基于LSTM循环神经网络的大地电磁工频干扰压制

Magnetotelluric power frequency interference suppression based on LSTM recurrent neural network

大地电磁测深方法观测天然电磁场信号,具有频带范围宽、探测深度大等优点,被广泛应用于油气资源勘探与地球深部结构探测.但天然场源信号微弱、易受电磁干扰,压制电磁干扰是大地电磁数据处理的关键问题之一.本文提出了一种基于长短时记忆循环神经网络的大地电磁工频干扰压制方法.首先构建双向长短时记忆循环神经网络模型,然后建立数据集对模型进行训练,最后将含有工频噪声的野外实测数据输入训练好的模型,模型的输出为工频噪声,输入与输出的差值即为消噪后的真实信号.模拟数据处理结果表明去噪前后时间序列相关系数达0.9691,实测数据处理结果表明该方法能够有效压制大地电磁信号中的工频干扰,提高数据处理质量.

The magnetotelluric sounding method observes the natural electromagnetic field signal and has the advantages of wide frequency band and large detection depth.It is widely used in oil and gas resource exploration and deep earth structure detection.However,the natural field source signal is weak and susceptible to electromagnetic interference.Electromagnetic interference suppression is one of the key issues in magnetotelluric data processing.In this paper,a magnetotelluric power frequency interference suppression method based on Long Short-Term Memory(LSTM)recurrent neural network is proposed.Firstly,build a Bidirectional long short-term neural network model,then build data sets to train the model,and finally input the measured data containing power frequency noise into the trained model.The output of the model is power frequency noise,and the difference between input and output is the true signal after denoising.The simulation data processing results show that the time series coefficient before and after denoising ups to 0.9691.The measured data processing results show that the method can effectively suppress the power frequency interference in the magnetotelluric signal and improve the data processing quality.