Study Of Load Identification And Damage Detection Based On Artificial Neural Networks

Recent years civil engineering accidents occurred frequently in the world, which resulted in great concern on the structural health. The actual conditions of engineering structures would change when it was used for a period of time, such as material aging and fatigue. This change might lead to reduction in bearing capacity. How to know the damage conditions of engineering structures before they failed to work was a hot problem in structural engineering that many experts put their emphasis on. This problem was called health monitoring. Damage identification and load monitoring were two important items of it. In this text, artificial neural networks was used to identify load acting on the structures and detect damage case in the structures.The self-adaptive algorithm of learning rate for BP neural networks was implemented in this context. Values for the key parameters of the networks such as initial learning rate and momentum factor were optimized by numerical imitation for several times. Training specimen capacity varied to obtain different output precision, and relationship between the specimen capacity and the identification effect was discussed.In load identification , displacement (or displacement curvature) and strain (or strain curvature) were used as input data of the BP networks respectively. It turned out that the identification effect of network using displacement curvature or strain curvature as input data was better than that of the networks using displacement or strain; the networks using the strain (strain curvature) as input data performs better than the networks using displacement (displacement curvature).In damage detection, the damage indices were divided into two parts, namely location index related to damage location only and extent index which was just related to damage extent. A new location index that was based on fusion of static and dynamic data was proposed. Two-step-method and one step-method using self-adaptive networks were adopted for damage detection respectively. It turned out that two-step-method could make an more accurate estimation for damage location and extent than one-step-method did. At last, the influence of the number of location index or extent index on the detection result was discussed.Through the numerical examples, the developed algorithm was proved to avoid oscillation phenomenon during training process and improve learning rate effectively. Optimum values of key parameters enhanced error convergence velocity and output precision of the BP networks.