Study On Structural Damage Identification Based On Independent Component Analysis

Identification of structural damage is always a topic in structural engineering domain. Structural stiffness decreases due to aging, damages, and other harmful effects. These adverse changes lead to abnormal dynamic characteristics in natural frequencies and mode shapes. Detection of structural damage is increasingly important for preventing catastrophic failures and prolonging the service life of structures. By instrumenting structures with a vibration sensor system, structural health monitoring (SHM) aims to provide reliable and economical approaches to monitor the performance of structural systems in the early stage so as to facilitate the decisions on structure maintenance, repair and rehabilitation.Structural damage identification can be regard as a pattern recognition problem, that is, there are two steps: one is feature abstraction from measured dynamic sensor data; the other is identifying the structure damage status based on selected features. This thesis used the dynamic data from sensors as study object, and took advantage of the theory of machine learning, data mining and statistics learning to build a structural damage identification frame based on independent component analysis (ICA), the detail is as following.(1) Principal component analysis (PCA) can keep uncorrelation among the principal components and reduce the influence of noise to the useful structural signals, so it can effectively abstract the structural features. In addition, the cumulate contribution rate decides the accuracy of signal abstraction, and it should satisfy the need of structures. In our experiments, it can work well as 99%.(2) The features can be abstracted by ICA, which overcomes PCA because it keeps the dependence among the components. Fast-ICA algorithm is valid for extracting the structural features, and works well and stably.(3) After analyzing the disadvantage of BP neural network, ICA-ANN model is built by integrating ICA with ANN (BP in this thesis): the processed feature signals by ICA will be input to a neural network for damage detection. The solutions for improving BP neural network were discussed in detail to speed up the convergence, reduce the local minimization and overfitting. The experiments showed that the model can effectively identify the structural damage and damage level, and they proved the validity of this model.