Magnetotelluric signal denoising method and system based on multi-path fusion network

Abstract

The invention provides a magnetotelluric signal denoising method and system based on a multi-path fusion network. The method comprises the following steps: constructing and training a noise identification model and a signal mapping model based on the multi-path fusion network; inputting a noisy signal segment identified based on the noise identification model into the signal mapping model to obtain a pure signal segment or a noise contour; and finally, splicing the pure signal segments in sequence to obtain the magnetotelluric signal after noise reduction. According to the multi-path fusion network provided by the invention, three convolution layers with convolution kernels of different sizes and a convolution path of an activation function are arranged in a multi-path convolution module, and the multi-path fusion network is used for multi-scale signal feature extraction; and in an attention-enhanced multi-path feature fusion module, outputs of different paths are respectively input into channel attention, space attention and a convolutional layer, and the corresponding outputs are spliced in a channel dimension, so that fusion of multi-path features is realized. In addition, physical constraints are further introduced, and deep optimization is carried out on the two models.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of geological exploration, and particularly relates to a magnetotelluric signal denoising method and system based on a multi-path fusion network. BACKGROUND

[0002] Magnetotellurics (MT) is a geophysical exploration method that measures the natural electromagnetic field on the surface to detect the electrical conductivity structure of the subsurface. This method can provide information on the electrical structure of the subsurface from the shallow crust to the lithosphere and even deeper, and thus has a wide range of applications in geological research, mineral resource exploration, and geothermal resource exploration. Natural magnetotelluric signals are very weak and easily disturbed by noise. At the same time, with the development of cities, human electromagnetic noise generated during production has a great influence on the collection of magnetotelluric signals, which seriously affects the interpretation and inversion of the subsurface geological structure.

[0003] With the efforts of many scholars, many effective methods have been applied in the field of magnetotelluric signal processing. Remote Reference (RR) is one of the most widely used magnetotelluric denoising methods. The noise suppression effect of the remote reference method is greatly related to the selection of the reference point, but with the continuous acceleration of urbanization, the selection of the remote reference point has become increasingly difficult, and the limitations of the remote reference method have also been magnified. In addition to the remote reference method, the Robust estimation method is also commonly used to suppress noise, but when the noise part accounts for a large proportion of the collected data, the denoising effect of this method is significantly reduced. These methods can suppress noise to some extent and ultimately restore the original signal. However, they all have different defects, such as the inability to effectively and flexibly handle interference, leading to a large gap between the restored signal and the actual signal, the introduction of new noise affecting the denoising accuracy, etc. In recent years, deep learning methods have also been widely used in magnetotelluric signal processing. Convolutional Neural Network (CNN) was initially used to identify magnetotelluric noise. After discovering the non-linear mapping ability of neural networks in noise profiles, CNN was also used for magnetotelluric signal denoising. However, the defects of CNN are also obvious, including the use of single-scale convolution, the failure to combine noise identification and signal mapping tasks, and the separate processing from the time domain. Therefore, further exploring the potential of deep learning in this application is still a research hotspot in this field. SUMMARY

[0004] To improve the accuracy and efficiency of magnetotelluric (MT) signal denoising, this invention proposes a MT signal denoising method and system based on a multi-path fusion network. The multi-path fusion network, capable of noise identification and signal mapping, serves as the backbone network for feature extraction. It utilizes convolutional kernels of varying sizes in multi-path convolution; large-scale kernels extract the global representation of the signal, while small-scale kernels extract local features. Channel attention and spatial attention further enhance the network's representation of global features and local details, respectively. The feature maps from different paths are concatenated along the channel dimension to achieve multi-scale feature fusion, enabling the network to learn noise types with different amplitudes and forms. Furthermore, this invention deeply optimizes this improvement by incorporating physical constraints (utilizing the frequency domain characteristics and sparsity of MT signals). Prior physical information about the MT signals guides the network training and constrains the network's output, making the network interpretable while achieving more accurate noise identification and signal reconstruction.

[0005] Therefore, the present invention provides the following technical solution:

[0006] On the one hand, the present invention provides a method for denoising magnetotelluric signals based on a multi-path fusion network, comprising the following steps:

[0007] Step S1: Sample library construction, constructing a noise recognition sample library and a signal mapping sample library;

[0008] Step S2: Model building and training. A noise recognition model based on a multi-path fusion network is built and trained based on a noise recognition sample library. A signal mapping model based on a multi-path fusion network is built and trained based on a signal mapping sample library.

[0009] The multi-path fusion network includes at least a multi-path convolution module and a multi-path feature fusion module. The multi-path convolution module has three parallel convolutional paths with different kernel sizes. The multi-path feature fusion module also has three parallel paths: channel attention, spatial attention, and convolutional channels. The outputs of the three paths are concatenated along the channel dimension. In the multi-path convolution module, the output of the path with the small convolutional kernel corresponds to the input channel attention path; the output of the path with the large convolutional kernel corresponds to the input spatial attention path; and the outputs of the paths with the large, medium, and small convolutional kernels are first concatenated along the channel dimension and then input into the convolutional channels.

[0010] Step S3: Use the noise identification model to identify the segmented signal of the magnetotelluric signal to be denoised as a noisy signal segment or a clean signal segment. Then, input the identified noisy signal segment into the signal mapping model to obtain a clean signal segment or a noise profile. If it is a noise profile, subtract the noise profile from the noisy signal segment to obtain a clean signal segment.

[0011] Step S4: Sequentially splice the clean signal segments to obtain the noise-reduced magnetotelluric signal.

[0012] More preferably, the multi-path convolution module first has an initial convolutional layer, batch normalization, and activation function; then it is divided into three convolutional paths, each of which has two sets of alternating convolutional layers and activation functions, and the kernel sizes of the three convolutional paths are 1×1, 3×1, and 5×1, respectively.

[0013] The input to the channel attention is divided into two branches. One branch first inputs into the max pooling layer, and the other branch first inputs into the average pooling layer. The outputs of both are then passed through a 3×1 convolution, a Swish activation function, and a 3×1 convolution. The outputs of the two branches are then superimposed and connected to a sigmoid function. Finally, the input is multiplied by the input of the channel attention.

[0014] The convolutional channels are sequentially configured with 1×1 convolution, batch normalization, and Swish activation function;

[0015] The input to the spatial attention is divided into two branches. One branch first inputs to the max pooling layer, and the other branch first inputs to the average pooling layer. The outputs of the two are concatenated in the channel dimension, and then passed through a 7×1 convolution, a sigmoid function, and a global max pooling layer. Finally, it is multiplied by the input of the spatial attention.

[0016] More preferably, the multi-path fusion network includes a dynamic gating module between the multi-path convolution module and the multi-path feature fusion module;

[0017] The dynamic gating module is equipped with a gating weight generator based on global pooling and fully connected layers. It outputs the weights of each convolutional path in the multi-path fusion network through the Sigmoid function, and then weights them with the outputs of the corresponding convolutional paths.

[0018] More preferably, physical constraints are introduced into the noise identification model in step S2, that is, the input of the noise identification model is set to a frequency domain signal containing the frequency domain characteristics of the magnetotelluric signal, specifically:

[0019] First, Fourier transform is introduced to convert the time series of the magnetotelluric signal into a frequency domain signal; then, the median of the amplitude spectrum is used to perform adaptive threshold filtering on the frequency domain signal to obtain the filtered frequency domain signal.

[0020] Among them, the median of the amplitude spectrum is used Adaptive threshold filtering is performed to obtain the filtered frequency domain signal. , For threshold parameters, For indicator functions, when The output is 1 if the frequency data is selected, and 0 if it is selected, meaning the frequency data is retained. This represents the kth complex component of the frequency domain output of the frequency domain signal obtained after the Fourier transform of the magnetotelluric signal segment.

[0021] More preferably, the threshold parameter For an adjustable dynamic threshold designed to target the characteristics of magnetotelluric noise, the following exists:

[0022]

[0023] in, , These are learnable parameters that are automatically optimized during the training process of a noise recognition network. amplitude spectrum The skewness, i.e., the third moment; Logarithmic amplitude spectrum The standard deviation exists:

[0024]

[0025]

[0026] Where N is the number of sampling points in the magnetotelluric signal segment. Logarithmic amplitude spectrum The mean, This represents the mean function.

[0027] More preferably, the output of the signal mapping model in step S2 is a pure signal segment, and the signal mapping model introduces physical constraints, namely, the loss function incorporates at least wavelet domain adaptive sparse loss. ,exist:

[0028]

[0029] Among them, the residual Perform wavelet packet transform to generate subband coefficients , This indicates a wavelet transform using the Symlets wavelet basis. Let i be the true value of the i-th sample. Let be the predicted value of the i-th sample, n be the number of samples, and j be the subband index. Let be the subband coefficient corresponding to the j-th subband index in the wavelet packet transform of the i-th sample. For sub-band adaptive thresholds, satisfying:

[0030] in, This is the subband noise standard deviation estimate, where N is the number of sampling points.

[0031] More preferably, the loss function of the signal mapping model is constructed by weighting the mean square error loss and the wavelet domain adaptive sparse loss, and is expressed as:

[0032]

[0033] in, The loss function of the signal mapping model. , All are adjustable weighting coefficients. This is the mean square error loss;

[0034] The loss function used in the training process of the noise recognition model is the cross-entropy loss function.

[0035] Secondly, the present invention provides a system based on a magnetotelluric signal denoising method, comprising:

[0036] The sample library construction module is used to build noise recognition sample libraries and signal mapping sample libraries;

[0037] The model building and training module is used to build and train a noise recognition model based on a multi-path fusion network based on a noise recognition sample library, and to build and train a signal mapping model based on a multi-path fusion network based on a signal mapping sample library.

[0038] The multi-path fusion network includes at least a multi-path convolution module and a multi-path feature fusion module. The multi-path convolution module has three parallel convolutional paths with different kernel sizes. The multi-path feature fusion module also has three parallel paths: channel attention, spatial attention, and convolutional channels. The outputs of the three paths are concatenated along the channel dimension. In the multi-path convolution module, the output of the path with the small convolutional kernel corresponds to the input channel attention path; the output of the path with the large convolutional kernel corresponds to the input spatial attention path; and the outputs of the paths with the large, medium, and small convolutional kernels are first concatenated along the channel dimension and then input into the convolutional channels.

[0039] The noise reduction module is used to identify the segmented signal of the magnetotelluric signal to be denoised as a noisy signal segment or a clean signal segment using the noise recognition model, and then input the identified noisy signal segment into the signal mapping model to obtain a clean signal segment or a noise profile. If it is a noise profile, the noise signal segment is subtracted from the noise profile to obtain a clean signal segment; and it is used to sequentially splice the clean signal segments to obtain the denoised magnetotelluric signal.

[0040] In three aspects, the computer device provided by the technical solution of the present invention includes:

[0041] One or more processors;

[0042] A memory that stores one or more computer programs;

[0043] The processor invokes a computer program to achieve the following:

[0044] The steps of a magnetotelluric signal denoising method based on multipath fusion network.

[0045] In four aspects, the present invention provides a computer-readable storage medium storing a computer program, which is invoked for execution:

[0046] The steps of a magnetotelluric signal denoising method based on multipath fusion network.

[0047] Beneficial effects

[0048] 1. This invention employs multi-path convolution combined with channel attention and spatial attention to extract noise features from magnetotelluric signals, successfully applying it to the field of magnetotelluric signal denoising. The multi-path convolution module comprises three convolutional paths with kernels of different sizes. Large-scale kernels extract the global representation of the signal, while small-scale kernels extract local features. The feature fusion process inputs the outputs of different paths to channel attention, spatial attention, and the convolutional layer respectively, concatenating the corresponding outputs along the channel dimension. The global features extracted by the large convolutional kernel are fed into the channel attention mechanism, which utilizes the interdependence between channel mappings to improve the feature representation of specific semantics. The local features extracted by the small convolutional kernel are input into the spatial attention mechanism to enhance local details and suppress irrelevant regions. This multi-path feature fusion module, to a certain extent, prevents problems such as gradient vanishing, gradient exploding, and network degradation, thereby effectively capturing global and local feature information at each layer of the network and achieving multi-path feature fusion.

[0049] 2. The technical solution of this invention is further optimized by introducing physical constraints of frequency domain features. It utilizes the frequency domain information of magnetotelluric signals to achieve noise identification. By converting the magnetotelluric signal to the frequency domain through a frequency domain feature reconstruction method, it solves the limitation of traditional noise identification methods that identify signal amplitude. It achieves accurate noise identification, removes most of the large-scale noise while retaining the interference-free effective part of the signal, reduces the overprocessing of useful signals by the network, and can retain more complete low-frequency information of the signal, thus restoring a more realistic interference-free magnetotelluric signal.

[0050] 3. The technical solution of this invention is further optimized by introducing physical constraints of signal sparsity characteristics. That is, by utilizing the sparsity characteristics of magnetotelluric signals, a wavelet adaptive sparse loss function is constructed to optimize the performance of the neural network, control the complexity of the model, reduce the risk of network overfitting, and make the network more interpretable. This ensures that the output signal has the physical meaning of magnetotelluric signals, and directly maps the corresponding denoised signal from the noisy signal, improving computational efficiency without causing loss of denoising effect. Attached Figure Description

[0051] Figure 1 A flowchart provided for an embodiment of the present invention;

[0052] Figure 2 This is a network structure diagram of a multipath fusion network;

[0053] Figure 3 This is a schematic diagram of a dynamic gating module;

[0054] Figure 4 A schematic diagram of an attention-enhanced multi-path feature fusion module;

[0055] Figure 5 The recognition effect of the multi-path fusion network noise recognition model on the measured data is shown in the figure.

[0056] Figure 6 The denoising effect of the multi-path fusion network signal mapping model on measured data is shown in the figure.

[0057] Figure 7 This is a schematic diagram showing the connection relationships of electronic components in a computer device. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other.

[0059] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0061] This invention provides a method and system for denoising magnetotelluric signals based on a multi-path fusion network. The method employs multi-path convolution and an attention-enhanced feature fusion approach combining channel attention and spatial attention to extract noise features from the magnetotelluric signal. Furthermore, it optimizes the denoising accuracy by introducing physical constraints, specifically through noise identification using frequency domain features and signal mapping optimized by wavelet adaptive sparse loss function.

[0062] To extract features of different amplitudes and morphologies from magnetotelluric signals, such as charging and discharging triangular waves, square waves, and impulse noise, this invention proposes a multi-scale signal feature extraction method comprising three convolutional kernel paths with different sizes. Large-scale convolutional kernels extract the global representation of the signal, while small-scale kernels extract local features. Simultaneously, a multi-path feature fusion module is constructed, consisting of channel attention, spatial attention, and convolutional channels operating in parallel. The global features extracted by the large convolutional kernels are input into the channel attention mechanism, utilizing the interdependence between channel mappings to improve feature representation. The local features extracted by the small convolutional kernels are input into the spatial attention mechanism to enhance local noise details and suppress irrelevant clean signal components. This multi-path feature fusion module prevents common problems in deep neural networks, such as gradient vanishing and network degradation, effectively capturing global and local feature information at each level. Based on this, the trained noise recognition model and signal mapping model are applied to measured magnetotelluric signals. Noise recognition effectively reduces overprocessing of useful signals by the signal mapping model, preserving better low-frequency trends and achieving higher-performance magnetotelluric signal denoising.

[0063] The present invention will be further described below with reference to embodiments.

[0064] Example 1

[0065] like Figure 1 As shown in the figure, the present invention provides a method for denoising magnetotelluric signals based on a multi-path fusion network, the technical idea of ​​which is as follows:

[0066] Step S1: Sample library construction, constructing a noise recognition sample library and a signal mapping sample library.

[0067] Step S2: Model building and training. A noise recognition model based on a multi-path fusion network is built and trained based on a noise recognition sample library. A signal mapping model based on a multi-path fusion network is built and trained based on a signal mapping sample library.

[0068] Step S3: Use the noise identification model to identify the segmented signal of the magnetotelluric signal to be denoised as a noisy signal segment or a clean signal segment. Then, input the identified noisy signal segment into the signal mapping model to obtain a clean signal segment or a noise profile. If it is a noise profile, subtract the noise profile from the noisy signal segment to obtain a clean signal segment.

[0069] Step S4: Sequentially splice the clean signal segments to obtain the noise-reduced magnetotelluric signal.

[0070] It should be noted that in this embodiment, both the noise recognition model and the signal mapping model use a multi-path fusion network. The multi-path fusion network provided in this embodiment is preferably a gated multi-path fusion network, that is, it sequentially includes a multi-path convolution module, a dynamic gating module, and a multi-path feature fusion module. In other feasible embodiments, the multi-path fusion module may not have a gating module; instead, it may consist of a multi-path convolution module and a multi-path feature fusion module, which is also feasible and falls within the protection scope of this invention.

[0071] The multi-path convolution module has three parallel convolutional paths consisting of alternating convolutional layers and activation functions. The kernel sizes of the convolutional layers in different paths vary, enabling multi-scale feature extraction. For example... Figure 2 As shown, in this embodiment, the structure of the multi-path convolution module is as follows: first, an initial convolutional layer, batch normalization, and activation function are set; then, it is divided into three convolutional paths, each with two sets of alternating convolutional layers and activation functions, with convolutional kernel sizes of 1×1, 3×1, and 5×1, respectively. It should be understood that in other feasible embodiments, provided that large, medium, and small convolutional kernels are set, the size of the convolutional kernels can be adjusted according to the desired effect.

[0072] The dynamic gating module includes a gating weight generator based on global pooling and fully connected layers. It outputs the weights of each path using the sigmoid function to achieve dynamic feature selection. For example... Figure 3 As shown, the dynamic gating module in this embodiment first concatenates the outputs of the multi-path convolution module, captures global statistics through global average pooling, then performs dimensionality reduction and activation by the fully connected layer, and finally generates normalized weights by the sigmoid function, which are then weighted with the outputs of the corresponding convolution path.

[0073] The multi-path feature fusion module has three parallel paths: channel attention, spatial attention, and convolutional channels. The outputs of the three paths are concatenated along the channel dimension to achieve multi-path feature fusion. In this embodiment, the weighted sum of the output of the path containing the small convolutional kernel and its weights is used as the input to the channel attention path; the weighted sum of the output of the path containing the large convolutional kernel and its corresponding weights is used as the input to the spatial attention path; and the weighted sum of the outputs of the paths containing the large, medium, and small convolutional kernels is first concatenated along the channel dimension and then used as the input to the convolutional channels. In other feasible embodiments, if there is no gating module, the output of the path containing the small convolutional kernel is directly input to the channel attention path; the output of the path containing the large convolutional kernel is directly input to the spatial attention path; and the outputs of the paths containing the large, medium, and small convolutional kernels are first concatenated along the channel dimension and then input to the convolutional channels.

[0074] like Figure 4 As shown, in this embodiment, the input to the channel attention layer is divided into two branches. One branch first inputs to a max pooling layer, and the other branch first inputs to an average pooling layer. The outputs of both branches are then passed through a 3×1 convolution, a Swish activation function, and another 3×1 convolution. The outputs of the two branches are then superimposed and connected to a sigmoid function, and finally multiplied by the input to the channel attention layer. The convolutional channels sequentially have a 1×1 convolution, batch normalization, and a Swish activation function. The input to the spatial attention layer first passes through max pooling and average pooling respectively, then is concatenated along the channel dimension, then passes through a 7×1 convolution, a sigmoid function, and a global max pooling layer, and finally multiplied by the input to the spatial attention layer.

[0075] Regarding step S1, both the noise recognition sample library and the signal mapping sample library are constructed based on noisy samples, noise samples, and clean signal samples. The noise recognition sample library contains at least noisy samples, clean signal samples, and their respective noise or non-noise labels; this sample library is used to train the noise recognition model. The signal mapping sample library contains at least noisy samples and their noise contours or clean signals; this sample library is used to train the signal mapping model.

[0076] In step S1, this embodiment analyzes the types of magnetotelluric noise and finds that the noise types include charging and discharging triangular waves, pulses, square wave noise, etc., and then uses mathematical functions to construct the corresponding noise patterns. Each noise segment is set to a length of 200, and the noise amplitude ranges from -8000 to 8000. Furthermore, operations such as flipping, translating, and superimposing are used to expand the noise sample library, improving the diversity of the final noisy signal sample library.

[0077] Given a piece of noise data Represented as:

[0078]

[0079] in, This represents a segment of noise data with a segment length of N. , , , Represent The first, i-th, j-th, and N-th elements, and the noise part from Start and continue until The remaining elements are set to 0.

[0080] The flip operation can be represented as follows:

[0081]

[0082]

[0083] in, This represents the noise after vertical flipping. This represents the noise after horizontal flipping. Flipub is a function in MATLAB used to achieve horizontal flipping.

[0084] The translation operation can be represented as follows:

[0085]

[0086] in, This refers to the noise after the translation operation. and These are adjustable parameters; by changing... + The size of the variable can be adjusted to change the position of the noise within the data segment, thus achieving a translation operation.

[0087] Specifically, the superposition operation can be represented as follows:

[0088]

[0089] in, It is a superimposed signal. and These are two random segments of noisy data from the sample library. and This is a scaling factor, typically ranging from 0.5 to 1.5 to prevent the amplitude from being too large or too small. By superimposing different noises, the noise types in the sample library can be enriched, enhancing the network's ability to handle complex noise.

[0090] It should be understood that in other feasible embodiments, it is not required to expand the noise contour sample library in the manner described above in terms of sample diversity; some steps may be performed or other techniques may be used to expand the samples.

[0091] Regarding step S2, in some embodiments, the noise recognition model and signal mapping model are trained using conventional methods. Specifically, the noise recognition model's input is limited to a signal segment, and its output is a classification result of noise or a clean signal; and the signal mapping model's input is limited to a noisy signal segment, and its output is a clean signal or noise contour. Conventional methods are used for model training. In this embodiment, physical constraints are further preferably introduced to optimize the noise recognition model and signal mapping model, as follows:

[0092] Regarding the noise identification model:

[0093] In some embodiments, the physical characteristics of the frequency domain features are fully considered, and frequency domain feature reconstruction is used to convert the magnetotelluric time series into a frequency domain signal. That is, Fourier transform is first introduced to convert the time series of the magnetotelluric signal into a frequency domain signal; then the median of the amplitude spectrum is used to perform adaptive threshold filtering on the frequency domain signal to obtain the filtered frequency domain signal.

[0094] For example, let the geomagnetic signal segment N is the number of sampling points. These are the signals from the 1st, 2nd, and Nth sampling points in the magnetotelluric signal segment, respectively. After Fourier transform, the frequency domain signal is obtained. , The k-th complex component of the frequency domain output has the following characteristics:

[0095]

[0096] in, This represents the nth sample of the magnetotelluric signal, where i is the imaginary unit. is the rotation factor, and represents the complex exponent.

[0097] Then use the median of the amplitude spectrum Adaptive threshold filtering is performed to obtain the filtered frequency domain features. ,, For indicator functions, when The output is 1 if the frequency point data is selected, and 0 otherwise if it is selected, indicating that the frequency point data will be filtered out. In some embodiments, The threshold parameter is set based on the effect. In this embodiment, an adjustable dynamic threshold parameter is further designed based on the characteristics of magnetotelluric noise to avoid extreme values ​​(such as impulse noise peaks) affecting the threshold reference, while balancing signal preservation and noise suppression. Specifically, it satisfies:

[0098]

[0099] in, , These are learnable parameters that are automatically optimized during the training process of a noise recognition network. amplitude spectrum The skewness, i.e., the third moment; Logarithmic amplitude spectrum The standard deviation exists:

[0100]

[0101]

[0102] Where N is the number of sampling points in the magnetotelluric signal segment. Logarithmic amplitude spectrum The mean.

[0103] In this embodiment, the loss function for the training process of the noise recognition model is the cross-entropy loss function, expressed as:

[0104]

[0105] Where y represents the real label. The model predicts the probability that a sample belongs to category 1, where n is the number of samples. It should be understood that in other feasible embodiments, the loss function of the noise recognition model can be other loss functions, which also meet the requirements of this invention.

[0106] Regarding signal mapping models:

[0107] In some embodiments, the signal mapping model outputs a noise profile, and the clean signal is obtained by subtracting the noise profile from the noisy signal; in other embodiments, the signal mapping model outputs a clean signal. In this embodiment, the output of the signal mapping model is a clean signal segment, and further technical optimization is performed by introducing physical constraints into the signal mapping model, namely, the loss function incorporates at least wavelet domain adaptive sparse loss. ,exist:

[0108]

[0109] Among them, the residual Perform wavelet packet transform to generate subband coefficients , This indicates a wavelet transform using the Symlets wavelet basis. Let i be the true value of the i-th sample. Let be the predicted value of the i-th sample, N be the number of samples, and j be the subband index. Let be the subband coefficient corresponding to the j-th subband index in the wavelet packet transform of the i-th sample. For sub-band adaptive thresholds, satisfying: Where j is the sub-band index, For subband noise standard deviation estimation, N is the number of sampling points. In this embodiment, Sym4 wavelet packet transform is selected, resulting in a wavelet decomposition level of 4, generating the corresponding 16 subband coefficients.

[0110] In this embodiment, the loss function of the signal mapping model is constructed by weighting the mean square error loss and the wavelet domain adaptive sparse loss, and is expressed as:

[0111]

[0112]

[0113] in, The loss function of the signal mapping model. , All are adjustable weighting coefficients. The mean squared error loss is used. It should be understood that in other feasible embodiments, the mean squared error loss can be replaced with other loss functions, and specific references can be made to existing technologies in the art. This invention does not impose specific limitations on this.

[0114] In steps S3 and S4, the identified noisy signal segments are mapped to the input signal model to obtain either a clean signal segment or a noise profile. If it is a noise profile, the noisy signal segment is subtracted from the noise profile to obtain a clean signal segment. Finally, the clean signal segments are sequentially spliced ​​together to obtain the denoised magnetotelluric signal. This process is a conventional technique in this field, therefore, it will not be described in detail.

[0115] For example, the measured magnetotelluric signal Y is uniformly divided into k segments with a segment length of 200, resulting in a magnetotelluric signal matrix with k segments. It is expressed as follows:

[0116]

[0117] In the formula, arrive express The first data segment with a length of 200, arrive express The k-th data segment of length 200 in Y is the k-th data segment of length 200 in Y.

[0118] Set the signal segment identified as a pure signal to 0, then set the matrix... The elements are input column-wise into the signal mapping model, and the network output is a matrix. The pure signal of the corresponding data , means as follows:

[0119]

[0120] In the formula, arrive express The first data segment with a length of 200, arrive express The k-th data segment of length 200 in Y is the clean signal corresponding to the k-th data segment of length 200 in Y.

[0121] Finally, the denoised signal matrix is ​​added to the signal matrix identified as the clean signal to reconstruct the complete denoised magnetotelluric matrix. This matrix is ​​then expanded column-wise to obtain the denoised magnetotelluric signal X'', which can be represented as... , These are the 1st, 2nd, and 200kth signals in the magnetotelluric signal X'', respectively.

[0122] Figure 5 To simulate the recognition performance of a noise recognition model based on a multi-path fusion network with introduced physical constraints in an example, where... Figure 5 Figure (a) shows the time-domain waveform of the measured noisy magnetotelluric signal. Figure 5 Figure (b) shows the noisy signal segment identified by the noise recognition model. Figure 5 Figure (c) shows the clean signal segment identified by the noise recognition model. Figure 6 To simulate the denoising effect of a signal mapping model based on a multi-path fusion network with physical constraints in a single instance, Figure 6 Figure (a) shows the time-domain waveform of the noisy magnetotelluric signal. Figure 6 Figure (b) shows the signal after denoising using the signal mapping model. From Figure 5 and Figure 6 It can be seen that the noise identification model provided by the technical solution of the present invention can accurately identify noisy signal segments, and the signal mapping model can effectively suppress large-scale magnetotelluric noise.

[0123] Example 2:

[0124] This embodiment also provides a system based on the magnetotelluric signal denoising method, including a sample library construction module, a noise identification model construction module, a signal mapping model construction module, and a signal reconstruction module that are connected sequentially or interconnected.

[0125] The sample library construction module is used to build a clean signal sample library and a noisy signal sample library, including a noise identification sample library and a sample library for signal mapping.

[0126] The noise recognition model building module is used to train a multi-path fusion network using a noise recognition sample library to achieve noise recognition from signal to corresponding label.

[0127] The signal mapping model building module is used to train a multi-path fusion network using a signal mapping sample library to achieve the signal mapping effect from noisy signals to corresponding clean signals.

[0128] The signal reconstruction module is used to input the measured magnetotelluric data into the noise identification model and the signal mapping model in sequence to obtain each segment of the denoised magnetotelluric signal, and then reassemble each segment of the signal to restore the complete denoised magnetotelluric signal.

[0129] It should also be understood that the specific implementation process of each module is described in the above method. This invention will not repeat it here. The division of the above functional modules is only for illustrative purposes. In some embodiments, some functional modules can be merged and some functional modules can be split. Each functional module can be implemented in software, hardware, or a combination of software and hardware. The software and hardware devices include, but are not limited to, general electronic terminals, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0130] Example 3:

[0131] This embodiment provides a computer device, including: one or more processors; and a memory storing a computer program; wherein the processor calls the computer program to implement: the steps of a magnetotelluric signal denoising method based on a multipath fusion network. Specifically, steps S1-S4 are executed.

[0132] For details on the implementation of each step, please refer to the description of the aforementioned denoising method embodiment.

[0133] In some embodiments, such as Figure 7 As shown, the electronic components of a computer device include:

[0134] The processor 1600 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 1600 is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0135] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700, and the processor 1600 calls and executes the algorithm program of the noise reduction method described in the embodiments of this invention.

[0136] The input / output interface 1800 is used to implement information input and output.

[0137] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0138] Bus 2000 transmits information between various components of the device, such as processor 1600, memory 1700, input / output interface 1800, and communication interface 1900.

[0139] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0140] Example 4:

[0141] This embodiment provides a computer-readable storage medium storing a computer program, which is called by a processor to implement the steps of a magnetotelluric signal denoising method based on a multipath fusion network. Specifically, steps S1-S4 are executed.

[0142] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the foregoing embodiments, such as the controller's hard drive or memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both the controller's internal storage unit and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0143] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0144] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A method for denoising magnetotelluric signals based on a multi-path fusion network, characterized in that: Includes the following steps: Step S1: Sample library construction, constructing a noise recognition sample library and a signal mapping sample library; Step S2: Model building and training. A noise recognition model based on a multi-path fusion network is built and trained based on a noise recognition sample library. A signal mapping model based on a multi-path fusion network is built and trained based on a signal mapping sample library. The multi-path fusion network includes at least a multi-path convolution module and a multi-path feature fusion module. The multi-path convolution module has three parallel convolutional paths with different kernel sizes. The multi-path feature fusion module also has three parallel paths: channel attention, spatial attention, and convolutional channels. The outputs of the three paths are concatenated along the channel dimension. In the multi-path convolution module, the output of the path with the small convolutional kernel corresponds to the input channel attention path; the output of the path with the large convolutional kernel corresponds to the input spatial attention path; and the outputs of the paths with the large, medium, and small convolutional kernels are first concatenated along the channel dimension and then input into the convolutional channels. Step S3: Use the noise identification model to identify the segmented signal of the magnetotelluric signal to be denoised as a noisy signal segment or a clean signal segment. Then, input the identified noisy signal segment into the signal mapping model to obtain a clean signal segment or a noise profile. If it is a noise profile, subtract the noise profile from the noisy signal segment to obtain a clean signal segment. Step S4: Sequentially splice the clean signal segments to obtain the noise-reduced magnetotelluric signal.

2. The magnetotelluric signal denoising method according to claim 1, characterized in that: The multi-path convolution module first has an initial convolutional layer, batch normalization, and activation function; then it is divided into three convolutional paths, each with two sets of alternating convolutional layers and activation functions. The kernel sizes of the three convolutional paths are 1×1, 3×1, and 5×1, respectively. The input to the channel attention is divided into two branches. One branch first inputs into the max pooling layer, and the other branch first inputs into the average pooling layer. The outputs of both are then passed through a 3×1 convolution, a Swish activation function, and a 3×1 convolution. The outputs of the two branches are then superimposed and connected to a sigmoid function. Finally, the input is multiplied by the input of the channel attention. The convolutional channels are sequentially configured with 1×1 convolution, batch normalization, and Swish activation function; The input to the spatial attention is divided into two branches. One branch first inputs to the max pooling layer, and the other branch first inputs to the average pooling layer. The outputs of the two are concatenated in the channel dimension, and then passed through a 7×1 convolution, a sigmoid function, and a global max pooling layer. Finally, it is multiplied by the input of the spatial attention.

3. The magnetotelluric signal denoising method according to claim 1, characterized in that: The multi-path fusion network includes a dynamic gating module between the multi-path convolution module and the multi-path feature fusion module. The dynamic gating module is equipped with a gating weight generator based on global pooling and fully connected layers. It outputs the weights of each convolutional path in the multi-path fusion network through the Sigmoid function, and then weights them with the outputs of the corresponding convolutional paths.

4. The magnetotelluric signal denoising method according to claim 1, characterized in that: In step S2, physical constraints are introduced into the noise identification model, that is, the input of the noise identification model is set to a frequency domain signal containing the frequency domain characteristics of the magnetotelluric signal, specifically: First, Fourier transform is introduced to convert the time series of the magnetotelluric signal into a frequency domain signal; then, the median of the amplitude spectrum is used to perform adaptive threshold filtering on the frequency domain signal to obtain the filtered frequency domain signal. Among them, the median of the amplitude spectrum is used Adaptive threshold filtering is performed to obtain the filtered frequency domain signal. , For threshold parameters, For indicator functions, when The output is 1 if the frequency data is selected, and 0 if it is selected, meaning the frequency data is retained. This represents the kth complex component of the frequency domain output of the frequency domain signal obtained after the Fourier transform of the magnetotelluric signal segment.

5. The magnetotelluric signal denoising method according to claim 4, characterized in that: Threshold parameter For an adjustable dynamic threshold designed to target the characteristics of magnetotelluric noise, the following exists: in, , These are learnable parameters that are automatically optimized during the training process of a noise recognition network. amplitude spectrum The skewness, i.e., the third moment; Logarithmic amplitude spectrum The standard deviation exists: Where N is the number of sampling points in the magnetotelluric signal segment. Logarithmic amplitude spectrum The mean, This represents the mean function.

6. The magnetotelluric signal denoising method according to claim 1, characterized in that: The output of the signal mapping model in step S2 is a clean signal segment. The signal mapping model introduces physical constraints, namely, the loss function incorporates at least wavelet domain adaptive sparse loss. ,exist: Among them, the residual Perform wavelet packet transform to generate subband coefficients , This indicates a wavelet transform using the Symlets wavelet basis. Let i be the true value of the i-th sample. Let be the predicted value of the i-th sample, n be the number of samples, and j be the subband index. Let be the subband coefficient corresponding to the j-th subband index in the wavelet packet transform of the i-th sample. For sub-band adaptive thresholds, satisfying: in, This is the subband noise standard deviation estimate, where N is the number of sampling points.

7. The magnetotelluric signal denoising method according to claim 6, characterized in that: The loss function of the signal mapping model is constructed by weighting the mean square error loss and the wavelet domain adaptive sparse loss, and is expressed as: in, Let be the loss function of the signal mapping model. , All are adjustable weighting coefficients. This is the mean square error loss; The loss function used in the training process of the noise recognition model is the cross-entropy loss function.

8. A system based on the magnetotelluric signal denoising method according to any one of claims 1-7, characterized in that: include: The sample library construction module is used to build noise recognition sample libraries and signal mapping sample libraries; The model building and training module is used to build and train a noise recognition model based on a multi-path fusion network based on a noise recognition sample library, and to build and train a signal mapping model based on a multi-path fusion network based on a signal mapping sample library. The multi-path fusion network includes at least a multi-path convolution module and a multi-path feature fusion module. The multi-path convolution module has three parallel convolutional paths with different kernel sizes. The multi-path feature fusion module also has three parallel paths: channel attention, spatial attention, and convolutional channels. The outputs of the three paths are concatenated along the channel dimension. In the multi-path convolution module, the output of the path with the small convolutional kernel corresponds to the input channel attention path; the output of the path with the large convolutional kernel corresponds to the input spatial attention path; and the outputs of the paths with the large, medium, and small convolutional kernels are first concatenated along the channel dimension and then input into the convolutional channels. The noise reduction module is used to identify the segmented signal of the magnetotelluric signal to be denoised as a noisy signal segment or a clean signal segment using the noise recognition model, and then input the identified noisy signal segment into the signal mapping model to obtain a clean signal segment or a noise profile. If it is a noise profile, the noise signal segment is subtracted from the noise profile to obtain a clean signal segment; and it is used to sequentially splice the clean signal segments to obtain the denoised magnetotelluric signal.

9. A computer device, characterized in that: include: One or more processors; A memory that stores one or more computer programs; The processor invokes a computer program to achieve the following: The steps of the magnetotelluric signal denoising method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer program is stored and is invoked to execute: The steps of the magnetotelluric signal denoising method according to any one of claims 1-7.

Images