Damage Identification of Structures Through Machine Learning Techniques with Updated Finite Element Models and Experimental Validations

Structural Health Monitoring (SHM) Techniques have recently started to draw significant attention in engineering applications due to the need of maintenance cost reductions and catastrophic failures prevention. Most of the current research on SHM focuses on developing either purely experimental models or stays on purely numerical data without real application validation. The potential of SHM methods however could be unlocked, having accurate enough numerical models and classifiers that not only recognize but also locate or quantify the structural damage. The present study focuses on the implementation of a methodology to bridge the gap between SHM models with numerically generated data and correspondence with measurements from the real structure to provide reliable damage predictions. The methodology is applied in a composite carbon fiber tube truss structure which is constructed, using aluminum elements and steel bolts for the connections. The composite cylindrical parts are produced on a spinning axis by winded carbon fibers, cascaded on specified number of plies, in various angles and directions. 3D FE models of the examined cylindrical parts are developed in robust finite element analysis software simulating each carbon fiber ply and resin matrix and analyzed against static and dynamic loading to investigate their linear and nonlinear response. In addition, experimental tests on composite cylindrical parts are conducted based on the corresponding analysis tests. The potential damage to the structure is set as loose bolts defining a multiclass damage identification problem. The SHM procedure starts with optimal modeling of the structure using an updated Finite Element (FE) model scheme, for the extraction of the most accurate geometrical and physical numerical model. To develop a high-fidelity FE model for reliable damage prediction, modal residuals and mode shapes are combined with response residuals and time-histories of strains and accelerations by using the appropriate updating algorithm. Next, the potential multiclass damage is simulated with the optimal model through a series of stochastic FE load cases for different excitation characteristics. The acceleration time series obtained through a network of optimally placed sensors are defined as the feature vectors of each load case, which are to be fed in a supervised Neural Network (NN) classifier. The necessary data processing, feature learning and limitations of the NN are discussed. Finally, the learned NN is tested against the real structure for different damage cases identification.