surace C. An application of genetic algorithms to identify damage in elastic structures. Journal of Sound and Vibration 1996;195(2):195215. [18] Ruotolo R, Surace C. Damage assessment of multiple cracked beams: numerical results and experimental vali- dation. Journal of Sound and Vibration 1997;206(4):567 88. [19] Cawley P, Adams RD, Pye CJ, Stone BJ. A vibration technique for non-destructively assessing the integrity of structures. Journal of Mechanical Engineering Science 1978;20(2):93100. [20] Cawley P, Adams RD. The location of defects in struc- tures from measurements of natural frequencies. Journal of Strain Analysis 1979;14(2):4957. [21] Cawley P, Adams RD. A vibration technique for non- destructive testing of bre composite structures. Journal of Composite Materials 1979;13:16174. M. Friswell et al. / Computers and Structures 69 (1998) 547556 555 [22] Friswell MI, Penny JET, Wilson DAL. Using vibration data and statistical measures to locate damage in struc- tures. Modal Analysis: The International Journal of Analytical and Experimental Modal Analysis 1994;9(4):23954. [23] Friswell MI, Penny JET, Garvey SD. Parameter subset selection in damage location. Inverse Problems in Engineering, 1997; 5(3):189215. [24] Wu X, Ghaboussi J, Garrett JH. Use of neural networks in detection of structural damage. Computers and Structures 1992;42:64959. [25] Tsou P, Shen MHH. Structural damage detection and identication using neural networks. AIAA Journal 1994;32(1):17683. [26] Elkordy MF, Chang KC, Lee GC. Application of neural networks in vibrational signature analysis. ASCE Journal of Engineering Mechanics 1994;120(2):25065. [27] Szewczyk ZP, Hajela P. Damage detection in structures based on feature-sensitive neural networks. ASCE Journal of Computing in Civil Engineering 1994;8(2):16378. [28] Worden K. Structural fault detection using a novelty measure. Journal of Sound and Vibration 1997;201(1):85101. [29] Elkordy MF, Chang KC, Lee GC. A structural damage neural network monitoring system. Microcomputers in Civil Engineering 1994;9:8396. [30] Kirkegaard P, Rytter A. Use of neural networks for damage assessment in a steel mast. 12th International Modal Analysis Conference, Honolulu, Hawaii, 1994:112834. [31] Allemang RJ, Brown DL. A correlation coecient for modal vector analysis. 1st International Modal Analysis Conference, Orlando, Florida, 1982:1106. [32] Friswell MI, Penny JET, Lindeld G. The location of damage from vibration data using genetic algorithms. 13th International Modal Analysis Conference, Nashville, Tennessee, USA, 1995:16406. [33] Friswell MI, Mottershead JE. Finite element model updating in structural dynamics. Dordrecht: Kluwer, 1995. M. Friswell et al. / Computers and Structures 69 (1998) 547556 556