Applications Of Artificial Neural Networks In The Prediction Of Structural Dynamic Stability Under Seismic Excitations

In structural engineering, increasing importance has been attached to the analysis and prediction of nonlinear dynamic stability of spatial skeletal structures under seismic excitations. Following the procedures of structural damage detection based on Artificial Neural Networks (ANN) and wavelet theory, this dissertation proposes two methods of prediction of structural dynamic stability based on ANN: (a) offline prediction of dynamic stability based on PNN (Probabilistic Neural Network); (b) online prediction of dynamic stability based on P-RAN (Resource-Allocating Network). The first method can be used to predict dynamic stability performances of the structure under different excitations by taking advantage of known characteristics. This method, combined with the known information, could enable us to identify the dynamic stability characteristic of structure according to the some properties of load. The second method is used for online prediction of structure response. With the information obtained from this method, the future tendency of structure dynamic stability can be predicted in accordance with a predetermined instability criterion.The dynamic stability characteristic of a 72-beam flat latticed shell under random seismic excitations is analyzed with the offline prediction of dynamic stability based on PNN. The coincidence rate of prediction is up to 75%, and the effective rate peaks in 95%. When the network predicts same type of seismic load, but the peak value, as its studied, the effective rate reaches 100%. On the other hand, if the network predicts the different type of seismic load, the effective rate lower a bit, reaches 92.3%, and the coincidence rate is 61.5%. More, this dissertation also carries out the study of effects of sample quantity on prediction accuracy. The investigation shows that the quantity and quality of sample is important for network performance.Three improvements of P-RAN, include (1) Adopting K-clustering method to initialize network, (2) Proposing Error-Stack learning strategy, (3) A method to optimize hidden neuron number, is proposed. Upon the analysis result of a 24-beam flat latticed shell experimental data, it is indicated that the improved P-RAN method can predict the future response of structure according to history data. The result shows that, P-RAN algorithm has the virtue of responsive and small error, and basing on it, it is effective to predict instability of structure with instability criterions.