Fast ground-air transient electromagnetic inversion method based on FiLM-U-net

Abstract

The invention discloses a fast air-ground transient electromagnetic inversion method based on FiLM-U-net, and belongs to the field of electromagnetic signal data processing, and the method comprises the steps: obtaining geological data and air-ground transient electromagnetic detection data, carrying out the preprocessing, and constructing a data set; dividing a training set, a verification set and a test set; the method comprises the following steps: constructing a FiLM-U-net network model in combination with a FiLM structure and a U-net structure; adopting the training set to select a proper training strategy to train the FiLM-U-net network model, and adopting the test set to evaluate the inversion effect of the model; and the trained FiLM-U-net network model is used to carry out inversion on ground-air transient electromagnetic measured data, transmit-receive positions and ground-air transient electromagnetic response are inputted, and an underground model resistivity value is outputted. According to the invention, the space information can be effectively utilized to carry out efficient and rapid inversion on the ground-air transient electromagnetic signal.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic signal data processing, and in particular to a fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net. Background Art

[0002] Ground-to-air transient electromagnetic (GATE) is an important underground exploration method. Unlike traditional terrestrial or airborne transient electromagnetic (ET) methods, GATE uses a ground-based source and drones for aerial reception. This allows for deeper detection depths and the use of drones to adapt to complex terrain. GATE is currently widely used in groundwater detection, potential mineral exploration, investigation of hidden disaster-causing factors, and urban underground space imaging.

[0003] After acquiring the ground-to-air transient electromagnetic detection data, inversion imaging is required to interpret the unknown underground structure and obtain a visual image that is easy to interpret. Since drones have excellent detection efficiency when flying in the field, a large amount of data is collected for each survey mission. At present, the traditional inversion method based on the least squares method has shown relatively accurate and high-resolution inversion results. However, since the solution process requires the use of computing equipment for repeated iterative calculations, the computational complexity and cost will increase rapidly with the increase in data volume. Limited by the available computing resources, the traditional inversion process usually needs to be carried out in the laboratory and may take several days to complete. In order to meet the rapid inversion requirements of field detection missions, a fast and high-precision inversion imaging method is urgently needed.

[0004] Deep learning methods can achieve rapid imaging of the detection area during field missions by pre-training models. However, current research only focuses on deep learning inversion methods for ground-based transient electromagnetic (TEM) and airborne transient electromagnetic (AE). However, these methods do not consider the transmitter and receiver locations. However, for ground-to-air TEM, transmitter and receiver locations are crucial, and integrating this spatial information into the EM response signal remains an unresolved issue.

[0005] U-net is a neural network structure that has been verified to have excellent performance. However, it is difficult for U-net to perform feature fusion of spatial information and electromagnetic response information at the input end. Feature-wise Linear Modulation (FiLM) is an effective modal fusion technology that can achieve feature fusion of spatial information and electromagnetic response information.

[0006] In summary, there is an urgent need to design a method for rapid inversion of ground-to-air transient electromagnetics to efficiently and accurately invert and interpret the collected data. Summary of the Invention

[0007] The application provides a fast ground-air transient electromagnetic inversion method based on a FiLM-U-net, aiming to fuse spatial information of transmitting and receiving positions, improve inversion efficiency and accuracy, and solve the problem of large-scale ground-air transient electromagnetic fast inversion.

[0008] According to an aspect of the application, a fast ground-air transient electromagnetic inversion method based on a FiLM-U-net is provided, including: acquiring geological data and ground-air transient electromagnetic detection data, and performing preprocessing to construct a data set; dividing the data set into a training set, a validation set and a test set; constructing a FiLM-U-net network model, which combines a FiLM structure and a U-net structure and is composed of a feature fusion layer, an encoding layer and a decoding layer, wherein the feature fusion layer is composed of the FiLM structure and is responsible for fusing different feature information; the encoding layer and the decoding layer are composed of a convolutional neural network and jointly constitute the U-net structure; training the FiLM-U-net network model by selecting a suitable training strategy using the training set, and evaluating the inversion effect of the model using the test set; and using the trained FiLM-U-net network model to perform inversion on ground-air transient electromagnetic measured data, and outputting corresponding underground model resistivity values according to the transmitting and receiving positions and the ground-air transient electromagnetic response of the measured data.

[0009] Optionally, the acquiring geological data and ground-air transient electromagnetic detection data, and performing preprocessing to construct a data set includes: acquiring stratum thickness, detection depth, transmitting and receiving positions, sampling time and ground-air transient electromagnetic response of an electromagnetic detection area; dividing the strata of the electromagnetic detection area into multiple levels according to equal thickness intervals, each level corresponding to a depth range; generating corresponding resistivity values for each divided stratum level; performing normalization processing on the transmitting and receiving positions, the ground-air transient electromagnetic response and the resistivity values of each layer; splicing the normalized transmitting and receiving positions and the corresponding ground-air transient electromagnetic response to form a fusion vector as the input of the model, and taking the normalized resistivity of each layer as the output of the model; and one-to-one corresponding each input and output to construct the data set.

[0010] Optionally, the normalization processing on the transmitting and receiving positions, the ground-air transient electromagnetic response and the resistivity values of each layer includes: maximum and minimum value normalization on the transmitting and receiving positions, which is expressed by the formula: ; In the formula, represents the original transmitting and receiving positions; represents the normalized transmitting and receiving positions; represents taking the maximum value; represents taking the minimum value; After the logarithmic transformation of the ground-to-air transient electromagnetic response, the maximum and minimum values ​​are normalized, and the formula is expressed as: ; Where, represents the unnormalized raw electromagnetic response; represents the normalized logarithmic response; Represents the logarithmic function with base 10; After the resistivity value of each layer is logarithmically transformed, the maximum and minimum values ​​are normalized, and the formula is expressed as: ; Where, represents the unnormalized raw resistivity value; Represents the normalized logarithmic resistivity value.

[0011] Optionally, the first part of the FiLM-U-net network model is a position feature fusion module, and the features output by the position feature encoding module are: ; in, is the feature after fusing position information; is the weight adjustment parameter of FiLM, , The mapping relationship between the weight adjustment parameters of FiLM and the position information; is the bias adjustment parameter of FiLM, , is the mapping relationship between the bias adjustment parameters of FiLM and the position information, ; ; is a learnable parameterized activation function that adjusts the weights and biases of each neuron. ; ; ; m is the function type, l is the neural network layer sequence, i is the output neuron sequence, j is the input neuron sequence, w is the scaling factor, is a spline function, is the B-spline basis function, represents the learnable coefficient that controls the weight of the basis function; After the output of the position feature encoding module, it enters the encoding module. First, the input data is subjected to a convolution operation with a kernel size of 7, a stride of 2, and a padding of 3. Then, a downsampling and three combined CNN structure operations are performed simultaneously. Each combined CNN structure consists of a downsampling and three convolution operations. The convolution kernel sizes of these three convolution operations are 1, 3, and 1, the strides are 1, 2, and 1, and the padding is 0, 1, and 0. After each combined CNN structure, a feature is output to the trunk and the bypass at the same time. Then, a 256 and 512 neuron are input. The 256 linear layer further extracts hidden features. feature; finally, the feature size output by the encoding layer is batch×128×8; then it enters the decoding layer, and the feature is first subjected to a convolution operation with a convolution kernel size of 3, a stride of 1, and a padding of 1, and then batch normalization and activation function activation are performed; then three consecutive upsamplings are performed, and each upsampling process is spliced ​​with the output of the combined CNN structure operation in the encoding module to simultaneously decode deep and shallow features and improve the generalization of the model; after three upsamplings, the final output feature size is batch×128×8; then it enters the output layer, and the output layer sets the number of feature channels to 1 and converts the output size to 10 to realize the inversion process.

[0012] Optionally, selecting an appropriate training strategy using the training set to train the FiLM-U-net network model includes: Based on the training set, Adam optimizer is used to select Use the loss function to train the FiLM-U-net network model; Where, is the loss function during model training, is the batch number, The batch size of the input model in each round of training, The output of the model is Batch resistivity value, For the actual Batch resistivity value; Set the initial learning rate to 10 -4 , and adopt a step-by-step attenuation strategy, attenuating the learning rate to 0.2 times the previous learning rate at training rounds of 200, 300, and 400, and setting the total training rounds to 500 rounds and the input batch size to 512; During the training process, the parameters of the model are continuously adjusted so that the loss function When the value of the loss function on the validation set is less than or equal to the first preset threshold, the model is considered to have converged.

[0013] Optionally, using the test set to evaluate the inversion effect of the model includes: The root mean square error (RMSE) is selected to evaluate the FiLM-U-net network model, where RMSE is defined as: ; Where, is the sample serial number, is the sample size of the test set, The output of the model is Group resistivity value, For the actual Group resistivity value; The RMSE is calculated on the test set. When the calculated RMSE is less than or equal to the second preset threshold, the FiLM-U-net network model is considered to be qualified.

[0014] The present invention provides a fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net, which introduces the FiLM structure into the field of geophysical inversion and combines it with U-net, solving the problem of efficient fusion of multi-source information (especially spatial position information) in a deep learning model. At the input end, feature fusion of ground-to-air transient electromagnetic response data and transceiver positions is achieved, and the fused features are further extracted using U-net, so that the U-net can better capture the complex mapping relationship between input information and output information. By fusing the spatial information with the transient electromagnetic response signal, the limitations of CNN and LSTM methods in utilizing spatial information are overcome. At the same time, compared with traditional inversion methods that require several seconds to complete the inversion of a measuring point, the present method can complete the inversion of a measuring line (thousands of points) in a few seconds by relying on a trained model. It has a very fast inversion speed and can achieve effective and rapid inversion of ground-to-air transient electromagnetic signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 This is a flow chart of a fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to an embodiment of the present invention; Figure 2 Schematic diagram of the architecture of the FiLM-U-net model in an embodiment of the present invention; Figure 3 The results of the inversion results and the re-forward modeling responses of the inversion results of the FiLM-U-net inversion method of the present invention and other inversion methods on the test set are compared; Figure 4 This is a comparison of the quasi-2D inversion results of the FiLM-U-net inversion method of the present invention; Figure 4(a) A pseudo-two-dimensional model generated using a simulation program; Figure 4 (b) is the inversion result of FiLM-U-net on the pseudo-two-dimensional model. DETAILED DESCRIPTION

[0016] In order 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 in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only embodiments of a part of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of this application. It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0017] Secondly, the term "one embodiment" or "embodiment" 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 various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0018] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0019] Example 1: Reference Figure 1 , Figure 1 This is a flow chart of a fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: S1, obtain geological data and ground-to-air transient electromagnetic detection data, perform preprocessing, and construct a data set; This step is to build a dataset for training and testing the FiLM-U-net model, including: S11, obtaining the stratum thickness, detection depth, transmitting and receiving positions, sampling time and ground-to-air transient electromagnetic response of the electromagnetic detection area; According to the known geological information of the detection area, the area is preliminarily evaluated to determine the resistivity value distribution interval of the data set, the stratum thickness division grid, the detection depth, the receiving and transmitting position, and the sampling time; The resistivity value distribution interval can be determined according to the existing information, for example, the vertical resistivity curve can be directly obtained through borehole logging, and the resistivity distribution interval of the data set can be determined by analyzing the vertical resistivity curve, or the resistivity interval can be determined by referring to the other electromagnetic detection results in the area; The stratum thickness division grid mainly considers the stratum thickness and the number of strata, and can be freely divided to meet the detection characteristics of the ground-air transient electromagnetic method, that is, the resolution is high in the shallow layer and low in the deep layer, and to ensure fast imaging of the underground under a certain resolution; The detection depth refers to the depth that is expected to be detected in the detection task, and is set according to the detection requirement and the maximum detection depth of the instrument, and is used to determine the layering of the stratum thickness division grid; The receiving and transmitting position refers to the relative position of the transmitting line source and the receiver, and the inversion region is determined by three coordinates, that is, x, y, and z (representing the horizontal coordinate, the vertical coordinate, and the flight height respectively), in the embodiment of the present application, single line inversion is fixed, y and z are fixed, and only x changes. One receiving and transmitting position will only collect one electromagnetic response signal.

[0020] S12, the stratum of the electromagnetic detection area is divided into multiple levels according to the equal thickness interval, and each level corresponds to a depth range; In an example, the stratum thickness is divided into 30, 40, 50, 80, 100, 120, 130, 150, 200, and 300, a total of ten layers; The resistivity range is set to 1-10000Ω·m, which has basically covered the stratum resistivity of the area to be explored; The detection depth is 0-1000m; The receiving and transmitting position is 0-3000m; The sampling time is 1μs-1s, and 63 times are sampled at equal intervals according to the logarithm, and the sampling mode is to take the logarithm of 1e-6 and 1 first, then take 63 numbers at equal intervals in the logarithmic domain, and then take the exponential.

[0021] S13, the corresponding resistivity value of each divided stratum level is generated; According to the above setting, the logarithmic resistivity value of each layer is generated according to the Gaussian random distribution with the mean value of 2 and the variance of 1.5.

[0022] The ground-air transient electromagnetic forward data generation is set according to the determined parameters (the above resistivity range, stratum thickness, detection depth, receiving and transmitting position, and sampling time).

[0023] S14, normalizing the transmitting and receiving positions, the ground-to-air transient electromagnetic response, and the resistivity value of each layer; For sending and receiving positions Perform maximum and minimum value normalization, the normalization formula is: ,in, Indicates the original sending and receiving position; Indicates the normalized sending and receiving position; Indicates taking the maximum value; Indicates taking the minimum value; the maximum and minimum normalization here is based on the determination of the detection range. As above, the x coordinate range is determined to be 0-3000m, and normalization can be performed directly according to the minimum value of 0 and the maximum value of 3000.

[0024] Furthermore, since electromagnetic responses and underground resistivity typically span multiple orders of magnitude, taking the logarithm can compress the dynamic range, making the data distribution more stable and facilitating neural network learning and convergence. Therefore, in this embodiment of the present invention, the (ground-to-air transient electromagnetic) response signal is logarithmized and then normalized to its maximum and minimum values, as expressed by the formula: ,in, represents the unnormalized raw response; represents the normalized logarithmic response; Represents the logarithmic function with base 10; Similarly, after the resistivity value of each layer is logarithmically transformed, the maximum and minimum values ​​are normalized, and the formula is expressed as: ,in, represents the unnormalized raw resistivity value; Represents the normalized logarithmic resistivity value; S15: Concatenate the normalized receiving and transmitting positions with the corresponding ground-to-air transient electromagnetic responses to form a fusion vector as the input of the model, and use the normalized resistivity of each layer as the output of the model; establish a one-to-one correspondence between each input and output to construct a data set.

[0025] Normalized transmit and receive positions The normalized logarithmic response Spliced ​​into a fusion vector as the input of the FiLM-U-net model, the model label is the normalized logarithmic resistivity value of the 10 layers , each such "input-output" pair constitutes a sample in the dataset.

[0026] S2, dividing the data set into a training set, a validation set, and a test set; After the above preprocessing, the large amount of sample data is divided into training set, validation set and test set according to the ratio of 90%, 5% and 5%.

[0027] The training set is used for model parameter learning; the validation set is used for hyperparameter tuning, preventing overfitting, and assisting model selection; the test set is used for final evaluation of the trained model.

[0028] S3, build the FiLM-U-net network model; The FiLM-U-net in the embodiment of the present invention combines the FiLM structure and the U-net structure. The preprocessed transceiver position passes through the FiLM structure and is feature-fused with the preprocessed electromagnetic response, and then enters 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 the 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 layer and the decoding layer. The hyperparameters include the learning rate, the optimizer, and the regularization strategy.

[0029] The FiLM-U-net model consists of a feature fusion layer, an encoding layer, and a decoding layer. The feature fusion layer uses a FiLM (Feature Linear Modulation) structure to fuse different feature information. (Due to the fact that directly splicing the inputs horizontally into the model forces the model to autonomously learn the relationship between the two during subsequent model training, and position information is easily lost in high-dimensional information, feature fusion modifies the electromagnetic response to better conform to the propagation laws of the electromagnetic field, amplifying this nonlinear relationship.) The encoding and decoding layers are primarily composed of convolutional neural networks (CNNs), and together they form a U-net structure. The model inputs are the preprocessed transmit and receive locations and the ground-to-air transient electromagnetic response, and the output is the preprocessed resistivity value of each subsurface model layer. The FiLM-U-net network structure is established based on the input information dimension (spatial information dimension and the number of transient electromagnetic signal samples) and the output information dimension (total number of stratum divisions) of the synthetic data set. In this embodiment, the input layer input vector size of the FiLM-U-net is set to 1×64 (63 sampling bits + 1 receiving and transmitting position) according to the number of sampling time points and stratum divisions selected in S1, and the output layer output vector size is set to 1×10.

[0030] Reference Figure 2 , Figure 2 FiLM-U-net structure diagram according to an embodiment of the present invention is shown in FIG. Figure 2 As shown in Figure 1, the architecture consists of three parts. The feature fusion layer fuses the position information with the response signal and then inputs it into the encoding layer. The encoding layer performs multiple downsampling to obtain 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.

[0031] The input of FiLM-U-net consists of ground-to-air transient electromagnetic data from a single transmitting and receiving location and 63 time sampling points. The final output is the resistivity value of a 10-layer subsurface model. The first part of FiLM-U-net is the position feature fusion module. The features output by the position feature encoding module are: ; in, is the feature after fusing position information; is the weight adjustment parameter of FiLM, , The mapping relationship between the weight adjustment parameters of FiLM and the position information; is the normalized logarithmic response signal without position information; is the bias adjustment parameter of FiLM, , is the mapping relationship between the bias adjustment parameters of FiLM and the position information, ; ; is a learnable parameterized activation function that adjusts the weights and biases of each neuron. ; ; ; m is the function type, l is the neural network layer sequence, i is the output neuron sequence, j is the input neuron sequence, w is the scaling factor, is a spline function, is the B-spline basis function, represents the learnable coefficient that controls the weight of the basis function; Indicates the normalized sending and receiving position; After the output of the position feature encoding module, it enters the encoding module. First, the input data is subjected to a convolution operation with a kernel size of 7, a stride of 2, and a padding of 3; then, a downsampling and three combined CNN structure operations are performed simultaneously. Each combined CNN structure consists of a downsampling and three convolution operations. The convolution kernel sizes of these three convolution operations are 1, 3, and 1, the strides are 1, 2, and 1, and the padding is 0, 1, and 0, respectively. After each combination of the CNN structure, a feature is output to the trunk and the bypass at the same time; then a linear layer with neurons of 256, 512, and 256 is input to further extract hidden features; finally, the feature size output by the encoding layer is batch×128×8.

[0032] Next, in the decoding layer, the features are convolved with a kernel size of 3, a stride of 1, and padding of 1, followed by batch normalization and activation. Three consecutive upsampling operations are then performed, each of which is concatenated with the output of the combined CNN structure in the encoding module to simultaneously decode deep and shallow features, improving model generalization. The final output feature size after three upsampling operations is batch × 128 × 8. Next, in the output layer, the number of feature channels is reduced to 1, and the output size is converted to 10 to facilitate the inversion process.

[0033] S4, using the training set to select an appropriate training strategy to train the FiLM-U-net network model; The training strategy includes the optimizer method, loss function type, learning rate, training cycle, and input batch size. During training, the model parameters are continuously adjusted to gradually reduce the loss function value on the training set. Simultaneously, the model performance is monitored using the validation set to prevent overfitting. For example, if the model's performance on the training set continues to improve but begins to decline on the validation set, this indicates that the model may be overfitting. In this case, adjustments to the training strategy may be necessary, such as stopping training early or adding regularization terms.

[0034] In the embodiment of the present invention, the Adam optimizer is used, and the The FiLM-U-net network model is trained as the loss function; is the batch number, The batch size of the input model in each round of training, The output of the model is Batch resistivity value, For the actual Batch resistivity values, the loss function is used as an indicator to guide the model to train normally until convergence.

[0035] The initial learning rate is set to 10 -4 , and adopt a step-by-step attenuation strategy, attenuating to 0.2 times the previous learning rate at training rounds of 200, 300, and 400; the total training rounds are set to 500 rounds; the input batch size is 512.

[0036] Observe the validation set loss curve. When the loss value on the validation set stabilizes and no longer decreases significantly (that is, the loss value is less than or equal to the first preset threshold, such as 0.01), the model is considered to have converged. Ensure that the model has converged normally and save the model.

[0037] S5, using the test set to evaluate the inversion effect of the model; The test set is used to verify whether the model achieves the expected results on this dataset and to assess its inversion capabilities. 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 capabilities for unknown data. If these metrics meet the expected requirements, the model demonstrates good performance and can be used in actual ground-to-air transient electromagnetic inversion tasks.

[0038] In the embodiment of the present invention, the root mean square error (RMSE) is selected as the error indicator to evaluate the inversion result of FiLM-U-net, where RMSE is defined as: ; Where, is the sample serial number, is the sample size of the test set, The output of the model is Group resistivity value, For the actual Group resistivity value.

[0039] By calculating the RMSE on the test set, when the calculated RMSE is less than or equal to the second preset threshold (generally, the second preset threshold is set to 0.1), it means that the overall prediction error of the model for the formation resistivity is within an acceptable range and meets the training qualification requirements.

[0040] S6, use the trained FiLM-U-net network model to invert the ground-to-air transient electromagnetic measured data.

[0041] By preprocessing the measured data collected in the field (normalization of the transmitting and receiving positions, and logarithmic normalization of the ground-to-air transient electromagnetic response) and inputting it into the trained FiLM-U-net model, the distribution of the underground model resistivity (normalized logarithmic resistivity) can be obtained quickly and accurately, providing an important reference basis for geological exploration, resource development, etc.

[0042] The present invention determines the parameters of the data set based on known geological data in the detection area, generates a large amount of sample data based on the ground-to-air transient electromagnetic forward modeling program, establishes a suitable FiLM-U-net model based on the input and output information dimensions of the synthetic data set, selects a suitable training strategy for model training to obtain the best training results, and inputs the preprocessed field detection 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 the present invention achieves fast and accurate ground-to-air transient electromagnetic data inversion and interpretation through reasonable data set generation and preprocessing, a carefully designed model structure, an appropriate training strategy, and comprehensive evaluation and verification. It also exhibits good inversion performance under complex terrain conditions, providing an effective technical means for fields such as geological exploration.

[0043] Example 2: To verify the effectiveness of the FiLM-U-net model in the present invention in the inversion task, this example randomly selects a set of data from the test set constructed in Example 1, and inverts 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, the underground resistivity models obtained by different inversion methods are forward modeled to obtain forward response curves, and the effect is further evaluated by observing the forward response curves and the root mean square error.

[0044] The inversion results are as follows Figure 3 As shown in the curve on the left. Wherein, the blue line is the inversion effect of the FiLM-U-net model of the present invention under the test set, and the other three colors represent the inversion effects of other inversion methods respectively, and True represents the true value. The inversion effect is evaluated by calculating the inversion root mean square error with the true value. The larger the root mean square error, the worse the inversion result of the model. Among them, the root mean square error of the FiLM-U-net model is 5.26%, the root mean square error of the CNN model is 55.24%, the root mean square error of the MLP model is 53.16%, and the root mean square error of the U-net model is 19.66%. The inversion result of the FiLM-U-net in the present invention is closer to the true underground resistivity.

[0045] The underground resistivity models obtained by different inversion methods are forward modeled and different forward response curves are obtained, such as Figure 3As shown in the right curve, it can be seen that the underground model response obtained by the FiLM-U-net is closest to the true model. According to the mean square error of the response signal and the true signal (the root mean square error of the FiLM-U-net model is 4.01%, the CNN model is 19.37%, the MLP model is 16.52%, and the U-net model is 10.23%), the response signal obtained from the re-interpretation is very close to the true signal.

[0046] In summary, the inversion effect and the comparison results of the forward response curve show that the FiLM-U-net model in the present application has high accuracy in the inversion task.

[0047] Example 3: This embodiment further verifies the accuracy and reliability of the FiLM-U-net in the inversion of the resistivity distribution under complex terrain by using a simulated pseudo two-dimensional complex model.

[0048] A simulated pseudo two-dimensional complex model is used as the true model reference standard in this experiment, and its specific form is shown in Figure 4 (a). The model fully considers the complex terrain factor and can simulate the diversity and complexity of the resistivity distribution in the real underground environment. Based on the generated pseudo two-dimensional complex model, its response signal is obtained, and the response signal of the model is inverted by the FiLM-U-net, and the finally obtained inversion result is shown in Figure 4 (b).

[0049] Comparing the true model Figure 4 (a) with the underground model Figure 4 (b) obtained by the present application, it can be found that the inverted model is almost consistent with the true model in overall shape, distribution position of different resistivity regions, and boundary shape, etc., indicating that the FiLM-U-net model in the present application can accurately invert the resistivity distribution result of the true underground model, and verifies the inversion performance of the present application in the two-dimensional complex underground model.

[0050] In summary, compared with other existing deep learning methods, the present application has better generalization and better inversion accuracy.

[0051] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net, characterized by: include: Acquire geological data and ground-to-air transient electromagnetic detection data, perform preprocessing, and construct a data set; Dividing the dataset into a training set, a validation set, and a test set; Constructing a FiLM-U-net network model. The FiLM-U-net network model combines the FiLM structure and the U-net structure and 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 and together form a U-net structure. The training set is used to select an appropriate training strategy to train the FiLM-U-net network model, and the test set is used to evaluate the inversion effect of the model; The trained FiLM-U-net network model is used to invert the measured ground-to-air transient electromagnetic data, and the corresponding underground model resistivity value is output according to the sending and receiving positions of the measured data and the ground-to-air transient electromagnetic response.

2. The fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to claim 1 is characterized in that: The acquisition of geological data and ground-to-air transient electromagnetic detection data, and preprocessing to construct a data set include: Obtain the stratum thickness, detection depth, transmitting and receiving positions, sampling time and ground-to-air transient electromagnetic response of the electromagnetic detection area; The strata in the electromagnetic detection area are divided into multiple layers according to equal thickness intervals, and each layer corresponds to a depth range; Generate corresponding resistivity values ​​for each divided stratigraphic layer; Normalize the transmitting and receiving positions, the ground-to-air transient electromagnetic response, and the resistivity value of each layer; 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 of the model, and the normalized resistivity of each layer is used as the output of the model. Each input is matched with the output one by one to construct a data set.

3. The fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to claim 2 is characterized in that: Normalization of the transmitting and receiving positions, the ground-to-air transient electromagnetic response, and the resistivity values ​​of each layer includes: The maximum and minimum values ​​of the transmitting and receiving positions are normalized, and the formula is expressed as: ； Where, Indicates the original sending and receiving position; Indicates the normalized sending and receiving position; Indicates taking the maximum value; Indicates taking the minimum value; After the logarithmic transformation of the ground-to-air transient electromagnetic response, the maximum and minimum values ​​are normalized, and the formula is expressed as: ； Where, represents the unnormalized raw electromagnetic response; represents the normalized logarithmic response; Represents the logarithmic function with base 10; After the resistivity value of each layer is logarithmically transformed, the maximum and minimum values ​​are normalized, and the formula is expressed as: ； Where, represents the unnormalized raw resistivity value; Represents the normalized logarithmic resistivity value.

4. The fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to claim 3 is characterized in that: The first part of the FiLM-U-net network model is the position feature fusion module. The features output by the position feature encoding module are: ； in, is the feature after fusing position information; is the weight adjustment parameter of FiLM, , The mapping relationship between the weight adjustment parameters of FiLM and the position information; is the bias adjustment parameter of FiLM, , is the mapping relationship between the bias adjustment parameters of FiLM and the position information, ； ； is a learnable parameterized activation function that adjusts the weights and biases of each neuron. ； ； ； m is the function type, l is the neural network layer sequence, i is the output neuron sequence, j is the input neuron sequence, w is the scaling factor, is a spline function, is the B-spline basis function, represents the learnable coefficient that controls the weight of the basis function; After the output of the position feature encoding module, it enters the encoding module. First, the input data is subjected to a convolution operation with a kernel size of 7, a stride of 2, and a padding of 3. Then, a downsampling and three combined CNN structure operations are performed simultaneously. Each combined CNN structure consists of a downsampling and three convolution operations. The convolution kernel sizes of these three convolution operations are 1, 3, and 1, the strides are 1, 2, and 1, and the padding is 0, 1, and 0. After each combined CNN structure, a feature is output to the trunk and the bypass at the same time. Then, a linear layer with 256 and 512 neurons is input. The hidden features are further extracted by the 256 linear layer. Finally, the feature size output by the encoding layer is batch×128×8. Then it enters the decoding layer, and the features are first convolved with a kernel size of 3, a stride of 1, and a padding of 1, and then batch normalization and activation function activation are performed; then three consecutive upsamplings are performed, and each upsampling process is spliced ​​with the output of the combined CNN structure operation in the encoding module to simultaneously decode deep and shallow features and improve the generalization of the model; after three upsamplings, the final output feature size is batch×128×8; then it enters the output layer, and the output layer sets the number of feature channels to 1 and converts the output size to 10 to realize the inversion process.

5. The fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to claim 1 is characterized in that: Selecting an appropriate training strategy using the training set to train the FiLM-U-net network model includes: Based on the training set, Adam optimizer is used to select Use the loss function to train the FiLM-U-net network model; Where, is the loss function during model training, is the batch number, The batch size of the input model in each round of training, The output of the model is Normalized resistivity value of the batch, For the actual Normalized resistivity value of the batch; Set the initial learning rate to 10 -4 , and adopt a step-by-step attenuation strategy, attenuating the learning rate to 0.2 times the previous learning rate at training rounds of 200, 300, and 400, and setting the total training rounds to 500 rounds and the input batch size to 512; During the training process, the parameters of the model are continuously adjusted so that the loss function When the value of the loss function on the validation set is less than or equal to the first preset threshold, the model is considered to have converged.

6. The fast ground-to-air transient electromagnetic inversion method based on FiLM-U-net according to claim 1, characterized in that: The inversion effect of the model evaluated using the test set includes: The root mean square error (RMSE) is selected to evaluate the FiLM-U-net network model, where RMSE is defined as: ； Where, is the sample serial number, is the number of samples in the test set, The output of the model is Group resistivity value, For the actual Group resistivity value; The RMSE is calculated on the test set. When the calculated RMSE is less than or equal to the second preset threshold, the FiLM-U-net network model is considered to be qualified.

Images