Method and system for magnetotelluric signal denoising based on multi-path fusion network

Abstract

The application provides a method and system for removing noise from a magnetotelluric signal based on a multi-path fusion network, the method constructs and trains a noise recognition model and a signal mapping model based on the multi-path fusion network; thereby inputting a noisy signal segment recognized based on the noise recognition model into the signal mapping model to obtain a pure signal segment or a noise profile; and finally splicing the pure signal segments in sequence to obtain a magnetotelluric signal after noise reduction. The multi-path fusion network provided by the application is provided with three convolution paths with convolution layers and activation functions of different size convolution kernels in a multi-path convolution module, which is used for multi-scale signal feature extraction; in an attention-enhanced multi-path feature fusion module, the outputs of different paths are respectively input into a channel attention, a spatial attention and a convolution layer, and the corresponding outputs are spliced in a channel dimension, so as to realize multi-path feature fusion. In addition, the application further introduces a physical constraint to perform deep optimization on the two models.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of geological exploration, and particularly relates to a magnetotelluric signal denoising method and system based on a multi-path fusion network. BACKGROUND

[0002] Magnetotellurics (MT) is a geophysical exploration method that measures the natural electromagnetic field on the surface to detect the electrical conductivity structure of the subsurface. This method can provide information on the electrical structure of the subsurface from the shallow crust to the lithosphere and even deeper, and thus has a wide range of applications in geological research, mineral resource exploration, and geothermal resource exploration. Natural magnetotelluric signals are very weak and easily disturbed by noise. At the same time, with the development of cities, human electromagnetic noise generated during production has a great influence on the collection of magnetotelluric signals, which seriously affects the interpretation and inversion of the subsurface geological structure.

[0003] With the efforts of many scholars, many effective methods have been applied in the field of magnetotelluric signal processing. Remote Reference (RR) is one of the most widely used magnetotelluric denoising methods. The noise suppression effect of the remote reference method is greatly related to the selection of the reference point, but with the continuous acceleration of urbanization, the selection of the remote reference point has become increasingly difficult, and the limitations of the remote reference method have also been magnified. In addition to the remote reference method, the Robust estimation method is also commonly used to suppress noise, but when the noise part accounts for a large proportion of the collected data, the denoising effect of this method is significantly reduced. These methods can suppress noise to some extent and ultimately restore the original signal. However, they all have different defects, such as the inability to effectively and flexibly handle interference, leading to a large gap between the restored signal and the actual signal, the introduction of new noise affecting the denoising accuracy, etc. In recent years, deep learning methods have also been widely used in magnetotelluric signal processing. Convolutional Neural Network (CNN) was initially used to identify magnetotelluric noise. After discovering the non-linear mapping capability of neural networks in noise profiles, CNN was also used for magnetotelluric signal denoising. However, the defects of CNN are also obvious, including the use of single-scale convolution, the failure to combine noise identification and signal mapping tasks, and the separate processing from the time domain. Therefore, further exploring the potential of deep learning in this application is still a research hotspot in this field. SUMMARY

[0004] In order to improve the precision and efficiency of the magnetotelluric signal denoising, the technical scheme of the present application provides a magnetotelluric signal denoising method and system based on a multi-path fusion network. The multi-path fusion network capable of noise recognition and signal mapping tasks is used as the backbone network for feature extraction. Different size convolution kernels are used in the multi-path convolution. The large-scale convolution kernel can extract the global representation of the signal, and the small-scale convolution kernel extracts the local features of the signal. Then, the channel attention and spatial attention are used to strengthen the representation of the global features and local details of the signal, respectively. Then, the feature maps of different paths are concatenated in the channel dimension to realize multi-scale feature fusion operation, so that the network can learn different amplitude and different morphological noise types. In addition, the present application is further optimized based on the improvement, and the physical constraint (using the frequency domain characteristics and sparsity characteristics of the magnetotelluric signal) is integrated. The physical prior information of the magnetotelluric signal is used to guide the training of the network, and the output of the network is constrained, so that the network has interpretability and realizes more accurate noise recognition and more accurate signal reconstruction effect.

[0005] Therefore, the present application provides the following technical scheme:

[0006] In one aspect, the present application provides a magnetotelluric signal denoising method based on a multi-path fusion network, comprising the following steps:

[0007] Step S1: Sample library construction, constructing a noise recognition sample library and a signal mapping sample library;

[0008] Step S2: model construction and training, constructing and training a noise recognition model based on a multi-path fusion network based on the noise recognition sample library, and constructing and training a signal mapping model based on a multi-path fusion network based on the signal mapping sample library;

[0009] The multi-path fusion network comprises at least a multi-path convolution module and a multi-path feature fusion module. The multi-path convolution module is provided with three parallel convolution paths, and the convolution kernel sizes of different convolution paths are different. The multi-path feature fusion module is provided with three parallel paths corresponding to channel attention, spatial attention and convolution channel. The output results of the three paths are finally spliced in the channel dimension. The output of the path with a small convolution kernel in the multi-path convolution module corresponds to the input channel attention path. The output of the path with a large convolution kernel corresponds to the input spatial attention path. The outputs of the paths with large, medium and small convolution kernels are first spliced in the channel dimension, and then input into the convolution channel.

[0010] Step S3: using the noise recognition model to identify the segmented signal of the magnetotelluric signal to be denoised as a noisy signal segment or a pure signal segment, and inputting the identified noisy signal segment into the signal mapping model to obtain a pure signal segment or a noise profile. If it is a noise profile, the noisy signal segment is subtracted from the noise profile to obtain a pure signal segment.

[0011] Step S4: sequentially splicing the pure signal segments to obtain the noise-reduced magnetotelluric signal.

[0012] Further preferably, an initial convolution layer, batch normalization and an activation function are first provided in the multi-path convolution module; then three convolution paths are sequentially divided, and two groups of convolution layers and activation functions are alternately arranged on each convolution path, and the convolution kernel sizes of the three convolution paths are 1x1, 3x1 and 5x1 respectively.

[0013] The input of the channel attention is divided into two branches, one branch first inputs a max-pooling layer, and the other branch first inputs an average-pooling layer, and the outputs of the two branches are then subjected to a 3x1 convolution, a Swish activation function, a 3x1 convolution, and then the outputs of the two branches are superimposed and connected to a sigmoid function, and finally multiplied by the input of the channel attention;

[0014] The convolution channel is sequentially provided with a 1x1 convolution, batch normalization and a Swish activation function.

[0015] The input of the spatial attention is divided into two branches, one branch first inputs a max-pooling layer, and the other branch first inputs an average-pooling layer, and the outputs of the two branches are then concatenated in the channel dimension, and then subjected to a 7x1 convolution, a sigmoid function and a global max-pooling layer, and finally multiplied by the input of the spatial attention.

[0016] Further preferably, a dynamic gating module is provided between the multi-path convolution module and the multi-path feature fusion module in the multi-path fusion network.

[0017] The dynamic gating module is provided with a gating weight generator based on a global pooling and a fully connected layer, which outputs the weights of each convolution path in the multi-path fusion network through a Sigmoid function, and then weights the output of the corresponding convolution path.

[0018] Further preferably, a physical constraint is introduced in the noise identification model of step S2, that is, the input of the noise identification model is set to contain a frequency domain signal containing the frequency domain characteristics of the magnetotelluric signal, specifically:

[0019] First, Fourier transform is introduced to convert the time series of the magnetotelluric signal into a frequency domain signal; then the median of the amplitude spectrum is used to perform adaptive threshold filtering on the frequency domain signal to obtain a filtered frequency domain signal;

[0020] wherein the median of the amplitude spectrum is used to perform adaptive threshold filtering to obtain a filtered frequency domain signal is a threshold parameter, is an indicator function, and ​​The output is 1, that is, the data of the frequency point is reserved, otherwise the output is 0, that is, the data of the frequency point is filtered out, represents the kth complex component of the frequency domain output obtained by Fourier transform on the magnetotelluric signal segment.

[0021] Further preferably, the threshold parameter is an adjustable dynamic threshold designed for the characteristics of magnetotelluric noise, and there is:

[0022]

[0023] wherein, , is a learnable parameter automatically optimized in the noise recognition network training process, is the skewness of the amplitude spectrum , that is, the third moment; is the standard deviation of the logarithmic amplitude spectrum , and there is:

[0024]

[0025]

[0026] wherein N is the number of sampling points of the magnetotelluric signal segment, is the mean of the logarithmic amplitude spectrum , represents the mean function.

[0027] Further preferably, the output of the signal mapping model in step S2 is a pure signal segment, and a physical constraint is introduced in the signal mapping model, that is, the loss function at least incorporates a wavelet domain adaptive sparse loss , and there is:

[0028]

[0029] wherein the wavelet packet transform is performed on the residual to generate subband coefficients , represents the wavelet transform using a Symlets wavelet basis, is the true value of the i th sample, is the predicted value of the i th sample, n is the number of samples, and j is the subband index, is the subband coefficient corresponding to the j th subband index in the wavelet packet transform of the i th sample, is a subband adaptive threshold, satisfying:

[0030] wherein, is the subband noise standard deviation estimate, and N is the number of sampling points.

[0031] Further preferably, the loss function of the signal mapping model is constructed by a weighted combination of a mean square error loss and a wavelet domain adaptive sparse loss, and is expressed as:

[0032]

[0033] wherein, is the loss function of the signal mapping model, , are adjustable weight coefficients, is the mean square error loss;

[0034] The loss function of the training process of the noise identification model is a cross-entropy loss function.

[0035] In the third aspect, the technical scheme of the present application provides a system based on the noise reduction method of the magnetotelluric signal, comprising:

[0036] A sample library construction module is configured to construct a noise identification sample library and a signal mapping sample library.

[0037] A model construction and training module is configured to construct and train a noise identification model based on a multi-path fusion network based on the noise identification sample library, and to construct and train a signal mapping model based on a multi-path fusion network based on the signal mapping sample library.

[0038] The multi-path fusion network comprises at least a multi-path convolution module and a multi-path feature fusion module, the multi-path convolution module is provided with three parallel convolution paths, and the convolution kernel sizes of different convolution paths are different, the multi-path feature fusion module is correspondingly provided with three parallel paths, which are channel attention, spatial attention and convolution channel, and the output results of the three paths are finally spliced in the channel dimension, the output of the path with a small convolution kernel in the multi-path convolution module corresponds to the input channel attention path, the output of the path with a large convolution kernel corresponds to the input spatial attention path, and the output of the path with a large, medium and small convolution kernel is first spliced in the channel dimension and then input to the convolution channel;

[0039] A noise reduction module is configured to identify the segmented signal of the noise reduction magnetotelluric signal as a noisy signal segment or a pure signal segment by using the noise identification model, input the identified noisy signal segment into the signal mapping model to obtain a pure signal segment or a noise profile, subtract the noise profile from the noisy signal segment to obtain a pure signal segment if the noise profile is a noise profile, and splice the pure signal segments in sequence to obtain the noise-reduced magnetotelluric signal.

[0040] In the third aspect, the technical scheme of the present application provides a computer device, comprising:

[0041] One or more processors;

[0042] a memory storing one or more computer programs;

[0043] wherein the processor invokes the computer program to implement:

[0044] Steps of the method for denoising of magnetotelluric signals based on a multi-path fusion network.

[0045] In four aspects, the technical scheme of the present application provides a computer readable storage medium which stores a computer program, and the computer program is invoked to perform:

[0046] Steps of the method for denoising of magnetotelluric signals based on a multi-path fusion network.

[0047] Advantages

[0048] 1. The technical scheme of the present application adopts a multi-path convolution and a combination of channel attention and spatial attention to extract noise features in magnetotelluric signals, and successfully applies it to the field of magnetotelluric signal denoising. Among them, the multi-path convolution module contains three convolution paths with different size convolution kernels, and the large-scale convolution kernel can extract the global representation of the signal, and the small-scale convolution kernel extracts the local features of the signal; the feature fusion process inputs the outputs of different paths into the channel attention, spatial attention and convolution layer respectively, and splices the corresponding outputs from the channel dimension, feeds the global features extracted by the large convolution kernel into the channel attention mechanism, which uses the mutual dependence between channel mappings to improve the feature representation of specific semantics; the local features extracted by the small convolution kernel are input into the spatial attention mechanism to enhance local details and suppress irrelevant areas. The multi-path feature fusion module prevents problems such as gradient disappearance, gradient explosion and network degradation to some extent, thereby effectively capturing global and local feature information at each level of the network and realizing the fusion of multi-path features.

[0049] 2. The technical scheme of the present application is further optimized, introduces the physical constraint of frequency domain features, uses the frequency domain information of magnetotelluric signals to realize the noise recognition task, and converts the magnetotelluric signals to the frequency domain through a frequency domain feature reorganization method, solves the limitation of traditional noise recognition methods that recognize signals based on amplitude, realizes accurate recognition of noise, removes most large-scale noise while retaining the effective part of the signal without interference, reduces the over-processing of useful signals by the network, can retain more complete signal low-frequency information, and restores more real magnetotelluric interference-free signals.

[0050] 3、The technical scheme of the present application is further optimized, physical constraints of signal sparsity characteristics are introduced, that is, the characteristics of magnetotelluric signals satisfying sparsity are utilized to construct a wavelet adaptive sparse loss function for optimizing the performance of the neural network, controlling the complexity of the model, reducing the risk of network overfitting, making the network more interpretable, ensuring that the output signal has the physical meaning of the magnetotelluric signal, directly mapping the corresponding denoised signal from the noisy signal, improving the computational efficiency without causing loss to the denoising effect. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flowchart provided for the embodiments of the present application;

[0052] Figure 2 The network structure diagram of the multi-path fusion network;

[0053] Figure 3 The schematic diagram of the dynamic gating module;

[0054] Figure 4 The schematic diagram of the attention-enhanced multi-path feature fusion module;

[0055] Figure 5 The recognition effect diagram of the multi-path fusion network noise recognition model for actual measurement data;

[0056] Figure 6 The denoising effect diagram of the multi-path fusion network signal mapping model for actual measurement data;

[0057] Figure 7 The schematic diagram of the electronic element connection relationship of the computer equipment. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. The technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0059] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification is for describing the embodiments of the application only and is not intended to limit the application.

[0061] The application provides a method and system for removing noise from a magnetotelluric signal based on a multi-path fusion network, wherein a multi-path convolution and a combination of channel attention and spatial attention are used to extract noise features in the magnetotelluric signal. On this basis, physical constraints are further introduced to optimize noise recognition and signal mapping based on a wavelet adaptive sparse loss function to improve the accuracy of noise removal.

[0062] In order to extract different amplitude and different morphological features in the magnetotelluric signal, such as charging and discharging triangular waves, square waves and pulse noise, etc., the technical scheme of the application proposes a multi-path feature fusion module composed of three paths with different size convolution kernels for multi-scale signal feature extraction, wherein a large-scale convolution kernel is used to extract the global representation of the signal, and a small-scale convolution kernel is used to extract the local features of the signal. At the same time, a multi-path feature fusion module composed of channel attention and spatial attention and convolution channels in parallel is constructed, the global features extracted by the large convolution kernel are input into the channel attention mechanism to improve the representation of the features by utilizing the interdependence between the channel mappings; the local features extracted by the small convolution kernel are input into the spatial attention mechanism to enhance the local noise details and suppress the irrelevant pure signal part; the multi-path feature fusion module prevents problems such as gradient disappearance and network degradation often occurring in deep neural networks, and effectively captures global and local feature information at each level. Based on this, the trained noise recognition model and signal mapping model are applied to the measured magnetotelluric signal, the noise recognition can effectively reduce the over-processing of the useful signal by the signal mapping model, and the low-frequency trend of the signal is better preserved, thereby realizing a higher performance of magnetotelluric signal denoising effect.

[0063] The application will be further described below with reference to the embodiments.

[0064] Embodiment 1

[0065] As shown in Figure 1 The technical idea of the method for removing noise from a magnetotelluric signal based on a multi-path fusion network provided by the application is as follows:

[0066] Step S1: sample library construction, constructing a noise recognition sample library and a signal mapping sample library.

[0067] Step S2: model construction and training, constructing and training a noise recognition model based on a multi-path fusion network based on the noise recognition sample library, and constructing and training a signal mapping model based on a multi-path fusion network based on the signal mapping sample library.

[0068] Step S3: identifying the segmented signal of the to-be-noise-reduced magnetotelluric signal as a noisy signal segment or a pure signal segment by using the noise identification model, and inputting the identified noisy signal segment into the signal mapping model to obtain a pure signal segment or a noise profile, and if it is a noise profile, subtracting the noise profile from the noisy signal segment to obtain a pure signal segment.

[0069] Step S4: splicing the pure signal segments in sequence to obtain a noise-reduced magnetotelluric signal.

[0070] It should be noted that in this embodiment, the noise identification model and the signal mapping model are both multi-path fusion networks. The multi-path fusion network provided in this embodiment is preferably a multi-path fusion network with gating, i.e., sequentially provided with a multi-path convolution module, a dynamic gating module and a multi-path feature fusion module. In other feasible embodiments, the multi-path fusion module does not have a gating module, and the multi-path convolution module and the multi-path feature fusion module are also feasible and fall within the protection scope of the present application.

[0071] The multi-path convolution module is provided with three parallel convolution paths alternately composed of convolution layers and activation functions, and the convolution kernel sizes of the convolution layers of different convolution paths are different, so as to realize multi-scale feature extraction. As shown in Figure 2 In this embodiment, the structure of the multi-path convolution module is: first provided with an initial convolution layer, batch normalization and an activation function; then divided into three convolution paths in sequence, and two groups of alternately arranged convolution layers and activation functions are provided on each path, and the convolution kernel sizes are 1x1, 3x1 and 5x1 respectively. It should be understood that in other feasible embodiments, the size of the convolution kernel can be adjusted according to the effect under the premise of setting large, medium and small convolution kernels.

[0072] The dynamic gating module is provided with a gating weight generator based on global pooling and a fully connected layer, and outputs the path weights through a Sigmoid function to realize dynamic feature selection. As shown in Figure 3 In this embodiment, the dynamic gating module first splices the output of the multi-path convolution module, captures global statistics through global average pooling, and then reduces the dimension and activates through a fully connected layer, and finally generates normalized weights through a sigmoid function to weight the output of the corresponding convolution path.

[0073] The multi-path feature fusion module is provided with three parallel paths, respectively, channel attention, spatial attention and convolution channel, and the output results of the three paths are finally spliced in the channel dimension to realize the fusion of multi-path features. In the embodiment, the output of the path where the small convolution kernel is located and the weighted result of the weight are taken as the input of the channel attention path; the output of the path where the large convolution kernel is located and the weighted result of the corresponding weight are taken as the input of the spatial attention path; the output of the path where the large, medium and small convolution kernels are located and the weighted result of the corresponding weight are first spliced in the channel dimension and then taken as the input of the convolution channel. In other feasible embodiments, if there is no gating module, the output of the path where the small convolution kernel is located is directly input into the channel attention path; the output of the path where the large convolution kernel is located is directly input into the spatial attention path; the output of the path where the large, medium and small convolution kernels are located is first spliced in the channel dimension and then input into the convolution channel.

[0074] As shown in Figure 4 The input of the channel attention of the embodiment is divided into two branches, one branch is first input into the maximum pooling layer, and the other branch is first input into the average pooling layer, and the outputs of the two are respectively subjected to a 3x1 convolution, a Swish activation function and a 3x1 convolution, and then the outputs of the two branches are superimposed and connected with a sigmoid function, and finally multiplied with the input of the channel attention. The convolution channel is provided with a 1x1 convolution, a batch normalization and a Swish activation function in sequence. The input of the spatial attention is first subjected to maximum pooling and average pooling and then spliced in the channel dimension, and then subjected to a 7x1 convolution, a sigmoid function and a global maximum pooling layer, and finally multiplied with the input of the spatial attention.

[0075] Regarding step S1, the noise recognition sample library and the signal mapping sample library are both constructed based on noisy samples, noise samples and pure signal samples. Among them, the noise recognition sample library at least contains noisy samples, pure signal samples and their respective noise labels or non-noise labels, and this sample library is used to train the noise recognition model. The signal mapping sample library at least contains noisy samples and their noise profiles or pure signals, and this sample library is used to train the signal mapping model.

[0076] In step S1, the noise type is analyzed to obtain noise types including charging and discharging triangular wave, pulse, square wave noise, etc., and then corresponding noise patterns are constructed by using mathematical functions. Among them, each piece of noise length is set to 200, the noise amplitude range is between-8000 and 8000, and the noise sample library is expanded through flipping, shifting and superposition operations to improve the diversity of the finally constructed noisy signal sample library.

[0077] A piece of noise data is expressed as:

[0078]

[0079] wherein, represents a segment of noise data with segment length N, , , , respectively represent the 1st, i-th, j-th and N-th element in , and the noise part starts from and lasts to , and the rest elements are set to 0.

[0080] wherein, the flipping operation can be represented as follows:

[0081]

[0082]

[0083] wherein, represents the noise after vertical flipping, represents the noise after horizontal flipping, Flipub is a function in MATLAB for realizing horizontal flipping.

[0084] The translation operation can be represented as follows:

[0085]

[0086] wherein, is the noise after translation operation, and are adjustable parameters, by changing the size of + , the position of the noise in this segment of data can be adjusted, and the translation operation is realized.

[0087] Specifically, the superposition operation can be represented as follows:

[0088]

[0089] wherein, is the signal after superposition, and are two random segments of noise data in the sample library, and are scaling factors, usually taking values in the range of 0.5-1.5 to prevent the amplitude from being too large or too small, and by superimposing different noises, the noise types in the sample library can be enriched, and the processing capability of the network for complex noise can be enhanced.

[0090] It should be understood that in other feasible embodiments, the noise profile sample library does not have to be expanded in the above-mentioned manner of sample diversity, and some steps can be performed or other techniques can be used to expand the samples.

[0091] Regarding step S2, in some embodiments, the noise recognition model and the signal mapping model are trained in a conventional manner, i.e., the input of the noise recognition model is defined as a signal segment, and the output is the classification result of noise or pure signal; and the input of the signal mapping model is defined as a noisy signal segment, and the output is a pure signal or a noise profile, and the model training is performed in a conventional manner. In the present embodiment, it is further preferred to introduce physical constraints to optimize the noise recognition model and the signal mapping model, as follows:

[0092] Regarding the noise recognition model:

[0093] In some embodiments, the physical characteristics of the frequency domain features are fully considered, and the frequency domain features are reorganized to convert the magnetotelluric time series into frequency domain signals, i.e., Fourier transform is first introduced to convert the magnetotelluric signal time series into frequency domain signals; and then the median of the amplitude spectrum is used for adaptive threshold filtering of the frequency domain signals to obtain filtered frequency domain signals.

[0094] For example, let the magnetotelluric signal segment be , N is the number of sampling points, are the 1st, 2nd, and Nth sampling point signals in the magnetotelluric signal segment, respectively, and the frequency domain signals obtained after Fourier transform are , denote the kth complex component of the frequency domain output, and there is:

[0095]

[0096] wherein, denotes the n-th sample of the magnetotelluric signal, i is the imaginary unit, is a rotation factor, and denotes a complex exponential.

[0097] The median of the amplitude spectrum is then used for adaptive threshold filtering to obtain filtered frequency domain features , is an indicator function, which outputs 1 when , i.e., the data at that frequency point is retained, otherwise it outputs 0, i.e., the data at that frequency point is filtered out. In some embodiments, is a threshold parameter set according to the effect. In the present embodiment, an adjustable dynamic threshold parameter is further designed according to the characteristics of the magnetotelluric noise, which avoids the influence of extreme values (such as impulse noise peak values) on the threshold reference, while taking into account the balance between signal retention and noise suppression. Specifically, it satisfies:

[0098]

[0099] wherein, , a learning parameter automatically optimized in a noise identification network training process, a skewness of the amplitude spectrum , i.e., a third moment; a standard deviation of the logarithmic amplitude spectrum , there is:

[0100]

[0101]

[0102] wherein N is a number of sampling points of a magnetotelluric signal segment, a mean value of the logarithmic amplitude spectrum

[0103] In this embodiment, the loss function of the training process of the noise identification model is a cross-entropy loss function, which is represented as:

[0104]

[0105] wherein y is a real label, is a probability that the model predicts that the sample belongs to class 1, and n is a number of samples. It should be understood that in other feasible embodiments, the loss function of the noise identification model can select other loss functions, which also meet the needs of the present application.

[0106] Regarding the signal mapping model:

[0107] In some embodiments, the signal mapping model outputs a noise profile, so that the noise profile is subtracted from the noisy signal to obtain a pure signal; in other embodiments, the signal mapping model outputs a pure signal. In this embodiment, the output of the signal mapping model is a pure signal segment, and further technical optimization is performed, and a physical constraint is introduced into the signal mapping model, i.e., a loss function at least incorporates a wavelet domain adaptive sparse loss , there is:

[0108]

[0109] wherein a wavelet packet transform is performed on the residual to generate subband coefficients , denotes a wavelet transform using a Symlets wavelet basis, is a real value of the i-th sample, is a predicted value of the i-th sample, N is a number of samples, and j is a subband index, is a subband coefficient corresponding to the j-th subband index in the wavelet packet transform corresponding to the i-th sample, is a subband adaptive threshold value, satisfying: wherein j is a subband index,​ N is the number of sampling points. In this embodiment, the Sym4 wavelet packet transform is selected, and the number of wavelet decomposition layers is 4, and 16 sub-band coefficients are generated.

[0110] In this embodiment, the loss function of the signal mapping model is constructed by the weighted mean square error loss and the wavelet domain adaptive sparse loss, and is represented as:

[0111]

[0112]

[0113] wherein, is the loss function of the signal mapping model, , are adjustable weight coefficients, is the mean square error loss. It should be understood that in other feasible embodiments, the mean square error loss can be replaced by other loss functions, which can be referred to the existing technologies in the art, and the present application does not make specific limitations thereto.

[0114] Regarding the input of the identified noise-containing signal segment into the signal mapping model in steps S3 and S4 to obtain the noise profile or the pure signal segment, if it is the noise profile, the noise profile is subtracted from the noise-containing signal segment to obtain the pure signal segment, and finally the pure signal segments are spliced in sequence to obtain the de-noised magnetotelluric signal. This process is a conventional technical means in the art, and therefore, it will not be described in detail.

[0115] For example, the measured magnetotelluric signal Y is uniformly divided into k segments with a segment length of 200, and the magnetotelluric signal matrix with a segment number of k is represented as follows:

[0116]

[0117] wherein, to represent the first data segment with a length of 200 in Y, to represent the kth data segment with a length of 200 in Y, i.e., the kth data segment with a length of 200 in Y.

[0118] The signal segment identified as the pure signal is set to 0, and then the elements in the matrix are sequentially input into the signal mapping model column by column, and the network output is the matrix corresponding to the data of the pure signal , represented as follows:

[0119]

[0120] In the formula, to indicates the first data segment with a length of 200 in the middle, to indicates the kth data segment with a length of 200 in the middle, that is, the kth data segment with a length of 200 in Y corresponding to the pure signal.

[0121] Finally, the denoised signal matrix and the signal matrix identified as the pure signal are added to restore the complete denoised magnetotelluric matrix, and the denoised magnetotelluric signal X'' is obtained by expanding once by column, which can be expressed as , respectively, the first, second, and 200k signals in the magnetotelluric signal X''.

[0122] Figure 5 The noise recognition model based on the multi-path fusion network with physical constraints is simulated in an example to recognize the effect, wherein Figure 5 Fig. (a) of the noise recognition model is the time-domain waveform of the measured noisy magnetotelluric signal, Figure 5 Fig. (b) of the noise recognition model is the noisy signal segment identified by the noise recognition model, Figure 5 Fig. (c) of the noise recognition model is the pure signal segment identified by the noise recognition model. Figure 6 The signal mapping model based on the multi-path fusion network with physical constraints is simulated in an example to simulate the denoising effect, Figure 6 Fig. (a) of the noise recognition model is the time-domain waveform of the measured noisy magnetotelluric signal, Figure 6 Fig. (b) of the noise recognition model is the signal after denoising by the signal mapping model. Figure 5 and Figure 6 It can be seen that the noise recognition model provided by the technical scheme can accurately identify the noisy signal segment, and the signal mapping model can effectively suppress large-scale magnetotelluric noise.

[0123] Embodiment 2:

[0124] The embodiment also provides a system based on the magnetotelluric signal denoising method, comprising a sample library construction module, a noise recognition model construction module, a signal mapping model construction module, and a signal reconstruction module connected in sequence or with each other.

[0125] The sample library construction module is used to construct a pure signal sample library and a noisy signal sample library, and contains a noise recognition sample library and a sample library for signal mapping.

[0126] The noise recognition model construction module is used to train a multi-path fusion network using the noise recognition sample library to achieve the noise recognition effect from the signal to the corresponding label.

[0127] The signal mapping model construction module is configured to train the multi-path fusion network by using the signal mapping sample library, so as to realize the signal mapping effect from the noisy signal to the corresponding pure signal.

[0128] The signal reconstruction module is configured to input the measured magnetotelluric data into the noise identification model and the signal mapping model in sequence to obtain each piece of denoised magnetotelluric signal, and reassemble each piece of signal to restore the complete denoised magnetotelluric signal.

[0129] It should also be understood that the specific implementation process of each module can refer to the above method content, and the present application will not be repeated here, and the segmentation of the above functional modules is only for illustration. In some embodiments, part of the functional modules can be combined, part of the functional modules can be split, and each functional module can be implemented in software, hardware or a combination of software and hardware. The software and hardware devices include but are not limited to general electronic terminals, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0130] Embodiment 3:

[0131] The embodiment provides a computer device, including: one or more processors; and a memory storing a computer program; wherein the processor invokes the computer program to implement: the steps of a magnetotelluric signal denoising method based on a multi-path fusion network. Specifically, the above steps S1-S4 are executed.

[0132] The specific implementation process of each step can refer to the description of the foregoing denoising method embodiment.

[0133] In some embodiments, as shown in Figure 7 The electronic components of the computer device include:

[0134] The processor 1600 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The processor 1600 is configured to execute related programs to implement the technical solutions provided by the embodiments of the present application.

[0135] The memory 1700 can be implemented in the form of a Read Only Memory (ROM), a static storage device, a dynamic storage device, or a Random Access Memory (RAM), etc. The memory 1700 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the present specification are implemented by software or firmware, the related program codes are stored in the memory 1700 and are called and executed by the processor 1600 to execute the algorithm program of the above-mentioned denoising method of the embodiments of the present application.

[0136] The input / output interface 1800 is configured to realize information input and output.

[0137] The communication interface 1900 is configured to realize the communication interaction between the device and other devices. The communication can be realized by a wired manner (for example, a USB, a network cable, etc.) or a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.).

[0138] The bus 2000 is configured to transmit information between various components (for example, the processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900) of the device.

[0139] The processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900 are connected to each other through the bus 2000 to realize the communication connection between them in the device.

[0140] Embodiment 4:

[0141] The embodiment provides a computer readable storage medium, which stores a computer program. The computer program is called by a processor to implement the steps of a magnetotelluric signal denoising method based on a multi-path fusion network. The above-mentioned steps S1-S4 are specifically executed.

[0142] The readable storage medium is a computer readable storage medium, which can be an internal storage unit of the controller, such as a hard disk or a memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the readable storage medium can include both the internal storage unit and the external storage device of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0143] Based on such understanding, the technical solutions of the present application essentially or say the part of the prior art that contributes to the present application, or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned readable storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0144] It should be emphasized that the examples described in the present application are illustrative rather than limiting, and therefore the present application is not limited to the examples described in the specific embodiments, and any other embodiments derived by those skilled in the art from the technical solutions of the present application without departing from the purpose and scope of the present application, whether modified or replaced, also belong to the protection scope of the present application.

Claims

1. A method for denoising of magnetotelluric signals based on a multi-path fusion network, characterized in that: Comprising the following steps: Step S1: sample library construction, constructing a noise recognition sample library and a signal mapping sample library; Step S2: model construction and training, constructing and training a noise recognition model based on a multi-path fusion network based on the noise recognition sample library, and constructing and training a signal mapping model based on a multi-path fusion network based on the signal mapping sample library; Wherein, the multi-path fusion network at least includes a multi-path convolution module and a multi-path feature fusion module, the multi-path convolution module is provided with three parallel convolution paths, and the convolution kernel sizes of different convolution paths are different; the multi-path feature fusion module is correspondingly provided with three parallel paths, which are channel attention, spatial attention and convolution channel respectively, and the output results of the three paths are finally spliced in the channel dimension; the output of the path where the small convolution kernel in the multi-path convolution module is located corresponds to the input channel attention path; the output of the path where the large convolution kernel is located corresponds to the input spatial attention path; the output of the path where the large, medium and small convolution kernels are located is first spliced in the channel dimension, and then the convolution channel is correspondingly input; Step S3: using the noise recognition model to identify the segmented signal of the to-be-de-noised magnetotelluric signal as a noisy signal segment or a pure signal segment, and then inputting the identified noisy signal segment into the signal mapping model to obtain a pure signal segment or a noise profile, if it is a noise profile, then subtracting the noise profile from the noisy signal segment to obtain a pure signal segment; Step S4: sequentially splicing the pure signal segments to obtain a de-noised magnetotelluric signal.

2. The method of removing noise from a magnetotelluric signal according to claim 1, wherein: The multi-path convolution module is first provided with an initial convolution layer, batch normalization and an activation function; then it is divided into three convolution paths in sequence, and each convolution path is provided with two groups of alternately arranged convolution layers and activation functions, and the convolution kernel sizes of the three convolution paths are 1x1, 3x1 and 5x1 respectively; The input of the channel attention is divided into two branches, one branch first inputs a max pooling layer, and the other branch first inputs an average pooling layer, and the outputs of the two are then subjected to a 3x1 convolution, a Swish activation function, and a 3x1 convolution, and then the outputs of the two branches are superimposed and connected with a sigmoid function, and finally multiplied with the input of the channel attention; The convolution channel is provided with a 1x1 convolution, a batch normalization and a Swish activation function in sequence; The input of the spatial attention is divided into two branches, one branch first inputs a max pooling layer, and the other branch first inputs an average pooling layer, and the outputs of the two are concatenated in the channel dimension, and then subjected to a 7x1 convolution, a sigmoid function and a global max pooling layer, and finally multiplied with the input of the spatial attention.

3. The method of removing noise from a magnetotelluric signal according to claim 1, wherein: The multi-path fusion network is provided with a dynamic gating module between the multi-path convolution module and the multi-path feature fusion module; The dynamic gating module is provided with a gating weight generator based on global pooling and a fully connected layer, which outputs the weights of each convolution path in the multi-path fusion network through a Sigmoid function, and then weights the output of the corresponding convolution path.

4. The method of removing noise from magnetotelluric signals according to claim 1, wherein: In the noise recognition model of step S2, a physical constraint is introduced, that is, the input of the noise recognition model is set to a frequency domain signal containing the frequency domain characteristics of the magnetotelluric signal, specifically: Firstly, the Fourier transform is introduced to convert the time series of the magnetotelluric signal into a frequency domain signal; then the median of the amplitude spectrum is used to perform adaptive threshold filtering on the frequency domain signal to obtain a filtered frequency domain signal; Wherein, the median of the amplitude spectrum is used The filtered frequency domain signal is obtained by adaptive threshold filtering , is a threshold parameter, is an indicator function, when output is 1, that is, the frequency point data is reserved, otherwise output is 0, that is, the frequency point data is filtered out, represents the kth complex component of the frequency domain output obtained by Fourier transform of the magnetotelluric signal segment.

5. The method of removing noise from a magnetotelluric signal according to claim 4, wherein: Threshold parameter There is an adjustable dynamic threshold designed for the characteristics of the magnetotelluric noise. in, , These are learnable parameters that are automatically optimized during the training process of a noise recognition network. amplitude spectrum The skewness, i.e., the third moment; Logarithmic amplitude spectrum The standard deviation exists: where N is the number of sampling points of the magnetotelluric signal segment, is the mean value of the logarithmic amplitude spectrum is the mean value of the logarithmic amplitude spectrum denotes the mean function.

6. The method of removing noise from magnetotelluric signals according to claim 1, wherein: The output of the signal mapping model in step S2 is a pure signal segment. Physical constraints are introduced in the signal mapping model, i.e. the loss function at least incorporates a wavelet domain adaptive sparse loss , there is: wherein the residual is subjected to wavelet packet transform to generate subband coefficients , denotes the wavelet transform using the Symlets wavelet basis, is the true value of the i-th sample, is the predicted value of the i-th sample, n is the number of samples, and j is the subband index, is the subband coefficient corresponding to the j-th subband index in the wavelet packet transform of the i-th sample, is the subband adaptive threshold value, satisfying: wherein is the sub-band noise standard deviation estimate, N is the number of samples.

7. The method of magnetotelluric signal denoising according to claim 6, characterized in that: The loss function of the signal mapping model is constructed by the mean square error loss and the wavelet domain adaptive sparse loss, and is expressed as: wherein, is a loss function for the signal mapping model, , are adjustable weight coefficients, is a mean square error loss; The loss function of the training process of the noise recognition model is a cross-entropy loss function.

8. A system for removing noise from a magnetotelluric signal based on the method of any one of claims 1-7, characterized in that: The method comprises the following steps: a sample library construction module, configured to construct a noise recognition sample library and a signal mapping sample library; a model construction and training module, configured to construct and train a noise recognition model based on a multi-path fusion network based on the noise recognition sample library, and to construct and train a signal mapping model based on a multi-path fusion network based on the signal mapping sample library; The multi-path fusion network comprises at least a multi-path convolution module and a multi-path feature fusion module, the multi-path convolution module is provided with three parallel convolution paths, and the convolution kernel sizes of different convolution paths are different; the multi-path feature fusion module is correspondingly provided with three parallel paths, which are channel attention, spatial attention and convolution channel respectively, and the output results of the three paths are finally spliced in the channel dimension; the output of the path where the small convolution kernel is located in the multi-path convolution module corresponds to the input channel attention path; the output of the path where the large convolution kernel is located corresponds to the input spatial attention path; the output of the path where the large, medium and small convolution kernels are located is first spliced in the channel dimension, and then corresponds to the input convolution channel; a noise reduction module, configured to identify the segmented signal of the magnetotelluric signal to be de-noised as a noisy signal segment or a pure signal segment by using the noise recognition model, input the identified noisy signal segment into the signal mapping model to obtain a pure signal segment or a noise profile, if it is a noise profile, subtract the noise profile from the noisy signal segment to obtain a pure signal segment, and sequentially splice the pure signal segments to obtain the de-noised magnetotelluric signal.

9. A computer device, comprising: The method comprises the following steps: one or more processors; a memory storing one or more computer programs; wherein the processor invokes the computer program to implement: the steps of the magnetotelluric signal de-noising method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer program is stored and invoked to perform: the steps of the magnetotelluric signal de-noising method of any one of claims 1-7.

Images