Audio magnetotelluric method exploration data analysis method

Abstract

The invention discloses an audio magnetotelluric method exploration data analysis method, and relates to the technical field of physical exploration, and the method comprises the steps: firstly, employing a deep neural network embedded with an electromagnetic field physical rule constraint, carrying out the intelligent deconstruction of original time series data, and guaranteeing the physical authenticity of purified data from the source; secondly, combining data-driven inversion modeling with man-machine collaborative active learning, and guiding a geologic model to iteratively converge towards the uncertainty reduction direction through intelligent questioning and quantitative feedback of expert knowledge; and finally, mathematizing an exploration decision process into a multi-objective optimization problem with risk consideration, and calculating and presenting a series of action scheme combinations which achieve optimal balance among information gain, cost and risk for the user based on a high-credibility model. According to the method, an intelligent analysis framework capable of managing uncertainty in the whole process and realizing decision closed loop is constructed, and the scientificity and the success rate of exploration decision are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of physical exploration technology, and in particular to an audio frequency magnetotelluric method exploration data analysis method. Background Art

[0002] Magnetotellurics (MT) is an important geophysical method used in resource exploration and engineering geological surveys. However, the current exploration data analysis process is essentially an open, linear information processing chain, which suffers from a fundamental technical flaw: the lack of a comprehensive, systematic mechanism for transmitting and controlling information uncertainty throughout the entire process.

[0003] This flaw is reflected in the entire analysis process. At the beginning of the process, the original data is subject to strong noise interference. Traditional denoising methods, while suppressing noise, often introduce new uncertainties or damage effective signals, leading to doubts about the physical authenticity of the output data. After these uncertain data are input into the inversion modeling link, the uncertainty of the model is further amplified due to the inherent multi-solution nature of geophysical inversion. Although constraints can be imposed by introducing expert prior knowledge, traditional manual and static constraints are inefficient, and the subjective judgment of experts themselves becomes a new source of uncertainty.

[0004] Therefore, when the analysis process reaches its final stage—the exploration decision phase—decision-makers are faced with a geological model that has accumulated all the uncertainties from the previous phase and is therefore questionable in its reliability. Based on such a model, any decision about the next high-cost exploration activity (such as drilling) must rely solely on qualitative, personal judgment. Existing technology lacks a scientific method to quantify the risks inherent in different decision options, nor does it systematically identify an action plan that maximizes information benefits and achieves the optimal risk-benefit balance within a limited budget.

[0005] In summary, this process, in which uncertainty accumulates step by step throughout and cannot be effectively managed, is the core technical bottleneck that restricts the current success rate and economic benefits of magnetotelluric exploration. Summary of the Invention

[0006] The present invention provides an audio magnetotelluric exploration data analysis method to solve the technical problem in the existing magnetotelluric data analysis process, that is, the lack of full-process transmission and control of information uncertainty, which leads to the reliability of geological cognition decreasing with the deepening of analysis, and ultimately makes it impossible to scientifically optimize the decision-making of exploration actions based on quantified risks and benefits.

[0007] In view of the above problems, the present invention provides an audio frequency magnetotelluric exploration data analysis method, comprising: Receive multi-channel magnetotelluric time series data collected in the exploration area; Deconstructing the time series data into a pure signal component and a noise component using a neural network based on physical information constraints to generate a cleansed data set, wherein the physical information constraints stipulate that the pure signal component must follow the physical laws of the magnetotelluric method; generating an initial underground three-dimensional electrical structure model based on the cleansed data set, and quantifying the uncertainty of the initial model to identify key areas where the model uncertainty is higher than a preset threshold; Generate and present structured geological questions to the user based on the key areas, receive user feedback and quantify it into mathematical constraints, and integrate the mathematical constraints into the model through iterative inversion to generate an optimized geological model that incorporates user knowledge; The target body circled by the user on the optimized geological model is received, and based on the two optimization objectives of information gain and exploration cost, a multi-objective optimization algorithm is used to calculate and generate a set of Pareto optimal action plan combinations for the next exploration activity.

[0008] The technical solution provided by this application has at least the following technical effects or advantages: The present invention achieves effective control of uncertainty in the entire exploration process and builds a closed loop from data analysis to decision feedback, solving the core problem of existing technologies where uncertainty accumulates and gets out of control due to process disruptions.

[0009] This method uses physical laws as constraints to train a neural network during data cleansing. This trained network is then used to intelligently deconstruct the original time series data. This ensures that the cleansed data generated retains the maximum physical authenticity of the geophysical field response.

[0010] This invention transforms the geological modeling process into an iterative, human-machine collaborative active learning cycle. It automatically quantifies and locates the uncertainty of the current model, intelligently questions experts accordingly, and efficiently and accurately quantifies the experts' qualitative knowledge into mathematical constraints. This cycle guides the inversion process to rapidly converge toward the dual constraints of data and geological laws, significantly improving the reliability of the resulting geological model and resolving the technical challenge of efficiently and dynamically integrating expert knowledge with data-driven models.

[0011] This invention transforms exploration decision-making, a complex, empirically based process, into a calculable, optimizable, and quantitative decision-making process. Based on a highly reliable geological model and its uncertainty distribution, it clearly calculates and presents the trade-offs between information gain, cost, and risk for different exploration action plans to decision makers, providing a series of Pareto-optimal action combinations. This transforms exploration decisions from qualitative, ambiguous empirical judgments to quantitative, clear optimization choices. By using the results of the optimal action plan as feedback for subsequent modeling iterations, a self-optimizing decision-making closed loop is established. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention provides a flow chart of an audio frequency magnetotelluric exploration data analysis method. DETAILED DESCRIPTION

[0013] The present invention relates to an audio magnetotelluric exploration data analysis method to solve the technical problem in the existing magnetotelluric data analysis process that the reliability of geological cognition decreases with the deepening of analysis due to the lack of full-process transmission and control of information uncertainty, and ultimately makes it impossible to scientifically optimize the decision-making of exploration actions based on quantified risks and benefits.

[0014] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the drawings.

[0015] like Figure 1 A method for analyzing audio frequency magnetotelluric exploration data is shown in a flow chart. The method comprises the following steps: Receive multi-channel magnetotelluric time series data collected in the exploration area; Deconstructing the time series data into a pure signal component and a noise component using a neural network based on physical information constraints to generate a cleansed data set, wherein the physical information constraints stipulate that the pure signal component must follow the physical laws of the magnetotelluric method; generating an initial underground three-dimensional electrical structure model based on the cleansed data set, and quantifying the uncertainty of the initial model to identify key areas where the model uncertainty is higher than a preset threshold; Generate and present structured geological questions to the user based on the key areas, receive user feedback and quantify it into mathematical constraints, and integrate the mathematical constraints into the model through iterative inversion to generate an optimized geological model that incorporates user knowledge; The target body circled by the user on the optimized geological model is received, and based on the two optimization objectives of information gain and exploration cost, a multi-objective optimization algorithm is used to calculate and generate a set of Pareto optimal action plan combinations for the next exploration activity.

[0016] Specifically, the overall technical process of the method described in this invention is divided into three stages. The first stage is data deconstruction and purification. This stage receives raw, noisy, multi-channel magnetotelluric time series data and uses a deep neural network model to intelligently decompose the data into physically realistic pure signal and noise components based on the fundamental laws of physics, thereby outputting a purified dataset.

[0017] Specifically, the second stage is human-machine collaborative iterative modeling. In this stage, the purified data set is used to perform an initial inversion to obtain a preliminary underground three-dimensional electrical structure model. Subsequently, the model is evaluated for uncertainty, and the weakest and most critical areas in the model that need improvement are automatically identified. Based on this, structured geological questions are generated, and knowledge input is actively requested from the user. The user's qualitative feedback is automatically quantified into precise mathematical constraints and applied to the next round of inversion. This iterative cycle of "inversion-questioning-feedback-re-inversion" continues until the model converges, and finally an optimized geological model that deeply integrates user knowledge is generated.

[0018] Specifically, the third stage involves quantitative decision-making and proactive path planning. Once the geological target volume is determined based on the optimized geological model, the exploration decision-making problem is mathematically formulated as a multi-objective optimization problem under the conflicting objectives of maximizing information gain and minimizing exploration costs. Using a multi-objective optimization algorithm, a series of exploration action plans that achieve the optimal balance between cost and benefit are solved and generated. Ultimately, the optimal combination of action plans is presented to the decision-maker in the form of a Pareto optimal front, transforming exploration decision-making from traditional qualitative empirical judgment to scientific quantitative optimization.

[0019] Furthermore, the training process of the neural network based on physical information constraints adopts a total loss function, which includes: A reconstruction loss term, used to constrain the consistency of the sum of the pure signal component and the noise component with the original time series data; A physical law residual loss term, used to penalize the deviation of the pure signal component from the preset physical law; The noise component constraint item is used to impose a priori characteristic constraints on the noise component.

[0020] Specifically, the three components of the total loss function work together to guide the neural network based on physical information constraints to perform unsupervised learning.

[0021] The reconstruction loss term ensures information conservation. Its mathematical form is usually the mean square error, which is used to measure the difference between the sum of the pure signal component and the noise component of the network output and the original input time series data. Minimizing this term is the basis of network learning.

[0022] The physical law residual loss term introduces the first principles of physics as a strong constraint into the neural network. This loss term quantifies the degree to which the pure signal components output by the network conform to the basic physical laws of magnetotellurics.

[0023] Noise component constraints are regularization terms imposed based on prior knowledge of noise statistics or physical properties. For example, based on the sparse nature of industrial harmonic noise in the frequency domain, the L1 norm can be imposed on the frequency domain representation of the noise component to encourage the network to identify sparse components as noise.

[0024] Furthermore, the calculation of the physical law residual loss term adopts a frequency adaptive mechanism, and for signal components in different frequency ranges, physical equations matching their physical essence are used to perform residual calculation; and within the frequency range where the quasi-static field approximation is applicable, the physical law residual loss term is jointly defined by fast one-dimensional impedance estimation and local Kramers-Kronig relationship constraints.

[0025] Specifically, in order to improve the accuracy of physical constraints at different frequencies, the physical law residual loss term of the present invention adopts a frequency adaptive mechanism. A switching frequency is preset, for example, 10 kHz. When processing data with a frequency higher than the switching frequency, the propagation effect of electromagnetic waves cannot be ignored. Therefore, the physical law residual loss term is mainly calculated based on the residual of the full-wave Maxwell equation to accurately describe the wave propagation effect. When processing data with a frequency lower than the switching frequency, a quasi-static field approximation can be used. At this time, the physical law residual loss term is switched to be calculated mainly based on the residual of the impedance relationship between the electric field and the magnetic field. This frequency-varying physical constraint mechanism ensures that the physical constraints imposed are closest to their physical essence throughout the exploration frequency band.

[0026] Furthermore, within the frequency range applicable to the quasi-static field approximation, and to avoid the logical loop caused by reliance on a non-existent accurate three-dimensional model, the present invention designs a fast one-dimensional impedance estimation. This estimation is a lightweight, differentiable calculation. At each time-frequency point in the time-spectrum graph, only the electric and magnetic field components at that point are used to rapidly calculate a local one-dimensional Cagniard impedance through complex division or local least-squares fitting. This process does not pursue absolute inversion accuracy, but the results are sufficient to provide a real-time, physically consistent constraint direction to guide network training.

[0027] Furthermore, to enhance the rigor of physical constraints, the present invention also introduces a local Kramers-Kronig relation constraint. Since a strict Kramers-Kronig relation requires data of infinite bandwidth, the present invention adopts a localized approach, performing a Hilbert transform on the imaginary part of the impedance within the observation frequency band and comparing it with the real part of the impedance to check the causal consistency of its response. This penalty term is intended to ensure that the impedance response curve corresponding to the pure signal generated by the network is smooth and conforms to the physical causal law, avoiding physically unreasonable violent oscillations, thereby improving the physical authenticity of the purified signal.

[0028] The step of quantifying the uncertainty of the initial model further generates an uncertainty distribution map of the model by calculating the posterior covariance matrix or the posterior entropy of the model.

[0029] Specifically, after obtaining the initial underground three-dimensional electrical structure model, the reliability of the model is evaluated unit by unit based on the Bayesian inversion theory. The diagonal elements of the posterior covariance matrix of the model directly quantify the variance of the resistivity value of each model grid unit. The larger the variance, the higher the uncertainty. Alternatively, the posterior entropy of each model parameter can be calculated as a measure of uncertainty. Based on this calculation result, a three-dimensional uncertainty distribution map is generated, and spatially connected areas whose uncertainty values ​​exceed a preset threshold are automatically identified and highlighted. These areas are defined as the key areas that most need external knowledge intervention to enhance model certainty.

[0030] Furthermore, the step of quantifying user feedback into mathematical constraints includes: Access a pre-defined rock physics knowledge base to automatically convert user semantic selections of geological attributes into electrical parameter boundary constraints applied to specific regions of the model; Confidence weights are introduced for the mathematical constraints, and a conflict resolution mechanism is established. The conflict resolution mechanism issues a prompt to the user when the introduction of the mathematical constraints causes the increase in the fitting difference between the model and the observed data to exceed a preset threshold.

[0031] Specifically, to automatically convert user qualitative judgments into computer-recognizable quantitative constraints, a configurable and extensible rock physics knowledge base is pre-defined. This knowledge base establishes a mapping between geological semantic labels and physical parameter ranges and can be stored in a structured data format (such as a JSON file). For example, a knowledge base entry might include: { "geological_label":"Water-rich sandstone", "resistivity_range_ohm_m":[5,50], "source_reference":"Industry standard chart" }; When the user selects the "water-rich sandstone" tag, the knowledge base is automatically queried and the corresponding [5,50] ohm·m range is extracted and applied as a boundary constraint to the relevant area of ​​the inversion model.

[0032] Furthermore, the present invention introduces a robust human-computer interaction process that includes confidence weights and conflict resolution mechanisms. It allows users to attach a qualitative confidence level (e.g., high, medium, low) when providing semantic feedback. An internal configurable confidence weight mapping table is maintained to convert these qualitative levels into quantitative confidence levels. In addition, a conflict resolution mechanism monitors in real time the change in the model's fit error to the observed data after the introduction of new user constraints. If the increase in the fit error exceeds a preset threshold, the constraint is deemed to be highly incompatible with the data, and a warning is issued to the user, suggesting that they review their feedback or lower the corresponding confidence weight. This mechanism ensures that the introduction of user knowledge does not significantly violate objective data.

[0033] Furthermore, in the iterative inversion, the mathematical constraints are integrated into the new objective function.

[0034] New objective function is defined as: in, is the data fitting term, is a conventional model smoothing regularization term, and α is its weight. The present invention adds a third term, namely the user knowledge constraint term. The mathematical expression representing the i-th user constraint (e.g., a function that penalizes resistivity values ​​that are outside the bounds), is the weight associated with the constraint, mapped from the confidence level, and β is a global hyperparameter used to balance data-driven and user-knowledge-driven approaches. By minimizing this new objective function, the inversion process will find the optimal balance between fitting the observed data, maintaining model smoothness, and satisfying user constraints.

[0035] Furthermore, the information gain target is defined by the expected reduction in overall uncertainty of the optimized geological model according to the action plan; and the exploration cost target is defined by the sum of the preset costs of each activity in the action plan.

[0036] Furthermore, before adopting the multi-objective optimization algorithm, the method also includes: automatically identifying and generating a candidate action pool consisting of multiple high-potential exploration actions based on the optimized geological model and its uncertainty distribution; and searching the discrete combination space defined by the multi-objective optimization algorithm in the candidate action pool.

[0037] Specifically, in order to solve the optimization problem caused by the infinite possible combinations of exploration actions, the present invention sets a step for generating the candidate action pool before performing multi-objective optimization. This step is an intelligent pre-screening process that automatically analyzes the optimized geological model and its uncertainty distribution to identify the most valuable potential actions. For example, "adding new magnetotelluric measurement points" is recommended in the center of the area where the model uncertainty is the largest, or "arranging verification drilling holes" is recommended on the boundary of the identified important electrical anomaly. This process converts a continuous, infinite search space into a discrete, finite but high-quality set of candidate actions. The search task of the multi-objective optimization algorithm is to find the optimal subset among all possible combinations of this candidate action pool, thereby greatly reducing the complexity of the optimization problem.

[0038] Furthermore, after generating the set of Pareto optimal action plan combinations, the method also includes: responding to a user's instruction to select a specific plan from the Pareto optimal action plan combination; and calculating and displaying the conditional risk value of the economic value associated with the target object through Monte Carlo simulation for the plan selected by the user.

[0039] Specifically, after the multi-objective optimization is completed, a two-dimensional Pareto front of cost and information gain is presented to the decision maker. On this front, the decision maker can interactively select any action plan that meets their budget and information needs. Once a plan is selected, a posterior quantitative risk assessment is initiated. Using the uncertainty distribution of the optimized geological model, thousands of Monte Carlo simulations are run, each generating a probabilistically plausible geological realization and calculating its corresponding economic value.

[0040] Furthermore, based on the value distribution derived from the Monte Carlo simulation, the conditional VaR is calculated. This metric quantifies the expected average loss of the portfolio at a preset confidence level (e.g., a worst-case scenario of 5%). By showing decision makers the change in the conditional VaR before and after implementing the selected option, they can intuitively understand the option's ability to mitigate extreme risks, providing a scientific and quantitative basis for the final business decision.

[0041] The neural network based on physical information constraints adopts an encoder-decoder architecture and integrates a self-attention mechanism in its feature extraction path to capture the long-distance dependencies of the time series data in the time domain and frequency domain.

[0042] Specifically, the architecture of the neural network based on physical information constraints has been specifically designed. Its main body adopts an encoder-decoder structure, which can effectively extract and reconstruct data features at different scales through a layer-by-layer downsampling encoding path and a layer-by-layer upsampling decoding path, combined with jump connections. On this basis, the present invention integrates a self-attention mechanism. By calculating the mutual correlation weights between any two points in the input feature map, global and long-distance dependencies can be effectively captured. This feature is extremely effective for processing structured noise in magnetotelluric data. For example, it can capture the long-term correlation of industrial harmonics at multiple discrete frequency points, or capture the synchronization of lightning noise covering a wide frequency band at the same time, thereby achieving accurate identification and separation of such complex noises.

[0043] The reconstruction loss term and the noise component constraint term are calculated in each training batch, while the physical law residual loss term adopts a periodic update strategy and is calculated once every several preset training batches or each training cycle.

[0044] Specifically, to address the engineering issue of potentially excessive computational complexity associated with the physical law residual loss term and to improve overall training efficiency, the present invention employs a periodic or delayed update optimization strategy. During neural network training, the relatively simple reconstruction loss term and noise component constraint term can be calculated and backpropagated within each training batch. However, since the physical law residual loss term may involve more complex calculations (such as impedance estimation and Hilbert transform), its update utilizes a more flexible periodic strategy. For example, it can be calculated once at the end of each training cycle (epoch) or once every predetermined number of training batches (e.g., every 10 batches). This step-by-step delayed update strategy significantly reduces the average computational load of the training process while effectively ensuring the constraints imposed by the physical law.

[0045] In summary, the audio magnetotelluric exploration data analysis method disclosed in the present invention creatively integrates intelligent signal deconstruction based on physical information constraints, iterative cognitive modeling of human-computer collaboration, and quantitative decision support based on multi-objective optimization into a logically closed-loop, self-optimizing systematic process. The present invention not only achieves significant innovation in various technical links, but more importantly, it fundamentally solves the core problem of the prior art of the gradual accumulation and loss of control of information uncertainty caused by process breaks. By constructing a new technical framework that can effectively transmit, manage, and control uncertainty throughout the entire process, the present invention elevates the traditional, experience-based exploration and analysis model to a new level of scientificity, quantification, and risk control.

[0046] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing audio frequency magnetotelluric exploration data, characterized in that: The method comprises the following steps: Receive multi-channel magnetotelluric time series data collected in the exploration area; Deconstructing the time series data into a pure signal component and a noise component using a neural network based on physical information constraints to generate a cleansed data set, wherein the physical information constraints stipulate that the pure signal component must follow the physical laws of the magnetotelluric method; generating an initial underground three-dimensional electrical structure model based on the cleansed data set, and quantifying the uncertainty of the initial model to identify key areas where the model uncertainty is higher than a preset threshold; Generate and present structured geological questions to the user based on the key areas, receive user feedback and quantify it into mathematical constraints, and integrate the mathematical constraints into the model through iterative inversion to generate an optimized geological model that incorporates user knowledge; The target body circled by the user on the optimized geological model is received, and based on the two optimization objectives of information gain and exploration cost, a multi-objective optimization algorithm is used to calculate and generate a set of Pareto optimal action plan combinations for the next exploration activity.

2. The method for analyzing audio magnetotelluric exploration data according to claim 1, wherein: The training process of the neural network based on physical information constraints adopts a total loss function, which includes: A reconstruction loss term, used to constrain the consistency of the sum of the pure signal component and the noise component with the original time series data; A physical law residual loss term, used to penalize the deviation of the pure signal component from the preset physical law; The noise component constraint item is used to impose a priori characteristic constraints on the noise component.

3. The method for analyzing audio magnetotelluric exploration data according to claim 2, wherein: The calculation of the physical law residual loss term adopts a frequency adaptive mechanism, and for signal components in different frequency ranges, physical equations matching their physical nature are used to perform residual calculations; and within the frequency range where the quasi-static field approximation is applicable, the physical law residual loss term is jointly defined by fast one-dimensional impedance estimation and local Kramers-Kronig relationship constraints.

4. The method for analyzing audio magnetotelluric exploration data according to claim 1, wherein: The step of quantifying the uncertainty of the initial model generates an uncertainty distribution map of the model by calculating the posterior covariance matrix or the posterior entropy of the model.

5. The method for analyzing audio magnetotelluric exploration data according to claim 1, wherein: The step of quantifying user feedback into mathematical constraints comprises: Access a pre-defined rock physics knowledge base to automatically convert user semantic selections of geological attributes into electrical parameter boundary constraints applied to specific regions of the model; Confidence weights are introduced for the mathematical constraints, and a conflict resolution mechanism is established. The conflict resolution mechanism issues a prompt to the user when the introduction of the mathematical constraints causes the increase in the fitting difference between the model and the observed data to exceed a preset threshold.

6. The method for analyzing audio magnetotelluric exploration data according to claim 1, wherein: The information gain target is defined by the expected reduction in overall uncertainty of the optimized geological model based on the action plan; the exploration cost target is defined by the sum of the preset costs of each activity in the action plan.

7. The method for analyzing audio magnetotelluric exploration data according to claim 1, wherein: Before adopting the multi-objective optimization algorithm, the method further includes: automatically identifying and generating a candidate action pool consisting of multiple high-potential exploration actions based on the optimized geological model and its uncertainty distribution; and searching the discrete combination space defined by the multi-objective optimization algorithm in the candidate action pool.

8. The method for analyzing audio magnetotelluric exploration data according to claim 1, wherein: After generating the set of Pareto optimal action plan combinations, the method further includes: responding to a user's instruction to select a specific plan from the Pareto optimal action plan combination; and calculating and displaying the conditional risk value of the economic value associated with the target object through Monte Carlo simulation for the plan selected by the user.

9. The method for analyzing audio frequency magnetotelluric exploration data according to claim 1, wherein: The neural network based on physical information constraints adopts an encoder-decoder architecture and integrates a self-attention mechanism in its feature extraction path to capture the long-distance dependencies of the time series data in the time domain and frequency domain.

10. The audio frequency magnetotelluric exploration data analysis method according to claim 2, characterized in that: The reconstruction loss term and the noise component constraint term are calculated in each training batch, while the physical law residual loss term adopts a periodic update strategy and is calculated once every several preset training batches or each training cycle.

Images