Learning Rate Effect in Neural Network for Damage Detection

Neural network is a powerful tool that can be utilized for structural damage detection and health monitoring. Since damage usually varies/reduces stiffness, frequency response variation can be used as indicator for damage occurrence. A well designed neural network can correlate frequency response variation to damage localization/severity without resorting to detailed structural modeling. While various neural network based approaches have been developed, their effectiveness, efficiency, and robustness oftentimes rely on the selection of several important parameters in the network construction. One of the key performance metrics for a neural network is the learning rate. Although the dynamic steepest descent algorithm (DSD) and fuzzy steepest descent algorithm (FSD) have shown promising possibility of improving the learning convergence speed significantly without increasing the computational effort, its performance still depends on the selection of control parameters and control strategy. In this paper, a tunable steepest descent algorithm (TSD) improving the performance of the dynamic steepest descent algorithm is proposed. A numerical benchmark example shows that the proposed algorithm significantly improves the convergence rates of the backpropagation algorithm. A structural health monitoring system incorporated with the neural network trained by the adaptive learning algorithm is developed for detecting the impact damage.