DAMAGE DETECTION OF STRUCTURES USING MODAL PARAMETERS AND NEURAL NETWORK

Abstract

In this thesis, the changes in the modal parameters due to the changes in the characteristics of the structure are studied which can be used for structural damage identification and also new directions in the area of damage detection are explored. Two models have been tested with the varying condition of induced damage in the two different laboratories. The first model tested was a one bay by one bay by three story steel non shear frame model with two shear panel in both the planes in the form of bracing parallel to the direction of motion. The structure was given an input motion of artificially generated pink noise at the base in the shake table with different structural configuration such as with removal of single bracing on both sides of plane and removal of all the bracing of the structure. Since the damage in the structure was given by removal of bracings which were connected to the roof of the structure hence the stiffness reduction identified in the third story damage with varying amount of damage as per the respective cases was significant. The idea of keeping the input motion same in all the cases of the experiments led us to conclude clearly that the various dynamic parameters such a frequency, mode shape, modal strain energy, flexibility matrix and Frequency Response Function (FRF) changed according to the varying degree of damage for a particular system. Thus the purpose of the test to be able to determine any change or to locate the change was achieved. To develop a substantially accurate one dimensional analytical model of the structure, sensitivity analysis was also carried out for the beam column fixity. The matching of the frequency was set as the benchmark for the comparative study of the analytical and the experimental results. The transfer functions of the different floors for various cases correlated well for the experimental and analytical model. To study the correlation of the degree of damage in a structure with the lateral displacement and to study the energy dissipation in a structure another test was carried out in the quasi-static laboratory on a one bay by one bay by one story concrete frame model without any infill. The damage was created by introducing crack in the model by applying quasi-static force and thus reducing the stiffness of elements which was measured in the form of decreasing frequency. Damage thus introduced through increased displacement led to the damage in the particular location which was determined through the clearly evident variations of the other derived modal parameters. The analysis results depict that the proposed empirical equation based on strain energy gives a good indication ofdamage location. Damage indices based on various parameters such as displacement, stiffness degradation given by various authors are compared and the result of which are in form the increasing or decreasing trends with the increase of damage. The result also shows that the state of the structure evolved after the energy dissipation of the input energy if determined from the free vibration test will be on a higher side than state of the structure obtained through the quasi static test. However the variation of stiffness defining the state of the structure decreased in the free vibration test and the quasi static test decreases as the damage is increased. The free vibration test carried on the structure was also used to locate the damaged element in the structure. The study of Neural Network based approach for damage detection was further carried out through two different approaches to determine the degree of damage in the structure. The first approach was to train the network with the frequency changes and mode shape changes while the second approach was to training the network with change in transfer function carried out for each floor independently. Through both the approach it was possible to obtain satisfactorily accurate degree of damage in floor of the case studies undertaken. The purpose of both approaches was to quantify the damage and to apply in different operational level buildings. After the earthquake event, the determination of frequency and mode shape of building would require ambient vibration test of the building again whereas in case of transfer function approach it has been assumed that the structure is instrumented during the earthquake. To validate both the approaches four storey and eight storey reinforced concrete building model was considered. The input data was artificially generated for the neural network from the program codes written. The training of the network was carried out for different combination of damage cases and the result showed that the accuracy of degree of damage detected in structure increased with the increase in the number of combination of damage considered for neural network training. Frequency and mode shapes being a linear system property was assumed to be determined through ambient vibration where as the transfer function also is assumed to be taken for that part of the vibration in which the system vibrates linearly after the structure has been damaged. From the two methodologies developed it has been found that the accuracy to determine severity of damage decreases with increase in the number of storey being damaged. Further the instrumentation of first floor of the building would give best result incase the damage is detected based on transfer function change approach.