Intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification

Abstract

The application discloses a method for intelligently selecting power spectrum of magnetotelluric sounding method based on deep learning classification, and specifically comprises the following steps: S1: segmenting MT electric field and magnetic field time series data collected in the field, calculating power spectrum of each time series, and obtaining MT sub-power spectrum; S2: respectively performing impedance estimation on all MT sub-power spectrum and total power spectrum after full stacking, and respectively calculating apparent resistivity and phase; S3: using apparent resistivity and phase after full stacking to perform Rhoplus inversion, and calculating fitting difference between resistivity and phase obtained by inversion and apparent resistivity and phase of the sub-power spectrum. The method for intelligently selecting power spectrum of magnetotelluric sounding method based on deep learning classification has the effects that in MT measured data processing, manpower cost can be greatly saved, uncertainty and randomness of manual selection of power spectrum can be avoided, and the influence of noise on impedance estimation of MT data can be effectively inhibited.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical exploration and application, and particularly relates to a magnetotelluric sounding method power spectrum intelligent selection method based on deep learning classification. BACKGROUND

[0002] The magnetotelluric sounding method (MT) is an electromagnetic exploration method using a natural field source. The natural field signal is unstable, irregular, weak and susceptible to electromagnetic signal interference. In order to reduce the influence of electromagnetic noise interference, magnetotelluric data processing needs to be subjected to a series of data processing means to obtain impedance, apparent resistivity, phase, and dip required for subsequent inversion, among which the core is magnetotelluric power spectrum calculation and tensor impedance element estimation.

[0003] In order to suppress noise, the magnetotelluric time series is often segmented, a plurality of power spectra are calculated, and the selection of the power spectrum is completed through human-computer interaction. This process is time-consuming and laborious, and is seriously dependent on personal experience and has great randomness. SUMMARY

[0004] The present application discloses a magnetotelluric sounding method power spectrum intelligent selection method based on deep learning classification, which aims to solve the technical problem that in order to suppress noise, the magnetotelluric time series is often segmented, a plurality of power spectra are calculated, and the selection of the power spectrum is completed through human-computer interaction. This process is time-consuming and laborious, and is seriously dependent on personal experience and has great randomness.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] The magnetotelluric sounding method power spectrum intelligent selection method based on deep learning classification specifically comprises the following steps:

[0007] S1: segmenting the MT electric field and magnetic field time series data collected in the field, calculating the power spectrum of each time series, and obtaining the MT sub-power spectrum;

[0008] S2: estimating the impedance of all MT sub-power spectra and the total power spectrum after full stacking, respectively, and calculating the apparent resistivity and phase;

[0009] S3: using the apparent resistivity and phase after full stacking to perform Rhoplus inversion, and calculating the fitting difference between the resistivity and phase obtained by inversion and the apparent resistivity and phase of the sub-power spectrum;

[0010] S4: performing deep neural network classification after data normalization;

[0011] S5: selecting the stacking spectrum of the selected power spectrum after classification to calculate the apparent resistivity and phase through impedance estimation;

[0012] S6: Calculate the apparent resistivity and phase of the selected power spectrum and perform Rhoplus inversion calculation to see if the fitting error meets the requirements;

[0013] S7: When the requirements are met, save the classification results, otherwise, calculate the fitting difference between the apparent resistivity and phase of the last selected power spectrum and the inversion, and perform deep neural network classification again until the fitting error meets the requirements.

[0014] S8: Output the intelligent selected MT standard format power spectrum EDI file;

[0015] The intelligent selection result of the MT sounding method power spectrum based on Rhoplus correction and deep learning classification can effectively improve the quality of the measured MT data compared with the full stacking result of the power spectrum. The analysis of the artificial data screening result shows that the deep neural network classification can avoid the uncertainty of manual selection of power spectrum and effectively suppress the influence of noise on impedance estimation, which is better than the manual editing result.

[0016] In a preferred scheme, the Rhoplus inversion is a Rhoplus method of inverting apparent resistivity and phase parameters, and the Rhoplus method is different from the conventional least squares iterative fitting inversion method. The Rhoplus method performs optimal fitting of observation data by constructing a physically effective ideal geoelectric model, and uses an optimization method to obtain a stable numerical solution. The Rhoplus method specifically includes the following calculation steps:

[0017] In the MT method, the parameter c(ω) is defined using the electric field and magnetic field components on the ground:

[0018] c(ω) = E(ω) / iωB(ω)

[0019] where ω is the circular frequency, E(ω) and B(ω) are the frequency spectra of the electric field and magnetic field components, respectively; the apparent resistivity ρ s (ω) and the phase Φ(ω) have the following relationship with the parameter c(ω):

[0020] ρ S (ω) = μ0ω|c(ω)| 2

[0021]

[0022] The relationship between c(ω) and ρ s (ω), Φ(ω) is further arranged as:

[0023]

[0024] For one-dimensional geoelectric section, c(ω) can be expressed as the following integral form:

[0025]

[0026] Where a0, λ, a(λ) are real numbers, and λ≥0, a(λ)≥0; the integral expression is discretized, and the above formula becomes:

[0027]

[0028] For two-dimensional geoelectric section, the parameter c(ω) defined by the electric field and magnetic field components corresponding to the TM mode has the above expression form corresponding to the one-dimensional geoelectric section, and the Rhoplus method is finally expressed as the following discrete expression:

[0029]

[0030]

[0031] Where ω=iv n is the solution of c(ω);

[0032] The right side of the integral expression can be regarded as the response corresponding to the mathematical model composed of a series of δ functions, which shows that for one-dimensional geoelectric section and two-dimensional geoelectric section of TM mode, the magnetotelluric response is always corresponding to a mathematical model composed of a series of δ functions. s And the phase always exists.

[0033] In a preferred scheme, in the S4 step, the deep learning classification includes a Logistic regression algorithm, and the Logistic regression algorithm is used to study the nonlinear superposition relationship of the prediction value y and each dimension of the sample x, the Logistic regression algorithm includes a Logistic regression model, and the formula of the Logistic regression model is

[0034]

[0035] The MT data screened by artificial data is used as the training set of the sample to train the classification model, the MT data sample without editing is used as the test set, the resistivity and phase and the fitting difference data of the RhoPlus inversion are normalized and used as the input of the network, the deep neural network can extract data features, and the Logistic regression algorithm can imitate the logical processing of neuron signal transmission in biological neuroscience through the nonlinear Sigmiod activation function, and classify the important features of the input.

[0036] As can be seen from the above, the power spectrum intelligent selection method of magnetotelluric sounding based on deep learning classification specifically includes the following steps:

[0037] S1: segment the MT electric field and magnetic field time series data collected in the field, calculate the power spectrum of each time series, and obtain the MT sub-power spectrum;

[0038] S2: impedance estimation is performed on all MT sub-power spectra and the total power spectrum after full stacking, and apparent resistivity and phase are calculated respectively;

[0039] S3: Rhoplus inversion is performed using the apparent resistivity and phase after full stacking, and the fitting error between the calculated resistivity and phase and the apparent resistivity and phase of the sub-power spectrum is calculated;

[0040] S4: after data normalization, deep neural network classification is performed;

[0041] S5: after classification, the stacking spectrum of the selected power spectrum is used for impedance estimation to calculate the apparent resistivity and phase;

[0042] S6: the apparent resistivity and phase calculated by the selected power spectrum are used for Rhoplus inversion to calculate whether the fitting error meets the requirements;

[0043] S7: when the requirements are met, the classification result is saved, otherwise the fitting error between the apparent resistivity and phase of the last selected power spectrum and the inversion is calculated, and the deep neural network classification is performed again until the fitting error meets the requirements.

[0044] S8: output the intelligent selected MT standard format power spectrum EDI file. The deep learning classification based MT sounding method power spectrum intelligent selection method provided by the present application has the power spectrum selection method of traditional MT data preprocessing, which is artificial selection of sub-power spectrum after segmentation of time series and then stacking. Although the overall data quality is improved, it is time-consuming and laborious, and it is seriously dependent on personal experience, and has great randomness. According to the characteristics of MT power spectrum impedance estimation, a deep learning classification method for MT sounding method power spectrum selection based on Rhoplus correction is proposed. The present application can greatly save labor cost in MT data processing, can avoid the uncertainty and randomness of artificial selection of power spectrum, and effectively suppresses the influence of noise on MT data impedance estimation. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The overall flowchart of the deep learning classification based MT sounding method power spectrum intelligent selection method provided by the present application.

[0046] Figure 2 The root mean square error comparison table of different data processing methods of the deep learning classification based MT sounding method power spectrum intelligent selection method provided by the present application.

[0047] Figure 3 The MT data before and after the deep learning classification of the intelligent power spectrum selection method for magnetotelluric sounding based on deep learning classification proposed in the application is compared.

[0048] Figure 4 The deep neural network classification and artificial data screening result comparison chart of the intelligent power spectrum selection method for magnetotelluric sounding based on deep learning classification proposed in the application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application.

[0050] The intelligent power spectrum selection method for magnetotelluric sounding based on deep learning classification disclosed in the application is mainly applied to solve the scene that manual selection of power spectrum is time-consuming and laborious, subjective and random in MT data processing.

[0051] In order to reduce the influence of electromagnetic interference on impedance tensor estimation of magnetotelluric data, solve the problem that manual selection of power spectrum is time-consuming and laborious, subjective and random in MT data processing, an intelligent power spectrum selection method for magnetotelluric sounding based on deep learning classification is proposed. The method calculates the difference between the impedance estimation result of the full stack of the segmented MT time sequence sub-power spectrum and the Rhoplus inversion, uses a deep neural network to remove the sub-power spectrum that deviates from the Rhoplus inversion result, re-impedance estimates the stacked power spectrum, calculates the fitting error between the selected and Rhoplus inversion, and repeats the above process to finally find the best power spectrum combination.

[0052] REFERENCE Figure 1 The intelligent power spectrum selection method for magnetotelluric sounding based on deep learning classification specifically comprises the following steps:

[0053] S1: Segment the MT electric field and magnetic field time sequence data collected in the field, calculate the power spectrum of each time sequence, and obtain the MT sub-power spectrum;

[0054] S2: Impedance estimation is performed on all MT sub-power spectra and the total power spectrum after full stack, and the apparent resistivity and phase are calculated respectively;

[0055] S3: Rhoplus inversion is performed using the apparent resistivity and phase after full stack, and the fitting difference between the resistivity and phase obtained by inversion and the apparent resistivity and phase of the sub-power spectrum is calculated;

[0056] S4: After data normalization, deep neural network classification is performed;

[0057] S5: After classification, the superimposed spectrum of the selected power spectrum is used for impedance estimation calculation of apparent resistivity and phase;

[0058] S6: The apparent resistivity and phase calculated from the selected power spectrum are used for Rhoplus inversion calculation again to determine whether the fitting error meets the requirements;

[0059] S7: When the requirements are met, the classification results are saved, otherwise the fitting difference between the apparent resistivity and phase of the last classification selected power spectrum and the inversion is calculated again for deep neural network classification until the fitting error meets the requirements.

[0060] S8: Output the intelligent selected MT standard format power spectrum EDI file;

[0061] The intelligent selection results of the MT sounding method power spectrum based on Rhoplus correction and deep learning classification can effectively improve the quality of the measured MT data compared with the full superimposed results of the power spectrum. The analysis of the artificial data selection results shows that the deep neural network classification can avoid the uncertainty of manual selection of power spectrum and effectively suppress the influence of noise on impedance estimation, which is better than the manual editing results.

[0062] Reference Figure 3 , Figure 3 The comparison chart of MT data before and after deep learning classification shows that deep learning classification can effectively improve the quality of MT data in the "dead frequency band" and is significantly better than the full superimposed impedance estimation of power spectrum, effectively reducing the influence of noise on impedance estimation.

[0063] In a preferred embodiment, the Rhoplus inversion is a Rhoplus method for inverting apparent resistivity and phase parameters, and the Rhoplus method is different from the conventional least squares iterative fitting inversion method. The Rhoplus method performs optimal fitting of observation data by constructing a physically effective ideal geoelectric model, and uses an optimization method to obtain a stable numerical solution;

[0064] Since the Rhoplus method considers the correlation of impedance apparent resistivity and phase data and the continuity of data in the frequency domain, and has a complete theoretical basis and clear physical background, it provides an important theoretical basis for detecting the distortion degree of the MT sounding curve and studying the distortion curve correction method.

[0065] In a preferred embodiment, the Rhoplus method specifically includes the following calculation steps:

[0066] In the MT method, the parameter c(ω) is defined using the electric field and magnetic field components on the ground surface:

[0067] c(ω) = E(ω) / iωB(ω)

[0068] where ω is the circular frequency, E(ω), B(ω) are the frequency spectrum of electric and magnetic field components respectively; apparent resistivity p s The following relationship exists between the amplitude A(ω) and phase Φ(ω) and the parameter c(ω):

[0069] p S (ω) = μ0ω|c(ω)| 2

[0070]

[0071] c(ω) and p s The relationship between A(ω), Φ(ω) and c(ω) is further arranged as:

[0072]

[0073] In a preferred embodiment, for a one-dimensional geoelectric section, c(ω) can be expressed in the following integral form:

[0074]

[0075] where a0, λ, a(λ) are real numbers, and λ≥0, a(λ)≥0; the integral expression is discretized, and the above formula becomes

[0076]

[0077] For a two-dimensional geoelectric section, the parameter c(ω) defined by the electric and magnetic field components corresponding to the TM mode has the above expression corresponding to the one-dimensional geoelectric section.

[0078] In a preferred embodiment, the Rhoplus method is finally expressed as:

[0079]

[0080]

[0081] where ω = iv n is the solution of c(ω);

[0082] The right side of the integral expression can be regarded as the response corresponding to the mathematical model composed of a series of δ functions, which indicates that for the one-dimensional geoelectric section and the TM mode of the two-dimensional geoelectric section, the apparent resistivity p s and the phase always have a mathematical model composed of a series of δ functions corresponding thereto.

[0083] In a preferred embodiment, in the S4 step, the deep learning classification includes a Logistic regression algorithm, and the Logistic regression algorithm is used to study the nonlinear superposition relationship of the predicted value y and each dimension of the sample x, the Logistic regression algorithm includes a Logistic regression model, and the formula of the Logistic regression model is

[0084]

[0085] In a preferred embodiment, the MT data screened by manual data screening is used as a training set of samples to train the classification model, the unedited MT data sample is used as a test set, the normalized resistivity and phase and RhoPlus inversion fitting difference data are used as the input of the network, the deep neural network can extract data features, and the Logistic regression algorithm can simulate the logical processing of neuron signal transmission in biological neural science through a nonlinear Sigmiod activation function to classify the important features of the input.

[0086] Referring to Figure 4 , Figure 4 The comparison chart of the deep neural network classification and the manual data screening result shows that the deep neural network classification has good noise suppression effect and achieves similar results to manual editing.

[0087] The above shows and describes the basic principles and main features of the present application and the advantages of the present application, and it is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application, therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application, and any reference signs in the claims should not be regarded as limiting the claims.

[0088] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description manner of the specification is only for clarity, those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be properly combined to form other embodiments that those skilled in the art can understand.

Claims

1. A method for intelligent selection of power spectrum in magnetotelluric sounding based on deep learning classification, characterized in that, Specifically, the following steps are included: S1: Segment the time series data of MT electric field and magnetic field collected in the field, calculate the power spectrum of each time series segment, and obtain the MT sub-power spectrum; S2: Impedance estimation is performed on the power spectra of all MT sub-subs and the total power spectrum after full superposition, and the apparent resistivity and phase are calculated respectively. S3: Perform Rhoplus inversion using the fully superimposed apparent resistivity and phase, and calculate the fitting difference between the inverted apparent resistivity and phase and the apparent resistivity and phase of the sub-power spectrum; S4: Data normalized by power spectrum, apparent resistivity, phase and fit difference are classified by deep neural network. S5: After classification, perform impedance estimation on the superimposed spectrum of the selected power spectrum, and calculate the apparent resistivity and phase; S6: Perform Rhoplus inversion again on the apparent resistivity and phase obtained from the selected power spectrum, and calculate whether the fitting difference meets the requirements. S7: If the requirements are met, save the classification results; otherwise, if the requirements are not met, calculate the apparent resistivity and phase of the power spectrum selected in the previous classification and the fitting difference between the inversion and the deep neural network classification again until the fitting difference meets the requirements. S8: Outputs the intelligently selected magnetotelluric standard format power spectrum EDI file.

2. The intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification according to claim 1, characterized in that, The Rhoplus inversion is a method for inverting apparent resistivity and phase parameters. Unlike conventional least-squares iterative fitting inversion methods, the Rhoplus method constructs a physically valid ideal geoelectric model to achieve optimal fitting of the observation data and obtains a stable numerical solution using optimization methods.

3. The intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification according to claim 2, characterized in that, The Rhoplus method specifically includes the following calculation steps: In magnetotelluric methods, parameters are defined using the electric and magnetic field components of the Earth's surface. : in It is the angular frequency. , The spectra of the electric and magnetic field components are shown respectively; apparent resistivity. and phase With parameters The following relationship exists: ； and , The relationship between them can be further summarized as follows: 。 4. The intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification according to claim 3, characterized in that, For a one-dimensional geoelectric cross section This can be expressed in the following integral form: ； in , , are real numbers, and , ; Discretizing the integral expression, the above equation becomes... ； For a two-dimensional geoelectric profile, the parameters are defined by the electric and magnetic field components corresponding to the TM mode. It has the above-mentioned expression form corresponding to a one-dimensional geoelectric cross-section.

5. The intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification according to claim 4, characterized in that, The Rhoplus method ultimately yields the following integral expression: ， in yes The solution; The right side of the integral can be viewed as consisting of a series of The response corresponding to the mathematical model composed of functions indicates that, for the TM models of one-dimensional and two-dimensional geoelectric sections, the apparent resistivity of the magnetotelluric response... There is always one phase consisting of a series The mathematical model composed of functions corresponds to this.

6. The intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification according to claim 1, characterized in that, In step S4, the deep learning classification includes a Logistic regression algorithm, which is used to study the nonlinear superposition relationship between the predicted value y and the sample x across various dimensions. The Logistic regression algorithm includes a Logistic regression model, and the formula for the Logistic regression model is: 。 7. The intelligent selection method for power spectrum of magnetotelluric sounding based on deep learning classification according to claim 6, characterized in that, The MT data, after being manually selected, is used as the training set to train the classification model. Unedited MT data samples are used as the test set. The apparent resistivity and phase, along with the fitting difference data obtained from the Rhoplus inversion, are normalized and used as the input to the network. The deep neural network can extract data features, and the Logistic Regression algorithm, through the non-linear Sigmoid activation function, can mimic the logical processing of neuronal signal transmission in biological neuroscience to classify the important input features.

Images