Method and system for mt deep learning cross-scale inversion based on physical simulation criteria

Abstract

The application discloses a MT deep learning cross-scale inversion method and system based on physical simulation criteria, which is based on electromagnetic physical characteristics to deduce the similarity criteria followed by electromagnetic physical simulation, and creatively introduces the similarity criteria into the MT deep learning inversion technology, realizes the data equivalent switching of the training space and the real space, and promotes the network trained by the MT data set of a certain frequency and scale to be applicable to the MT data of other scenes in the same field, greatly enhances the generalization ability of the network, and saves the training time of the network. Specifically, the similarity criteria are used to switch the spatial physical parameters of the real space to the training space, and then based on the switched spatial physical parameters, the MT forward response data of the real space is converted into the forward response data of the training space and input into the MT inversion model based on the neural network to obtain the inversion result, and finally, the similarity criteria are used again to switch the obtained inversion result to the real space.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of exploration geophysics, and in particular to an MT deep learning cross-scale inversion method and system based on physical simulation criteria. BACKGROUND

[0002] Magnetotelluric (MT) is a frequency domain natural field source electromagnetic sounding method, which has the advantages of low cost, high data acquisition efficiency, large exploration depth, etc., and plays an important role in oil and gas, hydrology, geothermal, metal mineral resources exploration and deep electrical structure research.

[0003] In fact, the response of MT observation data to underground electrical structure has the characteristics of high nonlinearity and non-uniqueness, which is difficult to infer complex geological structure by direct interpretation, and the inversion result obtained by processing the observation data by inversion technology can be better used for data interpretation. At present, the mainstream MT inversion methods such as Occam method, quasi-Newton method, conjugate gradient method and nonlinear conjugate gradient method all belong to local optimization algorithm, which is easily affected by initial model and regularization method, and has large calculation amount and required memory. Deep learning technology provides a new way to solve the MT inversion problem. The method can realize fast inversion without initial model assumption through the global approximation ability of deep neural network to the complex nonlinear relationship between data and model, and significantly improves the calculation efficiency. However, there are still a series of problems in the application of deep learning technology in actual MT exploration, one of which is the limited generalization ability of the network. Deep learning usually establishes an end-to-end mapping through training, but MT exploration usually designs different scale lines, measuring points and acquisition frequencies according to the exploration depth and accuracy requirements. For MT data of different scales and frequencies, a network of corresponding scale needs to be trained, which is undoubtedly a very time-consuming work whether to establish a training set or to train it. This further makes the deep learning inversion method difficult to be popularized and practical, which seriously restricts its popularization space. SUMMARY

[0004] The purpose of the present application is to overcome the technical obstacle of limited network generalization ability of MT inversion technology based on deep learning, and to provide an MT deep learning cross-scale inversion method and system based on physical simulation criteria. The method enables the network trained on a certain scale and frequency of MT data set to be applicable to the inversion of MT data of other measurement frequencies, enhances the generalization ability of the network, and helps to improve the practicality of deep learning in the field of MT inversion.

[0005] To this end, the present application provides the following technical solutions:

[0006] In one aspect, the application provides a MT deep learning cross-scale inversion method based on physical simulation criteria, comprising the following steps:

[0007] Step 1: constructing a similarity criterion between a training space and a real space based on electromagnetic physical properties, wherein the training space is a sampling space in which forward response data used by a MT inversion model based on a neural network in a training stage is located; and the real space is a sampling space in which forward response data to be inverted is located;

[0008] Step 2: using the similarity criterion, first switching the spatial physical parameters of the real space to the training space, and then converting the MT forward response data corresponding to the frequency points of the real space into MT forward response data corresponding to the frequency points of the training space;

[0009] The frequency points refer to discrete frequency points selected in a working frequency range; the MT forward response data corresponding to the frequency points of the real space refer to electromagnetic responses calculated for the discrete frequency points under the conditions of the physical parameters of the real space; and the MT forward response data corresponding to the frequency points of the training space refer to electromagnetic responses calculated for the discrete frequency points under the conditions of the physical parameters of the training space;

[0010] Step 3: using the MT inversion model based on the neural network trained by the training space data to invert the MT forward response data corresponding to the frequency points of the training space after switching, to obtain a geoelectric model;

[0011] Step 4: using the similarity criterion again to switch the model scale of the obtained geoelectric model to the model scale of the real space. For example, for the resistivity model inversion result obtained in the training space, the model scale thereof is switched to the model scale of the real space. , to describe the inversion model corresponding to the MT data of the real space. The model scale refers to the size of the model, and also includes the size of the grid division.

[0012] Preferably, the spatial physical parameters representing the sampling space include magnetic permeability, working frequency, model conductivity and model scale; and the conversion formula of the similarity criterion is:

[0013]

[0014] In the formula, the ratio of the model scales is , , , and are the magnetic permeability, working frequency, model conductivity and model scale of the real space, respectively. , , and ​respectively, are the magnetic permeability of the training space, the working frequency, the conductivity of the model and the scale of the model.

[0015] Preferably, after converting the working frequency of the real space into the working frequency of the training space by using the similarity criterion in step 2, the following is further performed:

[0016] determining whether the frequency width of the converted forward response data meets the frequency width of the training data;

[0017] wherein, if the converted forward response data has a frequency missing or the frequency width is less than the frequency width of the training data, a mask needs to be added to the converted data to make the frequency width flush with the training frequency width;

[0018] if the frequency width of the converted forward response data exceeds the frequency width of the training data, a sliding window type data segmentation is performed, and the data after segmentation with the frequency width equal to the frequency width of the training data is taken as a frequency unit;

[0019] for each frequency unit, the MT forward response data corresponding to the real space frequency point is converted into the MT forward response data corresponding to the training space frequency point with the frequency range of the training space as the standard, so as to perform the network inversion in step 3, and the frequency width of the sliding window is equal to the frequency width of the training data;

[0020] if the forward response data is segmented by the sliding window type, step 3 integrates the geoelectric model obtained by the inversion of each frequency unit data to obtain an integrated geoelectric model.

[0021] It should be understood that during the data acquisition process, part of the frequency band data is often distorted due to factors such as dead frequency band and low frequency electromagnetic field signal. The processing method for distorted data is often to exclude it, and part of the frequency of the MT data is missing, which makes the application of the training network face greater technical obstacles. The technical scheme of the present application effectively overcomes the above technical obstacles by the above technical means. It should be noted that the frequency width refers to the span of the frequency range on the logarithmic scale, for example: 0.001~100 Hz, and the frequency width is 5 orders of magnitude.

[0022] Preferably, the process of integrating the geoelectric model obtained by the network inversion of each frequency unit data sets a fitting objective function and solves to obtain the optimal geoelectric model parameters, and the fitting objective function includes a data fitting term, a smoothing constraint term and a trend consistency term.

[0023] Preferably, if the geoelectric model parameter is resistivity, the fitting objective function is represented as:

[0024]

[0025] In the formula, is the resistivity of the i-th layer of the integrated model, is the resistivity of the i-th layer of the integrated model, is the number of output models of different scales, , , represents the resistivity of the i-th layer, i+1-th layer and i+2-th layer of the integrated model, represents the resistivity of the i-th layer of the integrated model interpolated by the j-th output model, is the weight corresponding to the j-th model and the i-th layer resistivity, is the resistivity of the i-th layer of the initial model.

[0026] Preferably, the ratio of the model scale is equal to the square root of the ratio of the training data frequency set in the training space to the forward response data frequency in the real space.

[0027] In fact, the working frequency in the field is known, and the training frequency and the model scale are also known, that is, the working frequency is known.

[0028] Preferably, the inference process of the similarity criterion is:

[0029] Under the premise of ignoring displacement current, the similarity criterion formula obeyed by electromagnetic method physical simulation is derived from Maxwell equations: ;

[0030] Based on the same class of material conductivity and permeability, it is inferred that other space physical parameters have a ratio relationship in the training space and the real space;

[0031] When the space physical parameters of the training space and the real space satisfy the similarity criterion, the wave equations satisfied by the magnetic fields of the two spaces will be exactly the same, producing the same electromagnetic anomalies.

[0032] Preferably, for the sampling area, the training sample construction process of the MT inversion model is:

[0033] First, a random one-dimensional geoelectric model with smooth characteristics is generated by using the method of random assignment of specific points and spline interpolation as a training label;

[0034] Then, the spatial parameters of the training space in the sampling area are obtained, and the forward response data is obtained based on the spatial parameters of the training space through forward calculation, and then random noise and designed low frequency band and medium frequency band missing frequencies are added, and 0 is used as a mask to replace the missing frequencies, to construct sample data;

[0035] Finally, based on the sample data and the training label, a neural network is introduced and network training is performed to construct an approximate MT quasi-inversion operator, wherein the forward response data is taken as the network input and the geoelectric model is taken as the network output.

[0036] If the inversion variable is the resistivity model The mathematical expression of the inversion model is as follows:

[0037]

[0038] In the formula, is the trained approximate quasi-inversion operator, is the forward response data related to the magnetic permeability of the training space, the working frequency, the conductivity of the model and the model scale.

[0039] In the second aspect, the technical scheme of the present application provides a system based on the above-mentioned MT deep learning cross-scale inversion method, comprising:

[0040] A similarity criterion construction module is configured to construct a similarity criterion between a training space and a real space based on electromagnetic physical characteristics, wherein the training space is a sampling space in which the forward response data used by the MT inversion model based on the neural network is located in the training stage; and the real space is a sampling space in which the forward response data to be inverted is located.

[0041] A space conversion module is configured to switch the space physical parameters of the real space to the training space first, and then convert the MT forward response data corresponding to the frequency points of the real space into the MT forward response data corresponding to the frequency points of the training space by using the similarity criterion.

[0042] A deep learning inversion module is configured to perform inversion on the MT forward response data corresponding to the frequency points of the training space after switching by using the MT inversion model based on the neural network trained by the training space data, so as to obtain a geoelectric model.

[0043] A scale switching module is configured to switch the model scale of the obtained geoelectric model to the model scale of the real space by using the similarity criterion.

[0044] In the third aspect, the technical scheme of the present application provides a computer device, comprising one or more processors and a memory storing one or more computer programs; wherein the processor invokes the computer program to realize the steps of the MT deep learning cross-scale inversion method based on the physical simulation criterion.

[0045] In the fourth aspect, the technical scheme of the present application provides a computer readable storage medium storing a computer program, wherein the computer program is invoked by a processor to realize the steps of the MT deep learning cross-scale inversion method based on the physical simulation criterion.

[0046] Beneficial effects

[0047] (1) The electromagnetic physical simulation method is used to derive the similarity criterion followed by the electromagnetic physical simulation, and the similarity criterion is creatively introduced into the neural network in the MT inversion technology, so that the data equivalent switching between the training space and the real space is realized, the network trained by the MT data set at a certain frequency and scale can also be applied to the MT data in other scenes in the same field, the generalization ability of the network is greatly enhanced, and the training time of the network is saved.

[0048] (2) The application further preferably generates a random one-dimensional geoelectric model with smooth characteristics by using the specific point random assignment and spline interpolation method in the deep learning process, so that the complexity of the generated model can be controlled, and the needs of different resolutions can be adapted. In addition, the MT inversion strategy of the sliding window type is proposed to cope with the case that the frequency width of the forward response data in the real space exceeds the training data. In addition, by adding noise to the training data and considering the case of frequency missing, the robustness of the network is effectively enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A basic flowchart of a MT deep learning cross-scale inversion method based on physical simulation criteria according to the application is provided.

[0050] Figure 2 A medium frequency and low frequency range diagram is provided according to the application.

[0051] Figure 3 A training sample generation scheme diagram is provided according to the application.

[0052] Figure 4 An FCNN network diagram is provided according to the application.

[0053] Figure 5 A loss curve diagram of network training is provided according to the application.

[0054] Figure 6 The inversion results of complete data and frequency missing data and their response diagrams are provided according to the application, (a), (b) and (c) are respectively the inversion model and the apparent resistivity, phase curve fitting of the complete data; (d), (e) and (f) are respectively the inversion model and the apparent resistivity, phase curve fitting of the missing data in the medium frequency band; (g), (h) and (i) are respectively the inversion model and the apparent resistivity, phase curve fitting of the missing data in the low frequency band.

[0055] Figure 7Figures (a), (b) and (c) are respectively the inversion model and the apparent resistivity and phase curve fitting condition of model one (the first layer of the model scale is 1.5m); (d), (e) and (f) are respectively the inversion model and the apparent resistivity and phase curve fitting condition of model two (the first layer of the model scale is 15m); (g), (h) and (i) are respectively the inversion model and the apparent resistivity and phase curve fitting condition of model three (the first layer of the model scale is 150m); (j), (k) and (l) are respectively the inversion model and the apparent resistivity and phase curve fitting condition of model four (the first layer of the model scale is 1500m);

[0056] Figure 8 Figure is a schematic diagram of the sliding window inversion strategy according to the present application;

[0057] Figure 9 Figures (a), (b) are respectively the apparent resistivity schematic diagram of model five and model six, (c), (d) are respectively the phase schematic diagram of model five and model six;

[0058] Figure 10 Figures (a) and (b) are respectively the sliding window inversion result of model five and the integrated model; (c) and (d) are respectively the sliding window inversion result of model six and the integrated model;

[0059] Figure 11 Figure is a schematic diagram of the hardware connection of the computer device. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application. The technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0061] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification and claims and the above-described drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application.

[0063] The present application explores the technical obstacle of limited network generalization ability of the MT inversion technology based on deep learning, that is, when there are different scales of survey lines and frequencies in different scenes, the network trained for MT observation data of different scales and frequencies does not have universality, resulting in that the trained network only has good inversion effect within the training scene, and cannot directly invert the geoelectric structure model of new frequency and scale. Therefore, the technical scheme of the present application provides an MT deep learning cross-scale inversion method based on physical simulation criteria, which explores the similarity criteria followed by electromagnetic physical simulation. When each physical parameter of two spaces satisfies the similarity criteria reasoned by the present application, the wave equations satisfied by the magnetic fields of the two spaces will be exactly the same, and the same electromagnetic anomalies will also be produced. Further, the present application creatively introduces the similarity criteria into deep learning to realize the switching of training space and real space, and solves the problem of limited network universality caused by the difference between training space and real space in space parameters. The MT forward response data referred to by the present application is a relatively broad statement, and the apparent resistivity and phase can be considered as the MT forward response.

[0064] The present application will be described below in conjunction with specific embodiments.

[0065] Embodiment 1

[0066] The MT deep learning cross-scale inversion method based on physical simulation criteria provided by the embodiment of the present application comprises the following steps:

[0067] Step 1: Construct the similarity criteria of training space and real space based on electromagnetic physical properties. The training space is the sampling space where the forward response data used by the MT inversion model based on neural network in the training stage is located; the real space is the sampling space where the forward response data to be inverted is located.

[0068] The space physical parameters representing the sampling space in this embodiment include magnetic permeability, working frequency, model conductivity and model scale. The process of reasoning the similarity criteria of training space and real space based on electromagnetic physical properties by the present application is as follows:

[0069] In the electromagnetic physical simulation experiment, the displacement current is ignored, and the similarity criteria formula followed by electromagnetic physical simulation is derived from Maxwell's equations:

[0070]

[0071] wherein, is the ratio of the model scale, and there are: are the magnetic permeability of the real space, the working frequency, the conductivity of the model and the scale of the model, respectively; are the magnetic permeability of the training space, the working frequency, the conductivity of the model and the scale of the model, respectively; when the respective physical parameters of the two spaces satisfy the above formula, the wave equations satisfied by the magnetic fields of the two spaces will be exactly the same, and the same electromagnetic anomalies will also be produced.

[0072] The ratio of the model scale is the square root of the ratio of the training data frequency set by the training space to the forward response data frequency of the real space. In fact, the working frequency in the field is known, and the training data frequency of the training space and the scale of the model are also known, so the ratio of the model scale between the two spaces and the model scale of the real space can be determined.

[0073] Step 2: Using the similarity criterion, the space physical parameters of the real space are switched to the training space, and the MT forward response data corresponding to the frequency points of the real space are converted to the MT forward response data corresponding to the frequency points of the training space.

[0074] Since the working frequency and the model scale of the real space and the training space are not consistent, the forward response data of the real space cannot be directly input into the network trained in the training space for inversion. Therefore, the technical scheme of the present application proposes to convert the forward response data of the real space to the forward response data of the training space according to the physical simulation criterion, and then use the network trained in the training space to perform inversion on it to obtain the geoelectric model.

[0075] Firstly, the space physical parameters of the real space are converted to the training space, and the conversion process is as follows:

[0076]

[0077] This process shows the frequency and scale change characteristics in the real space. Under the premise of meeting the change relationship, the two spaces will produce the same electromagnetic response anomaly. It should be understood that the technical scheme of the present application takes the same magnetic permeability and conductivity of the two spaces as the premise, so the switching of the two spaces with the same magnetic permeability and conductivity is also applicable to the technical scheme of the present application.

[0078] Then, the MT forward response data corresponding to the frequency points of the real space are converted to the forward response data corresponding to the frequency points of the training space.​​​​​​​

[0079] The frequency point refers to a discrete frequency point selected in a working frequency range; the MT forward response data corresponding to the real space frequency point refers to a theoretical electromagnetic response calculated for the discrete frequency point under the real space physical parameter condition; and the MT forward response data corresponding to the training space frequency point refers to a theoretical electromagnetic response calculated for the discrete frequency point under the training space physical parameter condition.

[0080] It should be understood that the process of switching the real space data to the training space is to switch the space according to the simulation criterion, and the range and scale of the frequency are changed at the same time. Therefore, according to the above criterion, the frequency range of the real space is switched to the frequency range of the training space first, wherein the frequency of the frequency point in the real space is also changed, but the numerical value of the MT forward response data of the frequency point is unchanged; since the frequency point in the training space frequency range is different from the frequency point in the real space, it is necessary to further interpolate the frequency value and the MT forward response data of the frequency point in the real space in the training space to convert the MT forward response data of the frequency point in the training space frequency range.

[0081] For example, the frequency range f1 of the real space is 0.1~10000Hz, and there are 30 frequency points in the frequency range; the frequency range f2 of the training space is 0.001~100Hz, and there are 20 frequency points. Based on the above simulation criterion for switching the frequency range, as long as f1 ÷100, the corresponding model scale is increased by 10 times. When switching, the frequency of the frequency point will of course change, but the number of frequency points will not change, that is, the frequency of the real space data is converted from 0.1~10000Hz 30 frequency points to 0.001~100Hz 30 frequency points, not 0.001~100Hz 20 frequency points in the training space, so it is necessary to interpolate in the training space frequency range 0.001~100Hz to interpolate 30 frequency points into 20 frequency points.

[0082] In this embodiment, the spline interpolation method is adopted, and the general form of spline interpolation is:

[0083]

[0084] So that:

[0085]

[0086] In the formula, is an interpolation function on the frequency interval , which corresponds to the MT forward response data in this application; is any frequency point in the frequency interval , which corresponds to the frequency of the training space, Frequency point The response value on the graph corresponds to the MT forward modeling response data; Frequency point The response value on; They represent the intervals respectively. and interval spline interpolation function at frequency points The first derivative at that point, They represent the intervals respectively. and interval spline interpolation function at frequency points The second derivative at that point, For this frequency range The initial value of the interpolation function; The coefficient of the first-order term affects the starting point of the "slope" of this segment of the curve; These are the coefficients of the second-order term, affecting the change in the "curvature" of this section of the curve; The coefficients of the third-order terms determine the curvature of this curve segment. Within this frequency range, a sparse linear system is formed; solving for all coefficients... , , , Once the coefficients of the cubic polynomial are obtained, the interpolation function can be used. Find the interval arbitrary frequency points within The corresponding response value.

[0087] The above describes the theoretical content of spline interpolation. When applied to this field, it involves converting the MT forward response data corresponding to the frequency points after switching the real space frequency range to the training space frequency range into the forward response data corresponding to the training space frequency points. The specific implementation process is as follows:

[0088] use After switching from the real space to the training space, the operating frequency range is determined into frequency intervals (or divided into multiple frequency intervals). Then, based on the frequency after the real space frequency point switching and the MT response value, the polynomial coefficients of each frequency interval are calculated. , , , ; The operating frequency of the training space Substitute the frequency range interpolation function Get response value This refers to the forward response data corresponding to frequency points within the training space. Specifically, it requires calculating the forward response (MT) data of the frequency points used during training within the operating frequency range of the training space. This means that not only must the number of frequency points be equal, but their positions must also correspond.

[0089] The technical solution of the present application further optimizes the solution. After the real space working frequency is converted into the training space working frequency by using the similarity criterion, it is found that the width of the converted working frequency is greater than or less than the training frequency width, or there is a frequency missing. Therefore, after the frequency conversion, the present application further performs:

[0090] determining whether the frequency width of the converted forward response data conforms to the frequency width of the training data (whether it conforms to the neural network input interface of the training space);

[0091] If the converted forward response data has a frequency missing or its frequency width is less than the training frequency width, a mask needs to be added to the converted data to make its frequency width flush with the training frequency width (which can be realized by the prior art); if the frequency width of the converted forward response data exceeds the training frequency width, a sliding window type data segmentation is performed to divide the data into several frequency units that conform to the training space frequency width, and then each segmented data continues to perform "frequency conversion", and for each frequency unit, the MT forward response data corresponding to the real space frequency point is converted into the MT forward response data corresponding to the training space frequency point. Or in some embodiments, during each frequency conversion, the frequency points within its range are converted and interpolated at the same time. That is, after switching the frequency range of the vacuum space to the training space, the real space is divided into several frequency units according to the training space frequency width at one time, and the MT data of the frequency points in the real space are switched at the same time; then for each frequency unit, the frequency conversion is performed according to the frequency range of the training space, and the MT data of the frequency points are converted at the same time. For example, the data with a frequency of 0.0001-100000Hz is directly divided into 0.0001-10Hz, 0.01-1000Hz, and 1-100000Hz three frequency units, and then the frequency conversion is continued for each frequency data.

[0092] Suppose the frequency range f1 of the real space is 0.01-100000Hz, and the frequency range f2 of the training space is 0.001-100Hz. Aligning with the high frequency, the real space frequency is converted to the training space, which is ÷1000, i.e. 0.00001-100Hz. Obviously, this frequency range is wider, so the data needs to be cut. Taking a sliding step frequency band of 2 orders of magnitude as an example, 0.00001-100Hz can be divided into 0.00001-1Hz and 0.001-100Hz. The frequency width of these two parts conforms to the input frequency width, but the frequency range of 0.00001-1Hz still needs to be converted according to the frequency range f2 (0.001-100Hz) of the training space, and the MT data of the frequency points are converted at the same time, and then the inversion is performed.

[0093] It should be understood that the frequency width of the sliding window is consistent with the training data frequency width, and the frequency width step of the sliding can be adjusted according to actual conditions, but not more than 2 orders of magnitude, such as the step of the sliding is an order of magnitude.

[0094] Step 3: Based on the switched spatial physical parameters and the processed real space forward response data, input the trained neural network MT inversion model to obtain the inversion result, that is, the geoelectric model parameters.

[0095] It should be understood that the cutting according to the width is the preferred embodiment of the application, and other feasible embodiments, reducing the precision requirement, not cutting and multi-model integration are also within the protection scope of the application.

[0096] If the forward response data is divided by the sliding window, the neural network prediction needs to be performed on each divided data segment, and then the obtained inversion result is fitted to obtain an integrated model. The objective function of the fitting includes a data fitting term, a smoothing constraint term and a trend consistency term. This embodiment takes resistivity as an example and is expressed as:

[0097]

[0098] In the formula, is the fitting objective function, is the resistivity layer number of the integrated model, is the number of output models of different scales, , , represents the resistivity of the i-th layer, the i+1-th layer and the i+2-th layer of the integrated model, represents the resistivity of the i-th layer of the integrated model scale interpolated by the j-th output model, is the weight corresponding to the j-th model and the i-th layer resistivity, is the resistivity of the i-th layer of the initial model. In some embodiments, the L-BFGS-B algorithm is used to minimize the objective function to obtain the optimal solution During the optimization process, Generally, the average value of the models of different scales at the same depth is taken, and the resistivity value of each layer of the integrated model is set in the physically acceptable range (for example: [1~10 4], as a boundary constraint. In the theoretical research stage, the MT observation data in real space is actually the data synthesized by the theoretical model, but the model scale and frequency condition of this synthesized data is different from that of the training data. Then the MT observation data in real space is converted into the training space according to the conversion formula of the similarity criterion. In fact, this conversion does not change the specific values of the observation data (only the conversion of the physical simulation criterion does not change the specific values of the observation data), but changes the scale and frequency of the data to be consistent with the training space. However, due to the change of the frequency points of the two spaces, the interpolated observation data will change, but the overall MT response curve in the frequency range will not change. Based on the theoretical statement, it is known that the network trained in the training space can also be used for inversion of real space data after conversion. Therefore, when converting, the magnetic permeability and electrical conductivity of the model do not change, and only the frequency and scale change. When the present application is proved to be effective in theory, the actual observation data is only needed to be replaced by the theoretical observation data in actual use.

[0099] It should be further pointed out that although the network input of the inversion network model is the MT forward response data, the scale and working frequency of the model are not input as variables to the network, but the training process of the inversion network model is based on the batch MT forward response data under the working frequency and scale of the training space, which determines what scale and frequency range of data the inversion network can invert. Therefore, the frequency and scale of the real space need to be switched to the training space, and then the frequency point MT forward response data is input into the inversion network model to obtain the geoelectric model, and then the geoelectric model is switched to the corresponding scale in the real space to obtain the geoelectric model in the real space. That is, the model scale that can actually describe the real space is obtained by multiplying the scale of the training space model by the scale reduction ratio. In actual application, the scale switching is not needed first, and the scale of the model is switched at the end.

[0100] For example: there is a group of frequency points in real space and the MT forward response data thereof , n is the number of frequency points in the real space; a group of frequency points in the training space and the MT forward response data thereof are obtained after the above switching, m is the number of frequency points in the training space.

[0101] The MT forward response data is input into the inversion network model to obtain the geoelectric model under the scale of the training space;

[0102] The model scale is switched to the real space by to obtain the geoelectric model in the real space.

[0103] In summary, this invention fully considers that after transformation using similarity criteria, the two spaces will produce the same electromagnetic anomalies. By transforming and interpolating spatial physical parameters, the data frequency and scale of the real space are transformed to be consistent with those of the training space, so that the network trained in the training space can also be used for the inversion of real space data.

[0104] It should be understood that the embodiments of the present invention do not have specific requirements for neural networks, and any deep learning that can be applied to the field for MT inversion falls within the protection scope of the present invention.

[0105] The following is just an example. Taking the FCNN network as an example, the process of constructing training samples in the training space is as follows:

[0106] For the sampling area, firstly, a stochastic one-dimensional geoelectric model with smoothing properties is generated using specific point random assignment and spline interpolation, serving as training labels. A one-dimensional geoelectric model refers to a one-dimensional geological layered structure model that includes the electrical conductivity (or resistivity) distribution of the subsurface medium. This model is typically a mathematical representation used to describe the electrical properties of the subsurface medium.

[0107] Then, the spatial parameters of the training space within the sampling area are obtained. Based on the spatial parameters of the training space, the apparent resistivity and phase are obtained through forward modeling. Random noise is added and low-frequency and mid-frequency missing frequencies are designed. 0 is used as a mask to replace the missing frequencies and the sample data is constructed.

[0108] Finally, an FCNN network is introduced to train the samples, approximating the MT pseudo-inversion operator, where apparent resistivity and phase are used as network inputs, and a random one-dimensional geoelectric model is used as network output.

[0109] Step 4: Apply the similarity criterion again to switch the inversion results obtained in Step 3 to the real space. Specifically, for the resistivity model inversion results obtained in the training space, scale the model... Transform into , This yields an inversion model that directly describes the real-space MT data.

[0110] Application Examples

[0111] like Figure 3 The training sample generation scheme shown below has the following specific steps:

[0112] 1. The underground area is divided into 20 layers. Resistivity values ​​are randomly assigned to layers 1, 5, 9, 13, 17, and 20. There are two methods for selecting resistivity values: one is to directly select a decimal from the range [0-3] as the logarithmic resistivity value; the other is to first randomly select a value from the range [1-1000] and then take its logarithm. Both methods are used to randomly select values ​​before generating each sample.

[0113] 2. Calculate the resistivity values ​​of the remaining layers using spline interpolation to obtain the resistivity of the smooth model. .

[0114] 3. Subtract the model resistivity obtained in step 2 from the following formula. All values ​​are greater than 0, and the final model resistivity is denoted as... :

[0115]

[0116] 4. In a 20-layer system, the first layer has a thickness of 150m, the intermediate layers increase in thickness in multiples of 1.2 downwards, and the bottom layer is a uniform half-space with infinite thickness. This is used to obtain the model scale vector of the training space. .

[0117] 5. Set the frequency range to 0.001~100Hz, and take 20 frequency points at logarithmic intervals to calculate the corresponding MT forward response, i.e., apparent resistivity and phase. To enhance the robustness of the training network, 1~10% random noise was added to the forward response data, and frequency missing was artificially introduced. This included randomly missing 0~4 frequencies consecutively in the mid-frequency band and replacing them with 0 as a mask, and randomly missing 0~8 frequencies consecutively in the low-frequency band and replacing them with 0 as a mask. The mid-frequency and low-frequency ranges are illustrated below. Figure 2 As shown.

[0118] Figure 4 A schematic diagram of the trained network (FCNN) is shown. Each hidden layer has 100 nodes. During training, the apparent resistivity and phase calculated using MT forward modeling are used as network inputs, and a stochastic one-dimensional geoelectric model is used as the network output. Figure 5 The diagram shows the loss curves during network training. A total of 70,000 samples were used for training, with a training, validation, and test set ratio of 8:1:1. The batch size was set to 64, the learning rate to 0.001, the activation function to ReLU, and the optimizer to Adam. The network underwent 500 iterations. It can be seen that as the number of iterations increases, the MSE values ​​of both the training and validation sets gradually decrease to a very small value, and there is no overfitting during training.

[0119] Figure 6The inversion results and responses of the complete data, medium frequency missing data and low frequency missing data of the same model are shown. It can be seen that even if 4 frequency points are missing, the inversion model has little difference with the result under the condition of complete data, and the stratum structure can be well restored, and the model responses under the three conditions are highly consistent with the theoretical synthetic data. Figure 6 It is shown that the network trained by the technical scheme of the application can still realize effective one-dimensional inversion under the condition of missing frequency within a certain range.

[0120] Table 1 shows four test sets of different scales and frequencies established by the same method as the sample generation scheme, which are named TS1, TS2, TS3 and TS4 respectively, and their frequency ranges almost cover the entire MT observation frequency band. Then, the same trained FCNN is used to perform inversion on these test sets one by one, and the Pearson correlation coefficient is introduced to understand the correlation between the true value and the predicted value of the neural network, and when the correlation coefficient is closer to 1, it indicates that the positive correlation between the true value and the predicted value is stronger. For example, the Pearson correlation degree is used.

[0121] Among them, TS3 uses the same modeling parameters as the training set and can be directly input to the trained network for prediction, while TS1, TS2 and TS4 need to be cross-scale inverted. Taking TS2 inversion as an example, first, TS2 is converted to change its frequency range to the frequency range of the training space, that is, from 10 4 ~10 -1 to 10 2 ~10 -3 , which is reduced by 100 times, and this conversion does not change the specific values of the inversion data. Then, the mask and spline interpolation method is used to align the data frequency points with the training frequency points; since the number of TS2 frequency points is set to be consistent with the training data, its frequency after conversion is already aligned with the training frequency points, so this step is not needed. Finally, the converted TS2 is input into the network to obtain the resistivity model, at this time, the scale of the resistivity model is determined by the frequency conversion multiple and the scale of the training space (i.e. resistivity model scale = × training space scale = 0.1 × training space scale).

[0122] Table 2 shows the average Pearson correlation degree of the same trained network for inversion of different scale and frequency verification data sets. It can be seen from the table that the final inversion result has a high correlation with the true value, and all the correlation degrees are above 98%.

[0123] Table 1

[0124]

[0125] Table 2

[0126]

[0127] One inversion result was randomly selected from each of the four test sets described in Table 1. Figure 7 As shown, the first layer of the model scale for Model 1 is 1.5m, for Model 2 it is 15m, for Model 3 it is 150m, and for Model 4 it is 1500m, corresponding to datasets TS1, TS2, TS3, and TS4. It can be seen that the inverted model response matches the observed data well, and the trends and patterns of the inverted model are consistent with the actual model. Figure 7 This shows that the same training network can effectively invert MT data at different scales and frequencies.

[0128] Figure 8 A schematic diagram of sliding window inversion is shown. The frequency width of the sliding window is consistent with the frequency width of the training data, and the step size of the frequency width can be adjusted according to the actual situation, but not exceeding two orders of magnitude. The objective function for fitting the cross-scale inversion results of the MT data segmented by the sliding window includes a data fitting term, a smoothing constraint term, and a trend consistency term. Figure 10 Two large-scale resistivity catastrophe models and their sliding window inversion results are presented (Model 5 and Model 6). Both models are divided into 7 layers, with alternating resistivity layers of 800 Ω·m and 80 Ω·m. The measurement frequency is set to 10... -5 ~10 6 Hz, a total of 40 frequency points, calculate the corresponding forward modeling data for the two models, such as Figure 9 As shown. Based on the sliding window data segmentation strategy, the frequency width step size is set to two orders of magnitude. The theoretically synthesized data is divided into four parts with frequency ranges of 10... 6 ~10 1 Hz, 10 4 ~10 -1 Hz, 10 2 ~10 -3 Hz, 10 0 ~10 -5 Hz. The segmented forward modeling data is subjected to spline interpolation to align with the training frequency points, and then input into the network for inversion prediction. The inversion result is as follows: Figure 10 (a) Figure 10 As shown in (c), it can be seen that the inversion results of these four parts can accurately reflect the resistivity information at the corresponding depth. Next, I established the objective function for model fitting and combined the piecewise inversion results into an integrated model, as shown in [example missing]. Figure 10 (b) Figure 10The results show that the integrated model is very close to the trend and shape of the true model, which indicates that the sliding window inversion method is effective.

[0129] Embodiment 2

[0130] The embodiment of the present application provides a system based on the above-mentioned MT deep learning cross-scale inversion method, which comprises similarity criterion construction modules, a space conversion module, a deep learning inversion module and a scale switching module connected with each other or in sequence.

[0131] The similarity criterion construction module is used for constructing a similarity criterion between a training space and a true space based on electromagnetic physical characteristics. The training space is a sampling space in which forward response data used by an MT inversion model based on a neural network in a training stage is located. The true space is a sampling space in which forward response data to be inverted is located.

[0132] The space conversion module is used for switching a space physical parameter of the true space to the training space first, and then converting MT forward response data corresponding to a frequency point of the true space into MT forward response data corresponding to a frequency point of the training space by using the similarity criterion. The switched space physical parameter at least includes a model scale and a working frequency.

[0133] The deep learning inversion module is used for inverting the MT forward response data corresponding to the frequency point of the training space after switching by using an MT inversion model based on a neural network trained by training space data, to obtain a geoelectric model.

[0134] The scale switching module is used for switching the model scale of the obtained geoelectric model to the model scale of the true space by using the similarity criterion.

[0135] In some embodiments, a model fitting module is further provided for fitting inversion results of each sliding window segmented data into an integrated model.

[0136] It should also be understood that the specific implementation process of each module can refer to the above method content, and the present application will not be repeated here, and the segmentation of the above functional modules is only for example, in some embodiments, part of the functional modules can be combined, part of the functional modules can be split, and each functional module can be realized in a software manner or in a hardware or software and hardware combination, wherein the software and hardware devices include but are not limited to general electronic terminals, programmable gate arrays, digital signal processors, microprocessors and corresponding programming or burning software thereof.

[0137] Embodiment 3

[0138] The embodiment of the present application provides a computer device, comprising: one or more processors; and a memory storing one or more computer programs; wherein the processor invokes the computer program to implement the steps of the MT deep learning cross-scale inversion method based on a physical simulation criterion.

[0139] Specific implementation:

[0140] Step 1: constructing a similarity criterion between a training space and a real space based on electromagnetic method physical characteristics, the training space being a sampling space in which forward response data used by an MT inversion model based on a neural network in a training stage is located; the real space being a sampling space in which forward response data to be inverted is located;

[0141] Step 2: using the similarity criterion, first switching the spatial physical parameters of the real space to the training space, and then converting the MT forward response data corresponding to the frequency points of the real space into the MT forward response data corresponding to the frequency points of the training space;

[0142] Step 3: using the MT inversion model based on the neural network trained by the training space data to invert the MT forward response data corresponding to the frequency points of the training space after switching, to obtain a geoelectric model;

[0143] Step 4: using the similarity criterion again, switching the model scale of the obtained geoelectric model to the model scale of the real space.

[0144] The specific implementation process of each step can refer to the description of the foregoing embodiment of the MT deep learning cross-scale inversion method based on a physical simulation criterion.

[0145] In some embodiments, as shown in Figure 11 The electronic components of the computer device include:

[0146] The processor 1600 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or can also be any conventional processor. The processor 1600 is configured to execute related programs to implement the technical solutions provided by the embodiments of the present application.

[0147] The memory 1700 can be implemented in the form of a Read Only Memory (ROM), a static storage device, a dynamic storage device, or a Random Access Memory (RAM), etc. The memory 1700 can store an operating system and other application programs, and when the technical solutions provided by the embodiments of the present specification are implemented by software or firmware, the related program codes are stored in the memory 1700 and are called and executed by the processor 1600 to implement the algorithm program of the MT deep learning cross-scale inversion method based on the physical simulation criterion.

[0148] The input / output interface 1800 is configured to realize information input and output.

[0149] The communication interface 1900 is configured to realize the communication interaction between the device and other devices, and the communication can be realized in a wired manner (for example, a USB, a network cable, etc.) or in a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.).

[0150] The bus 2000 is configured to transmit information between various components (for example, the processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900) of the device.

[0151] The processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900 are connected to each other in the device through the bus 2000.

[0152] Embodiment 4

[0153] The embodiment of the present application also provides a computer readable storage medium storing a computer program, and the computer program is called by a processor to implement the steps of the MT deep learning cross-scale inversion method based on the physical simulation criterion.

[0154] Specific implementation:

[0155] Step 1: constructing a similarity criterion between a training space and a real space based on electromagnetic method physical characteristics, the training space is a sampling space in which forward response data used by a neural network-based MT inversion model in a training stage is located; the real space is a sampling space in which forward response data to be inverted is located;

[0156] Step 2: using the similarity criterion, first switching the spatial physical parameters of the real space to the training space, and then converting the MT forward response data corresponding to the frequency points of the real space into the MT forward response data corresponding to the frequency points of the training space;

[0157] Step 3: The MT inversion model based on the neural network trained by the training space data is used to perform inversion on the MT forward response data corresponding to the training space frequency point after switching, to obtain a geoelectric model;

[0158] Step 4: The model scale of the obtained geoelectric model is switched to the model scale of the real space again by using the similarity criterion.

[0159] The specific implementation process of each step can refer to the description of the foregoing MT deep learning cross-scale inversion method based on physical simulation criterion.

[0160] The readable storage medium is a computer readable storage medium, which can be an internal storage unit of the software and hardware device in any of the foregoing embodiments, for example, a hard disk or a memory of the controller. The readable storage medium can also be an external storage device of the controller, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like. Further, the readable storage medium can include both the internal storage unit and the external storage device of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0161] Based on such understanding, the technical solutions of the present application, essentially or in the contribution to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing readable storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0162] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer readable code. The present application is directed to a method, apparatus (system) and computer program product at any one of the method embodiments according to the present application, and any combinations of the method embodiments according to the present application, the processor- implemented process producing the apparatus with the functionality to perform the functions specified in one or more of the flow diagrams and / or block diagrams according to the present application. These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in one or more of the flow diagrams and / or block diagrams according to the present application. The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the flow diagrams and / or block diagrams according to the present application.

[0163] It should be emphasized that the above-described embodiments of the present application are merely illustrative of the present application and are not intended to be limiting thereof. Therefore, the present application is not limited to the embodiments described in the specific embodiments, but any other embodiments derived by those skilled in the art from the technical solutions of the present application, without departing from the spirit and scope of the present application, whether modified or replaced, also belong to the protection scope of the present application.

Claims

1. A method for MT deep learning cross-scale inversion based on physical simulation criteria, characterized in that: The method comprises the following steps: Step 1: constructing a similarity criterion between a training space and a real space based on electromagnetic physical characteristics, wherein the training space is a sampling space in which forward response data used by a neural network-based MT inversion model in a training phase is located; and the real space is a sampling space in which forward response data to be inverted is located; Step 2: using the similarity criterion, first switching the spatial physical parameters of the real space to the training space, and then converting the MT forward response data corresponding to the frequency points of the real space to the MT forward response data corresponding to the frequency points of the training space; wherein it is determined whether the frequency width of the converted forward response data meets the frequency width of the training data; if the converted forward response data has a frequency missing or the frequency width is less than the frequency width of the training data, a mask needs to be added to the converted data to make the frequency width flush with the training frequency width; if the frequency width of the converted forward response data exceeds the frequency width of the training data, a sliding window type data segmentation is performed, and the data after the segmentation and having a frequency width equal to the frequency width of the training data is taken as a frequency unit; for each frequency unit, the MT forward response data corresponding to the frequency points of the real space is converted to the MT forward response data corresponding to the frequency points of the training space with the frequency range of the training space as a standard, so as to perform network inversion in step 3, and the frequency width of the sliding window is equal to the frequency width of the training data; if the forward response data is segmented by the sliding window, step 3 integrates the geoelectric models obtained by the data inversion of each frequency unit to obtain an integrated geoelectric model; Step 3: using the neural network-based MT inversion model trained by the training space data to invert the MT forward response data corresponding to the frequency points of the training space after the switching to obtain a geoelectric model; wherein the process of integrating the geoelectric models obtained by the network inversion of each frequency unit data is set to solve an optimal geoelectric model parameter by setting a fitting objective function, and the fitting objective function comprises a data fitting term, a smoothing constraint term and a trend consistency term; if the geoelectric model parameter is resistivity, the fitting objective function is expressed as: ； where, is the fitting objective function, is the resistivity of the i-th layer of the integrated model, is the number of output models at different scales, , , represents the resistivity of the i-th layer, i+1-th layer, i+2-th layer of the integrated model, represents the resistivity of the i-th layer of the integrated model interpolated by the j-th output model, is the weight corresponding to the j-th model and the i-th layer resistivity, is the resistivity of the i-th layer of the initial model; Step 4: using the similarity criterion again to switch the model scale of the obtained geoelectric model to the model scale of the real space.

2. The MT deep learning cross-scale inversion method of claim 1, wherein: The spatial physical parameters representing the sampling space include magnetic permeability, working frequency, model conductivity and model scale; the conversion formula of the similarity criterion is: ； wherein is the ratio of the model scale, , , and are the magnetic permeability of the real space, the operating frequency, the electrical conductivity of the model and the scale of the model, respectively; , , and are the magnetic permeability of the training space, the operating frequency, the electrical conductivity of the model and the scale of the model, respectively.

3. The method of claim 2, wherein: The ratio of the model scales is equal to the square root of the ratio of the training data frequency set in the training space to the frequency of the forward response data in the real space.

4. The method of claim 2, wherein: The reasoning process of the similarity criterion is: The similarity criterion formulae obeyed by the electromagnetic method physical simulation are derived from the Maxwell equations under the premise of ignoring displacement current: ; Based on the unchanged conductivity and magnetic permeability of the same type of material, it is inferred that the other spatial physical parameters have a ratio relationship in the training space and the real space; wherein when the spatial physical parameters of the training space and the real space meet the similarity criterion, the wave equations of the magnetic fields of the two spaces will be completely the same, and the same electromagnetic anomalies are produced.

5. A system for MT deep learning cross-scale inversion based on the physical analog criterion according to any one of claims 1-4, characterized in that: The method comprises: a similarity criterion construction module, configured to construct a similarity criterion between a training space and a real space based on electromagnetic physical characteristics, the training space being a sampling space in which forward response data used by a neural network-based MT inversion model in a training phase is located, and the real space being a sampling space in which forward response data to be inverted is located; a space conversion module, configured to switch a space physical parameter of the real space to the training space and convert MT forward response data corresponding to a frequency point of the real space to MT forward response data corresponding to a frequency point of the training space by using the similarity criterion; a deep learning inversion module, configured to perform inversion on the MT forward response data corresponding to the frequency point of the training space after switching by using a neural network-based MT inversion model trained by training space data, to obtain a geoelectric model; a scale switching module, configured to switch a model scale of the obtained geoelectric model to a model scale of the real space by using the similarity criterion.

6. A computer device, comprising: comprise: one or more processors; a memory storing one or more computer programs; wherein the processor invokes the computer program to implement: steps of the MT deep learning cross-scale inversion method based on the physical simulation criterion according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that: store a computer program, and the computer program is invoked by a processor to implement: steps of the MT deep learning cross-scale inversion method based on the physical simulation criterion according to any one of claims 1-4.

Images