A data denoising method, device and medium for magnetotelluric signals

Abstract

The application discloses a data denoising method and device for magnetotelluric signals and a medium, and belongs to the technical field of magnetotelluric signals, and is used for solving the technical problem that the signal-to-noise separation of the existing magnetotelluric signals is not accurate enough, effective magnetotelluric signals are difficult to obtain, and the main effective components of the magnetotelluric signals cannot be accurately extracted. The method comprises the following steps: performing signal superposition processing between the effective magnetotelluric signals and the simulated synthetic noise waveform to generate a magnetotelluric signal sample pair; sequentially performing feature denoising processing on the characteristic variables in the magnetotelluric signal sample pair in the noise intensity perception, weight dynamic adjustment and self-attention weighted output; performing two-stage denoising feature sequence generation strategy calculation on the original signal time sequence in the magnetotelluric signal sample pair based on a convolution layer; and performing denoising control and feature extraction on the current original magnetotelluric signal to obtain effective signal features.

Description

Technical Field

[0001] This application relates to the field of magnetotelluric signals, and in particular to a method, device and medium for data noise reduction of magnetotelluric signals. Background Technology

[0002] Magnetotelluric (MT) sounding is widely used in resource exploration and deep electrical structure imaging due to its large detection depth and simple measurement method. However, the natural MT signals obtained from the surface are inherently weak and highly susceptible to various types of anthropogenic noise interference, leading to distorted response results and affecting the accuracy of subsequent inversion and geological interpretation. With the continuous development of modern industrial activities, the types and intensities of environmental and cultural noise are becoming increasingly complex, making MT signal quality issues increasingly prominent. Therefore, improving the denoising capability of MT data has become a key aspect in promoting its application and development.

[0003] Besides the inherent weakness of magnetotelluric (MT) signals, another major challenge in MT signal processing lies in their extremely wide frequency range. Within this band, almost all types of noise contain effective components, making signal-to-noise separation extremely difficult and increasing the difficulty of fully acquiring effective information at different frequencies. Existing methods often attempt to distinguish effective signals from noise simultaneously across the entire frequency band, but this inevitably results in the loss of some effective information. Furthermore, MT signals exhibit high randomness in composition and intensity; when the effective signal and man-made noise are not significantly different in energy or waveform, the separation difficulty increases significantly. This is especially true in the low-frequency band, where signal energy fluctuations are greater, making denoising more complex than in the mid-to-high-frequency bands.

[0004] Although existing technologies have achieved automatic recognition of high-quality signals based on convolutional neural networks and combined adaptive dictionary learning to determine the learning termination condition, reducing the loss of low-frequency signals in high-quality data segments, the loss of low-frequency information in noisy segments is still inevitable; moreover, the long-period characteristics of low-frequency signals have not been fully considered, resulting in poor low-frequency processing performance. Summary of the Invention

[0005] This application provides a data noise reduction method, device, and medium for magnetotelluric signals to solve the following technical problems: the signal-to-noise separation of existing magnetotelluric signals is not accurate enough, making it difficult to obtain effective magnetotelluric signals and making it difficult to accurately extract the main effective components of magnetotelluric signals.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] On one hand, embodiments of this application provide a data denoising method for magnetotelluric signals, comprising: superimposing effective magnetotelluric signals with simulated synthetic noise waveforms, and generating magnetotelluric signal sample pairs based on the original magnetotelluric signals; performing feature denoising processing on the feature variables in the magnetotelluric signal sample pairs sequentially according to the dynamic weight self-attention mechanism in the DnCNN network framework, involving noise intensity perception, dynamic weight adjustment, and self-attention weighted output, to obtain a denoising feature sequence generation strategy; calculating the original signal time series in the magnetotelluric signal sample pairs using the two-level denoising feature sequence generation strategy based on convolutional layers through a pre-configured DnCNN-Dynamic-SAM network; generating and obtaining a trained DnCNN-Dynamic-SAM network model based on the feature denoising model training strategy of the original signal time series; and performing denoising control and feature extraction on the current original magnetotelluric signals based on the DnCNN-Dynamic-SAM network model to obtain effective signal features after local noise suppression and global feature smoothing.

[0008] This application's embodiments utilize a pre-configured DnCNN-Dynamic-SAM network to accurately extract low-frequency components from noisy time series, effectively avoiding the loss of valid MT signals. Compared to traditional methods, this approach relies on a deep learning model, reducing manual intervention and achieving higher processing efficiency, making it particularly suitable for batch processing of large-scale MT data. Experimental results on synthetic and measured data demonstrate that this method exhibits excellent denoising performance when handling various noise types, including pulses, square waves, and triangular waves. Applying this method not only significantly improves the stability and reliability of apparent resistivity and phase curves but also outperforms traditional robust estimation methods in overall performance. Furthermore, it enables accurate extraction of low-frequency dominant components, effectively solving the problem of efficient and high-precision denoising of low-frequency MT signals, thereby improving the inversion effect in noise-affected areas.

[0009] In one feasible implementation, a signal superposition process is performed between the effective magnetotelluric signal and the simulated synthetic noise waveform, and magnetotelluric signal sample pairs are generated based on the original magnetotelluric signal. Specifically, this includes: extracting the effective magnetotelluric signal in the noise-free area based on the original magnetotelluric signal and determining it as an effective signal sample; simulating various types of noise waveforms and obtaining simulated synthetic noise waveforms based on the noise characteristics corresponding to the noise pollution signals; superimposing the synthetic noise waveforms onto the effective signal samples to generate noise samples; and pairing the noise samples with the effective signal samples to obtain several magnetotelluric signal sample pairs.

[0010] In one feasible implementation, before performing noise intensity perception, dynamic weight adjustment, and self-attention weighted output-based feature denoising processing on the feature variables of the magnetotelluric signal sample pair according to the dynamic weight self-attention mechanism in the DnCNN network framework to obtain a denoising feature sequence generation strategy, the method further includes: extracting intermediate feature variables from the magnetotelluric signal sample pair using the DnCNN network framework; calculating the noise intensity of the input feature sequence under local energy perception based on the intermediate feature variables, time series length, and feature dimension to obtain local noise energy; calculating the local interference degree of the signal under different noise environments using the Sigmoid activation function and the bias term of the one-dimensional convolutional layer in the DnCNN network framework to obtain a noise intensity coefficient; calculating the sensitivity of the attention response to the noise intensity coefficient based on the dynamic adjustment amplitude term to obtain dynamic scaling weights; and using the dynamic scaling weights, performing noise dynamic change tracking calculations on the query matrix, key matrix, and value matrix in the self-attention mechanism related to the dynamic scaling factor, and constructing a noise-aware attention model for noise intensity perception.

[0011] In one feasible implementation, based on the dynamic weight self-attention mechanism in the DnCNN network framework, the feature variables in the magnetotelluric signal sample pair are sequentially subjected to noise intensity perception, dynamic weight adjustment, and self-attention weighted output feature denoising processing to obtain a denoising feature sequence generation strategy. Specifically, this includes: extracting intermediate feature sequences from the magnetotelluric signal sample pair through multi-layer convolution and normalization processing in the DnCNN network framework; performing local energy calculation on the intermediate feature sequences using the noise-aware attention model to generate a noise intensity vector; performing weight adaptive adjustment calculation on the noise intensity vector according to the noise level adaptive mechanism to obtain a dynamic scaling coefficient; fusing the dynamic scaling coefficient with the self-attention mechanism and obtaining an optimized feature sequence under the self-attention weighted output based on the adjusted weight distribution; and obtaining the denoising feature sequence generation strategy for magnetotelluric signal denoising based on the processing and generation model of the optimized feature sequence.

[0012] In one feasible implementation, before constructing the noise-aware attention model for the noise intensity perception, the method further includes: according to This yields a comprehensive loss function used for joint optimization of the dynamic adjustment amplitude term and the bias term. ;in, This is the mean square error term, used to measure the noise reduction accuracy; This is the envelope entropy term, used to constrain signal smoothness; The balance coefficient is used; through the comprehensive loss function, the noise-aware attention model is subjected to real-time attention distribution update and adjustment processing related to signal-to-noise ratio, energy distribution and spectral smoothness, and a joint optimization strategy for the noise-aware attention model is generated.

[0013] In one feasible implementation, a pre-configured DnCNN-Dynamic-SAM network is used to perform a two-level noise reduction feature sequence generation strategy based on convolutional layers on the original signal time series of the magnetotelluric signal sample pair. Specifically, this includes: obtaining the original signal time series of the magnetotelluric signal sample pair using the pre-configured DnCNN-Dynamic-SAM network; wherein the DnCNN-Dynamic-SAM network is a fusion neural network based on deep convolution, dynamic weights, and a self-attention mechanism; extracting the basic temporal features of the pre-processed original signal time series through the first convolutional layer of the DnCNN-Dynamic-SAM network; and then converting the basic temporal features after initial convolution... The first intermediate feature sequence output from the 16-layer convolutional block is input into a 16-layer convolutional block. Through the Dynamic-SAM module corresponding to the denoising feature sequence generation strategy, the first intermediate feature sequence is processed by noise perception and self-attention weighting to obtain a first optimized feature sequence. Through a second convolutional layer, the first optimized feature sequence is processed by feature mapping and channel mapping to obtain a second intermediate feature sequence. Through the Dynamic-SAM module corresponding to the denoising feature sequence generation strategy, the second intermediate feature sequence is processed by dynamic weight calculation and self-attention weighting optimization to obtain a second optimized feature sequence. The second optimized feature sequence is then input into the third convolutional layer of the DnCNN-Dynamic-SAM network to obtain a feature denoising model training processing strategy for magnetotelluric signals.

[0014] In one feasible implementation, a feature denoising model training and processing strategy based on the original signal time series is used to generate and obtain a trained DnCNN-Dynamic-SAM network model. Specifically, this includes: iteratively training the model for feature denoising on several pairs of magnetotelluric signal samples based on the feature denoising model training and processing strategy based on the original signal time series; mining and processing the deep features in the magnetotelluric signal according to the feature denoising algorithm trained by the model iterative training, and capturing the global dependencies of the signal in the time and frequency dimensions, thereby generating and obtaining the trained DnCNN-Dynamic-SAM network model.

[0015] In one feasible implementation, based on the DnCNN-Dynamic-SAM network model, the current raw magnetotelluric signal is subjected to noise reduction control and feature extraction to obtain effective signal features after local noise suppression and global feature smoothing. Specifically, this includes: performing local noise suppression calculations on the current raw magnetotelluric signal using the DnCNN-Dynamic-SAM network model under a relevant noise reduction feature sequence generation strategy; performing global feature smoothing control on the noise-reduced features after local noise suppression calculations using the DnCNN-Dynamic-SAM network model under a relevant secondary noise reduction feature sequence generation strategy; and outputting the effective signal features in the current raw magnetotelluric signal based on the model noise reduction calculation results under the local noise suppression and global feature smoothing processing.

[0016] Secondly, embodiments of this application also provide a data noise reduction device for magnetotelluric signals, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to execute a data noise reduction method for magnetotelluric signals as described in any of the above embodiments.

[0017] Thirdly, embodiments of this application also provide a non-volatile computer storage medium, wherein the storage medium is a non-volatile computer-readable storage medium, the non-volatile computer-readable storage medium stores at least one program, each program including instructions, and the instructions, when executed by a terminal, cause the terminal to execute a data noise reduction method for magnetotelluric signals as described in any of the above embodiments.

[0018] This application provides a method, device, and medium for data noise reduction of magnetotelluric signals. Compared with the prior art, the embodiments of this application have the following beneficial technical effects:

[0019] This application's embodiments utilize a pre-configured DnCNN-Dynamic-SAM network to accurately extract low-frequency components from noisy time series, effectively avoiding the loss of valid MT signals. Compared to traditional methods, this approach relies on a deep learning model, reducing manual intervention and achieving higher processing efficiency, making it particularly suitable for batch processing of large-scale MT data. Experimental results on synthetic and measured data demonstrate that this method exhibits excellent denoising performance when handling various noise types, including pulses, square waves, and triangular waves. Applying this method not only significantly improves the stability and reliability of apparent resistivity and phase curves but also outperforms traditional robust estimation methods in overall performance. Furthermore, it enables accurate extraction of low-frequency dominant components, effectively solving the problem of efficient and high-precision denoising of low-frequency MT signals, thereby improving the inversion effect in noise-affected areas. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0021] Figure 1 A flowchart of a data noise reduction method for magnetotelluric signals provided in this application embodiment;

[0022] Figure 2 This is a schematic diagram of the time-domain denoising effect of a measurement data segment provided in an embodiment of this application; wherein, (a) is a time series segment image of impulse noise; (d) is a time series segment image of impulse noise after denoising; (b) is a time series segment image of triangular wave noise; (e) is a time series segment image of triangular wave noise after denoising; (c) is a time series segment image of mixed noise contamination; and (f) is a time series segment image of mixed noise contamination after denoising.

[0023] Figure 3 This is a schematic diagram comparing the data segments before and after denoising at measurement point E1-141 provided in the embodiments of this application; wherein, (a) is the noise interference waveform in the Ex channel; (e) is the denoised waveform in the Ex channel; (b) is the noise interference waveform in the Ey channel; (f) is the denoised waveform in the Ey channel; (c) is the complex noise waveform with pulse signals mixed in the Hx magnetic channel; (g) is the denoised waveform in the Hx magnetic channel; (d) is the complex noise waveform with pulse signals mixed in the Hy magnetic channel; and (h) is the denoised waveform in the Hy magnetic channel.

[0024] Figure 4This is a schematic diagram of a data noise reduction device for magnetotelluric signals provided in an embodiment of this application. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0026] This application provides a data noise reduction method for magnetotelluric signals, such as... Figure 1 As shown, the data denoising method for magnetotelluric signals specifically includes steps S101-S104:

[0027] S101. Perform signal superposition processing between the effective magnetotelluric signal and the simulated synthetic noise waveform, and generate magnetotelluric signal sample pairs based on the original magnetotelluric signal.

[0028] Specifically, the first step is to extract the effective magnetotelluric signals from the noise-free area based on the original magnetotelluric signals and identify them as effective signal samples.

[0029] Furthermore, various types of noise waveforms are simulated, and based on the noise characteristics corresponding to the noise pollution signal, simulated synthetic noise waveforms are obtained.

[0030] Furthermore, the synthesized noise waveform is superimposed onto the effective signal sample to generate a noise sample.

[0031] Furthermore, the noise samples are paired with the valid signal samples to obtain several pairs of magnetotelluric signal samples.

[0032] In one embodiment, this application focuses on developing a deep learning model to accurately extract the main effective components from MT (Gallium Earth Electromagnetic Signal) data. High-precision signal-to-noise separation is achieved, ultimately obtaining complete and effective MT data. To achieve accurate extraction of the effective components, a precise sample library is required. To enable the model to learn the mapping relationship between noise signals and effective signals, paired samples containing noise-contaminated samples and corresponding noise-free samples are generated. Therefore, high-quality signals collected from areas without strong noise are considered effective signal samples. Subsequently, various types of noise waveforms are simulated to simulate the observed noise characteristics of noise-contaminated signals in field measurements. Finally, these synthesized noise waveforms are superimposed on the effective signal samples to generate noise samples. Noisy signals are paired with their corresponding noise-free original signals to form sample pairs. After preliminary preprocessing, the prepared samples are fed into the DnCNN-SAM deep learning model for training. Subsequently, the trained model is used to process actual MT data, where DnCNN is a deep convolutional neural network specifically designed for image denoising; SAM is Dynamic Self-Attention Mechanism (Dynamic-SAM).

[0033] S102. Based on the dynamic weight self-attention mechanism in the DnCNN network framework, the feature variables in the magnetotelluric signal sample pair are sequentially subjected to noise intensity perception, dynamic weight adjustment, and feature denoising under self-attention weighted output to obtain the denoising feature sequence generation strategy.

[0034] It should be noted that traditional SAM assigns weights solely based on feature similarity, failing to consider the dynamic changes in noise intensity. This leads to a bias in the attention distribution towards high-amplitude outliers in high-noise scenarios, while it tends towards averaging in low-noise scenarios, resulting in problems such as low-frequency principal component distortion and loss of detail. To address these shortcomings, this application proposes a Dynamic Self-Attention Mechanism (Dynamic-SAM). This mechanism introduces a noise intensity sensing and dynamic weight adjustment module into the standard self-attention framework, which can adaptively adjust the attention distribution according to the local energy of the signal, achieving differentiated modeling of features under different noise intensity conditions. This application embeds this module into the DnCNN network framework for high-fidelity noise reduction of magnetotelluric signals, achieving a significant performance improvement.

[0035] The Dynamic-SAM module mainly consists of three parts: a noise intensity sensing module (NEM); a dynamic weight adjustment module (DWM); and an enhanced self-attention mechanism module (Enhanced SAM Unit). Specifically, the intermediate features to be extracted by DnCNN are first evaluated for local noise intensity by the NEM module. Then, the DWM generates dynamic scaling coefficients based on the noise intensity and adjusts the attention sensitivity. Finally, the enhanced self-attention mechanism implements noise-aware weighted output, thereby enhancing the robustness of signal feature representation.

[0036] Specifically, it is also necessary to extract intermediate feature variables from the magnetotelluric signal sample pairs using the DnCNN network framework.

[0037] Furthermore, based on the intermediate feature variables, time series length, and feature dimension, the noise intensity of the input feature sequence is calculated under local energy perception to obtain the local noise energy.

[0038] In one embodiment, NEM aims to sense the noise strength in an input feature sequence through local energy calculation. Let the input features be... Where T represents the time series length and F represents the feature dimension, the local noise energy is defined as: .

[0039] Furthermore, by using the Sigmoid activation function and the bias term of the one-dimensional convolutional layer in the DnCNN network framework, the local interference degree of the dynamic characterization signal under different noise environments is calculated to obtain the noise intensity coefficient.

[0040] In one embodiment, local statistical features are extracted through a one-dimensional convolutional layer, and noise intensity coefficients are generated using a sigmoid activation function. ;in, ∈[0,1] represents a time series Noise intensity factor at b n The bias term is used to adjust the baseline level of the NEM output and determines the initial perception threshold of the network for noise intensity. This application sets an initial value b. n =0, during training b n As learnable parameters, they are automatically optimized via gradient descent. This design can dynamically characterize the degree of local interference in signals under different noise environments.

[0041] Furthermore, based on the dynamic adjustment amplitude term, the sensitivity of the attention response to the noise intensity coefficient is calculated to obtain the dynamic scaling weight.

[0042] In one embodiment, DWM can calculate a dynamic scaling weight based on the noise intensity coefficient α(t) to adjust the sensitivity of the attention response: Where β∈(0,1) is used to control the dynamic adjustment amplitude. When β is large (0.6–0.8), the model responds quickly to noise changes and is suitable for strong noise scenarios; when β is small (0.2–0.4), the weight changes are relatively smooth and are suitable for weak noise environments. This application selects β=0.5 as the initial value and sets it as a learnable parameter so that it can be automatically adjusted through backpropagation during training. When the noise intensity is high ( When it tends towards 1), The value tends towards 1, thus reducing the high-frequency interference response; when the noise is weak ( When it approaches 0, A value greater than 1 is used to enhance detailed features. This mechanism effectively balances the relationship between noise suppression and feature preservation.

[0043] Furthermore, it is necessary to dynamically scale the weights to perform noise dynamic changes tracking calculations on the query matrix, key matrix, and value matrix in the self-attention mechanism, and construct a noise-aware attention model for noise intensity perception.

[0044] In one embodiment, the calculation formula for the traditional self-attention mechanism is: Where Q=CWQ, K=CWK, and V=CWV. Dynamic-SAM introduces a dynamic scaling factor based on this. Construct a noise-aware attention model: By utilizing the self-attention mechanism described above, the attention distribution can be dynamically changed according to the noise level, thereby achieving adaptive information focusing and interference suppression.

[0045] Furthermore, through multi-layer convolution and normalization processing within the DnCNN network framework, intermediate feature sequences are extracted from the magnetotelluric signal sample pairs. Then, a noise-aware attention model is used to perform local energy calculations on the intermediate feature sequences, generating a noise intensity vector. Next, based on a noise level adaptive mechanism, the noise intensity vector undergoes dynamic weight adjustment to obtain dynamic scaling coefficients. These dynamic scaling coefficients are then fused with a self-attention mechanism, and based on the adjusted weight distribution, an optimized feature sequence under self-attention weighted output is obtained. Finally, based on the processing and generation model of the optimized feature sequence, a denoising feature sequence generation strategy for magnetotelluric signal denoising is derived.

[0046] In one embodiment, in the construction step of the noise reduction feature sequence generation strategy: First, a DnCNN network with multiple convolutional and batch normalization operations extracts an intermediate feature sequence C; then, a noise intensity sensing module (NEM) performs local energy calculation on the input feature sequence C and generates a noise intensity vector. Then the Dynamic Weighting (DWM) module receives the noise intensity vector. Adaptively generate dynamic scaling factors based on noise level Finally, in self-attention computation, we introduce... Adjusting the weight distribution allows for the output of the dynamically weighted optimized feature sequence Cout, which can then be used by subsequent convolution or reconstruction modules.

[0047] As a possible implementation method, according to This yields the comprehensive loss function used for joint optimization of the dynamic adjustment amplitude term and the bias term. .in, This is the mean square error term, used to measure the noise reduction accuracy; This is the envelope entropy term, used to constrain signal smoothness; This serves as a balance coefficient. Then, through a comprehensive loss function, the noise-aware attention model undergoes real-time attention distribution update adjustment processing related to signal-to-noise ratio, energy distribution, and spectral smoothness, generating a joint optimization strategy for the noise-aware attention model. That is, during the network training phase, β and b in the Dynamic-SAM module... n A joint adaptive optimization strategy is adopted, dynamically adjusting its values ​​through a gradient descent algorithm to achieve an optimal balance between signal-to-noise ratio, energy distribution, and spectral smoothness. Based on the joint optimization strategy corresponding to the objective function of the joint optimization process, the Dynamic-SAM module can update the attention distribution, β, and b in real time according to noise characteristics. n As a learnable parameter, it participates in gradient descent updates and is automatically adjusted through backpropagation, thereby achieving an optimal balance between signal-to-noise ratio, spectral fidelity, and energy distribution.

[0048] S103. Using a pre-configured DnCNN-Dynamic-SAM network, the original signal time series from the magnetotelluric signal sample pairs are processed using a two-level denoising feature sequence generation strategy based on convolutional layers. The denoising model is then trained based on the feature denoising model of the original signal time series to generate and obtain the trained DnCNN-Dynamic-SAM network model.

[0049] It should be noted that DnCNN networks possess strong feature extraction capabilities and have been widely applied in fields such as image denoising. Self-attention mechanisms (SAM) can model and capture dependencies between positions in a sequence globally, thereby capturing long-distance correlations and improving feature representation capabilities. Based on this, this application proposes a novel deep learning model, DnCNN-SAM, which combines the deep feature extraction advantages of DnCNN with the global feature modeling capabilities of self-attention mechanisms. This model can not only fully mine the deep features in the MT signal but also effectively capture the global dependencies of the signal in the time and frequency dimensions, thereby improving the robustness and accuracy of noise suppression. During denoising, the DnCNN-SAM network can accurately extract the effective component features of the MT signal, maintain the overall structural integrity of the signal, and achieve efficient signal-to-noise separation.

[0050] The DnCNN-Dynamic-SAM network of this application uses the traditional DnCNN as the main framework and embeds two-level Dynamic-SAM modules between its feature extraction and reconstruction to realize noise perception and dynamic weighted modeling of features, and finally generate and obtain the trained DnCNN-Dynamic-SAM network model.

[0051] Specifically, the original signal time series in the magnetotelluric signal sample pairs are first obtained through a pre-configured DnCNN-Dynamic-SAM network. The DnCNN-Dynamic-SAM network is a fusion neural network with deep convolution, dynamic weights, and a self-attention mechanism.

[0052] Furthermore, the basic temporal features of the preprocessed original signal time series are extracted through the first convolutional layer in the DnCNN-Dynamic-SAM network; and the basic temporal features after initial convolution are input into the 16-layer convolutional block.

[0053] Furthermore, the first intermediate feature sequence output by the 16-layer convolutional block is subjected to noise perception and self-attention weighting processing through the Dynamic-SAM module corresponding to the noise reduction feature sequence generation strategy to obtain the first optimized feature sequence.

[0054] Furthermore, through the second convolutional layer, the first optimized feature sequence is subjected to feature mapping and channel mapping to obtain the second intermediate feature sequence.

[0055] Furthermore, it is also necessary to use the Dynamic-SAM module corresponding to the noise reduction feature sequence generation strategy, that is, to use the Dynamic-SAM module twice. After the two-level noise reduction feature sequence generation strategy is calculated, the second intermediate feature sequence is dynamically weighted and self-attention weighted to obtain the second optimized feature sequence.

[0056] Furthermore, by inputting the second optimized feature sequence into the third convolutional layer of the DnCNN-Dynamic-SAM network, a feature denoising model training and processing strategy for magnetotelluric signals can be obtained.

[0057] In one embodiment, the computational processing of the DnCNN-Dynamic-SAM network is as follows: 1) First, the original MT time series is acquired and preprocessed; 2) The preprocessed MT signal is processed by the first convolutional layer (Conv1) to extract basic temporal features; 3) The features after the initial convolution are input into the 16th convolutional block. Each CBR module uses nonlinear activation to enhance the network's ability to represent complex noise and signal patterns, while batch normalization ensures the stability of training and the consistency of feature distribution; 4) After the output of the 16th CBR block, the first Dynamic-SAM module is introduced to process the intermediate feature sequence. Noise perception and self-attention weighting are performed; 5) Feature sequences optimized by Dynamic-SAM1 in the first stage. The features are integrated and channel mapped after the second convolutional layer (Conv2); 6) The feature sequence output by Conv2 is input into the second Dynamic-SAM module to perform the same dynamic weight calculation and attention optimization process as in the first stage; 7) Finally, the feature sequence after weighted optimization by the two-stage Dynamic-SAM module is input into the third convolutional layer (Conv3) to map the high-dimensional features back to the time domain signal space; finally, the feature denoising model training and processing strategy for magnetotelluric signals is obtained.

[0058] Furthermore, based on the feature denoising model training strategy of the original signal time series, the feature denoising model is iteratively trained on several pairs of magnetotelluric signal samples.

[0059] Furthermore, based on the feature denoising algorithm trained by model iteration, deep features in the magnetotelluric signal are mined and processed, and the global dependencies of the signal in the time and frequency dimensions are captured. Finally, the trained DnCNN-Dynamic-SAM network model is generated and obtained.

[0060] As a feasible implementation, the entire DnCNN-Dynamic-SAM network achieves a trained DnCNN-Dynamic-SAM network model through a three-level collaborative structure of "convolutional feature extraction—noise-aware attention—dynamic weighted reconstruction," and also realizes an adaptive balance between signal denoising and feature preservation. The two Dynamic-SAM modules respectively undertake the roles of local noise suppression and global feature smoothing, enabling the network to exhibit stronger robustness and generalization ability in complex interference environments.

[0061] S104. Based on the DnCNN-Dynamic-SAM network model, the current original magnetotelluric signal is subjected to noise reduction control and feature extraction to obtain effective signal features after local noise suppression and global feature smoothing.

[0062] Specifically, firstly, the DnCNN-Dynamic-SAM network model is used to perform local noise suppression calculations on the current raw magnetotelluric signal under a denoising feature sequence generation strategy. Then, the DnCNN-Dynamic-SAM network model is used to perform global feature smoothing control on the denoised features after local noise suppression calculations under a secondary denoising feature sequence generation strategy.

[0063] Furthermore, based on the model denoising calculation results under local noise suppression and global feature smoothing, the effective signal features in the current original magnetotelluric signal are output.

[0064] In one embodiment, based on denoising analysis using measured data:

[0065] (1) In the experimental analysis in the time domain:

[0066] For field data processing in a certain mining area: Data acquisition was conducted using a V5-2000 MT sounder, with sampling rates of 2400Hz, 150Hz, and 15Hz. Single measurements lasted 18 to 24 hours, aiming to study the geological structure within a 10km depth range. However, the area suffers from widespread electromagnetic interference from mining activities, high-voltage power lines, and vehicles, resulting in a very low signal-to-noise ratio (SNR) for the MT data. High-amplitude or prolonged interference signals cause significant distortion in the MT response in the low- to mid-frequency range, severely affecting the accurate inversion of deep electrical structures. To improve data quality, low SNR signals subjected to continuous noise interference were selected, and the selected signals were processed using the algorithm included in this application.

[0067] Regarding the time-domain denoising effect on actual measurement data segments, Figure 2 This is a schematic diagram illustrating the temporal denoising effect of a measurement data segment provided in an embodiment of this application; as shown... Figure 2 As shown in the diagram, (a) represents a time series segment affected by impulse noise, (b) by triangular wave noise, and (c) by mixed noise. These segments are contaminated by impulse noise, triangular wave noise, and mixed noise, respectively, resulting in interference and a significant decrease in the signal-to-noise ratio. After denoising with DNCNN-SAM, the method demonstrated good noise resistance. Figure 2As shown in (d), (e), and (f), which are time series segments of impulse noise after noise reduction, respectively, and triangular wave noise after noise reduction, the main trend of the signal remains unchanged, indicating that the low-frequency dominant component is not destroyed. This application can effectively remove complex noise such as strong impulses while accurately preserving the dominant signal.

[0068] Among them, measuring point E1-141 was significantly affected by mining noise, resulting in continuous abnormal waveforms in the MT time series. Figure 3 This is a schematic diagram comparing the data segments before and after denoising at measurement point E1-141 provided in this application embodiment; the time series segment recorded by this measurement point is as follows: Figure 3 Figure (a) shows the noise interference waveform in the Ex channel, (b) shows the noise interference waveform in the Ey channel, (c) shows the complex noise waveform with mixed pulse signals in the Hx magnetic channel, and (d) shows the complex noise waveform with mixed pulse signals in the Hy magnetic channel. Similar and continuous noise interference can be observed in the Ex and Ey channels, while complex noise with mixed pulse signals exists in the Hx and Hy magnetic channels. The amplitudes of these interference signals far exceed the effective MT signal, resulting in a low signal-to-noise ratio for the MT data at this measurement point. DNCNN-SAM denoising was performed on the data from this measurement point, and a partial denoising result is shown below. Figure 3 As shown in (e), (f), (g), (h), and (h) of the graphs, the denoised waveforms in the Ex, Ey, Hx, and Hy magnetic channels are respectively, it can be seen that the denoised low-frequency information basically matches the time series trend. The effective slow low-frequency changes are preserved after denoising, without significant loss of low-frequency information due to processing.

[0069] (2) In the measured analysis of the resistivity-phase curve:

[0070] The resistivity component curves in the 0.6–0.07 Hz frequency band are severely downward shifted, and the resistivity component curves in the 0.5–0.06 Hz frequency band exhibit violent, sawtooth-like fluctuations. The power spectrum apparent resistivity curve is generally discontinuous and not smooth, and abrupt changes are visible in its phase curve. Furthermore, waveform observation in the time domain indicates that the original signal contains a large amount of interference resembling triangular waves and pulses. Compared to the original curves, the apparent resistivity-phase curves obtained by the DWT algorithm show improved convergence and divergence, but due to the presence of pulse signals, the phase curves still exhibit some flying points. After processing with the CNN method, the resistivity-frequency and phase-frequency curves obtained by the CNN method show significant improvement in the high-frequency range, while severe jumps occur at single frequency points below 0.5 Hz in the low-frequency range. After systematic processing using the method proposed in this application, the apparent resistivity curves are significantly improved, the data distribution is more stable, and the phase curves tend to be continuous and smooth, resulting in a more accurate revelation of the electrical structural characteristics of the subsurface medium.

[0071] In addition, embodiments of this application also provide a data noise reduction device for magnetotelluric signals, such as... Figure 4 As shown, the data noise reduction device 400 for magnetotelluric signals specifically includes:

[0072] At least one processor 401; and a memory 402 communicatively connected to the at least one processor 401; wherein the memory 402 stores instructions executable by the at least one processor 401 to enable the at least one processor 401 to execute:

[0073] The effective magnetotelluric signal is superimposed with the simulated synthetic noise waveform, and magnetotelluric signal sample pairs are generated based on the original magnetotelluric signal.

[0074] Based on the dynamic weight self-attention mechanism in the DnCNN network framework, the feature variables in the magnetotelluric signal sample pair are sequentially subjected to noise intensity perception, dynamic weight adjustment and self-attention weighted output feature denoising processing to obtain the denoising feature sequence generation strategy.

[0075] Using a pre-configured DnCNN-Dynamic-SAM network, a two-stage denoising feature sequence generation strategy based on convolutional layers is calculated for the original signal time series from magnetotelluric signal sample pairs. The processing strategy is then trained based on the feature denoising model of the original signal time series, generating and obtaining the trained DnCNN-Dynamic-SAM network model.

[0076] Based on the DnCNN-Dynamic-SAM network model, the current raw magnetotelluric signal is subjected to noise reduction control and feature extraction to obtain effective signal features after local noise suppression and global feature smoothing.

[0077] This application's embodiments utilize a pre-configured DnCNN-Dynamic-SAM network to accurately extract low-frequency components from noisy time series, effectively avoiding the loss of valid MT signals. Compared to traditional methods, this approach relies on a deep learning model, reducing manual intervention and achieving higher processing efficiency, making it particularly suitable for batch processing of large-scale MT data. Experimental results on synthetic and measured data demonstrate that this method exhibits excellent denoising performance when handling various noise types, including pulses, square waves, and triangular waves. Applying this method not only significantly improves the stability and reliability of apparent resistivity and phase curves but also outperforms traditional robust estimation methods in overall performance. Furthermore, it enables accurate extraction of low-frequency dominant components, effectively solving the problem of efficient and high-precision denoising of low-frequency MT signals, thereby improving the inversion effect in noise-interference areas.

[0078] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0079] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0080] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0085] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0086] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0087] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0088] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of this specification.

Claims

1. A data noise reduction method for magnetotelluric signals, characterized in that, The method includes: The effective magnetotelluric signal is superimposed with the simulated synthetic noise waveform, and magnetotelluric signal sample pairs are generated based on the original magnetotelluric signal. Intermediate feature variables were extracted from the magnetotelluric signal sample pairs using the DnCNN network framework. Based on the intermediate feature variables, time series length, and feature dimension, the noise intensity of the input feature sequence is calculated under local energy perception to obtain the local noise energy. By using the Sigmoid activation function and the bias term of the one-dimensional convolutional layer in the DnCNN network framework, the local noise energy is used to calculate the local interference degree of the dynamic characterization signal under different noise environments, and the noise intensity coefficient is obtained. Based on the dynamic adjustment amplitude term, the sensitivity of the attention response to the noise intensity coefficient is calculated to obtain the dynamic scaling weight; By using the dynamic scaling weights, the query matrix, key matrix, and value matrix in the self-attention mechanism are subjected to noise dynamic change tracking calculations related to the dynamic scaling factors, and a noise-aware attention model for noise intensity perception is constructed. Based on the dynamic weight self-attention mechanism in the DnCNN network framework, the feature variables in the magnetotelluric signal sample pairs are sequentially subjected to noise intensity sensing, dynamic weight adjustment, and feature denoising under self-attention weighted output to obtain a denoising feature sequence generation strategy, which specifically includes: The intermediate feature sequences in the magnetotelluric signal sample pairs are extracted through multi-layer convolution and normalization processing in the DnCNN network framework. The noise-aware attention model is used to perform local energy calculation on the intermediate feature sequence to generate a noise intensity vector. According to the noise level adaptive mechanism, the noise intensity vector is subjected to weight adaptive adjustment calculation under the dynamic adjustment of the weight to obtain the dynamic scaling coefficient; The dynamic scaling factor is fused with the self-attention mechanism for calculation, and the optimized feature sequence under the self-attention weighted output is obtained based on the adjusted weight distribution. Based on the optimized feature sequence processing and generation model, the noise reduction feature sequence generation strategy for noise reduction processing of magnetotelluric signals is obtained; Using a pre-configured DnCNN-Dynamic-SAM network, the original signal time series from the magnetotelluric signal sample pairs are processed using a two-level noise reduction feature sequence generation strategy based on convolutional layers, specifically including: The original signal time series in the magnetotelluric signal sample pair is obtained by using a pre-configured DnCNN-Dynamic-SAM network; wherein, the DnCNN-Dynamic-SAM network is a fusion neural network with deep convolution, dynamic weights and self-attention mechanism. The basic temporal features of the preprocessed original signal time series are extracted through the first convolutional layer of the DnCNN-Dynamic-SAM network; and the basic temporal features after initial convolution are input into the 16-layer convolutional block. The first optimized feature sequence is obtained by using the Dynamic-SAM module corresponding to the noise reduction feature sequence generation strategy to perform noise perception and self-attention weighting on the first intermediate feature sequence output by the 16-layer convolutional block. The first optimized feature sequence is subjected to feature mapping and channel mapping through the second convolutional layer to obtain the second intermediate feature sequence; The second optimized feature sequence is obtained by using the Dynamic-SAM module corresponding to the noise reduction feature sequence generation strategy to perform dynamic weight calculation and self-attention weighted optimization on the second intermediate feature sequence. The second optimized feature sequence is input into the third convolutional layer of the DnCNN-Dynamic-SAM network to obtain a feature denoising model training and processing strategy for magnetotelluric signals. Based on the feature denoising model training strategy of the original signal time series, the trained DnCNN-Dynamic-SAM network model is generated and obtained; Based on the DnCNN-Dynamic-SAM network model, the current raw magnetotelluric signal is subjected to noise reduction control and feature extraction to obtain effective signal features after local noise suppression and global feature smoothing.

2. The data noise reduction method for magnetotelluric signals according to claim 1, characterized in that, The effective magnetotelluric signal is superimposed with the simulated synthetic noise waveform, and magnetotelluric signal sample pairs are generated based on the original magnetotelluric signal, specifically including: Based on the original magnetotelluric signal, the effective magnetotelluric signal in the noise-free area is extracted and identified as the effective signal sample; Various types of noise waveforms are simulated, and the simulated synthetic noise waveforms are obtained based on the noise characteristics corresponding to the noise pollution signals. The synthesized noise waveform is superimposed onto the effective signal sample to generate a noise sample; The noise samples are paired with the valid signal samples to obtain several pairs of magnetotelluric signal samples.

3. The data noise reduction method for magnetotelluric signals according to claim 1, characterized in that, Before constructing the noise-aware attention model for noise intensity perception, the method further includes: according to This yields a comprehensive loss function used for joint optimization of the dynamic adjustment amplitude term and the bias term. ;in, This is the mean square error term, used to measure the noise reduction accuracy; This is the envelope entropy term, used to constrain signal smoothness; This is the balance coefficient; The noise-aware attention model is subjected to real-time attention distribution update and adjustment processing related to signal-to-noise ratio, energy distribution and spectral smoothness through the comprehensive loss function, and a joint optimization strategy for the noise-aware attention model is generated.

4. The data noise reduction method for magnetotelluric signals according to claim 1, characterized in that, Based on the feature-based denoising model training strategy of the original signal time series, a trained DnCNN-Dynamic-SAM network model is generated and obtained, which specifically includes: Based on the feature denoising model training and processing strategy of the original signal time series, the feature denoising model is iteratively trained on several pairs of magnetotelluric signal samples. Based on the feature denoising algorithm trained by model iteration, deep features in magnetotelluric signals are mined and processed, and the global dependencies of the signals in the time and frequency dimensions are captured to generate and obtain the trained DnCNN-Dynamic-SAM network model.

5. A data noise reduction method for magnetotelluric signals according to claim 1, characterized in that, Based on the DnCNN-Dynamic-SAM network model, the current raw magnetotelluric signal is subjected to noise reduction control and feature extraction to obtain effective signal features after local noise suppression and global feature smoothing, specifically including: Using the DnCNN-Dynamic-SAM network model, local noise suppression calculations are performed on the current original magnetotelluric signal under the relevant noise reduction feature sequence generation strategy. Using the DnCNN-Dynamic-SAM network model, the denoised features after local noise suppression calculation are subjected to global feature smoothing control under a secondary denoising feature sequence generation strategy. Based on the model denoising calculation results under the local noise suppression and global feature smoothing processing, the effective signal features in the current original magnetotelluric signal are output.

6. A data noise reduction device for magnetotelluric signals, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform a data noise reduction method for magnetotelluric signals according to any one of claims 1-5.

7. A non-volatile computer storage medium, characterized in that, The storage medium is a non-volatile computer-readable storage medium that stores at least one program, each program including instructions that, when executed by a terminal, cause the terminal to perform a data noise reduction method for magnetotelluric signals according to any one of claims 1-5.

Images