Keywords: inverse problem, load identification, genetic algorithms, neural network, finite element.

Recently the interest in transferring techniques developed in one field to the analysis of problems of other fields has increased. The determination of loads and stresses using structural response (strain) could be defined as an inverse problem [1]. Solving this inverse problem means accurately calculating the external loads or boundary conditions that can generate a known amount of stress or displacement at predetermined locations in a structure. Of course, in these problems, we assume that we have available appropriate data about the problem under consideration. Such data are usually gathered by the classical methods of experimental stress analysis such as strain gauge, photo-elasticity or data generated by numerical methods such as finite element analysis (FEA) and boundary element analysis (BEA) [2,3].

This research addresses an innovative methodology based on a hybrid artificial neural network (ANN) and genetic algorithms (GA), for the solution of inverse problems pertaining to load identification. Using this methodology, we express the inverse problems as a minimization problem with the objective function being the differences between the actual outputs and the corresponding computed outputs from a candidate ANN. A GA uses this objective function to measure the fitness of individuals in a population of candidate ANNs (solutions). Possible approximate solutions found by the GA gradually lead to an increase in the average fitness of the population and thus better estimates of load identification. A significant benefit of using an ANN model is its ability to learn relationships between variables through an adaptive training process. The main advantage of GAs is that they do not require any mathematical augmentation to the numerical solution methods used to represent ill-posed objects.

Some of the advantages of the hybrid GA and NN, which combines both genetic search with neural networks algorithm, are that:

1. The global search approach is less likely to obtain a local minimum.
2. There is a higher probability of convergence as a result of the global search.
3. The global near-optimum solutions (obtained by the genetic search) solution is improved by local gradient descent algorithm (back propagation) to obtain the true global optimal solution.
4. The parallel genetic search approach offers several potential solutions for a holdout sample.

The results show that the models produced by the proposed algorithm achieve a smaller prediction error compared with standard training methods. The successful network predictions on the testing data set illustrated the efficiency of this method and clearly showed that it is a reliable tool for solving inverse problems, by extracting valuable information from input-output data. The algorithm uses GAs to determine the fittest neural network, contrary to most standard ANN training methods, where the network is selected by a time-consuming trial and error procedure. Another advantage of this new technique is that it employs the input-output training data during the entire training procedure. The data gathered from finite element analysis (FEA). The selection of training data plays large role in the ability of the hybrid system to accurately predict the loading. It is important that the hybrid system is trained with data that represents the full spectrum of expected loading conditions On the basis of achieved results the following general conclusions can be stated:

• A novel approach to the simulation of load-strain response of a component in service has been demonstrated using the hybrid genetic algorithm, neural networks (GANN) and data from finite element analysis.
• The neural network trained by genetic algorithms is able to predict the load value and position for a simplified wing model, after its successful training.
• The neural networks trained by genetic algorithms provide a very fast and computationally efficient economical analysis tool to simulate the load-strain relationship.

1

Shkarayev S., Krashantisa R., Tessler A., "An inverse interpolation method utilising in-flight strain measurements for determining loads and structural response of aerospace vehicles", NASA Langley Research Center, 2002

2

Cao X., Sugiyama Y., Mitsui Y., "Application of artificial neural networks to load identification", Computer & Structures, Pergamon, p63-78,1998. doi:10.1016/S0045-7949(98)00085-6

3

Xu G., Noroozi. S., Littlefair G., Penson R., Callan R., "A Finite Element based neural network model", MESM 99, Society for computer Simulation International European Council , University of Jordan, March 1999

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £105 +P&P)