A method and system for denoising magnetotelluric signals based on a cross-scale connection network

Abstract

The present invention discloses a method and system for denoising a magnetotelluric signal based on a cross-scale connection network, comprising: using a noisy signal sample library and a noise profile sample library to train a signal-to-noise mapping model based on a cross-scale connection network model, the dense connection network comprises a feature extraction module based on a lightweight dense connection network backbone, a cross-scale connection layer and a noise feature fusion module connected in sequence; inputting the measured noisy magnetotelluric signal into the signal-to-noise mapping model to obtain a corresponding noise profile, and subtracting the noise profile from the measured noisy magnetotelluric signal to obtain a denoised magnetotelluric signal. The present invention optimizes the network architecture and introduces a cross-scale connection network for denoising, and further introduces sliding multi-scale distribution entropy for signal-to-noise identification, which can more effectively guide the network to accurately suppress strong interference in magnetotelluric signals, retain the slowly changing information of the low-frequency band of the signal, achieve good denoising effect, and can be applied to other signal noise reduction fields.

Description

Technical Field

[0001] The present invention belongs to the technical field of magnetotelluric signal processing, and in particular relates to a magnetotelluric signal denoising method and system based on a cross-scale connection network. Background Art

[0002] In the 1950s, Tikhon and Cagniard proposed the magnetotelluric (MT) method, which is a geophysical exploration method that uses natural alternating electromagnetic fields to study the electrical structure of the earth. This method has the advantages of low cost, simple construction, and high resolution, and has been widely used in mineral exploration and resource exploration. However, natural magnetotelluric signals are weak and have a wide bandwidth, and are easily affected by various noises. Especially in resource-rich mining areas, due to the interference of many complex field communication facilities and human noise, the collected MT data is often accompanied by very complex and extremely high-energy noise, and useful information is almost completely annihilated. This seriously affects the quality of the collected data, resulting in the inability of the collected data to truly and accurately reflect the underground electrical structure.

[0003] In order to improve the quality of magnetotelluric data, modern signal processing techniques such as Hilbert-Huang transform, mathematical morphological filtering, wavelet transform, sparse decomposition, K-SVD dictionary learning, etc. have been applied to this field, which can improve the quality of MT data to a certain extent. However, the existing methods mainly use the overall processing method to suppress the noise of the data, which often results in excessive processing of magnetotelluric data, loss of low-frequency and slowly changing information, and thus elimination of useful signals. In addition, more and more researchers are exploring the application and effects of various neural networks in magnetotelluric signal noise reduction technology, such as BP neural network, CNN-LSTM and Bi-LSTM. Therefore, studying the application and effects of different types of neural networks and further improving the noise reduction accuracy of magnetotelluric signals are still the current hot research directions in this field. The present invention is committed to exploring the noise reduction technology based on neural networks so that it can effectively improve the quality of magnetotelluric data. Summary of the invention

[0004] The purpose of the present invention is to improve the precision of magnetotelluric signal denoising, provide a magnetotelluric signal denoising method and system based on a cross-scale connection network, that is, to explore the application of the cross-scale connection network in the field of magnetotelluric signal denoising, and provide a new deep learning method to achieve magnetotelluric denoising. The core is the innovative establishment of a cross-scale connection network architecture including a feature extraction module based on a lightweight dense connection network backbone, a cross-scale connection layer, and a noise feature fusion module CSLPP (Cross Stage Layer Pyramid Pooling). Among them, the feature extraction module is used to extract rich and deep magnetotelluric signal features; the cross-scale connection layer is used to avoid the network from reusing a large amount of repeated gradient information, and the redundant gradient information in the network is optimized; the noise feature fusion module is used to fuse the noise features of different scales learned by the network to achieve the fusion of local features and global features; furthermore, the technical solution of the present invention has greatly improved efficiency and performance compared with traditional CNN, ResNet, and densely connected networks, and can extract low-frequency and slowly changing trends from electromagnetic data containing complex magnetotelluric noise and large-scale strong interference. At the same time, the denoising performance and generalization ability are also better than other networks, which effectively improves the network's magnetotelluric signal noise reduction ability.

[0005] On the one hand, the present invention provides a method for denoising a magnetotelluric signal based on a cross-scale connection network, comprising the following steps:

[0006] Step S1: construct a noise profile sample library and a noisy signal sample library;

[0007] Step S2: training a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy magnetotelluric signal, and the output is a noise profile;

[0008] The cross-scale connection network at least comprises a feature extraction module based on a lightweight dense connection network backbone, a cross-scale connection layer and a noise feature fusion module arranged in sequence, wherein the feature extraction module is used to extract noise features, and the cross-scale connection layer divides the input feature data into several branches for convolution and linear rectification operations respectively, and then superimposes the feature maps of each branch; the noise feature fusion module comprises a multi-scale two-dimensional maximum pooling layer of different window sizes, and finally fuses the feature maps of different scales through a cascade operation;

[0009] Step S3: inputting the measured noisy magnetotelluric signal into the signal-to-noise mapping model to obtain a corresponding noise contour, and subtracting the noise contour from the measured noisy magnetotelluric signal to obtain a denoised magnetotelluric signal.

[0010] The technical solution of the present invention introduces a cross-scale connection network into the magnetotelluric signal denoising technology, uses the feature extraction module of the cross-scale connection network to extract the complex / detailed features of the magnetotelluric noise, and the cross-scale connection layer of the cross-scale connection network optimizes the obtained complex feature map, and the noise feature fusion module of the cross-scale connection network fuses the optimized feature map, thereby further utilizing the feature information of the magnetotelluric noise. Among them, the cross-scale connection layer is divided into several branches, and the obtained complex feature maps are all convolved and linearly rectified, and then the feature maps of different branches are connected in series to form a new feature map; finally, the noise feature fusion layer performs several maximum pooling operations of different window sizes on the input features to extract the maximum value in the window, and finally the extracted feature values ​​are connected in series to map deeper and more complex magnetotelluric signal feature representations, and finally improve the magnetotelluric signal denoising accuracy.

[0011] Further optionally, the feature extraction module comprises dense blocks and transition layers arranged alternately, and the output of each dense block is connected to the next dense block through the transition layer;

[0012] Each dense block is provided with N1 convolution blocks and N1 series layers, the convolution blocks and the series layers are alternately arranged and densely connected through the series layers, and N1 is a positive integer;

[0013] Each series layer except the last series layer cascades the output of the previous convolution block and the feature input of the dense block, and then outputs it to the next convolution block in the dense block; the last series layer cascades the output of each convolution block and the feature input of the dense block.

[0014] like Figure 2 For example, each dense block is provided with 3 convolution blocks and 3 series layers, and the convolution blocks and the series layers are alternately arranged and densely connected through the series layers;

[0015] Among them, the output of the first convolution block is the input of the first and third series layers; the output of the second convolution block is the input of the second and third series layers respectively; the output of the third convolution block is directly connected to the third series layer; and the original feature input of the dense block is input to the first, second and third series layers respectively, and the third series layer cascades the input feature output to obtain a merged feature map, and inputs it to the next transition layer as a new feature map.

[0016] Among them, the technical solution of the present invention utilizes the dense blocks of the cross-scale connection network to preliminarily extract noise features; the transition layer of the cross-scale connection network adjusts the scale and depth of the noise feature map extracted by the dense block to facilitate the network downsampling operation; by cascading different feature maps, the feature map learned by the dense block layer of any feature extraction module can be accessed by all subsequent layers, further improving the feature extraction efficiency of the dense block.

[0017] The technical solution of the present invention uses a lightweight densely connected network backbone to extract features. The core of the densely connected network is composed of multiple dense blocks and transition layers, which serves as a time domain feature extractor for magnetotelluric signals. The feature extraction module improves the traditional densely connected network and uses the improved dense blocks to extract richer and deeper magnetotelluric signal features, that is, by introducing dense connections from the previous layer to all subsequent layers, the network can extract deeper time domain, frequency domain, and nonlinear features of magnetotelluric signals. In many complex field application scenarios, experimental results show that compared with traditional CNN, ResNet and other networks, the dense connections in the densely connected network further improve the information flow between layers, retain the features of all layers, and aggregate multiple feature maps to the last layer through jump connections. This method can effectively reduce the number of network parameters while maintaining the richness of data information, and improve the efficiency and performance of the network. At the same time, the model uses a cross-scale connection layer to optimize the extracted complex feature input and a noise feature fusion module to enable the model to process large-scale and complex data through a multi-scale noise feature fusion method, so that the network can better handle the complex noise environment and large-scale strong interference of magnetotelluric. The model can show better performance in both speed and accuracy, and can effectively extract small target interference while retaining the low-frequency trend of pure magnetotelluric signals.

[0018] Further optionally, the convolution block in the dense block is a batch normalization layer, a linear correction unit, a 1x1 convolution layer operation, a batch normalization layer, a linear correction unit, and a 2x2 convolution layer connected in sequence;

[0019] The transition layer includes a batch normalization layer and a 2x2 average pooling layer connected in sequence.

[0020] The traditional dense block includes three convolution blocks and three series layers. The convolution block consists of a batch normalization layer, a linear correction unit, and a 3x3 convolution layer. The three convolution blocks are connected together in a dense connection manner, and the input of each convolution block comes from the output of all previous layers. In the feature extraction module improved by the technical solution of the present invention, each dense block is connected by a transition layer, and the transition layer consists of a batch normalization layer and a 2x2 average pooling layer. Among them, by improving the batch normalization layer, the linear correction unit, and the 3x3 convolution layer before the batch normalization layer, the linear correction unit, and the 1x1 convolution layer operation, and replacing the original 3x3 convolution layer with a 2x2 convolution layer, and then connecting the convolution blocks in a dense connection manner in sequence to reduce the number of feature maps, thereby improving the computational efficiency of the network, according to the above-mentioned convolution kernel setting, the layer includes 128 1x1 convolution kernels and 64 2x2 convolution kernels. It should be noted that, on the basis of protecting the present network architecture, the present invention is not limited to the above-mentioned convolution kernel setting, and it can be adaptively adjusted on the basis of this network architecture according to experience and application requirements.

[0021] Further optionally, the cross-scale connection layer includes two parallel branches, the first branch is provided with a 1x1 convolution layer, a leaky rectified linear unit, a 2x2 convolution layer, and a leaky rectified linear unit in sequence, and the second branch is provided with a 2x2 convolution layer and a leaky rectified linear unit in sequence; the feature outputs of the last two branches are aggregated through an addition layer.

[0022] Further optionally, the noise feature fusion module includes a CBL layer, a maximum pooling layer and a series layer, the CBL layer is composed of a 3x1 convolution layer, a batch normalization layer, and a leaky rectified linear unit, and the output of the CBL layer is divided into four branches, the first branch is directly input to the series layer, the second branch is input to the series layer through a 1x1 maximum pooling layer, the third branch is input to the series layer through a 3x3 maximum pooling layer, and the fourth branch is input to the series layer through a 5x5 maximum pooling layer, and the feature map inputs of the four branches merge all the feature maps through a series operation.

[0023] In summary, the technical solution of the present invention extracts the complex features of the magnetotelluric noise through the feature extraction module. Compared with other networks with single-path connection structures, the model parameters are further reduced and the network training efficiency is further improved due to the improvement of the dense blocks of the traditional dense connection network. At the same time, the cross-scale connection layer and the noise feature fusion module are introduced, and the feature map learned by the cross-scale connection layer network is further optimized. The noise feature fusion module fuses the feature map, so that the network can perform more complex magnetotelluric signal-to-noise mapping to extract the complex and numerous noise types contained in the magnetotelluric signal, such as: charging and discharging triangle waves, pulses, and square wave noises. The dense connection better extracts the complex noise while retaining the low-frequency trend of the original pure signal. The cross-scale connection layer branches the extracted feature input to improve the network's generalization ability for different noises and alleviate the problem of repeated calculation of gradient information in the network. The multi-feature map fusion of the noise feature fusion module realizes the fusion of local noise features and global noise features, effectively improving the noise reduction effect of the network.

[0024] Further optionally, the execution process of step S3 is:

[0025] The measured noisy magnetotelluric signal segments that have been divided into equal lengths are classified into signal-to-noise categories using sliding multi-scale distribution entropy. The data segments classified as noise are input into the signal-to-noise mapping model based on the cross-scale connection network. The model output is the noise profile corresponding to the noise segment.

[0026] The corresponding noise contour is subtracted from the noise segment of the measured noisy magnetotelluric signal, and then the noise-free segment is spliced ​​to obtain the denoised magnetotelluric signal.

[0027] Further optionally, the process of using sliding multi-scale distribution entropy to classify the measured noisy magnetotelluric signals that have been segmented into equal lengths is as follows:

[0028] A Chebyshev distance matrix L is constructed for each data segment in the measured noisy magnetotelluric signal, wherein the Chebyshev distance between the sampling points in each data segment is calculated;

[0029] Calculate the empirical probability density function f of the Chebyshev distance matrix L by presetting the number of histogram columns q k , the histogram frequency distribution is N = [n1,n2,...,n k ,...,n q ], f k It is expressed as follows:

[0030]

[0031] In the above formula, n1, n2, n k ,n qThey represent the number of Chebyshev distance values ​​in the first, second, kth and last intervals respectively; by calculating the Chebyshev distance values ​​in each interval, the empirical probability density function f of the Chebyshev distance matrix L can be obtained. k ;

[0032] Through the empirical probability density function f k Calculate the distribution entropy Dist corresponding to each data segment in a sliding manner en , which is expressed as follows:

[0033]

[0034] Using the Dist of each h data segment of adjacent time en Calculate the sliding multi-scale distribution entropy Multidist en , the formula is as follows:

[0035]

[0036] Among them, the sliding multi-scale distribution entropy Multidist en h data segments larger than the set threshold are separated into noise segments.

[0037] The technical solution of the present invention uses sliding multi-scale distribution entropy to identify the signal-noise of the magnetotelluric signal, so that the network only processes the noise data segment, avoids over-processing the pure magnetotelluric data segment, and makes the network more focused on the processing of the noise data. Among them, the set threshold is an empirical value, and the present invention does not make a specific limitation on this.

[0038] Sliding Multiscale Distribution Entropy (SMDE) has some key differences from Multiscale Distribution Entropy (MDE) in processing data and analyzing signals. MDE is a method that considers the behavior of signals at different time scales. It first converts the original signal into signals of different scales through a coarse-graining process, and then calculates the distribution entropy of the signal at each scale. In the coarse-graining process, the original signal is divided into continuous non-overlapping segments, and the signal at each scale is obtained by calculating the average of these segments. MDE provides an entropy value for each scale of data, which can be used to analyze the complexity changes of signals at different scales. SMDE is an extension of MDE that not only considers signals at different scales, but also considers the changes of signals over time. In SMDE, instead of calculating the distribution entropy of each scale only once, the distribution entropy is calculated continuously by sliding the window over the entire signal. In this way, for each scale, a series of entropy values ​​that change over time are obtained. SMDE provides a method to monitor the complexity of signals changing over time at different scales, and can be used to identify non-stationary and transient features in signals. In summary, both MDE and SMDE are tools for analyzing signal complexity, but SMDE provides a dynamic way to capture the changes of signals over time through sliding windows. This is particularly useful when dealing with non-stationary noise in magnetotelluric data, identifying transient features of signals, or performing signal-to-noise classification. It can reveal differences in signal behavior at different time points, which may not be obtained when using MDE.

[0039] Further optionally, the error function of the cross-scale connection network training is The formula is:

[0040]

[0041] In the formula, Represents the noise profile in the noise profile sample library corresponding to the output data of the cross-scale connection network, X′ i They respectively represent the noise contour sample corresponding to the i-th input magnetotelluric signal sample and the noise contour output by the cross-scale connection network, and k is the total number of samples participating in the training.

[0042] Further optionally, the noise profile sample library is constructed in the following process:

[0043] Step S11: According to the noise contained in the actual measurement, different mathematical functions are used to construct typical noises with the same segment length as the noise profile sample sub-library; wherein the noise types include any combination of charge-discharge triangle waves (including sawtooth triangle waves, step triangle waves), pulse waves, and square waves.

[0044] Step S12: for the noise profile sample sub-libraries corresponding to various types of noise, the temporal position and amplitude of the noise profile in the noise profile sample sub-libraries are changed by translation, scaling and flipping to obtain several noise profile sample libraries with the same length but different noise shapes and waveform positions.

[0045] Among them, the formula of the charge and discharge triangle wave (including sawtooth triangle wave and step triangle wave) can be expressed as:

[0046] StepTriangle=B(e -1.3×t -e -2.2×t ),t=[0:0.01:n]

[0047] SawtoothTriangle=B×Sawtooth(t),t=[0:0.01:n]

[0048] The formula of pulse-like wave can be expressed as:

[0049] Pulse=[zeros(1,n1),B,zeros(1,n2)],n1+n2+1=n-type square wave formula can be expressed as:

[0050] Square=[zeros(1,n1),B+zeros(1,n2),zeros(1,n3)],n1+n2+n3=n

[0051] The above formula contains two parameters, B and n, which are used to control the amplitude and length of the noise profile waveform respectively. The zeros function and sawtooth function in Matlab are used in the formula to generate an array of all zeros and a triangular waveform. By generating noise profiles of different amplitudes, the amount of data in the network training sample library can be increased, so that the model can learn richer noise features, thereby further improving the generalization ability of the model. This method can generate different types of profiles similar to measured noise, such as triangular waves, square waves, and pulses, and its shape and size can be controlled by adjusting the parameters B and n, which has a certain degree of flexibility.

[0052] Step S13: sequentially changing the amplitudes of the profile samples of different types of noise to obtain quasi-measured noise sample libraries with different amplitudes; wherein the noise profile sample libraries corresponding to the various types of noise obtained are combined to form the quasi-measured noise sample library.

[0053] Further optionally, the translation process of the noise contour samples of different types of measured noise lengths in step S11 is: fixing the starting position in the noise contour with a segment length of n, randomly changing the ending position of the interference data contour, and obtaining a number of noise contour samples with a length of n and different noise widths. The scaling process is: fixing the length of the noise contour sample, scaling it by different scaling factors, and obtaining a number of type-measured noise contour samples with a length of n and different noise amplitudes. The flipping process is: fixing the length of the noise contour sample, by moving the starting position in the noise contour with a length of n to the ending position of the contour, and moving the ending position in the noise contour with a length of n to the starting position of the contour, such as proceeding backward from the starting position in sequence with a step size of 1, and proceeding forward from the ending position in sequence, so that the noise contour is flipped to obtain a new type-measured noise contour sample.

[0054] Further optionally, the process of acquiring noise profile samples at various noise waveform positions in step S11 is as follows:

[0055] By shifting the position of the noise profile sample library X of various types of noise and changing the starting position of the interference data profile in the noise profile, a sample library X1' of the diversity of the measured noise profile is obtained, which is expressed as follows:

[0056]

[0057] Where, the length of the measured noise profile sample library X1' is N, which consists of a number of noise profile samples; the first to the βth elements are composed of 0, and the β+1th to the N-β-αth elements are composed of x1 to x N-β-α The elements N-β-α+1 to N-th are composed of 0. β+α represents the number of 0s, and 0≤β+α<N.

[0058] The values ​​of β and α are changed several times in sequence to obtain several groups of noise profiles with different noise starting positions.

[0059] By scaling the sample library X of the quasi-measured noise profiles of various types of noise and changing the amplitude of the interference data profile in the noise profile, a sample library X2' of quasi-measured noise profile diversity is obtained, which is expressed as follows:

[0060]

[0061] Wherein, the length of the measured noise profile sample library X2' is N, which is composed of a number of noise profile samples, and δ is the scaling factor with a value range of 0.8 to 1.2.

[0062] By flipping the position of the measured noise profile sample library X of various types of noise and changing the position of the interference data profile in the noise profile, a diversity sample library X3' of the measured noise profile is obtained, which is expressed as follows:

[0063]

[0064] Wherein, the length of the measured noise profile sample library X3' is N, which is composed of a number of noise profile samples. Flipud is a Matlab function, which is often used to perform matrix flipping operations.

[0065] Further optionally, the process of constructing the noisy signal sample library in step S1 is:

[0066] A pure signal sample library is constructed, wherein measured magnetotelluric interference-free data with the same length as the samples of the quasi-measured noise profile sample library and an amplitude two to three orders of magnitude smaller than the noise amplitude are selected to construct the pure signal sample library; and then the samples of the pure signal sample library and the quasi-measured noise profile sample library are added one by one to obtain a noisy signal sample library.

[0067] The technical solution of the invention is based on the characteristics of measured noisy data, and divides it into two categories: interfered data and non-interfered data. On this basis, three data sets are constructed: a pure signal sample library, a quasi-measured noise profile sample library, and a noisy signal sample library. Among them, the pure signal sample library contains non-interfered signal data, the quasi-measured noise profile sample library contains quasi-measured noise profile data, and the noisy signal sample library contains interfered signal data. Through the pure signal sample library, there is a one-to-one correspondence between the data of the two sample libraries, the noise profile sample library and the noisy signal sample library. This construction method can accurately simulate the noisy signal data in the real world, and can provide more realistic training data for subsequent signal processing algorithms, while improving the reliability and generalization ability of the algorithm.

[0068] In a second aspect, the present invention provides a time domain signal denoising method based on a cross-scale connection network, comprising the following steps:

[0069] Step S1: construct a noise profile sample library and a noisy signal sample library;

[0070] Step S2: training a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy time domain signal, and the output is a noise profile;

[0071] The cross-scale connection network at least comprises a feature extraction module, a cross-scale connection layer and a noise feature fusion module arranged in sequence. The feature extraction module adopts a network backbone based on lightweight dense connection to extract noise features. The cross-scale connection layer divides the input feature data into several branches and performs convolution and linear rectification operations respectively, and then superimposes the feature maps of each branch. The noise feature fusion module comprises a multi-scale two-dimensional maximum pooling layer with different window sizes, and finally fuses the feature maps of different scales through a cascade operation.

[0072] Step S3: inputting the measured noisy time domain signal into the signal-to-noise mapping model to obtain a corresponding noise profile, and subtracting the noise profile from the measured noisy time domain signal to obtain a denoised time domain signal.

[0073] In a third aspect, the present invention provides a system based on the magnetotelluric signal denoising method, comprising:

[0074] A sample library construction module is used to construct a noisy signal sample library and a noise profile sample library;

[0075] The noise profile samples of the noise profile sample library are noise profiles corresponding to magnetotelluric signal samples in the noisy signal sample library;

[0076] A signal-to-noise mapping model construction module is used to train a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy magnetotelluric / time domain signal, and the output is a noise profile;

[0077] The cross-scale connection network comprises a feature extraction module based on a lightweight dense connection network backbone, a cross-scale connection layer and a noise feature fusion module which are sequentially connected. The feature extraction module is used to extract noise features. The cross-scale connection layer divides the input feature data into several branches and performs convolution and linear rectification operations respectively, and then superimposes the feature maps of each branch. The noise feature fusion module comprises multi-scale two-dimensional maximum pooling layers of different window sizes which are sequentially connected in series, and finally the feature maps of different scales are fused by cascading operation.

[0078] A noise contour extraction module, used for inputting the measured noisy magnetotelluric / time domain signal into the signal-to-noise mapping model to obtain the corresponding noise contour;

[0079] The reconstruction module is used to subtract the noise profile from the measured noisy magnetotelluric / time domain signal to obtain a denoised magnetotelluric / time domain signal.

[0080] In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program is called by a processor to implement: steps of a method for denoising magnetotelluric signals based on a cross-scale connection network or steps of a method for denoising time domain signals based on a cross-scale connection network.

[0081] Beneficial Effects

[0082] 1. In the basic scheme of the method for denoising magnetotelluric signals provided by the technical solution of the present invention, a cross-scale connection network including a feature extraction module based on a lightweight dense connection network backbone and a cross-scale connection layer and a noise feature fusion module is constructed, and it is successfully applied to the magnetotelluric signal denoising technology. Among them, the feature extraction module based on the lightweight dense connection network backbone can extract more complex and diverse magnetotelluric signal noise features by using dense blocks; the feature branch module is constructed by using the cross-scale connection layer, and the noise feature fusion of the multi-branch structure allows the model to learn more signal features, enhances the adaptability to complex magnetotelluric signals and optimizes the feature map; the noise feature fusion module is used to fuse multiple feature maps to realize the noise feature fusion model of local features and global features. The network structure of this model can effectively extract the noise features of magnetotelluric signals while retaining the low-frequency slow-changing trend, and fuses the feature information of multiple layers to enable the network to map more complex and deep magnetotelluric signal features, and finally realize the magnetotelluric signal denoising.

[0083] 2. In some embodiments, the preferred feature extraction module of the present invention is based on the dense block of the backbone of the improved lightweight dense connection network for feature extraction, and its network core is composed of multiple improved dense block modules. The feature extraction module connects different feature maps in series so that the feature maps learned by the dense block layer of the feature extraction module can be accessed by all subsequent layers, further improving the feature extraction efficiency of the dense block, thereby realizing the extraction of complex nonlinear features in the time domain of the magnetotelluric signal; and by improving the dense block, its structure is transformed into a batch normalization layer, a linear correction unit, a 1x1 convolution layer operation, a batch normalization layer, a linear correction unit, and a 2x2 convolution layer. Compared with the traditional dense connection network, the improved lightweight backbone dense connection network reduces the number of feature maps by introducing a 1x1 convolution layer, thereby improving the computational efficiency of the network. The cross-scale connection network has greatly improved efficiency and performance, can better handle complex magnetotelluric noise and large-scale strong interference, and has a stronger small target extraction capability, and the speed and accuracy are better than traditional CNN, ResNet and dense connection networks.

[0084] 3. In some embodiments, the present invention preferably divides the complex feature input extracted by the feature extraction module into two branches by the cross-scale connection layer, and performs convolution and linear rectification operations respectively. Compared with other single-path networks, due to its noise feature fusion strategy of multiple parallel branches, the cross-scale connection layer avoids the network from reusing a large amount of repeated gradient information and optimizes the redundant gradient information in the network.

[0085] 4. In some embodiments, the preferred noise feature fusion module of the present invention adopts a multi-scale magnetotelluric noise feature map fusion method, uses a stacked network to fuse the features, and at the same time, through the series operation of the noise feature map, the network realizes the fusion of noise features of different scales of local noise features and global noise features, thereby further retaining the noise signal characteristics, so that the network can more accurately extract the noise contour in the magnetotelluric signal.

[0086] 5. In some embodiments, the present invention preferably introduces sliding multi-scale distribution entropy to effectively distinguish between noisy data segments and pure data segments in magnetotelluric signals, thereby performing more refined processing of magnetotelluric signals, allowing the network to focus on processing noisy signals while avoiding over-processing of pure signals, further optimizing the denoising process and improving the denoising effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 A flowchart provided for an embodiment of the present invention.

[0088] Figure 2 Network structure diagram for dense block-based feature extraction.

[0089] Figure 3 This is the structural diagram of the cross-scale connection layer.

[0090] Figure 4 This is the structural diagram of the noise feature fusion module.

[0091] Figure 5 This is a diagram of the cross-scale connection network structure.

[0092] Figure 6-9 This is the noise suppression effect diagram of the simulated data, where Figure 6 Corresponding to the simulated data, Figure 7 Corresponding to the noise profile, Figure 8 Corresponding to the reconstructed signal / denoised signal; Fig. 9 It is the original pure signal.

[0093] Figure 10-11 The signal-noise separation effect of sliding multi-scale distribution entropy of measured data, Fig.10 Corresponding to the data segment separated into noise data, Fig.11 It is the data segment identified as pure signal in the measured data.

[0094] Figure 12-13 is the noise suppression effect diagram of the measured data, where Fig.12 is the measured data, Fig.13 Corresponding to the noise contour.

[0095] Fig.14 Corresponding to the reconstructed signal / denoised signal. DETAILED DESCRIPTION

[0096] The present invention provides a method and system for denoising magnetotelluric signals based on a cross-scale connection network, which makes full use of the advantages of the cross-scale connection network to improve the denoising accuracy, and further uses the sliding multi-scale distribution entropy to perform refined classification processing on magnetotelluric signals. In order to improve the denoising effect, the traditional dense connection network is optimized, and it is proposed to construct a dense block structure based on the backbone of a lightweight dense connection network, and to construct a cross-scale connection layer and a noise feature fusion module. Compared with the traditional dense connection network, the improved lightweight backbone dense connection network can greatly reduce the number of feature maps, improve the computational efficiency of the network, and extract more complex and diverse magnetotelluric signal noise features, which can better meet the requirements of magnetotelluric signal-to-noise mapping. The cross-scale connection layer adopts a multi-parallel branch noise feature fusion strategy, which avoids the network from reusing a large amount of repeated gradient information and optimizes the redundant gradient information in the network. The noise feature fusion module performs maximum pooling of different windows on the multi-scale feature map, and finally fuses the feature maps of different scales through cascade operation, so that the network realizes the noise feature fusion of local features and global features. The above network improvements enhance the adaptability of the network to the complex and diverse noise of magnetotelluric signals, improve the generalization ability of the network, and have stronger noise feature extraction capabilities when facing the complex noise environment and large-scale strong interference of magnetotelluric signals. The speed and accuracy are better than the traditional densely connected network. At the same time, the present invention also proposes to use sliding multi-scale distribution entropy to classify the signal-noise of magnetotelluric signals, retain the signal segments classified as pure data, and denoise the signal segments classified as noise, further optimizing the denoising of magnetotelluric signals.

[0097] The present invention will be further described below with reference to the embodiments.

[0098] like Figure 1 As shown, a method for denoising a magnetotelluric signal based on a cross-scale connection network provided by an embodiment of the present invention comprises the following steps:

[0099] Step 1: Construct three types of sample libraries: pure signal sample library, noise profile sample library and noisy signal sample library. First, construct the noise profile sample library and the pure signal sample library, and then use the noise profile sample library and the pure signal sample library to construct the noisy signal sample library.

[0100] This embodiment analyzes the noise type of magnetotelluric data and uses different mathematical functions to construct a group of noise profile sample sub-libraries of charging and discharging triangle waves, pulse waves, and square waves with the same segment length and different noise widths. For example, the length is 200 and the noise amplitude range is between -10000 and 10000; for the corresponding noise profile sample sub-libraries corresponding to each type of noise, the time sequence position of the noise profile in the noise profile sample sub-libraries is changed successively to obtain a number of noise profile sample sub-libraries with the same length but different noise waveform positions; for the corresponding noise profile sample sub-libraries corresponding to each type of noise, the sample amplitude of each type of noise is changed by using different scaling factors δ for the generated noise to obtain a noise profile sample sub-library with different amplitudes of the same type of noise; for the corresponding noise profile sample sub-libraries corresponding to each type of noise, the sample position of each type of noise is changed by flipping the generated noise with a Flipud function to obtain a noise profile sample sub-library with different positions of the same type of noise; finally, the obtained noise profile sample sub-libraries corresponding to each type of noise are merged to form a new noise profile sample library.

[0101] Then, the measured magnetotelluric interference-free signals with a length difference of two to three orders of magnitude from the noise profile sample library are selected, with an amplitude range of -100 to 100, to construct a pure signal sample library. Finally, the pure signal sample library and the noise profile sample library are added one by one to obtain the noisy signal sample library.

[0102] Specifically, assume that a single noise profile X1 is expressed as:

[0103] X1=[x1,x2,...,x i ,x i+1 ,...,x j ,x j+1 ,...,x N ]

[0104] Where X1 represents the noise profile data with a segment length of N, i and j represent the i-th and j-th elements in X1, respectively, and x i (1≤i<j) is the starting position of the interference data contour in the noise contour, x j (1≤i<j≤N) is the end position of the interference data contour in the noise contour.

[0105] Then fix the i-th element of the starting position of the interfering data contour in the noise contour, and space it by p elements, that is, j=i+p, j is the end position of the interfering data contour in the noise contour, and change the position of j t times in sequence to obtain t noise contours with a length of 200, that is, construct data with different noise contour widths, which can be expressed as follows:

[0106] X=[X1,X2,X3,...,X t ]=[x1,x2,x3,...,xN ]

[0107] Wherein, L represents a noise profile sample library constructed by noise profile samples with t segment lengths of 200 and different noise widths, wherein N = t × 200. Noise profile samples with different noise widths can be obtained in the above manner. It should be understood that for each type of noise, the noise profile sample sub-library of multiple types of noise with the same segment length and different noise widths can be obtained in the above method.

[0108] Then, the noise profile sample library diversity is further enriched by changing the position of the noise waveform. Specifically, the noise profile sample library X is shifted, and the starting position of the interference data profile in the noise profile is changed to obtain the noise profile diversity sample library X1', which is expressed as follows:

[0109]

[0110] Where, the length of the measured noise profile sample library X1' is N, which consists of a number of noise profile samples; the first to the βth elements are composed of 0, and the β+1th to the N-β-αth elements are composed of x1 to x N-β-α The N-β-α+1th to Nth elements are composed of 0. β+α represents the number of 0s, and 0≤β+α<N. By changing the values ​​of β and α several times in sequence, several groups of noise profiles with different noise starting positions are obtained.

[0111] The diversity of the noise profile sample library is enriched by changing the amplitude. Specifically, changing the amplitude of the sample library means multiplying the sample library X by different scaling factors δ, and selecting different constants δ to construct sample libraries X2' with different amplitudes, which can be expressed as follows:

[0112]

[0113] In the formula, δ is a constant with a value range of 0.8 to 1.2.

[0114] Finally, by flipping the position of the measured noise profile sample library X of various types of noise and changing the position of the interference data profile in the noise profile, we can obtain the measured noise profile diversity sample library X3', which is expressed as follows:

[0115]

[0116] Wherein, the length of the measured noise profile sample library X3' is N, which is composed of a number of noise profile samples. Flipud is a Matlab function, which is often used to perform matrix flipping operations.

[0117] It should be understood that in other feasible embodiments, the noise profile sample library is not constrained to be expanded in the manner of sample diversity described above, and some steps may be performed or other techniques may be used to expand the samples.

[0118] Step 2: Divide the three types of sample libraries into equal-length segments, and use the noisy signal sample library and the noise contour sample library with equal-length segments as the input and output of the cross-scale connection network for training, respectively, to obtain a signal-to-noise mapping model based on the cross-scale connection network.

[0119] like Figure 5 As shown, the cross-scale connection network of this embodiment includes at least an input terminal, a convolution layer, a feature extraction module, a maximum pooling layer, a cross-scale connection layer, a noise feature fusion module and an output terminal connected in sequence. The input terminal is used to input noisy magnetotelluric signal data, and the output terminal outputs noise profile data.

[0120] like Figure 2 The feature extraction model network architecture based on the improved lightweight dense connection network backbone provided by the present embodiment is shown. The feature extraction model provided by the present embodiment mainly includes 3 dense blocks and 2 transition layers, specifically: dense block 1, transition layer 1, dense block 2, transition layer 2, dense block 3, and maximum pooling layer connected after the convolution layer. In other feasible embodiments, the number and order of dense blocks and transition layers are not limited to the above examples, and their number can be adjusted according to the effect of the network model. The dense blocks and transition layers are alternately arranged, that is, the output of each dense block is connected to the next dense block through the transition layer.

[0121] Each dense block mainly contains 3 convolution blocks and 3 series layers, which are densely connected through the series layers; the internal structure of the convolution block of the dense block is a batch normalization layer, a linear correction unit, a 1x1 convolution layer operation, a batch normalization layer, a linear correction unit, and a 2x2 convolution layer connected in sequence. Every two dense blocks are connected in sequence through a transition layer; the structure of each transition layer is a batch normalization layer and a 2x2 average pooling layer connected in sequence, so that the network can downsample in the structure. In this embodiment, the number of convolution kernels in each dense block is 128, and the sizes of the convolution kernels are 1×1 and 2×2 respectively; in other feasible embodiments, the number of convolution kernels is set to an empirical value, which is adjusted according to the network requirements and the network model effect.

[0122] like Figure 3The figure shows the cross-scale connection layer in the cross-scale connection network provided by the embodiment of the present invention. The structure mainly includes two parallel branches. The first branch is provided with a 1x1 convolution layer, a leaky rectified linear unit, a 2x2 convolution layer, and a leaky rectified linear unit in sequence. The second branch is provided with a 2x2 convolution layer and a leaky rectified linear unit in sequence. The number of all convolution kernels is 128, of which the step length of the 1x1 convolution is 1 and the step length of the 2x2 convolution is 2. The feature outputs of the two branches are aggregated through the addition layer.

[0123] like Figure 4 The figure shows the noise feature fusion module in the cross-scale connection network provided by the embodiment of the present invention. The noise feature fusion module mainly includes a CBL (Convolution Batch LeakyReLU) layer, a maximum pooling layer, and a series layer (Concat layer). The CBL layer is composed of a 3x1 convolution layer, a batch normalization layer, and a leaky rectified linear unit. The output of CBL is divided into four branches. The first branch is directly input to the series layer, the second branch is input to the series layer through a 1x1 maximum pooling layer, the third branch is input to the series layer through a 3x3 maximum pooling layer, and the fourth branch is input to the series layer through a 5x5 maximum pooling layer. The feature map inputs of the four branches merge all the feature maps through a series operation. It should be understood that the four branches in this embodiment are the best settings for this embodiment, and adaptive adjustments can be made in other feasible embodiments.

[0124] In this embodiment, the training process of the cross-scale connection network is as follows:

[0125] Step 2.1: Divide the simulated sample library D into m segments of equal length of 200, that is, m = n / 200, m is an integer, and obtain m segments of data. The simulated sample library D′ is as follows:

[0126]

[0127] Where d1,...,d 200m It represents the mth data segment with a length of 200 in the simulated sample library D′, and is also the mth data segment with a length of 200 in D. The simulated sample library D′ comes from the noisy signal sample library.

[0128] Step 2.2: Set the feature extraction module parameters, such as convolution kernel, convolution step size, etc.

[0129] Step 2.3: Set the cross-scale connection layer parameters, such as convolution kernel, leakage slope, etc.

[0130] Step 2.4: Set the noise feature fusion module parameters, such as convolution kernel, convolution step size, pooling window size, and concatenation dimension.

[0131] Step 2.5: Set the training parameters of the cross-scale connection network: the optimizer is Adam, the initial learning rate is 0.001, the maximum number of iterations is 1000, the minimum training batch size is set to 32, the learning rate decay factor is 0.5, and the learning rate decay period is 10. It should be noted that the present invention does not restrict the specific values ​​of the parameter settings, and the execution order of steps 2.2-2.5 is not limited to the above rules, and the order can be adjusted or executed simultaneously.

[0132] Step 2.6: Input the simulated sample library D′ into the cross-scale connection network in sequence to train and extract the feature x l , the feature extraction process is expressed as follows:

[0133] x l =H l ([x0,x1,...,x l-1 ])

[0134] In the formula, x l For the first th The characteristic output of the magnetotelluric signal of the layer, where [x0,x1,...,x l-1 ] refers to the concatenation of feature maps generated in layers 0,…,l-1. For ease of implementation, H in the above formula is replaced by l (·) are connected into one tensor, H l (·) can be a composite function of the corresponding layer operation, such as batch normalization (BN), rectified linear unit (ReLU), pooling (Pool) or convolution (Conv).

[0135] For example, the error function of network training for a cross-scale connection network is J(), and the specific formula is as follows:

[0136]

[0137] In the formula, Represents the noise profile in the noise profile sample library corresponding to the output data of the cross-scale connection network, X′ i They represent the noise contour sample corresponding to the input i-th magnetotelluric signal sample and the noise contour output by the cross-scale connection network, respectively, and k is the total number of samples involved in the training. It should be understood that the above error function is not the only optional error function of the technical solution of the present invention, and the present invention has no special restrictions on the error function.

[0138] Step 3: Use sliding multi-scale distribution entropy to classify the measured noisy magnetotelluric signals that have been divided into equal length segments, and input the data segments classified as noise into the signal-to-noise mapping model based on the cross-scale connection network. The model output is the noise contour corresponding to the noise segment.

[0139] Step 3.1: Take the segment length as 200 to evenly divide the measured magnetotelluric signal E into m short time periods, that is: m = n / 200, m is an integer.

[0140] Step 3.2: Construct a Chebyshev distance matrix L for each data segment in the magnetotelluric signal matrix E′, where the Chebyshev distance between samples in each short time period is calculated, denoted as d ij , such as two sampling points e in a data segment 200i and e 200j Between ij The Chebyshev distance matrix L is composed as follows:

[0141] L = {d ij}(1≤i,i≤j≤m)

[0142] Step 3.3: Calculate the empirical probability density function f of the Chebyshev distance matrix L by presetting the number of histogram bins q k , the histogram frequency distribution is N = [n1,n2,...,n k ,...,n q ], f k It is expressed as follows:

[0143]

[0144] Step 3.4: Through f k Calculate the distribution entropy Dist in a sliding manner en , which is expressed as follows:

[0145]

[0146] Step 3.5: Use the Dist between each h vectors (data segments) en Calculate the sliding multi-scale distribution entropy Multidist en , the formula is as follows:

[0147]

[0148] Among them, the value of h is set according to the application requirements, and the length of the data segment is selected as a whole to calculate the sliding multi-scale distribution entropy.

[0149] Step 3.6: Sliding multi-scale distribution entropy Multidist en The h segments that are larger than a set threshold (such as an empirical value) are separated into noise segments, which are sequentially input into the signal-to-noise mapping model to obtain the noise profile Z.

[0150] Step 4: Subtract the corresponding noise contour from the noise segment of the measured noisy magnetotelluric signal, and then splice the noise-free segment to obtain the denoised magnetotelluric signal.

[0151] It should be noted that in this embodiment, the sliding multi-scale distribution entropy is preferably used to classify the signal-noise of the magnetotelluric signal, the signal segment classified as pure data is retained, and the signal segment classified as noise is denoised, which further optimizes the denoising of the magnetotelluric signal. In other feasible embodiments, on the basis of constructing the signal-noise mapping model, signal-noise classification is not performed, each data segment is directly regarded as a noisy signal segment, the signal-noise mapping model is input to obtain the noise profile, and then the magnetotelluric signal matrix E′ is subtracted from the noise profile matrix Z corresponding to the output of the signal-noise mapping model to obtain the denoised magnetotelluric signal matrix T, which is expressed as follows:

[0152]

[0153] In the formula, e 200m -z 200m represents the mth sampling point in the magnetotelluric signal matrix E′ minus the mth sampling point in Z, that is: t 200m =e 200m -z 200m .t 200m Represents the mth sampling point in the sample library T after denoising.

[0154] Among them, the denoised magnetotelluric signal matrix T is expanded column by column to form a denoised one-dimensional matrix T′, which is expressed as follows:

[0155] T′=[t1,t2,t3,...,t 200m ]

[0156] Where t,...,t 200m It represents the mth data segment with a length of 200 in T′, and is also the mth data segment with a length of 200 in T.

[0157] like Figure 6-9 as well as Figure 12-14 As shown in the figure, they are the noise suppression effect diagram of the simulated data and the noise suppression effect diagram of the measured data, Figure 10-11 The figure shows the effect of signal-noise separation using sliding multi-scale distribution entropy. Figure 6 The simulated data is effectively eliminated by the denoising method of the present invention. Figure 7 The noise profile shown is reconstructed to obtain Figure 8 The denoised signal shown is Fig. 9 is the original pure signal, which is used to compare the reconstruction effect. Fig.10 is the data segment identified as noise in the measured data, Fig.11 It is the data segment identified as pure signal in the measured data. Fig.12 The measured data is effectively eliminated by the denoising method of the present invention. Fig.13 The noise profile shown is reconstructed to obtain Fig.14Based on the above figures, it can be seen that the denoising method provided by the present invention effectively suppresses strong interference in the magnetotelluric signal.

[0158] In some embodiments, the above technical ideas of the present invention are also applicable to other time domain signal denoising. Therefore, the present invention also provides a time domain signal denoising method based on a cross-scale connection network, comprising the following steps:

[0159] Step S1: construct a noise profile sample library and a noisy signal sample library;

[0160] Step S2: training a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy time domain signal, and the output is a noise profile;

[0161] The cross-scale connection network at least comprises a feature extraction module, a cross-scale connection layer and a noise feature fusion module arranged in sequence. The feature extraction module adopts a network backbone based on lightweight dense connection to extract noise features. The cross-scale connection layer divides the input feature data into several branches and performs convolution and linear rectification operations respectively, and then superimposes the feature maps of each branch. The noise feature fusion module comprises a multi-scale two-dimensional maximum pooling layer with different window sizes, and finally fuses the feature maps of different scales through a cascade operation.

[0162] Step S3: inputting the measured noisy time domain signal into the signal-to-noise mapping model to obtain a corresponding noise profile, and subtracting the noise profile from the measured noisy time domain signal to obtain a denoised time domain signal.

[0163] The implementation of each specific step refers to the steps in the aforementioned embodiment.

[0164] Based on the above method, the present invention also provides a denoising system based on the magnetotelluric signal denoising method, comprising:

[0165] Sample library construction module: used to construct the pure signal sample library, noise profile sample library and noisy signal sample library of magnetotelluric / time domain signals. The pure signal sample library is a sample library without interference data, and the noise profile sample library is a noise profile sample library with interference data.

[0166] Signal-to-noise mapping model construction module: input the training data into the cross-scale connection network for training, save the model parameters after training, and obtain the signal-to-noise mapping model.

[0167] Noise contour extraction module: Use sliding multi-scale distribution entropy to separate the signal-noise of magnetotelluric / time domain noisy signals, input the noise data segment into the signal-noise mapping model of the cross-scale connection network, and output the noise contour.

[0168] Reconstruction module: used for subtracting the corresponding noise profile from the magnetotelluric / time domain signal to be denoised to obtain the denoised magnetotelluric / time domain signal.

[0169] The specific application process of each module is implemented with reference to the above description, and the combination and splitting number between modules are also applicable to the present invention, and the integration and disassembly and separation between modules can also be achieved in other ways. It should be understood that the specific implementation method described is only a preferred mode of the present invention, and any modification or replacement within the knowledge of those skilled in the art that does not depart from the purpose and scope of the present invention is also included in the protection scope of the present invention.

[0170] The technical solution of the present invention also provides a computer-readable storage medium, characterized in that: a computer program is stored, and the computer program is called by a processor to implement: the steps of a method for denoising magnetotelluric signals based on a cross-scale connection network. Specifically, steps S1 to S4 are performed.

[0171] For the specific implementation process of each step, please refer to the description of the above method.

[0172] The readable storage medium is a computer-readable storage medium, which may be an internal storage unit of the controller described in any of the foregoing embodiments, such as a hard disk or memory of the controller. The readable storage medium may also be an external storage device of the controller, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the controller. Further, the readable storage medium may also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium may also be used to temporarily store data that has been output or is to be output.

[0173] Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0174] It should be emphasized that the examples described in the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. A method for denoising magnetotelluric signals based on a cross-scale connection network, characterized in that: The following steps are involved: Step S1: construct a noise profile sample library and a noisy signal sample library; Step S2: training a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy magnetotelluric signal, and the output is a noise profile; The cross-scale connection network at least comprises a feature extraction module, a cross-scale connection layer and a noise feature fusion module arranged in sequence. The feature extraction module adopts a network backbone based on lightweight dense connection to extract noise features. The feature extraction module comprises dense blocks and transition layers arranged alternately. The output of each dense block is connected to the next dense block through the transition layer. Each dense block is provided with N1 convolution blocks and N1 series layers. The convolution blocks and the series layers are arranged alternately and densely connected through the series layers. N1 is a positive integer. The cross-scale connection layer divides the input feature data into several branches, which are respectively subjected to convolution and linear rectification operations, and then the feature maps of each branch are superimposed. The noise feature fusion module comprises a multi-scale two-dimensional maximum pooling layer with different window sizes, and finally the feature maps of different scales are fused through a cascade operation. The cross-scale connection layer includes two parallel branches, the first branch is provided with a 1x1 convolution layer, a leaky rectified linear unit, a 2x2 convolution layer, and a leaky rectified linear unit in sequence, and the second branch is provided with a 2x2 convolution layer and a leaky rectified linear unit in sequence; the feature outputs of the last two branches are aggregated through an addition layer; The noise feature fusion module includes a CBL layer, a maximum pooling layer and a series layer. The CBL layer is composed of a 3x1 convolution layer, a batch normalization layer, and a leaky rectified linear unit. The output of the CBL layer is divided into four branches. The first branch is directly input to the series layer, the second branch is input to the series layer through a 1x1 maximum pooling layer, the third branch is input to the series layer through a 3x3 maximum pooling layer, and the fourth branch is input to the series layer through a 5x5 maximum pooling layer. The feature map inputs of the four branches merge all the feature maps through a series operation. Step S3: inputting the measured noisy magnetotelluric signal into the signal-to-noise mapping model to obtain a corresponding noise profile, and subtracting the noise profile from the measured noisy magnetotelluric signal to obtain a denoised magnetotelluric signal; Among them, the sliding multi-scale distributed entropy is used to classify the signal-to-noise of the measured noisy magnetotelluric signal segments that have been divided into equal length segments, and the data segments classified as noise are input into the signal-to-noise mapping model based on the cross-scale connection network. The model output is the noise contour corresponding to the noise segment; the noise segment of the measured noisy magnetotelluric signal is subtracted from the corresponding noise contour, and then the noise-free segment is spliced ​​to obtain the denoised magnetotelluric signal.

2. The method according to claim 1, characterized in that: Each series layer in the dense block except the last series layer cascades the output of the previous convolution block and the feature input of the dense block, and then outputs it to the next convolution block in the dense block; the last series layer cascades the output of each convolution block and the feature input of the dense block.

3. The method according to claim 2, characterized in that: The convolutional block in the dense block is a batch normalization layer, a linear correction unit, a 1x1 convolutional layer operation, a batch normalization layer, a linear correction unit, and a 2x2 convolutional layer connected in sequence; The transition layer includes a batch normalization layer and a 2x2 average pooling layer connected in sequence.

4. The method according to claim 1, characterized in that: The process of using sliding multi-scale distribution entropy to classify the measured noisy magnetotelluric signal segments that have been divided into equal length segments is as follows: A Chebyshev distance matrix L is constructed for each data segment in the measured noisy magnetotelluric signal, wherein the Chebyshev distance between the sampling points in each short-time data segment is calculated; Calculate the empirical probability density function f of the Chebyshev distance matrix L by presetting the number of histogram columns q k , the histogram frequency distribution interval is N = [n1, n2, ..., n k ,...,n q ], f k It is expressed as follows: In the above formula, n1, n2, n k ,n q Respectively represent the number of Chebyshev distance values ​​in the first, second, kth and last intervals; Using the empirical probability density function f k Calculate the distribution entropy Dist corresponding to each data segment in a sliding manner en , which is expressed as follows: Using the Dist of each h data segment of adjacent time en Calculate the sliding multi-scale distribution entropy Multidist en , the formula is as follows: Among them, the sliding multi-scale distribution entropy Multidist en h data segments larger than the set threshold are separated into noise segments.

5. A time domain signal denoising method based on a cross-scale connection network, characterized in that: The following steps are involved: Step S1: construct a noise profile sample library and a noisy signal sample library; Step S2: training a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy time domain signal, and the output is a noise profile; Among them, the cross-scale connection network at least includes a feature extraction module, a cross-scale connection layer and a noise feature fusion module arranged in sequence, the feature extraction module adopts a network backbone based on lightweight dense connection to extract noise features; the feature extraction module includes dense blocks and transition layers arranged alternately, the output of each dense block is connected to the next dense block through the transition layer, each dense block is provided with N1 convolution blocks and N1 series layers, the convolution blocks and the series layers are arranged alternately and densely connected through the series layers, N1 is a positive integer; the cross-scale connection layer divides the input feature data into several branches, performs convolution and linear rectification operations respectively, and then superimposes the feature maps of each branch; the noise feature fusion module includes a multi-scale two-dimensional maximum pooling layer with different window sizes, and finally fuses the feature maps of different scales through a cascade operation; The cross-scale connection layer includes two parallel branches, the first branch is provided with a 1x1 convolution layer, a leaky rectified linear unit, a 2x2 convolution layer, and a leaky rectified linear unit in sequence, and the second branch is provided with a 2x2 convolution layer and a leaky rectified linear unit in sequence; the feature outputs of the last two branches are aggregated through an addition layer; The noise feature fusion module includes a CBL layer, a maximum pooling layer and a series layer. The CBL layer is composed of a 3x1 convolution layer, a batch normalization layer, and a leaky rectified linear unit. The output of the CBL layer is divided into four branches. The first branch is directly input to the series layer, the second branch is input to the series layer through a 1x1 maximum pooling layer, the third branch is input to the series layer through a 3x3 maximum pooling layer, and the fourth branch is input to the series layer through a 5x5 maximum pooling layer. The feature map inputs of the four branches merge all the feature maps through a series operation. Step S3: inputting the measured noisy time domain signal into the signal-to-noise mapping model to obtain a corresponding noise profile, and subtracting the noise profile from the measured noisy time domain signal to obtain a denoised time domain signal; Among them, the sliding multi-scale distributed entropy is used to classify the signal-noise of the measured noisy time domain signal that has been divided into equal length segments, and the data segments classified as noise are input into the signal-noise mapping model based on the cross-scale connection network. The model output is the noise contour corresponding to the noise segment; the noise segment of the measured noisy time domain signal is subtracted from the corresponding noise contour, and then the noise-free segment is spliced ​​to obtain the denoised time domain signal.

6. A noise reduction system based on the method according to any one of claims 1 to 5, characterized in that: Include: A sample library construction module is used to construct a noise profile sample library and a noisy signal sample library; A signal-to-noise mapping model construction module is used to train a signal-to-noise mapping model based on a cross-scale connection network using a noisy signal sample library and a noise profile sample library; wherein the input of the signal-to-noise mapping model is a noisy magnetotelluric / time domain signal, and the output is a noise profile; The cross-scale connection network comprises a feature extraction module based on a lightweight dense connection network backbone, a cross-scale connection layer and a noise feature fusion module which are sequentially connected. The feature extraction module is used to extract noise features. The feature extraction module comprises dense blocks and transition layers which are arranged alternately. The output of each dense block is connected to the next dense block through the transition layer. Each dense block is provided with N1 convolution blocks and N1 series layers. The convolution blocks and the series layers are arranged alternately and densely connected through the series layers. N1 is a positive integer. The cross-scale connection layer divides the input feature data into several branches, which are respectively subjected to convolution and linear rectification operations, and then the feature maps of each branch are superimposed. The noise feature fusion module comprises multi-scale two-dimensional maximum pooling layers of different window sizes which are sequentially connected in series. Finally, the feature maps of different scales are fused by cascading operation. A noise contour extraction module, used for inputting the measured noisy magnetotelluric / time domain signal into the signal-to-noise mapping model to obtain the corresponding noise contour; The reconstruction module is used to subtract the noise profile from the measured noisy magnetotelluric / time domain signal to obtain a denoised magnetotelluric / time domain signal.

7. A computer-readable storage medium, characterized in that: A computer program is stored, which is called by a processor to implement: The steps of the method according to any one of claims 1 to 5.

Images