Magnetotelluric signal noise suppression method and system based on complexity driving
By introducing the complexity perception mechanism and the multi-path attention residual denoising network, the modeling strategy of the signal segments is dynamically adjusted, which solves the problem of poor noise suppression in traditional magnetotelluric signal processing and achieves efficient signal recovery and accurate reconstruction.
Patent Information
- Application Number
- CN202511106149.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-09-09
- Estimated Expiration
- 2045-08-08
AI Technical Summary
Among existing magnetotelluric signal processing technologies, traditional methods have limited effectiveness in suppressing strong interference noise and ignore the differences in complexity between different data segments, resulting in uneven signal processing and insufficient extraction of key information, which can easily cause effective signal distortion and residual noise.
By introducing a complexity-aware mechanism and constructing a multi-path, structurally heterogeneous feature recovery network, and using complexity-aware gating units and a multi-path attention residual denoising network, the modeling strategy of signal fragments is dynamically adjusted to achieve deep modeling of high-complexity signals and rapid recovery of low-complexity signals.
It effectively removes strong interference components, improves signal recovery quality and reconstruction accuracy, maintains target information, and significantly improves the recovery quality and reconstruction accuracy of magnetotelluric signals.
Smart Images

Figure CN120610324A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of magnetotelluric signal processing, and in particular relates to a complexity-driven magnetotelluric signal noise suppression method and system. Background Art
[0002] Magnetotelluric (MT) sounding is a mainstream geophysical exploration method that studies the electrical properties and distribution of underground rock formations by observing natural alternating electromagnetic fields. It is used to analyze geological structures and detect the location of deep ore bodies. However, due to the numerous complex field communication facilities and interference from human noise, coupled with the wide frequency range and weak signal amplitude of MT signals, MT signals are highly susceptible to various noise interferences, which seriously affect the quality of the collected data and prevent them from accurately reflecting the underground electrical structure.
[0003] Emerging modern signal processing techniques, such as the Hilbert-Huang transform, mathematical morphological filtering, synchronous time series dependency, signal subspace enhancement, wavelet transform, sparse decomposition, fractal-entropy and clustering, and K-SVD dictionary learning, have been applied to this field, improving the quality of magnetotelluric data to some extent. However, traditional methods have limited effectiveness against strong interfering noise and can easily distort valid signals. Existing deep learning denoising models mostly adopt a unified processing strategy, ignoring the differences in complexity between different data segments, resulting in uneven processing effects. In the feature extraction process, key information is insufficiently extracted, and the inter-layer information transmission mechanism is poorly designed, which can easily lead to the loss of useful information or residual noise. Summary of the Invention
[0004] The present invention aims to address the problems of over-processing and low precision in traditional magnetotelluric denoising methods, and provides a method and system for suppressing magnetotelluric signal noise based on complexity driving. The technical solution of the present invention introduces a complexity perception mechanism and sets a complexity perception gating unit, integrates the complexity evaluation mechanism, and constructs a multi-path, structurally heterogeneous feature recovery network. By introducing a complexity perception gating module, adaptive scheduling of the path activation depth and modeling strategy of different signal segments is achieved. This strategy is different from the traditional symmetric diffusion modeling method. The present invention dynamically regulates the participating modules based on the signal complexity score to form a complexity driving mechanism, so that high-complexity signal segments can obtain deeper modeling and fine-grained reconstruction, while low-complexity signal segments can only be quickly restored through lightweight paths, thereby achieving precise adaptation of path structure and modeling capabilities, and ultimately effectively removing strong interference components, avoiding over-processing, while enhancing the ability to retain target information, and significantly improving the recovery quality and reconstruction accuracy of magnetotelluric signals.
[0005] To this end, the present invention provides the following technical solutions:
[0006] In one aspect, the present invention provides a complexity-driven method for suppressing noise in magnetotelluric signals, comprising the following steps:
[0007] S1: Build a sample library, build a sample library for model training and perform segmentation processing on the sample signal, wherein the sample library at least includes a noisy signal sample set and its corresponding target noise profile sample set;
[0008] S2: Calculate the complexity feature CS of each noisy signal segment, and the complexity feature CS is used to control the path scheduling mechanism in the multi-path attention residual denoising network;
[0009] S3: Construct and train a denoising model based on a multi-path attention residual denoising network, wherein the noisy signal segment and the complexity feature CS are input into the multi-path attention residual denoising network;
[0010] The multi-path attention residual denoising network is composed of multiple structurally heterogeneous feature recovery modules in parallel to form different main branches. The output of each main branch is coordinated to input the output layer, and then outputs the noise reduction signal or noise profile; the feature recovery modules corresponding to different main branches are respectively the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT, and each main branch is provided with a complexity perception gating unit with the complexity feature CS as input, and its output signal is coordinated with the feature recovery module of the main branch where the noisy signal segment is input;
[0011] The path scheduling mechanism uses the output results of the complexity perception gating unit on each main branch to guide the feature recovery module of the main branch to participate in data processing;
[0012] S4: Use the trained denoising model based on multi-path attention residual denoising network to denoise the measured magnetotelluric signals.
[0013] Preferably, the path scheduling mechanism is: based on the module activation factor output by the complexity-aware gating unit , adaptively determine whether the feature recovery module of the main branch participates in data processing according to the hard activation strategy;
[0014] The mathematical model of the complexity-aware gating unit is: ,in, and are weight and bias constants, used to control the output range and response curve shape of the sigmoid function; is the complexity feature of the input;
[0015] The hard activation strategy is: if the module activation factor of the complexity-aware gating unit in the main branch is If the value is greater than or equal to the preset threshold, the feature recovery module of the main branch is activated and participates in data processing; otherwise, it does not participate in data processing.
[0016] Preferably, the complexity-aware gating unit determines its module parameters after training so that:
[0017] when , the module activation factor of the three main branches are all greater than or equal to a preset threshold, the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT are all activated;
[0018] when , the shallow convolution module CBL and the channel structure modeling module MACNet are activated;
[0019] when , only the shallow convolution module CBL is activated.
[0020] As can be further seen from the above, the technical solution of the present invention further designs and optimizes the multi-layer sparse structure: the first sparse layer is a CBL module, which is particularly adapted for low-complexity segments and is used for preliminary capture of large-scale noise features. It has a wide convolution kernel and a small number of channels to ensure rapid perception of the overall noise profile. The second sparse layer is an improved MACNet structure (composed of 1×1 bypass convolution, DSConv, and a fine-grained multi-path feature refinement branch), which mainly processes medium-complexity regions, focusing on channel recalibration and spatial detail extraction. The multi-path structure introduces rich gradient flows during the training phase, significantly enhancing feature expression capabilities. The third sparse layer, the global attention modeling module DHVT, is targeted at high-complexity segments. Through the above technical solution, the present invention combines a complexity-driven path adaptation strategy to achieve precise suppression of complex non-Gaussian magnetotelluric noise and high-fidelity recovery of weak effective signals.
[0021] Preferably, the complexity feature CS is the normalized result of five complexity indicators: local change rate, spectral entropy, sparsity, energy density, and autocorrelation, as follows:
[0022] Define the i-th noisy signal segment , s is the sliding window step size, L is the window length, j is the time, is the signal sampling point corresponding to time point j;
[0023] The i-th noisy signal segment The local rate of change : , is the signal sampling point corresponding to time point j+1;
[0024] The i-th noisy signal segment Spectral entropy : ,in, , is the i-th noisy signal segment The frequency domain coefficient after Fourier transformation, k is the frequency index, Indicates the energy proportion of the corresponding frequency point. Spectral entropy reflects the complexity of the spectrum energy distribution. The larger the value, the more dispersed the spectrum and the more uniform the energy distribution.
[0025] The i-th noisy signal segment Sparsity : , 、 Represents signal segments The L1 norm and L2 norm of ;
[0026] The i-th noisy signal segment Energy density : ;
[0027] The i-th noisy signal segment Autocorrelation : ,in, Represents a noisy signal segment The mean of
[0028] The formula for complexity feature CS is: , ;
[0029] satisfy , is the weight of each complexity index.
[0030] The technical solution of the present invention preferably uses five complexity indicators of local change rate, spectral entropy, sparsity, energy density, and autocorrelation to construct the complexity feature CS. This is mainly based on the following considerations: this group of indicators has good complementarity in the time domain, frequency domain and statistical feature levels, and can reflect the differences between noise and effective components in the magnetotelluric signal from different angles. Among them, the local change rate and autocorrelation characterize the temporal continuity and structure of the signal, the spectral entropy and sparsity reflect the frequency domain distribution and information complexity, and the energy density measures the signal strength characteristics. The combined use can improve the accuracy and robustness of complexity estimation, and provide a reliable basis for the dynamic structure control of subsequent modules.
[0031] Preferably, the output layer is to sum the outputs of the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT in proportion and weighted, and fuse them into a unified feature representation, or splice the outputs of the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT in the channel dimension and then input them into a 1×1 convolution fusion module.
[0032] Preferably, the channel structure modeling module MACNet consists of three parallel sub-branches;
[0033] The first sub-branch is equipped with a 1×1 bypass convolution;
[0034] The second sub-branch is the depth-wise separable convolution branch DSConv, which has 1×1 normal convolution, K×K depth-wise convolution, and 1×1 point-wise convolution in sequence;
[0035] The third sub-branch is the fine-grained multi-path feature refinement branch FMFRB;
[0036] Among them, each sub-branch in the module MACNet is embedded with a complexity-aware gating sub-unit, and the sub-branch activation factor is adaptively calculated according to the complexity feature CS. The sub-branch activation factor adopts a soft activation mechanism to indicate the degree of participation of each sub-branch in the current complexity region. That is, after each sub-branch output is multiplied by the corresponding sub-branch activation factor, it is spliced in the channel dimension and the sub-branch feature fusion is achieved through 1×1 convolution.
[0037] Preferably, the convolution module and the maximum pooling layer in the shallow convolution module CBL are alternately arranged in sequence, and the convolution module is sequentially provided with a convolution layer, a batch normalization layer and a LeakyReLU activation layer.
[0038] Preferably, the multi-head attention interaction module in the global attention modeling module DHVT assigns activation factors to three types of token labels according to the complexity feature CS. The three types of token labels are channel guide label vector (Head Token), local fragment representation vector (Patch Token) and class label vector (Class Token). The activation factor is set as:
[0039] ;
[0040] in, is the complexity segmentation threshold; is a smoothing function used to achieve soft activation, Control the slope of the activation function, They are the activation factors of the local fragment representation vector Patch Token, the category token vector ClassToken and the channel guidance token vector Head Token respectively;
[0041] Then, based on the activation factor, three types of Token tags are activated and replace the original three types of Token tags of the multi-head attention interaction module, and then the original algorithm of the multi-head attention interaction module is executed.
[0042] In another aspect, the technical solution of the present invention provides a noise reduction system based on the above-mentioned method for suppressing noise of magnetotelluric signals, comprising:
[0043] A sample library construction module is used to construct a sample library for model training and perform segmentation processing on the sample signal, wherein the sample library at least includes a noisy signal sample set and a corresponding target noise profile sample set;
[0044] A complexity feature CS calculation module is used to calculate the complexity feature CS of each noisy signal. The complexity feature CS is used in the path scheduling mechanism in the multi-path attention residual denoising network;
[0045] The denoising module constructs a training module for constructing and training a denoising model based on a multi-path attention residual denoising network, wherein the noisy signal segment and the complexity feature CS are input into the multi-path attention residual denoising network;
[0046] The multi-path attention residual denoising network is composed of multiple structurally heterogeneous feature recovery modules in parallel to form different main branches. The output of each main branch is coordinated to input the output layer, and then outputs the noise reduction signal or noise profile; the feature recovery modules corresponding to different main branches are respectively the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT, and each main branch is provided with a complexity perception gating unit with the complexity feature CS as input, and its output signal is coordinated with the feature recovery module of the main branch where the noisy signal segment is input;
[0047] The path scheduling mechanism uses the output results of the complexity perception gating unit on each main branch to guide the feature recovery module of the main branch to participate in data processing;
[0048] The denoising module is used to denoise the measured magnetotelluric signals using a trained denoising model based on a multi-path attention residual denoising network.
[0049] In a third aspect, the technical solution of the present invention provides a computer storage medium storing a computer program, which is called by a processor to implement the steps of the above-mentioned method for suppressing noise of magnetotelluric signals.
[0050] Compared with the prior art, the present invention achieves the following technical effects:
[0051] The present invention introduces complexity features as the core basis for dynamic structural regulation, mainly based on its ability to effectively characterize the degree of signal disturbance in the time domain, frequency domain and statistical level. This feature has regional sensitivity and modeling guidance, which can be used to realize soft / hard activation control of each module of the main branch, so that the network can adaptively adjust the modeling path and structural response when facing inputs of different complexities, taking into account both modeling capabilities and computational efficiency. The overall network consists of three sparse modules, namely the lightweight convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT. The three complement each other in the receptive field range and modeling granularity, and realize multi-path dynamic fusion through complexity driving.
[0052] Furthermore, the technical solution of the present invention also preferably sets up multiple sub-branches within MACNet, introducing complexity to regulate their participation level, so that the network activates fine-structure branches in high-complexity areas and retains efficient paths in low-complexity areas, improving the adaptability and expressiveness of channel modeling. Furthermore, the preferred DHVT adjusts the participation intensity of various tokens based on complexity, constructing a phased attention response mechanism, effectively reducing redundant computation while maintaining global modeling capabilities. Through the collaborative design of these mechanisms, the model exhibits superior structural flexibility and denoising performance when processing magnetotelluric signals in a strong noisy background. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 It is a schematic diagram of a process provided by an embodiment of the present invention.
[0054] Figure 2 It is a schematic diagram of the network architecture provided by an embodiment of the present invention.
[0055] Figure 3 This is a diagram showing the denoising effect of the present invention on simulated noisy data.
[0056] Figure 4 This is a diagram showing the denoising effect of the present invention on measured data.
[0057] Figure 5 It is the original apparent resistivity-phase curve of the measured data and the apparent resistivity-phase curve diagram after being processed by the method of the present invention, wherein, Figure a is the original apparent resistivity curve diagram of the measured data in the xy direction and the apparent resistivity curve diagram after being processed by the method of the present invention, Figure b is the original apparent resistivity curve diagram of the measured data in the yx direction and the apparent resistivity curve diagram after being processed by the method of the present invention, Figure c is the phase curve diagram of the measured data in the xy direction and the phase curve diagram after being processed by the method of the present invention, and Figure d is the phase curve diagram of the measured data in the yx direction and the phase curve diagram after being processed by the method of the present invention. DETAILED DESCRIPTION
[0058] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to illustrate the present invention and are not intended to limit the present invention. The technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0059] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.
[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.
[0061] This paper aims to address the complex and intense noise interference in magnetotelluric signals. It proposes a complexity-driven multipath attention residual denoising network (MARD-Net) for magnetotelluric data denoising. This method integrates a complexity assessment mechanism to construct a multipath, structurally heterogeneous feature recovery network. By introducing a complexity-aware gating module, it achieves adaptive scheduling of path activation depths and modeling strategies for different signal segments. The MARD-Net network also draws on the modeling principles of forward noisiness and backward recovery in diffusion models, preferably implementing a forward diffusion process during the training phase to generate a gradually perturbed supervisory signal.
[0062] To this end, an embodiment of the present invention provides a complexity-driven method for suppressing noise in magnetotelluric signals, comprising the following steps:
[0063] S1: Construct a sample library, construct a sample library for model training and perform segmentation processing on the sample signal, the sample library at least includes a noisy signal sample set and its corresponding target noise profile sample set.
[0064] In some embodiments, by analyzing the noise type of magnetotelluric data, different mathematical functions are used to construct a set of noise profile time series sample libraries, each with the same segment length (e.g., 200), noise amplitudes ranging from -100,000 to 100,000, and different noise widths, including triangle waves, charge-discharge triangle waves, pulse waves, and square waves. The time series position of each noise profile is then sequentially altered to obtain several noise profile sample libraries with the same segment length of 200 and different noise waveform positions. The amplitudes of all noise profile sample libraries for each noise type are then altered to obtain noise profile sample libraries with different amplitudes for the same noise type. A Gaussian white noise sample library with the same length as the noise profile sample library and an amplitude substantially close to that of the measured magnetotelluric interference-free data is selected, with an amplitude ranging from -600 to 600, to construct a clean signal sample library. The clean signal sample library and the noise profile sample library are then added one-to-one to obtain a noisy signal sample library.
[0065] It should be understood that the above-constructed clean signal sample library, noise profile sample library, and noisy signal sample library are merely examples of the present invention. In other feasible embodiments, the clean signal sample library, noise profile sample library, and noisy signal sample library constructed in other ways also meet the technical requirements of the present invention and fall within the scope of protection of the present invention.
[0066] For example, in some embodiments, a forward perturbation mechanism based on a diffusion model is employed to construct a multi-stage noise reduction task by gradually injecting noise into the original signal, thereby improving the generalization and robustness of the network. Specifically, starting from the target noise profile, perturbations of the pure signal are gradually introduced according to a preset ratio to generate a multi-stage noisy signal sequence. This process simulates the generation mechanism of the noisy signal and is used for step-by-step reconstruction supervision during the training phase to construct an inverse diffusion mapping relationship from the noisy signal to the target noise profile.
[0067] Assume the original noise profile is , a series of noisy samples are gradually generated through a predefined Gaussian diffusion process ,satisfy ,in In the diffusion process The decay factor of the step, Represents standard Gaussian noise. Then, the network training process is based on noisy samples. As input, the original noise profile To supervise the target, we train its ability to predict noise perturbations. This training mechanism is equivalent to learning a mapping from any diffusion stage back to the original signal, providing prior support for the subsequent multi-path attention residual denoising process. It also works in conjunction with the complexity-driven dynamic path control strategy to achieve joint modeling of different noise stages and path structures.
[0068] After constructing the sample library, this embodiment selects the noisy signal as input and the noise profile as output; in other feasible embodiments, the noise-reduced signal can also be selected as output. This is a technical means that can be implemented in this field, and the present invention does not specifically limit this.
[0069] S2: Calculate the complexity feature CS of each noisy signal. The complexity feature CS is used for the path scheduling mechanism in the multi-path attention residual denoising network.
[0070] In the embodiment of the present invention, the complexity feature CS is set to be the normalized result of five complexity indicators: local change rate, spectral entropy, sparsity, energy density, and autocorrelation, as follows:
[0071] Define the i-th noisy signal segment , s is the sliding window step size, L is the window length, j is the time, is the signal sampling point corresponding to time point j.
[0072] The i-th noisy signal segment The local rate of change : .
[0073] The i-th noisy signal segment Spectral entropy : ,in, , is the i-th noisy signal segment The frequency domain coefficient after Fourier transformation, k is the frequency index, It represents the energy proportion of the corresponding frequency point. The spectrum entropy reflects the complexity of the spectrum energy distribution. The larger the value, the more dispersed the spectrum and the more uniform the energy distribution.
[0074] The i-th noisy signal segment Sparsity : , , Represents signal segments The L1 norm and L2 of .
[0075] The i-th noisy signal segment Energy density : .
[0076] The i-th noisy signal segment Autocorrelation : ,in, Represents a noisy signal segment The mean of .
[0077] The formula for complexity feature CS is: , Among them, satisfy , is the weight of each complexity indicator, which is set empirically or obtained through validation set tuning.
[0078] It should be understood that in other feasible embodiments, the complexity feature can also integrate other complexity indicators or use other complexity indicators to replace some of the above complexity indicators. The technical ideas and principles remain unchanged, and the present invention does not make specific limitations on this.
[0079] S3: Construct and train a denoising model based on a multi-path attention residual denoising network, wherein the noisy signal segment and the complexity feature CS are input into the multi-path attention residual denoising network. The network body of the multi-path attention residual denoising network provided in this embodiment is composed of three structurally heterogeneous feature recovery paths in parallel, which are respectively adapted to signal characteristics of different complexity levels. The feature recovery modules (sparse modules) corresponding to different main branches are the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT, respectively, and each main branch is provided with a complexity perception gating unit with the complexity feature CS as input, and its output signal cooperates with the main branch where the noisy signal segment is input.
[0080] Among them, the low-complexity path (shallow convolution module CBL) is an interlaced convolution module and maximum pooling layer, which is used to quickly extract smooth background features; the medium-complexity path introduces the MACNet module, combining 1×1 convolution, depth-wise separable convolution and multi-branch feature refinement units to enhance structural modeling capabilities; the high-complexity path configures a dual-branch hybrid visual Transformer (DHVT) structure to improve the modeling effect of long-distance dependence and channel-level correlation.
[0081] The mathematical model of the complexity-aware gating unit is: ,in, and are weight and bias constants, used to control the output range and response curve shape of the sigmoid function; Is the complexity characteristic value of the input. If the module activation factor of the complexity-aware gate control unit of the main branch is If the value is greater than or equal to the preset threshold, the feature recovery module of the main branch is activated and participates in data processing; otherwise, it does not participate in data processing.
[0082] During network training, the module parameters of the complexity-aware gating unit are adaptively adjusted and , affecting module activation factor , which promotes the complexity characteristic value of different main branches in the input Different module activation factors are obtained under , and then guide the activation of the feature recovery module of the main branch. , the module activation factor of the three main branches are greater than or equal to the preset threshold, the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT are all activated; when , the shallow convolution module CBL and the channel structure modeling module MACNet are activated; when , only the shallow convolution module CBL is activated. Experimental results show that the proposed method is superior to traditional methods in terms of signal-to-noise ratio improvement, structure preservation and signal detail recovery, showing good generalization ability and practical application value. It should be understood that the above complexity eigenvalue The value range and the corresponding activation results of each module are feasible ways to conduct experimental deduction in this application. In other feasible embodiments, the module activation factor is directly used. Alternatively, in different applications, adaptive adjustment of the above value range may be made according to accuracy requirements, which also falls within the scope of protection of the present invention.
[0083] like Figure 2 In the left part of the diagram, the convolution modules and maximum pooling layers in the shallow convolution module CBL are alternately arranged. In some embodiments, the number of convolution modules and maximum pooling layers is determined based on accuracy requirements and experimental results. The convolution module is sequentially provided with a convolution layer, a batch normalization layer, and a LeakyReLU activation layer. To more clearly illustrate the processing process, the mathematical model is illustrated as follows:
[0084] First, the input data is processed by the convolution layer to obtain the convolution layer output , , represents the input of the convolutional layer, and denote weight and bias respectively, Represents the convolution operation;
[0085] Secondly, the output of the convolutional layer is input into the batch normalization layer. The mathematical model of the batch normalization layer is: ;in, The convolution layer output corresponding to the i-th sample The output after the batch normalization layer, and is a learnable parameter in the batch normalization layer BN, Represents the batch normalization layer (which can accelerate the convergence of the network and avoid the phenomenon of gradient dispersion), parameter satisfy:
[0086] ;
[0087] Where, is the input batch sample size, , Respectively represent the output of the batch sample The mean and standard deviation of is a constant, usually a very small constant to prevent the denominator from being zero.
[0088] Finally, the output of the batch normalization layer is processed by the activation function LeakyReLU, specifically:
[0089]
[0090] Where, is the input of the activation function LeakyReLU, that is, the output of the batch normalization layer of the batch sample constitute, It should be understood that the preferred activation function LeakyReLU in this embodiment is an improvement on ReLU. In other feasible embodiments, typical activation functions include sigmoid, tanh, and rectified linear unit (ReLU).
[0091] In summary, the convolutional layer of the shallow convolutional module (CBL) in the first sparse layer of the present invention performs local linear mapping on the input signal to extract spatial or temporal features. The batch normalization layer performs mean-variance normalization on the activation values of each mini-batch, reducing internal covariate shift, stabilizing training, and accelerating convergence. The nonlinear mapping layer introduces nonlinear expression capabilities to the network, enabling the model to fit complex functions and weak signals. The combination of these three significantly improves the network's feature extraction quality, training stability, and convergence speed.
[0092] like Figure 2 In the middle part shown, the channel structure modeling module MACNet consists of three parallel sub-branches; the first sub-branch is equipped with 1×1 bypass convolution; the second sub-branch is the depth-separable convolution branch DSConv, which is equipped with 1×1 normal convolution, K×K depth-wise convolution, and 1×1 point-by-point convolution in sequence; the third sub-branch is the fine-grained multi-path feature refinement branch FMFRB. First, the input features are split into several sub-groups along the channel, and each group independently refines fine-grained information in the lightweight ConvNeck, and then the outputs of all sub-groups are spliced.
[0093] The mathematical model of the module MACNet is expressed as:
[0094] ;
[0095] Where, is the input feature of the module MACNet, Perform a convolution on the input to obtain an intermediate feature representation, which is the starting point of all subsequent paths; It is a bypass convolution to extract basic features (corresponding to a sub-branch); For the middle branch, first do ordinary convolution and then separable convolution to increase model efficiency and receptive field; for the third sub-branch, For the features after the branch layer operation, the intermediate features are split into multiple channels or sub-features to construct sparse convolution paths. is the result obtained after sparse convolution branch, It is a lightweight convolution module. The residual connection enhances the training stability. Its subscript is used to distinguish different lightweight convolution modules. ; is the output of the last convolution branch. The number of channels is 2c. And the output 、 、…、 、 Each of them has c channels. Finally, through concatenation and then 1×1 convolution, the information of these three features is fused and compressed to generate a channel number of 2c , as shown below:
[0096] .
[0097] It should be understood that in some embodiments, the above-mentioned module MACNet has met the technical requirements of the present invention, and the above-mentioned mathematical model is only an example. For example, the number of paths in the FMFRB branch can be adjusted according to actual accuracy and application requirements.
[0098] In other embodiments, the technical solution of the present invention also preferably optimizes the module MACNet, namely, each sub-branch is embedded with a complexity-aware gating sub-unit, and the sub-branch activation factor is adaptively calculated based on the complexity feature CS. The sub-branch activation factor uses a soft activation mechanism to represent the degree of participation of each sub-branch in the current complexity region. That is, after each sub-branch output is multiplied by the corresponding sub-branch activation factor, the sub-branch features are spliced in the channel dimension and fused through 1×1 convolution. This can adaptively adjust the fusion and splicing process of the above mathematical model.
[0099] like Figure 2As shown on the right side, the global attention modeling module (DHVT) primarily comprises four stages: embedding, attention interaction, feature aggregation, and residual fusion. This module architecture is currently available and will not be elaborated on in detail in this invention. It should be understood that the existing global attention modeling module (DHVT) also meets the requirements of the technical solution of this invention. However, to enhance noise reduction, this invention optimizes the attention interaction stage in some embodiments: the multi-head attention interaction module further adjusts the response of various tokens involved in attention modeling based on the specific values of the activation factors, forming a token-based soft activation mechanism. The multi-head attention interaction module includes a channel guidance token vector (HeadToken), a local segment representation vector (Patch Token), and a class token vector (Class Token). Patch Tokens extract feature representations of local segments by segmenting or sliding windowing the original one-dimensional signal, serving as the basis for modeling the local structure of the signal. Class Tokens are global summary vectors used to represent the overall characteristics of the input signal. They serve as aggregation and guidance in the attention mechanism, enhancing global perception. After the introduction of the multi-head attention mechanism, Head Tokens are structural guidance vectors constructed for modeling the interaction between channel groups or feature paths, controlling the interaction strength of information between channels. This part is prior art and will not be described in detail. It is briefly described as follows:
[0100] Assume that the input feature dimension of the multi-head attention interaction module is D, and the number of heads is h. First, the HTG module generates h Head Token vectors to represent the channel features of each attention head. The generation formula is:
[0101]
[0102] Among them, AvgSplit means dividing the input into h groups according to the channel and then calculating the average to generate h Dimension's HeadToken; and represents the activation function and the fully connected transformation, and t is the input feature vector.
[0103] The one-dimensional signal fragments input to the global attention modeling module DHVT are divided into sliding windows or segments to extract local fragments, each fragment is fed into a convolutional or fully connected network encoding, and the result is as follows: Represented as a Patch Token sequence, used to model local features.
[0104] The one-dimensional signal fragments input to the global attention modeling module DHVT are pooled, globally encoded, or specially initialized to obtain a global summary vector C, represented as a Class Token, which participates in aggregating global information and guiding attention calculation.
[0105] The improvement of the embodiment of the present invention is that after the DHVT module is activated, the token type and response weight involved in the attention calculation are further controlled according to the CS value of the signal segment. That is, according to the value range of the characteristic quantity CS, activation factors are assigned to the three types of tokens respectively:
[0106] ;
[0107] in, is the complexity segmentation threshold; is a smoothing function used to achieve soft activation, Control the slope of the activation function. According to the above activation factor, when CS is between 0.7 and 0.85, only PatchToken and Class Token are enabled, and the participation of Head Token is limited; when CS is higher than 0.85, all tokens are enabled and the weight contribution of Head Token in attention calculation is significantly enhanced, achieving structural adaptive control of token granularity.
[0108] The combination T of the last three types of tokens input into the original multi-head attention interaction module is: All activated tokens are labeled and replace the original three-class token labels of the multi-head attention interaction module. The original algorithm of the multi-head attention interaction module is then executed. It should be understood that the original algorithm of the multi-head attention interaction module is existing technology and will not be explained in detail. Generally speaking, it establishes global interactions between tokens, allowing each token to update itself based on "global information" and autonomously aggregate information based on their interrelationships. In the attention module, all tokens are first mapped into three sets of vectors: query, key, and value. The multi-head attention mechanism then calculates attention weights for each of these vectors and aggregates the corresponding information. Ultimately, the output of the head token is integrated into the class token representation to enhance its global modeling capabilities, while the output of the patch token serves as the local representation input for subsequent feature extraction or classification, completing the structure-adaptive feature modeling process. After completing the multi-head attention interaction, the DHVT module performs feature aggregation through a feedforward network and fuses the updated results with the original input using a residual mechanism to improve model stability and expressiveness. The final output is a structure-enhanced global feature representation.
[0109] In order to cooperate with the complexity-driven dynamic path control mechanism, the network can pay more attention to the error modeling of high-complexity areas during training, and further improve the fidelity and structural consistency of the output noise contour map. In this embodiment of the present invention, a complexity-aware and structure-preserving fusion loss function CAD-Loss (Complexity-Aware and Detail-Preserving Loss) is preferably used. It consists of three parts: basic reconstruction error term, structural gradient preservation term, and complexity weighted error term. The overall definition is as follows:
[0110]
[0111] Where, The reconstruction error term , structural gradient preservation term and the complexity-weighted error term The weight of .
[0112] Among them, the basic reconstruction error term Used to measure network prediction output With the target noise profile The basic difference between , N represents the total number of training samples.
[0113] The structural gradient preservation term is used to enhance the model's ability to retain signal mutation areas (such as edges or high-frequency components). The first-order gradient difference is introduced as a structural preservation indicator, which is defined as: ,in The gradient of a one-dimensional signal, such as the difference between adjacent points, can effectively reflect the trend of signal changes.
[0114] The complexity weighted error term takes into account the complexity differences of different signal segments and designs a weighting mechanism to score the complexity. In order to guide the model to focus on high-complexity samples based on adaptive weighting of sample errors, the definition is as follows: , where the complexity score , calculated by the above steps. Weight function The definition is as follows: ,in is a tunable hyperparameter.
[0115] The back propagation of MARD-Net training updates the weights and biases. The specific formula is as follows:
[0116] ;
[0117] Where W represents the weight of MARD-Net network training, represents the bias of MARD-Net network training, η represents the learning rate, Represents the error function, that is, .
[0118] S4: Use the trained denoising model based on multi-path attention residual denoising network to denoise the measured magnetotelluric signals. Figure 3 and Figure 4 The denoising effects obtained after using MARD-Net to process the simulated data and the measured data are shown. Figure 5 The comparison between the original apparent resistivity-phase curve and the apparent resistivity-phase curve processed by the method of the present invention is shown in FIG. Figure 5 As can be seen, the proposed method can significantly suppress strong interference in magnetotelluric signals, effectively reducing the obvious near-source effect in the original curve and improving the phase response. This method not only has high data processing efficiency but also demonstrates excellent denoising capabilities even in strong interference environments, demonstrating promising application potential in magnetotelluric data processing.
[0119] In other feasible embodiments, the technical solution of the present invention provides a noise reduction system based on the above-mentioned magnetotelluric signal noise suppression method, including: a sample library construction module, a complexity feature CS calculation module, a noise reduction module construction training module and a noise reduction module.
[0120] Among them, the sample library construction module constructs a sample library for model training and performs segmented processing on the sample signal. The sample library at least includes a noisy signal sample set and its corresponding target noise profile sample set; the complexity feature CS calculation module is used to calculate the complexity feature CS of each segment of the noisy signal; the denoising module constructs a training module for constructing and training a denoising model based on a multi-path attention residual denoising network, wherein the noisy signal segment and the complexity feature CS are input into the multi-path attention residual denoising network; the multi-path attention residual denoising network is composed of multiple structurally heterogeneous feature recovery modules in parallel to form different main branches, and the output of each main branch is collaboratively input into the output layer, and the output Denoising signal or noise contour; the feature recovery modules corresponding to different main branches are the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT, and each main branch is equipped with a complexity-aware gating unit with the complexity feature CS as input, and its output signal is coordinated with the noisy signal segment to input the feature recovery module of the main branch; the path scheduling mechanism is to guide whether the feature recovery module of the main branch participates in data processing based on the output results of the complexity-aware gating unit on each main branch; the denoising module is used to denoise the measured magnetotelluric signal using the trained denoising model based on the multi-path attention residual denoising network.
[0121] It should also be understood that the specific implementation process of each module please refer to the above method content, the present invention will not go into details here, and the division of the above functional modules is only for example illustration. In some embodiments, some functional modules can be merged, and some functional modules can be split. Each functional module can be implemented in software or hardware or a combination of software and hardware. Among them, software and hardware equipment include but are not limited to general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.
[0122] In other feasible embodiments, the technical solution of the present invention provides a computer storage medium that stores a computer program, and the computer program is called by a processor to implement the steps of the method for suppressing noise of magnetotelluric signals.
[0123] Specifically, the following are implemented:
[0124] S1: Build a sample library, build a sample library for model training and perform segmentation processing on the sample signal, the sample library at least includes a noisy signal sample set and its corresponding target noise profile sample set.
[0125] S2: Calculate the complexity feature CS of each noisy signal. The complexity feature CS is used for the path scheduling mechanism in the multi-path attention residual denoising network.
[0126] S3: Construct and train a denoising model based on a multi-path attention residual denoising network, in which the noisy signal segment and complexity feature CS are input into the multi-path attention residual denoising network. The network body of the multi-path attention residual denoising network is composed of three structurally heterogeneous feature recovery paths in parallel, which are adapted to signal characteristics of different complexity levels.
[0127] S4: Use the trained denoising model based on multi-path attention residual denoising network to denoise the measured magnetotelluric signals.
[0128] For the specific implementation process of each step, please refer to the description of the above method.
[0129] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the software and hardware device of any of the aforementioned embodiments, such as a hard disk or memory of a controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk equipped on the controller, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. Furthermore, the readable storage medium can also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used to store computer programs 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 is to be output.
[0130] Based on this understanding, the technical solution of the present invention, or the portion 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 instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0131] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is a flow chart according to the method, device (system), and computer program product of the embodiment of the present application and / or the instructions executed by the processor to generate a device for realizing the function specified in one flow chart or multiple flows and / or one box or multiple boxes of the block diagram. These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a product comprising an instruction device, which realizes the function specified in one flow chart or multiple flows and / or one box or multiple boxes of the block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.
[0132] In other feasible embodiments, the technical solution of the present invention provides an electronic device, comprising: one or more processors and a memory storing one or more computer programs, wherein the processor calls the computer program to implement: the steps of the method for suppressing noise of magnetotelluric signals.
[0133] Specifically, the following are implemented:
[0134] S1: Build a sample library, build a sample library for model training and perform segmentation processing on the sample signal, the sample library at least includes a noisy signal sample set and its corresponding target noise profile sample set.
[0135] S2: Calculate the complexity feature CS of each noisy signal. The complexity feature CS controls the path scheduling mechanism in the multi-path attention residual denoising network.
[0136] S3: Construct and train a denoising model based on a multi-path attention residual denoising network, in which the noisy signal segment and complexity feature CS are input into the multi-path attention residual denoising network. The network body of the multi-path attention residual denoising network is composed of three structurally heterogeneous feature recovery paths in parallel, which are adapted to signal characteristics of different complexity levels.
[0137] S4: Use the trained denoising model based on multi-path attention residual denoising network to denoise the measured magnetotelluric signals.
[0138] For the specific implementation process of each step, please refer to the description of the above method.
[0139] It should be understood that in the embodiments of the present invention, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.
[0140] It should be emphasized that the examples of the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the examples in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.
Claims
1. A complexity-driven method for suppressing noise in magnetotelluric signals, characterized by: The following steps are involved: S1: Build a sample library, build a sample library for model training and perform segmentation processing on the sample signal, wherein the sample library at least includes a noisy signal sample set and its corresponding target noise profile sample set; S2: Calculate the complexity feature CS of each noisy signal, which is used for the path scheduling mechanism in the multi-path attention residual denoising network; S3: Construct and train a denoising model based on a multi-path attention residual denoising network, wherein the noisy signal segment and the complexity feature CS are input into the multi-path attention residual denoising network; The multi-path attention residual denoising network is composed of multiple structurally heterogeneous feature recovery modules in parallel to form different main branches. The output of each main branch is coordinated to input the output layer, and then outputs the noise reduction signal or noise profile; the feature recovery modules corresponding to different main branches are respectively the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT, and each main branch is provided with a complexity perception gating unit with the complexity feature CS as input, and its output signal is coordinated with the feature recovery module of the main branch where the noisy signal segment is input; The path scheduling mechanism uses the output results of the complexity perception gating unit on each main branch to guide the feature recovery module of the main branch to participate in data processing; S4: Use the trained denoising model based on multi-path attention residual denoising network to denoise the measured magnetotelluric signals.
2. The method for suppressing magnetotelluric signal noise according to claim 1, wherein: The path scheduling mechanism is: based on the module activation factor output by the complexity-aware gating unit , adaptively determine whether the feature recovery module of the main branch participates in data processing according to the hard activation strategy; The mathematical model of the complexity-aware gating unit is: ,in, and are weight and bias constants, used to control the output range and response curve shape of the sigmoid function; is the complexity feature of the input; The hard activation strategy is: if the module activation factor of the complexity-aware gating unit in the main branch is If the value is greater than or equal to the preset threshold, the feature recovery module of the main branch is activated and participates in data processing; otherwise, it does not participate in data processing.
3. The method for suppressing magnetotelluric signal noise according to claim 2, wherein: The complexity-aware gating unit determines its module parameters after training, such that: when , the module activation factor of the three main branches are all greater than or equal to a preset threshold, the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT are all activated; when , the shallow convolution module CBL and the channel structure modeling module MACNet are activated; when , only the shallow convolution module CBL is activated.
4. The method for suppressing magnetotelluric signal noise according to claim 1, wherein: The complexity feature CS is the normalized result of five complexity indicators: local change rate, spectral entropy, sparsity, energy density, and autocorrelation, as follows: Define the i-th noisy signal segment , s is the sliding window step size, L is the window length, j is the time, is the signal sampling point corresponding to time point j; The i-th noisy signal segment The local rate of change : , is the signal sampling point corresponding to time point j+1; The i-th noisy signal segment Spectral entropy : ,in, , is the i-th noisy signal segment The frequency domain coefficient after Fourier transformation, k is the frequency index, Indicates the energy proportion of the corresponding frequency point. Spectral entropy reflects the complexity of the spectrum energy distribution. The larger the value, the more dispersed the spectrum and the more uniform the energy distribution. The i-th noisy signal segment Sparsity : , 、 Represents signal segments The L1 norm and L2 norm of ; The i-th noisy signal segment Energy density : ; The i-th noisy signal segment Autocorrelation : ,in, Represents a noisy signal segment The mean of The formula for complexity feature CS is: , ; satisfy , is the weight of each complexity index.
5. The method for suppressing magnetotelluric signal noise according to claim 1, wherein: The output layer is to sum the outputs of the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT in proportion and fuse them into a unified feature representation, or to splice the outputs of the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT in the channel dimension and then input them into a 1×1 convolution fusion module.
6. The method for suppressing magnetotelluric signal noise according to claim 1, wherein: The channel structure modeling module MACNet consists of three parallel sub-branches; The first sub-branch is equipped with a 1×1 bypass convolution; The second sub-branch is the depth-wise separable convolution branch DSConv, which has 1×1 normal convolution, K×K depth-wise convolution, and 1×1 point-wise convolution in sequence; The third sub-branch is the fine-grained multi-path feature refinement branch FMFRB; Among them, each sub-branch in the module MACNet is embedded with a complexity-aware gating sub-unit, and the sub-branch activation factor is adaptively calculated according to the complexity feature CS. The sub-branch activation factor adopts a soft activation mechanism to indicate the degree of participation of each sub-branch in the current complexity region. That is, after each sub-branch output is multiplied by the corresponding sub-branch activation factor, it is spliced in the channel dimension and the sub-branch feature fusion is achieved through 1×1 convolution.
7. The method for suppressing magnetotelluric signal noise according to claim 1, wherein: The convolution module and the maximum pooling layer in the shallow convolution module CBL are alternately arranged in sequence, and the convolution module is sequentially provided with a convolution layer, a batch normalization layer and a LeakyReLU activation layer.
8. The method for suppressing magnetotelluric signal noise according to claim 1, wherein: The multi-head attention interaction module in the global attention modeling module DHVT assigns activation factors to three types of token labels according to the complexity feature CS. The three types of token labels are the channel guidance label vector Head Token, the local fragment representation vector PatchToken, and the category label vector Class Token. The activation factor is set as: ; in, is the complexity segmentation threshold; is a smoothing function used to achieve soft activation, Control the slope of the activation function, They are the activation factors of the local segment representation vector Patch Token, the category tag vector Class Token, and the channel guidance tag vector Head Token respectively; Then, based on the activation factor, three types of Token tags are activated and replace the original three types of Token tags of the multi-head attention interaction module, and then the original algorithm of the multi-head attention interaction module is executed.
9. A noise reduction system based on the method for suppressing magnetotelluric signal noise according to any one of claims 1 to 8, characterized in that: include: A sample library construction module is used to construct a sample library for model training and perform segmentation processing on the sample signal, wherein the sample library at least includes a noisy signal sample set and a corresponding target noise profile sample set; A complexity feature CS calculation module is used to calculate the complexity feature CS of each noisy signal. The complexity feature CS is used in the path scheduling mechanism in the multi-path attention residual denoising network; The denoising module constructs a training module for constructing and training a denoising model based on a multi-path attention residual denoising network, wherein the noisy signal segment and the complexity feature CS are input into the multi-path attention residual denoising network; The multi-path attention residual denoising network is composed of multiple structurally heterogeneous feature recovery modules in parallel to form different main branches. The output of each main branch is coordinated to input the output layer, and then outputs the noise reduction signal or noise profile; the feature recovery modules corresponding to different main branches are respectively the shallow convolution module CBL, the channel structure modeling module MACNet and the global attention modeling module DHVT, and each main branch is provided with a complexity perception gating unit with the complexity feature CS as input, and its output signal is coordinated with the feature recovery module of the main branch where the noisy signal segment is input; The path scheduling mechanism uses the output results of the complexity perception gating unit on each main branch to guide the feature recovery module of the main branch to participate in data processing; The denoising module is used to denoise the measured magnetotelluric signals using a trained denoising model based on a multi-path attention residual denoising network.
10. A computer storage medium, characterized in that: A computer program is stored, which is called by a processor to implement: The steps of the method for suppressing magnetotelluric signal noise according to any one of claims 1 to 8.
Citation Information
Patent Citations
RAW domain night scene image denoising method based on improved Unet
CN115393212A
Image denoising method based on lightweight U-shaped structure network
CN116229083A
Magnetic resonance signal power frequency noise suppression method based on recurrent neural network
CN116561515A
Belt conveyor carrier roller fault diagnosis method and system based on multi-scale feature fusion and residual mask convolution attention algorithm
CN117421581A
Magnetotelluric signal denoising method and system based on cross-scale connection network
CN118260535A
Cited By
Magnetotelluric signal denoising method and system based on multi-path fusion network
CN120972273A
Magnetic anomaly data de-noising method and de-noising system based on step-by-step U-Net
CN121257332A
Multi-source electromagnetic noise suppression method based on noise classification and deep learning
CN122153262A
A Multi-Source Electromagnetic Noise Suppression Method Based on Noise Classification and Deep Learning
CN122153262B
A Magnetotelluric Denoising Method and System Based on Physics-Driven and Conditional Diffusion Models
CN122362521A
