A fast airborne transient electromagnetic inversion method based on FiLM-U-net

Abstract

This invention discloses a fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net, belonging to the field of electromagnetic signal data processing. The method includes: acquiring geological data and ground-to-space transient electromagnetic detection data; preprocessing the data to construct a dataset; dividing the dataset into training, validation, and test sets; constructing a FiLM-U-net network model that combines the FiLM and U-net structures; training the FiLM-U-net network model using a suitable training strategy selected from the training set; evaluating the model's inversion performance using the test set; and using the trained FiLM-U-net network model to invert the measured ground-to-space transient electromagnetic data. The inputs are the transmitter and receiver locations and the ground-to-space transient electromagnetic response, and the output is the resistivity value of the subsurface model. This invention effectively utilizes spatial information for efficient and rapid inversion of ground-to-space transient electromagnetic signals.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic signal data processing technology, and in particular to a fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net. Background Technology

[0002] Ground-to-air transient electromagnetic (GTE) is an important method for underground exploration. Unlike traditional ground-based or airborne GTE methods, GTE employs ground-based data sources and utilizes unmanned aerial vehicles (UAVs) for aerial reception. This allows for deeper exploration depths while adapting to complex terrain. Currently, GTE is widely used in groundwater exploration, potential mineral resource surveys, investigation of hidden disaster-causing factors, and imaging of urban underground spaces.

[0003] After acquiring transient electromagnetic detection data from the ground and air, inversion imaging is required to interpret unknown subsurface structures and obtain easily interpretable visual images. Due to the excellent detection efficiency of UAVs during field flights, each survey mission collects a large amount of data. Currently, traditional inversion methods based on the least squares method have shown relatively accurate and high-resolution inversion results. However, because the solution process requires iterative calculations using computing equipment, the computational complexity and cost increase rapidly with the amount of data. Limited by available computing resources, traditional inversion processes typically need to be conducted in a laboratory and may take several days to complete. Therefore, a fast and high-precision inversion imaging method is urgently needed to meet the rapid inversion requirements of field exploration missions.

[0004] Deep learning methods can enable rapid imaging of the target area during field exploration by pre-training models before the mission. However, current research only offers deep learning inversion methods for ground-based and airborne transient electromagnetic systems (TEMS). None of these methods consider the transmitter and receiver locations. For ground-to-air TEMS, however, the transmitter and receiver locations are crucial; therefore, integrating this spatial information into the electromagnetic response signal remains a challenge.

[0005] U-net is a proven neural network architecture with excellent performance. However, U-net still struggles to fuse spatial information and electromagnetic response information at the input. Feature-wise Linear Modulation (FiLM), as an effective modal fusion technique, can achieve feature fusion of spatial information and electromagnetic response information.

[0006] In conclusion, there is an urgent need to design a method for rapid inversion of ground-to-space transient electromagnetic data in order to efficiently and accurately interpret the collected data. Summary of the Invention

[0007] This invention provides a fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net, which aims to integrate spatial information of the transmitting and receiving locations to improve inversion efficiency and accuracy, and solve the problem of fast inversion of large-scale ground-to-space transient electromagnetic inversion.

[0008] According to one aspect of the present invention, a fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net is provided, comprising: acquiring geological data and ground-to-air transient electromagnetic detection data, preprocessing them, and constructing a dataset; dividing the dataset into a training set, a validation set, and a test set; constructing a FiLM-U-net network model, wherein the FiLM-U-net network model combines a FiLM structure and a U-net structure, and consists of a feature fusion layer, an encoding layer, and a decoding layer, wherein the feature fusion layer is composed of a FiLM structure and is responsible for fusing different feature information; the encoding layer and the decoding layer are composed of convolutional neural networks, which together form a U-net structure; using the training set to select an appropriate training strategy to train the FiLM-U-net network model, and using the test set to evaluate the inversion effect of the model; using the trained FiLM-U-net network model to invert ground-to-air transient electromagnetic measured data, and outputting the corresponding subsurface model resistivity value according to the transmission and reception location of the measured data and the ground-to-air transient electromagnetic response.

[0009] Optionally, the acquisition of geological data and ground-to-air transient electromagnetic detection data, followed by preprocessing to construct a dataset, includes: acquiring the stratigraphic thickness, detection depth, transmission and reception locations, sampling time, and ground-to-air transient electromagnetic response of the electromagnetic detection area; dividing the stratigraphy of the electromagnetic detection area into multiple layers according to equal thickness intervals, with each layer corresponding to a depth range; generating corresponding resistivity values ​​for each layer; normalizing the transmission and reception locations, ground-to-air transient electromagnetic response, and resistivity values ​​of each layer; concatenating the normalized transmission and reception locations with the corresponding ground-to-air transient electromagnetic response to form a fusion vector as the input to the model, and using the normalized resistivity of each layer as the output of the model; and establishing a one-to-one correspondence between each input and output to construct the dataset.

[0010] Optionally, normalization of the transmit / receive location, the ground-to-air transient electromagnetic response, and the resistivity value of each layer includes: normalizing the transmit / receive location by its maximum and minimum values, expressed by the formula:

[0011] ;

[0012] In the formula, Indicates the original send / receive location; Indicates the normalized send / receive position; This indicates taking the maximum value; This indicates taking the minimum value;

[0013] After performing a logarithmic transformation on the ground-to-air transient electromagnetic response, and then normalizing it to its maximum and minimum values, the formula is expressed as:

[0014] ;

[0015] In the formula, This represents the unnormalized raw electromagnetic response; This represents the normalized logarithmic response; Represents the logarithmic function to base 10;

[0016] After performing a logarithmic transformation on the resistivity values ​​of each layer, and then performing a minima-normalization, the formula is expressed as:

[0017] ;

[0018] In the formula, This represents the unnormalized raw resistivity value; This represents the normalized logarithmic resistivity value.

[0019] Optionally, the first part of the FiLM-U-net network model is a location feature fusion module, and the features output by the location feature encoding module are:

[0020] ;

[0021] in, Features derived from fused location information; These are the weighting parameters for FiLM. , This represents the mapping relationship between the weight adjustment parameters of FiLM and the location information. These are the bias adjustment parameters for FiLM. , This represents the mapping relationship between the bias adjustment parameters of FiLM and position information.

[0022] ; ;

[0023] It is a learnable parameterized activation function used to adjust the weights and biases of each neuron.

[0024] ;

[0025] ;

[0026] ;

[0027] m represents the function type, l represents the neural network layer sequence, i represents the output neuron sequence, j represents the input neuron sequence, and w represents the scaling factor. It is a spline function. It is a B-spline basis function. These represent the learnable coefficients that control the weights of the basis functions;

[0028] After the output from the location feature encoding module, the data enters the encoding module. First, a convolution operation with a kernel size of 7, a stride of 2, and padding of 3 is performed on the input data. Then, a downsampling operation and three combined CNN structure operations are performed simultaneously. Each combined CNN structure consists of one downsampling operation and three convolution operations with kernel sizes of 1, 3, and 1, strides of 1, 2, and 1, and padding of 0, 1, and 0, respectively. Each combined CNN structure outputs a feature that is simultaneously fed to both the main branch and the side branches. Next, a 256- and 512-neuron linear layer is input; the 256-neuron linear layer further extracts hidden features. The features are first processed by the encoding layer, and then output by the decoding layer. The features are first convolved with a kernel size of 3, a stride of 1, and padding of 1. Then batch normalization and activation function are performed. Then, three consecutive upsampling operations are performed. Each upsampling process is concatenated with the output of the combined CNN structure operation in the encoding module to decode deep and shallow features simultaneously and improve the generalization of the model. After three upsampling operations, the final output features are batch × 128 × 8. Then, the features are output by the output layer, which sets the number of feature channels to 1 and converts the output size to 10 to achieve the inversion process.

[0029] Optionally, training the FiLM-U-net network model using the training set by selecting an appropriate training strategy includes:

[0030] Based on the training set, the Adam optimizer is used to select... The FiLM-U-net network model is trained using this loss function;

[0031] In the formula, This is the loss function used during model training. For batch number, The batch size for inputting the model in each round of training. The first output of the model Batch resistivity values For the actual first Batch resistivity values;

[0032] Set the initial learning rate to 10. -4A step decay strategy is adopted, which reduces the learning rate to 0.2 times the learning rate of the previous round at training rounds of 200, 300 and 400, and sets the total number of training rounds to 500 and the input batch size to 512.

[0033] During training, the model parameters are continuously adjusted to optimize the loss function. When the value of the loss function on the validation set decreases, the model is considered to have converged when the value of the loss function is less than or equal to the first preset threshold.

[0034] Optionally, evaluating the inversion performance of the model using the test set includes:

[0035] The root mean square error (RMSE) is used to evaluate the FiLM-U-net network model, where RMSE is defined as:

[0036] ;

[0037] In the formula, The sample number. The size of the test set. The first output of the model Group resistivity value, For the actual first Group resistivity values;

[0038] The RMSE is calculated on the test set. If the calculated RMSE is less than or equal to the second preset threshold, the FiLM-U-net network model is considered to be successfully trained.

[0039] This invention provides a fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net. It introduces the FiLM structure into the field of geophysical inversion and combines it with U-net to solve the problem of efficient fusion of multi-source information (especially spatial location information) in deep learning models. At the input end, it achieves feature fusion of ground-to-space transient electromagnetic response data and transmit / receive location, and further extracts the fused features using U-net, enabling U-net to better capture the complex mapping relationship between input and output information. By fusing spatial information with transient electromagnetic response signals, it overcomes the limitations of CNN and LSTM methods in utilizing spatial information. Furthermore, compared to traditional inversion methods that require several seconds to complete the inversion of a single measurement point, this method, with its trained model, can complete the inversion of a measurement line (thousands of points) in several seconds, demonstrating a very fast inversion speed and enabling effective and rapid inversion of ground-to-space transient electromagnetic signals. Attached Figure Description

[0040] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0041] Figure 1 This is a flowchart of a fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the architecture of the FiLM-U-net model in an embodiment of the present invention;

[0043] Figure 3 This is a comparison of the inversion results of the FiLM-U-net inversion method of this invention with those of other inversion methods on the test set, and the results of re-firing the inversion results to the forward response.

[0044] Figure 4 This is a comparison of the results of the FiLM-U-net inversion method of the present invention in the pseudo-two-dimensional inversion. Figure 4 (a) is a pseudo-two-dimensional model generated using a simulation program; Figure 4 (b) shows the inversion results of FiLM-U-net on a quasi-two-dimensional model. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0048] Example 1: Refer to Figure 1 , Figure 1 This is a flowchart of a fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0049] S1: Acquire geological data and ground-to-air transient electromagnetic detection data, perform preprocessing, and construct a dataset;

[0050] This step involves building a dataset for training and testing the FiLM-U-net model, specifically including:

[0051] S11, acquire the stratigraphic thickness, detection depth, transmission and reception location, sampling time, and ground-to-space transient electromagnetic response of the electromagnetic detection area;

[0052] Based on the known geological data of the exploration area, a preliminary assessment of the area is made to determine the resistivity value distribution range of the dataset, the stratigraphic thickness grid division, the exploration depth, the transmission and reception location, and the sampling time;

[0053] The resistivity distribution range can be determined based on existing data. For example, the vertical resistivity curve can be obtained directly through borehole logging. By analyzing the vertical resistivity curve, the resistivity distribution range of the dataset can be roughly determined. Alternatively, the resistivity range can be determined by referring to the results of other electromagnetic detection methods carried out in this area.

[0054] The formation thickness grid mainly considers the formation thickness and the number of formations. It can be freely divided to meet the detection characteristics of ground-to-space transient electromagnetic fields with high resolution for shallow layers and lower resolution for deep layers, so as to achieve rapid imaging of the subsurface while ensuring a certain resolution.

[0055] The detection depth refers to the depth that is desired to be detected in the detection mission. It is set according to the detection requirements and the maximum detection depth of the instrument, and is used to determine the layering of the stratigraphic thickness grid.

[0056] The transmit / receive position refers to the relative position of the transmitting line source and the receiver. The inversion region is determined by three coordinates: x, y, and z (representing the horizontal, vertical, and flight altitude, respectively). In this embodiment of the invention, a single measuring line inversion is used, fixing y and z, and only changing x. Only one electromagnetic response signal will be collected from a single transmit / receive position.

[0057] S12, the strata of the electromagnetic detection area are divided into multiple layers according to equal thickness intervals, and each layer corresponds to a depth range;

[0058] In one example, the strata thickness is divided into ten layers: 30, 40, 50, 80, 100, 120, 130, 150, 200, and 300.

[0059] The resistivity range was set to 1-10000Ω·m, which basically covers the resistivity of the strata in the area to be explored.

[0060] The detection depth is 0-1000m;

[0061] The transmission and reception range is 0-3000m;

[0062] The sampling time was 1μs-1s, and samples were taken 63 times at equal logarithmic intervals. The sampling method was to first take the logarithm of 1e-6 and 1, then take 63 numbers at equal intervals in the logarithmic field, and then take the exponent.

[0063] S13 generates the corresponding resistivity value for each divided stratigraphic layer;

[0064] Based on the above settings, the logarithmic resistivity value of each layer is generated according to a Gaussian random distribution with a mean of 2 and a variance of 1.5.

[0065] Ground-to-air transient electromagnetic forward modeling data is generated according to the determined parameters (resistivity range, formation thickness, detection depth, transmission and reception location, and sampling time).

[0066] S14, normalizes the transmit and receive positions, ground-to-air transient electromagnetic response, and resistivity values ​​of each layer.

[0067] For sending and receiving locations To perform local minima normalization, the normalization formula is: ,in, Indicates the original send / receive location; Indicates the normalized send / receive position; This indicates taking the maximum value; This indicates taking the minimum value; the minimax normalization here is based on the determination of the detection range. As shown above, the coordinate range of x is determined to be 0-3000m, so normalization can be performed directly with the minimum value being 0 and the maximum value being 3000.

[0068] Furthermore, since electromagnetic response and underground resistivity typically span multiple orders of magnitude, taking the logarithm can compress the dynamic range, resulting in a more stable data distribution, which is beneficial for the learning and convergence of neural networks. Therefore, in this embodiment of the invention, the logarithm of the (ground-to-air transient electromagnetic) response signal is taken before minima-maximum normalization, as expressed by the formula: ,in, This represents the unnormalized raw response; This represents the normalized logarithmic response; Represents the logarithmic function to base 10;

[0069] Similarly, after performing a logarithmic transformation on the resistivity values ​​of each layer, and then performing a minima-normalization, the formula is expressed as: ,in, This represents the unnormalized raw resistivity value; This represents the normalized logarithmic resistivity value;

[0070] S15: The normalized transmit and receive positions are concatenated with the corresponding ground-to-air transient electromagnetic responses to form a fusion vector as the input to the model, and the resistivity of each layer after normalization is used as the output of the model; each input and output is mapped one-to-one to construct the dataset.

[0071] Normalized send and receive positions With the normalized logarithmic response The vectors are concatenated into a single fusion vector, which serves as the input to the FiLM-U-net model. The model's labels are the normalized logarithmic resistivity values ​​of the 10 layers. Each such "input-output" pair constitutes a sample in the dataset.

[0072] S2, divide the dataset into a training set, a validation set, and a test set;

[0073] The large amount of sample data after the above preprocessing is divided into training set, validation set and test set according to the ratio of 90%, 5% and 5% respectively.

[0074] The training set is used for learning model parameters; the validation set is used for hyperparameter tuning, preventing overfitting, and assisting in model selection; and the test set is used for the final evaluation of the trained model.

[0075] S3, construct the FiLM-U-net network model;

[0076] In this embodiment of the invention, FiLM-U-net combines the FiLM structure and the U-net structure. The preprocessed transmit and receive positions are fused with the preprocessed electromagnetic response after passing through the FiLM structure, and then enter the U-net structure for inversion. At the same time, the U-net network parameters are optimized for this task (i.e., ground-to-air transient electromagnetic inversion) to achieve better performance. The optimized U-net network parameters include network structure and hyperparameters. The network structure includes the feature input structure designed in conjunction with FiLM, the skip connection method, and the optimization of the encoding and decoding layers. The hyperparameters include the learning rate, optimizer, and regularization strategy.

[0077] The FiLM-U-net model consists of a feature fusion layer, an encoding layer, and a decoding layer. The feature fusion layer, composed of a FiLM (Feature Linear Modulation) structure, is used to fuse different feature information. (Directly concatenating the inputs horizontally into the model relies solely on the model's autonomous learning of the relationship between the two features during subsequent training, and location information is easily obscured by high-dimensional information. Feature fusion modifies the electromagnetic response, better conforming to the propagation laws of the electromagnetic field, and amplifies this nonlinear relationship.) The encoding and decoding layers are primarily constructed using convolutional neural networks (CNNs), together forming the U-net structure. The model input consists of preprocessed transmit / receive positions and ground-to-air transient electromagnetic responses; the output is the preprocessed resistivity values ​​of each subsurface model layer. The FiLM-U-net network structure is established based on the input information dimension (spatial information dimension and number of samples of transient electromagnetic signals) and output information dimension (total number of stratigraphic divisions) of the synthetic dataset. In this embodiment, the input vector size of the FiLM-U-net input layer is set to 1×64 (63 sampling bits + 1 transmit / receive position) and the output vector size of the output layer is set to 1×10 according to the number of sampling time points and stratigraphic divisions selected in S1.

[0078] Reference Figure 2 , Figure 2 This is a schematic diagram of the FiLM-U-net structure in an embodiment of the present invention, as shown below. Figure 2 As shown, the structure consists of three parts: a feature fusion layer fuses the location information with the response signal, which is then input into the coding layer; the coding layer performs multiple downsampling to acquire features at different depths and outputs them to the decoding layer; finally, the decoding layer aggregates the features at multiple depths and uses multiple upsampling to obtain the inverted resistivity distribution.

[0079] The input to FiLM-U-net consists of ground-to-air transient electromagnetic data from a single transceiver location and 63 time-sampling points. The final output is the resistivity values ​​of a 10-layer underground model. The first part of FiLM-U-net is the location feature fusion module. The features output by the location feature encoding module are as follows:

[0080] ;

[0081] in, Features derived from fused location information; These are the weighting parameters for FiLM. , This represents the mapping relationship between the weight adjustment parameters of FiLM and the location information. The signal is a normalized logarithmic response signal that does not contain location information; These are the bias adjustment parameters for FiLM. , This represents the mapping relationship between the bias adjustment parameters of FiLM and position information.

[0082] ; ;

[0083] It is a learnable parameterized activation function used to adjust the weights and biases of each neuron.

[0084] ;

[0085] ;

[0086] ;

[0087] m represents the function type, l represents the neural network layer sequence, i represents the output neuron sequence, j represents the input neuron sequence, and w represents the scaling factor. It is a spline function. It is a B-spline basis function. These represent the learnable coefficients that control the weights of the basis functions; Indicates the normalized send / receive location;

[0088] After the output from the location feature encoding module, the data enters the encoding module. First, a convolution operation with a kernel size of 7, a stride of 2, and padding of 3 is performed on the input data. Then, a downsampling operation and three combined CNN structure operations are performed simultaneously. Each combined CNN structure consists of one downsampling operation and three convolution operations with kernel sizes of 1, 3, and 1, strides of 1, 2, and 1, and padding of 0, 1, and 0, respectively. After each combined CNN structure, a feature is output, simultaneously contributing one feature to both the main branch and the side branches. Next, a 256-neuron and a 512-neuron linear layer are input to further extract hidden features. Finally, the feature size output by the encoding layer is batch × 128 × 8.

[0089] The model then enters the decoding layer, where the features are first subjected to convolutional operations with a kernel size of 3, a stride of 1, and padding of 1, followed by batch normalization and activation. Next, three consecutive upsampling operations are performed, each concatenated with the output of the combined CNN structure operations in the encoding module, to simultaneously decode deep and shallow features and improve model generalization. After three upsampling operations, the final output feature size is batch×128×8. The model then enters the output layer, where the number of feature channels is reduced to 1, and the output size is converted to 10 to achieve the inversion process.

[0090] S4, Use the training set to select an appropriate training strategy to train the FiLM-U-net network model;

[0091] The training strategy includes optimizer selection, loss function type, learning rate, training period, and input batch size. During training, the model's parameters are continuously adjusted to gradually reduce the loss function value on the training set. Simultaneously, the validation set is used to monitor model performance and prevent overfitting. For example, if the model's performance on the training set continuously improves but its performance on the validation set begins to decline, it indicates that the model may be overfitting. In this case, the training strategy needs to be adjusted, such as stopping training early or adding a regularization term.

[0092] In this embodiment of the invention, the Adam optimizer is used, and the selection... The FiLM-U-net network model is trained using this loss function; where... For batch number, The batch size for inputting the model in each round of training. The first output of the model Batch resistivity values For the actual first The batch resistivity values ​​and loss function are used as indicators to guide the model's normal training until convergence.

[0093] The initial learning rate is set to 10. -4 A stepped decay strategy was adopted, with the learning rate decaying to 0.2 times the previous learning rate at training rounds 200, 300, and 400; the total number of training rounds was set to 500; and the input batch size was 512.

[0094] Observe the loss curve on the validation set. When the loss value on the validation set tends to stabilize and no longer decreases significantly (i.e., the loss value is less than or equal to the first preset threshold, such as 0.01), the model is considered to have converged. Save the model after ensuring that it has converged normally.

[0095] S5, Evaluate the inversion effect of the model using the test set;

[0096] The model is validated using a test set to determine whether it achieves the expected results on the dataset and to evaluate its inversion capability. Various evaluation metrics, such as root mean square error (RMSE) and mean absolute error (MAE), are calculated on the test set to determine the model's predictive and generalization abilities for unknown data. If the evaluation metrics meet the expected requirements, it indicates that the model has good performance and can be used for practical ground-to-air transient electromagnetic inversion tasks.

[0097] In this embodiment of the invention, the root mean square error (RMSE) is selected as the error index to evaluate the inversion results of FiLM-U-net, wherein RMSE is defined as:

[0098] ;

[0099] In the formula, The sample number. The size of the test set. The first output of the model Group resistivity value, For the actual first Group resistivity values.

[0100] By calculating the RMSE on the test set, if the calculated RMSE is less than or equal to the second preset threshold (generally, the second preset threshold is set to 0.1), it indicates that the model's overall prediction error of formation resistivity is within an acceptable range and meets the training qualification requirements.

[0101] S6. The trained FiLM-U-net network model is used to invert the measured ground-to-air transient electromagnetic data.

[0102] By preprocessing the measured data collected in the field (normalizing the transmission and reception locations and logarithmic normalization of the ground-to-air transient electromagnetic response) and inputting it into the trained FiLM-U-net model, the distribution of resistivity (normalized logarithmic resistivity) of the underground model can be obtained quickly and accurately, providing important reference for geological exploration, resource development, and other purposes.

[0103] This invention determines the parameters of the dataset based on known geological data of the exploration area, generates a large amount of sample data using a ground-to-air transient electromagnetic forward modeling program, establishes a suitable FiLM-U-net model based on the input and output dimensions of the synthetic dataset, selects an appropriate training strategy for model training to obtain optimal training results, and inputs the preprocessed field exploration data into the trained FiLM-U-net to achieve fast and accurate ground-to-air transient electromagnetic data inversion and interpretation. The fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net proposed in this invention, through reasonable dataset generation and preprocessing, a carefully designed model structure, an appropriate training strategy, and comprehensive evaluation and verification, achieves fast and accurate ground-to-air transient electromagnetic data inversion and interpretation, exhibiting good inversion performance even in complex terrain conditions, and providing an effective technical means for geological exploration and other fields.

[0104] Example 2: To verify the effectiveness of the FiLM-U-net model in the inversion task of this invention, this example randomly selects a set of data from the test set constructed in Example 1, and performs inversion on it using the FiLM-U-net model and three other inversion methods (CNN model, MLP model, and U-net model). The inversion effect is evaluated by calculating the root mean square error. In addition, forward modeling is performed on the underground resistivity models obtained by different inversion methods to obtain forward response curves. The effect is further evaluated by observing the forward response curves and the root mean square error.

[0105] The effect obtained by inversion is as follows Figure 3 The curve on the left is shown. The blue line represents the inversion performance of the FiLM-U-net model of this invention on this test set. The other three colors represent the inversion performance of other inversion methods, with True indicating the true value. The inversion performance is evaluated by calculating the root mean square error (RMSE) of the inversion compared to the true value. A larger RMSE indicates a worse inversion result. The RMSE of the FiLM-U-net model is 5.26%, the CNN model is 55.24%, the MLP model is 53.16%, and the U-net model is 19.66%. The inversion result of FiLM-U-net in this invention is closer to the true underground resistivity.

[0106] Forward modeling was performed on the subsurface resistivity models obtained by different inversion methods, and different forward response curves were obtained, as shown below. Figure 3As shown by the curve on the right, it can be seen that the underground model response obtained by FiLM-U-net is closest to the real model. At the same time, based on the root mean square error between the response signal and the real signal (the root mean square error of the FiLM-U-net model is 4.01%, that of the CNN model is 19.37%, that of the MLP model is 16.52%, and that of the U-net model is 10.23%), the response signal obtained from the re-forward modeling is very close to the real signal.

[0107] The combined results of the inversion and the comparison of the forward response curves show that the FiLM-U-net model in this invention has high accuracy in the inversion task.

[0108] Example 3: This example further utilizes a pseudo-two-dimensional complex model generated by a simulation program to verify the accuracy and reliability of FiLM-U-net in retrieving resistivity distribution under complex terrain.

[0109] A quasi-two-dimensional complex model was generated using a simulation program. This model serves as the reference standard for the actual model in this experiment, and its specific form is as follows: Figure 4 As shown in (a), this model fully considers complex terrain factors and can simulate the diversity and complexity of resistivity distribution in real underground environments. Based on the generated quasi-two-dimensional complex model, its response signal is obtained, and the response signal of the model is inverted using FiLM-U-net. The final inversion result is shown in (a). Figure 4 As shown in (b).

[0110] Comparison with real models Figure 4 (a) The underground model obtained by the present invention Figure 4 (b) It can be found that the inverted model is almost identical to the real model in terms of overall shape, distribution of different resistivity regions and boundary shape, indicating that the FiLM-U-net model in this invention can accurately invert the resistivity distribution results of the real underground model, verifying the inversion performance of this invention in two-dimensional complex underground models.

[0111] In summary, compared with other existing deep learning methods, this invention has better generalization and better inversion accuracy.

[0112] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net, characterized in that, include: Geological data and transient electromagnetic detection data from the ground and air were acquired, preprocessed, and used to construct a dataset. The dataset is divided into a training set, a validation set, and a test set; A FiLM-U-net network model is constructed, which combines the FiLM structure and the U-net structure. It consists of a feature fusion layer, an encoding layer, and a decoding layer. The feature fusion layer is composed of a FiLM structure and is responsible for fusing different feature information. The encoding layer and the decoding layer are composed of convolutional neural networks, which together form the U-net structure. The FiLM-U-net network model is trained using the training set by selecting an appropriate training strategy, and the inversion effect of the model is evaluated using the test set. The trained FiLM-U-net network model is used to invert the measured ground-to-air transient electromagnetic data. Based on the transmission and reception locations of the measured data and the ground-to-air transient electromagnetic response, the corresponding underground model resistivity value is output.

2. The fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to claim 1, characterized in that, The acquisition of geological data and ground-to-air transient electromagnetic detection data, followed by preprocessing and dataset construction, includes: The data includes the stratigraphic thickness, detection depth, transmission and reception location, sampling time, and transient electromagnetic response between the ground and the air in the electromagnetic detection area. The strata in the electromagnetic detection area are divided into multiple layers according to equal thickness intervals, with each layer corresponding to a depth range. Generate corresponding resistivity values ​​for each of the divided stratigraphic layers; Normalize the transmit / receive location, ground-to-air transient electromagnetic response, and resistivity values ​​of each layer. The normalized transmit and receive locations are concatenated with the corresponding ground-to-air transient electromagnetic responses to form a fusion vector, which serves as the input to the model. The resistivity of each layer after normalization is used as the output of the model. By mapping each input to its corresponding output, a dataset is constructed.

3. The fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to claim 2, characterized in that, Normalization of transmit / receive location, ground-to-air transient electromagnetic response, and resistivity values ​​for each layer includes: The transmit and receive positions are normalized by their maximum and minimum values, as expressed by the following formula: ； In the formula, Indicates the original send / receive location; Indicates the normalized send / receive position; This indicates taking the maximum value; This indicates taking the minimum value; After performing a logarithmic transformation on the ground-to-air transient electromagnetic response, and then normalizing it to its maximum and minimum values, the formula is expressed as: ； In the formula, This represents the unnormalized raw electromagnetic response; This represents the normalized logarithmic response; Represents the logarithmic function to base 10; After performing a logarithmic transformation on the resistivity values ​​of each layer, and then performing a minima-normalization, the formula is expressed as: ； In the formula, This represents the unnormalized raw resistivity value; This represents the normalized logarithmic resistivity value.

4. The fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to claim 3, characterized in that, The first part of the FiLM-U-net network model is a location feature fusion module. The features output by the location feature encoding module are: ； in, Features derived from fused location information; These are the weighting parameters for FiLM. , This represents the mapping relationship between the weight adjustment parameters of FiLM and the location information. These are the bias adjustment parameters for FiLM. , This represents the mapping relationship between the bias adjustment parameters of FiLM and position information. ； ； It is a learnable parameterized activation function used to adjust the weights and biases of each neuron. ； ； ； m represents the function type, l represents the neural network layer sequence, i represents the output neuron sequence, j represents the input neuron sequence, and w represents the scaling factor. It is a spline function. It is a B-spline basis function. These represent the learnable coefficients that control the weights of the basis functions; After the output from the location feature encoding module, the data enters the encoding module. First, a convolution operation with a kernel size of 7, a stride of 2, and padding of 3 is performed on the input data. Then, a downsampling operation and three combined CNN structure operations are performed simultaneously. Each combined CNN structure consists of one downsampling operation and three convolution operations with kernel sizes of 1, 3, and 1, strides of 1, 2, and 1, and padding of 0, 1, and 0, respectively. After each combined CNN structure, a feature is output, simultaneously contributing one feature to both the main branch and the side branches. Next, a 256-neuron linear layer with 512 neurons is input to further extract hidden features. Finally, the feature size output by the encoding layer is batch × 128 × 8. The model then enters the decoding layer, where the features are first subjected to convolution operations with a kernel size of 3, a stride of 1, and padding of 1, followed by batch normalization and activation function activation. Next, three consecutive upsampling operations are performed, each concatenated with the output of the combined CNN structure operations in the encoding module, to simultaneously decode deep and shallow features and improve model generalization. After three upsampling operations, the final output feature size is batch×128×8. The model then enters the output layer, where the feature channel number is reduced to 1 and the output size is converted to 10 to achieve the inversion process.

5. The fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to claim 1, characterized in that, Training the FiLM-U-net network model using the training set and selecting an appropriate training strategy includes: Based on the training set, the Adam optimizer is used to select... The FiLM-U-net network model is trained using this loss function; In the formula, This is the loss function used during model training. For batch number, The batch size for inputting the model in each round of training. The first output of the model Normalized resistivity values ​​for the batch For the actual first Normalized resistivity value for the batch; Set the initial learning rate to 10. -4 A step decay strategy is adopted, which reduces the learning rate to 0.2 times the learning rate of the previous round at training rounds of 200, 300 and 400, and sets the total number of training rounds to 500 and the input batch size to 512. During training, the model parameters are continuously adjusted to optimize the loss function. When the value of the loss function on the validation set decreases, the model is considered to have converged when the value of the loss function is less than or equal to the first preset threshold.

6. The fast ground-to-space transient electromagnetic inversion method based on FiLM-U-net according to claim 1, characterized in that, The inversion performance of the model was evaluated using the test set, including: The root mean square error (RMSE) is used to evaluate the FiLM-U-net network model, where RMSE is defined as: ； In the formula, The sample number. The size of the sample size in the test set. The first output of the model Group resistivity value, For the actual first Group resistivity values; The RMSE is calculated on the test set. If the calculated RMSE is less than or equal to the second preset threshold, the FiLM-U-net network model is considered to be successfully trained.

Images