Research On Blind Source Separation Method For Surface NMR Using FastICA

China is with a shortage of fresh water resources.Efficient and accurate detection of underground available fresh water resources is an effective way to solve the lack of water resources.Surface Nuclear Magnetic Resonance(SNMR)can directly detect the Larmor precession of hydrogen protons in water,which is non-invasive and can directly detect the advantage of groundwater,and can invert the hydrogeological parameters such as water content,depth,porosity and permeability.The SNMR has been involved in a widely-used method.However,the intensity of the SNMR signal is at the order of nano-volt,and it is easily disturbed by electromagnetic interference and embedded in the noise.If the noise cannot be effectively removed,the late inversion interpretation of the results will be inaccurate.The typical noise source in SNMR includes spikes,powerline harmonics and random noise.In most of state-of-the-art literature,noise is removed when the signal and noise characteristics are known.But there is no uniform characteristics in some of this noise,and the mixing mechanism with uncertainty,which will lead to the efficiency of these Noise reduction algorithms are low,the performance is not optimal.To address the above problem,blind source separation(BSS)is applied to separate signal from noise for the SNMR.Since BSS can separate and recover the independent components of the source signal from the observed multi-source mixed signal in the case where both the source signal and the mixing process are unknown.In the many BSS methods,FastICA is used because FastICA algorithm has good stability and fast convergence speed,and can batch data.In this paper,we first use the data reconstruction method to restore the amplitude of the independent components,and compare the method with the spectral correction method.The simulation results show that the error of the data reconstruction method is less than that of the spectral correction method.Especially in the case of large random noise,such as signal to noise ratio below-37.1d B,the spectral correction method of the error is greater,even greater than 10%,which is out of the error range allowed.The data reconstruction error is less than 5% in the entire experiment,which meets the practical application requirements.Secondly,the powerline harmonics are removed and the simulation experiment is conducted.Removal powerline harmonics noise based on FastICA for SNMR is divided into two cases,one is the frequency harmonic noise frequency and the NMR signal Larmor frequency has a offset,that implies the correlation between the two is not strong;the other is the harmonic noise frequency and Larmor frequency is coincidence,that implies the correlation between the signal and harmonic is strong.In the second case,estimation of the powerline harmonic frequency is implemented by Gaussian interpolation method,and we construct the sinusoidal signal and cosine signal as the reference signal,which have the same frequency with powerline harmonic.Then,SNMR signal and powerline harmonics are separated by FastICA.And the separated results were compared with the notch filter result.The improvement of signal to noise ratio can be increased to 43.42 d B by FastICA.The improvement of signal to noise ratio can be increased to 36.75 d B by notch filter.The compared results show that the FastICA algorithm not only removes the noise,but also retain the NMR signal.The powerline harmonics with Larmor frequency is difficult to remove.In this case,the FastICA algorithm is used to remove the power frequency harmonics with other frequencies different from Larmor frequency.And then we use the late half part of the data,because the nuclear signal has been attenuated to zero in this part of data.The data contains the same frequency powerline harmonic noise and random noise,and the Gaussian interpolation method is used to estimate the frequency of the powerline harmonic.Correlation detection method is used to calculate the amplitude and phase of the harmonic noise.The estimated frequency,amplitude and phase can be used to construct the harmonic frequency waveform of the poweline harmonic waveform,and yields a small error.Finally,the estimated power frequency harmonic noise is subtracted and the removal of powerline harmonic with Larmor frequency is achieved.There is still some random noise remained in the data after removing the harmonic noise.FastICA is applied to suppress the random noise for SNMR.Only two datasets are sued in the simulation experiment to implement the separation of Gaussian random noise,which can not only improve the signal-to-noise ratio of data,but also greatly shorten the time of field survey time.Finally,the field detection is carried out in the two places of Beihu Park and Shaoguo Town respectively.The FastICA algorithm is used to process the power frequency harmonic noise and random noise respectively.The improvement of signal to noise ratio can be increased to 27.18 d B in Beihu Park.The improvement of signal to noise ratio can be increased to 33.53 d B in Shaoguo Town.It verifies the validity and practicality of the algorithm.