A magnetotelluric denoising method and device based on enhanced deep learning

Abstract

The present application relates to the technical field of magnetotelluric signal processing, in particular to a magnetotelluric denoising method and device based on enhanced deep learning; a noise-containing signal sample library and a noise profile sample library are constructed; based on the magnetotelluric signal samples of the noise-containing signal sample library and the sample training of the noise profile sample library, a signal-to-noise mapping model of different deep learning networks is trained; the magnetotelluric signal to be denoised is input into the signal-to-noise mapping model to learn the noise profile, and the denoised magnetotelluric signal is obtained through a lightweight integrated learning method, which is composed of pipeline segmentation, integrated learning, noise threshold, three-layer DBSCAN and restoration decision model; the fitting denoising ability of a single deep learning model is improved.

Description

Technical Field

[0001] This invention relates to the field of magnetotelluric signal processing technology, specifically to a magnetotelluric denoising method and apparatus based on enhanced deep learning. Background Technology

[0002] Magnetotelluric (MT) methods are widely used in geophysical exploration. By monitoring the temporal variations of the Earth's natural electromagnetic field at the Earth's surface, it helps characterize the Earth's electrical structure across a wide range of depths. However, the inherent magnetotelluric signal exhibits significant randomness, finite amplitude, and a broad spectrum, making it susceptible to interference from environmental noise. This interference significantly distorts the apparent resistivity-phase curve, subsequently affecting electromagnetic inversion procedures and geological interpretation, potentially leading to misinterpretations of subsurface configurations. Therefore, effectively reducing noise in magnetotelluric data is a crucial prerequisite for accurately determining subsurface structures.

[0003] Traditional noise suppression methods include wavelet denoising, singular value decomposition (SVD), sparse representation, empirical mode decomposition (EMD), and mathematical morphological filtering. However, these methods have limitations, such as difficulty in handling complex and nonlinear noise, the need for careful selection of parameters or foundations, and challenges in achieving accurate signal decomposition. In recent years, much research has focused on finding superior deep learning models to optimize denoising performance, making the evaluation of these models' effectiveness challenging. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a magnetotelluric denoising method and apparatus based on enhanced deep learning.

[0005] On the one hand, the present invention provides a magnetotelluric denoising method based on enhanced deep learning, the method comprising the following steps:

[0006] S1: Construct a noisy signal sample library and a noise contour sample library;

[0007] S2: Based on the magnetotelluric signal samples in the noisy signal sample library and the samples in the noise contour sample library, signal-to-noise mapping models of different deep learning networks are trained. The input of the signal-to-noise mapping model is the magnetotelluric signal sample, and the output is the noise contour sample. Different deep learning networks are deep learning network models designed based on different applications.

[0008] S3: Input the magnetotelluric signal to be denoised into the signal-to-noise mapping model to learn the noise profile, and obtain the denoised magnetotelluric signal through a lightweight ensemble learning method. The lightweight ensemble learning method consists of pipeline segmentation, ensemble learning, noise thresholding, three-layer DBSCAN, and restoration decision model.

[0009] Optionally, the steps of constructing the noise contour sample library include: constructing noise contour sample sub-libraries of multiple types of noise with the same segment length and different noise widths using different mathematical functions, wherein the noise categories include any combination of triangular waves, charging and discharging triangular waves, pulse waves, and square waves; successively changing the temporal position of the noise contours in the noise contour sample sub-libraries corresponding to each type of noise to obtain several noise contour sample sub-libraries with the same length but different noise waveform positions; sequentially changing the amplitude of all noise contour samples of each type of noise in the noise contour sample sub-libraries to obtain noise contour sample sub-libraries of the same type of noise with different amplitudes; and combining the noise contour sample sub-libraries corresponding to each type of noise to form the noise contour sample library.

[0010] Optionally, the step of constructing the noisy signal sample library includes: selecting Gaussian white noise of the same length as the noise profile sample library and whose amplitude is close to the measured magnetotelluric undulation terrain fluctuation, and selecting several segments without obvious interference from the measured magnetotelluric data through DBSCAN, thereby constructing a clean signal sample library; adding the latter half of the clean signal sample library and the randomly selected segments to the noise profile sample library of the same length as the latter half of the clean signal sample library to obtain the noisy signal sample library.

[0011] Optionally, the lightweight ensemble learning method includes the following steps: combining the pipeline segmentation with the ensemble learning to fit a temporal noise profile; eliminating the fitting error occurring in each fitted temporal noise profile using the noise threshold; averaging all the obtained noise profiles at corresponding time points to obtain the final noise profile; using a three-layer DBSCAN to determine whether the segment contains noise, and if no noise is detected, no processing is performed; if noise is detected, the noise profile is subtracted from the magnetotelluric signal to be denoised to process the noise segment; recording the start and end points when the three-layer DBSCAN detects noise until no more noise is detected, and smoothing the data period by restoring the decision model to obtain the denoised magnetotelluric signal.

[0012] Optionally, the pipeline segmentation step includes: substituting the overlapping area of ​​the magnetotelluric signal to be denoised into the uniform segmentation, and adopting a segmentation method with a coverage area ratio of 50%; and classifying the segments of the input model into two ends, with the first 50% being the prior segments and the latter 50% being the unprocessed segments referred to as noise segments.

[0013] Optionally, the ensemble learning step includes: using a single deep learning model to backtrack the noise segment point by point in time and fitting the noise profile of the noise segment until we scan the entire noise segment and obtain the same number of segments as the noise segment.

[0014] Optionally, the noise threshold is an adaptive threshold used to clear the approximately zero fitting error in the noise contour fitted by the deep learning model, and its formula is expressed as: T = 10 <M(A-B)> Where T represents the noise threshold, <·> represents the order of magnitude of the data after rounding up, M(·) represents the global average value, A is the original MT signal, and B represents the noise profile fitted by the model.

[0015] Optionally, the three-layer DBSCAN step includes: the three-layer DBSCAN scans time scales of 50, 25, and 1 respectively to identify noise profiles of different time lengths, and determines whether noise exists by scanning the noise profiles of three time regions from global to local.

[0016] Optionally, the network structure of the restoration decision model is the same as that of the model for fitting the noise profile, except that the input and output are changed. The input is a noise segment with spikes after removing the noise profile, and the output is the corresponding smoothed magnetotelluric signal.

[0017] On the other hand, the present invention provides a magnetotelluric denoising device based on enhanced deep learning, the device comprising:

[0018] The sample library construction module is used to build a noisy signal sample library and a noise contour sample library;

[0019] The model training module is used to train signal-to-noise mapping models of different deep learning networks based on the magnetotelluric signal samples of the noisy signal sample library and the samples of the noise contour sample library. The input of the signal-to-noise mapping model is the magnetotelluric signal sample, and the output is the noise contour sample. The different deep learning networks are deep learning network models designed based on different applications.

[0020] The denoising and restoration module is used to input the magnetotelluric signal to be denoised into the signal-to-noise mapping model to learn the noise profile, and to obtain the denoised magnetotelluric signal through a lightweight ensemble learning method. The lightweight ensemble learning method consists of pipeline segmentation, ensemble learning, noise thresholding, three-layer DBSCAN, and restoration decision model.

[0021] The beneficial effects of this invention are as follows:

[0022] (1) Fully utilize the advantages of ensemble learning to improve the denoising accuracy of the model, and more importantly, to enhance the denoising effect. Lightweight ensemble learning is proposed, effectively saving the resource cost of training multiple models. Compared to traditional ensemble learning, while training multiple models can effectively improve the accuracy of fitting noise contours, the training and optimization of multiple models often wastes a lot of time. Lightweight ensemble learning, however, achieves the ensemble effect of multiple models using a single model, fully utilizing the model's capabilities. Then, facing multiple noise contours fitted by a single model, we use a noise threshold to identify non-noise segments and set them to zero in the noise contours, further refining the noise contours and avoiding the impact of fitting errors. Next, DBSCAN is used to discriminate the noise contours of the finally fitted noise segments. We only process the noise segments, retaining the non-noise segments, further protecting the effective signal. Finally, the initial denoised segment with the noise contours subtracted is input into the restoration decision model to further restore the signal's inherent trend, obtaining the denoised magnetotelluric signal.

[0023] (2) Ensemble learning was chosen to further enhance the model's performance. This method aims to improve the ability of a single model to extract noise contours. The main goal of this ensemble learning approach is to address the problem of limited fitting ability of a single model that performs well in different aspects after training. The basic concept behind ensemble learning is that even if a single model makes a wrong prediction, other individual models can correct the error. However, training multiple models in an ensemble is very time-consuming, making it challenging to meet the timeliness requirements of data collection and processing in practical applications. Lightweight ensemble learning is combined with a pipeline segmentation method. Although the input and output of a single model are one-to-one, changes in the input will also lead to adjustments in the output. In this method, we use a single model to continuously backtrack to past moments in the noise segment and fit the noise contour of the noise segment until we scan the entire noise segment. Therefore, multiple noise distributions are obtained. Subsequently, we average the noise contours at the corresponding time points to finally fit the noise contour. This effectively improves the noise contour fitting ability of a single model.

[0024] (3) Pipeline segmentation is introduced into magnetotelluric signal denoising technology, replacing the traditional uniform segmentation. When dealing with noise periods longer than the segment length, the uniform segmentation method may fail to detect the noise. Pipeline segmentation adds the overlapping area of ​​the previous segment to the uniform segmentation and adjusts the proportion of the covered area to achieve a pipeline-like segmentation method, which effectively ensures the continuity in the time domain and avoids the problem of noise not being identified.

[0025] (4) DBSCAN is chosen to determine the noise profile within the segment in order to reduce interference to non-noise areas. Current methods for noise identification using deep learning are a form of supervised learning, requiring human intervention to determine the noise profile, which is highly subjective. In contrast, Density-Based Noise Applied Spatial Clustering (DBSCAN) is an unsupervised machine learning clustering algorithm that learns statistical patterns or inherent structures from unlabeled data. DBSCAN excels at identifying irregularly shaped and differently dense clusters in MT noise, making it robust to outliers. Unlike traditional K-Means clustering algorithms, DBSCAN does not require a predetermined number of clusters, enabling it to divide high-density data into smaller clusters, thereby enhancing the identification of both MT noise and valid signals.

[0026] (5) The restoration decision model is introduced into the magnetotelluric signal denoising technology. The decision model is a concept of ensemble learning, which makes the final decision on the effects obtained by multiple models. After denoising the noise segment, in order to smooth the spikes after denoising the noise segment, we also introduced the restoration decision model. This model takes the noise segment filtered out by the lightweight ensemble learning method and the original noise-free segment as input and output, and trains them to further refine the denoising of MT data in the noise region. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0028] Figure 1 This is a flowchart illustrating a magnetotelluric denoising method based on enhanced deep learning provided in an embodiment of the present invention.

[0029] Figure 2 A flowchart of a magnetotelluric denoising method based on enhanced deep learning provided in an embodiment of the present invention;

[0030] Figure 3 This is a structural diagram of the lightweight ensemble learning method proposed in this invention;

[0031] Figure 4 The apparent resistivity-phase curve changes of the synthesized data of this invention before and after noise suppression;

[0032] Figure 5 The apparent resistivity-phase curve changes before and after noise suppression are measured according to the present invention.

[0033] Figure 6This is a schematic diagram of a magnetotelluric denoising device based on enhanced deep learning provided in an embodiment of the present invention. Detailed Implementation

[0034] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0035] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0036] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] In recent years, many studies have focused on finding superior deep learning models to optimize denoising performance, which makes evaluating the effectiveness of these models challenging. To address this issue, it is necessary to develop a magnetotelluric denoising method and device based on augmented deep learning.

[0038] This invention designs a general method to enhance the performance of various deep learning models in denoised magnetotelluric data, and provides a lightweight ensemble learning method to improve the problem of insufficient fitting and denoising ability of individual deep learning models.

[0039] This invention provides a magnetotelluric denoising method based on enhanced deep learning, as follows: Figure 1-5 As shown, it includes the following steps:

[0040] In step S1, a noisy signal sample library and a noise contour sample library are constructed.

[0041] In this embodiment of the invention, by analyzing the noise types of magnetotelluric data, a set of noise profile sample sub-libraries with the same segment length (each segment being 100), noise amplitude ranging from -10000 to 10000, and different noise widths (triangular waves, charging / discharging triangular waves, pulse waves, and square waves) are constructed using different mathematical functions. For each type of noise, the temporal position of the noise profile in the corresponding noise profile sample sub-library is changed sequentially to obtain several noise profile samples with the same length but different waveform positions. The sample amplitudes of all noise profile sample libraries for each type of noise are changed sequentially to obtain noise profile samples of the same noise with different amplitudes. Gaussian white noise of the same length as the noise profile sample library and with undulating terrain fluctuations that are basically similar to the measured magnetotelluric interference-free data are selected, along with segments of the same length without obvious noise in the measured data selected using DBSCAN, to construct a clean signal sample library. The noise profile samples and the noise segments of the clean signal sample library are added one-to-one to obtain a noisy signal sample library.

[0042] In step S2, signal-to-noise mapping models for different deep learning networks are trained based on the magnetotelluric signal samples from the noisy signal sample library and the samples from the noise contour sample library.

[0043] In this embodiment of the invention, the input of the signal-to-noise mapping model is a magnetotelluric signal sample, and the output is a noise contour sample; different deep learning networks are deep learning network models designed based on different applications.

[0044] In step S3, the magnetotelluric signal to be denoised is input into the signal-to-noise mapping model to learn the noise profile, and the denoised magnetotelluric signal is obtained through a lightweight ensemble learning method. The lightweight ensemble learning method consists of pipeline segmentation, ensemble learning, noise thresholding, three-layer DBSCAN, and restoration decision model.

[0045] In this embodiment of the invention, the pipeline segmentation is combined with the ensemble learning to fit the temporal noise profile. The overlapping area of ​​the magnetotelluric signal to be denoised is substituted into the uniform segmentation, and a segmentation method with a coverage area ratio of 50% is adopted. The segments of the input model are classified into two ends: the first 50% are prior segments, and the latter 50% of unprocessed segments are called noise segments. When processing noise periods longer than the segmentation length, the uniform segmentation method may not be able to detect the noise. However, the pipeline segmentation adds the overlapping area of ​​the previous segment to the uniform segmentation and adjusts the coverage area ratio to achieve a pipeline-like segmentation method, which effectively ensures the continuity in the time domain and avoids the problem of noise not being identified. The steps of the ensemble learning include: using a single deep learning model to backtrack the noise segment point by point in time and fitting the noise contour of the noise segment until we scan the entire noise segment and obtain the same number of segments as the noise segment. The ensemble learning method aims to solve the problem of the limited fitting ability of a single model by having a single model learn the relationship between the noisy signal and the noise contour. However, after a large number of model trainings, sometimes we can only obtain multiple biased models (weakly supervised models), and different models perform better in different aspects. We call these individual models weak learners, and the effect of ensemble multiple models is called a strong learner. The core idea of ​​ensemble learning is to combine multiple models in order to obtain a better and more comprehensive one. Strongly supervised models, where one weak classifier makes a wrong prediction, can be corrected by other weak classifiers. However, training multiple weak learners in ensemble learning requires a significant amount of time, making it difficult to meet the needs of timely data acquisition and processing applications in MT (Mean Transmission Model) applications. Lightweight ensemble learning combined with pipelined segmentation methods addresses this. Although the input and output of a single model are one-to-one, changes in the input will lead to adjustments in the output. In this method, we use a single model to continuously backtrack to past moments in the noise segment and fit the noise profile of the noise segment until we scan the entire noise segment. We then make an average decision for each time point of the obtained different noise profiles to fit the final noise profile, effectively improving the noise profile fitting ability of a single model.

[0046] The noise threshold eliminates fitting errors in the noise contours at each fitting time. Through a lightweight ensemble learning denoising method, we effectively fit the noise contours on the noisy segments. However, deep learning models always have errors; the fitted noise contours fluctuate around moments due to excessive noise, and these fluctuations are not due to the noise contours themselves but rather to incomplete fitting. Although this error may be small, removing this fitted noise contour from the noise segment will distort the effective signal at noise-free moments. To preserve the effective signal as much as possible, we only process moments containing noise, thus proposing a noise thresholding method, similar to wavelet thresholding. Its introduction aims to reduce the influence of fitting non-noise contours by zeroing out poorly fitted areas. Before training the deep learning model, we perform maximum value normalization on the input and labels, normalizing the output noise contour amplitude to [-1, 1]. This not only improves training effectiveness but also fixes the output within a range, facilitating threshold selection. The formula is: T = 10 <M(A-B)> Where T represents the noise threshold, <·> represents the order of magnitude of the data after rounding up, M(·) represents the global average value, A is the original MT signal, and B represents the noise profile fitted by the model.

[0047] The obtained noise profiles are averaged at their corresponding time points to obtain the final noise profile. A three-layer DBSCAN is used to determine if noise exists in the segment; if no noise is detected, no processing is performed; if noise is detected, the noise profile is subtracted from the magnetotelluric signal to be denoised to process the noise segment. Current methods for noise identification using deep learning are a form of supervised learning, requiring human determination of the noise profile, which is highly subjective. In contrast, Density-Based Noise Applied Spatial Clustering (DBSCAN) is an unsupervised machine learning clustering algorithm that learns statistical patterns or inherent structures from unlabeled data. DBSCAN excels at identifying irregularly shaped and differently dense clusters in MT noise, making it robust to outliers. Unlike traditional K-Means clustering algorithms, DBSCAN does not require a predetermined number of clusters, enabling it to divide high-density data into smaller clusters, thereby enhancing the identification of MT noise and valid signals. The two parameters of the DBSCAN algorithm, the density threshold Minimum Points (MinPts) and the radius epsilon (ε), play a crucial role in MT data denoising. In MT data denoising, the density threshold MinPts specifies the minimum number of data points in the ε-neighborhood required for a core point. The setting of MinPts affects the definition of core points and the formation of clusters. When the MinPts value is large, more data points are needed to meet the core point condition, which means that the cluster density requirement is higher, potentially leading to smaller clusters being merged or considered noise. Conversely, when the MinPts value is small, it is easier to form more clusters, but noise points may also be incorrectly classified as part of a cluster. For time series, only the two consecutive time points are adjacent to the current time point. Therefore, in this study, we fixed MinPts at 3, indicating that the current time point is only related to the two consecutive time points. The radius ε determines the neighborhood range of a data point. Specifically, for a given valid data point, its ε-neighborhood includes all data points within ε of that point, forming a cluster. In MT data, if ε is set too large, it will cause a large number of noise points to be misclassified as valid data points. Therefore, the selection of the radius ε in DBSCAN is easily affected by undulating waveforms, leading to misjudgments. To address this, we first use model fitting to obtain the noise profile that removes undulating terrain, effectively improving the accuracy of DBSCAN in identifying noise. In MT time series, the proportion of effective signals is often much larger than these human noise points. Furthermore, analysis of the acquired MT signals shows that the amplitude of noise points is usually much larger than the effective signal itself, often exceeding ten times or more. From these two points, we can see that noise points are characterized by their small number and large amplitude. We use a data arrangement and integration method to organize all data points in the MT data, using the amplitude and frequency of each point as criteria to adaptively determine the selection of the radius ε in different situations.Under complex human interference, the generated noise signals are random and diverse. The sparsity of human noise effectively illustrates the morphological characteristics of noise in time series. A single-layer DESCAN is insufficient to handle different types of noise. Therefore, we propose a three-layer DESCAN structure. Each layer scans at different time scales to identify noise frames at different time lengths. First, we perform a global scan of the time series to determine if large-area noise segments exist. Then, we scan sequentially from local to detailed areas. This effectively scans the time series from the global to the local, improving the efficiency of noise identification.

[0048] The starting and ending points of noise detection by the three-layer DBSCAN were recorded until no more noise was detected. The data period was smoothed using a restoration decision model to obtain the denoised magnetotelluric signal. Ensemble learning generally recognizes that the results of the strong model's decisions are still inaccurate, meaning that errors in noise profile estimation may lead to the loss of the original signal trend and some effective signals. Therefore, to better improve and smooth the signal at noisy locations, a restoration decision model was also introduced in the ensemble learning. It was trained using the noisy segment filtered by the strong model and the original, noise-free segment as input and output to further smooth the MT signal at the denoised locations.

[0049] like Figure 4 as well as Figure 5 The figures show comparisons of apparent resistivity-phase diagrams before and after noise suppression for synthetic data and for measured data, respectively. Figure 4 It can be seen that although the single model effectively removes most of the artificial noise, the apparent resistivity curve shows that the ρ in the mid-frequency range... xy and ρ yx Problems still existed, but by using our method to enhance the model's denoising capabilities, noise was further removed, effectively restoring the original apparent resistivity curve. From Figure 5 It is also evident that lightweight ensemble learning optimizes the denoising performance of a single model, better handles mid-frequency noise, and effectively protects the valid low-frequency signal from being destroyed. The denoising method provided by this invention effectively enhances the denoising capability of deep learning network models. It has broad application prospects for magnetotelluric signal data processing.

[0050] This invention fully leverages the advantages of ensemble learning to improve the denoising accuracy of the model, and more importantly, to enhance the denoising effect. It employs a lightweight approach to ensemble learning, proposing the concept of lightweight ensemble learning, which effectively saves the resource costs of training multiple models. Compared to traditional ensemble learning, while training multiple models can effectively improve the accuracy of fitting noise contours, the training and optimization of multiple models often wastes a lot of time. The proposed lightweight ensemble learning achieves the ensemble effect of multiple models using a single model, fully utilizing the capabilities of the models. Then, faced with multiple noise contours fitted by a single model, we use a noise threshold to identify non-noise segments and set them to zero in the noise contours, further refining the noise contours and avoiding the impact of fitting errors. Next, DBSCAN is used to discriminate the noise contours of the finally fitted noise segments. We only process the noise segments, retaining the non-noise segments, further protecting the effective signal. Finally, the preliminarily denoised segments, after subtracting the noise contours, are input into the restoration decision model to further restore the signal's inherent trend, obtaining the denoised magnetotelluric signal.

[0051] See Figure 6 , Figure 6 A magnetotelluric denoising device based on enhanced deep learning is provided. The device includes:

[0052] The sample library construction module 21 is used to construct a noisy signal sample library and a noise contour sample library.

[0053] The model training module 22 is used to train signal-to-noise mapping models of different deep learning networks based on the magnetotelluric signal samples of the noisy signal sample library and the samples of the noise contour sample library. The input of the signal-to-noise mapping model is the magnetotelluric signal sample, and the output is the noise contour sample. The different deep learning networks are deep learning network models designed based on different applications.

[0054] The denoising and restoration module 23 is used to input the magnetotelluric signal to be denoised into the signal-to-noise mapping model to learn the noise profile, and to obtain the denoised magnetotelluric signal through a lightweight ensemble learning method. The lightweight ensemble learning method consists of pipeline segmentation, ensemble learning, noise thresholding, three-layer DBSCAN, and restoration decision model.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions 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, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A magnetotelluric denoising method based on augmented deep learning, characterized in that, The method includes the following steps: S1: Construct a noisy signal sample library and a noise contour sample library; S2: Based on the magnetotelluric signal samples in the noisy signal sample library and the samples in the noise contour sample library, signal-to-noise mapping models of different deep learning networks are trained. The input of the signal-to-noise mapping model is the magnetotelluric signal sample, and the output is the noise contour sample. Different deep learning networks are deep learning network models designed based on different applications. S3: Input the magnetotelluric signal to be denoised into the signal-to-noise mapping model to learn the noise profile, and obtain the denoised magnetotelluric signal through a lightweight ensemble learning method. The lightweight ensemble learning method consists of pipeline segmentation, ensemble learning, noise threshold, three-layer DBSCAN, and restoration decision model. The steps for constructing the noise contour sample library include: Using different mathematical functions, noise profile sample sub-libraries of various noise types with the same segment length but different noise widths are constructed. The noise types include any combination of triangular waves, charging and discharging triangular waves, pulse waves, and square waves. By successively changing the temporal position of the noise contour sample sub-library corresponding to various types of noise, several noise contour sample word libraries with the same length but different noise waveform positions are obtained. By sequentially changing the amplitude of all noise profile samples of each type of noise in the noise profile sample sub-library, a noise profile sample sub-library with different amplitudes of the same noise is obtained. The noise profile sample library is composed of the sub-libraries of noise profile samples corresponding to various types of noise. The steps for constructing the noisy signal sample library include: A clean signal sample library is constructed by selecting Gaussian white noise of the same length and amplitude as the noise profile sample library and the measured magnetotelluric undulation terrain fluctuations, and by selecting several segments without obvious interference from the measured magnetotelluric data through DBSCAN. The noisy signal sample library is obtained by adding the latter half of the pure signal sample library and the randomly selected segments to the noise contour sample library with the same length as the latter half of the pure signal sample library. The steps of the lightweight ensemble learning method include: The pipeline segmentation is combined with the ensemble learning to fit a temporal noise profile; The fitting error occurring in the noise profile at each fitting time is eliminated by using the noise threshold; All the obtained noise profiles are averaged at the corresponding time points to obtain the final noise profile. The presence of noise in the segment is determined using a three-layer DBSCAN. If no noise is detected, no processing is performed. If noise is detected, the noise segment is processed by subtracting the noise profile from the magnetotelluric signal to be denoised. Record the start and end points when the three-layer DBSCAN detects noise until no more noise is detected. Smooth the data period by restoring the decision model to obtain the denoised magnetotelluric signal. The pipeline segmentation steps include: substituting the overlapping area of ​​the magnetotelluric signal to be denoised into the uniform segmentation, and using a segmentation method with a coverage area ratio of 50%; and classifying the segments of the input model into two ends, with the first 50% being the prior segment and the latter 50% being the unprocessed segment referred to as the noise segment. The steps of the ensemble learning include: using a single deep learning model to backtrack the noise segment point by point in time and fitting the noise profile of the noise segment until we scan the entire noise segment and obtain the same number of segments as the noise segment; The steps of the three-layer DBSCAN include: scanning the three-layer DBSCAN at time scales of 50, 25, and 1 respectively to identify noise profiles of different time lengths, and determining whether noise exists by scanning the noise profiles of three time regions from global to local.

2. The method according to claim 1, characterized in that, The noise threshold is an adaptive threshold used to clear the approximately zero fitting error in the noise profile fitted by the deep learning model, and its formula is expressed as: Where T represents the noise threshold. This represents the order of magnitude of the data after rounding up. A represents the global average value, A is the original MT signal, and B represents the noise profile fitted by the model.

3. The method according to claim 1, characterized in that, The network structure of the restoration decision model is the same as that of the model for fitting noise contours, except that the input and output have been changed. The input is a noise segment with spikes after removing the noise contour, and the output is the corresponding smoothed magnetotelluric signal.

4. A magnetotelluric denoising device based on enhanced deep learning, characterized in that, The device includes: A sample library construction module is used to construct a noisy signal sample library and a noise contour sample library. The steps for constructing the noise contour sample library include: constructing sub-libraries of noise contour samples for multiple noise types with the same segment length but different noise widths using different mathematical functions, wherein the noise types include any combination of triangular waves, charging / discharging triangular waves, pulse waves, and square waves; sequentially changing the temporal position of the noise contour samples in the sub-libraries corresponding to each type of noise to obtain several noise contour sample libraries with the same length but different noise waveform positions; and sequentially changing the temporal position of all noise contour samples for each type of noise in the sub-libraries. Amplitude values ​​are used to obtain noise profile sample sub-libraries of the same type of noise with different amplitude values; the noise profile sample sub-libraries corresponding to various types of noise are combined to form the noise profile sample library; the steps of constructing the noisy signal sample library include: selecting Gaussian white noise of the same length as the noise profile sample library and whose amplitude is close to the measured magnetotelluric undulation terrain fluctuations, and selecting several segments without obvious interference from the measured magnetotelluric data through DBSCAN, thereby constructing a clean signal sample library; the latter half of the clean signal sample library and the randomly selected segments are added one-to-one with the noise profile sample library of the same length as the latter half of the clean signal sample library to obtain the noisy signal sample library; The model training module is used to train signal-to-noise mapping models of different deep learning networks based on the magnetotelluric signal samples of the noisy signal sample library and the samples of the noise contour sample library. The input of the signal-to-noise mapping model is the magnetotelluric signal sample, and the output is the noise contour sample. The different deep learning networks are deep learning network models designed based on different applications. The denoising and restoration module is used to input the magnetotelluric signal to be denoised into the signal-to-noise mapping model to learn the noise profile, and obtain the denoised magnetotelluric signal through a lightweight ensemble learning method. The lightweight ensemble learning method consists of pipeline segmentation, ensemble learning, noise thresholding, a three-layer DBSCAN, and a restoration decision model. The steps of the lightweight ensemble learning method include: combining pipeline segmentation with ensemble learning to fit a temporal noise profile; eliminating the fitting error in each fitted temporal noise profile using the noise threshold; averaging all obtained noise profiles at corresponding time points to obtain the final noise profile; using a three-layer DBSCAN to determine if noise exists in the segment; if no noise is detected, no processing is performed; if noise is detected, the noise profile is subtracted from the magnetotelluric signal to be denoised to process the noise segment; and recording the start and end points when the three-layer DBSCAN detects noise until no more noise is detected. The noise is smoothed by restoring the decision model to obtain the denoised magnetotelluric signal. The pipeline segmentation step includes: substituting the overlapping area of ​​the magnetotelluric signal to be denoised into the uniform segmentation, and using a segmentation method with a coverage ratio of 50%; classifying the segments of the input model into two ends, the first 50% as prior segments, and the latter 50% of unprocessed segments as noise segments; the ensemble learning step includes: using a single deep learning model to backtrack the noise segments one time point at a time and fitting the noise profile of the noise segments until we scan the entire noise segment and obtain the same number of segments as the noise segments; the three-layer DBSCAN step includes: the three-layer DBSCAN scans the time scales of 50, 25, and 1 respectively to identify the noise profiles of different time lengths, and by scanning the noise profiles of the three time regions from global to local, it is determined whether noise exists.

Images