Shear velocity estimation by the combined use of supervised and unsupervised neural networks

A neural estimator composed of two different neural networks, namely, a perceptron an a Kohonen map, is proposed. The training algorithm of the resulting combined network is very simple since the component networks are trained in succession, starting from the Kohonen network, and the algorithms originally developed for them are used. The estimator is fully developed in connection with a determination of the shear velocity of the formation surrounding a fluid-filled borehole. The neural estimator is applied to synthetic seismograms out of the training set. In most of the cases a considerable improvement in estimation accuracy is obtained with respect to the Kohonen map. However, only a slight improvement in estimation accuracy is obtained with respect to the Kohonen map. However, only a slight improvement was noted for perceptrons which suffered convergence problems during the training phase. >