Method and device for magnetotelluric data denoising based on deep fusion model

Abstract

The application provides a method and device for removing noise from magnetotelluric data based on a deep fusion model, which is applied to a magnetotelluric data noise removal system and relates to the technical field of magnetotelluric data processing. A deep fusion model with a 'noise feature recognition-signal selective reconstruction' architecture is established, pre-classification of input signals is performed through a front noise detection module, the problem of effective signal false attenuation in traditional end-to-end noise removal is effectively avoided, a noise removal autoencoder containing a residual structure is used, the deep feature extraction capability of the residual structure is combined with the signal reconstruction advantage of the noise removal autoencoder, the fidelity and integrity of the reconstructed signal are significantly improved, the noise removal process does not depend on manual intervention, and the efficiency and accuracy of data processing are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetotelluric data processing, and in particular to a magnetotelluric data denoising method and device based on a deep fusion model. Background Art

[0002] Magnetotelluric sounding (MT) is a geophysical exploration method based on natural field sources. After more than half a century of development, it has formed a complete system including high-resolution sensors, large dynamic range receivers, high-precision impedance estimation and forward and inversion interpretation technologies.

[0003] However, in practical applications, MT faces many technical challenges, the core of which stems from the inherent characteristics of its signal source. Natural electromagnetic fields have ultra-wideband (10 -5 Hz to 10 4 Hz), non-periodic, random, and weak. During data acquisition, signals are highly susceptible to various interferences, including 50Hz power frequency and its harmonics in industrial areas, electromagnetic noise generated by power facilities, and natural electromagnetic interference such as lightning. Especially in today's rapidly developing urbanization and the increasingly complex human electromagnetic environment, acquiring high-quality MT data poses greater challenges. These interferences can cause signal amplitude and phase distortion. When the signal-to-noise ratio falls below a critical threshold, this significantly affects the accuracy of impedance tensor estimation, ultimately leading to deviations in the resistivity model inversion results. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and device for denoising magnetotelluric data based on a deep fusion model, establish a deep fusion model with a "noise feature recognition-signal selective reconstruction" architecture, pre-classify the input signal through a front-end noise detection module, effectively avoid the problem of false attenuation of effective signals in traditional end-to-end denoising, and utilize a denoising autoencoder containing a residual structure, combining the deep feature extraction capability of the residual structure with the signal reconstruction advantage of the denoising autoencoder, significantly improving the fidelity and integrity of the reconstructed signal. The denoising process does not rely on human intervention, greatly improving the efficiency and accuracy of data processing.

[0005] In a first aspect, the present invention provides a method for denoising magnetotelluric data based on a deep fusion model, which is applied to a magnetotelluric data denoising system. The method comprises:

[0006] Construct a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted;

[0007] A pre-built basic denoising model is trained based on a signal sample library until preset conditions are met, and the trained basic denoising model is determined as the magnetotelluric data denoising model. The basic denoising model includes: a recognition network and a denoising network; the recognition network is a deep residual network with a channel spatial attention mechanism; the denoising network is a denoising autoencoder with a residual structure;

[0008] The magnetotelluric data is input into the magnetotelluric data denoising model, and the denoised magnetotelluric data is output.

[0009] In some preferred embodiments of the present invention, the signal sample library includes: an identification sample library and a denoising sample library; the step of constructing the signal sample library based on measured high-quality magnetotelluric data with preset noise inserted includes:

[0010] Obtain high-quality measured magnetotelluric data;

[0011] Sliding sampling of measured high-quality magnetotelluric data is performed based on a time window of preset length to obtain high-quality data segments;

[0012] After inserting preset noise into the high-quality data segment, a noisy data segment is obtained;

[0013] Determine the noisy data segment assigned with the first label and the high-quality data segment assigned with the second label as samples in the recognition sample library;

[0014] The noisy data segment and the corresponding high-quality data segment are determined as samples in the denoising sample library; wherein the noisy data segment is a feature and the high-quality data segment is a label.

[0015] In some preferred embodiments of the present invention, the preset noise includes: triangular wave noise, square wave noise and pulse noise; the amplitude, width and position of the preset noise inserted into the high-quality data segment are all randomly generated within a preset range.

[0016] In some preferred embodiments of the present invention, the recognition network includes four base layers and a pooling layer connected in sequence; each base layer includes two residual blocks connected in sequence; and each residual block includes a convolutional layer, a channel attention layer, and a spatial attention layer connected in sequence.

[0017] In some preferred embodiments of the present invention, the denoising network includes an encoder and a decoder; the encoder includes three consecutive residual blocks, each residual block includes two parallel channels, the first channel includes two consecutive layers of 3×1 convolution layers, a batch normalization layer and a linear rectification activation function; the second channel is a residual structure, including a layer of 1×1 convolution layer; the decoder includes two consecutive layers of 3×1 convolution layers and a flattening operation layer.

[0018] In some preferred embodiments of the present invention, the loss function of the denoising network is based on an average constraint of the weighted loss function of each data sample.

[0019] In some preferred embodiments of the present invention, the magnetotelluric data denoising model includes: a trained recognition network and a trained denoising network; the steps of inputting magnetotelluric data into the magnetotelluric data denoising model and outputting denoised magnetotelluric data include:

[0020] Input the magnetotelluric data into the trained recognition network to obtain high-quality data segments and noisy data segments;

[0021] Input the noisy data segment into the trained denoising network to obtain the denoised data segment;

[0022] The data obtained by integrating the high-quality data segments and the denoised data segments are determined as the denoised magnetotelluric data.

[0023] In a second aspect, the present invention provides a magnetotelluric data denoising device based on a deep fusion model, which is applied to a magnetotelluric data denoising system. The device includes:

[0024] A signal sample library determination module is used to construct a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted;

[0025] The denoising model training module is used to train a pre-built basic denoising model based on the signal sample library until the preset conditions are met, and the trained basic denoising model is determined as the magnetotelluric data denoising model. The basic denoising model includes: a recognition network and a denoising network; the recognition network is a deep residual network with a channel spatial attention mechanism; the denoising network is a denoising autoencoder with a residual structure;

[0026] The denoising module is used to input the magnetotelluric data into the magnetotelluric data denoising model and output the denoised magnetotelluric data.

[0027] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the magnetotelluric data denoising method based on the deep fusion model provided in the first aspect above.

[0028] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the magnetotelluric data denoising method based on the deep fusion model provided in the first aspect above.

[0029] The present invention brings the following beneficial effects:

[0030] The present invention provides a method and device for denoising magnetotelluric data based on a deep fusion model, which is applied to a magnetotelluric data denoising system. The method comprises: constructing a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted; training a pre-constructed basic denoising model based on the signal sample library until a preset condition is met, and determining the trained basic denoising model as the magnetotelluric data denoising model; wherein the basic denoising model comprises: a recognition network and a denoising network; the recognition network is a deep residual network including a channel spatial attention mechanism; and the denoising network is a denoising autoencoder including a residual structure; The magnetotelluric data is input into the magnetotelluric data denoising model, and the denoised magnetotelluric data is output; a deep fusion model with a "noise feature recognition-signal selective reconstruction" architecture is established, and the input signal is pre-classified through a front-end noise detection module, effectively avoiding the problem of false attenuation of effective signals in traditional end-to-end denoising. The denoising autoencoder containing a residual structure is used, combining the deep feature extraction capability of the residual structure with the signal reconstruction advantage of the denoising autoencoder, which significantly improves the fidelity and integrity of the reconstructed signal. The denoising process does not rely on human intervention, greatly improving the efficiency and accuracy of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 A flowchart of a method for denoising magnetotelluric data based on a deep fusion model provided in an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of obtaining high-quality data segments by intercepting measured high-quality magnetotelluric data through a time window provided by an embodiment of the present invention;

[0034] Figure 3 A sample schematic diagram of a signal sample library provided by an embodiment of the present invention;

[0035] Figure 4 A schematic diagram of the structure of an identification network provided by an embodiment of the present invention;

[0036] Figure 5 A schematic diagram of the structure of an attention mechanism layer provided by an embodiment of the present invention;

[0037] Figure 6A schematic diagram of the structure of a denoising network provided by an embodiment of the present invention;

[0038] Figure 7 A schematic diagram of identifying network output data provided by an embodiment of the present invention;

[0039] Figure 8 Schematic diagram of a noisy magnetotelluric Ex channel data before and after denoising provided by an embodiment of the present invention;

[0040] Figure 9 A schematic structural diagram of a magnetotelluric data denoising device based on a deep fusion model provided in an embodiment of the present invention;

[0041] Figure 10 A schematic structural diagram of an electronic device provided by an embodiment of the present invention.

[0042] Icon: 310 - signal sample library determination module; 320 - denoising model training module; 330 - denoising module; 400 - memory; 401 - processor; 402 - bus; 403 - communication interface. DETAILED DESCRIPTION

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0044] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0045] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0046] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" and the like indicate positions or locations based on the positions shown in the accompanying drawings, or the positions or locations in which the inventive product is typically placed when in use. These terms are intended solely to facilitate the description of the present invention and to simplify the description, and are not intended to indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third," etc., are used solely to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0047] Furthermore, terms such as "horizontal," "vertical," and "overhanging" do not necessarily imply that a component must be absolutely horizontal or overhanging, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but rather that it can be slightly tilted.

[0048] In the description of the present invention, it should also be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0049] Magnetotelluric sounding (MT method) is a geophysical exploration method based on natural field sources. After more than half a century of development, it has formed a complete system including high-resolution sensors, large dynamic range receivers, high-precision impedance estimation and forward and inversion interpretation technology. This method inverts the electrical structure of the underground medium by measuring the naturally existing alternating electromagnetic field. Its unique advantages are mainly reflected in three aspects: first, the exploration depth range is large, from tens of meters in the shallows to hundreds of kilometers in the upper mantle; second, there is no need to deploy artificial transmitting sources, and it is suitable for various complex terrain conditions; third, the equipment is light and suitable for large-scale regional detection. These characteristics make it unique in the fields of oil and gas resource exploration, geothermal surveys, metal mine exploration, and lithosphere structure research. However, in practical applications, MT faces many technical challenges, and the core problem stems from the inherent characteristics of its signal source. Natural electromagnetic fields have ultra-wideband (10⁻ 5 Hz to 10 4Hz), non-periodic, random, and weak. During data acquisition, signals are highly susceptible to various interferences, including 50Hz power frequency and its harmonics in industrial areas, electromagnetic noise generated by power facilities, and natural electromagnetic interference such as lightning. Especially in today's rapidly developing urbanization and the increasingly complex human electromagnetic environment, acquiring high-quality MT data poses greater challenges. These interferences can cause signal amplitude and phase distortion. When the signal-to-noise ratio falls below a critical threshold, this significantly affects the accuracy of impedance tensor estimation, ultimately leading to deviations in the resistivity model inversion results.

[0050] To address this challenge, a variety of data processing methods have been developed. Traditional approaches include least squares-based power spectrum estimation, robust estimation, remote reference techniques, and wavelet transform denoising. While these methods can suppress noise to a certain extent, they still have limitations, such as being effective against certain interferences but ineffective against others, requiring complex manual intervention, or being computationally inefficient and unable to process massive amounts of data.

[0051] To address the problems of poor adaptability, strong manual dependence, insufficient anti-interference ability, and low computational efficiency in traditional methods in the existing technology, the present invention proposes a magnetotelluric data denoising method and device based on a deep fusion model. The fusion model is established based on an improved residual neural network (ResNet) and a denoising autoencoder (DAE). This fusion model is suitable for denoising magnetotelluric (MT) signals. Through a two-stage strategy of "identification first, denoising later", it effectively identifies and suppresses noise in MT signals in complex noisy environments while retaining valid signals.

[0052] The following embodiments of the present invention are described in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments may be combined with each other.

[0053] Example 1

[0054] The embodiment of the present invention provides a method for denoising magnetotelluric data based on a deep fusion model, which is applied to a magnetotelluric data denoising system. Figure 1 The flowchart of a method for denoising magnetotelluric data based on a deep fusion model provided by an embodiment of the present invention is shown, and the method includes:

[0055] Step S102 : constructing a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted.

[0056] Specifically, we collect a large amount of high-quality, field-measured magnetotelluric time series data that has been rigorously screened and preliminarily processed by experts and can be considered "relatively pure." High-quality field-measured magnetotelluric data is observational data collected using field survey lines, with a high signal-to-noise ratio and complete waveform information.

[0057] Then, to simulate the various complex noise scenarios likely encountered in real-world environments, it is necessary to systematically artificially inject various preset noise types into these "relatively pure" signal samples. These preset noises are not randomly added but are based on in-depth research and modeling of actual MT noise sources. These include, but are not limited to: simulating power frequency harmonics and their corresponding harmonics of varying intensities and frequency characteristics; generating random noise sequences with specific statistical properties (such as Gaussian, impulse, and colored noise); simulating transient noise generated by moving sources; and even attempting to synthesize complex geological noise patterns. Injection can be additive, multiplicative, or convolutional to cover a wider range of noise pollution scenarios.

[0058] Finally, these generated samples undergo necessary preprocessing (such as standardization, normalization, and framing / segmentation) and are organized into a structured dataset, known as a signal sample library. This library contains not only the original, relatively pure target signal but, more importantly, the corresponding, controllable, noisy input signal. This provides solid data support for subsequent deep learning models to learn the complex mapping relationship between noise and signal features. The size of the sample library, the coverage of noise types, and the accuracy of the noise-signal pairings directly determine the robustness and generalization capabilities of the resulting trained denoising model.

[0059] Step S104: training a pre-built basic denoising model based on the signal sample library until a preset condition is met, and determining the trained basic denoising model as the magnetotelluric data denoising model; wherein the basic denoising model includes: a recognition network and a denoising network; the recognition network is a deep residual network including a channel spatial attention mechanism; and the denoising network is a denoising autoencoder including a residual structure.

[0060] Specifically, the "noisy" data from the sample library is fed into the basic denoising model (first through the recognition network and then into the denoising network). The model outputs a predicted "clean" signal. The predicted result is compared with the actual "relatively clean" target signal from the sample library, and a loss function (such as mean squared error (MSE), L1 loss, or a combination of perceptual loss) is calculated. The network weights are updated using backpropagation algorithms (such as SGD and Adam) and gradient descent, with the goal of continuously minimizing the gap between the predicted output and the actual target. Training continues until the model's performance on an independent validation set stabilizes and a preset stopping criterion is met (such as reaching the maximum number of iterations, no significant decrease in validation set loss, or reaching a preset performance threshold). Once training converges and the conditions are met, the trained and parameter-optimized basic denoising model is officially designated as the final, usable MT data denoising model. It has learned the powerful ability to recover real MT signals from various complex noise backgrounds.

[0061] The core task of the recognition network is to deeply mine and precisely focus on the most critical feature information in noisy input data for distinguishing noise from useful signals. Its core architecture utilizes a deep residual network (ResNet) with a channel-wise spatial attention mechanism. By introducing skip connections, deep residual networks effectively alleviate the vanishing / exploding gradient problems common in deep network training, allowing for deeper networks and the learning of more complex feature representations. The channel-wise spatial attention mechanism (CBAM) typically consists of two parallel or sequential submodules: channel attention and spatial attention. This helps the model focus on channels and time segments that are more discriminative for classification tasks, thereby improving robustness and classification accuracy. The channel-wise attention module adaptively learns the importance weights of the four electromagnetic channels (Ex, Ey, Hx, and Hy) and the different feature channels after convolution and residual modules, enabling the model to identify which channels in the current input are most helpful in distinguishing signal quality (high quality or noisy). The spatial attention module models attention weights along the temporal dimension within each channel. By assigning different weights to different sampling points along the time axis, it emphasizes time segments in the signal that are critical for classification (such as abnormal waveforms or signal mutation points), thereby improving the model's ability to capture temporal information. Guided by the channel-wise spatial attention mechanism, the recognition network possesses stronger feature selection capabilities, effectively improving the classification accuracy and robustness for both high-quality and noisy data segments. Compared to models without the attention mechanism, models integrating CBAM exhibit significant advantages in classification performance.

[0062] The core task of the denoising network is to reconstruct denoised data from noisy data segments, based on the classification results provided by the recognition network, that is as close as possible to the original, relatively clean signal. Its architecture primarily utilizes a denoising autoencoder (DAE) with a residual structure. The autoencoder inherently possesses powerful data reconstruction capabilities. Its encoder compresses the input into low-dimensional features, while the decoder reconstructs the original input from these features. The denoising autoencoder (DAE) is a variant specifically designed for denoising tasks. Its training objective is to learn the mapping from noisy input to clean output. Embedding the residual structure into the DAE network framework combines the DAE's powerful reconstruction capabilities with its excellent feature learning and gradient flow properties. The denoising network input is the noisy data segment identified by the recognition network. Through multiple layers of nonlinear transformations, the network learns to remove the residual noise embedded in the features and accurately restore the original waveform, spectral characteristics, and phase information of the magnetotelluric signal. Residual connections help preserve shallow-level information in the input features and directly transmit it to deeper layers, preventing over-compression and enabling feature reuse. The DAE's learning goal is to correct high-frequency details.

[0063] Step S106: input the magnetotelluric data into a magnetotelluric data denoising model, and output the denoised magnetotelluric data.

[0064] Specifically, the new data undergoes similar preprocessing as the training data, such as format conversion, time series segmentation, and necessary normalization or standardization, to ensure that its format and scale are consistent with the input data used during model training. The preprocessed noisy magnetotelluric data is then directly input into the trained magnetotelluric data denoising model. The recognition network within the model operates first, leveraging its embedded channel-space attention mechanism to automatically analyze the input data and focus on data segments with significant noise characteristics. Next, data segments identified as noisy are fed into the denoising network. Based on the learned noise distribution characteristics and signal reconstruction rules, the denoising network deeply processes the input data segments, removes residual noise interference, and meticulously reconstructs the structure and details of the original signal. Finally, the model outputs the corresponding denoised magnetotelluric data. The output data significantly reduces the levels of various types of noise contained in the original data and improves the signal-to-noise ratio. Its apparent resistivity curve and phase curve are smoother and more reasonable, and more in line with the response of the actual underground electrical structure. It provides higher quality and more reliable basic data input for subsequent data interpretation, inversion imaging and geological modeling, thereby greatly improving the accuracy and effectiveness of magnetotelluric exploration.

[0065] The present invention provides a magnetotelluric data denoising method based on a deep fusion model, which is applied to a magnetotelluric data denoising system. The method comprises: constructing a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted; training a pre-constructed basic denoising model based on the signal sample library until a preset condition is met, and determining the trained basic denoising model as the magnetotelluric data denoising model; wherein the basic denoising model comprises: a recognition network and a denoising network; the recognition network is a deep residual network including a channel spatial attention mechanism; the denoising network is a denoising autoencoder including a residual structure; The geomagnetic data is input into the magnetotelluric data denoising model, and the denoised magnetotelluric data is output; a deep fusion model with a "noise feature recognition-signal selective reconstruction" architecture is established, and the input signal is pre-classified through a front-end noise detection module, effectively avoiding the problem of false attenuation of effective signals in traditional end-to-end denoising. The denoising autoencoder with a residual structure is used, combining the deep feature extraction capability of the residual structure with the signal reconstruction advantage of the denoising autoencoder, which significantly improves the fidelity and integrity of the reconstructed signal. The denoising process does not rely on human intervention, greatly improving the efficiency and accuracy of data processing.

[0066] Example 2

[0067] To improve the model's training accuracy and generalization capabilities, a recognition sample library and a denoising sample library were constructed from the MT signal sample library, which contains diverse noise characteristics. The recognition sample library is used to train the recognition network to automatically distinguish between high-quality data segments and noisy data segments, while the denoising sample library is used to train the denoising network, focusing on recovering clear signals from noisy data segments.

[0068] Specifically, to balance data authenticity and accuracy, we selected high-quality data at multiple sampling rates collected at different measurement points without obvious artificial interference, including four common sampling rates: 5Hz, 500Hz, 8kHz, and 128kHz. We also selected simulated high-quality data generated by forward modeling. Considering that most noise is between 10 and 150 sampling points in length, we selected a time window of 256 sampling points to slide this data, obtaining tens of thousands of high-quality data segments. Figure 2 The embodiment of the present invention provides a schematic diagram of obtaining high-quality data segments by intercepting measured high-quality magnetotelluric data through a time window. The embodiment of the present invention adopts a four-channel (Ex channel, Ey channel, Hx channel, Hy channel) time window. Figure 2 The four-channel data shown in a are slid to obtain Figure 2 The four-channel data segment with 256 sampling points is shown in b.

[0069] Compared with most existing deep learning methods, which usually assume the independence of each channel when constructing a sample library and fail to fully consider the potential connections between orthogonal channels, the embodiment of the present invention uses four-channel data and introduces correlated noise when constructing a recognition sample library, thereby ensuring that the neural network can learn the mutual relationship between orthogonal channels and the spatiotemporal correlation of noise.

[0070] Furthermore, in some preferred embodiments of the present invention, the preset noise includes: triangular wave noise, square wave noise and pulse noise; the amplitude, width and position of the preset noise inserted into the high-quality data segment are all randomly generated within a preset range.

[0071] Extensive analysis of noise in field-measured data at various sampling rates reveals that the types and forms of noise present in MT signals vary. Therefore, for high-quality data segments at different sampling rates, we insert simulated noise of varying types and forms, including triangular, square, and pulse waves. To ensure the authenticity of the simulated noise, the amplitude, width, and position of the noise are randomly generated within a preset range using code.

[0072] For identification networks, see Figure 3A sample schematic diagram of a signal sample library provided by an embodiment of the present invention is shown. For the recognition network, the recognition sample library ensures that the ratio of samples containing four-channel noisy data segments to four-channel high-quality data segments is 1:1, and assigns them a first label of 0 and a second label of 1, respectively, so that the network can learn the difference between noisy data and high-quality data. For the denoising network, in a single denoising sample, a single-channel noisy data segment is used as a feature, and its corresponding high-quality data segment is used as a label, which respectively constitute the input and output of the network.

[0073] During data preprocessing, we first performed minimum and maximum normalization to improve the comparability of different data sets and the stability of model training, avoiding training bias caused by differences in numerical ranges. At the same time, we divided the data into training, validation, and test sets in a ratio of 8:1:1.

[0074] Furthermore, in some preferred embodiments of the present invention, the recognition network includes four base layers and a pooling layer connected in sequence; each base layer includes two residual blocks connected in sequence; and each residual block includes a convolutional layer, a channel attention layer, and a spatial attention layer connected in sequence.

[0075] For details, see Figure 4 The diagram shows a structure of a recognition network provided by an embodiment of the present invention. The training data in the recognition sample library is input through a 4×256 input layer, subjected to a 3×3 convolution, and then sequentially input into a first convolution layer with a feature channel of 256, a second convolution layer with a feature channel of 128, a third convolution layer with a feature channel of 64, a fourth convolution layer with a feature channel of 32, and a pooling layer with a feature channel of 32, and outputs a result with a label of 0 or 1; wherein, the first layer includes residual blocks 1 and residual blocks 2 with the same structure, the input and output of each residual block are linked by a shortcut, and any residual block includes two layers of 3×3 convolution layers, a channel attention mechanism layer, and a spatial attention mechanism layer connected in sequence.

[0076] The noise recognition network uses a deep residual network structure based on ResNet, and embeds a Convolutional Block Attention Module (CBAM) in each residual block. Through this attention mechanism, the CBAM module automatically focuses on key features in the signal, enhancing the model's sensitivity to noise. This is particularly true when dealing with strong noise backgrounds, effectively distinguishing between noise and valid signals.

[0077] See also Figure 5 The schematic diagram of the structure of an attention mechanism layer provided by an embodiment of the present invention is shown in FIG. Figure 5 a in the figure is the structure of the channel attention mechanism layer. Figure 5 In the figure, b is the structure of the spatial attention mechanism layer.

[0078] For channel attention, we first perform channel attention on the input feature map Perform Adaptive Average Pooling (AvgPool) and Adaptive Max Pooling (MaxPool) operations to obtain and Global information representation:

[0079] ;

[0080] Among them, X avg is the adaptive average pooling output, X max is the adaptive max pooling output.

[0081] Then, these two parts of features are transformed by shared multi-layer perceptron (Shared MLP) to obtain weighted feature weights and :

[0082] ;

[0083] in, and is the weight matrix of the two-layer fully connected network, r is the channel dimensionality reduction ratio, is the ReLU activation function, is the Sigmoid activation function.

[0084] Finally, the channel attention weight is obtained by adding the two parts of features and normalizing them :

[0085] .

[0086] For spatial attention, first, calculate the input feature map Mean mapping in channel dimension and maximum mapping :

[0087] ;

[0088] Then, these two parts of information are concatenated along the channel dimension and transformed through a k×1 one-dimensional convolution to obtain the spatial attention weight :

[0089] ;

[0090] in, is the Sigmoid activation function, and Convld[·,·] represents the concatenation operation on the channel dimension.

[0091] Finally, channel attention and spatial attention are used in tandem, i.e.:

[0092] ;

[0093] Among them, X out The output of the basic block.

[0094] During the training of the basic denoising model in this embodiment of the present invention, the recognition network extracts multi-level features of the input signal through convolutional layers and then focuses on the characteristics of the noise signal through an attention mechanism. Furthermore, the network is optimized using a backpropagation algorithm to gradually improve the accuracy of noise recognition. The cross-entropy loss function is used to measure the model's classification performance, and network performance is optimized by continuously adjusting hyperparameters.

[0095] The noise recognition network provided by the present invention achieves a paradigm shift in noise recognition, moving from single-channel independent processing to four-channel joint analysis, by deeply coupling spatial-channel dual attention with four-channel electromagnetic data. Given that noise is often correlated across multiple channels—that is, when noise occurs in one channel during a certain period, other channels are often also affected at the same time—the present invention utilizes four-channel joint modeling. As long as a data segment in any channel exhibits significant noise characteristics, all four-channel data within the corresponding time period are uniformly labeled as noisy, effectively improving the overall accuracy of noise recognition. During the channel attention phase, the network automatically establishes a cross-channel correlation model of Ex-Ey-Hx-Hy. The spatial attention mechanism, by focusing on key locations in the temporal dimension, enhances the model's responsiveness to temporal anomalies such as local mutations and short-term interference. Experiments have shown that CBAM can improve the accuracy of MT signal-to-noise recognition by 0.82%.

[0096] Furthermore, in some preferred embodiments of the present invention, the denoising network includes an encoder and a decoder.

[0097] For details, see Figure 6The diagram shows the structure of a denoising network provided by an embodiment of the present invention. The encoder comprises three consecutive residual blocks, each of which includes two parallel channels. The first channel comprises two consecutive 3×1 convolutional layers, batch normalization, and a ReLU activation function; the second channel, the residual structure, comprises a single 1×1 convolutional layer. The decoder comprises two consecutive 3×1 convolutional layers and a flattening operation. The denoising network is based on a network architecture (ResDAE) that combines a DAE (denoising autoencoder) with a residual structure. The DAE network can effectively reconstruct clear signals from noisy data, with low computational complexity and strong denoising capabilities. By combining the residual structure and DAE, the network can not only extract multi-level features from noisy data but also perform efficient signal reconstruction.

[0098] The ResDAE network extracts deep features from noisy data through convolution and residual structures, and maps these features to the clean signal space using the decoding portion of the denoising autoencoder. During training, the network learns the mapping relationship for removing noise from noisy data, ensuring that the denoised signal retains as much of the original effective information as possible.

[0099] Furthermore, in some preferred embodiments of the present invention, the loss function of the denoising network is based on an average constraint of the weighted loss function of each data sample.

[0100] Specifically, the noise is determined by the amplitude threshold, and then a higher weight is applied to the loss value of these areas to enhance the model's attention to the noisy areas. The specific process is as follows:

[0101] Calculate the absolute error between the model output and the target, and set a magnitude threshold , to determine which area is noise. The calculation formula is as follows:

[0102] ;

[0103] ;

[0104] Where y is the model output, is the target output, e is the absolute error between the two, and pulse_mask represents the set of absolute errors greater than the amplitude threshold.

[0105] The Huber loss function is used to calculate the Huber loss. When the absolute error is small (error≤𝛿), the Huber loss uses square loss, which has better smoothness; when the absolute error is large (error>𝛿), the linear loss is used to avoid the influence of outliers. The calculation formula is as follows:

[0106] ;

[0107] Among them, L(e) represents the loss function of a single sample.

[0108] For the loss of the noisy areas, a higher weight (pulse_weight) is applied so that the model pays more attention to these areas during training. The calculation formula is as follows:

[0109] ;

[0110] Among them, L weighted Represents the weighted loss function of a single sample with weights.

[0111] Calculate the average of the weighted losses of all samples. This step ensures that the loss function can be updated stably during training. The calculation formula is as follows:

[0112] ;

[0113] Among them, L final Represents the average of the weighted losses of all samples, and N is the number of samples.

[0114] In view of the particularity of MT noise, the weighted Huber Loss loss function is adopted to enhance the robustness to outliers while ensuring the retention of valid signals.

[0115] The parameters of the trained basic denoising model are saved as the parameters of the magnetotelluric data denoising model.

[0116] The magnetotelluric data denoising model provided by this embodiment integrates the Convolutional Block Attention Module (CBAM) channel-by-channel spatial attention mechanism into the ResNet backbone network to identify noise segments before feature extraction. CBAM utilizes weighted spatial-channel collaborative attention to enhance the network's focus on key noise features while also suppressing interference from redundant information, enabling the recognition network to achieve a 99.3% recognition accuracy.

[0117] The magnetotelluric data denoising model provided by this embodiment of the present invention implements a dual protection mechanism for valid signal regions. On the one hand, the two-stage architecture ensures that non-noise segments are fully preserved, preventing misoperation. On the other hand, the residual learning mechanism of ResDAE can restore valid signals within noisy segments with high fidelity.

[0118] Furthermore, in some preferred embodiments of the present invention, the magnetotelluric data denoising model includes: a trained recognition network and a trained denoising network; the steps of inputting the magnetotelluric data into the magnetotelluric data denoising model and outputting the denoised magnetotelluric data include: inputting the magnetotelluric data into the trained recognition network to obtain high-quality data segments and noisy data segments; inputting the noisy data segments into the trained denoising network to obtain denoised data segments; and determining the data obtained by integrating the high-quality data segments and the denoised data segments as the denoised magnetotelluric data.

[0119] Specifically, after the model training is completed, the MT time series data to be processed is input into the trained noise recognition network. The noise recognition network determines whether each time series segment is contaminated by noise. If the data segment is identified as noise interference, it is input into the denoising network for denoising; if the data segment is identified as noise-free, the original data is directly retained. Figure 7 The diagram shown is a schematic diagram of a recognition network output data provided by an embodiment of the present invention, wherein the black portion is a high-quality data segment and the gray portion is a noisy data segment.

[0120] See also Figure 8 The diagram below shows a schematic diagram of noisy magnetotelluric (Ex) channel data before and after denoising, as provided by an embodiment of the present invention. The data processed by the denoising network is reassembled into a complete time series, generating a denoised MT signal. This process not only effectively removes noise but also ensures the accuracy and integrity of the valid signal.

[0121] The MT data denoising method based on a deep fusion model provided by the present invention directly uses unconverted time-series MT signals as input to the recognition network, compared to some other methods that use images as input. This avoids potential signal distortion during the conversion process and effectively preserves the original signal characteristics. It supports adaptive adjustment of low, medium, and high frequency sampling rates and is applicable to a variety of mainstream data formats and acquisition platforms.

[0122] The magnetotelluric data denoising method based on the deep fusion model provided in the embodiment of the present invention adopts a two-stage strategy of "identification first, denoising later" to effectively identify and suppress noise for MT signals in complex noise environments while retaining valid signals.

[0123] Example 3

[0124] On the basis of the above embodiments, the present invention provides a device for denoising magnetotelluric data based on a deep fusion model, which is applied to a magnetotelluric data denoising system. Figure 9 The structure diagram of a magnetotelluric data denoising device based on a deep fusion model provided by an embodiment of the present invention is shown, and the device includes:

[0125] The signal sample library determination module 310 is configured to construct a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted.

[0126] The denoising model training module 320 is used to train a pre-built basic denoising model based on the signal sample library until the preset conditions are met, and determine the trained basic denoising model as the magnetotelluric data denoising model; wherein the basic denoising model includes: a recognition network and a denoising network; the recognition network is a deep residual network including a channel spatial attention mechanism; the denoising network is a denoising autoencoder including a residual structure.

[0127] The denoising module 330 is configured to input the magnetotelluric data into a magnetotelluric data denoising model and output the denoised magnetotelluric data.

[0128] Furthermore, in some preferred embodiments of the present invention, the signal sample library includes: an identification sample library and a denoising sample library; a signal sample library determination module 310 is used to obtain measured high-quality magnetotelluric data; wherein, the measured high-quality magnetotelluric data is observation data collected by field-laid survey lines, having a high signal-to-noise ratio and complete waveform information; the measured high-quality magnetotelluric data is slid through a time sliding window to obtain a high-quality data segment of a preset time length; after inserting preset noise into the high-quality data segment, a noisy data segment is obtained; the noisy data segment assigned to the first label and the high-quality data segment assigned to the second label are determined as samples in the identification sample library; the noisy data segment and the corresponding high-quality data segment are determined as samples in the denoising sample library; wherein, the noisy data segment is a feature and the high-quality data segment is a label.

[0129] Generally, the first label is 0 and the second label is 1.

[0130] Furthermore, in some preferred embodiments of the present invention, the preset noise includes: triangular wave noise, square wave noise and pulse noise; the amplitude, width and position of the preset noise inserted into the high-quality data segment are all randomly generated within a preset range.

[0131] Furthermore, in some preferred embodiments of the present invention, the recognition network includes four base layers and a pooling layer connected in sequence; each base layer includes two residual blocks connected in sequence; and each residual block includes a convolutional layer, a channel attention layer, and a spatial attention layer connected in sequence.

[0132] Furthermore, in some preferred embodiments of the present invention, the denoising network (ResDAE) combines a residual structure with a DAE (denoising autoencoder), comprising an encoder and a decoder. The encoder consists of three consecutive residual blocks, each of which includes two parallel channels. The first channel comprises two consecutive 3×1 convolutional layers, a batch normalization layer, and a ReLU activation function (rectified linear unit activation function); the second channel is a residual structure with a single 1×1 convolutional layer; and the decoder comprises two consecutive 3×1 convolutional layers and a flattening layer.

[0133] Furthermore, in some preferred embodiments of the present invention, the loss function of the denoising network is based on an average constraint of the weighted loss function of each data sample.

[0134] Furthermore, in some preferred embodiments of the present invention, the magnetotelluric data denoising model includes: a trained recognition network and a trained denoising network; a denoising module 330, which is used to input the magnetotelluric data into the trained recognition network to obtain high-quality data segments and noisy data segments; input the noisy data segments into the trained denoising network to obtain denoised data segments; and determine the data obtained by integrating the high-quality data segments and the denoised data segments as denoised magnetotelluric data.

[0135] Those skilled in the art can clearly understand that, for the convenience and conciseness of description, the specific working process of the magnetotelluric data denoising device based on the deep fusion model described above can refer to the corresponding process in the aforementioned embodiment of the magnetotelluric data denoising method based on the deep fusion model, and will not be repeated here.

[0136] Example 4

[0137] The embodiment of the present invention further provides an electronic device for executing a method for denoising magnetotelluric data based on a deep fusion model; see Figure 10 The embodiment of the present invention shown is a structural schematic diagram of an electronic device provided, which includes a memory 400 and a processor 401, wherein the memory 400 is used to store one or more computer instructions, and the one or more computer instructions are executed by the processor 401 to implement the above-mentioned magnetotelluric data denoising method based on the deep fusion model.

[0138] Furthermore, Figure 10 The electronic device shown further includes a bus 402 and a communication interface 403 , and the processor 401 , the communication interface 403 and the memory 400 are connected via the bus 402 .

[0139] The memory 400 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage. The communication connection between the system network element and at least one other network element is achieved through at least one communication interface 403 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used. The bus 402 may be an ISA bus, a PCI bus, or an EISA bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 10 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0140] The processor 401 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the processor 401 or by software instructions. The above processor 401 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in conjunction with the embodiments of the present invention can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium well-known in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in memory 400, and processor 401 reads the information in memory 400 and, in conjunction with its hardware, completes the steps of the method of the aforementioned embodiment.

[0141] An embodiment of the present invention also provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the above-mentioned magnetotelluric data denoising method based on the deep fusion model. The specific implementation can be found in the method embodiment and will not be repeated here.

[0142] The computer program product of the magnetotelluric data denoising method, device and electronic device based on the deep fusion model provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method in the previous method embodiment. The specific implementation can be referred to the method embodiment and will not be repeated here.

[0143] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process of the system and / or device described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0144] In addition, in the description of the embodiments of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0145] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for denoising magnetotelluric data based on a deep fusion model, characterized in that: Applied to a magnetotelluric data denoising system, the method comprises: Construct a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted; A pre-built basic denoising model is trained based on the signal sample library until a preset condition is met, and the trained basic denoising model is determined as the magnetotelluric data denoising model; wherein, the basic denoising model includes: a recognition network and a denoising network; the recognition network is a deep residual network including a channel spatial attention mechanism; the denoising network is a denoising autoencoder including a residual structure; the recognition network includes four base layers and a pooling layer connected in sequence; each base layer includes two residual blocks connected in sequence; each residual block includes a convolution layer, a channel attention layer and a spatial attention layer connected in sequence; the denoising network includes an encoder and a decoder; the encoder includes three consecutive residual blocks, each residual block includes two parallel channels, the first channel includes two consecutive 3×1 convolution layers, a batch normalization layer and a linear rectification activation function; the second channel is a residual structure, including a 1×1 convolution layer; the decoder includes two consecutive 3×1 convolution layers and a flattening operation layer; The magnetotelluric data is input into the magnetotelluric data denoising model, and the denoised magnetotelluric data is output.

2. The method for denoising magnetotelluric data based on a deep fusion model according to claim 1, characterized in that: The signal sample library includes: an identification sample library and a denoising sample library; the steps of constructing the signal sample library based on measured high-quality magnetotelluric data with preset noise inserted include: Acquiring the measured high-quality magnetotelluric data; Sliding sampling is performed on the measured high-quality magnetotelluric data based on a time window of a preset length to obtain a high-quality data segment; After inserting the preset noise into the high-quality data segment, a noisy data segment is obtained; Determining the noisy data segment assigned with the first label and the high-quality data segment assigned with the second label as samples in the recognition sample library; The noisy data segment and the corresponding high-quality data segment are determined as samples in the denoising sample library; wherein the noisy data segment is a feature and the high-quality data segment is a label.

3. The method for denoising magnetotelluric data based on a deep fusion model according to claim 2, characterized in that: The preset noise includes: triangular wave noise, square wave noise and pulse noise; the amplitude, width and position of the preset noise inserted into the high-quality data segment are all randomly generated within a preset range.

4. The method for denoising magnetotelluric data based on a deep fusion model according to claim 1, characterized in that: The loss function of the denoising network is based on the mean constraint of the weighted loss function of each data sample.

5. The method for denoising magnetotelluric data based on a deep fusion model according to claim 1, characterized in that: The magnetotelluric data denoising model includes: the trained recognition network and the trained denoising network; the steps of inputting magnetotelluric data into the magnetotelluric data denoising model and outputting the denoised magnetotelluric data include: Inputting the magnetotelluric data into the trained recognition network to obtain high-quality data segments and noisy data segments; Inputting the noisy data segment into the trained denoising network to obtain a denoised data segment; The data obtained by integrating the high-quality data segment and the denoised data segment is determined as the denoised magnetotelluric data.

6. A device for denoising magnetotelluric data based on a deep fusion model, characterized in that: Applied to a magnetotelluric data denoising system, the device comprises: A signal sample library determination module is used to construct a signal sample library based on measured high-quality magnetotelluric data with preset noise inserted; A denoising model training module is used to train a pre-built basic denoising model based on the signal sample library until a preset condition is met, and the trained basic denoising model is determined as the magnetotelluric data denoising model; wherein, the basic denoising model includes: a recognition network and a denoising network; the recognition network is a deep residual network including a channel spatial attention mechanism; the denoising network is a denoising autoencoder including a residual structure; the recognition network includes four base layers and a pooling layer connected in sequence; each base layer includes two residual blocks connected in sequence; each residual block includes a convolution layer, a channel attention layer and a spatial attention layer connected in sequence; the denoising network includes an encoder and a decoder; the encoder includes three consecutive residual blocks, each residual block includes two parallel channels, the first channel includes two consecutive 3×1 convolution layers, a batch normalization layer and a linear rectification activation function; the second channel is a residual structure including a 1×1 convolution layer; the decoder includes two consecutive 3×1 convolution layers and a flattening operation layer; The denoising module is used to input the magnetotelluric data into the magnetotelluric data denoising model and output the denoised magnetotelluric data.

7. An electronic device, characterized in that: It includes a processor and a memory, the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the magnetotelluric data denoising method based on the deep fusion model as described in any one of claims 1 to 5 above.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by the processor, the computer-executable instructions prompt the processor to implement the magnetotelluric data denoising method based on the deep fusion model described in any one of claims 1 to 5.

Images