Decoupling representation learning based semi-airborne transient electromagnetic data denoising method

Abstract

The application discloses a semi-airborne transient electromagnetic data denoising method based on decoupling feature learning, and comprises the following steps: acquiring a known geophysical constraint electromagnetic resistivity model database DL-RMD, and performing forward processing to obtain forward simulation data; adding noise to the forward simulation data, obtaining a plurality of sample pairs, and splitting the sample pairs into a training set, a verification set and a test set; building a decoupling feature learning network, and outputting a pair of decoupling analysis by using the decoupling feature learning network; training the decoupling feature learning network by using the training set, and performing parameter adjustment fitting to obtain a trained decoupling feature learning network; verifying and testing the trained decoupling feature learning network by using the verification set and the test set respectively; acquiring actual collected semi-airborne transient electromagnetic data, and performing denoising processing by using the trained decoupling feature learning network. By the above scheme, the application has the advantages of simple logic, accuracy and reliability and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electromagnetic data processing, in particular to a semi-airborne transient electromagnetic data denoising method based on decoupled feature learning. BACKGROUND

[0002] Semi-airborne transient electromagnetic method is a geophysical exploration technology developed on the basis of magnetotelluric method, which extracts and measures the geoelectric structure by observing the received secondary field signal, and can realize rapid coverage of some large area and inaccessible measurement area by placing the observation system on a flying vehicle such as a fixed-wing aircraft or a helicopter. Since the system is observed in motion (during the flight of the fixed-wing aircraft or the helicopter), it inevitably introduces a lot of noise, and the noise condition is complex, which brings great difficulty to subsequent data processing. In order to ensure the accuracy and reliability of detection, noise removal is needed.

[0003] Among them, semi-airborne transient electromagnetic data denoising refers to removing noise from noise data collected in the field to obtain clean data after denoising. However, most traditional methods rely on the selection of some empirical parameters, which puts high requirements on the experience of denoising personnel, and these methods are more suitable for data with small noise levels, and it is difficult to denoise complex field data with large noise. Among them, in the disclosed technology with publication number CN113655532A and the name of a non-full-time semi-airborne transient electromagnetic data motion noise removal method and system, it adopts: obtaining time domain secondary field sample data, and its sampling frequency and sample data period; the multiple half-cycle sample data contained in the secondary field sample data are extended to full length; the motion noise baseline of the full length is preliminarily fitted, and the expected frequency is obtained through frequency spectrum analysis; according to the expected frequency, the inverse discrete Fourier transform is used to obtain the full-length time-domain motion noise; the obtained motion noise is removed from the secondary field sample data. This technology only removes the motion noise, and the noise brought by complex terrain and the like cannot be processed.

[0004] And the existing deep learning-based methods all output the prediction of noise or signal, which is a single learning scheme, not decoupled. This single representation is difficult to separate signal and noise, and the representation result often contains both noise and signal, which leads to their difficulty in effectively separating noise and signal when processing complex actual signals, and they can only focus on part of the information in the signal.

[0005] Therefore, it is urgent to propose a semi-airborne transient electromagnetic data denoising method based on decoupled feature learning, which is simple in logic, accurate and reliable. SUMMARY

[0006] In view of the above problems, the present invention aims to provide a semi-airborne transient electromagnetic data denoising method based on decoupled representation learning. The technical solutions adopted by the present invention are as follows:

[0007] A semi-airborne transient electromagnetic data denoising method based on decoupled representation learning includes the following steps:

[0008] Obtain the known geophysical constrained electromagnetic resistivity model database DL-RMD and perform forward modeling to obtain forward simulation data;

[0009] Noise is added to the forward simulation data to obtain several sample pairs, which are then split into training set, validation set, and test set;

[0010] Build a decoupled representation learning network and use it to output a pair of decoupled parsing;

[0011] The decoupled representation learning network is trained using the training set, and parameter adjustment and fitting are performed to obtain the trained decoupled representation learning network;

[0012] The trained decoupled representation learning network is verified and tested using the validation set and test set respectively;

[0013] The actual collected semi-airborne transient electromagnetic data are obtained and denoised using the trained decoupled representation learning network.

[0014] Furthermore, noise is added to the forward modeling data, including the following steps:

[0015] Divide the forward simulation data into two parts at a ratio of 1:1 to obtain the first forward simulation data and the second forward simulation data;

[0016] Add Gaussian noise to the first forward simulation data, and its expression is:

[0017] s n1 =s1+N(0,1)

[0018] Among them, s n1 represents the data after adding Gaussian noise; s1 represents the first forward simulation data; N(0,1) represents Gaussian noise;

[0019] Add mixed noise to the second forward simulation data, and its expression is:

[0020]

[0021] Among them, s n2 represents the data after adding mixed noise; s2 represents the second forward simulation data; STD represents linear noise; n represents the background noise distribution.

[0022] Further, the expression of the background noise distribution n is:

[0023]

[0024] Wherein, b represents the noise level at 1ms; t represents time.

[0025] Further, the decoupling representation learning network comprises an encoder and a decoder;

[0026] The encoder comprises a first input layer, a first embedding layer and a first backbone network connected in sequence; the embedding layer performs convolution operation on the forward modeling data with added noise obtained by the first input layer to obtain embedding output; and the first backbone network encodes the embedding output to obtain encoding data.

[0027] The decoder comprises a second backbone network and a first linear layer connected in sequence; the second backbone network decodes the encoding data to obtain decoding data; and the first linear layer decouples and analyzes the decoding data output by the second backbone network.

[0028] Further, the first backbone network and the second backbone network have the same structure, and the first backbone network comprises a second input layer, a second embedding layer, a mixing layer, a second linear layer and a second output layer connected in sequence; the mixing layer comprises a signal mixing layer and a channel mixing layer connected in sequence; the second input layer obtains the forward modeling data with added noise; the second embedding layer performs one-dimensional convolution operation on the data input by the second input layer to obtain embedding data; the signal mixing layer performs signal mixing on the embedding data to obtain first data; the channel mixing layer performs channel mixing on the first data after signal mixing to obtain second data; and the second linear layer performs full connection operation on the second data after channel mixing to obtain third data.

[0029] Further, the signal mixing layer comprises a first token transfer layer, a WKV operation layer and a first normalization layer; the first token transfer layer obtains the embedding data output by the second embedding layer, and places the last data in the embedding data at the forefront of the embedding data, and splits the embedding data in the channel dimension to obtain the value, key and received value of the embedding data; the WKV operation layer performs WKV addition attention operation on the value and key of the embedding data to obtain fourth data; the fourth data output by the WKV operation layer is subjected to matrix multiplication processing with the received value of the embedding data, and is sent to the first normalization layer for normalization processing to obtain fifth data.

[0030] Further, the channel mixing layer comprises a second token transfer layer and a second normalization layer; the second token transfer layer obtains fifth data and places the last data in the fifth data at the forefront of the fifth data to obtain a value corresponding to the fifth data and a received value; the value corresponding to the fifth data and the received value output by the second token transfer layer are mixed by using a matrix multiplication method to obtain sixth data; and the second normalization layer performs normalization processing on the sixth data.

[0031] Further, the decoupling representation learning network is trained by using a training set, and parameter adjustment fitting is performed, and a total loss function L is added to the decoupling representation learning network oss , which is expressed as:

[0032] L oss = L s + L no

[0033] Wherein, L s represents a signal loss function; and L no represents a noise loss function.

[0034] The expression of the signal loss function L s is:

[0035]

[0036] Wherein, represents data without adding noise in a sample pair; represents a decoupling representation of a signal output by the decoupling representation learning network; and n 1 represents a total number of signals.

[0037] The expression of the noise loss function L no is:

[0038]

[0039] Wherein, represents data after adding noise in a sample pair; represents a decoupling representation of noise output by the decoupling representation learning network; and n 2 represents a total number of data after adding noise.

[0040] Further, the actually collected semi-airborne transient electromagnetic data is obtained, and the decoupling representation learning network after training is used for denoising processing, comprising the following steps:

[0041] The actually collected semi-airborne transient electromagnetic data is input into the decoupling representation learning network after training to obtain double-channel network output data of the decoupling representation learning network;

[0042] The dual-channel network output data is split in an array splitting manner to obtain noise prediction values and signal prediction values of the decoupled representation learning network;

[0043] The noise prediction values of the decoupled representation learning network are subtracted from the actually collected semi-airborne transient electromagnetic data to obtain denoised data.

[0044] Compared with the prior art, the present application has the following beneficial effects:

[0045] (1) The present application uses a known geophysical constraint electromagnetic resistivity model database DL-RMD, adds noise to the forward simulation data, and combines deep learning for data denoising, effectively solving the problem that traditional methods mostly rely on the selection of empirical parameters, and are more suitable for data with small noise levels, and it is difficult to denoise complex field data with large noise.

[0046] (2) The present application combines decoupled representation learning, so that the network outputs a pair of decoupled representations of signals and noise, solving the problem that in existing deep learning, a single representation is difficult to separate signals and noise, and the representation result often contains both noise and signals, which makes it difficult to effectively separate noise and signals when processing complex actual signals, and can only focus on part of the information in the signal.

[0047] (3) The present application uses noise data to subtract the network's prediction of noise when obtaining the final denoising result, and this residual learning method has better performance and better denoising results.

[0048] (4) The present application sets up a decoupled representation learning network, which has the advantages of improving model interpretability, balancing data representation and decoupling ability, being suitable for complex data sets, improving sample diversity, and promoting cross-domain decoupled representation, and can reliably output representations of noise and signals.

[0049] In summary, the present application has the advantages of simple logic, accuracy and reliability, and has high practical value and promotional value in the field of electromagnetic data processing technology. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments, and it should be understood that the following drawings only show some embodiments of the present application, and should not be regarded as limiting the scope of protection, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0051] Figure 1 The present application is a logical flowchart.

[0052] Figure 2A schematic diagram of data noise adding in the present application.

[0053] Figure 3 A schematic diagram of decoupling the structure of the feature learning network in the present application.

[0054] Figure 4 A structure diagram of the backbone network in the present application.

[0055] Figure 5 A schematic diagram of training loss in the present application.

[0056] Figure 6 A denoising result diagram in the present application. DETAILED DESCRIPTION

[0057] To make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described below in combination with the drawings and embodiments, and the embodiments of the present application include but are not limited to the following embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor shall fall within the scope of protection of the present application.

[0058] In the present embodiment, the term "and / or" is merely used to describe the association relationship of the associated objects, and can represent the existence of three relationships, for example, A and / or B can represent the existence of A alone, the existence of A and B together, and the existence of B alone.

[0059] The terms "first" and "second" and the like in the description and claims of the present embodiment are used to distinguish different objects, and are not used to describe the specific order of the objects. For example, the first target object and the second target object are used to distinguish different target objects, and are not used to describe the specific order of the target objects.

[0060] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design schemes. Rather, the use of "exemplary" or "for example" is intended to present relevant concepts in a concrete manner.

[0061] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more. For example, a plurality of processing units means two or more processing units; a plurality of systems means two or more systems.

[0062] As Figures 1 to 6 shown, the present embodiment provides a semi-airborne transient electromagnetic data denoising method based on decoupling feature learning, which comprises the following steps:

[0063] Firstly, a known geophysical constraint electromagnetic resistivity model database DL-RMD is obtained, and forward processing is performed to obtain forward simulation data. In this embodiment, 5000 models randomly selected from the geophysical constraint electromagnetic resistivity model database DL-RMD are subjected to forward processing.

[0064] Secondly, noise is added to the forward simulation data, and a plurality of sample pairs are obtained, and the plurality of sample pairs are divided into a training set, a verification set and a test set according to a ratio of 7:2:1. As shown in FIG. 2, wherein, Figure 2 Figure 2 (a) in FIG. 2 is an example of a model in the DL-RMD model set, Figure 2 (b) in FIG. 2 is forward data obtained by the model (a), Figure 2 (c) in FIG. 2 is a result of adding noise to the forward data in (b). In this embodiment, the forward simulation data is divided into two parts according to a ratio of 1:1, to obtain first forward simulation data and second forward simulation data.

[0065] Gaussian noise is added to the first forward simulation data, and an expression of the Gaussian noise is as follows:

[0066] s n1 =s1+N(0,1)

[0067] wherein s n1 represents data after Gaussian noise is added; s1 represents the first forward simulation data; and N(0,1) represents Gaussian noise with a mean of 0 and a variance of 1.

[0068] Mixed noise is added to the second forward simulation data, and an expression of the mixed noise is as follows:

[0069]

[0070] wherein s n2 represents data after mixed noise is added; s2 represents the second forward simulation data; STD represents linear noise; and n represents background noise distribution.

[0071] An expression of the background noise distribution n is as follows:

[0072]

[0073] wherein b represents a noise level at 1 ms; and t represents time.

[0074] Thirdly, a decoupling feature learning network is built, and the decoupling feature learning network outputs a pair of decoupling solutions. As shown in FIG. 3, Figures 3 to 4 ​As shown, the input of the encoder is single-channel data, and the output is 256-channel data for decoding by the decoder; the input of the decoder is 256 channels, and the output is two channels, the first channel being the decoupling representation of the noise by the decoder, and the second channel being the decoupling representation of the signal by the decoder. Specifically, the decoupling representation learning network comprises an encoder and a decoder. The encoder comprises a first input layer, a first embedding layer and a first backbone network connected in sequence; the embedding layer performs convolution operation on the forward modeling data with added noise obtained by the first input layer to obtain embedding output; and the first backbone network encodes the embedding output to obtain encoded data. In addition, the decoder comprises a second backbone network and a first linear layer connected in sequence; the second backbone network decodes the encoded data to obtain decoded data; and the first linear layer decouples and analyzes the decoded data output by the second backbone network.

[0075] In this embodiment, the first backbone network and the second backbone network have the same structure, and the first backbone network comprises a second input layer, a second embedding layer, a mixing layer, a second linear layer and a second output layer connected in sequence; the mixing layer comprises a signal mixing layer and a channel mixing layer connected in sequence; the second input layer obtains the forward modeling data with added noise; the second embedding layer performs one-dimensional convolution operation on the data input by the second input layer to obtain embedding data; the signal mixing layer performs signal mixing on the embedding data to obtain first data; the channel mixing layer performs channel mixing on the first data after signal mixing to obtain second data; and the second linear layer performs full connection operation on the second data after channel mixing to obtain third data.

[0076] The signal mixing layer comprises a first token transfer layer, a WKV operation layer and a first normalization layer; the first token transfer layer obtains the embedding data output by the second embedding layer, and places the last data in the embedding data at the forefront of the embedding data, and splits the embedding data in the channel dimension to obtain the value, key and received value of the embedding data; the WKV operation layer performs WKV addition attention operation on the value and key of the embedding data to obtain fourth data; the fourth data output by the WKV operation layer is subjected to matrix multiplication processing with the received value of the embedding data, and is sent to the first normalization layer for normalization processing to obtain fifth data.

[0077] In addition, the channel mixing layer comprises a second token transfer layer and a second normalization layer; the second token transfer layer obtains the fifth data and places the last data in the fifth data at the forefront of the fifth data to obtain the value and received value corresponding to the fifth data; the value and received value corresponding to the fifth data output by the second token transfer layer are mixed by matrix multiplication to obtain sixth data; and the second normalization layer performs normalization processing on the sixth data.

[0078] Fourthly, the decoupling representation learning network is trained by using the training set, and the parameters are adjusted and fitted to obtain the trained decoupling representation learning network. In this regard, the decoupling representation learning network is adjusted and fitted to make the loss of the test set fluctuate in a certain range. Figure 5 The change of the loss value during the network training is shown, and it can be seen that the loss gradually converges.

[0079] In this embodiment, a total loss function L oss is added to the decoupling representation learning network, and the expression is as follows:

[0080] L oss = L s + L no

[0081] Wherein, L s represents a signal loss function; L no represents a noise loss function.

[0082] The expression of the signal loss function L s is as follows:

[0083]

[0084] Wherein, represents the data without adding noise in the sample pair; represents the decoupled representation of the signal output by the decoupling representation learning network; and n 1 represents the total number of signals.

[0085] The expression of the noise loss function L no is as follows:

[0086]

[0087] Wherein, represents the data after adding noise in the sample pair; represents the decoupled representation of the noise output by the decoupling representation learning network; and n 2 represents the total number of data after adding noise.

[0088] Fifthly, the trained decoupling representation learning network is verified and tested by using the verification set and the test set respectively.

[0089] Sixthly, the actually collected semi-airborne transient electromagnetic data is obtained, and the trained decoupling representation learning network is used for denoising. Specifically:

[0090] (1) The actually collected semi-airborne transient electromagnetic data is input into the trained decoupling representation learning network to obtain the dual-channel network output data of the decoupling representation learning network.

[0091] (2) The two-channel network output data is split by array splitting to obtain the noise prediction value and the signal prediction value of the decoupling representation learning network; here, it is a pair of decoupled representations, the first channel is a representation of noise, and the second channel is a representation of signal.

[0092] (3) The actual collected semi-airborne transient electromagnetic data is subtracted by the noise prediction value of the decoupling representation learning network to obtain the denoised data.

[0093] As can be seen from Figure 6 , the actual data Figure 6 (a) The latter part is greatly interfered by noise, and information is difficult to use, Figure 6 (b) is the data after processing by the present application, the data noise is obviously removed, the data trend is smooth, and conforms to the actual law. It can be seen that the present application effectively improves the efficiency and accuracy of actual data denoising, and provides an efficient general framework template for future similar geological engineering actual noise data denoising.

[0094] The above examples are only preferred embodiments of the present application, and are not a limitation on the scope of protection of the present application, but any change made by using the design principle of the present application and on this basis without creative labor shall belong to the protection scope of the present application.

Claims

1. A semi-aeronautical transient electromagnetic data denoising method based on decoupled representation learning, characterized by: The following steps are involved: Obtain the known geophysical constrained electromagnetic resistivity model database DL-RMD and perform forward modeling to obtain forward simulation data; Noise is added to the forward simulation data to obtain several sample pairs, which are then split into training set, validation set, and test set; Build a decoupled representation learning network and use it to output a pair of decoupled parsing; The decoupled representation learning network is trained using the training set, and parameter adjustment and fitting are performed to obtain the trained decoupled representation learning network; The trained decoupled representation learning network is verified and tested using the validation set and test set respectively; Obtain the actual collected semi-airborne transient electromagnetic data and perform denoising using the trained decoupled representation learning network, including the following steps: Input the actually collected semi-aeronautical transient electromagnetic data into the trained decoupling representation learning network to obtain the dual-channel network output data of the decoupling representation learning network; The dual-channel network output data is split using an array splitting method to obtain the noise prediction value and signal prediction value of the decoupled representation learning network; The noise prediction value of the decoupling representation learning network is subtracted from the actual collected semi-airborne transient electromagnetic data to obtain the denoised data.

2. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 1 is characterized in that: Adding noise to the forward modeling data includes the following steps: Dividing the forward simulation data into two parts at a ratio of 1:1 to obtain first forward simulation data and second forward simulation data; Add Gaussian noise to the first forward simulation data, and its expression is: in, represents the data after adding Gaussian noise; represents the first forward simulation data; represents Gaussian noise; Add mixed noise to the second forward simulation data, and its expression is: in, Represents the data after adding mixed noise; represents the second forward simulation data; represents linear noise; n represents background noise distribution.

3. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 2 is characterized in that: The expression of the background noise distribution n is: in, Indicates the noise level at 1ms; Indicates time.

4. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 1 is characterized in that: The decoupled representation learning network includes an encoder and a decoder; The encoder includes a first input layer, a first embedding layer, and a first backbone network connected in sequence; the embedding layer performs a convolution operation on the forward simulation data with added noise obtained by the first input layer to obtain an embedded output; the first backbone network encodes the embedded output to obtain encoded data; The decoder includes a second backbone network and a first linear layer connected in sequence; the second backbone network decodes the encoded data to obtain decoded data; The first linear layer decouples and analyzes the decoded data output by the second backbone network.

5. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 4 is characterized in that: The first backbone network and the second backbone network have the same structure, and the first backbone network includes a second input layer, a second embedding layer, a mixing layer, a second linear layer, and a second output layer connected in sequence; the mixing layer includes a signal mixing layer and a channel mixing layer connected in sequence; the second input layer obtains forward simulation data with added noise; the second embedding layer performs a one-dimensional convolution operation on the data input by the second input layer to obtain embedded data; The signal mixing layer performs signal mixing on the embedded data to obtain first data; the channel mixing layer performs channel mixing on the first data after signal mixing to obtain second data; and the second linear layer performs a full connection operation on the second data after channel mixing to obtain third data.

6. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 5 is characterized in that: The signal mixing layer includes a first token transfer layer, a WKV operation layer and a first normalization layer; the first token transfer layer obtains the embedded data output by the second embedding layer, and puts the last data in the embedded data at the front end of the embedded data, and splits it in the channel dimension to obtain the value, key and received value of the embedded data; the WKV operation layer performs a WKV addition attention operation on the value and key of the embedded data to obtain the fourth data; the fourth data output by the WKV operation layer is matrix multiplied with the received value of the embedded data, and is sent to the first normalization layer for normalization processing to obtain the fifth data.

7. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 6 is characterized in that: The channel mixing layer includes a second token transfer layer and a second normalization layer; the second token transfer layer obtains the fifth data and puts the last data in the fifth data at the front of the fifth data to obtain the value corresponding to the fifth data and the received value; The value corresponding to the fifth data output by the second token transfer layer and the received value are mixed using a matrix multiplication method to obtain sixth data; The second normalization layer performs normalization processing on the sixth data.

8. The semi-airborne transient electromagnetic data denoising method based on decoupled representation learning according to claim 4, 5, 6 or 7, characterized in that: Use the training set to train the decoupled representation learning network, adjust the parameters and fit them, and add the total loss function to the decoupled representation learning network. , which is expressed as: in, represents the signal loss function; represents the noise loss function; the signal loss function The expression is: in, represents the data without adding noise to the sample pair; Representing the decoupled representation of the signal output by the decoupled representation learning network; Indicates the total number of signals; The noise loss function The expression is: in, Represents the data after adding noise to the sample pair; Representation of the decoupled representation of the noise output of the decoupled representation learning network; Indicates the total number of data after adding noise.

Images