Modal frequencies are one of the most easily quantifiable damage features for damage detection. However, the fact that changes in modal parameters caused by temperature variations usually mask those induced by structural damage is one the main challenges for the reliability of modal parameter-based damage identification methods. Hence, removing the adverse temperature effects from the damage detection process is a crucial need to prevent false positives or negatives. Although autoassociative neural network (AANN) has been extensively used in this context, the fact that temperature values can be measured precisely and easily due to maturing sensing technologies has hardly been exploited in the AANN-based damage identification methods. To fill this void, a modified AANN-based method, which add the representational temperature values to the inputs, is explored in this paper. Numerical models based on an experimental grid structure, with varying temperature conditions and different extents of simulated damage, are used to test the performance of the proposed technique. After training the modified AANN model using the frequency and temperature data sets from models in intact states, corresponding data sets from models in unseen states are used to verify the detectability and robustness of this method by taking the Euclidean Norm as a damage index. Results show that this unsupervised learning method can discern the presence and severity of damage reliably in all of the cases investigated. Various temperature effects and increasing levels of noises introduced into the models hardly influence the damage identification, revealing the high robustness of this method.
doi: 10.12783/SHM2015/103