A magnetotelluric signal processing method based on adaptive noise modulation

Abstract

This invention discloses a magnetotelluric signal processing method based on adaptive noise modulation. It acquires multi-band base signal (MT) time series data and segments it into base signal data segments. Each base signal data segment is analyzed, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives. The noise primitive data is preprocessed to generate composite noise profiles, which are then superimposed onto the base signal data segments to obtain noisy data segments. Using the noisy data segments as samples and the composite noise profiles as labels, a denoising dataset closely resembling real-world scenarios is constructed, providing high-quality data support for the subsequent training of the denoising model. A TransUnet model is constructed, and the denoised dataset is input into the model for training to obtain a noise profile extraction model. Finally, measured magnetotelluric signals are input into the noise profile extraction model to obtain predicted noise profiles. This method is suitable for high-precision MT signal denoising at different sampling rates, effectively suppressing strong non-stationary interference and improving signal quality and the reliability of geological interpretation.

Description

Technical Field

[0001] This invention belongs to the field of magnetotelluric signal processing technology, and particularly relates to a magnetotelluric signal processing method based on adaptive noise modulation. Background Technology

[0002] Magnetotellurics (MT) is a passive source exploration technique that relies on changes in natural electromagnetic fields to study underground electrical structures. It has wide applications in mineral resource exploration, geothermal energy development, and crustal structure analysis. However, in actual observations, MT signals are often affected by complex electromagnetic interference, especially in mining areas, industrialized regions, or electrified environments. Strong non-stationary noise, such as periodic rectangular waves, transient waves, and various forms of power frequency superposition interference, is frequently mixed into the observed signals. This noise not only destroys the time and frequency domain characteristics of the signal but also severely affects the inversion accuracy of key parameters such as apparent resistivity and impedance phase, thereby reducing the reliability of geological interpretation results.

[0003] Existing signal denoising methods mainly include robust estimation, distant reference methods, morphological filtering, wavelet transform, and deep learning based on convolutional neural networks, but they still have significant limitations in practical applications. For example, robust estimation and wavelet transform lack adaptability when noise characteristics are unknown or dynamically changing; distant reference methods struggle to select suitable reference points in high-noise environments; time-domain methods such as morphological filtering heavily rely on human experience and are prone to accidentally omitting valid signals; while deep learning methods based on convolutional neural networks (CNNs), such as U-Net, ResNet, and stacked autoencoders, although capable of automatically extracting local spatiotemporal features of signals, have limited receptive fields and struggle to effectively model long-range dependencies between distant points in a time series. For common rectangular waves or periodic noise with large spans in magnetotelluric signals, simple CNN models may not be able to fully capture their overall structure, resulting in distortion or residue at the noise start and end boundaries after denoising. Especially in the low-frequency band, the identification and suppression of non-stationary strong noise is not ideal.

[0004] In summary, existing denoising techniques still have shortcomings in suppressing non-stationary and strong interference noise, making it difficult to guarantee the purity of MT signals across the entire frequency band and the reliability of inversion parameters. Therefore, how to enable deep learning models to simultaneously possess excellent local feature extraction capabilities and global context modeling capabilities is a key challenge in improving the denoising accuracy of magnetotelluric signals, especially low-frequency signals. There is an urgent need to provide a magnetotelluric signal processing method based on adaptive noise modulation to solve the aforementioned technical problems. Summary of the Invention

[0005] In view of this, the present invention provides an adaptive noise modulation magnetotelluric signal processing method. By constructing an adaptive noise generation mechanism that matches the characteristics of real magnetotelluric signals, an adaptive noise sequence that highly matches the characteristics of the original magnetotelluric signal can be generated. After superimposing the noise sequence onto the original signal, a denoised dataset that closely resembles the real scene can be constructed, providing high-quality data support for the training of subsequent magnetotelluric signal denoising models. The specific technical solution adopted is as follows.

[0006] This invention provides a magnetotelluric signal processing method based on adaptive noise modulation, comprising the following steps: The system acquires a preset standard geoelectric structure model to generate multi-band base signal MT time series data, and uses a preset fixed sliding window to segment the base signal MT time series data of all frequency bands to obtain base signal data segments. Pre-analysis is performed on each base signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the base signal data segment. At least three noise mathematical models corresponding to the waveforms are constructed based on the noise characteristics of the measured MT data. The noise primitives are preprocessed to generate a composite noise profile, and the composite noise profile is superimposed on the base signal data segment to obtain a noisy data segment. A denoised dataset is constructed using the noisy data segment as a sample and the composite noise profile as a label. A TransUnet model with U-Net as the backbone is constructed, and the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model. The TransUnet model includes a feature extraction path, a signal reconstruction path, and a global context modeling module. The measured magnetotelluric signal is input into the noise profile extraction model to obtain the predicted noise profile, and the target magnetotelluric signal is obtained based on the measured magnetotelluric signal and the predicted noise profile.

[0007] As a preferred embodiment of the above technical solution, multi-band base signal MT time series data are generated by obtaining a preset standard geoelectric structure model, and the base signal MT time series data of all frequency bands are segmented using a preset fixed sliding window to obtain base signal data segments, including: Acquire synthetic MT time series data of the measurement point in three frequency bands: low frequency, medium frequency, and high frequency, wherein the low frequency, medium frequency, and high frequency are 15Hz, 150Hz, and 2400Hz, respectively; Set the window length for multiple sampling points in the synthesized MT time series data, and perform equal-interval sliding extraction on the channel data of each measurement point to obtain the segmented data; The segmented data is Z-score normalized to obtain multiple base signal data segments.

[0008] As a preferred embodiment of the above technical solution, at least three noise mathematical models corresponding to the measured MT data are constructed, including: The typical noise characteristics in the measured MT time series data are analyzed, and the mathematical models of transient waves, rectangular waves, and triangular waves are expressed as follows: (1) (2) (3) in, A mathematical model of noise representing a transient wave. A mathematical model of noise representing a rectangular wave. The mathematical model of noise represents a triangular wave; t represents the time variable or sampling point index, used to determine the amplitude of the noise at each time point; K represents the number of independent noise events superimposed in the time series; This represents the peak value or amplitude of the k-th noise. Indicates the left half width of the transient wave. Indicates the right half width of the transient wave. Indicates the transient center position. Indicates the center position of the triangular wave. This indicates the space affected by noise disturbance.

[0009] As a preferred embodiment of the above technical solution, a pre-analysis is performed on each base signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the base signal data segment, including: The local energy envelope and dominant frequency of each base signal data segment are calculated using short-time Fourier transform. The local energy envelope and the dominant frequency are respectively smoothed by moving average to obtain a smoothed energy sequence and a smoothed dominant frequency sequence. The smoothed energy sequence is normalized to obtain a normalized energy sequence, and the peak-to-peak value is obtained by calculating the difference between the 95th percentile and the 5th percentile of the base signal data segment. Based on the normalized energy sequence, the peak-to-peak value, and the smoothed main frequency sequence, adaptive noise amplitude parameters, noise width parameters, and noise occurrence rate parameters are calculated respectively, and upper and lower limits are applied to each parameter. The parameters after upper and lower limit amplitude limiting are assigned to the model parameters of the noise mathematical model to generate noise primitives; The calculation expressions for the noise amplitude parameter, the noise width parameter, and the noise occurrence rate parameter are as follows: (4) (5) (6) Where i represents the number of base signal data segments. Indicates the noise amplitude parameter. Indicates peak-to-peak value. Represents the normalized energy sequence. This represents a standard normally distributed random sequence after smoothing and normalization using a moving average. Indicates the noise width parameter. Indicates the noise occurrence rate parameter; for implement Limiting operation, for implement Limiting operation.

[0010] As a preferred embodiment of the above technical solution, the noise primitives are preprocessed to generate a composite noise profile, and the composite noise profile is superimposed on the base signal data segment to obtain a noisy data segment. A denoised dataset is constructed using the noisy data segment as samples and the composite noise profile as labels, including: The noise primitives generated by formulas (1), (2) and (3) are combined, transformed and nested to generate the composite noise profile, wherein the data preprocessing includes the combination, transformation and nesting of the noise primitives; The composite noise profile is injected into multiple base signal data segments to generate noisy data segments; The generation of the composite noise profile through the combination, transformation, and nesting of the noise primitives includes: Based on the local energy envelope, a smooth energy sub-envelope env1 is generated, and a smooth random sub-envelope env2 is generated based on a standard normal distribution random sequence. The two sub-envelopes are then weighted and fused to obtain a composite modulation envelope. The noise primitives corresponding to transient waves, rectangular waves, and triangular waves are combined in a random ratio, and the combined noise primitives are then subjected to amplitude scaling and time axis offset transformation operations. The noise primitives transformed at different scales are nested and superimposed, and the nested and superimposed noise primitives are multiplied with the composite modulation envelope to obtain the composite noise profile. Before superimposing the composite noise profile onto the base signal data segment, the following steps are also included: Energy control is applied to the composite noise profile to calculate the energy of the base signal data segment. Initial energy of the composite noise profile According to the ratio of target noise energy to signal energy Calculate the scaling factor Multiply the composite noise profile by the scaling factor. And superimposed on the base signal data segment.

[0011] As a preferred embodiment of the above technical solution, a TransUnet model with U-Net as the backbone is constructed, and the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model, including: A global context modeling module composed of TransformerEncoders is embedded in the bottleneck layer where the feature extraction path and signal reconstruction path of U-Net intersect. The global context modeling module uses the self-attention mechanism of TransformerEncoder to explicitly model the long-range dependency between each time sampling point in the noisy data segment. The TransUnet model extracts multi-scale local features through the aforementioned feature extraction path, which includes several one-dimensional convolutional layers. Each convolutional layer is followed by a batch normalization function and a ReLU activation function. The mathematical expression for the output of the l-th layer is: (7) in, Indicates the first Layer output features, , These represent the convolution kernel and bias parameters, respectively. Indicates the first Layer output features; This indicates batch normalization, used to normalize the output of convolutions. This represents a one-dimensional convolution operation used to extract local features from time series data. This represents the modified linear unit activation function; The TransUnet model's signal reconstruction path gradually restores the time series resolution corresponding to the time series through upsampling or transposed convolution, and concatenates the high-resolution features of the corresponding layers in the feature extraction path through skip connections to achieve fine reconstruction of the noise contour. The corresponding mathematical expression is: (8) in, Indicates the first Layer recovery features, Indicates the first Layer recovery features, This indicates a feature concatenation operation. This indicates an upsampling operation.

[0012] As a preferred embodiment of the above technical solution, the execution process of the global context modeling module includes: The global context modeling module is used to receive the multi-channel feature maps output by the feature extraction path, and flatten and rearrange the feature maps along the time dimension to obtain a two-dimensional feature sequence. Introducing learnable positional encoding into the two-dimensional feature sequence To characterize the location information of each sampling point in the time series. The feature sequence after positional encoding is then input into at least one Transformer Encoder to obtain the self-attention output feature. The Transformer Encoder includes a multi-head self-attention layer and a feedforward neural network layer. The corresponding mathematical expression for calculating the attention output feature is: (9) in, This represents the attention output features. , , These represent the query vector, key vector, and value vector obtained by linear mapping of the two-dimensional feature sequence, respectively. The dimension of the key vector is represented by T, which represents the transpose operation. According to formula (9), the global correlation between the features corresponding to any two time sampling points in the two-dimensional feature sequence is modeled to obtain the global features containing long-range dependencies in the time dimension.

[0013] As a preferred embodiment of the above technical solution, the global features are fused with the multi-scale local features of the feature extraction path to obtain fused features, and the corresponding mathematical expression is: (10) in, Indicates the first Layer fusion features This indicates a feature concatenation operation. Indicates the first Layer output features, Represents global features; The fused feature is input to the signal reconstruction path and the time series resolution corresponding to the time series is gradually restored through multi-level upsampling operations. Based on the time series resolution, a noise profile with the same data segment length as the fused feature is output at the end of the signal reconstruction path.

[0014] As a preferred embodiment of the above technical solution, the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model, including: Using composite noise contours as labels, and employing a weighted mean square error (MSE) that distinguishes between noise foreground and signal background as the loss function, the TransUnet model is trained under supervision to obtain a noise contour extraction model. The mathematical expression for the loss function is as follows: (11) (12) in, Let y[t] represent the loss function, and y[t] represent the true value of the sample. This represents the network's predicted output value; Indicates the sample point index; This represents the foreground mask, taking the position of the foreground with a true value that is not zero; This indicates that the background mask is taken from a position outside the foreground. The weights of the background penalty term are used; a stochastic gradient descent optimizer is used to train the low-frequency, mid-frequency, and high-frequency bands independently. The initial learning rate for independent training is set to 1e-3, and multiple rounds of training are performed.

[0015] As a preferred embodiment of the above technical solution, the measured magnetotelluric signal is input into the noise profile extraction model to obtain a predicted noise profile, and the target magnetotelluric signal is obtained based on the measured magnetotelluric signal and the predicted noise profile, including: The measured time series data of the collected magnetotelluric signals corresponding to the measured MT signals were subjected to the same data preprocessing as the TransUnet model. The preprocessed MT measured time series data is input into the noise profile extraction model and the predicted noise profile is output. The measured magnetotelluric signal is subtracted from the predicted noise profile to obtain the denoised clean magnetotelluric signal, wherein the target magnetotelluric signal is the denoised clean magnetotelluric signal.

[0016] This invention provides a magnetotelluric signal processing method based on adaptive noise modulation. It generates multi-band base signal (MT) time series data by acquiring a preset standard geoelectric structure model, and segments the MT time series data of all frequency bands using a preset fixed sliding window to obtain base signal data segments. Pre-analysis is performed on each base signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the base signal data segments. Data preprocessing is performed on the noise primitives to generate composite noise profiles, which are then superimposed onto the base signal data segments to obtain noisy data segments. A denoised dataset is constructed using the noisy data segments as samples and the composite noise profiles as labels. This invention constructs a TransUnet model with U-Net as the backbone and trains it using a denoised dataset to obtain a noise contour extraction model. Measured magnetotelluric signals are input into the noise contour extraction model to obtain predicted noise contours. Based on the measured magnetotelluric signals and the predicted noise contours, the target magnetotelluric signal is obtained. An adaptive noise generation mechanism matching the features of the real magnetotelluric signal is constructed, generating an adaptive noise sequence highly matching the features of the original magnetotelluric signal. This noise sequence is superimposed on the original signal to construct a denoised dataset closely resembling the real scene, providing high-quality data support for the training of subsequent magnetotelluric signal denoising models. This invention is suitable for high-precision MT signal denoising of signals with different sampling rates, effectively suppressing non-stationary strong interference, improving signal quality and the reliability of geological interpretation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of the magnetotelluric signal processing method based on adaptive noise modulation provided by the present invention; Figure 2 A schematic diagram of the synthesized time series data of measurement point B in the low-frequency, mid-frequency, and high-frequency bands provided by the present invention; Figure 3 This is a schematic diagram of the network structure of the TransUnet model provided by the present invention; Figure 4 A diagram illustrating the adaptive noise modulation process provided by this invention; Figure 5 A flowchart of composite noise contour processing provided by the present invention; Figure 6 Waveform diagrams for the fundamental signal analysis provided by this invention; Figure 7 The waveform diagram generated by the noise primitive provided by this invention; Figure 8 A waveform diagram constructed for the composite noise profile provided by the present invention; Figure 9 Waveform diagram generated for the noisy sample provided by this invention; Figure 10 The time series waveform diagram of measurement point L010S020 provided by this invention; Figure 11 A time series comparison of the denoising effects of different models on the Ey channel of the L010S020 measurement point provided by this invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] See Figure 1 and Figure 4 This invention provides a magnetotelluric signal processing method based on adaptive noise modulation, comprising the following steps: S1: Obtain the multi-band base signal MT time series data from the preset standard geoelectric structure model, and use a preset fixed sliding window to segment the base signal MT time series data of all frequency bands to obtain base signal data segments; S2: Perform pre-analysis on each base signal data segment and adaptively modulate the model parameters of the noise mathematical model to generate noise primitives corresponding to the base signal data segment. Among them, at least three noise mathematical models corresponding to waveforms are constructed based on the noise characteristics of the measured MT data. S3: Perform data preprocessing on the noise primitives to generate a composite noise profile, and superimpose the composite noise profile onto the base signal data segment to obtain a noisy data segment. Construct a denoised dataset using the noisy data segment as a sample and the composite noise profile as a label. S4: Construct a TransUnet model with U-Net as the backbone, and input the denoised dataset into the TransUnet model for training to obtain a noise contour extraction model. The TransUnet model includes a feature extraction path, a signal reconstruction path, and a global context modeling module. S5: Input the measured magnetotelluric signal into the noise profile extraction model to obtain the predicted noise profile, and obtain the target magnetotelluric signal based on the measured magnetotelluric signal and the predicted noise profile.

[0021] In this embodiment, multi-band base signal time series data are generated from a preset standard geoelectric structure model. A preset fixed sliding window is used to segment the base signal time series data across all frequency bands to obtain base signal data segments. This includes: acquiring synthetic MT time series data from the measurement point in three frequency bands: low frequency, mid frequency, and high frequency, where the low frequency, mid frequency, and high frequency are 15Hz, 150Hz, and 2400Hz, respectively; setting the window length for multiple sampling points in the synthetic MT time series data; and performing equidistant sliding extraction on the data from each channel of the measurement point to obtain segmented data; and performing Z-score normalization on the segmented data to obtain multiple base signal data segments. Taking measurement point B as the data extraction point, synthetic MT time series data in three frequency bands (low frequency 15Hz, mid frequency 150Hz, and high frequency 2400Hz) are acquired, representing the main operating frequency ranges in magnetotelluric signal processing. Figure 2 The synthesized time series waveforms of measurement point B in the low-frequency, mid-frequency, and high-frequency bands are shown, with sampling rates of 15Hz, 150Hz, and 2400Hz from top to bottom.

[0022] Specifically, Figure 2 The base signal magnetotelluric data is obtained through the COMMEMI3D-2A model. Magnetotelluric data itself is irregular, and noise typically manifests as spikes, rectangular waves, or other regular waveforms superimposed on the base signal data. The base signal data segments are obtained by segmenting the base signal data using a sliding window. To reduce dimensional differences between different channels and improve the numerical stability of model training, the segmented data undergoes Z-score standardization, which adjusts the mean of each feature dimension to 0 and the standard deviation to 1. The standardized base signal data segments serve as the input for subsequent noise simulation and model training. Figure 2 In this context, Ex, Ey, Hx, and Hy represent the main physical quantities in magnetotelluric observations. Ex represents the east-west electric field component of the Earth's surface, and Ey represents the north-south electric field component, reflecting the changes in the Earth's electric field in different directions. Hx and Hy represent the east-west and north-south magnetic field components of the Earth's surface, respectively, used to describe the intensity of changes in the magnetic field in two horizontal directions.

[0023] It should be noted that, considering the input requirements of subsequent deep learning models and the integrity of signal coverage, this embodiment uses a fixed-length sliding window method to segment the above-mentioned synthetic time series. Specifically, the window length is set to 4096 sampling points. This length can fully cover the longest duration signal (such as the rectangular wave interference period) and also meet the input dimension requirements of the fully convolutional network (FCN) structure. At each sampling rate, the data of each channel of measuring point B is extracted at equal intervals to obtain 632*4 basic signal data segments, which constitute the basic sample set. Among them, "channel data" refers to the time series of multiple physical components recorded by measuring point B in magnetotelluric observation, including the east-west electric field (Ex), north-south electric field (Ey), east-west magnetic field (Hx), and north-south magnetic field (Hy) of the Earth's surface, which respectively reflect the response of the underground electrical structure to the natural electromagnetic field.

[0024] It should be understood that multi-band fundamental signal time series data are generated by acquiring a preset standard geoelectric structure model, and the fundamental signal time series data of all frequency bands are segmented using a preset fixed sliding window to obtain fundamental signal data segments; pre-analysis is performed on each fundamental signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the fundamental signal data segments; data preprocessing is performed on the noise primitives to generate composite noise profiles, and the composite noise profiles are superimposed on the fundamental signal data segments to obtain noisy data segments; a denoised dataset is constructed using the noisy data segments as samples and the composite noise profiles as labels; and a U-Net backbone is constructed. The TransUnet model is used, and a noise contour extraction model is trained by inputting the denoised dataset into the TransUnet model. Measured magnetotelluric signals are then input into the noise contour extraction model to obtain predicted noise contours. Based on the measured magnetotelluric signals and the predicted noise contours, the target magnetotelluric signal is obtained. An adaptive noise generation mechanism matching the features of the real magnetotelluric signal is constructed, which can generate an adaptive noise sequence that highly matches the features of the original magnetotelluric signal. This noise sequence is superimposed on the original signal to construct a denoised dataset that closely resembles the real scene, providing high-quality data support for the training of subsequent magnetotelluric signal denoising models. This invention is suitable for high-precision MT signal denoising of signals with different sampling rates, effectively suppressing non-stationary strong interference, improving signal quality and the reliability of geological interpretation.

[0025] Optionally, at least three noise mathematical models corresponding to the measured MT data are constructed based on the noise characteristics, including: The typical noise characteristics in the measured MT time series data are analyzed, and the mathematical models of transient waves, rectangular waves, and triangular waves are expressed as follows: (1) (2) (3) in, A mathematical model of noise representing a transient wave. A mathematical model of noise representing a rectangular wave. The mathematical model of noise represents a triangular wave; t represents the time variable or sampling point index, used to determine the amplitude of the noise at each time point; K represents the number of independent noise events superimposed in the time series; This represents the peak value or amplitude of the k-th noise. Indicates the left half width of the transient wave. Indicates the right half width of the transient wave. Indicates the transient center position. Indicates the center position of the triangular wave. This indicates the space affected by noise disturbance.

[0026] In this embodiment, as Figure 4 As shown, a pre-analysis is performed on each base signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the base signal data segment. This includes: calculating the local energy envelope and dominant frequency of each base signal data segment using short-time Fourier transform; performing moving average smoothing on the local energy envelope and the dominant frequency to obtain a smoothed energy sequence and a smoothed dominant frequency sequence; normalizing the smoothed energy sequence to obtain a normalized energy sequence, and calculating the difference between the 95th percentile and the 5th percentile of the base signal data segment to obtain the peak-to-peak value; calculating adaptive noise amplitude parameters, noise width parameters, and noise occurrence rate parameters based on the normalized energy sequence, the peak-to-peak value, and the smoothed dominant frequency sequence, and performing upper and lower limit amplitude limiting on each parameter; and assigning the parameters after upper and lower limit amplitude limiting to the noise model parameters of the noise mathematical model to generate noise primitives. The calculation expressions for the noise amplitude parameter, the noise width parameter, and the noise occurrence rate parameter are as follows: (4) (5) (6) Where i represents the number of base signal data segments. Indicates the noise amplitude parameter. Indicates peak-to-peak value. Represents the normalized energy sequence. This represents a standard normally distributed random sequence after smoothing and normalization using a moving average. Indicates the noise width parameter. Indicates the noise occurrence rate parameter; for implement Limiting operation, for implement Limiting operation.

[0027] It should be noted that the process of preprocessing the noise primitives to generate a composite noise profile, and then superimposing the composite noise profile onto the base signal data segment to obtain a noisy data segment, and constructing a denoised dataset using the noisy data segment as a sample and the composite noise profile as a label, includes: generating the composite noise profile by combining, transforming and nesting the noise primitives generated by formulas (1), (2) and (3), wherein the data preprocessing includes combining, transforming and nesting the noise primitives; injecting the composite noise profile into multiple base signal data segments to generate a noisy data segment; wherein generating the composite noise profile by combining, transforming and nesting the noise primitives includes: based on the local energy envelope A smooth energy sub-envelope env1 is generated, and a smooth random sub-envelope env2 is generated based on a standard normal distribution random sequence. The two sub-envelopes are weighted and fused to obtain a composite modulation envelope. Noise primitives corresponding to transient waves, rectangular waves, and triangular waves are combined according to random proportions, and amplitude scaling and time axis offset transformations are performed on the combined noise primitives. The transformed noise primitives at different scales are nested and superimposed, and the nested and superimposed noise primitives are multiplied by the composite modulation envelope to obtain a composite noise profile. Before superimposing the composite noise profile onto the base signal data segment, energy control is performed on the composite noise profile, and the energy of the base signal data segment is calculated. Initial energy of the composite noise profile According to the ratio of target noise energy to signal energy Calculate the scaling factor Multiply the composite noise profile by the scaling factor. And superimposed on the base signal data segment.

[0028] Specifically, taking 15Hz data as an example, the noise addition process is illustrated. A single sample data segment (4096 sampling points in length) is obtained through sliding window segmentation. The noise contour serves as the sample label (Table) for network output, while the noisy data segment serves as the sample feature (Feature) for network input. The noise amplitude, width, and location parameters are generated through adaptive modulation to simulate the diversity and uncertainty of real noise. The generated noise contour is superimposed on the base signal data segment obtained in step S1 to generate the noisy data segment, thus constructing a supervised learning sample library with the noisy data segment as input and the noise contour as the output label.

[0029] Optionally, a TransUnet model with U-Net as the backbone is constructed, and the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model, including: A global context modeling module composed of TransformerEncoders is embedded in the bottleneck layer where the feature extraction path and signal reconstruction path of U-Net intersect. The global context modeling module uses the self-attention mechanism of TransformerEncoder to explicitly model the long-range dependency between each time sampling point in the noisy data segment. The TransUnet model extracts multi-scale local features through the aforementioned feature extraction path, which includes several one-dimensional convolutional layers. Each convolutional layer is followed by a batch normalization function and a ReLU activation function. The mathematical expression for the output of the l-th layer is: (7) in, Indicates the first Layer output features, , These represent the convolution kernel and bias parameters, respectively. Indicates the first Layer output features; This indicates batch normalization, used to normalize the output of convolutions. This represents a one-dimensional convolution operation used to extract local features from time series data. This indicates that the activation function of the linear unit is modified to introduce nonlinearity, ensuring that the network can learn complex mappings; The TransUnet model's signal reconstruction path gradually restores the time series resolution corresponding to the time series through upsampling or transposed convolution, and concatenates the high-resolution features of the corresponding layers in the feature extraction path through skip connections to achieve fine reconstruction of the noise contour. The corresponding mathematical expression is: (8) in, Indicates the first Layer recovery features, Indicates the first Layer recovery features, This indicates a feature concatenation operation. This indicates an upsampling operation.

[0030] Specifically, Figure 3The network structure of the TransUnet model is shown. The TransUnet model first extracts multi-scale local features through an encoder, and then reconstructs the signal through a decoder. The encoder consists of multiple layers of one-dimensional convolutions, batch normalization, and activation functions, and pooling is used to reduce the temporal resolution. The TransUnet model includes a feature extraction path, a signal reconstruction path, and a global context modeling module. U-Net has a U-shaped structure, with the encoder (feature extraction path) on the left, the bottleneck layer in the middle, and the decoder (signal reconstruction path) on the right. TransUnet essentially replaces the bottleneck layer of U-Net with the global context modeling module.

[0031] In this embodiment, the execution process of the global context modeling module includes: the global context modeling module receiving multi-channel feature maps output by the feature extraction path, and flattening and rearranging the feature maps along the time dimension to obtain a two-dimensional feature sequence; and introducing learnable positional encoding into the two-dimensional feature sequence. To characterize the location information of each sampling point in the time series. The feature sequence after positional encoding is then input into at least one Transformer Encoder to obtain the self-attention output feature. The Transformer Encoder includes a multi-head self-attention layer and a feedforward neural network layer. The corresponding mathematical expression for calculating the attention output feature is: (9) in, This represents the attention output features. , , These represent the query vector, key vector, and value vector obtained by linear mapping of the two-dimensional feature sequence, respectively. The dimension of the key vector is represented by T, which represents the transpose operation. According to formula (9), the global correlation between the features corresponding to any two time sampling points in the two-dimensional feature sequence is modeled to obtain the global features containing long-range dependencies in the time dimension.

[0032] Specifically, the time dimension refers to T in [B, C, T]. The core idea is to cut a long signal into small segments, flatten each segment into a vector, and arrange them in a row. The Transformer Encoder then analyzes the relationships between these segments. This segment-by-segment approach, rather than point-by-point, reduces computational cost while preserving local continuity information, allowing the Transformer Encoder to learn local patterns and global dependencies of the signal more efficiently. For specific implementation details, refer to [link to implementation details]. Figure 3The bottom left corner. Step 1: Patch along the time dimension. For example, if the patch size is 16, then 256 ÷ 16 = 16 patches, [B, 512, 256] becomes [B, 512, 16, 16], cutting the time series into 16 segments, each 16 segments long; Step 2: Rearrange, changing "data arranged continuously by time" to "organized by blocks", for example, from [B, 512, 16, 16] to [B, 16, 16, 512], the original data was: (channel first, time last), now it becomes: (first by "patch number", time within the patch, channel); Step 3: Flatten, for example, from [B, 16, 16, 512] to [B, 16, 8192], each patch becomes a token vector, which is the input format required by the Transformer Encoder.

[0033] It should be noted that the execution process of the global context modeling module also includes: The global features are fused with the multi-scale local features of the feature extraction path to obtain the fused features, and the corresponding mathematical expression is: (10) in, Indicates the first Layer fusion features This indicates a feature concatenation operation. Indicates the first Layer output features, Represents global features; The fused feature is input to the signal reconstruction path and the time series resolution corresponding to the time series is gradually restored through multi-level upsampling operations. Based on the time series resolution, a noise profile with the same data segment length as the fused feature is output at the end of the signal reconstruction path.

[0034] Specifically, such as Figure 6 , Figure 7 , Figure 8 and Figure 9 As shown, the original noise primitives are first processed and combined to generate a more complex "composite noise profile." This noise is then superimposed onto the clean base signal data segment to obtain a noisy data segment. Next, the "noisy data segment" is used as the model input, and the corresponding "composite noise profile" is used as the label (i.e., the part the model needs to learn to predict or remove). Each pair (noisy data, composite noise profile) constitutes a training sample, and all samples together form the dataset used to train the denoising model. The input to the network is the noisy data segment, and its corresponding composite noise profile serves as the label, guiding the network to learn in the direction of the label.

[0035] Optionally, the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model, including: Using composite noise contours as labels, and employing a weighted mean square error (MSE) that distinguishes between noise foreground and signal background as the loss function, the TransUnet model is trained under supervision to obtain a noise contour extraction model. The mathematical expression for the loss function is as follows: (11) (12) in, Let y[t] represent the loss function, and y[t] represent the true value of the sample. This represents the network's predicted output value; Indicates the sample point index; This represents the foreground mask, taking the position of the foreground with a true value that is not zero; This indicates that the background mask is taken from a position outside the foreground. The weights are for the background penalty term. This loss design effectively alleviates the overfitting problem of the network model in the background region, improving the overall denoising quality and structural consistency.

[0036] In this embodiment, a stochastic gradient descent (SGD) optimizer can be used during training, meaning that the SGD optimizer is used to train the low-frequency, mid-frequency, and high-frequency bands independently. The initial learning rate for each independent training band is set to 1e-3, and multiple epochs of training are performed. To obtain optimal performance, dedicated models are trained separately for different sampling rates (e.g., 15Hz, 150Hz, and 2400Hz).

[0037] Optionally, the measured magnetotelluric signal is input into the noise profile extraction model to obtain a predicted noise profile, and the target magnetotelluric signal is obtained based on the measured magnetotelluric signal and the predicted noise profile, including: The measured time series data of the collected magnetotelluric signals corresponding to the measured MT signals were subjected to the same data preprocessing as the TransUnet model. The preprocessed MT measured time series data is input into the noise profile extraction model and the predicted noise profile is output. The measured magnetotelluric signal is subtracted from the predicted noise profile to obtain the denoised clean magnetotelluric signal, wherein the target magnetotelluric signal is the denoised clean magnetotelluric signal.

[0038] In this embodiment, the noisy MT measured time series data (L010S020 measuring point, including Ex, Ey, Hx and Hy components) collected in the field is preprocessed in the same way as the training data (such as normalization), and then input into the trained network model; the model automatically infers and outputs the predicted noise profile; finally, the original noisy signal is subtracted from the predicted noise profile to obtain the denoised pure magnetotelluric field signal. Figure 10 The original waveform of the L010S020 measuring point. Figure 11 A comparison of the denoising performance of different models on the Ey channel of the L010S020 measurement point is presented, from top to bottom: the denoising performance of the noisy signal, the model of this invention, U-Net, ResNet, and TCN. ResNet and TCN are insufficient in suppressing transient noise with small amplitudes, and their signal restoration in the rectangular wave plateau region is poor. While U-Net shows some improvement, the rectangular wave region is still not fully recovered. In contrast, the network model of this invention provides the most thorough noise suppression and the highest signal integrity.

[0039] The core logic of the "adaptive modulation" in the magnetotelluric signal processing method based on adaptive noise modulation provided by this invention is as follows: First, based on the characteristics of the MT base signal (local energy envelope, dominant frequency), the core parameters (amplitude, width, duration) of a single type of noise are adaptively adjusted. Then, the complexity is increased through combination, transformation, and nesting, ensuring both "adaptive fit to MT characteristics" and "noise complexity matching actual measurements." In other words, the core of the magnetotelluric signal processing method based on adaptive noise modulation proposed in this invention lies in constructing an adaptive noise generation mechanism that matches the characteristics of real magnetotelluric signals. The following example uses the processing of the first segment of time-series data (file ID: B.NE1-STS3) from the NE1 channel with a sampling frequency of 15Hz. Figure 5 The process of generating adaptive noise is explained in detail below: I. Data Preprocessing Stage This stage involves reading and standardizing the original magnetotelluric time-series data, laying the foundation for subsequent feature extraction and noise generation. The specific steps are as follows: Data Reading: Read the timing data file of channel NE1 stored in `. / data / clean / 15HZ / B.NE1-STS3`. The sampling frequency of this data is 15Hz, the total length of the original data is 2,591,997 sampling points, the data format is CSV, and the first column of the file is extracted as the valid timing signal data. Segmentation: A fixed-length segmentation strategy was adopted, setting the segment length to 4096 sampling points. The original time-series data was divided into equal-length segments, retaining only complete data segments with a length of 4096 sampling points and discarding incomplete segments with fewer than 4096 sampling points at the end. After segmentation, a total of 632 valid data segments were obtained from the original data. The first data segment, consisting of 4096 sampling points corresponding to indices 0 to 4095 in the original data, is denoted as the data segment to be processed. .

[0040] II. Feature Extraction Stage This stage extracts the time-frequency features of the data segment to be processed through Short Time Fourier Transform (STFT) to achieve adaptive correlation between noise parameters and the characteristics of the original signal. The specific steps are as follows: STFT parameter configuration and transformation: For the first 4096-point data segment A short-time Fourier transform is performed, with the analysis window length set to 128 sampling points. A rectangular window is used as the analysis window function. After STFT transformation, a time-frequency matrix with a dimension of 65×65 (65 points in the frequency dimension and 65 points in the time dimension) is obtained. This matrix represents the amplitude characteristics of the data segment at different times and frequencies. Energy characteristic calculation: The power spectrum matrix is ​​obtained by squaring the amplitude of the time-frequency matrix. Then, the power spectrum matrix is ​​summed in the frequency dimension to obtain the energy sequence E in the time dimension. This sequence reflects the energy distribution characteristics of the data segment at different time points. In this embodiment, the value range of the energy sequence E is 0.000037~0.156747. Dominant frequency feature calculation: For each time point in the power spectrum matrix, the frequency corresponding to the maximum power spectrum value is extracted as the dominant frequency at that time point, thus obtaining the dominant frequency sequence. This sequence reflects the dominant frequency characteristics of the data segment at different time points. In this embodiment, the dominant frequency sequence is... The value range is 0.000Hz to 0.938Hz; Feature interpolation matching: Due to the energy sequence E and the dominant frequency sequence obtained by STFT transformation The length is 65 points, and the data segment to be processed The 4096 points have a length mismatch, so a linear interpolation method is used to connect the energy sequence E and the main frequency sequence. Interpolation from 65 points to 4096 points ensures that the length of the feature sequence is completely consistent with the length of the data segment to be processed, guaranteeing dimensionality matching for subsequent parameter calculations.

[0041] III. Adaptive Noise Parameter Generation Stage Based on the extracted time-frequency features, this stage generates adaptive noise parameters that dynamically match the signal characteristics. The specific steps are as follows: Basic feature smoothing: To eliminate high-frequency fluctuations in the feature sequence, the energy sequence E and the dominant frequency sequence are smoothed respectively. A 200-point moving average smoothing process is performed to obtain the smoothed energy sequence. and smoothed main frequency sequence The window length of the moving average can be adjusted according to the actual signal fluctuation. In this embodiment, the window length is set to 200 sampling points. Energy normalization: This process normalizes the smoothed energy sequence. Perform a normalization operation to map its value range to the interval 0~1, resulting in a normalized energy sequence. The purpose of normalization is to eliminate the influence of differences in absolute energy values, making it a core basis for adaptive parameter adjustment. The normalization formula is: Among them, Used to avoid cases where the denominator is 0; Peak-to-peak value calculation: Calculate the data segment to be processed The difference between the 95th percentile and the 5th percentile (i.e., peak-to-peak value) reflects the amplitude fluctuation range of the data segment and can be used as a benchmark scale for noise amplitude. In this embodiment, the peak-to-peak value of the data segment to be processed is 0.536068. Adaptive parameter output: Noise amplitude parameter A: It is determined by the peak-to-peak value, normalized energy sequence and random smoothing component, and reflects the magnitude of the noise event. In this embodiment, the value range of noise amplitude parameter A is 0.0238~0.1973. The noise width parameter W is obtained by inverse adjustment of the main frequency sequence, and its value range is limited to 20 to 200 sampling points, reflecting the time width characteristics of the noise event. In this embodiment, the value range of the noise width parameter W is 67.9 to 100.0 sampling points. The noise occurrence rate parameter λ is positively correlated with the normalized energy sequence, and its value range is limited to 0.001~0.02. It reflects the probability of noise events occurring in the time series. In this embodiment, the noise occurrence rate parameter λ is in the range of 0.001~0.01.

[0042] IV. Noise Generation and Modulation Stage This stage employs a composite envelope modulation and multi-waveform random selection strategy to generate a noise sequence with time-varying and non-stationary characteristics. The specific steps are as follows: Composite envelope generation: A composite envelope is generated by combining a 300-point smoothed energy sequence and a 500-point smoothed random sequence. This envelope is used to achieve time-varying modulation of the noise amplitude, so that the noise amplitude changes dynamically with the signal energy and random components. In this embodiment, the amplitude range of the composite envelope is 0.3472~1.9760. Noise event sampling: Random sampling is performed on the time series based on the noise occurrence rate parameter λ. Among the 4096 sampling points in the data segment to be processed, the positions that meet the sampling conditions are selected as noise event trigger positions. In this embodiment, a total of 15 noise event trigger positions are generated. Multi-waveform noise generation: For each noise event trigger location, a noise primitive is generated by randomly selecting one of the following waveform types: pulse wave, square wave, and triangle wave. In this embodiment, the waveforms and parameters of the first 5 typical noise events are as follows: Event 0 (trigger position 230): Uses square wave noise with an amplitude of 0.0386 and a width of 81 sampling points; Event 1 (Trigger position 566): Pulse noise with an amplitude of 0.0542 and a width of 93 sampling points is used; Event 2 (Trigger location 746): Uses square wave noise with an amplitude of 0.0371 and a width of 100 sampling points; Event 3 (Trigger position 803): Uses triangular wave noise with an amplitude of 0.0450 and a width of 100 sampling points; Event 4 (Trigger position 2364): Uses square wave noise with an amplitude of 0.1040 and a width of 98 sampling points; After all noise primitives are generated, amplitude modulation is performed through a composite envelope, and the noise primitives are superimposed to obtain the initial noise sequence.

[0043] V. Energy Control Stage This stage uses energy normalization to ensure that the energy ratio between the noise and the original signal meets the preset signal-to-noise ratio requirement. The specific steps are as follows: Energy calculation: Calculate the data segments to be processed separately. The energy (mean square value) and the energy of the initial noise sequence are used to determine the energy of the signal. In this embodiment, the energy of the original signal is 0.026870 and the energy of the initial noise sequence is 0.035464. Scaling factor calculation: Set the target signal-to-noise ratio (noise energy / signal energy) to 7. Calculate the scaling factor based on the original signal energy and the initial noise sequence energy. The calculation formula is as follows: , where r is the target signal-to-noise ratio. The original signal energy, The initial noise sequence energy is represented by a scaling factor of 2.3030 in this embodiment. Noise sequence adjustment: The initial noise sequence is multiplied by the scaling factor so that the ratio of the energy of the adjusted noise sequence to the energy of the original signal meets the preset target signal-to-noise ratio requirement, and finally an adaptive noise sequence that can be superimposed on the original signal is obtained.

[0044] Through the above five stages of processing, an adaptive noise sequence that highly matches the characteristics of the original magnetotelluric signal can be generated. After superimposing this noise sequence onto the original signal, a denoised dataset that closely resembles the real scene can be constructed, providing high-quality data support for the training of subsequent magnetotelluric signal denoising models.

[0045] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0046] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0047] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for magnetotelluric signal processing based on adaptive noise modulation, characterized in that, Includes the following steps: The system acquires a preset standard geoelectric structure model to generate multi-band base signal MT time series data, and uses a preset fixed sliding window to segment the base signal MT time series data of all frequency bands to obtain base signal data segments. Pre-analysis is performed on each base signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the base signal data segment. Among them, at least three noise mathematical models corresponding to the measured MT data noise characteristics are constructed. The noise primitives are preprocessed to generate a composite noise profile, and the composite noise profile is superimposed on the base signal data segment to obtain a noisy data segment. A denoised dataset is constructed using the noisy data segment as a sample and the composite noise profile as a label. A TransUnet model with U-Net as the backbone is constructed, and the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model. The TransUnet model includes a feature extraction path, a signal reconstruction path, and a global context modeling module. The measured magnetotelluric signal is input into the noise profile extraction model to obtain the predicted noise profile, and the target magnetotelluric signal is obtained based on the measured magnetotelluric signal and the predicted noise profile.

2. The method of processing magnetotelluric signals based on adaptive noise modulation according to claim 1, characterized in that, The system acquires multi-band base signal time series data from a preset standard geoelectric structure model, and then uses a preset fixed sliding window to segment the base signal time series data of all frequency bands to obtain base signal data segments, including: Acquire synthetic MT time series data of the measurement point in three frequency bands: low frequency, medium frequency, and high frequency, wherein the low frequency, medium frequency, and high frequency are 15Hz, 150Hz, and 2400Hz, respectively; Set the window length for multiple sampling points in the synthesized MT time series data, and perform equal-interval sliding extraction on the channel data of each measurement point to obtain the segmented data; The segmented data is Z-score normalized to obtain multiple base signal data segments.

3. The method of processing magnetotelluric signals based on adaptive noise modulation according to claim 2, characterized in that, Based on the noise characteristics of measured MT data, at least three noise mathematical models corresponding to the waveforms are constructed, including: The typical noise characteristics in the measured MT time series data are analyzed, and the mathematical models of transient waves, rectangular waves, and triangular waves are expressed as follows: （1） （2） （3） in, A mathematical model of noise representing a transient wave. A mathematical model of noise representing a rectangular wave. The mathematical model of noise represents a triangular wave; t represents the time variable or sampling point index, used to determine the amplitude of the noise at each time point; K represents the number of independent noise events superimposed in the time series; This represents the peak value or amplitude of the k-th noise. Indicates the left half width of the transient wave. Indicates the right half width of the transient wave. Indicates the transient center position. Indicates the center position of the triangular wave. This indicates the space affected by noise disturbance.

4. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 3, characterized in that, Pre-analysis is performed on each base signal data segment, and the model parameters of the noise mathematical model are adaptively modulated to generate noise primitives corresponding to the base signal data segment, including: The local energy envelope and dominant frequency of each base signal data segment are calculated using short-time Fourier transform. The local energy envelope and the dominant frequency are respectively smoothed by moving average to obtain a smoothed energy sequence and a smoothed dominant frequency sequence. The smoothed energy sequence is normalized to obtain a normalized energy sequence, and the peak-to-peak value is obtained by calculating the difference between the 95th percentile and the 5th percentile of the base signal data segment. Based on the normalized energy sequence, the peak-to-peak value, and the smoothed main frequency sequence, adaptive noise amplitude parameters, noise width parameters, and noise occurrence rate parameters are calculated respectively, and upper and lower limits are applied to each parameter. The parameters after upper and lower limit amplitude limiting are assigned to the model parameters of the noise mathematical model to generate noise primitives; The calculation expressions for the noise amplitude parameter, the noise width parameter, and the noise occurrence rate parameter are as follows: （4） （5） （6） Where i represents the number of base signal data segments. Indicates the noise amplitude parameter. Indicates peak-to-peak value. Represents the normalized energy sequence. This represents a standard normally distributed random sequence after smoothing and normalization using a moving average. Indicates the noise width parameter. Indicates the noise occurrence rate parameter; for implement Limiting operation, for implement Limiting operation.

5. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 3, characterized in that, The noise primitives are preprocessed to generate a composite noise profile, and the composite noise profile is superimposed on the base signal data segment to obtain a noisy data segment. A denoised dataset is constructed using the noisy data segment as samples and the composite noise profile as labels, including: The noise primitives generated by formulas (1), (2) and (3) are combined, transformed and nested to generate the composite noise profile, wherein the data preprocessing includes the combination, transformation and nesting of the noise primitives; The composite noise profile is injected into multiple base signal data segments to generate noisy data segments; The generation of the composite noise profile through the combination, transformation, and nesting of the noise primitives includes: Based on the local energy envelope, a smooth energy sub-envelope env1 is generated, and a smooth random sub-envelope env2 is generated based on a standard normal distribution random sequence. The two sub-envelopes are then weighted and fused to obtain a composite modulation envelope. The noise primitives corresponding to transient waves, rectangular waves, and triangular waves are combined in a random ratio, and the combined noise primitives are then subjected to amplitude scaling and time axis offset transformation operations. The noise primitives transformed at different scales are nested and superimposed, and the nested and superimposed noise primitives are multiplied with the composite modulation envelope to obtain the composite noise profile. Before superimposing the composite noise profile onto the base signal data segment, the following steps are also included: Energy control is applied to the composite noise profile to calculate the energy of the base signal data segment. Initial energy of the composite noise profile According to the ratio of target noise energy to signal energy Calculate the scaling factor Multiply the composite noise profile by the scaling factor. And superimposed on the base signal data segment.

6. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 1, characterized in that, A TransUnet model with U-Net as the backbone is constructed, and the denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model, including: A global context modeling module composed of Transformer Encoders is embedded in the bottleneck layer where the feature extraction path and signal reconstruction path of U-Net intersect. The global context modeling module uses the self-attention mechanism of Transformer Encoder to explicitly model the long-range dependency between each time sampling point in the noisy data segment. The TransUnet model extracts multi-scale local features through the aforementioned feature extraction path, which includes several one-dimensional convolutional layers. Each convolutional layer is followed by a batch normalization function and a ReLU activation function. The mathematical expression for the output of the l-th layer is: （7） in, Indicates the first Layer output features, , These represent the convolution kernel and bias parameters, respectively. Indicates the first Layer output features; This indicates batch normalization, used to normalize the output of convolutions. This represents a one-dimensional convolution operation used to extract local features from time series data. This represents the modified linear unit activation function; The TransUnet model's signal reconstruction path gradually restores the time series resolution corresponding to the time series through upsampling or transposed convolution, and concatenates the high-resolution features of the corresponding layers in the feature extraction path through skip connections to achieve fine reconstruction of the noise contour. The corresponding mathematical expression is: （8） in, Indicates the first Layer recovery features, Indicates the first Layer recovery features, This indicates a feature concatenation operation. This indicates an upsampling operation.

7. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 6, characterized in that, The execution process of the global context modeling module includes: The global context modeling module is used to receive the multi-channel feature maps output by the feature extraction path, and flatten and rearrange the feature maps along the time dimension to obtain a two-dimensional feature sequence. Introducing learnable positional encoding into the two-dimensional feature sequence To characterize the location information of each sampling point in the time series. The feature sequence after positional encoding is then input into at least one Transformer Encoder to obtain the self-attention output feature. The Transformer Encoder includes a multi-head self-attention layer and a feedforward neural network layer. The corresponding mathematical expression for calculating the attention output feature is: （9） in, This represents the attention output features. , , These represent the query vector, key vector, and value vector obtained by linear mapping of the two-dimensional feature sequence, respectively. The dimension of the key vector is represented by T, which represents the transpose operation. According to formula (9), the global correlation between the features corresponding to any two time sampling points in the two-dimensional feature sequence is modeled to obtain the global features containing long-range dependencies in the time dimension.

8. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 7, characterized in that, Also includes: The global features are fused with the multi-scale local features of the feature extraction path to obtain the fused features, and the corresponding mathematical expression is: （10） in, Indicates the first Layer fusion features This indicates a feature concatenation operation. Indicates the first Layer output features, Represents global features; The fused feature is input to the signal reconstruction path and the time series resolution corresponding to the time series is gradually restored through multi-level upsampling operations. Based on the time series resolution, a noise profile with the same data segment length as the fused feature is output at the end of the signal reconstruction path.

9. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 8, characterized in that, The denoised dataset is input into the TransUnet model for training to obtain a noise contour extraction model, including: Using composite noise contours as labels, and employing a weighted mean square error (MSE) that distinguishes between noise foreground and signal background as a loss function, the TransUnet model is trained under supervision to obtain a noise contour extraction model. The mathematical expression for the loss function is as follows: （11） （12） in, Let y[t] represent the loss function, and y[t] represent the true value of the sample. This represents the network's predicted output value; Indicates the sample point index; This represents the foreground mask, taking the position of the foreground with a true value that is not zero; This indicates that the background mask is taken from a position outside the foreground. The weights of the background penalty term are used; a stochastic gradient descent optimizer is used to train the low-frequency, mid-frequency, and high-frequency bands independently. The initial learning rate for independent training is set to 1e-3, and multiple rounds of training are performed.

10. The magnetotelluric signal processing method based on adaptive noise modulation according to claim 1, characterized in that, The measured magnetotelluric signal is input into the noise profile extraction model to obtain a predicted noise profile. Based on the measured magnetotelluric signal and the predicted noise profile, the target magnetotelluric signal is obtained, including: The measured time series data of the collected magnetotelluric signals corresponding to the measured MT signals were subjected to the same data preprocessing as the TransUnet model. The preprocessed MT measured time series data is input into the noise profile extraction model and the predicted noise profile is output. The measured magnetotelluric signal is subtracted from the predicted noise profile to obtain the denoised clean magnetotelluric signal, wherein the target magnetotelluric signal is the denoised clean magnetotelluric signal.

Images