Three-dimensional resistivity and polarizability data joint inversion method and system

Abstract

The invention belongs to the technical field of electromagnetic exploration, and discloses a three-dimensional resistivity and polarizability data joint inversion method and system. Designing and generating a real geologic model capable of reflecting underground structure characteristics; performing forward modeling calculation on the real geologic model based on a finite element method to obtain a resistivity observation value, a polarizability observation value and resistivity-polarizability coupling data; performing denoising, normalization and missing value completion processing on the forward modeling data, randomly dividing the forward modeling data into a training sample and a test sample, and adding noise disturbance into the training sample to enhance robustness; constructing and training an adjusted U-Net deep learning model, and performing feature extraction and reconstruction through an encoder and a decoder in combination with jump connection; inputting the preprocessed resistivity, polarizability and coupling data thereof, and training a deep learning model to predict the three-dimensional electrical characteristics of the underground medium; and based on the Dice coefficient optimization model, outputting three-dimensional resistivity and polarizability distribution, and evaluating inversion precision and stability.

Description

Technical Field

[0001] The invention belongs to the technical field of electromagnetic exploration, and in particular relates to a three-dimensional resistivity and polarizability data joint inversion method and system. Background Art

[0002] Once the soil is contaminated, the pollutants are difficult to remove, so prevention of heavy metal pollution is more important than remediation. At present, although traditional soil pollution monitoring methods (such as chemical tracer method, sampling analysis method, borehole monitoring method, etc.) are relatively mature in theory, they still have great limitations in practical application. Chemical tracer method can only qualitatively point out the source of pollution and may inadvertently cause secondary pollution. Sampling and analysis method cannot reveal the specific leakage path of the pollution source. On the other hand, borehole sampling may destroy the pollution distribution structure and is not suitable for long-term comprehensive monitoring. Therefore, traditional methods are insufficient in achieving comprehensive, rapid and low-interference monitoring and investigation of soil heavy metal pollution. In contrast, geophysical monitoring technology, which uses the difference in physical properties between contaminated wastewater and surrounding media, provides comprehensive spatial coverage, on-site non-destructive detection, low cost and rapidity. They are widely used to infer the spatial and physical distribution of underground contaminated wastewater leakage. Methods including electromagnetic and ground penetrating radar (GPR) have been applied to detect specific pollution sources. However, traditional geophysical methods encounter problems such as weak signals and susceptibility to environmental interference when dealing with weak magnetic anomalies and electromagnetic effects caused by heavy metal contaminated wastewater. GPR also struggles to provide unambiguous property inferences. So despite the potential advantages of geophysical methods, there are still clear disadvantages in terms of accuracy and application effectiveness.

[0003] In response to these challenges, the contact induced polarization method (MALM-IP) is an innovative monitoring technology with relatively significant advantages. The design of this contact power supply is based on the fact that the leakage pollution source is a man-made body, such as mining wastewater pools and tailings ponds, industrial wastewater treatment boilers, oil and gas reservoir leakage areas, and leakage halos above oil and gas reservoirs. Since the leakage source is visible, it is possible to directly connect one end of the power supply electrode to the leakage source. Since the contact power supply is that one end of the electrode is powered inside the leakage body and the other end is placed at infinity, the environmental requirements for the site are relatively low and easy to implement; the internal direct power supply, the current is mainly concentrated in the leakage body range with fluidity and continuity, which can form the highest intensity of primary field excitation. The primary electric field of the contact power supply is similar to the point current source field, and there is no problem of fit for the distribution of the leakage body. In theory, the distribution range of the leakage body is consistent with the power supply current concentration range, and the best primary field excitation can be performed on any form of leakage body. The contact power supply equipment can directly use the existing traditional induced polarization method high-power induced polarization power supply system, so the implementation process is feasible. The method has low requirements on environmental standards, is easy to implement, and effectively suppresses interference signals, enhancing the credibility of the observation results. However, the application of MALM-IP also faces challenges, especially in data processing during inversion imaging. Early inversion methods mainly focused on the separate inversion of resistivity or polarizability, such as one-dimensional and two-dimensional least squares methods, but these methods have problems of low computational efficiency and multiple solutions. With the development of computer technology, three-dimensional DC resistivity inversion has received increasing attention, and methods such as linear and nonlinear conjugate gradient methods have been used to improve inversion efficiency and stability. However, the above-mentioned gradient-based inversion algorithms rely on the imposed constraints, such as minimum structural constraints, inequality constraints, and prior structural constraints. In recent years, considering the significant dependence of traditional polarizability inversion on resistivity inversion results, resistivity and polarizability joint or simultaneous inversion algorithms have been continuously developed, such as simultaneous inversion of induced polarization data based on cross-gradient constraints.

[0004] The exploration method based on charge field source does not involve the change of electrode spacing, and the observation data collected on the ground or in the well are far less than the unknowns required for calculation. Therefore, the obtained inversion Jacobian matrix and its inverse matrix have a large condition number, and the ill-conditioned nature of the inversion process needs to be considered. Therefore, it is necessary to consider the joint inversion of resistivity and polarizability to avoid solving the ill-conditioned Jacobian matrix and solve the problem of charge equivalent field source. The joint inversion of three-dimensional resistivity and polarizability based on charge field source will be a highly potential imaging technology.

[0005] With the rapid development of artificial intelligence (AI) and deep learning (DL) technologies, traditional geophysical inversion imaging methods have ushered in new opportunities. AI technology can improve the accuracy and resolution of geophysical inversion, especially when dealing with complex underground data, showing great application potential. Although AI has made remarkable achievements in various fields, traditional AI algorithms are often limited in dealing with the nonlinear characteristics of geophysical problems because they usually only deal with a single data feature. This approach often fails to fully explore the intrinsic relationship between different data sets, limiting the improvement of inversion accuracy.

[0006] In the geophysical inversion process, key parameters such as resistivity and polarizability show significant spatial correlation and jointly reflect the physical characteristics of the underground medium. Therefore, inversion based on a single feature often ignores the coupling relationship between resistivity and polarizability, making it difficult to fully capture the true characteristics of the underground structure. Therefore, effectively combining two different data types (such as resistivity and polarizability data) for joint inversion of the same underground model has become the key to improving the accuracy and reliability of geophysical inversion. Traditional methods have not yet effectively solved this problem, limiting the application of resistivity or polarizability information in geophysical inversion.

[0007] In view of the above analysis, the existing technology has the following technical problems that need to be solved urgently:

[0008] (1) Limitations of traditional pollution monitoring methods:

[0009] Chemical tracer method: can only point out the source of pollution qualitatively and may cause secondary pollution.

[0010] Sampling and analysis method: cannot reveal the specific leakage path of the pollution source.

[0011] Drilling monitoring method: may damage soil structure and is not suitable for long-term and comprehensive monitoring.

[0012] Urgently needed: a more comprehensive, rapid, low-intrusive, and long-term monitoring method.

[0013] (2) Limitations of traditional geophysical methods:

[0014] Electromagnetic and ground penetrating radar technology: When dealing with weak magnetic anomalies and electromagnetic effects caused by heavy metal contaminated wastewater, the signal is weak and easily affected by environmental interference.

[0015] Ground penetrating radar: It is difficult to provide accurate inference of the properties of pollution sources.

[0016] Urgently needed: A monitoring technology with higher signal strength, lower environmental interference, and the ability to accurately infer the properties of pollution sources.

[0017] (3) Issues regarding the efficiency and accuracy of data processing during inversion imaging:

[0018] Early inversion methods focused on inverting resistivity or polarizability alone, but they suffered from low computational efficiency and multiple solutions.

[0019] Gradient-based inversion methods rely on imposed constraints but are sometimes unstable and have difficulty in imaging complex subsurface structures.

[0020] Urgent need: A more efficient and stable three-dimensional resistivity and polarizability joint inversion method that can handle ill-conditioned Jacobian matrix problems.

[0021] (4) Problems of joint inversion of resistivity and polarizability data:

[0022] Traditional inversion methods fail to effectively combine resistivity and polarizability data and ignore the coupling relationship between the two, resulting in insufficient accuracy of the inversion results.

[0023] Urgently needed: A joint inversion algorithm that can simultaneously invert resistivity and polarizability data to provide more comprehensive and accurate underground pollution detection results.

[0024] (5) Insufficient application of traditional artificial intelligence methods in geophysical inversion:

[0025] Traditional AI algorithms usually only process single data features and find it difficult to effectively capture the spatial correlation between key parameters such as resistivity and polarizability.

[0026] Urgently needed: An intelligent imaging method that can integrate multiple data features and capture the coupling relationship between data to improve inversion accuracy and reliability.

[0027] (6) Data processing and imaging accuracy issues:

[0028] In the process of geophysical inversion, we often face challenges such as large data volume, strong nonlinearity, and complex underground structure. Traditional algorithms show low accuracy when dealing with such problems.

[0029] Urgently needed: A deep learning algorithm that can effectively process complex subsurface data and combine multiple data types (such as resistivity and polarizability) to improve imaging accuracy and resolution. Summary of the invention

[0030] In view of the problems existing in the prior art, the present invention provides an innovative intelligent imaging algorithm, which adopts the U-Net neural network framework to jointly process the resistivity and polarizability data obtained by the MALM-IP method for contaminated site detection. The U-Net network has shown excellent performance in tasks such as medical image segmentation and semantic segmentation. Its unique encoder-decoder structure can efficiently extract multi-scale feature information while retaining the spatial structure. Taking advantage of the U-Net network, the present invention designs a deep learning method for resistivity and polarizability data based on the MALM-IP method. Through this deep learning model, the method effectively integrates the two data types, which not only makes up for the limitations of a single data source, but also can achieve more accurate contaminated site detection results in a wider underground environment.

[0031] The present invention is implemented as follows: a novel resistivity and polarizability joint inversion imaging method, comprising:

[0032] Step 1: Construct the data set required for deep learning. First, you need to design and generate a real geological model that can truly reflect the underground structural characteristics of the target area. The core content of the data set includes: forward calculation results based on these real geological models, specifically the observations of resistivity and polarizability. These observations should cover a variety of scenarios under different geological conditions to ensure the breadth and representativeness of the data. In addition, observation data at different depths and different spatial resolutions should also be included to enhance the generalization ability and accuracy of the model, and ultimately provide sufficient and high-quality input data for the training of deep learning algorithms.

[0033] Step 2, the preliminary processing of input data, the construction of network model, the output of loss curve and the output of model results, constitute the core steps of the entire deep learning process. In the preliminary processing stage, the original input data is first cleaned and standardized, including denoising, normalization, missing value filling, etc., to ensure data quality and consistency. In addition, according to the needs of specific problems, it may be necessary to perform feature extraction, dimensionality reduction and other operations on the data to improve the training efficiency and accuracy of the model. In the construction stage of the network model, according to the characteristics of the data and the task objectives, select the appropriate network architecture (such as convolutional neural network CNN, recurrent neural network RNN, etc.), and configure related hyperparameters such as learning rate, number of layers, activation function, etc. Through reasonable network design, ensure that the model can effectively learn meaningful features from the input data and has strong generalization ability. During the training process, the performance of the model is evaluated by calculating the loss function, and the network parameters are continuously optimized through the back propagation algorithm. In this process, the output of the loss curve can intuitively reflect the convergence of the model training process, as well as the performance of the model on the training set and validation set, and thus provide a reference for model tuning. Finally, after sufficient training, the model can output prediction results and perform subsequent processing or application according to actual needs.

[0034] Step 3: Save the best training weights for subsequent model inversion. During the model training process, through continuous iterative optimization, select the training weights with the smallest loss function and the best performance and save them. These weight parameters will be used as initial values ​​in the subsequent inversion process to ensure that the inversion model can perform efficient inference and prediction based on a fully trained network.

[0035] In the inversion stage, the resistivity and polarizability data are input and three-dimensional inversion is performed through the trained network model. Based on the input observation data, the network gradually derives the distribution of electrical properties of the underground medium and generates corresponding three-dimensional resistivity and polarizability inversion results. Subsequently, the inversion results are compared and analyzed with the real geological model to evaluate the accuracy and stability of the inversion model. By comparing the error between the inversion results and the true value, the performance of the network can be effectively evaluated, including the accuracy of the inversion, the convergence speed, and the generalization ability of the model on unknown data. This process helps to further optimize the model and provide a reliable prediction and analysis tool for geological exploration in practical applications.

[0036] Further, the step 1 specifically includes:

[0037] The finite element method is used for forward modeling, aiming to obtain the resistivity observations, polarizability observations and resistivity-polarizability coupling data of representative contaminated sites, so as to establish a sample data set. First, the underground space of the exploration area is divided into rectangular grid cells, each of which is 32×32×32 in size. Each cell has a known position, side length, and is assigned a resistivity or polarizability value. Figure 2 a and Figure 2 As shown in (b), the complex contaminated target model is constructed by three ellipsoids whose axis lengths and positions are randomly assigned. The non-uniform background model consists of five ellipsoids in the same layer, whose axis lengths and spatial inclinations are randomly assigned. These ellipsoids divide the uniform background into three layers. Finally, the contaminated target model is placed at the center of the background model to generate the layered and non-layered medium models required for training. Figure 2 c and Figure 2 As shown in Figure d, the apparent resistivity and apparent polarizability on the ground are calculated using the finite element method.

[0038] Further, the step 2 specifically includes:

[0039] The resistivity-polarizability coupling value is calculated by coupling the apparent resistivity and the apparent polarizability, and can be expressed as:

[0040]

[0041] In the formula, Z0 represents the observed value of resistivity-polarizability coupling, and Z Z represents the observed value of resistivity, Z j Represents the observed value of the polarizability. The data obtained by forward calculation is preprocessed and randomly divided into training samples (80%) and test samples (20%) for training with known labels. In addition, in order to enhance the robustness of the network, the present invention adds 5% noise perturbation samples to each training sample. Then, the model is tested and optimized based on the error loss output by the calculation model. Finally, the trained inversion model is tested using synthetic data to verify its effectiveness and adaptability.

[0042] In order to reconstruct the 3D model from the spatial features of the electrical data, the U-Net architecture was adjusted and modified, such as Figure 3 shown. Figure 3 A classic fully convolutional network (FCN) is shown, which consists of two main parts. In this FCN, the left side acts as an encoder and the right side acts as a decoder. The encoder consists of four submodules, each of which contains two convolutional layers. Each submodule is downsampled by performing a convolution operation using a 3×3 convolution kernel with a stride of 2. In addition, a dropout layer is added to minimize overfitting.

[0043] Specifically, the present invention inputs a single resistivity anomaly map with a resolution of 32×32, which means that 1,024 data sets are evenly distributed on the surface. The resolutions of modules 1 to 5 are 32×32, 16×16, 8×8, 4×4, and 2×2, respectively, and the number of channels of each subsequent module is half of the previous module. The operation of this part of the network is similar to a standard convolutional neural network (CNN), using a 3×3 convolution kernel to capture the implicit pixel relationships in the image.

[0044] The decoder is mirror-symmetric to the encoder and contains four modules. Through upsampling, the features are gradually enlarged until the output resolution matches the input image. At the same time, the network uses jump connections to connect the upsampling results with the outputs of the corresponding resolution submodules in the encoder. These connections serve as inputs to subsequent submodules in the decoder to extract more accurate contextual information and achieve better segmentation effects. Finally, a Sigmoid function with a value range of (0,1) is used to predict each pixel. The present invention will train three independent neural networks simultaneously to predict resistivity, polarizability, and resistivity-polarizability coupling, respectively.

[0045] In this process, the reasonable design of the loss function is crucial. It ensures that each network can effectively learn its specific task while collaborating with other networks for optimization. The loss function must comprehensively consider the target task of each network, balance the errors between different networks, and avoid overfitting through appropriate regularization strategies. This enhances the generalization ability and practical application effect of the model. Therefore, in the experiment, the present invention uses the Dice coefficient for segmentation and outlines the sharp edges of small targets, especially when the number of voxels in the target and background is unbalanced. The Dice coefficient ranges from 0 to 1 and measures the similarity between two binary voxel sets. When the two sets are exactly the same, the Dice coefficient reaches a maximum value of 1. The Dice coefficient can be calculated by the following formula:

[0046]

[0047] It is characterized in that m i and m0 represent the predicted model, the true model and the specified values ​​of the corresponding electrical data respectively.

[0048] Further, the step three specifically includes:

[0049] Figure 4 The loss curve of the inverted network is shown, showing that the loss of the validation dataset begins to slow down when the number of training rounds reaches 250, thus satisfying the condition for early stopping training. Subsequently, the randomly generated models are inverted and tested.

[0050] like Figure 5As shown in the figure, the algorithm proposed in this study verifies the prediction results of the random homogeneous medium model. Each pixel value in the inversion result is scored in the range of 0 to 1, indicating the prediction value related to the surrounding medium (marked as 0) and the target body (marked as 1). Overall, the 3D volume distribution obtained by inversion is highly consistent with the 3D real model in terms of shape, burial depth and tilt direction. Figure 5 Detailed plane cross sections comparing the true model with the predicted model are shown in (c) and 5(d), demonstrating the good performance of the proposed algorithm.

[0051] Figure 6 The prediction results of the random layered medium model are shown. In the inversion results, the value of each pixel represents the predicted resistivity or polarizability at that specific location. Overall, the anomaly shape, depth, and tilt direction in the inversion model are roughly consistent with the true model, and the boundaries between different layers are clearly visible. The predicted resistivity and polarizability values ​​are also very accurate.

[0052] Another object of the present invention is to provide an innovative inversion method based on artificial intelligence, which aims to optimize the inversion process through an intelligent algorithm, thereby significantly reducing the computing resources and time costs required for inversion, and effectively reducing the economic losses in the inversion process. This method not only improves the inversion efficiency, but also reduces the high expenses brought by traditional inversion methods on the basis of ensuring the inversion accuracy, and has high practical application value and economic benefits.

[0053] Another purpose of the present invention is to provide an innovative idea that combines artificial intelligence with geophysical inversion, and breaks through the limitations of existing inversion methods through the organic integration of intelligent algorithms and traditional geophysical inversion techniques. This innovative idea can give full play to the advantages of artificial intelligence in data processing, pattern recognition and predictive analysis, improve inversion accuracy and efficiency, and effectively respond to challenges under complex geological conditions, promote the advancement of geophysical inversion technology, and provide more accurate and efficient solutions for related fields.

[0054] Another object of the present invention is to provide a new method for joint inversion of apparent resistivity and apparent polarizability. This method overcomes the limitations of single physical parameter inversion by comprehensively considering the relationship between apparent resistivity and apparent polarizability, effectively combining the advantages of both. Through joint inversion, the electrical characteristics of the underground medium can be more comprehensively obtained, and the accuracy and reliability of the inversion results can be improved. This method not only provides new ideas for inversion under complex geological conditions, but also provides more accurate technical support for underground detection and resource exploration in practical applications.

[0055] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0056] First, the present invention provides an innovative intelligent imaging algorithm that uses a U-Net neural network framework to jointly process the resistivity and polarizability data obtained by the MALM-IP method for contaminated site detection. The U-Net network has shown excellent performance in tasks such as medical image segmentation and semantic segmentation, and its unique encoder-decoder structure can efficiently extract multi-scale feature information while retaining spatial structure. Taking advantage of the U-Net network, we designed a deep learning method for resistivity and polarizability data based on the MALM-IP method. Through this deep learning model, the method effectively integrates the two data types, which not only makes up for the limitations of a single data source, but also enables more accurate contaminated site detection results in a wider range of underground environments.

[0057] The present invention fills the gap that the resistivity and polarizability data based on the MALM-IP method cannot be processed using traditional three-dimensional inversion methods. Traditional three-dimensional inversion methods are often unable to process resistivity and polarizability data at the same time, while the present invention innovatively combines the two data types and makes full use of their respective spatial characteristics to achieve joint inversion, thereby greatly improving the accuracy and reliability of the inversion results.

[0058] Compared with the prior art, the method provided by the present invention can more accurately capture the electrical characteristics of underground media when processing resistivity and polarizability data, the inversion results are more reliable, and the underground geological structure characteristics can be better explained. Through this innovative joint inversion method, we can more comprehensively identify the location, depth and distribution of underground pollution sources, significantly improving the accuracy and efficiency of contaminated site detection.

[0059] Second, the difficulty of solving the technical problems existing in the existing technology:

[0060] (1) Although traditional monitoring methods are relatively mature in theory, they have great limitations in practical applications. Developing a comprehensive, rapid, low-interference, and long-term monitoring technology solution requires innovation and optimization of existing technologies.

[0061] (2) Traditional electromagnetic and ground-penetrating radar technologies face problems such as weak signals and severe environmental interference, which involve multiple complex factors in geophysics and engineering technology.

[0062] (3) The inversion problem itself is a complex problem in mathematics and computation, especially the joint inversion of three-dimensional resistivity and polarizability, which involves a large amount of data processing and high-performance computing.

[0063] (4) The coupling relationship between resistivity and polarizability is complex, and the existing inversion algorithms cannot effectively solve this problem.

[0064] (5) Artificial intelligence, especially deep learning technology, has achieved remarkable achievements in many fields. However, the application of AI in geophysical inversion still faces many challenges.

[0065] (6) Large data volumes, strong nonlinearity, and complex underground structures are fundamental problems in geophysical inversion. In order to improve imaging accuracy, it is necessary to process large-scale complex data and solve the problem of multiple solutions during the inversion process.

[0066] The significance of solving the above technical problems: underground pollution (such as heavy metals, chemical leakage, etc.) has a great long-term impact on the environment, especially in agriculture and water resources, which will bring long-term negative consequences. By solving technical problems such as the joint inversion of resistivity and polarizability, the spatial distribution and depth of underground pollution can be determined more accurately, and then accurate pollution source tracking and monitoring can be carried out.

[0067] Third, traditional geological property inversion methods usually rely on manual experience and inefficient forward-inversion iterative processes, which leads to low model reconstruction accuracy and long calculation time. In addition, resistivity and polarizability data under complex geological conditions are often interfered by noise, making it difficult to accurately extract underground structural features. Especially in the field of pollution monitoring, existing technologies have difficulty in achieving high-precision three-dimensional predictions of pollution range and concentration, limiting the application of technology in environmental governance.

[0068] The present invention constructs a three-dimensional geological inversion model for underground medium characteristics by combining the finite element method and deep learning technology. Its data preprocessing step effectively cleans up noise and outliers, improving the quality of observation data; the optimized U-Net deep learning model can extract key features from resistivity and polarizability data through encoder and decoder structures, and achieve high-precision three-dimensional inversion under complex geological conditions. In addition, the multi-task deep learning network can simultaneously predict resistivity, polarizability and their coupling values, significantly improving the inversion efficiency and comprehensiveness of the results.

[0069] Compared with traditional methods, the present invention significantly improves the inversion accuracy and computational efficiency. By optimizing the model loss through the Dice coefficient, the inversion results can more accurately restore the details of the underground structure, especially showing excellent segmentation effects in the detection of small volume targets. In the field of heavy metal pollution monitoring, the present invention can accurately identify the scope and concentration distribution of the polluted area through three-dimensional inversion, providing a scientific basis for environmental remediation. In addition, through the early stopping mechanism and weight preservation, the model training time is further reduced, and the application efficiency in industrial scenarios is improved.

[0070] The technical solution of the present invention has broad industrial application prospects in the fields of environmental pollution monitoring, resource exploration, mining area management, etc. In particular, in the prediction of pollution diffusion paths and heavy metal pollution monitoring, the present invention can quickly provide a reliable three-dimensional model of underground media, greatly improving the scientificity and efficiency of pollution control decisions. In addition, the present invention can also be extended to oil and gas exploration and groundwater resource assessment, providing technical support for the intelligence and efficiency of geological related industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of a method for joint inversion of three-dimensional resistivity and polarizability data provided by an embodiment of the present invention;

[0072] Figure 2 The embodiment of the present invention provides: observation system, model setting and model forward model data. (a) MALM-IP measurement of irregular anomaly; (b) inhomogeneous medium background model; (c) and (d) apparent resistivity and apparent polarizability data respectively;

[0073] Figure 3 The embodiment of the present invention provides: a schematic diagram of the structure of a U-Net convolutional neural network;

[0074] Figure 4 The embodiments of the present invention provide: loss curves of the invention on the training set and the test set;

[0075] Figure 5 The embodiment of the present invention provides: prediction results of an object in a uniform background medium. (a) 3D real model, (b) 3D predicted model, (c) plane section of the real model at a depth of 20 meters, (d) plane section of the predicted model at a depth of 20 meters;

[0076] Figure 6 The embodiment of the present invention provides: prediction results of an object in a random layered medium. (a) true resistivity model, (b) predicted resistivity model, (c) true polarization model, (d) predicted polarization model;

[0077] Figure 7 The embodiment of the present invention provides: a physical model of the initial state of heavy metal pollution and a schematic diagram of the observation system: (a) model space; (b) pollution sample filling; (c) plane view of the observation system; (d) cross-sectional view of the ground observation system;

[0078] Figure 8 The embodiments of the present invention provide: (a) apparent polarizability and (b) apparent resistivity of a physical model;

[0079] Fig. 9The embodiment of the present invention provides: the predicted polarizability results of the physical model. (a) Plan view of the predicted model at 4cm and 7cm depth, (b) xoz cross-sectional view of the predicted model, (c) reconstructed 3D polarizability model;

[0080] Fig.10 The embodiments of the present invention provide: the predicted resistivity results of the physical model. (a) Plan view of the predicted model at 4 cm and 7 cm depth, (b) xoz cross-sectional view of the predicted model, (c) reconstructed 3D resistivity model;

[0081] Fig.11 It is a structural diagram of a three-dimensional geological characteristic inversion system based on deep learning provided by an embodiment of the present invention;

[0082] Fig.12 is a structural diagram of a data construction module provided by an embodiment of the present invention;

[0083] In the figure: 1. Data construction module; 2. Data preprocessing module; 3. Deep learning model module; 4. Optimization and output module; 5. Geological modeling unit; 6. Forward calculation unit; 7. Coupling calculation unit. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0085] like Figure 1 As shown, an embodiment of the present invention provides a novel resistivity and polarizability joint inversion imaging method, which specifically includes:

[0086] S1, to build the data set required for deep learning, it is first necessary to design and generate a real geological model, which should be able to truly reflect the underground structural characteristics of the target area. The core content of the data set includes: the forward calculation results based on these real geological models, specifically the observation values ​​of resistivity and polarizability. These observations should cover a variety of scenarios under different geological conditions to ensure the breadth and representativeness of the data. In addition, observation data at different depths and different spatial resolutions should also be included to enhance the generalization ability and accuracy of the model, and ultimately provide sufficient and high-quality input data for the training of deep learning algorithms.

[0087] S2, the preliminary processing of input data, the construction of network models, the output of loss curves and the output of model results constitute the core steps of the entire deep learning process. In the preliminary processing stage, the original input data is first cleaned and standardized, including denoising, normalization, missing value filling, etc., to ensure data quality and consistency. In addition, according to the needs of specific problems, it may be necessary to perform feature extraction, dimensionality reduction and other operations on the data to improve the training efficiency and accuracy of the model. In the construction stage of the network model, according to the characteristics of the data and the task objectives, select the appropriate network architecture (such as convolutional neural network CNN, recurrent neural network RNN, etc.), and configure related hyperparameters such as learning rate, number of layers, activation function, etc. Through reasonable network design, ensure that the model can effectively learn meaningful features from the input data and has strong generalization ability. During the training process, the performance of the model is evaluated by calculating the loss function, and the network parameters are continuously optimized through the back propagation algorithm. In this process, the output of the loss curve can intuitively reflect the convergence of the model training process and the performance of the model on the training set and validation set, thereby providing a reference for model tuning. Finally, after sufficient training, the model can output prediction results and perform subsequent processing or application according to actual needs.

[0088] S3, save the best training weights for subsequent model inversion. During the model training process, through continuous iterative optimization, the training weights with the smallest loss function and the best performance are selected and saved. These weight parameters will be used as initial values ​​in the subsequent inversion process to ensure that the inversion model can perform efficient inference and prediction based on a fully trained network.

[0089] In the inversion stage, the resistivity and polarizability data are input and three-dimensional inversion is performed through the trained network model. Based on the input observation data, the network gradually derives the distribution of electrical properties of the underground medium and generates corresponding three-dimensional resistivity and polarizability inversion results. Subsequently, the inversion results are compared and analyzed with the real geological model to evaluate the accuracy and stability of the inversion model. By comparing the error between the inversion results and the true value, the performance of the network can be effectively evaluated, including the accuracy of the inversion, the convergence speed, and the generalization ability of the model on unknown data. This process helps to further optimize the model and provide a reliable prediction and analysis tool for geological exploration in practical applications.

[0090] Further, the S1 specifically includes:

[0091] The finite element method is used for forward modeling, aiming to obtain the resistivity observations, polarizability observations and resistivity-polarizability coupling data of representative contaminated sites, so as to establish a sample data set. First, the underground space of the exploration area is divided into rectangular grid cells, each of which is 32×32×32 in size. Each cell has a known position, side length, and is assigned a resistivity or polarizability value. Figure 2 a and Figure 2 As shown in (b), the complex contaminated target model is constructed by three ellipsoids whose axis lengths and positions are randomly assigned. The non-uniform background model consists of five ellipsoids in the same layer, whose axis lengths and spatial inclinations are randomly assigned. These ellipsoids divide the uniform background into three layers. Finally, the contaminated target model is placed at the center of the background model to generate the layered and non-layered medium models required for training. Figure 2 c and Figure 2 As shown in Figure d, the apparent resistivity and apparent polarizability on the ground are calculated using the finite element method.

[0092] Further, the S2 specifically includes:

[0093] The resistivity-polarizability coupling value is calculated by coupling the apparent resistivity and the apparent polarizability, and can be expressed as:

[0094]

[0095] In the formula, Z0 represents the observed value of resistivity-polarizability coupling, and Z Z represents the observed value of resistivity, Z j Represents the observed value of the polarizability. The data obtained by forward calculation is preprocessed and randomly divided into training samples (80%) and test samples (20%) for training with known labels. In addition, in order to enhance the robustness of the network, the present invention adds 5% noise perturbation samples to each training sample. Then, the model is tested and optimized based on the error loss output by the calculation model. Finally, the trained inversion model is tested using synthetic data to verify its effectiveness and adaptability.

[0096] In order to reconstruct the 3D model from the spatial features of the electrical data, the U-Net architecture was adjusted and modified, such as Figure 3 shown. Figure 3 A classic fully convolutional network (FCN) is shown, which consists of two main parts. In this FCN, the left side acts as an encoder and the right side acts as a decoder. The encoder consists of four submodules, each of which contains two convolutional layers. Each submodule is downsampled by performing a convolution operation using a 3×3 convolution kernel with a stride of 2. In addition, a dropout layer is added to minimize overfitting.

[0097] Specifically, the present invention inputs a single resistivity anomaly map with a resolution of 32×32, which means that 1,024 data sets are evenly distributed on the surface. The resolutions of modules 1 to 5 are 32×32, 16×16, 8×8, 4×4, and 2×2, respectively, and the number of channels of each subsequent module is half of the previous module. The operation of this part of the network is similar to a standard convolutional neural network (CNN), using a 3×3 convolution kernel to capture the implicit pixel relationships in the image.

[0098] The decoder is mirror-symmetric to the encoder and contains four modules. Through upsampling, the features are gradually enlarged until the output resolution matches the input image. At the same time, the network uses jump connections to connect the upsampling results with the outputs of the corresponding resolution submodules in the encoder. These connections serve as inputs to subsequent submodules in the decoder to extract more accurate contextual information and achieve better segmentation effects. Finally, a Sigmoid function with a value range of (0,1) is used to predict each pixel. The present invention will train three independent neural networks simultaneously to predict resistivity, polarizability, and resistivity-polarizability coupling, respectively.

[0099] In this process, the reasonable design of the loss function is crucial. It ensures that each network can effectively learn its specific task while collaborating with other networks for optimization. The loss function must comprehensively consider the target task of each network, balance the errors between different networks, and avoid overfitting through appropriate regularization strategies. This enhances the generalization ability and practical application effect of the model. Therefore, in the experiment, the present invention uses the Dice coefficient for segmentation and outlines the sharp edges of small targets, especially when the number of voxels in the target and background is unbalanced. The Dice coefficient ranges from 0 to 1 and measures the similarity between two binary voxel sets. When the two sets are exactly the same, the Dice coefficient reaches a maximum value of 1. The Dice coefficient can be calculated by the following formula:

[0100]

[0101] It is characterized in that m i and m0 represent the predicted model, the true model and the specified values ​​of the corresponding electrical data respectively.

[0102] Further, the S3 specifically includes:

[0103] Figure 4 The loss curve of the inverted network is shown, showing that the loss of the validation dataset begins to slow down when the number of training rounds reaches 250, thus satisfying the condition for early stopping training. Subsequently, the randomly generated models are inverted and tested.

[0104] like Figure 5As shown in the figure, the algorithm proposed in this invention verifies the prediction results of the random homogeneous medium model. Each pixel value in the inversion result is scored in the range of 0 to 1, indicating the prediction value related to the surrounding medium (marked as 0) and the target body (marked as 1). In general, the 3D volume distribution obtained by inversion is highly consistent with the 3D real model in terms of shape, burial depth and tilt direction. Figure 5 (c) and Figure 5 A detailed plane cross section comparing the true model with the predicted model is shown in (d), demonstrating the good performance of the proposed algorithm.

[0105] Figure 6 The prediction results of the random layered medium model are shown. In the inversion results, the value of each pixel represents the predicted resistivity or polarizability at that specific location. Overall, the anomaly shape, depth, and tilt direction in the inversion model are roughly consistent with the true model, and the boundaries between different layers are clearly visible. The predicted resistivity and polarizability values ​​are also very accurate.

[0106] The present invention will be further described below in conjunction with specific embodiments.

[0107] Example 1

[0108] The effectiveness of the invention in monitoring heavy metal pollution was verified by conducting a chamber experiment on soil samples collected from tailings pond wastewater in southwest China. Non-geophysical methods were used in the experiment to sample and detect soils with different degrees of pollution, including heavily polluted, moderately polluted, lightly polluted, slightly polluted and unpolluted soils. Six types of heavy metal contaminated soil samples and their surrounding unpolluted soils were collected and analyzed. The resistivity and polarizability of the contaminated soil samples and the soil around them were tested by the resistivity profile method, and the results showed significant differences between the six types of contaminated samples. For heavily polluted, highly polluted, moderately polluted and lightly polluted soils, the polarizability exceeded 2%, among which the average polarizability of heavily polluted, moderately polluted and lightly polluted soils was 11.84%, 6.83% and 3.19%, respectively. Therefore, in theory, the pollution level can be determined based on the observed polarizability.

[0109] Figure 7 A simulated physical model of the initial state of heavy metal contamination is shown. The contaminated liquid leaks and spreads through narrow channels, forming fan-shaped contaminated soil areas. In this model, the leakage channel corresponds to heavily contaminated soil (average polarizability is 11.84%), simulating the initial heavily contaminated area. The fan-shaped area is filled with lightly contaminated soil (average polarizability is 3.19%), representing the area where the concentration decreases after the pollution spreads. The surrounding uncontaminated soil in the water tank is filled with water to dilute the contaminants (average polarizability is 2%). The contaminated body is 12 cm long, and the horizontal dimensions in the depth direction are 2 cm and 4 cm (each is half of the total length of the contaminated body, as shown in Figure 2). Figure 7 d). The top surface of the contaminated body is 4 cm above the ground. The forward modeling data was obtained using MALM-IP. We used MALM-IP to conduct experiments on the simulated physical model. The experimental data obtained are shown in Figure 8 shown.

[0110] like Fig. 9 (c) and Fig.10 As shown in (c), the present invention uses the described method to reconstruct a 3D underground model of polarizability and resistivity. As the depth of the formation increases, the polarizability increases accordingly, while the resistivity decreases. This may be due to the downward penetration of water during the dilution process, resulting in a higher water content at the bottom, which leads to a higher polarizability and lower resistivity. Fig. 9 As shown in (a), when the depth is 6 cm, the leakage channel designed by us (average polarizability is 10%) can be clearly observed in the xoy cross section. At a depth of 7 cm, the shape of the channel changes to a fan shape, which is consistent with the outward leakage and diffusion of the narrow channel in the simulated physical model (average polarizability is 5%). Fig. 9 (b) and Fig.10 (b) shows the Figure 7 (d) The cross-sections at the same positions of the simulated physical model, especially L1-L1', L2-L2' and L3-L3'. It can be seen that the reconstructed 3D model is in good agreement with the simulated physical model at all positions, showing good accuracy.

[0111] like Fig.11 As shown, an embodiment of the present invention provides a three-dimensional geological characteristic inversion system based on deep learning, characterized in that it includes the following modules:

[0112] Data construction module 1: used to design a real geological model and generate resistivity observations, polarizability observations and resistivity-polarizability coupling data based on the finite element method;

[0113] Data preprocessing module 2: used to denoise, normalize and fill missing values ​​of forward modeling data, randomly divide training samples and test samples, and add noise perturbation samples to enhance robustness;

[0114] Deep learning model module 3: used to construct an adjusted U-Net network to extract and reconstruct the three-dimensional electrical characteristics distribution of the underground medium through encoders, decoders and skip connections;

[0115] Optimization and output module 4: used to train deep learning models, optimize model parameters and output three-dimensional resistivity and polarizability distribution results.

[0116] like Fig.12 As shown, the data construction module 1 includes:

[0117] Geological modeling unit 5: used to divide the underground space into rectangular grid cells, assign resistivity or polarizability values ​​to each cell, and generate target models and background models within the grid;

[0118] Forward calculation unit 6: used to simulate the apparent resistivity and apparent polarizability distribution under complex geological conditions using the finite element method and generate a forward modeling data set;

[0119] Coupling calculation unit 7: used to calculate the resistivity-polarizability coupling value based on the resistivity and polarizability data.

[0120] The deep learning model module 3 includes:

[0121] Encoder submodule: downsamples the input data through multiple convolutional layers and convolution operations with a stride of 2, and extracts features;

[0122] Decoder submodule: gradually restores feature resolution by upsampling and combines the output of the corresponding module of the encoder with the skip connection;

[0123] Loss calculation submodule: used to calculate the loss function based on the Dice coefficient and optimize the model parameters through back propagation.

[0124] The optimization and output module 4 includes:

[0125] Early stopping monitoring unit: used to monitor the loss curve of the validation data set during model training, and trigger the early stopping mechanism when the loss decrease rate slows down;

[0126] Weight saving unit: used to save the best training weight when the loss of the validation data set is minimal;

[0127] Result analysis unit: used to output the predicted three-dimensional resistivity and polarizability distribution, and compare and analyze with the real geological model to evaluate the accuracy and stability of the system.

[0128] Example 2: Mining area pollution monitoring and 3D inversion system based on deep learning

[0129] In a mining area in southwest China, soil areas affected by tailings pond wastewater leakage were contaminated with heavy metals. In order to monitor and assess the scope and concentration of contamination, the following steps were implemented:

[0130] 1. Data construction

[0131] Soil samples with six different contamination levels (including heavily contaminated, moderately contaminated, lightly contaminated, slightly contaminated and uncontaminated soil) were collected from the mining area and their resistivity and polarizability were tested.

[0132] The polluted area is simulated as a rectangular grid model, and the pollution source and diffusion area are simulated by combining the ellipsoid shape. The apparent resistivity and polarizability of the polluted area are calculated based on the finite element method.

[0133] Generate a forward modeling data set that includes resistivity, polarizability, and coupling data at different contamination concentrations.

[0134] 2. Deep learning model training

[0135] The collected forward modeling data were denoised, normalized and missing values ​​filled, and the data were randomly divided into a training set (80%) and a test set (20%).

[0136] The adjusted U-Net deep learning model was constructed, the preprocessed resistivity and polarizability data were input, and the Dice coefficient was used as the loss function for optimization.

[0137] 5% noise perturbation samples are added to the training data to enhance the robustness and generalization ability of the model.

[0138] 3. Result inversion and verification

[0139] The trained model is used to perform a three-dimensional inversion of the randomly generated mining area pollution model to predict the resistivity and polarizability distribution of the contaminated area.

[0140] By comparing the inversion results with the true model, it was found that the inversion model can accurately reconstruct the shape, depth and concentration distribution of the polluted area. The predicted value of the polarizability in the heavily polluted area is about 11.8%, which is basically consistent with the true value (11.84%).

[0141] 4. Application Analysis

[0142] The inversion results are used to evaluate the pollution diffusion path and concentration, and specific plans for pollution remediation are proposed.

[0143] Example 3: Three-dimensional inversion and prediction of tailings pond leakage and diffusion model

[0144] In a tailings pond leakage experiment, a physical model was designed to simulate the heavy metal pollution diffusion process, and a three-dimensional inversion analysis was performed.

[0145] The simulated tailings pond seepage channel is a narrow rectangular channel, through which the contaminated liquid leaks into the soil to form a fan-shaped diffusion area.

[0146] The model area is 12 cm long, the top of the channel is 4 cm from the ground surface, and the polluted area includes a heavily polluted area (average polarizability 11.84%) and a lightly polluted area (average polarizability 3.19%).

[0147] The MALM-IP method is used to test the apparent resistivity and apparent polarizability and generate a forward modeling data set.

[0148] A deep learning model was constructed and the resistivity, polarizability, and coupling data of the test were input.

[0149] The model extracts deep features through a four-layer encoder and uses a four-layer decoder combined with jump connections to restore the three-dimensional structure of the pollution distribution.

[0150] The Dice coefficient is used to optimize the model to ensure that the contaminated area can be clearly segmented even when the target volume is small.

[0151] The three-dimensional model obtained by inversion shows that the shape of the leakage channel is clearly observed at a depth of 6 cm, the contaminated area around the channel diffuses in a fan shape, and the predicted polarizability and resistivity values ​​are highly consistent with the simulation values.

[0152] The inversion cross sections at different depths show good consistency with the real model, and the predicted shape, depth and tilt direction are consistent with the actual situation of the physical model.

[0153] The diffusion path of tailings pond leakage is predicted based on the inversion results, providing a basis for remediation of polluted areas.

[0154] Verify that this method is applicable to heavy metal pollution monitoring and assessment in other similar contaminated sites.

[0155] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0156] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A three-dimensional geological property inversion method based on deep learning, characterized in that: The following steps are involved: Design and generate realistic geological models that reflect the characteristics of underground structures; Based on the finite element method, forward calculation is performed on the real geological model to obtain resistivity observation values, polarizability observation values ​​and resistivity-polarizability coupling data; The forward modeling data is denoised, normalized, and missing value filled, randomly divided into training samples and test samples, and noise disturbance is added to the training samples to enhance robustness; Build and train the adjusted U-Net deep learning model, and perform feature extraction and reconstruction through encoder and decoder combined with skip connections; Input preprocessed resistivity, polarizability and their coupling data to train a deep learning model to predict the three-dimensional electrical properties of the underground medium; Based on the Dice coefficient optimization model, the three-dimensional resistivity and polarizability distribution are output to evaluate the inversion accuracy and stability.

2. The method according to claim 1, characterized in that The forward calculation steps adopt the finite element method, which specifically includes the following contents: The underground space is divided into rectangular grid cells, and each cell is assigned a resistivity or polarizability value; The target model and background model are constructed within the grid, and complex geological conditions are simulated by randomly assigning the axis length, position and spatial inclination of the ellipsoid; The apparent resistivity and apparent polarizability data are calculated.

3. The method according to claim 1, characterized in that The division ratio of training samples and test samples is 80% and 20%, and 5% of noise perturbation samples are added to the training samples to enhance the robustness and generalization ability of the model.

4. The method according to claim 1, characterized in that The deep learning model is an adjusted U-Net network, including: The encoder part on the left is downsampled through multiple convolutional layers and convolution operations with a stride of 2; The decoder part on the right gradually restores the feature resolution by upsampling; Skip connections are used to combine the outputs of the corresponding modules of the encoder and decoder to extract more contextual information.

5. The method according to claim 1, characterized in that The Dice coefficient is used as the loss function in the model optimization process.

6. The method according to claim 1, characterized in that The method comprises the following early stopping training steps: Monitor the loss curve of the validation dataset during model training; When the rate of decrease of the validation set loss slows down, the early stopping mechanism is triggered; The best training weights with the minimum loss on the validation set are saved for subsequent geological property inversion tasks.

7. A three-dimensional geological property inversion system based on deep learning according to any one of claims 1 to 6, characterized in that: Includes the following modules: Data construction module: used to design a real geological model and generate resistivity observations, polarizability observations and resistivity-polarizability coupling data based on the finite element method; Data preprocessing module: used to denoise, normalize and fill missing values ​​on forward modeling data, randomly divide training samples and test samples, and add noise perturbation samples to enhance robustness; Deep learning model module: used to build an adjusted U-Net network to extract and reconstruct the three-dimensional electrical characteristics distribution of underground media through encoders, decoders and skip connections; Optimization and output module: used to train deep learning models, optimize model parameters and output three-dimensional resistivity and polarizability distribution results.

8. The system according to claim 7, characterized in that The data construction module includes: Geological modeling unit: used to divide the underground space into rectangular grid cells, assign resistivity or polarizability values ​​to each cell, and generate target models and background models within the grid; Forward calculation unit: used to simulate the apparent resistivity and apparent polarizability distribution under complex geological conditions using the finite element method and generate forward modeling data sets; Coupling calculation unit: used to calculate the resistivity-polarizability coupling value based on resistivity and polarizability data.

9. The system according to claim 7, characterized in that The deep learning model module includes: Encoder submodule: downsamples the input data through multiple convolutional layers and convolution operations with a stride of 2, and extracts features; Decoder submodule: gradually restores feature resolution by upsampling and combines the output of the corresponding module of the encoder with the skip connection; Loss calculation submodule: used to calculate the loss function based on the Dice coefficient and optimize the model parameters through back propagation.

10. The system according to claim 7, characterized in that The optimization and output module includes: Early stopping monitoring unit: used to monitor the loss curve of the validation data set during model training, and trigger the early stopping mechanism when the loss decrease rate slows down; Weight saving unit: used to save the best training weight when the loss of the validation data set is minimal; Result analysis unit: used to output the predicted three-dimensional resistivity and polarizability distribution, and compare and analyze with the real geological model to evaluate the accuracy and stability of the system.

Images