Research On The Application Of Relevance Vector Machine In Airborne Transient Electromagnetic Denoising

Airborne transient electromagnetic method(ATEM)is an airborne geophysical exploration method which carries transient electromagnetic instruments on an flight platform.ATEM is widely used in mineral exploration,groundwater resources exploration,basic geological survey,engineering,environmental exploration,marine exploration and other fields because of its advantages of large exploration depth,high exploration efficiency,relatively low cost,large-scale exploration and overcoming complex terrain.The secondary field signal of ATEM is a kind of broadband signal.In practical detection,it is vulnerable to the influence of noise in different frequency bands,such as sferics noise,random noise,man-made EM disturbances,etc.,which leads to serious degradation of data quality and affects the later data processing.Therefore,it is of great theoretical significance and application value to study the method of airborne transient electromagnetic de-noising.Conventional noise suppression methods are mainly based on time-frequency domain filtering and wavelet decomposition,often aiming at a single noise,but there is overlap between airborne transient electromagnetic signal and noise in time-frequency domain,and conventional processing methods often fail to achieve the desired results.In this paper,with the support of the 13 th Five-Year National Key Research and Development Project "Research and Development of Data Processing and Interpretation Software System for Heilicopter-based Airborne Electromagnetic Method ",the denoising method of airborne transient electromagnetic signal is studied.Because of the randomness of airborne transient electromagnetic noise,it can be treated as a random signal.Therefore,based on statistical machine learning theory,this paper introduces relevance vector machine method and support vector machine method based on statistical machine learning theory into airborne transient electromagnetic signal denoising,and analyzes the feasibility and effectiveness of applying them to airborne electromagnetic signal filtering.The main contents of the dissertation are described as follows.(1)One-dimensional forward and 2.5-dimensional forward processes are analyzed to generate forward data.According to the Gaussian characteristics and the characteristics of sferics noise in measured data,the Gaussian white noise model is used to simulate the signal disturbances with Gaussian characteristics,and the simulated sferics noise is generated by autoregressive moving average model(ARMA).The simulated measured noise data are synthesized to provide data basis for further de-noising analysis.(2)Airborne transient electromagnetic noise reduction method based on support vector machine(SVM).Firstly,the kernel training sample matrix is constructed by using the generated simulated noisy response,and the model and parameters reflecting the real response are obtained by using the optimization method of support vector machine.The noise samples are predicted to achieve the purpose of denoising.Then the denoising effect is evaluated by signal-to-noise ratio(SNR),and the influence of different kernel functions and parameters on SNR is also analyzed.Finally,the denoising performance of two-dimensional simulated profile data,inversion results and measured data is further evaluated.The experimental results show that the SNR after denoising is improved to a certain extent,and the SVM denoising method based on the Gauss kernel function has a larger SNR,which indicates that the SVM method can be applied to the airborne transient electromagnetic denoising.The simulated two-dimensional profile denoising and inversion results as well as the measured data denoising show that the SVM method is feasible in the airborne transient electromagnetic denoising.(3)Airborne transient electromagnetic de-noising method based on relevance vector machine(RVM).Firstly,the training sample matrix of base function is constructed by using the generated simulated noisy response.The posterior distribution of parameters and hyper-parameters is obtained by optimizing learning according to the maximum posterior criterion of RVM.The probability distribution of test samples is obtained by predicting and estimating the noise samples.Then the performance of denoising is evaluated by SNR,and the effects of different kernel functions and parameters on SNR are analyzed.In addition,the performance of denoising is compared with that of traditional wavelet method.Finally,the denoising performance of two-dimensional simulated profile data,inversion results and measured data is further evaluated.The experimental results show that the SNR after denoising is greatly improved,and the RVM method based on Laplacian kernel function has a higher SNR.The simulated two-dimensional profile denoising and inversion results as well as the measured data denoising show that the RVM method is feasible in the airborne transient electromagnetic denoising.Compared with the SVM denoising method and the wavelet denoising method,the RVM is better applied to the airborne transient electromagnetic denoising.In this paper,SVM and RVM based on statistical learning theory are applied to the experiment and processing of airborne transient electromagnetic signal de-noising,and good results are obtained,which proves the feasibility and validity of using them to de-noising.