Deep Learning Based Noise Identification for CSAMT Data Processing

Abstract

Controlled-source audio-frequency magnetotellurics (CSAMT) has been seriously affected by strong electromagnetic interferences including large-scale drift, durative outbreak interference, and impulsive outliers. To improve the efficiency of noise reduction, a deep learning strategy was proposed to identify the type of noise interference, so as to select the appropriate noise reduction method. First, a CSAMT time series simulation algorithm was developed based on current decomposition and one-dimensional (1D) forward modeling. Three kinds of noise interferences were also generated by simulation and randomly added to the pure signals. A total of 30000 groups of simulated noisy electromagnetic signals were generated. Together with 210 sets of practically measured data, these samples were used to train a long-short-term memory network (LSTM) noise classifier. Then, three targeted de-noising algorithms were adopted to separate the three interferences in the CSAMT time series according to the identifications. The test results by simulated data showed that the identification accuracy of LSTM for noise interferences can reach more than 95%. Finally, the noise identification and suppression methods were applied to a practical CSAMT dataset and the effect was further verified.