Method and system for suppressing strong interference in artificial current source electromagnetic data

Abstract

A method and system for efficiently suppressing strong interference in artificial current source electromagnetic data with high accuracy is provided. [Solution] The method inputs the artificial current source electromagnetic data to be processed into an artificial current source electromagnetic data classification model to obtain first noise-free artificial current source electromagnetic data and noise-containing artificial current source electromagnetic data, converts the noisy artificial current source electromagnetic data into two-dimensional image data using a dimension conversion function and inputs it into a DWTSC-UNet noise removal network model to obtain the noise-removed data result, performs inverse dimension conversion on the noise-removed data result using the dimension conversion function to obtain second noise-free artificial current source electromagnetic data, and combines the first noise-free artificial current source electromagnetic data and the second noise-free artificial current source electromagnetic data to obtain complete noise-free artificial current source electromagnetic data.

Description

[Technical Field]

[0001] The present invention relates to the intersecting fields of electronic information, artificial intelligence, and geophysics, and more particularly to a method and system for suppressing strong interference in controlled-source electromagnetic method (CSEM) data using a SC-UNet (Discrete Wavelet Transform optimized Spatial and Channel Attention Mechanism U-shaped Neural Network, DWTSC-UNet) network optimized by dimensional transformation and discrete wavelet transform. [Background technology]

[0002] The artificial current source electromagnetic method is an electromagnetic exploration method widely used in fields such as near-surface geophysical exploration and engineering exploration. It is currently increasingly used in fields such as shale gas exploration, metal ore exploration, and engineering exploration. Its essence is to generate different responses depending on the electrical structure of the underground medium to derive the distribution of the underground medium. The artificial current source electromagnetic method has a known frequency and controllable electric field source, and has higher anti-interference capabilities than the natural electric field source electromagnetic field method. With the increase in urbanization and industrialization, the distribution of human noise is becoming increasingly widespread, and artificial current source electromagnetic signals are highly susceptible to human noise. Therefore, people's requirements for exploration depth and accuracy are becoming increasingly higher, and suppressing noise in artificial current source electromagnetic signals is becoming increasingly important.

[0003] Facing severe human-caused noise contamination, data filtering methods are typically used. Given the periodic characteristics of electromagnetic signals from artificial current sources, the simplest approach is to screen and remove segments containing strong human-caused interference on a periodic basis. Data screening methods can improve data quality to some extent and extract valid signals. However, they still have limitations. First, traditional manual data filtering not only requires a certain amount of a priori knowledge, but is also inefficient and prone to subjective influence. Second, a single data statistical feature has difficulty distinguishing between noise segments and valid signals in complex noise environments. Using multiple sets of data statistical features requires considering the mutual influence of feature parameters. Third, residual noise (e.g., harmonics and Gaussian white noise) still exists in the filtered data.

[0004] In recent years, deep learning technology has developed rapidly, and its applications in electromagnetic fields have become increasingly widespread. Deep learning algorithms are increasingly being applied in electromagnetic signal processing. However, due to the complexity and polygonal shape of artificial current source electromagnetic signals, when the signals are subjected to strong noise interference, the learning model has difficulty accurately capturing the complete characteristics of the useful signal, which can lead to overfitting or underfitting. In addition, deep learning networks used in traditional methods, which were proposed decades ago, suffer from problems such as vanishing gradients, excessive parameter amounts, excessive memory consumption, and insufficient noise removal accuracy. Summary of the Invention

[0005] In response to the deficiencies of the prior art, the object of the present invention is to provide a method and system for suppressing strong interference in controlled-source electromagnetic data (also known as "artificial signal source electromagnetic data" or "controlled source electromagnetic data"), which aims to resolve the complexity and polygonal shape of artificial source electromagnetic signals, and to solve the problem that when a signal is subjected to strong noise interference, it is difficult for a learning model to accurately capture the complete characteristics of the useful signal, resulting in overfitting or underfitting and insufficient noise removal accuracy.

[0006] To achieve the above object, according to a first aspect, there is provided a method for suppressing strong interference in artificial current source electromagnetic data, comprising: inputting the artificial current source electromagnetic data to be processed into an artificial current source electromagnetic data classification model to obtain first noise-free artificial current source electromagnetic data and noise-containing artificial current source electromagnetic data; converting the noisy artificial current source electromagnetic data into two-dimensional image data by a dimension transformation function, and inputting the data into the DWTSC-UNet denoising network model to obtain the denoised data result; performing a dimensional inverse transformation on the denoised data result using a dimensional transformation function to obtain second noise-free artificial current source electromagnetic data; combining the first noise-free artificial-source electromagnetic data and the second noise-free artificial-source electromagnetic data to obtain complete noise-free artificial-source electromagnetic data; Including, The DWTSC-UNet denoising network model is a U-structured network that combines a discrete wavelet transform, a spatial attention mechanism, and a channel attention mechanism; A discrete wavelet transform is used to provide high-frequency and low-frequency signals, and weight learning is performed on different frequency components of the two-dimensional image data to extract multi-scale features from the two-dimensional image data. A channel attention mechanism is used to apply weights to different dimensions of the feature map, and a spatial attention mechanism is used to apply weights to different spatial locations of the feature map. A method for suppressing strong interference in artificial current source electromagnetic data is provided.

[0007] More preferably, the method for training the DWTSC-UNet denoising network model includes: selecting noisy artificial current source electromagnetic data samples from the actual measurement data, and performing noise removal using a dictionary learning method to separate the actual measurement noise signals; adding the measured noise signal and the analog noise signal to the noise-free artificial current source electromagnetic data to form noisy artificial current source electromagnetic data, which are combined with the noise-free artificial current source electromagnetic data samples to form sample pairs, thereby constructing a one-dimensional denoised sample library; Segmenting the one-dimensional denoised sample library and performing a dimensional transformation on each sample using a dimensional transformation function to obtain a two-dimensional image sample library; inputting a 2D image sample library into a DWTSC-UNet denoising network model for training; Includes.

[0008] More preferably, the DWTSC-UNet denoising network model includes an input layer, a hidden layer and an output layer, The hidden layer includes a discrete wavelet transform layer, a first shape reconstruction layer, a first convolution block, a first spatial attention mechanism module, a second convolution block, a second spatial attention mechanism module, a third convolution block, a first up-sample block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second up-sample block, a second skip connection layer, a second channel attention mechanism module, and a third up-sample block.

[0009] More preferably, the discrete wavelet transform layer decomposes the two-dimensional image data into a low-frequency approximation portion, a horizontal high-frequency portion, a vertical high-frequency portion, and a diagonal high-frequency portion, wherein the low-frequency approximation portion includes the basic contour of the two-dimensional image data, the horizontal high-frequency portion includes horizontal edge information of the two-dimensional image data, the vertical high-frequency portion includes vertical edge information of the two-dimensional image data, and the diagonal high-frequency portion includes diagonal edge information of the two-dimensional image data.

[0010] More preferably, the discrete wavelet transform layer uses a wavelet base whose wavelet base type is db5, and the boundary processing mode used is periodic boundary conditions.

[0011] More preferably, the artificial current source electromagnetic data classification model is the IncepTCN classification network model.

[0012] According to a second aspect, there is provided a system for suppressing strong interference in artificial current source electromagnetic data, comprising: a data classification module for inputting the artificial current source electromagnetic data to be processed into an artificial current source electromagnetic data classification model to obtain first noise-free artificial current source electromagnetic data and noise-containing artificial current source electromagnetic data; a data denoising module for converting the noisy artificial current source electromagnetic data into two-dimensional image data by a dimension transformation function and inputting the data into the DWTSC-UNet denoising network model to obtain a denoised data result; a dimension transformation module for performing a dimension inverse transformation on the denoised data result using a dimension transformation function to obtain second noise-free artificial current source electromagnetic data; a data combination module for combining the first noise-free artificial-source electromagnetic data and the second noise-free artificial-source electromagnetic data to obtain complete noise-free artificial-source electromagnetic data; Including, The DWTSC-UNet denoising network model is a U-structured network that combines a discrete wavelet transform, a spatial attention mechanism, and a channel attention mechanism; The discrete wavelet transform provides high-frequency and low-frequency signals, and is used to perform weighting learning on different frequency components of the two-dimensional image data and extract multi-scale features from the two-dimensional image data, a channel attention mechanism is used to perform weighting on different dimensions of the feature map, and a spatial attention mechanism is used to perform weighting on different spatial locations of the feature map.A system for suppressing strong interference in artificial current source electromagnetic data is provided.

[0013] More preferably, the method further includes a denoising model training module, The denoising model training module: a measured noise signal acquisition unit for selecting noisy artificial current source electromagnetic data samples from the measured data, and performing noise removal using a dictionary learning method to separate the measured noise signals; a one-dimensional denoising sample library construction unit for adding the measured noise signal and the analog noise signal to the noise-free artificial current source electromagnetic data to form noisy artificial current source electromagnetic data, which is combined with the noise-free artificial current source electromagnetic data sample to form a sample pair, thereby constructing a one-dimensional denoising sample library; a dimension transformation unit for segmenting the one-dimensional denoising sample library and performing a dimension transformation on each sample according to a dimension transformation function to obtain a two-dimensional image sample library; a model training unit for inputting a 2D image sample library into the DWTSC-UNet denoising network model for training; Includes.

[0014] More preferably, the DWTSC-UNet denoising network model in the data denoising module includes an input layer, a hidden layer and an output layer, and the hidden layer includes a discrete wavelet transform layer, a first shape reconstruction layer, a first convolution block, a first spatial attention mechanism module, a second convolution block, a second spatial attention mechanism module, a third convolution block, a first up-sampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second up-sampling block, a second skip connection layer, a second channel attention mechanism module and a third up-sampling block; The discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part and a diagonal high-frequency part, wherein the low-frequency approximation part contains the basic contour of the two-dimensional image data, the horizontal high-frequency part contains horizontal edge information of the two-dimensional image data, the vertical high-frequency part contains vertical edge information of the two-dimensional image data, and the diagonal high-frequency part contains diagonal edge information of the two-dimensional image data.

[0015] More preferably, the discrete wavelet transform layer uses a wavelet base whose wavelet base type is db5, and the boundary processing mode used is periodic boundary conditions.

[0016] More preferably, the artificial current source electromagnetic data classification model in the data classification module is the IncepTCN classification network model.

[0017] According to a third aspect, there is provided an electronic device, the electronic device comprising at least one memory for storing a program and at least one processor for executing the program stored in the memory, wherein execution of the program stored in the memory causes the processor to be used for performing a method according to the first aspect or any one possible embodiment of the first aspect.

[0018] According to a fourth aspect, there is provided a computer readable storage medium having stored thereon a computer program which, when running on a processor, causes the processor to perform a method according to the first aspect or any one possible embodiment thereof.

[0019] According to a fifth aspect, there is provided a computer program product, which when running on a processor causes the processor to perform a method according to the first aspect or any one possible embodiment of the first aspect.

[0020] For the beneficial effects of the second to fifth aspects, please refer to the related explanation of the first aspect.

[0021] Compared with the prior art, the above technical solution conceived in the present invention has the following beneficial effects:

[0022] The method for suppressing strong interference from artificial current source electromagnetic data according to the present invention introduces a dimension transformation method to transform the problem of artificial current source electromagnetic signal denoising into an image denoising problem, and through the proposed signal-to-image transformation method, convert the one-dimensional artificial current source electromagnetic time series into two-dimensional data, thereby preserving the structural features of the original artificial current source electromagnetic signal.

[0023] The method for suppressing strong interference from artificial current source electromagnetic data according to the present invention combines the discrete wavelet transform with a UNet network, fully utilizing the advantages of the discrete wavelet transform and UNet architecture. First, the discrete wavelet transform performs multi-scale feature extraction on the artificial current source electromagnetic data, extracting different frequency information and helping the network better understand the content of the artificial current source electromagnetic data. Second, the combination of the discrete wavelet transform with the UNet architecture enables more effective feature fusion between the encoder and decoder of the model, improving network performance. Furthermore, UNet itself has good local feature capture capabilities, and features extracted by the encoder can be directly transmitted to the decoder using skip connections. The discrete wavelet transform further enhances this capability, helping the model better capture local structural information of the image, especially at the image detail level. Furthermore, the multi-scale representation in which the discrete wavelet transform compresses the high dimensionality of the image into a low dimensional one reduces computational complexity and improves network processing efficiency.

[0024] The present invention provides a method for suppressing strong interference in electromagnetic data from artificial current sources. It makes full use of the attention mechanism to improve the model feature extraction ability, the multi-scale feature extraction ability of the discrete wavelet transform, and the unique encoding-decoding structure of the UNet network, skip connection, and effective integration ability of local and global information. It provides a new type of denoising network, namely, the DWTSC-UNet denoising network, which realizes strong noise suppression of electromagnetic data from artificial current sources without circulating the effective signal. The apparent resistivity curve calculated from the denoised data is obviously improved, and the curve is more smoothly continuous.

[0025] The present invention provides a method for suppressing strong interference in electromagnetic data from artificial current sources. After the model training is completed, the processes of data processing such as identification and noise removal are all completed fully automatically by the computer, without any manual intervention, and do not require the data processing operator to have experience. This not only solves the problem of subjective deviation caused by the need to manually set the threshold in the traditional method, but also improves adaptability to different types of noise. [Brief explanation of the drawings]

[0026] [Figure 1] 2 is a flowchart of a method for suppressing strong interference in artificial current source electromagnetic data according to an embodiment of the present invention. [Figure 2] FIG. 2 is a diagram showing a one-dimensional noise removal sample library according to an embodiment of the present invention. [Figure 3] FIG. 10 is a schematic diagram illustrating a dimension transformation function according to an embodiment of the present invention. [Figure 4] FIG. 1 is an architecture diagram of a WDTSC-UNet denoising network model according to an embodiment of the present invention. [Figure 5] FIG. 2 is an exploded view of discrete wavelet transform data according to an embodiment of the present invention. [Figure 6] FIG. 10 is a graph showing the change in the loss curve during the training process of the WDTSC-UNet denoising network model according to an embodiment of the present invention. [Figure 7] FIG. 1 is a diagram of a classification sample library according to an embodiment of the present invention. [Figure 8] FIG. 1 is an architecture diagram of an IncepTCN classification network model according to an embodiment of the present invention. [Figure 9] FIG. 10 is a graph showing the change in accuracy and loss curves during the training process of the IncepTCN classification network model according to an embodiment of the present invention. [Figure 10] FIG. 1 is a confusion matrix diagram of an IncepTCN classification network model according to an embodiment of the present invention. [Figure 11]FIG. 2 is a diagram illustrating the classification effect of the IncepTCN classification network model according to an embodiment of the present invention; [Figure 12] FIG. 10 is a diagram illustrating the noise reduction effect of the DWTSC-UNet model according to the embodiment of the present invention. [Figure 13] FIG. 10 is a diagram showing a U / I curve and apparent resistivity curve of noise-removed data and data calculation after noise removal according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0027] In order to clarify the objectives, technical solutions and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only for the purpose of illustrating the present invention and are not intended to limit the present invention.

[0028] In this specification, "and / or" is a related relationship that describes related objects, and indicates that three types of relationships may exist. For example, A and / or B can indicate three cases: A exists alone, A and B exist simultaneously, or B exists alone. Here, the symbol " / " means that the related object is either A or B.

[0029] In this specification and claims, the terms "first" and "second" and the like are not intended to specify a particular order for identifying objects, but are intended to distinguish between different objects.

[0030] In the embodiments of the present invention, terms such as "exemplary" or "for example" are used to exemplify, illustrate, or explain. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present invention should not be construed as preferred or superior over other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present related concepts.

[0031] In describing embodiments of the present invention, unless otherwise specified, "plurality" means two or more than two.

[0032] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0033] As shown in FIG. 1 , an embodiment of the present invention provides a method for suppressing strong interference in artificial current source electromagnetic data of SC-UNet network based on dimension transformation and discrete wavelet transform optimization, which specifically includes the following steps: Step 1: Selecting highly discriminative, typical, and noisy artificial current source electromagnetic data from a large amount of actual measurement data. Step 2: Create a denoising sample library, denoise the measured noisy samples using a dictionary learning method, add the separated noise to the high-quality artificial current source electromagnetic data to create some noisy data samples, and create some noise samples by adding artificially synthesized analog noise to the high-quality data, and construct a one-dimensional denoising sample library using the noisy artificial current source electromagnetic data samples and the corresponding high-quality artificial current source electromagnetic data samples as sample pairs. Step 3: Segment the denoised sample library constructed in step 2, and perform dimensional transformation on each sample using a dimensional transformation function to convert the one-dimensional data into a two-dimensional matrix, i.e., obtain a two-dimensional sample library. Step 4: Construct a DWTSC-UNet denoising network model, and input the 2D sample library obtained in step 3 into the DWTSC-UNet denoising network model for training to obtain an artificial current source electromagnetic data denoising model. Step 5: Build a classification sample library with each data segment in step 1 as one sample and a category tag for the mark sample. The category tag indicates whether the sample is a noise data segment or a high-quality data segment. Step 6: Construct an IncepTCN classification network, input the classification sample library into the IncepTCN classification network, and train it to obtain an artificial current source electromagnetic data classification model. Step 7: The artificial current source electromagnetic data to be processed is input and segmented, and then input into the artificial current source electromagnetic data classification model to obtain the classification results for each data segment, and the data segments recognized by the model as being of high quality are retained. Step 8: The artificial current source electromagnetic data segments recognized as containing noise by the classification model in Step 7 are input into the artificial current source electromagnetic data denoising model to obtain the data denoising results for each data segment. Step 9: Using the dimension transformation function in step 3, a dimension inverse transformation is performed on the noise removal result obtained in step 8, and the two-dimensional image data is transformed into a one-dimensional time series. Step 10: The high-quality data retained in step 7 is combined with the data obtained after dimension transformation of the denoised data in step 9 to obtain a complete artificial current source electromagnetic noise denoised sequence.

[0034] More preferably, the dimension transformation function is a data dimension transformation method, and the input one-dimensional data length is 1600, so

number

number

number

number

[0035] More preferably, wavelet transform (WT) is a mathematical tool used in signal analysis and processing, which can decompose a signal into different frequency components. Specifically, the present invention uses a two-dimensional discrete wavelet transform to extract local information of a signal by decomposing the signal using a set of wavelet basis functions. For a two-dimensional matrix, the two-dimensional discrete wavelet transform decomposes it into the following four components: LL (low-frequency approximation part): contains the low-frequency information of the matrix, i.e., the approximate contour of the matrix; LH (horizontal high-frequency part): contains the horizontal edge information of the matrix; HL (vertical high-frequency part): contains the vertical edge information of the matrix; and HH (diagonal high-frequency part): contains the diagonal edge information of the matrix. One-dimensional discrete wavelet transform (1D-DWT) is a convolution of a one-dimensional signal with a wavelet function.

number

number

[0036] The 2D discrete wavelet transform is an extension of the 1D transform. In the 2D case, first, a 1D discrete wavelet transform is performed on each row of the matrix, and then a 1D-DWT is performed on each column. To convert 1D data into a 2D matrix using a dimensional transformation function, we denote it as f(x,y), where x and y represent spatial coordinates in the horizontal and vertical directions, respectively. To perform a wavelet transform on f(x,y), we use

number

number

number

[0037] More preferably, the DWTSC-UNet denoising network includes an input layer, a hidden layer, and an output layer, where the hidden layer is composed of one discrete wavelet transform layer, two spatial attention mechanism layers, three convolution blocks, two channel attention mechanism layers, three upsampling convolution blocks, and two connection layers.

[0038] More preferably, the two-dimensional convolution layers in the three convolution blocks use convolution kernels with different numbers of channels and the same size. The 4-channel 3x3 convolution path extracts low-level features of an image using 4 channels and 3x3 convolutions. The 32-channel 3x3 convolution path extracts higher-level features at a more abstract level using 32 channels and 3x3 convolutions. The 64-channel 3x3 convolution path extracts more complex high-level features using 64 channels and 3x3 convolutions.

[0039] 2x2 max pooling path: Downsampling is performed by max pooling with a 2x2 pooling window, stride 2, to reduce the spatial size (width and height).

[0040] More preferably, the three upsampling blocks are upsampled using different numbers of channels, the same size convolution kernel, 128 channels, and 2x2 convolution paths (128 channels, stride 2, size 2x2 convolution) to restore the spatial resolution of the image.

[0041] 128-channel 2x2 convolution path: Upsample using a 128-channel, stride 2, size 2x2 convolution to further restore spatial resolution of the image. 96-channel 2x2 convolution path: Upsample using a 96-channel, stride 2, size 2x2 deconvolution to improve the spatial resolution of the image and recover the image size.

[0042] More preferably, in the two spatial attention mechanism block layers, spatial information is captured by the same size convolution kernel, 7x7 convolution path (7x7 convolution).

[0043] More preferably, two 1x1 convolutions are used in the two channel attention mechanism block layer: the first 1x1 convolution reduces the number of input channels, and the second 1x1 convolution restores the number of channels.

[0044] The architecture of the denoising network is an improved UNet network, i.e., the DWT-CSUNet denoising network. Compared to conventional UNet networks, the DWT-CSUNet denoising network is improved by combining the advantages of discrete wavelet transform, spatial attention mechanism, channel attention mechanism, and conventional U-shaped network. This invention introduces the discrete wavelet transform. First, the discrete wavelet transform can simultaneously provide high- and low-frequency information of a signal through multi-scale feature extraction, helping the network extract richer hierarchical information from the data. Second, weighted learning for different frequency components of the image improves sensitivity to details and edges, improving the network's ability to represent complex signals. Furthermore, the discrete wavelet transform serves as a feature preprocessing means, optimizing data before network training, thereby reducing the network's computational complexity during the training process and accelerating network training. Finally, for complex tasks, the discrete wavelet transform helps neural networks capture a wider range of modes and enhance the model's generalization ability by extracting multi-scale features of the signal. The present invention also introduces a channel attention mechanism module and a spatial attention mechanism module. The internal structure of the modules is simple, involving a small amount of convolution, pooling, and feature fusion operations, resulting in low computational complexity and saving computational resources. The channel attention mechanism utilizes weighting for different channels (i.e., different dimensions of the feature map), assigning larger weights to important channels to highlight important features and strengthen the network's attention to key features, while assigning smaller weights to unimportant channels to suppress unimportant features and reduce redundant information. The spatial attention mechanism weights different spatial locations of the feature map, helping the network focus on important regions or targets in the image and ignoring background or irrelevant parts.Second, by giving higher weights to specific spatial regions, finer spatial information can be captured, further improving the network's denoising accuracy. In denoising artificial source electromagnetic data, this means that the DWTSC-UNet can more closely observe and understand the characteristics of the artificial source electromagnetic signal. This is very beneficial for accurately denoising artificial source electromagnetic data, since noise may exist at different scales.

[0045] The UNet network plays an important role in the DWTSC-UNet network. The DWTSC-UNet network as a whole employs an "encoder-decoder" model. The encoder (downsampling part) gradually reduces image resolution through convolution and pooling operations to extract increasingly abstract features. The decoder (upsampling part) gradually restores the spatial resolution and image size of the image. Skip connections are also introduced into the UNet network, allowing feature maps from the encoder stage to be directly transmitted to the corresponding layers in the decoder stage. This directly transmits high-resolution feature maps, avoiding the loss of spatial information caused by pooling operations and helping the model better process information such as edges, details, and textures in images. The skip connections allow gradients to be more easily propagated to earlier levels of the network during backpropagation, accelerating the training process and avoiding the problem of vanishing gradients. The DWTSC-UNet network also fully utilizes the advantages of the UNet backbone network, discrete wavelet transform, and attention mechanism modules. By combining the UNet backbone network, discrete wavelet transform, and attention mechanism module, the DWTSC-UNet network can capture important information and extract local information and multi-scale features, thereby achieving a more comprehensive understanding of artificial current source electromagnetic data. By integrating and utilizing different modules in this way, the DWTSC-Unet network is more powerful and effective in extracting important features, retaining detailed features, and adapting and representing complex signals. Therefore, the DWTSC-UNet network of the present invention is advantageous in that it combines the attention mechanism module and discrete wavelet transform according to the structure of the UNet backbone network to extract features of artificial current source electromagnetic data and realize noise suppression of artificial current source electromagnetic data.

[0046] Based on the multi-scale feature extraction ability of discrete wavelet transform, the feature concentration property of attention mechanism and the unique advantages of two-dimensional data processing by UNet network, this invention combines discrete wavelet transform, attention mechanism and UNet network to optimize the UNet network structure, improve the network performance, and introduces an improved UNet network to solve the strong noise suppression problem of artificial current source electromagnetic, thereby improving the noise suppression accuracy.

[0047] More preferably, the discrete wavelet transform layer in the DWTSC-Unet denoising network adopts a wavelet base type of “db5”, i.e., Daubechies wavelet base, and the adopted boundary processing mode is “periodic” boundary processing, i.e., periodic boundary condition.

[0048] Example 1 The present invention provides a method and system for suppressing strong interference in artificial current source electromagnetic data based on SC-UNet (DWTSC-UNet) network optimized by dimension transformation and discrete wavelet transform, and solves the method for suppressing strong interference in artificial current source electromagnetic data by introducing a deep learning method optimized by dimension transformation and discrete wavelet transform, specifically including the following steps:

[0049] Step 1: Acquire artificial current source electromagnetic data and segment them using time windows, with each data segment being one sample.

[0050] Step 2: For typical real noisy data segments, data denoising is performed using dictionary learning (SISC) algorithm to obtain a reliable denoising effect, and the noisy data segments are converted into a reliable noise-free state, which is used to create a high-quality one-dimensional sample library for the denoising model.

[0051] As shown in FIG. 2, an embodiment of the present invention provides a one-dimensional denoising sample library, showcasing four typical samples. The one-dimensional sample library includes noise samples of different types and amplitudes. The noise samples include Gaussian noise signals, pulse noise signals, square wave noise signals, and mixed noise signals. To effectively distinguish between useful signals and noise and ensure good generalization ability of the network, the present invention requires providing as many useful signal and noise samples as possible during the training process to allow the network to learn more features. The sample library of the present invention contains 48,000 sample pairs, where the training set and validation set contain 42,000 sample pairs and 6,000 sample pairs, respectively, with a training set to validation set ratio of 7:1.

[0052] In an embodiment of the present invention, prior to model training, the sample data is normalized to reduce errors introduced by sample amplitude differences.

number

number

[0053] Step 3: Segment the one-dimensional sample library obtained in step 2, and use the dimension transformation function to transform each sample into a two-dimensional matrix to obtain a two-dimensional sample library.

[0054] FIG. 3 is a schematic diagram of dimension conversion using a dimension conversion function. The length of the sequence of sampling points of each sample is 1600, and the sequence is converted into a 40×40 matrix shape using the dimension conversion function.

[0055] In an embodiment of the present invention, the dimension transformation function performs a dimension transformation in an "S" shape, i.e., for odd-numbered rows (e.g., row 1, row 3, etc.), data is filled in the forward direction, and for even-numbered rows (e.g., row 2, row 4, etc.), data is filled in the reverse direction.

[0056] Step 4: Construct a DWTSC-UNet denoising network, input the constructed 2D sample library into the DWTSC-UNet denoising network, and train it to obtain an artificial current source electromagnetic data denoising model.

[0057] Figure 4 shows the DWTSC-UNet denoising network, which consists of three parts: input layer, hidden layer, and output layer.

[0058] In the embodiment of the present invention, the input layer undergoes dimensional transformation using a dimensional transformation function, converting the 1x1600 size to 1x40x40, and adding a batch size of 256, resulting in an input layer shape of (256,1,40,40).

[0059] In an embodiment of the present invention, the hidden layer is composed of a total of 15 parts, including a discrete wavelet transform layer, a first shape reconstruction layer, a first convolution block, a first spatial attention mechanism module, a second convolution block, a second spatial attention mechanism module, a third convolution block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block. The calculation process of the discrete wavelet transform layer, convolution block, spatial attention mechanism module, channel attention mechanism module, upsampling block, skip connection layer, and shape reconstruction layer in an embodiment of the present invention is as follows:

[0060] In the discrete wavelet transform layer step, the shape (256,1,40,40) is first input, and each image is decomposed into a low-frequency part (LL) and three high-frequency parts (LH, HL, HH), i.e., four channels, by discrete wavelet transform. Then, the image shape is transformed into (256,4,24,24) using the "db5" wavelet base and "periodic" boundary processing mode, and the result of the discrete wavelet transform layer is output.

[0061] In the first shape reconstruction layer step, the result of the discrete wavelet transform layer is input, and the output size of the discrete wavelet transform layer is adjusted to the same size as the input layer, i.e., (256, 4, 40, 40), using the bilinear interpolation method, and the result of the shape reconstruction layer is output.

[0062] In the first convolution block step, first, the result of the first shape reconstruction layer is input, and 32 convolution kernels with a size of 3x3, a stride of 1, and a padding of 1 are used to perform two-dimensional convolution on the result of the first shape reconstruction layer, and then activated using the "ReLU" activation function; second, a batch normalization layer is used to perform normalization on each channel data; third, two-dimensional max pooling is performed on the normalized result using a pooling kernel with a pooling window size of 2x2 and a stride of 2 to halve the spatial size, i.e., (256, 32, 20, 20), and the result of the first convolution block is output.

[0063] The steps of the first spatial attention mechanism module are as follows: First, input the results of the first convolution block and use average pooling to calculate the channel mean value for each position in the results of the first convolution block, resulting in a result of shape (256,1,20,20). Second, use max pooling to calculate the channel maximum value for each position in the results of the first convolution block, resulting in a result of shape (256,1,20,20). Third, stitch the results of average pooling and max pooling along the channel dimension, resulting in a result of shape (256,2,20,20). Fourth, use two convolution kernels with size 7x7, stride 1, and padding 3 to stitch the results, resulting in a 2D convolution with shape (256,1,20,20), and activate using the "Sigmoid" activation function. Finally, we perform element-wise multiplication of the result of the first convolution block and the result after "Sigmoid" activation, resulting in a shape of (256, 32, 20, 20). We output the result of the first spatial attention mechanism module.

[0064] In the second convolution block step, first, the result of the first spatial attention mechanism module is input, and 64 convolution kernels with a size of 3x3, a stride of 1, and a padding of 1 are used to perform two-dimensional convolution on the result of the first spatial attention mechanism module, and activation is performed using the "ReLU" activation function; second, a batch normalization layer is used to normalize each channel data; third, a pooling kernel with a pooling window size of 2x2 and a stride of 2 is used to perform two-dimensional max pooling on the normalized result, which halves the spatial size, i.e., (256,64,10,10), and outputs the result of the second convolution block.

[0065] In the spatial attention mechanism module step, first, the result of the second convolution block is input, and the channel average value of each position in the result of the second convolution block is calculated using average pooling, resulting in a shape of (256,1,10,10); second, the channel maximum value of each position in the result of the second convolution block is calculated using max pooling, resulting in a shape of (256,1,10,10); third, the results of average pooling and max pooling are stitched along the channel dimension, resulting in The shape is (256,2,10,10). Fourth, after stitching using two convolution kernels with size 7x7, stride 1, and padding 3, 2D convolution is performed on the result, which has shape (256,1,10,10). It is activated using the "Sigmoid" activation function. Finally, the result of the second convolution block and the result after "Sigmoid" activation are multiplied element-wise, which has shape (256,64,10,10). This outputs the result of the second spatial attention mechanism module.

[0066] In the third convolution block step, first, the result of the second spatial attention mechanism module is input, and 128 convolution kernels with a size of 3x3, a stride of 1, and a padding of 1 are used to perform two-dimensional convolution on the result of the second spatial attention mechanism module, and activation is performed using the "ReLU" activation function; second, a batch normalization layer is used to normalize each channel data; third, a pooling kernel with a pooling window size of 2x2 and a stride of 2 is used to perform two-dimensional max pooling on the normalized result, thereby halving the spatial size, i.e., (256,128,5,5), and output the result of the third convolution block.

[0067] In the step of the first upsampling block, first, the result of the third convolution block is input, and two-dimensional deconvolution is performed on the result of the third convolution block using 64 convolution kernels with a size of 3x3 and a stride of 2 to double the spatial size, i.e., to (256,64,10,10), and the result of the upsampling block is output.

[0068] In the step of the second shape reconstruction layer, the result of the first upsampling block is input, and the output size of the upsampling block is adjusted to the same size as the input layer, i.e., (256, 64, 10, 10), using the bilinear interpolation method, and the result of the second shape reconstruction layer is output.

[0069] In the step of the first skip connection layer, the results of the first shape reconstruction layer and the first upsampling block are input, and the result of stitching the results of the second shape reconstruction layer and the first upsampling block is a shape (256, 128, 10, 10), and the result of the first skip connection layer is output.

[0070] In the step of the first channel attention mechanism layer, first, the result of the first skip connection layer is input, and average pooling is used to perform spatial average pooling on the result of the first skip connection layer, i.e., pooling is performed in the height and width dimensions, resulting in a shape of (256,128,1,1); second, max pooling is used to perform max pooling in the height dimension on the result of the first skip connection layer, resulting in a shape of (256,128,1,10); further, max pooling is performed in the width dimension on the result of max pooling in the height dimension, resulting in a shape of (256,128,1,1); third, the results of average pooling and max pooling are stitched in the channel dimension, resulting in a shape of (256 ,256,1,1), fourth, after stitching using 256 convolution kernels with a size of 1x1 and a stride of 1, 2D convolution is performed on the result after stitching, resulting in a shape of (256,8,1,1) and activation is performed using the "ReLU" activation function; fifth, after "ReLU" activation using 128 convolution kernels with a size of 1x1 and a stride of 1, 2D convolution is performed on the result after "ReLU" activation, resulting in a shape of (256,128,1,1), and activation is performed using the "Sigmoid" activation function; finally, the result of the first skip connection layer is multiplied by the result after "Sigmoid" activation, resulting in a shape of (256,128,10,10), which outputs the result of the first channel attention mechanism layer.

[0071] In the second upsampling block step, the result of the first channel attention mechanism layer is input, and 64 convolution kernels with a size of 3x3 and a stride of 2 are used to perform two-dimensional deconvolution on the result of the first channel attention mechanism layer to double the spatial size, i.e., (256,64,20,20), and output the result of the second upsampling block.

[0072] In the step of the second skip connection layer, the results of the first convolution block and the second upsampling block are input, and the result of stitching the results of the first convolution block and the second upsampling block has a shape of (256, 96, 20, 20), and the result of the second skip connection layer is output.

[0073] In the step of the second channel attention mechanism module, first, the result of the second skip connection layer is input, and average pooling is used to perform spatial average pooling on the result of the first skip connection layer, i.e., pooling in the height and width dimensions results in a shape of (256,96,1,1); second, max pooling is used to perform max pooling in the height dimension on the result of the second skip connection layer, resulting in a shape of (256,96,1,20); and further max pooling is performed in the width dimension on the result of the max pooling in the height dimension, resulting in a shape of (256,96,1,1); third, the results of average pooling and max pooling are stitched in the channel dimension to obtain The shape is (256,192,1,1). Fourth, stitching is performed using 256 convolution kernels with a size of 1x1 and a stride of 1, resulting in a shape of (256,6,1,1). Activation is performed using the "ReLU" activation function. Fifth, 2D convolution is performed on the result after "ReLU" activation using 128 convolution kernels with a size of 1x1 and a stride of 1, resulting in a shape of (256,96,1,1). Activation is performed using the "Sigmoid" activation function. Finally, the result of the first skip connection layer is multiplied by the result after "Sigmoid" activation, resulting in a shape of (256,96,20,20). This outputs the result of the second channel attention mechanism layer.

[0074] In the third upsampling block step, the result of the second channel attention mechanism layer is input, and 64 convolution kernels with a size of 3x3 and a stride of 2 are used to perform two-dimensional deconvolution on the result of the second channel attention mechanism layer to double the spatial size, i.e., (256, 1, 40, 40), and output the result of the third upsampling block.

[0075] It should be noted that the above number of upsampling blocks, the number of spatial attention mechanism modules, the number of channel attention mechanism modules, the number of skip connection layers, and the number and size of convolution kernels are set based on the effect of model training. Therefore, the above embodiments are merely illustrative. The number of sample blocks, the number of spatial attention mechanism modules, the number of channel attention mechanism modules, the number of skip connection layers, and the size and number of convolution kernels can be adjusted without departing from the concept of the present invention.

[0076] More preferably, the Adam optimizer is used in the training process of the denoising model, where the training and validation batch sizes are both 256, and the initial learning rate is 1×10 -5 The learning rate decay method is used to decay the learning rate by 0.1 every 10 training rounds, for a total of 500 training rounds. In other possible embodiments, other optimizers may be selected without being specifically limited thereto.

[0077] Figure 5 shows a schematic diagram of the decomposition of the discrete wavelet transform multi-scale features. The two-dimensional matrix is ​​decomposed into four frequency components by the discrete wavelet transform, including a low-frequency part (LL) and three high-frequency parts (LH, HL, HH).

[0078] Figure 6 shows the loss change during the training process of the DWTSC-UNet denoising network model. Here, the solid line represents the loss change during model training on the validation set, and the dotted line represents the loss change on the training set after the model is trained on the training set. The shape of the curve shows that as the number of training sessions increases, the model's loss value gradually decreases and eventually approaches stability. This indicates that the data features learned by the model are trending from increase to convergence, the model's adaptability to the data is increasing, and the model error is decreasing.

[0079] Step 5: Construct a classification sample library with each data segment in step 1 as one sample and a category tag of the mark sample, where the category tag indicates whether the sample is a noise data segment or a high-quality data segment.

[0080] 7 shows a classification sample library according to the present invention, showing 16 typical samples in the sample library. The first column is a measured sample of relatively high quality, the second column is a measured noise-containing sample, the third column is a simulated noise-removed sample, and the fourth column is a simulated noise-containing sample. There are many types of noise in the control source signal noise, such as Gaussian noise, harmonic noise, square wave noise, impulse noise, and baseline drift noise.

[0081] Step 6: Construct an IncepTCN classification network, input the classification sample library into the IncepTCN classification network, and train it to obtain an artificial current source electromagnetic data classification model.

[0082] Figure 8 shows the structure of the IncepTCN data classification model constructed in this invention. The IncepTCN data classification model consists of three parts: an input layer, a hidden layer, and an output layer. The input layer has a shape equal to the number of channels multiplied by the sample length, i.e., 1 × 1600.

[0083] In an embodiment of the present invention, the hidden layer is composed of a total of 13 parts, including a convolution layer, a first multi-scale convolution module, a first time-domain convolution block, a first max pooling layer, a second multi-scale convolution module, a second time-domain convolution block, a second max pooling layer, a third multi-scale convolution module, a third time-domain convolution block, a third max pooling layer, a fourth multi-scale convolution module, a fourth time-domain convolution block, and a fourth max pooling layer.

[0084] In the convolutional layer step, one-dimensional convolution is performed on the training samples using 16 convolution kernels of size 16x1, and activation is performed using "ReLU", and the results of the convolutional layer are output.

[0085] In the first multi-scale convolution module step, the results of the convolution layer are input, and the four channels of the first multi-scale convolution module are respectively 13 convolution paths (using 13 convolutions to extract local temporal and spatial features), 15 convolution paths (using 15 convolutions to extract medium-range temporal and spatial features), 17 convolution paths (using 17 convolutions to extract wider-range temporal and spatial features), and 13 max-pooling channels (using 13 max-pooling to fuse local-range features). Finally, the outputs of the four channels are connected, and the final output of the first multi-scale convolution module realizes multi-scale feature extraction, and is the result of the first multi-scale convolution module.

[0086] In the first time-domain convolution block step, first, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution and one-dimensional convolution on the results of the first multi-scale convolution module, respectively. The extended causal convolution is activated using ReLU, and the one-dimensional convolution is not activated. Second, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution on the activation value of the first extended causal convolution of the residual block. Third, the one-dimensional convolution of the residual block and the one-dimensional extended causal convolution of the residual block that is not activated for the second time are summed, activated using ReLU, and the result of the first time-domain convolution block is output.

[0087] In the first max pooling layer step, the result of the first time-domain convolution block is input, and a max pooling operation with size 2 and stride 2 is used for screening, and the result of the first max pooling layer is output.

[0088] In the second multi-scale convolution module step, the results of the first max pooling layer are input, and the four channels of the second multi-scale convolution module are respectively 13 convolution paths (using 13 convolutions to extract local temporal and spatial features), 15 convolution paths (using 15 convolutions to extract medium-range temporal and spatial features), 17 convolution paths (using 17 convolutions to extract wider-range temporal and spatial features), and 13 max pooling paths (using 13 max pooling to fuse local ranges). Finally, the outputs of the four paths are connected, and the final output of the second multi-scale convolution module realizes multi-scale feature extraction, and is the result of the second multi-scale convolution module.

[0089] In the second time-domain convolution block step, first, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution and one-dimensional convolution on the results of the second multi-scale convolution module, respectively. The extended causal convolution is activated using ReLU, and the one-dimensional convolution is not activated. Second, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution on the activation value of the first extended causal convolution of the residual block. Third, the one-dimensional convolution of the residual block and the one-dimensional extended causal convolution of the residual block that is not activated for the second time are summed, activated using ReLU, and the result of the second time-domain convolution block is output.

[0090] In the second max pooling layer step, the result of the second time-domain convolution block is input, and a max pooling operation with a size of 2 and a stride of 2 is used for screening, and the result of the max pooling layer is output.

[0091] In the third multi-scale convolution module step, the results of the second max pooling layer are input, and the four paths of the third multi-scale convolution module are respectively 13 convolution path (using 13 convolutions to extract local temporal and spatial features), 15 convolution path (using 15 convolutions to extract medium-range temporal and spatial features), 17 convolution path (using 17 convolutions to extract wider-range temporal and spatial features), and 13 max pooling path (using 13 max pooling to fuse local-range features). Finally, the outputs of the four paths are connected, and the final output of the third multi-scale convolution module realizes multi-scale feature extraction, and is the result of the third multi-scale convolution module.

[0092] In the third time-domain convolution block step, first, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution and one-dimensional convolution on the results of the third multi-scale convolution module, respectively. The extended causal convolution is activated using ReLU, and the one-dimensional convolution is not activated. Second, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution on the activation value of the first extended causal convolution of the residual block. Third, the one-dimensional convolution of the residual block and the one-dimensional extended causal convolution of the second residual block that is not activated are summed, activated using ReLU, and the result of the third time-domain convolution block is output.

[0093] In the third max pooling layer step, the result of the third time-domain convolution block is input, and screening is performed using a max pooling operation with a size of 2 and a stride of 2, and the result of the third max pooling layer is output.

[0094] In the fourth multi-scale convolution module step, the results of the third max pooling layer are input, and the four paths of the fourth multi-scale convolution module are respectively 13 convolution path (using 13 convolutions to extract local temporal and spatial features), 15 convolution path (using 15 convolutions to extract medium-range temporal and spatial features), 17 convolution path (using 17 convolutions to extract wider-range temporal and spatial features), and 13 max pooling path (using 13 max pooling to fuse local range features). Finally, the outputs of the four paths are connected, and the final output of the fourth multi-scale convolution module realizes multi-scale feature extraction and is the result of the fourth multi-scale convolution module.

[0095] In the fourth time-domain convolution block step, first, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution and one-dimensional convolution on the results of the third multi-scale convolution module, respectively. The extended causal convolution is activated using "ReLU" and the one-dimensional convolution is not activated. Second, 64 convolution kernels with a size of 3x1 are used to perform one-dimensional extended causal convolution on the activation value of the first extended causal convolution of the residual block. Third, the one-dimensional convolution of the residual block and the one-dimensional extended causal convolution of the residual block that is not activated for the second time are summed, activated using "ReLU", and the result of the fourth time-domain convolution block is output.

[0096] In the fourth max pooling layer step, the result of the residual module 12 is input, and screening is performed using a max pooling operation with a size of 2 and a stride of 2, and the result of the max pooling layer 13 is output.

[0097] It should be noted that the number of multi-scale convolutional modules, the number of residual modules, and the number and size of convolutional kernels are set based on the model training effect. Therefore, the above embodiment is merely illustrative. The number of multi-scale convolutional modules, the number of residual modules, and the number and size of convolutional kernels can be adjusted without departing from the concept of the present invention.

[0098] Finally, the output layer of IncepTCN consists of one flattened layer and one fully connected layer. The flattened layer unfolds the result of the max pooling layer into a one-dimensional array of 196,000. The fully connected layer contains two neurons and is activated using the "Softmax" function. The output layer obtains the classification result of the sample (i.e., noise or high quality) by outputting the category probability.

[0099] In the training process of the classification model in this embodiment, the Adam optimizer is selected, where the batch size for training and validation is both 256, and the initial learning rate is 1×10 -5and train for a total of 30 epochs. In other possible embodiments, other optimizers may be selected without being specifically limited thereto.

[0100] Figure 9 shows the change in accuracy and loss during the training process of the IncepTCN classification network model. Here, the solid line shows the change in accuracy and loss during the model training process for the validation set. The dotted line shows the change in accuracy and loss after the model is trained on the training set. Looking at the shape of the curves, as the number of training sessions increases, the model accuracy gradually increases and the model loss value gradually decreases, eventually stabilizing. This indicates that the data features learned by the model are trending from increase to convergence, the model's adaptability to the data is increasing, and the model error is decreasing.

[0101] Step 7: Input the artificial current source electromagnetic data to be processed, segment it, and then input it into the artificial current source electromagnetic data classification model to obtain the classification results for each data segment, and retain the data segments recognized by the model as being of high quality.

[0102] Step 8: The artificial current source electromagnetic data segments recognized as containing noise by the classification model in Step 7 are input into the artificial current source electromagnetic data denoising model to obtain the data denoising results for each data segment.

[0103] Step 9: Using the dimension transformation function in step 3, perform a dimension inverse transformation on the noise removal result obtained in step 8, converting the two-dimensional data into a one-dimensional time series.

[0104] Step 10: The high-quality data retained in step 7 is combined with the data obtained by dimension transformation of the denoised data in step 9 to obtain a complete artificial current source electromagnetic noise denoised sequence.

[0105] Figure 10 shows the confusion matrix (matrix) for the IncepTCN classification model. Here, the abscissa represents the prediction results of the IncepTCN classification model, and the ordinate represents the results of manual marking. A label equal to 0 indicates high quality, while a label equal to 1 indicates the presence of noise. The upper left, upper right, lower left, and lower right parts of each subfigure represent the number of PT, NF, PF, and NT samples, respectively. Darker colors represent larger values. From left to right, they represent signal-dominant, well-balanced signal and noise, and noise-dominant situations, respectively. When high-quality samples dominate, the IncepTCN classification model achieves excellent performance, with an F1 score exceeding 0.99. When the number of high-quality samples and noise samples is balanced, the accuracy of the IncepTCN classification model exceeds 0.99. When noise samples dominate, the IncepTCN classification model also performs well, with an F1 score exceeding 0.95.

[0106] Figure 11 shows the classification effect of the IncepTCN classification model, which shows, from top to bottom, a situation where the signal is dominant, a situation where the signal and noise are well balanced, and a situation where the noise is dominant. As can be seen from Figure 11, the recognition effect of the IncepTCN classification model was accurate, and all noise-containing segments were recognized.

[0107] Figure 12 shows the denoising effect of the DWTSC-UNet model. The dotted line indicates the synthesized noise sample, and the solid line indicates the denoising result by the DWTSC-UNet model. As can be seen from Figure 12, the noise is effectively removed by the DWTSC-UNet model, the entire curve is smooth, and no obvious noise is observed.

[0108] FIG. 13 shows the apparent resistivity and U / I curves before and after data processing. The dashed lines show the apparent resistivity and U / I curves calculated based on the time series of the synthetic noise signal, and the solid lines show the apparent resistivity and U / I curves calculated based on the time series after processing using the method of the present invention. From left to right, the figures show the noise-containing signals with signal-to-noise ratios of 5 dB, -5 dB, -10 dB, and -15 dB. As can be seen from FIG. 13, when the signal-to-noise ratio is 5 dB, the original noise curve is relatively smooth, with only slight distortion at 3 Hz. When the signal-to-noise ratios are -55 dB, -10 dB, and -15 dB, the original noise curve shows significant distortion in the frequency range below 10 Hz. After processing using the method of the present invention, the distorted curve becomes smooth and continuous.

[0109] Example 2 The present invention provides a system for suppressing strong interference in artificial current source electromagnetic data.

[0110] The system includes a data classification module, a data denoising module, a dimensional transformation module, and a data combination module.

[0111] The data classification module is used to input the artificial-source electromagnetic data to be processed into the artificial-source electromagnetic data classification model to obtain first noise-free artificial-source electromagnetic data and noise-containing artificial-source electromagnetic data.

[0112] The data denoising module is used to convert the noise-containing artificial current source electromagnetic data into two-dimensional image data through a dimension transformation function and input it into the DWTSC-UNet denoising network model to obtain the denoised data results.

[0113] The dimension transformation module is used to perform a dimension inverse transformation on the denoised data result using a dimension transformation function to obtain second noise-free artificial current source electromagnetic data.

[0114] The data combination module is used to combine the first noise-free artificial-source electromagnetic data and the second noise-free artificial-source electromagnetic data to obtain complete noise-free artificial-source electromagnetic data.

[0115] Here, the DWTSC-UNet denoising network model is a U-shaped structure network that combines discrete wavelet transform, spatial attention mechanism and channel attention mechanism.

[0116] The discrete wavelet transform provides high-frequency and low-frequency signals, which are used to perform weighting training for different frequency components of the 2D image data and perform multi-scale feature extraction from the 2D image data. The channel attention mechanism is used to assign weights to different dimensions of the feature map. The spatial attention mechanism is used to assign weights to different spatial locations of the feature map.

[0117] More preferably, the system for suppressing strong interference in artificial current source electromagnetic data further includes a denoising model training module.

[0118] The denoising model training module includes an actual noise signal acquiring unit, a one-dimensional denoising sample library building unit, a dimension transforming unit, and a model training unit.

[0119] The measured noise signal acquisition unit selects the artificial current source electromagnetic data samples containing noise from the measured data, and performs noise removal using a dictionary learning method to separate the measured noise signal.

[0120] The one-dimensional denoising sample library construction unit adds the measured noise signal and the analog noise signal to the noise-free artificial current source electromagnetic data to form noise-containing artificial current source electromagnetic data, which is combined with the noise-free artificial current source electromagnetic data sample to form a sample pair, thereby constructing a one-dimensional denoising sample library.

[0121] The dimensional transformation unit segments the one-dimensional denoising sample library and performs a dimensional transformation on each sample using a dimensional transformation function to obtain a two-dimensional image sample library.

[0122] The model training unit inputs a 2D image sample library into the DWTSC-UNet denoising network model for training.

[0123] More preferably, the DWTSC-UNet denoising network model in the data denoising module includes an input layer, a hidden layer, and an output layer, where the hidden layer includes a discrete wavelet transform layer, a first shape reconstruction layer, a first convolution block, a first spatial attention mechanism module, a second convolution block, a second spatial attention mechanism module, a third convolution block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block.

[0124] Here, the discrete wavelet transform layer decomposes the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part, and a diagonal high-frequency part. The low-frequency approximation part contains the basic contour of the two-dimensional image data, the horizontal high-frequency part contains horizontal edge information of the two-dimensional image data, the vertical high-frequency part contains vertical edge information of the two-dimensional image data, and the diagonal high-frequency part contains diagonal edge information of the two-dimensional image data.

[0125] More preferably, the wavelet base type used by the discrete wavelet transform layer is the db5 wavelet, and the boundary processing mode adopted is the periodic boundary condition.

[0126] More preferably, the artificial current source electromagnetic data classification model in the data classification module is the IncepTCN classification network model.

[0127] As described above, the present invention has the following advantages over the prior art. This invention introduces a dimension transformation method to transform the artificial source electromagnetic signal denoising problem into an image denoising problem, and through the proposed signal-to-image transformation method, convert the one-dimensional artificial source electromagnetic time series into two-dimensional data, so as to preserve the structural features of the original artificial source electromagnetic signal.

[0128] This invention introduces a method that combines the discrete wavelet transform with a UNet network, fully utilizing the advantages of the discrete wavelet transform and the UNet architecture. First, the discrete wavelet transform performs multi-scale feature extraction on artificial current source electromagnetic data, extracting different frequency information and helping the network better understand the content of the artificial current source electromagnetic data. Second, combining the discrete wavelet transform with the UNet architecture enables more effective feature fusion between the model's encoder and decoder, improving network performance. Furthermore, UNet itself has good local feature capture capabilities, and features extracted by the encoder can be directly transmitted to the decoder using skip connections. The discrete wavelet transform further enhances this capability, helping the model better capture local structural information of the image, especially at the image detail level. Furthermore, the multi-scale representation in which the discrete wavelet transform compresses the high dimensionality of the image into a low dimensional representation reduces computational complexity and improves the network's processing efficiency.

[0129] The present invention makes full use of the attention mechanism to improve the model feature extraction ability, the multi-scale feature extraction ability of the discrete wavelet transform, and the unique encoding-decoding structure of the UNet network, skip connection, and effective integration ability of local and global information, and provides a new type of denoising network, namely the DWTSC-UNet denoising network, which realizes strong noise suppression of artificial current source electromagnetic data without circulating the effective signal. The apparent resistivity curve calculated from the denoised data is obviously improved, and the curve is more smoothly continuous.

[0130] In the present invention, after the model training is completed, the processes of data processing such as identification and noise removal are all completed fully automatically by the computer, without any manual intervention, and there is no need for experience on the part of the data processing operator. This not only solves the problem of subjective deviation caused by the need to manually set the threshold in the conventional method, but also improves adaptability to different types of noise.

[0131] For detailed functions of each of the above units / modules, please refer to the description of the above method embodiment, and the description will be omitted here.

[0132] It should be noted that the above system is used to execute the methods and corresponding program modules in the system in the above embodiments, and its realization principles and technical effects are similar to those described in the above methods. The operating process of the system can refer to the corresponding process in the above methods, and the description will be omitted here.

[0133] Based on the method in the above embodiment, an embodiment of the present invention provides an electronic device. The electronic device includes a processor, a communications interface, a memory, and a communications bus. The processor, the communications interface, and the memory communicate with each other via the communications bus. The processor can call the logic instructions in the memory to execute the method in the above embodiment.

[0134] Those skilled in the art will recognize that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims

1. 1. A method for suppressing strong interference in artificial current source electromagnetic data, comprising: inputting the artificial current source electromagnetic data to be processed into an artificial current source electromagnetic data classification model to obtain first noise-free artificial current source electromagnetic data and noise-containing artificial current source electromagnetic data; converting the noisy artificial current source electromagnetic data into two-dimensional image data by a dimension transformation function, and inputting the data into the DWTSC-UNet denoising network model to obtain the denoised data result; performing a dimensional inverse transformation on the denoised data result using a dimensional transformation function to obtain second noise-free artificial current source electromagnetic data; combining the first noise-free artificial-source electromagnetic data and the second noise-free artificial-source electromagnetic data to obtain complete noise-free artificial-source electromagnetic data; Including, The DWTSC-UNet denoising network model is a U-structured network that combines a discrete wavelet transform, a spatial attention mechanism, and a channel attention mechanism; A method, characterized in that a discrete wavelet transform is used to provide high-frequency and low-frequency signals, and weights are applied to different frequency components of the two-dimensional image data to extract multi-scale features from the two-dimensional image data, a channel attention mechanism is used to apply weights to different dimensions of the feature map, and a spatial attention mechanism is used to apply weights to different spatial locations of the feature map.

2. The training method of the DWTSC-UNet denoising network model is as follows: selecting noisy artificial current source electromagnetic data samples from the actual measurement data, and performing noise removal using a dictionary learning method to separate the actual measurement noise signals; adding the measured noise signal and the analog noise signal to the noise-free artificial current source electromagnetic data to form noisy artificial current source electromagnetic data, which are combined with the noise-free artificial current source electromagnetic data samples to form sample pairs, thereby constructing a one-dimensional denoised sample library; Segmenting the one-dimensional denoised sample library and performing a dimensional transformation on each sample using a dimensional transformation function to obtain a two-dimensional image sample library; training a DWTSC-UNet denoising network model using a 2D image sample library as input; 2. The method of claim 1, comprising:

3. The DWTSC-UNet denoising network model includes an input layer, a hidden layer, and an output layer.

3. The method of claim 1 or 2, wherein the hidden layer includes a discrete wavelet transform layer, a first shape reconstruction layer, a first convolution block, a first spatial attention mechanism module, a second convolution block, a second spatial attention mechanism module, a third convolution block, a first up-sample block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second up-sample block, a second skip connection layer, a second channel attention mechanism module, and a third up-sample block.

4. 4. The method of claim 3, wherein the discrete wavelet transform layer decomposes the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part, and a diagonal high-frequency part, wherein the low-frequency approximation part contains the basic contour of the two-dimensional image data, the horizontal high-frequency part contains horizontal edge information of the two-dimensional image data, the vertical high-frequency part contains vertical edge information of the two-dimensional image data, and the diagonal high-frequency part contains diagonal edge information of the two-dimensional image data.

5. The method according to claim 4, characterized in that the discrete wavelet transform layer uses a wavelet base with a wavelet base type of db5, and the boundary processing mode used is periodic boundary conditions.

6. The method of claim 1 , wherein the artificial current source electromagnetic data classification model is an InceptTCN classification network model.

7. 1. A system for suppressing strong interference in artificial current source electromagnetic data, comprising: a data classification module for inputting the artificial current-source electromagnetic data to be processed into an artificial current-source electromagnetic data classification model to obtain first noise-free artificial current-source electromagnetic data and noise-containing artificial current-source electromagnetic data; a data denoising module for converting the noisy artificial current source electromagnetic data into two-dimensional image data by a dimension transformation function and inputting the data into the DWTSC-UNet denoising network model to obtain a denoised data result; a dimension transformation module for performing a dimension inverse transformation on the denoised data result using a dimension transformation function to obtain second noise-free artificial current source electromagnetic data; a data combination module for combining the first noise-free artificial-source electromagnetic data and the second noise-free artificial-source electromagnetic data to obtain complete noise-free artificial-source electromagnetic data; Including, The DWTSC-UNet denoising network model is a U-structured network that combines a discrete wavelet transform, a spatial attention mechanism, and a channel attention mechanism; The discrete wavelet transform provides high-frequency and low-frequency signals, and is used to perform weighting training on different frequency components of the two-dimensional image data and extract multi-scale features from the two-dimensional image data, wherein a channel attention mechanism is used to perform weighting on different dimensions of the feature map, and a spatial attention mechanism is used to perform weighting on different spatial locations of the feature map.

8. further comprising a denoising model training module; The denoising model training module: a measured noise signal acquisition unit for selecting noisy artificial current source electromagnetic data samples from the measured data, and performing noise removal using a dictionary learning method to separate the measured noise signals; a one-dimensional denoising sample library construction unit for adding the measured noise signal and the analog noise signal to the noise-free artificial current source electromagnetic data to form noisy artificial current source electromagnetic data, which is combined with the noise-free artificial current source electromagnetic data sample to form a sample pair, thereby constructing a one-dimensional denoising sample library; a dimension transformation unit for segmenting the one-dimensional denoising sample library and performing a dimension transformation on each sample according to a dimension transformation function to obtain a two-dimensional image sample library; a model training unit for inputting a 2D image sample library into the DWTSC-UNet denoising network model for training; The system of claim 7, comprising:

9. The DWTSC-UNet denoising network model in the data denoising module includes an input layer, a hidden layer and an output layer, and the hidden layer includes a discrete wavelet transform layer, a first shape reconstruction layer, a first convolution block, a first spatial attention mechanism module, a second convolution block, a second spatial attention mechanism module, a third convolution block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module and a third upsampling block; 9. The system according to claim 7 or 8, wherein the discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part and a diagonal high-frequency part, wherein the low-frequency approximation part contains the basic contour of the two-dimensional image data, the horizontal high-frequency part contains horizontal edge information of the two-dimensional image data, the vertical high-frequency part contains vertical edge information of the two-dimensional image data, and the diagonal high-frequency part contains diagonal edge information of the two-dimensional image data.

10. 10. The system of claim 9, wherein the discrete wavelet transform layer uses a wavelet base with a wavelet base type of db5 and the boundary processing mode used is periodic boundary conditions.

Images