Abstract：The stability of division pier influences the safety of the permanent shiplock of Three Gorges project. The safety coefficient is usually used to evaluate the stability of division pier. Wavelet network combines the time-frequency domain localization properties of wavelet transform and self-learning ability of traditional feed-forward neural network. Wavelet network was used to approximate the nonlinear relation between safety coefficient and control factors of the division pier stability of the permanent shiplock. The stability of division pier can be analyzed and forecasted with wavelet network trained by engineering case as learning samples. The numerical example shows that the precision of approximation and prediction of the proposed model can satisfy the quantitative evaluation for the stability of division pier of the permanent shiplock. The model is of good capability of noise-resistance. The wavelet net work is trained by the hydrid learning algorithm of Levenbery-Marquardt algorithm and Least-Squares algorithm. Levenbery-Marquardt algorithm can train the nonlinear parameters. Least-Squares algorithm can train the linear parameters. Error objective function is minimized in subspace of whole parameter space of wavelet network with the presented algorithm. Wavelet network is obviously superior to conventional algorithms.
