Study On The Structural Damage Identification Based On The Time Series Analysis And Neural Network

During the servicing period of engineering structures, due to various complicated factors such as natural environment erosions and disasters etc, the structure will suffer different degrees of damage. So, the structural damage identification, evaluation and repairing has great significance to the loss of lives and properties.To improve the accuracy and robustness of the damage identification. The structural damage identification method based on the time series analysis and neural network have been used to identify the damage offshore platform. The main content of the thesis includes:(1) Various structural damage identification methods based on vibration response analysis have been summarized, and the scopes and limitations of thire application are pointed out.(2)For characteristics of offshore platform, the damage identification method based on AR (Auto-Regressive) model and BP neural network is proposed. The acceleration responses are used to create the dynamic model with the AR model of time series, the changes of the first 3-order model parameters are extracted and composed as damage characteristic vectors which are put into BP neural network to identify damage.(3) The damage identification of single damaged component and two damaged components are studied with numerical model of offshore platform, the white noise and ambient excitation are used to motivate the structure respectively, and the influences of measurement noise are also considered. The simulation results show that, when the degree of noise is not more than 3%, the method can identify the damage accurately, however, when the degree of noise is more than 3%, some conditions can not be identified correctly, some misjudgment phenomenons appear.(4) The data of the impact response experiment and the shaking table experiment for offshore platform are employed to test the proposed method. The results show that 6 of the 9 kinds of damage cases can be identified by the method with the impact response experiment data, and 2 of the 3 kinds of damage cases can be identified by the method with the shaking table experiment data. Because there are few damage cases in the shaking table experiment data, the numerical simulation data is used to train the BP network, and the experimental data is only used as test data. The results of the experiment show the proposed menthod is valid.