Use of artificial neural networks for optimal sensing in complex structures analysis

Among the advantages offered by the Artificial Neural Networks (ANNs), in the analysis and active control of structures characterized by high modal densities and complexity, it should be mentioned the possibility of optimizing the number and the position of sensors and actuators. This feature can result in a sensible reduction of cost in the analysis and control of large structures and could be of great advantage when some of the points under investigation of the structure are not physically accessible. The first step of research is the validation of the ANN approach to a satisfactory analysis of the structure. Previous papers have highlighted the encouraging results obtained from an ANN implementation for the description of a stiffened aluminum panel behaviour. In that case the ANN was trained with data obtained by a validated FEM (Finite Element Method) structural model. In the simulation phase the ANN estimated the behaviour of all the grid points with data referred only to a portion of points with interesting performances in terms of accuracy and computing time. The present paper describes a measurement system in which the ANN is trained with real data, experimentally obtained by accelerometers. The investigation was carried out considering a sinusoidal excitation, typically produced by the rotating engines of turboprob aircrafts. Real data allowed us to rest the ability of the ANN to learn the structural dynamics taking into the right account the influence of noise, not considered in the previous phase of the research. The experimental results highlight again the good ANN behaviour accuracy in the estimation of all the rest points considered for the analysis. The present research is the necessary premise for ANNs based active control applications in aeronautics and aerospace