The neural network (NN) processing of magnetotelluric (MT) multi-channel time series data based on Hopfield neural network is proposed. The problem of contamination of the measured MT signal by different types of noise requires accurate estimation of the MT transfer function. We propose NN optimization scheme in order to subdivide the multi-channel time series into sections and extract the smoothed-MT transfer functions from the good of them. Good time series data segmentation and smoothing- can be achieved by finding the optimum solution to an appropriate Hopfield-like energy function. We automatically split time series data into data sections in the process of energy minimization. The minimum of energy is received as a result of gradient descent iterative procedure. As a consequence the initial window of time series data converges to the different stable states. These stable states characterize corresponding data section and represent smoothed MT transform functions. We repeat the procedure of energy minimization recursively downward in hierarchy applied to data sections received from the initial window of data.
展开▼
