| Audio Magnetotelluric(AMT)is an important geophysical exploration method.Its basic principle is roughly the same as that of Magnetotelluric(MT).Both use the natural alternating electromagnetic field as the field source to explore the underground buried body by receiving the response of the underground electrical medium.However,AMT is different from MT in the coverage frequency range.AMT has many advantages,for example,its exploration depth is deep,and the instruments used are portable and easy for geophysical workers to carry.Therefore,AMT is often used to find various geological structures.However,with the development of industrialization,when conducting AMT exploration,the interference caused by various artificial sources such as signal towers,traffic vehicles,high-voltage lines,etc.can not be ignored.The cultural noise generated by these artificial sources will distort apparent resistivity and phase curves.If geophysical workers do not filter out these cultural noises during data processing,they will likely make a wrong geophysical interpretation,which will affect the positioning and reserves evaluation of underground ore bodies and cause economic losses that cannot be ignored.Machine learning is often used in signal-noise separation.It enables the computer to recognize the basic difference between effective signal and noise through some mathematical methods and then establishes a denoising model.Based on the denoising model learned,the computer can achieve the purpose of suppressing noise.This paper analyzes cultural noise in AMT from the perspective of machine learning and studies the separation of signal and noise in AMT from three aspects: sparse representation,mathematical decomposition,and neural network.The innovation of this paper is as follows:(1)In sparse representation,this paper proposes a double sparse denoising dictionary to overcome the shortcomings of fixed basis transformation that cannot adapt to sparse representation and learning basis that will introduce pseudo-noise.This method cascades the wavelet transform in the fixed base transform and the Data-Driven Tight Frame(DDTF)in the learning base transform to build a two-layer sparse structure,so that the double sparse dictionary has the characteristics of both the fixed base and the learning base,and makes up for the shortcomings of the fixed base and the learning base.(2)In mathematical decomposition,this paper proposes a denoising method based on adaptive IMF dictionary learning,aiming at the disadvantage that the intrinsic mode function(IMF)of these methods will leave impulse noise.This method learns a dictionary for each IMF to denoise after variational mode decomposition(VMD)of noisy AMT signals.Compared with the simple use of a threshold,this method can sparsely represent the residual noise in IMF and extract the cultural noise contained in IMF by adaptively setting the threshold,to achieve more accurate AMT denoising.(3)In the neural network,this paper proposes three different ways to improve the characteristics of the neural network and AMT signals:(1)This method does not change the neural network but improves the denoising process according to the local concentration of cultural noise.By first using the residual networks(Res Net)to identify the cultural noise,then reserving the non-noise part,and only using Res Net to denoise the noise part again,the purpose of protecting the effective signal and improving the denoising accuracy is achieved.In addition,during noise identification,a two-dimensional transformation is introduced: Gramian Angular Field(GAF)for preprocessing,which can expand the one-dimensional time series into two-dimensional space without destroying the signal timing relationship,thus increasing the spatial information of the one-dimensional signal,which is conducive to neural network identification of data segments containing cultural noise.(2)The structure of the neural network is improved.The U-shaped Convolutional Neural Network(UNet)in the convolutional network and the Gate Recurrent Unit(GRU)in the cyclic network are jointly built so that the newly built neural network can not only have good noise suppression ability but also protect the timing relationship of the time series from being damaged during network training.(3)The layer attribute in the neural network is improved by replacing the pooled layer with wavelet transform and changing the spatial dimension with linear interpolation.Compared with the maximum amplitude value,the wavelet maximum sparse coefficient can better represent the characteristics of cultural noise and is conducive to the feature information of noise being concerned by the neural network in the network training process.Finally,the synthetic and field data denoising experiments prove the superiority and effectiveness of the five improved AMT denoising methods proposed in this paper. |
