A magnetotelluric data processing method, system, electronic device and storage medium

Abstract

This invention discloses a magnetotelluric data processing method, system, electronic device, and storage medium, relating to the field of electromagnetic data processing technology. The method includes: acquiring open-source data; the open-source data includes magnetotelluric time series and corresponding calibration results, as well as user-shared data uploaded through an open-source platform; the calibration results include noise intensity levels from strongest to weakest: Level I, Level II, Level III, Level IV, and Level V; using the open-source data and a semi-supervised learning method, optimizing the model parameters of the generative model and the discriminative model respectively to obtain a trained model; using the trained model for noise identification; and processing the noise identification results output by the trained model using a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series. This invention enables the application of semi-supervised learning to magnetotelluric data, fully utilizing existing data resources and solving the problems of difficult or costly data annotation.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic data processing technology, and in particular to a magnetotelluric data processing method, system, electronic device, and storage medium. Background Technology

[0002] The magnetotelluric field (MTF) refers to the time-varying electric and magnetic fields within the Earth's natural electromagnetic field. Its period range is extremely wide, and it is caused by external factors, specifically the electromagnetic effects resulting from the continuous stream of particles emitted by the sun and electromagnetic radiation in the space surrounding the Earth. The MTF can be categorized into several forms, including electromagnetic storms, electromagnetic diurnal variations, electromagnetic pulsations, and lightning. These are the field sources for methods such as magnetotelluric sounding. Since the field source of the MTF can be considered infinitely large compared to geological bodies, it is generally viewed as a plane electromagnetic wave incident vertically from high altitudes to the ground, greatly simplifying the complex field source problem. When using electromagnetic fields with periods greater than 2500 seconds to study the Earth's deep structure, the size of the field source and the shape of the Earth must be considered. Therefore, the research and study of magnetotellurics is crucial. However, existing semi-supervised learning methods for processing MTF data still have many problems during operation. Because the supervision and learning of magnetotelluric data is extremely difficult, and is limited by various factors such as research capabilities, geographical environment, and technological constraints, as well as the huge investment of equipment, manpower, and material resources, the long-term supervised learning effect obtained from closed-source data processing is unlikely to achieve the expected results. Summary of the Invention

[0003] The purpose of this invention is to provide a magnetotelluric data processing method, system, electronic device, and storage medium that can apply semi-supervised learning to magnetotelluric data, make full use of existing data resources, and solve the problems of difficult or costly data annotation.

[0004] To achieve the above objectives, the present invention provides the following solution: A magnetotelluric data processing method includes: The data obtained includes open-source data, which comprises labeled and unlabeled datasets. The labeled datasets include magnetotelluric time series under various environments and corresponding calibration results. These environments include different regions, geological structures, and industrial environments. The calibration results include different levels of noise intensity, categorized into five levels from strongest to weakest: Level I, Level II, Level III, Level IV, and Level V. The unlabeled datasets are user-shared data uploaded through an open-source platform. Using the aforementioned open-source data and semi-supervised learning methods, the model parameters of the generative model and the discriminative model are optimized respectively to obtain a trained model; the trained model is used for noise recognition. The noise identification results output by the trained model are processed by a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series; the corresponding algorithm includes elimination, data processing, time-frequency conversion and robust regression; the data processing includes, but is not limited to, wavelet analysis and cepstral analysis.

[0005] Optionally, before obtaining the open-source data, the method further includes: An open-source platform is established; the open-source platform is used to share data, enabling users to perform semi-supervised magnetotelluric data processing by freely creating training sets and setting calibration data.

[0006] Optionally, the method for determining the labeled dataset is as follows: Acquire magnetotelluric time series under multiple environments; The magnetotelluric time series is subjected to feature preprocessing to determine each feature parameter; Based on the aforementioned characteristic parameters, the noise intensity level is calibrated using interactive methods to obtain the calibration results corresponding to the magnetotelluric time series.

[0007] Optionally, the step of optimizing the model parameters of the generative model and the discriminative model using the open-source data and the semi-supervised learning method to obtain a trained model specifically includes: Based on the aforementioned open-source data, the EM algorithm is used to optimize the parameters of the learning objective of the generative model, resulting in a well-trained generative model. The learning objective of the generative model is constructed based on the joint probability distribution. The parameter optimization is performed by maximizing the likelihood of the unlabeled dataset. Based on the aforementioned open-source data, the learning objective of the discriminative model is optimized using a self-training algorithm to obtain a well-trained discriminative model; the learning objective of the discriminative model is constructed based on a conditional probability distribution.

[0008] Optionally, the step of optimizing the parameters of the learning objective of the generative model using the EM algorithm based on the open-source data to obtain a well-trained generative model specifically includes: In the E-step of the EM algorithm, the posterior probability P(Y|X,θt) of the unlabeled dataset is estimated using the current parameter θt. Based on Bayes' theorem, the posterior probability is expressed as the ratio of the product of the likelihood probability and the prior probability to the marginal probability. P(Y|X,θt) = P(X|Y,θt) P(Y|θt) / P(X) Where P(Y|X,θt) represents the posterior probability of noise intensity Y given input X; P(X|Y,θt) represents the likelihood probability of input X given noise intensity Y; P(Y|θt) represents the prior probability of noise intensity Y; and P(X) represents the marginal probability. In the M-step of the EM algorithm, the model parameters θt+1 are updated by maximizing the likelihood of the unlabeled data, and the likelihood of the unlabeled data is represented as the sum of log-likelihood probabilities: θ_{t+1} = argmax_θ Σ_{X∈Du} Σ_{Y∈{I,II,III,IV,V}} P(Y|X,θt) log P(X,Y|θ) In this process, the outer summation iterates through all samples X in the unlabeled dataset Du, and the inner summation iterates through all possible noise intensity levels Y; P(Y|X,θt) represents the posterior probability obtained in step E, and P(X,Y|θ) is the joint probability distribution of the generative model for the data.

[0009] Optionally, the step of optimizing the parameters of the discriminative model's learning objective using a self-training algorithm based on the open-source data to obtain a well-trained discriminative model specifically includes: Suppose we have a labeled dataset D_l={(x_1, y_1), (x_2, y_2), ..., (x_n, y_n)}, where n is the total number of samples in the labeled dataset. We use maximum likelihood estimation to estimate the conditional probabilities: Θ^* = argmax_θ Σ_{i=1}^{n} log P(Y=y_i | X=x_i, θ) Where Θ^* represents the optimal parameters of the discriminant model; θ represents the parameters of the discriminant model; the summation iterates through all samples in the labeled dataset, where i is the sample index from 1 to n; P(Y=y_i|X=x_i, θ) represents the conditional probability that the noise intensity is y_i given the input x_i; The optimal parameters of the discriminant model are obtained by optimizing the model parameters by maximizing the likelihood of the labeled data; then, the decision boundary of the model is adjusted using information from the unlabeled dataset to obtain the trained discriminant model.

[0010] Optionally, the step of processing the noise identification results output by the trained model using a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series specifically includes: When the noise recognition result output by the trained model is level I, the data is directly discarded; When the noise identification result output by the trained model is level II or level III, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated using the data processing method, the time-frequency conversion and the robust regression. When the noise identification results output by the trained model are level IV and level V, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated using the time-frequency conversion and the robust regression.

[0011] The present invention also provides a magnetotelluric data processing system, comprising: The data acquisition unit is used to acquire open-source data, which includes labeled and unlabeled datasets. The labeled datasets include magnetotelluric time series under various environments and corresponding calibration results. The environments include different regions, geological structures, and industrial environments. The calibration results include different levels of noise intensity. The noise intensity includes five levels, from strongest to weakest: Level I, Level II, Level III, Level IV, and Level V. The unlabeled datasets are user-shared data uploaded through the open-source platform. The model training unit is used to optimize the model parameters of the generative model and the discriminative model using the open-source data and semi-supervised learning methods, respectively, to obtain a trained model; the trained model is used for noise recognition. The apparent resistivity curve and phase curve determination unit is used to process the noise identification results output by the trained model using a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series. The corresponding algorithm includes elimination, data processing, time-frequency conversion, and robust regression. The data processing includes, but is not limited to, wavelet analysis and cepstral analysis.

[0012] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to cause the electronic device to perform the magnetotelluric data processing method described above.

[0013] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the magnetotelluric data processing method described above.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a magnetotelluric data processing method, system, electronic device, and storage medium. The method includes acquiring open-source data, including labeled and unlabeled datasets, and using these two datasets to optimize the model parameters of a generative model and a discriminative model through semi-supervised learning. The noise intensity recognition results output by the trained models are then used for corresponding algorithmic calculations to ultimately obtain the apparent resistivity and phase curves of the magnetotelluric time series. This invention enables model training using both labeled and unlabeled data to achieve better performance and generalization ability, and fully utilizes existing data resources, solving the problems of difficult or costly data labeling. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the magnetotelluric data processing method of the present invention. Figure 2 This is a logical schematic diagram of the magnetotelluric data processing method in this embodiment. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The purpose of this invention is to provide a magnetotelluric data processing method, system, electronic device, and storage medium that can apply semi-supervised learning to magnetotelluric data, make full use of existing data resources, and solve the problems of difficult or costly data annotation.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figures 1-2 As shown, the present invention provides a magnetotelluric data processing method, comprising: Step 100: Obtain open-source data; the open-source data includes labeled and unlabeled datasets; the labeled datasets include magnetotelluric time series under various environments and corresponding calibration results; the environments include different regions, geological structures, and industrial environments; the calibration results include different levels of noise intensity; the noise intensity includes 5 levels, from strongest to weakest, namely Level I, Level II, Level III, Level IV, and Level V; the unlabeled datasets are user-shared data uploaded through the open-source platform; Step 200: Using the open-source data and semi-supervised learning method, optimize the model parameters of the generative model and the discriminative model respectively to obtain a trained model; the trained model is used for noise recognition. Step 300: The noise identification results output by the trained model are processed by a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series; the corresponding algorithm includes elimination, data processing, time-frequency conversion and robust regression; the data processing includes, but is not limited to, wavelet analysis and cepstral analysis.

[0021] As a further implementation, before obtaining the open-source data, the following steps are also included: An open-source platform is established; the open-source platform is used to share data, enabling users to perform semi-supervised magnetotelluric data processing by freely creating training sets and setting calibration data.

[0022] The method for determining the labeled dataset in step 100 is as follows: Acquire magnetotelluric time series under multiple environments; perform feature preprocessing on the magnetotelluric time series to determine each feature parameter; based on each feature parameter, use interactive methods to calibrate the noise intensity level to obtain the calibration result corresponding to the magnetotelluric time series.

[0023] As one specific implementation of step 200, it specifically includes: Based on the aforementioned open-source data, the EM algorithm is used to optimize the parameters of the learning objective of the generative model, resulting in a well-trained generative model. The learning objective of the generative model is constructed based on the joint probability distribution. The parameter optimization is performed by maximizing the likelihood of the unlabeled dataset. Based on the aforementioned open-source data, the learning objective of the discriminative model is optimized using a self-training algorithm to obtain a well-trained discriminative model; the learning objective of the discriminative model is constructed based on a conditional probability distribution.

[0024] As one specific implementation of step 300, it specifically includes: When the noise identification result output by the trained model is Level I, the data is directly discarded; when the noise identification result output by the trained model is Level II or Level III, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated using the data processing method, the time-frequency conversion, and the robust regression; when the noise identification result output by the trained model is Level IV or Level V, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated using the time-frequency conversion and the robust regression.

[0025] Based on the above implementation scheme, a design idea for semi-supervised learning is presented. This idea utilizes limited labeled data and abundant unlabeled data, improving model performance by incorporating information from the unlabeled data into the model's learning process. Its core assumption is that there is a certain correlation between the concepts represented by unlabeled and labeled data, and this correlation can be used to enhance the model's generalization ability.

[0026] Semi-supervised learning methods can be divided into two types: generative models and discriminative models. Generative models are based on probabilistic models, modeling the distribution of data and using this model to estimate unlabeled data. Discriminative models, on the other hand, directly model the decision boundary, adjusting it by utilizing information from unlabeled data.

[0027] In summary, applying semi-supervised learning to magnetotelluric data involves training models using both labeled and unlabeled data to achieve better performance and generalization ability. By utilizing information from unlabeled data, existing data resources can be fully utilized, making it particularly suitable for problems where data labeling is difficult or costly. Based on this idea, the specific data processing steps are as follows: S1, Open Source Processing: For magnetotelluric data supervision and learning, instead of the traditional approach of independent institutions or organizations, an open-source platform is established to collaboratively promote and accelerate the development of the magnetotelluric data field. The role of shared data is to help users establish their own magnetotelluric semi-supervised data processing workflows, giving users greater freedom in creating training sets and configuring calibration data.

[0028] S2, Shared Data: Once the open-source platform is established, it can be used for collective supervised learning or operate independently but provide supervision and learning results. Organizations and individuals within the open-source platform can upload and share the supervision and learning data they generate.

[0029] S3, Supervised Training Set (i.e., Labeled Dataset) Labeling: When numerous organizations and individuals collaborate on open-source projects, the supervision and learning of the Earth's magnetic field will no longer be limited by factors such as funding, geographical environment, and technological constraints. It can be freely and diversely learned through open-source network platforms. The algorithm flow is as follows: (1) Preprocess the data: calculate the spectral characteristics (power spectral density, band energy ratio, frequency peak, etc.), time-frequency characteristics (wavelet packet coefficients, short-time Fourier transform coefficients, time-frequency graph, etc.), polarization characteristics (polarization transfer matrix or polarization impulse response, etc.), and manually mark the noise type and time window; (2) Noise content calibration of data: Based on the parameters in (1), combined with human interaction, noise intensity calibration of magnetotelluric time series is performed to form a supervised learning training set with magnetotelluric time series as input and noise intensity as output; the noise intensity is divided into 5 levels, with I being the strongest, II and III being in the middle, and IV and V being the weakest.

[0030] S4, Establishing the generative model: Regarding the generative model, this embodiment assumes that there is a labeled dataset D. l ={(x1, y1), ..., (x n ,y n )}, where x i This represents the input magnetotelluric time series, y i The labels represent the corresponding noise intensities. The unlabeled dataset is D. u ={x1, ..., x m}, where x i This represents the input magnetotelluric time series. The unlabeled data for generating the model is provided collaboratively by users sharing the data.

[0031] The goal of the generative model is to learn the joint probability distribution P(X,Y), which is the joint distribution of magnetotelluric time series and noise intensity.

[0032] Generative models can optimize their parameters by maximizing the likelihood of unlabeled data. The semi-supervised EM algorithm is used for optimization, and the specific process is as follows: In the E-step of the EM algorithm, the posterior probability P(Y|X,θt) of the unlabeled dataset is estimated using the current parameter θt. Based on Bayes' theorem, the posterior probability is expressed as the ratio of the product of the likelihood probability and the prior probability to the marginal probability. P(Y|X,θt) = P(X|Y,θt) P(Y|θt) / P(X) Where P(Y|X,θt) represents the posterior probability of noise intensity Y given input X; P(X|Y,θt) represents the likelihood probability of input X given noise intensity Y; P(Y|θt) represents the prior probability of noise intensity Y; and P(X) represents the marginal probability. In the M-step of the EM algorithm, the model parameters θt+1 are updated by maximizing the likelihood of the unlabeled data, and the likelihood of the unlabeled data is represented as the sum of log-likelihood probabilities: θ_{t+1} = argmax_θ Σ_{X∈Du} Σ_{Y∈{I,II,III,IV,V}} P(Y|X,θt) log P(X,Y|θ) In this process, the outer summation iterates through all samples X in the unlabeled dataset Du, and the inner summation iterates through all possible noise intensity levels Y; P(Y|X,θt) represents the posterior probability obtained in step E, and P(X,Y|θ) is the joint probability distribution of the generative model for the data.

[0033] The optimal parameters for the generative model can be obtained through iterative optimization.

[0034] S5, Establishment of the discriminant model: Given an input magnetotelluric time series X, the goal of the discriminant model is to learn the conditional probability distribution P(Y|X), which is the conditional probability that the noise intensity is Y given the input X.

[0035] Discriminative models can use labeled datasets for supervised learning and adjust decision boundaries by utilizing information from unlabeled data. Semi-supervised learning methods, such as self-training and co-training, can also be used.

[0036] In supervised learning, we can use maximum likelihood estimation to estimate conditional probabilities.

[0037] Suppose we have a labeled dataset D_l={(x_1, y_1), (x_2, y_2), ..., (x_n, y_n)}, where n is the total number of samples in the labeled dataset. We use maximum likelihood estimation to estimate the conditional probabilities: Θ^* = argmax_θ Σ_{i=1}^{n} log P(Y=y_i | X=x_i, θ) Where Θ^* represents the optimal parameters of the discriminant model; θ represents the parameters of the discriminant model; the summation iterates through all samples in the labeled dataset, where i is the sample index from 1 to n; P(Y=y_i|X=x_i, θ) represents the conditional probability that the noise intensity is y_i given the input x_i.

[0038] By optimizing the model parameters by maximizing the likelihood of the labeled data, the optimal parameters for the discriminant model can be obtained. Then, information from the unlabeled data can be used to adjust the model's decision boundary to better fit the data distribution.

[0039] In summary, noise identification in magnetotelluric time series can be achieved through the optimization process of generative and discriminative models. Furthermore, by using semi-supervised learning algorithms and combining information from both labeled and unlabeled data, the model's performance and generalization ability can be improved.

[0040] S6. For magnetotelluric time series with excessively high noise levels (Level I), this data segment is directly discarded. For magnetotelluric time series with relatively high noise levels (Levels II and III), wavelet analysis, cepstral analysis, and other data processing methods are applied, followed by time-frequency transformation and robust regression to estimate their apparent resistivity and phase curves. For magnetotelluric time series with relatively low noise levels (Levels IV and V), time-frequency transformation and robust regression are directly applied to estimate their apparent resistivity and phase curves. This forms a semi-supervised learning-based data processing system that outputs apparent resistivity and phase curves from magnetotelluric time series input.

[0041] Real-world examples: (1) Users upload magnetotelluric data A(t) collected by one or more stations to the platform. The data is required to contain two components of the horizontal electric field, Ex and Ey, and two components of the horizontal magnetic field, Hx and Hy. The vertical magnetic field Hz is not required.

[0042] (2) Other users have the right to use this data and can download, perform online calculations and other processing on the platform.

[0043] (3) Users can select data related to A(t) as the basis for supervising the training set. Conditions such as proximity of the collection area, similar geological structure, and similar industrial environment can all be used as the basis for selecting the training set.

[0044] The training set is preprocessed to obtain features such as spectrum, time-frequency characteristics, and polarization. Based on these features, the noise content of the training set is calibrated, with noise intensity calibrated into five levels. Here, the time series can be a segment of the entire time series collected by each station.

[0045] (4) In the step of establishing the generative model, the joint probability distribution of magnetotelluric time series and noise intensity in the training set is calculated.

[0046] (5) In the step of establishing the discriminant model, unlabeled data from the training set is introduced, and the decision boundary of the discriminant model is adjusted through self-training technology. This part realizes the calculation of the probability distribution of noise intensity based on a time series. Thus, we have completed the semi-supervised learning process for discriminant noise intensity of magnetotelluric time series. By applying the discriminant model to data A(t), the noise intensity of each time series segment in the data can be obtained.

[0047] (6) For time series segments with noise intensity level I, they are directly removed; for time series segments with noise intensity levels II and III, wavelet analysis and other data processing methods are used, followed by time-frequency conversion and robust regression to estimate their apparent resistivity curves and phase curves; for time series segments with noise intensity levels IV and V, time-frequency conversion and robust regression are directly used to estimate their apparent resistivity curves and phase curves.

[0048] Ultimately, we achieved noise intensity discrimination of magnetotelluric time series through semi-supervised learning. Different strategies were adopted for data with different intensities, resulting in apparent resistivity curves and phase curves.

[0049] Furthermore, the present invention also provides a magnetotelluric data processing system, comprising: The data acquisition unit is used to acquire open-source data, which includes labeled and unlabeled datasets. The labeled datasets include magnetotelluric time series under various environments and corresponding calibration results. The environments include different regions, geological structures, and industrial environments. The calibration results include different levels of noise intensity. The noise intensity includes five levels, from strongest to weakest: Level I, Level II, Level III, Level IV, and Level V. The unlabeled datasets are user-shared data uploaded through the open-source platform. The model training unit is used to optimize the model parameters of the generative model and the discriminative model using the open-source data and semi-supervised learning methods, respectively, to obtain a trained model; the trained model is used for noise recognition. The apparent resistivity curve and phase curve determination unit is used to process the noise identification results output by the trained model using a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series. The corresponding algorithm includes elimination, data processing, time-frequency conversion, and robust regression. The data processing includes, but is not limited to, wavelet analysis and cepstral analysis.

[0050] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to cause the electronic device to perform the magnetotelluric data processing method described above.

[0051] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the magnetotelluric data processing method described above.

[0052] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0053] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method of processing magnetotelluric data, characterized by, include: The data obtained includes open-source data, which comprises labeled and unlabeled datasets. The labeled datasets include magnetotelluric time series under various environments and corresponding calibration results. These environments include different regions, geological structures, and industrial environments. The calibration results include different levels of noise intensity, categorized into five levels from strongest to weakest: Level I, Level II, Level III, Level IV, and Level V. The unlabeled datasets are user-shared data uploaded through an open-source platform. Using the aforementioned open-source data and semi-supervised learning methods, the model parameters of the generative model and the discriminative model are optimized respectively to obtain a trained model; the trained model is used for noise recognition, specifically including: Based on the aforementioned open-source data, the EM algorithm is used to optimize the parameters of the learning objective of the generative model, resulting in a well-trained generative model. The learning objective of the generative model is constructed based on the joint probability distribution. The parameter optimization is performed by maximizing the likelihood of the unlabeled dataset. Based on the aforementioned open-source data, the learning objective of the discriminative model is optimized using a self-training algorithm to obtain a well-trained discriminative model; the learning objective of the discriminative model is constructed based on a conditional probability distribution. The noise identification results output by the trained model are processed by a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series, specifically including: When the noise recognition result output by the trained model is level I, the data is directly discarded; When the noise identification result output by the trained model is level II and level III, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated by using data processing methods, time-frequency conversion and robust regression. When the noise identification results output by the trained model are level IV and level V, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated by using the time-frequency conversion and the robust regression. The corresponding algorithm includes elimination, data processing, time-frequency conversion, and robust regression; the data processing includes, but is not limited to, wavelet analysis and cepstral analysis.

2. The magnetotelluric data processing method of claim 1, wherein, Before obtaining the open-source data, the following is also included: An open-source platform is established; this platform is used to share data, enabling users to perform semi-supervised magnetotelluric data processing by freely creating training sets and setting calibration data.

3. The magnetotelluric data processing method of claim 1, wherein, The method for determining the labeled dataset is as follows: Acquire magnetotelluric time series under multiple environments; The magnetotelluric time series is subjected to feature preprocessing to determine each feature parameter; Based on the aforementioned characteristic parameters, the noise intensity level is calibrated using interactive methods to obtain the calibration results corresponding to the magnetotelluric time series.

4. The magnetotelluric data processing method according to claim 1, characterized in that, The step of optimizing the learning objective parameters of the generative model using the EM algorithm based on the open-source data to obtain a well-trained generative model specifically includes: In the E-step of the EM algorithm, the posterior probability P(Y|X,θt) of the unlabeled dataset is estimated using the current parameter θt. Based on Bayes' theorem, the posterior probability is expressed as the ratio of the product of the likelihood probability and the prior probability to the marginal probability. P(Y|X,θt) = P(X|Y,θt) P(Y|θt) / P(X) Where P(Y|X,θt) represents the posterior probability of noise intensity Y given input X; P(X|Y,θt) represents the likelihood probability of input X given noise intensity Y; P(Y|θt) represents the prior probability of noise intensity Y; and P(X) represents the marginal probability. In the M-step of the EM algorithm, the model parameters θt+1 are updated by maximizing the likelihood of the unlabeled data, and the likelihood of the unlabeled data is represented as the sum of log-likelihood probabilities: θ_{t+1} = argmax_θ Σ_{X∈Du} Σ_{Y∈{I,II,III,IV,V}} P(Y|X,θt) log P(X,Y|θ) In this process, the outer summation iterates through all samples X in the unlabeled dataset Du, and the inner summation iterates through all possible noise intensity levels Y; P(Y|X,θt) represents the posterior probability obtained in step E, and P(X,Y|θ) is the joint probability distribution of the generative model for the data.

5. The magnetotelluric data processing method of claim 1, wherein, Based on the open-source data, the self-training algorithm is used to optimize the parameters of the learning objective of the discriminative model to obtain a well-trained discriminative model, specifically including: Suppose we have a labeled dataset D_l={(x_1, y_1), (x_2, y_2), ..., (x_n, y_n)}, where n is the total number of samples in the labeled dataset. We use maximum likelihood estimation to estimate the conditional probabilities: Θ^* = argmax_θ Σ_{i=1}^{n} log P(Y=y_i | X=x_i, θ) Where Θ^* represents the optimal parameters of the discriminant model; θ represents the parameters of the discriminant model; the summation iterates through all samples in the labeled dataset, where i is the sample index from 1 to n; P(Y=y_i|X=x_i, θ) represents the conditional probability that the noise intensity is y_i given the input x_i; The optimal parameters of the discriminant model are obtained by optimizing the model parameters by maximizing the likelihood of the labeled data; then, the decision boundary of the model is adjusted using information from the unlabeled dataset to obtain the trained discriminant model.

6. A magnetotelluric data processing system characterized by, include: The data acquisition unit is used to acquire open-source data, which includes labeled and unlabeled datasets. The labeled datasets include magnetotelluric time series under various environments and corresponding calibration results. The environments include different regions, geological structures, and industrial environments. The calibration results include different levels of noise intensity. The noise intensity includes five levels, from strongest to weakest: Level I, Level II, Level III, Level IV, and Level V. The unlabeled datasets are user-shared data uploaded through the open-source platform. The model training unit is used to optimize the model parameters of the generative model and the discriminative model using the open-source data and semi-supervised learning methods, respectively, to obtain a trained model. The trained model is used for noise recognition, specifically including: Based on the aforementioned open-source data, the EM algorithm is used to optimize the parameters of the learning objective of the generative model, resulting in a well-trained generative model. The learning objective of the generative model is constructed based on the joint probability distribution. The parameter optimization is performed by maximizing the likelihood of the unlabeled dataset. Based on the aforementioned open-source data, the learning objective of the discriminative model is optimized using a self-training algorithm to obtain a well-trained discriminative model; the learning objective of the discriminative model is constructed based on a conditional probability distribution. The apparent resistivity curve and phase curve determination unit is used to process the noise identification results output by the trained model using a corresponding algorithm to obtain the apparent resistivity curve and phase curve of the magnetotelluric time series, specifically including: When the noise recognition result output by the trained model is level I, the data is directly discarded; When the noise identification result output by the trained model is level II and level III, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated by using data processing methods, time-frequency conversion and robust regression. When the noise identification results output by the trained model are level IV and level V, the apparent resistivity curve and phase curve of the magnetotelluric time series are estimated by using the time-frequency conversion and the robust regression. The corresponding algorithm includes elimination, data processing, time-frequency conversion, and robust regression; the data processing includes, but is not limited to, wavelet analysis and cepstral analysis.

7. An electronic device, comprising: The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the magnetotelluric data processing method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the magnetotelluric data processing method as described in any one of claims 1-5.

Images