Adaptive network for automated first break picking of seismic refraction

Abstract

A method of operating an adaptive, or neural, network is disclosed for performing first break analysis for seismic shot records. The adaptive network is first trained according to the generalized delta rule. The disclosed training method includes selection of the seismic trace with the highest error, where the backpropagation is performed according to the error of this worst trace. The learning and momentum factors in the generalized delta rule are adjusted according to the value of the worst error, so that the learning and momentum factors increase as the error decreases. The training method further includes detection of slow convergence regions, and methods for escaping such regions including restoration of previously trimmed dormant links, renormalization of the weighting factor values, and the addition of new layers to the network. The network, after the addition of a new layer, includes links between nodes which skip the hidden layer. The error value used in the backpropagation is reduced from that actually calculated, by adjusting the desired output value, in order to reduce the growth of the weighting factors. After the training of the network, data corresponding to an average of the graphical display of a portion of the shot record, including multiple traces over a period of time, is provided to the network. The time of interest of the data is incremented until such time as the network indicates that the time of interest equals the first break time. The analysis may be repeated for all of the traces in the shot record.