Decoupled representation learning based multi-stage inversion method for semi-airborne electromagnetic noise data

Abstract

The application discloses a semi-airborne electromagnetic noise data multi-stage inversion method based on decoupling representation learning, and comprises the following steps: obtaining a geoelectric model in DL-RMD to perform forward modeling, and pairing the forward modeling data of each time of forward modeling with the geoelectric model; adding environmental noise to the paired forward modeling data to obtain first simulation data, and pairing the forward modeling data, the geoelectric model and the first simulation data to form a first data set; building a first-stage encoder and a second-stage decoder of a decoupling representation learning network, extracting first data factors in the first data set, and performing inversion by using the first data factors; constructing a loss function; training the decoupling representation learning network to obtain a trained decoupling representation learning network; inputting actually collected noise data into the trained decoupling representation learning network to obtain second data factors, and performing inversion in the second-stage decoder of the trained decoupling representation learning network by using the second data factors.

Description

Technical Field

[0001] The present invention relates to the technical field of semi-aeronautical electromagnetic noise data processing, in particular to a multi-stage inversion method for semi-aeronautical electromagnetic noise data based on decoupling representation learning. Background Art

[0002] Semi-airborne transient electromagnetic (SATEM) is a geophysical exploration technique developed from magnetotellurics. It extracts and measures geoelectrical structure by observing received secondary field signals. By placing the observation system on an aerial vehicle, such as a fixed-wing aircraft or helicopter, it can rapidly cover large, inaccessible areas. Inverting the data directly reveals the subsurface electrical structure. SATEM data inversion involves applying an inversion algorithm to derive the subsurface electrical structure from field-collected data. However, using traditional methods, SATEM data inversion can take days, slowing understanding of subsurface structure and relying heavily on empirical parameter selection. [Research on Three-Dimensional Inversion of Ground and Semi-Airborne Transient Electromagnetic Data from Electrical Sources] Furthermore, complex environmental noise can mask the induced signals, increasing uncertainty and leading to a waste of data resources. Existing deep learning methods cannot handle noisy data and require the use of pre-processed data for inversion, further increasing the workload. For example, Li Siping's "Research on Deep Learning-Based Airborne Electromagnetic Inversion Methods" reports.

[0003] Therefore, it is urgent to propose a logically simple, accurate and reliable multi-stage inversion method for semi-aeronautical electromagnetic noise data based on decoupled representation learning. Summary of the Invention

[0004] In response to the above problems, the present invention aims to provide a multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning. The technical solutions adopted by the present invention are as follows:

[0005] A multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning includes the following steps:

[0006] Obtain the geoelectric model in the known geoelectric model set DL-RMD for forward modeling to obtain forward simulation data; pair the forward simulation data of any forward modeling with the geoelectric model;

[0007] Adding environmental noise to the paired forward simulation data to obtain first simulation data with the added environmental noise, and pairing the forward simulation data, the geoelectric model, and the first simulation data to form a first data set;

[0008] Build the first-stage encoder of the decoupled representation learning network and extract the first data factor in the first data set;

[0009] Build the second-stage decoder of the decoupled representation learning network and use the first data factor for inversion in the second-stage decoder;

[0010] Constructing a loss function; the loss function includes a data loss function and an inversion loss function;

[0011] Using the first data set to train the decoupled representation learning network, and adjusting parameters to obtain a trained decoupled representation learning network;

[0012] The actually collected noise data is input into the trained decoupled representation learning network to obtain a second data factor, and the second data factor is used to perform inversion in the second stage decoder of the trained decoupled representation learning network.

[0013] Furthermore, environmental noise is added to the paired forward simulation data to obtain first simulation data with environmental noise added, including

[0014] Take out the paired forward simulation data one by one and add environmental noise. The expression is:

[0015] d n =d c +N(0,1)

[0016] Among them, d n represents the first simulation data after adding environmental noise; d c represents the forward simulation data without adding environmental noise; N(0,1) represents Gaussian noise.

[0017] Furthermore, the first stage encoder performs data encoding to obtain a first data factor corresponding to the first data set and a second data factor corresponding to the actually collected noise data; the first data factor and the second data factor are both composed of a noise factor and a signal factor.

[0018] Furthermore, the second stage decoder includes a data decoder and an inversion decoder; the data decoder decodes the arithmetic sum of the noise factor and the signal factor to obtain estimated data; the inversion decoder inverts the signal factor to obtain resistivity.

[0019] Furthermore, the first-stage encoder, data decoder, and inversion decoder have the same structure, consisting of an embedding layer and a backbone network; the embedding layer performs a one-dimensional convolution operation on the data obtained from the input layer to obtain embedded data;

[0020] The backbone network includes a first linear layer, a second linear layer, and a multilayer perceptron connected in sequence. The first linear layer obtains the embedded data output by the embedding layer and performs full-connection processing to obtain the value, key, and query value corresponding to the embedded data. Matrix multiplication is performed on the key and query value output by the first linear layer to obtain third data. Matrix multiplication is performed on the value output by the first linear layer and the third data to obtain fourth data, which is input into the second linear layer. The second linear layer outputs fifth data, which is arithmetic summed with the embedded data output by the embedding layer to obtain sixth data. The multilayer perceptron performs a full-connection operation on the sixth data and then activates it using a sigmoid activation function to obtain seventh data. The sixth data and the seventh data are arithmetic summed to obtain the output of the backbone network.

[0021] Furthermore, the loss function L oss , whose expression is:

[0022] L oss =L d +L I

[0023] Among them, L d represents the data loss function; L I represents the inversion loss function;

[0024] The data loss function L d The expression is:

[0025]

[0026] in, represents the forward simulation data without adding environmental noise; represents the estimated data output by the second-stage decoder; n1 represents the total number of data in the forward simulation data without adding environmental noise.

[0027] The inversion loss function L I The expression is:

[0028]

[0029] in, represents a geoelectric model in the first data set; represents the inversion result output by the inversion decoder; n2 represents the total number of resistivities in the geoelectric model.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The present invention sets up a decoupled representation learning network, uses a first-stage encoder to extract a first data factor from a first data set, and uses the first data factor for inversion in a second-stage decoder. The decoupled representation learning network has the advantages of improving model interpretability, balancing data representation and decoupling capabilities, being applicable to complex data sets, increasing sample diversity, and promoting cross-domain decoupled representation. It can reliably output the first data factor.

[0032] (2) The present invention combines decoupled representation learning to perform multi-stage inversion of semi-airborne electromagnetic data, effectively solving the problems that traditional methods are time-consuming, mostly rely on the selection of empirical parameters, and are unable to invert complex field data with large noise.

[0033] (3) The present invention uses noise data for inversion, performs noise suppression in the first stage, and extracts data factors, which effectively solves the problem that existing deep learning methods cannot process noise data and need to use pre-processed data for inversion, which increases the workload.

[0034] (4) The encoder of the present invention encodes data into data factors and noise factors, and the inversion decoder uses the data factors for inversion, making the inversion more accurate and interpretable.

[0035] In summary, the present invention has the advantages of simple logic, accuracy and reliability, and has high practical value and promotion value in the field of semi-aeronautical electromagnetic noise data processing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope of protection. For those skilled in the art, other relevant drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 It is a logic flow chart of the present invention.

[0038] Figure 2 This is an example model and forward simulation data diagram in the present invention.

[0039] Figure 3 Schematic diagram for adding noise in the present invention

[0040] Figure 4 This is a network architecture diagram of the present invention.

[0041] Figure 5 This is a diagram of the backbone network structure of the encoder and decoder in the present invention.

[0042] Figure 6It is a schematic diagram of training loss in the present invention.

[0043] Figure 7 2 is an inversion result diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] To make the purpose, technical solutions, and advantages of this application more clear, the present invention is further described below with reference to the accompanying drawings and examples. Implementation methods of the present invention include, but are not limited to, the following examples. All other embodiments obtained by persons of ordinary skill in the art based on the examples in this application without creative effort are within the scope of protection of this application.

[0045] In this embodiment, the term "and / or" is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone.

[0046] In the description and claims of this embodiment, the terms "first" and "second" are used to distinguish different objects rather than to describe a specific order of objects. For example, a first target object and a second target object are used to distinguish different objects rather than to describe a specific order of objects.

[0047] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0048] In the description of the embodiments of this application, unless otherwise specified, "multiple" means two or more. For example, "multiple processing units" means two or more processing units; "multiple systems" means two or more systems.

[0049] like Figures 1 to 7 As shown, this embodiment provides a multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning, which includes the following steps:

[0050] The first step is to obtain the geoelectric model in the known geoelectric model set DL-RMD for forward modeling and obtain forward simulation data; the forward simulation data of any forward modeling is paired with the geoelectric model. Among them, 10,000 geoelectric models are randomly selected from the geoelectric model set DL-RMD. One of the geoelectric models and its forward simulation results are as follows Figure 2 As shown. Among them, Figure 2 (a) in the figure represents the geoelectric model.Figure 2 (b) in the figure shows the forward modeling result.

[0051] In the second step, environmental noise is added to the paired forward simulation data to obtain first simulation data with added environmental noise, and the forward simulation data, the geoelectric model and the first simulation data are paired to form a first data set.

[0052] In this embodiment, noise addition and data pairing both include the following steps:

[0053] (1) Take a row of data from the forward response data file and add the environmental noise. The expression is:

[0054] d n =d c +N(0,1)

[0055] Among them, d n represents the first simulation data after adding environmental noise; d c represents the forward simulation data without adding environmental noise; N(0,1) represents Gaussian noise with mean 0 and variance 1. The data before and after adding noise are as follows Figure 3 As shown, Figure 3 (a) in the figure represents the data before noise is added. Figure 3 (b) in the figure shows the data after noise is added.

[0056] (2) The forward modeling response data after adding the environmental noise is written into a text file, and the line number of the written file is consistent with the line number of the forward modeling response data.

[0057] The third step is to build the first-stage encoder of the decoupled representation learning network and extract the first data factor from the first dataset. The second-stage decoder of the decoupled representation learning network is built and the first data factor is used for inversion in the second-stage decoder.

[0058] Here, the first-stage encoder performs data encoding to obtain a first data factor corresponding to the first data set and a second data factor corresponding to the actually collected noise data; both the first and second data factors are composed of a noise factor and a signal factor. Furthermore, the second-stage decoder includes a data decoder and an inversion decoder. The data decoder decodes the arithmetic sum of the noise factor and the signal factor to obtain estimated data, while the inversion decoder inverts the signal factor to obtain resistivity.

[0059] The first-stage encoder, data decoder, and inversion decoder of this embodiment have the same structure, consisting of an embedding layer and a backbone network; the embedding layer performs a one-dimensional convolution operation on the data obtained from the input layer to obtain embedded data;

[0060] The backbone network includes a first linear layer, a second linear layer, and a multilayer perceptron connected in sequence. The first linear layer obtains the embedded data output by the embedding layer and performs full-connection processing to obtain the value, key, and query value corresponding to the embedded data. Matrix multiplication is performed on the key and query value output by the first linear layer to obtain third data. Matrix multiplication is performed on the value output by the first linear layer and the third data to obtain fourth data, which is input into the second linear layer. The second linear layer outputs fifth data, which is arithmetic summed with the embedded data output by the embedding layer to obtain sixth data. The multilayer perceptron performs a full-connection operation on the sixth data and then activates it using a sigmoid activation function to obtain seventh data. The sixth data and the seventh data are arithmetic summed to obtain the output of the backbone network.

[0061] The fourth step is to construct the loss function. The loss function includes the data loss function and the inversion loss function. Here, the loss function L oss , whose expression is:

[0062] L oss =L d +L I

[0063] Among them, L d represents the data loss function; L I represents the inversion loss function.

[0064] The data loss function L d The expression is:

[0065]

[0066] in, represents the forward simulation data without adding environmental noise; represents the estimated data output by the second-stage decoder; n1 represents the total number of data in the forward simulation data without adding environmental noise.

[0067] The inversion loss function L I The expression is:

[0068]

[0069] in, represents a geoelectric model in the first data set; represents the inversion result output by the inversion decoder; n2 represents the total number of resistivities in the geoelectric model.

[0070] The fifth step is to use the first data set to train the decoupled representation learning network and adjust the parameters to obtain the trained decoupled representation learning network. Here, the first data set is divided into training set, validation set, and test set in a ratio of 7:2:1; the divided data set is input into the decoupled representation learning network for training. During training, the ADAM optimizer is used, the learning rate is set to 0.0001, the batch size is set to 16, and the number of training rounds is set to 100 rounds. In addition, the parameters of the decoupled representation learning network are fine-tuned so that the validation set loss fluctuates within a certain range and the network reaches the optimal state. Among them, the loss of network training is as follows: Figure 6 shown.

[0071] The sixth step is to input the actual noise data collected into the trained decoupled representation learning network to obtain the second data factor, and use the second data factor to perform inversion in the second stage decoder of the trained decoupled representation learning network. The inversion result is as follows: Figure 7 As shown, Figure 7 (a) is the noisy data. Figure 7 (b) is the inversion result.

[0072] The above embodiments are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes that adopt the design principles of the present invention and any changes made through non-creative work on this basis should fall within the scope of protection of the present invention.

Claims

1. A multi-stage inversion method for semi-aeronautical electromagnetic noise data based on decoupled representation learning, characterized by: The following steps are involved: Obtain the geoelectric model in the known geoelectric model set DL-RMD for forward modeling to obtain forward simulation data; pair the forward simulation data of any forward modeling with the geoelectric model; Adding environmental noise to the paired forward simulation data to obtain first simulation data with the added environmental noise, and pairing the forward simulation data, the geoelectric model, and the first simulation data to form a first data set; Build the first-stage encoder of the decoupled representation learning network and extract the first data factor in the first data set; Build the second-stage decoder of the decoupled representation learning network and use the first data factor for inversion in the second-stage decoder; Constructing a loss function; the loss function includes a data loss function and an inversion loss function; Using the first data set to train the decoupled representation learning network, and adjusting parameters to obtain a trained decoupled representation learning network; The actually collected noise data is input into the trained decoupled representation learning network to obtain a second data factor, and the second data factor is used to perform inversion in the second stage decoder of the trained decoupled representation learning network.

2. The multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning according to claim 1 is characterized in that: Adding environmental noise to the paired forward simulation data to obtain the first simulation data with added environmental noise, including Take out the paired forward simulation data one by one and add environmental noise. The expression is: d n =d c +N(0,1) Among them, d n represents the first simulation data after adding environmental noise; d c represents the forward simulation data without adding environmental noise; N(0,1) represents Gaussian noise.

3. The multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning according to claim 1 or 2, characterized in that: The first stage encoder performs data encoding to obtain a first data factor corresponding to the first data set and a second data factor corresponding to the actually collected noise data; the first data factor and the second data factor are both composed of a noise factor and a signal factor.

4. The multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning according to claim 3 is characterized in that: The second stage decoder includes a data decoder and an inversion decoder; the data decoder decodes the arithmetic sum of the noise factor and the signal factor to obtain estimated data; the inversion decoder inverts the signal factor to obtain resistivity.

5. The multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning according to claim 4 is characterized in that: The first-stage encoder, data decoder, and inversion decoder have the same structure, consisting of an embedding layer and a backbone network; the embedding layer performs a one-dimensional convolution operation on the data obtained from the input layer to obtain embedded data; The backbone network includes a first linear layer, a second linear layer, and a multilayer perceptron connected in sequence; the first linear layer obtains the embedded data output by the embedding layer, performs full connection processing to obtain the value, key, and query value corresponding to the embedded data; Perform matrix multiplication on the key and query value output by the first linear layer to obtain third data, perform matrix multiplication on the value output by the first linear layer and the third data to obtain fourth data, and input it into the second linear layer; the second linear layer outputs fifth data, and performs an arithmetic sum operation on the embedded data output by the embedding layer to obtain sixth data; The multilayer perceptron performs a full connection operation on the sixth data and activates it using a sigmoid activation function to obtain seventh data, and performs arithmetic summation on the sixth data and the seventh data to obtain the output of the backbone network.

6. The multi-stage inversion method for semi-airborne electromagnetic noise data based on decoupled representation learning according to claim 5 is characterized in that: The loss function L oss , whose expression is: L oss =L d +L I Among them, L d represents the data loss function; L I represents the inversion loss function; The data loss function L d The expression is: in, represents the forward simulation data without adding environmental noise, represents the estimated data output by the second-stage decoder; n1 represents the total number of data in the forward simulation data without adding environmental noise; The inversion loss function L I The expression is: in, represents a geoelectric model in the first data set; represents the inversion result output by the inversion decoder; n2 represents the total number of resistivities in the geoelectric model.

Images