A two-dimensional magnetotelluric inversion method and system based on deep learning

Abstract

The invention discloses a two-dimensional magnetotelluric inversion method and system based on deep learning, which relates to the field of magnetotelluric inversion technology, including forward modeling to obtain response data for magnetotelluric data to be inverted; constructing an improved U-Net network model, replacing the encoder of the original U-net network model with a deformable convolution and a multi-scale residual module, and replacing the upsampling of the decoder of the original U-net network model with a deformable convolution and a bilinear interpolation method; based on a magnetotelluric data set, training the improved U-Net network model to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion; inputting the response data obtained by magnetotelluric forward modeling into the two-dimensional inversion model to obtain the resistivity parameters of the two-dimensional inversion model. The invention is based on a deformable convolution and multi-scale residual module neural network, so as to effectively improve the efficiency of magnetotelluric inversion.

Description

Technical Field

[0001] The present invention relates to the field of magnetotelluric inversion technology, and in particular to a two-dimensional magnetotelluric inversion method and system based on deep learning. Background Art

[0002] Magnetotelluric inversion is a method that uses electric field observation data on the earth to infer the underground medium model. This method is not affected by high-resistance shielding and has the advantages of low cost, convenient construction, and a wide range of exploration depths. It has now become one of the mainstream electromagnetic exploration methods and is widely used in many fields such as mineral resource exploration, electrical structure research, oil and gas energy exploration, and deep earth structure research.

[0003] In recent years, deep learning has been widely used in the field of inverse problems of partial differential equations. Compared with traditional numerical methods, deep learning-based methods for solving inverse problems of partial differential equations have many advantages, such as being more robust to noise, less dependent on initial conditions and boundary conditions, and parallelizable training. However, when dealing with certain complex geological structures, traditional inversion methods currently have difficulty in achieving end-to-end mapping from observation data to underground electrical structures, that is, it is difficult to directly invert the distribution of underground electrical structures from surface observation data, which in turn affects the accuracy of magnetotelluric inversion results. Summary of the invention

[0004] The present invention provides a two-dimensional magnetotelluric inversion method and system based on deep learning, which is used to solve the above-mentioned problem existing in the prior art, that is, the problem of poor accuracy of magnetotelluric inversion in the prior art. The present invention provides a two-dimensional magnetotelluric inversion method based on deep learning, and the method comprises:

[0005] Treat the inverted magnetotelluric data and perform forward modeling to obtain the response data;

[0006] Construct an improved U-Net network model, replace the encoder of the original U-Net network model with a deformable convolution and a multi-scale residual module, and replace the upsampling of the decoder of the original U-Net network model with a deformable convolution and bilinear interpolation method;

[0007] Based on magnetotelluric data, the improved U-Net network model is trained to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion.

[0008] The response data obtained by magnetotelluric forward modeling is input into the two-dimensional inversion model, and the image features of the response data at different scales are extracted by deformable convolution to obtain the feature map. Then, the feature map is transferred to the multi-scale residual block by maximum pooling to obtain the multi-scale feature map with reduced size. The enlarged multi-scale feature map is obtained by adopting deformable convolution and bilinear interpolation method, and the feature map is predicted through the output layer to obtain the resistivity parameters of the two-dimensional inversion model.

[0009] Optionally, the forward modeling of the magnetotelluric data to be inverted to obtain response data specifically includes:

[0010] The Gaussian smoothing operator is used to smooth the magnetotelluric data to be inverted, and the magnetotelluric data are obtained through forward modeling.

[0011] Optionally, the smoothing process of the magnetotelluric data to be inverted using a Gaussian smoothing operator specifically includes:

[0012] The two-dimensional Gaussian operator is obtained using the following formula:

[0013]

[0014] Among them, G 2D (x, y; s) is a two-dimensional Gaussian operator, s is the standard deviation, and x and y represent the two directions of the two-dimensional Gaussian function respectively;

[0015] The two-dimensional Gaussian operator is discretized, and the discrete Gaussian function value is used as the coefficient of the Gaussian kernel. The single coefficient value in the two-dimensional Gaussian kernel is obtained by the following formula:

[0016]

[0017] Among them, Gauss_kernel i,j is a single coefficient value in the two-dimensional Gaussian kernel, i and j represent the horizontal and vertical coordinates in the two-dimensional grid respectively, and s and k are constants.

[0018] Optionally, the Gaussian noise is added in a range of 1%, 3% and 5%.

[0019] The present invention provides a two-dimensional magnetotelluric inversion system based on deep learning, comprising:

[0020] An acquisition module is used to perform forward modeling on the inverted magnetotelluric data to obtain response data;

[0021] A construction module is used to construct an improved U-Net network model, replacing the encoder of the original U-Net network model with a deformable convolution and a multi-scale residual module, and replacing the upsampling of the decoder of the original U-Net network model with a deformable convolution and a bilinear interpolation method;

[0022] A training module, used for training the improved U-Net network model based on magnetotelluric data to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion;

[0023] The determination module is used to input the response data obtained by magnetotelluric forward modeling into the two-dimensional inversion model, extract the image features of the response data at different scales through deformable convolution to obtain a feature map, and then perform maximum pooling to transfer the feature map to the multi-scale residual block to obtain a multi-scale feature map of reduced size; by adopting deformable convolution and bilinear interpolation methods, the enlarged multi-scale feature map is obtained, and the feature map is predicted through the output layer to obtain the resistivity parameters of the two-dimensional inversion model.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention provides a two-dimensional magnetotelluric inversion method based on deep learning. The method replaces the encoder of the original U-net network model with a deformable convolution and a multi-scale residual module, and replaces the upsampling of the decoder of the original U-net network model with a deformable convolution and a bilinear interpolation method to construct an improved U-Net network model, which can directly invert the distribution of underground electrical structures from surface observation data, thereby enhancing the inversion effect of the network; then, according to the magnetotelluric data set, the improved U-Net network model is trained, and when processing complex geological structures, a two-dimensional inversion model that is more in line with the actual situation of field detection can be obtained. In addition, the inverted magnetotelluric data is forward modeled to obtain response data, thereby enhancing the authenticity of the inversion training samples; finally, the response data obtained by the magnetotelluric forward modeling is input into the two-dimensional inversion model for data processing to obtain the resistivity parameters of the two-dimensional inversion model, thereby obtaining a more accurate magnetotelluric inversion result. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0026] Figure 1 A flowchart of a two-dimensional magnetotelluric inversion method based on deep learning provided in an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of the effect of using a Gaussian smoothing operator to process a data set sample provided by an embodiment of the present invention;

[0028] Figure 3 A schematic diagram of the structure of an inversion network model of a two-dimensional magnetotelluric inversion method based on deep learning provided in an embodiment of the present invention;

[0029] Figure 4A network model training loss curve diagram of a two-dimensional magnetotelluric inversion method based on deep learning provided in an embodiment of the present invention;

[0030] Figure 5 An inversion effect diagram of a two-dimensional magnetotelluric inversion method based on deep learning provided in an embodiment of the present invention with 1%, 3% and 5% noise added;

[0031] Figure 6 A smoothed inversion result diagram of a two-dimensional magnetotelluric inversion method based on deep learning provided in an embodiment of the present invention.

[0032] The above drawings have shown clear embodiments of the present invention, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present invention in any way, but to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0035] Figure 1 is a flow chart of a two-dimensional magnetotelluric inversion method based on deep learning provided by an embodiment of the present invention, such as Figure 1 As shown, this embodiment shows a two-dimensional magnetotelluric inversion method based on deep learning, including:

[0036] S101: Perform forward modeling on the inverted magnetotelluric data to obtain response data.

[0037] Optionally, anomalies of different resistivities, sizes and shapes are predefined, and disturbances are performed based on the resistivity to determine resistivity models with different anomaly characteristics, thereby obtaining a magnetotelluric data set to be inverted.

[0038] Exemplarily, the shape of the anomaly may be, for example, a rectangular anomaly, a trapezoidal anomaly, or a rhombus anomaly; wherein the resistivity of the anomaly may range from, for example, 10Ωm to 10000Ωm.

[0039] Optionally, a Gaussian smoothing operator is used to smooth the magnetotelluric data to be inverted, and response data is obtained by forward modeling; wherein the response data are apparent resistivity and phase data of TE and TM polarization modes.

[0040] The inverted magnetotelluric data are smoothed by using a Gaussian smoothing operator, including:

[0041] The two-dimensional Gaussian operator is obtained using the following formula:

[0042]

[0043] Among them, G 2D (x, y; s) is a two-dimensional Gaussian operator, s is the standard deviation, and x and y represent the two directions of the two-dimensional Gaussian function respectively.

[0044] Then the two-dimensional Gaussian operator is discretized, and the discrete Gaussian function value is used as the coefficient of the Gaussian kernel. The specific calculation formula of a single coefficient value in the two-dimensional Gaussian kernel is as follows:

[0045]

[0046] Among them, Gauss_kernel i,j is a single coefficient value in the two-dimensional Gaussian kernel, i and j represent the horizontal and vertical coordinates in the two-dimensional grid respectively, and s and k are constants.

[0047] S102: construct an improved U-Net network model, replace the encoder of the original U-net network model with a deformable convolution and a multi-scale residual module, and replace the upsampling of the decoder of the original U-net network model with a deformable convolution and a bilinear interpolation method.

[0048] Exemplarily, the U-net network model is used as the design basis, and the network coding layer is set to four layers. In the encoding stage, the magnetotelluric forward response data is first used as input, and the features are extracted through 3*3 deformable convolution to obtain the feature map; then the feature map is transferred to the improved multi-scale residual block through 2*2 maximum pooling operation, so that the image size becomes half of the original size; then the residual learning and maximum pooling operations are performed alternately until the feature map is downsampled to a size of 2*2; finally, in the decoding stage, 1*1 deformable convolution and bilinear interpolation with doubled magnification factor are used as the upsampling method to double the image size, and residual learning and bilinear interpolation are performed alternately until the feature map is sampled to the resistivity model size of 32*32 (the size of the grid in the study area). In each downsampling or upsampling process, overfitting during network training can be prevented by adding random dropout layers of varying degrees. In the two-dimensional inversion model, the long connection of the basic U-Net is still used to integrate different feature maps in the encoding and decoding stages. The input channel k is set to 4, and the model input is the apparent resistivity and phase of the TE and TM polarization modes. The network structure is set as follows: Figure 3 shown.

[0049] S103: Based on the magnetotelluric data, the improved U-Net network model is trained to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion.

[0050] Exemplarily, the Adam optimization algorithm can be used for training, wherein the learning rate can be set to the segmented type, the initial learning rate is set to 0.001, the learning rate reduction factor is set to 0.5, the learning rate reduction cycle is set to 10, the shuffle method is set to every-epoch, and the L2 regularization factor is set to 0.005.

[0051] Optionally, MSE is used as the loss function to obtain the training status of the model. The specific calculation formula is:

[0052]

[0053] Among them, N is the total model parameter, is the model output value, m i is the sample value.

[0054] The training status of the model can be observed according to the decrease of the loss curve. The specific training curve is as follows: Figure 4 shown.

[0055] S104: Input the response data obtained by magnetotelluric forward modeling into the two-dimensional inversion model, extract the image features of the response data at different scales through deformable convolution to obtain a feature map, and then perform maximum pooling to transfer the feature map to the multi-scale residual block to obtain a multi-scale feature map of reduced size; obtain the enlarged multi-scale feature map by adopting deformable convolution and bilinear interpolation method, and predict the feature map through the output layer to obtain the resistivity parameters of the two-dimensional inversion model.

[0056] Since there is a small amount of noise in the measured data in actual exploration, in order to improve the accuracy of the inversion results, Gaussian noise of different sizes is added to the model data.

[0057] Optionally, the added range of Gaussian noise may be 1%, 3%, and 5%.

[0058] In the experiment, the inversion effects of adding 1%, 3% and 5% noise are shown in Figure 2. Figure 5 As shown, a1-a6 are theoretical models, b1-b6 are noise-free inversion results, c1-c6 are inversion results with 1% noise increase, d1-d6 are inversion results with 3% noise increase, and e1-e6 are inversion results with 5% noise increase. In actual geological environments, anomalies are generally irregular in shape, so we invert the smoothed anomalies. In the inversion, we compared the inversion effects of different input modes, using TE mode, TM mode, and TE combined with TM mode as model input for inversion.

[0059] like Figure 6 As shown, Figure 6 a1, a2 and a3 are the anomaly models after smoothing at different depths; b1, b2 and b3 are the inversion results of the anomaly in TE mode after smoothing at different depths; c1, c2 and c3 are the inversion results of the anomaly in TM mode after smoothing at different depths; d1, d2 and d3 are the inversion results of the anomaly in TE combined with TM mode after smoothing at different depths; in order to more intuitively reflect the inversion effect, e1, e2 and e3 are the fitting effects of the real model and the inversion results in different modes, respectively. It can be seen that the inversion effect in the TE combined with TM mode is better.

[0060] The above is a two-dimensional magnetotelluric inversion method provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding two-dimensional magnetotelluric inversion system based on deep learning, including:

[0061] An acquisition module is used to perform forward modeling on the inverted magnetotelluric data to obtain response data;

[0062] A construction module is used to construct an improved U-Net network model, replacing the encoder of the original U-Net network model with a deformable convolution and a multi-scale residual module, and replacing the upsampling of the decoder of the original U-Net network model with a deformable convolution and a bilinear interpolation method;

[0063] A training module, used for training the improved U-Net network model based on magnetotelluric data to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion;

[0064] The determination module is used to input the response data obtained by magnetotelluric forward modeling into the two-dimensional inversion model, extract the image features of the response data at different scales through deformable convolution to obtain a feature map, and then perform maximum pooling to transfer the feature map to the multi-scale residual block to obtain a multi-scale feature map of reduced size; by adopting deformable convolution and bilinear interpolation methods, the enlarged multi-scale feature map is obtained, and the feature map is predicted through the output layer to obtain the resistivity parameters of the two-dimensional inversion model.

[0065] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0066] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A two-dimensional magnetotelluric inversion method based on deep learning, characterized in that: include: Treat the inverted magnetotelluric data and perform forward modeling to obtain the response data; Construct an improved U-Net network model, replace the encoder of the original U-Net network model with a deformable convolution and a multi-scale residual module, and replace the upsampling of the decoder of the original U-Net network model with a deformable convolution and bilinear interpolation method; Based on magnetotelluric data, the improved U-Net network model is trained to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion. The response data obtained by magnetotelluric forward modeling is input into the two-dimensional inversion model, and the image features of the response data at different scales are extracted by deformable convolution to obtain the feature map. Then, the feature map is transferred to the multi-scale residual block by maximum pooling to obtain the multi-scale feature map with reduced size. The enlarged multi-scale feature map is obtained by adopting deformable convolution and bilinear interpolation method, and the feature map is predicted through the output layer to obtain the resistivity parameters of the two-dimensional inversion model.

2. The two-dimensional magnetotelluric inversion method based on deep learning as claimed in claim 1, characterized in that: The method of forward modeling the inverted magnetotelluric data to obtain response data specifically includes: The Gaussian smoothing operator is used to smooth the magnetotelluric data to be inverted, and the magnetotelluric data are obtained through forward modeling.

3. The two-dimensional magnetotelluric inversion method based on deep learning as claimed in claim 2, characterized in that: The Gaussian smoothing operator is used to smooth the magnetotelluric data to be inverted, specifically including: The two-dimensional Gaussian operator is obtained using the following formula: Among them, G 2D (x, y; s) is a two-dimensional Gaussian operator, s is the standard deviation, and x and y represent the two directions of the two-dimensional Gaussian function respectively; The two-dimensional Gaussian operator is discretized, and the discrete Gaussian function value is used as the coefficient of the Gaussian kernel. The single coefficient value in the two-dimensional Gaussian kernel is obtained by the following formula: Among them, Gauss_kernel i,j is a single coefficient value in the two-dimensional Gaussian kernel, i and j represent the horizontal and vertical coordinates in the two-dimensional grid respectively, and s and k are constants.

4. A two-dimensional magnetotelluric inversion system based on deep learning, characterized in that: include: An acquisition module is used to perform forward modeling on the inverted magnetotelluric data to obtain response data; A construction module is used to construct an improved U-Net network model, replacing the encoder of the original U-Net network model with a deformable convolution and a multi-scale residual module, and replacing the upsampling of the decoder of the original U-Net network model with a deformable convolution and a bilinear interpolation method; A training module, used for training the improved U-Net network model based on magnetotelluric data to obtain a two-dimensional inversion model for two-dimensional magnetotelluric inversion; The determination module is used to input the response data obtained by magnetotelluric forward modeling into the two-dimensional inversion model, extract the image features of the response data at different scales through deformable convolution to obtain a feature map, and then perform maximum pooling to transfer the feature map to the multi-scale residual block to obtain a multi-scale feature map of reduced size; by adopting deformable convolution and bilinear interpolation methods, the enlarged multi-scale feature map is obtained, and the feature map is predicted through the output layer to obtain the resistivity parameters of the two-dimensional inversion model.

Images