Data Denoising Method For Magnetic Resonance Sounding Based On Convolutional Neural Network

Magnetic Resonance Sounding(MRS)is a geophysical method for direct detection of groundwater.This method can quantitatively characterize the underground hydrogen proton abundance and the occurrence state of water bodies in the formation.Compared with the traditional geophysical water exploration technology,the MRS method has the advantages of rich information,unique interpretation,fast and efficient,so it is widely used in groundwater exploration,tunnel/mine water hazard detection and dam seepage.In the actual detection process,the MRS signal generated by the groundwater body is extremely weak,with an intensity of only nanovolts,and cannot be shielded during the measurement process.noise and random noise,etc.These noises result in a low signal-to-noise ratio of the data,which affects the accuracy of the inversion interpretation results.Therefore,it is crucial to denoise the measurement data and extract the MRS signal.The existing denoising methods can only remove a single type of noise,so it is necessary to use different denoising methods to process the noise in the MRS data multiple times.This process is not only complicated,the signal processing takes a long time,and Useful information is lost in each denoising process.Therefore,there is an urgent need for a fast and accurate denoising method for MRS data that can remove a variety of noises.Convolutional Neural Network(CNN)is one of the deep learning methods that has developed rapidly in recent years.By learning the image or signal features in the training set,CNN can realize the process of extracting useful images or signals from noisy data,so this method has been widely used in image denoising,speech separation,medical image processing and so on.Based on these studies,this paper designs a neural network structure and network learning configuration suitable for denoising MRS data.On this basis,this paper proposes a denoising method for MRS data based on Time-frequency Convolutional Neural Network(TFCN),and compares the TFCN method with other denoising methods based on simulated and measured data.By comparison,the effectiveness and reliability of the proposed method are verified.The network structure and learning parameters are the key factors that affect the denoising of MRS data by TFCN.In this paper,the root mean square error and signal-to-noise ratio improvement after denoising are used as evaluation indicators to compare different network structures and learning parameters.Neural network configuration suitable for denoising MRS data.Aiming at the random noise in the MRS data,a TFCN network is designed to suppress the random noise,and the denoising performance of the network is verified by using the simulation data.The results show that the TFCN has a good removal effect on the random noise in the MRS data.On this basis,the TFCN network for all types of noise removal was further designed and trained,and compared with the traditional MRS data denoising algorithm.In the comparison process,NEO+HMC+TFPF and NEO+HMC+EMD were selected to remove and Suppress spike noise,power frequency harmonic noise and random noise in MRS data.The results show that compared with NEO+HMC+TFPF,the SNR of TFCN can be improved by up to 10 d B,and compared with NEO+HMC+TFPF,the SNR of TFCN can be improved by up to 6 d B.The actual test was carried out in Taipingchi Reservoir in Changchun City,and the TFCN method was used to denoise the measured data to verify the effectiveness of the method.The results show that the denoising method based on TFCN can remove most of the noise components in the MRS data at one time,and by combining with The comparison of traditional denoising methods further verifies the superiority and efficiency of the TFCN method.The innovative work accomplished in this paper is as follows:1.For the denoising problem of MRS data,the time spectrum of the MRS signal is used instead of the time domain MRS signal as the training set of TFCN for signal feature extraction,which lays the foundation for subsequent TFCN training.2.The feature of pure MRS signal is learned by TFCN,and MRS signal is extracted from MRS data with random noise,which realizes effective suppression of random noise in simulated MRS data.3.On the basis of TFCN that suppresses random noise in MRS data,pure MRS signal and various noise data are used as training sets,and TFCN that removes various noises in noisy MRS data at one time is obtained by training.Both the data and the measured data show good results.