A method and system for denoising controllable source electromagnetic data based on covariance-driven gating

Abstract

This invention discloses a method and system for denoising controllable source electromagnetic data based on covariance-driven gating, belonging to the field of electromagnetic data processing technology. This method introduces a covariance-driven spatiotemporal gating module to replace the gating mechanism module of the original GRU network in the temporal convolutional gating network TCN-GRU, thus proposing to input noisy controllable source electromagnetic data into the TCN-GRU for denoising. Specifically, gating coefficients are generated based on the cross-covariance matrix of the signal feature representation S and the noise feature N, enabling the network to explicitly know the "degree of signal-noise aliasing" in the current time period, achieving more refined "spatiotemporal gating" and significantly improving the denoising effect. Furthermore, it proposes to introduce Dropout-based uncertainty estimation for the final features output by TCN-GRU, using periodic statistics to quantify point-by-point uncertainty, achieving highly refined signal-to-noise separation in controllable source electromagnetic methods.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic data processing technology, and in particular to a method and system for denoising controllable source electromagnetic data based on covariance-driven gating. Background Technology

[0002] Deep-earth scientific research is transitioning from resource exploration to a three-pronged approach integrating resources, space, and cutting-edge science, urgently requiring the development of high-precision, interference-resistant detection technologies and equipment. Controlled-source electromagnetic methods (CMEs), with their controllable field sources and known excitation signals, effectively overcome the problems of random and weak magnetotelluric (MT) field sources, exhibiting strong anti-interference capabilities. However, CME signals are deterministic signals with known frequencies, while aperiodic noise is a random, broadband signal. In practical applications, CME data is often affected by persistent aperiodic noise, making it difficult to guarantee data quality and limiting its effectiveness and application prospects in deep-earth exploration.

[0003] Denoising methods for controllable source electromagnetic data based on deep learning networks have become a hot topic in this field, leading to numerous related technological applications, such as Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), and Gated Recurrent Units (GRUs). However, manually selecting network hyperparameters in deep learning can result in suboptimal data processing performance.

[0004] This invention aims to explore more feasible applications of deep learning networks in this field to improve the noise reduction performance of controllable source electromagnetic data. Summary of the Invention

[0005] The purpose of this invention is to propose a novel deep learning technique for the field of controlled-source electromagnetic signal processing, thereby improving the denoising effect of controlled-source electromagnetic data and solving the problem of low data quality in existing controlled-source electromagnetic methods, especially reducing the impact of persistent non-periodic noise on deep exploration. Specifically, this invention provides a method and system for denoising controlled-source electromagnetic data based on covariance-driven gating. It integrates a temporal convolutional network (TCN) and a gated recurrent unit network (GRU). The TCN extracts the temporal, frequency, and deep local features of the controlled-source electromagnetic data; the GRU extracts the temporal features and identifies valid signals. More importantly, a covariance-driven spatiotemporal gating module is introduced to replace the original gating mechanism module in the GRU. After introducing the spatiotemporal gating module, the temporal convolutional gating network TCN-GRU can use the cross-covariance of S and N as the gating input, enabling the network to explicitly know the "degree of signal and noise aliasing" in the current time period. The designed cross-covariance matrix contains the correlation information between multiple channels, which can achieve more refined "spatiotemporal gating" compared with the vector dot product gating of the original GRU.

[0006] Therefore, the present invention provides the following technical solution:

[0007] On one hand, the present invention provides a controllable source electromagnetic data denoising method based on covariance-driven gating, the method comprising:

[0008] A covariance-driven spatiotemporal gating module is introduced to replace the gating mechanism module in the original gated recurrent unit network GRU in the temporal convolutional gating network TCN-GRU, which includes a temporal convolutional network TCN and a gated recurrent unit network GRU connected in sequence.

[0009] The noisy controllable source electromagnetic method data is input into the temporal convolutional gated network TCN-GRU for noise reduction;

[0010] The processing of the covariance-driven spatiotemporal gating module is as follows:

[0011] First, obtain the signal feature representation S and noise feature N cross-covariance matrix of the output of the temporal convolutional network TCN;

[0012] Then, the gating coefficients are generated using the cross-covariance matrix. This replaces the update gate of the original gated recurrent unit network (GRU).

[0013] Optionally, the gating coefficient The model is:

[0014]

[0015] In the formula, Let be the gating coefficient for time step t corresponding to time segment i; It is the hidden state of the previous time step t-1, and the subscript i represents the marker of the time segment; It is the cross covariance matrix The vector obtained by expanding by rows or columns; R represents the set of real numbers, and d represents the dimension; and Learnable weight matrix; This represents the Sigmoid activation function; For bias; The input features of the current gate at time step t;

[0016] The hidden state is updated based on gated sparse updates, specifically as follows:

[0017]

[0018] In the formula, The hidden state at time step t, Let be the candidate states at time step t; where GRU uses the gating coefficients proposed in this application. And reset gate control information flow, candidate state The calculation formula is:

[0019]

[0020] In the formula, This indicates a gate reset (refer to the original gated loop unit network GRU). Represents the weight matrix. Indicates the bias term. Represents the hyperbolic tangent activation function;

[0021] Cross covariance matrix satisfy:

[0022]

[0023] In the formula, This represents the number of time sampling points in time segment i. Let S represent the mean vector of the signal feature representation S corresponding to time segment i. T represents the mean vector of the noise feature N corresponding to time segment i, and T is the matrix transpose symbol; , Let S represent the signal feature representation of time segment i and N represent the feature corresponding to time sampling point t, respectively.

[0024] Optionally, the method further includes: for the final feature Z output by the temporal convolutional gated network TCN-GRU, introducing uncertainty estimation based on Monte Carlo Dropout to enhance the denoising effect, as follows:

[0025] First, for the final feature Z, M random forward propagations are performed on the final feature Z using Monte Carlo Dropout to obtain M sets of output feature tensors. ;

[0026] Then, for each of the M feature tensors corresponding to each time sampling point, the mean and standard deviation are calculated respectively. The mean is used as the denoised signal for the corresponding time sampling point, and the standard deviation is used as the confidence level of the denoised signal for the corresponding time sampling point.

[0027] Finally, a window function and a weighted overlapping summation method are introduced to process the denoised signal, resulting in the final denoised controllable source electromagnetic method data.

[0028] Optionally, for each of the M feature tensors corresponding to each time sampling point, the mathematical model for calculating the mean and standard deviation is as follows:

[0029]

[0030]

[0031] In the formula, For characteristic tensors The feature tensor of the m-th group The feature value of the t-th time sampling point within window b. This is the denoised signal at the t-th time sampling point within window b. Let f(t) be the confidence level of the denoised signal at the t-th time sampling point within window b.

[0032] Optionally, the mathematical model for processing denoised signals by introducing a window function and a weighted overlap summation method is as follows:

[0033]

[0034]

[0035] In the formula, This is the denoised signal at the t-th time sampling point within window b after windowing modulation using a window function; Let L be the window function, and L be the sequence length of the window function. This refers to the data at the t-th time sampling point in the final controlled-source electromagnetic method data for noise reduction; m is the frame number, and R is the frame length. The denoised signal at the t-th time sampling point within window b; In window function Signal at time sampling points.

[0036] Optionally, for each of the M feature tensors corresponding to each time sampling point, the mathematical model for calculating the mean and standard deviation is as follows:

[0037]

[0038]

[0039] In the formula, For characteristic tensors The feature tensor of the m-th group The feature value of the t-th time sampling point within window b. This is the denoised signal at the t-th time sampling point within window b. Let be the standard deviation of the denoised signal at the t-th time sampling point within window b.

[0040] Optionally, when noisy controllable source electromagnetic method data is input into the temporal convolutional gated network TCN-GRU for noise reduction, the processing procedure of the temporal convolutional gated network TCN-GRU is as follows:

[0041] S1. Input layer preprocessing: The trend term of the noisy controllable source electromagnetic method data is subtracted to obtain residual time series data, and the residual time series data is cut into equal length segments to obtain time series segments.

[0042] S2. Data processing of the temporal convolutional network TCN, wherein the TCN residual blocks of the temporal convolutional network TCN are executed sequentially:

[0043] Multi-scale feature extraction involves inputting each time segment into a parallel dilated convolution branch to extract multi-scale feature maps.

[0044] Temporal weighting involves generating temporal weight vectors based on the multi-scale feature maps of time segments, and then multiplying the temporal weight vectors with the corresponding multi-scale feature maps point by point to obtain the temporal weighted features of each time segment.

[0045] The residual connection inputs the time-domain weighted features into a one-dimensional convolutional layer, outputs noise features, and then subtracts the noise features from the time-domain weighted features to obtain the signal feature representation;

[0046] S3. Data processing of the gated recurrent unit network (GRU), wherein the signal feature representation S and the noise feature N or the cross-covariance matrix of the signal feature representation S and the noise feature N output by the temporal convolutional network (TCN) are input into the gated recurrent unit network (GRU) to obtain the final feature Z, which is the output of the temporal convolutional gated network (TCN-GRU).

[0047] Optionally, the parallel dilated convolution branch is composed of three or more parallel convolutional layers with different dilation rates to extract feature maps at different time scales. Finally, the feature maps at different time scales are stitched together in the channel dimension to obtain a multi-scale feature map.

[0048] The present invention also provides a system based on the above method, comprising:

[0049] The network construction module is used to construct the temporal convolutional gated network TCN-GRU, that is, to introduce a covariance-driven spatiotemporal gated module to replace the gated mechanism module in the original gated recurrent unit network GRU in the temporal convolutional gated network TCN-GRU. The temporal convolutional gated network TCN-GRU includes a temporal convolutional network TCN and a gated recurrent unit network GRU connected in sequence.

[0050] The noise reduction module is used to input noisy controllable source electromagnetic data into the temporal convolutional gated network TCN-GRU for noise reduction.

[0051] The present invention also provides a computer device, comprising:

[0052] One or more processors;

[0053] A memory that stores one or more computer programs;

[0054] The processor invokes a computer program to achieve the following:

[0055] Steps of a controllable source electromagnetic data denoising method based on covariance-driven gating.

[0056] The present invention also provides a computer-readable storage medium storing a computer program that is invoked by a processor to implement:

[0057] Steps of a controllable source electromagnetic data denoising method based on covariance-driven gating.

[0058] Compared with the prior art, the present invention achieves the following progress and effects:

[0059] 1. The technical solution of this invention uses a temporal convolutional gated network (TCN-GRU), constructed by hybridizing a temporal convolutional network (TCN) and a gated recurrent unit network (GRU), as the main structure of the denoising model. This fully leverages the respective advantages of the two networks, achieving complementarity and mutual assistance to improve the denoising effect of controllable source electromagnetic data. Specifically, a covariance-driven spatiotemporal gating module replaces the gating mechanism module in the original gated recurrent unit network (GRU), generating gating coefficients based on the cross-covariance matrix of the signal feature representation S and the noise feature N. This enables the network to explicitly know the "degree of signal-noise aliasing" in the current time period, achieving more refined "spatiotemporal gating" and significantly improving the noise reduction effect.

[0060] Among them, the signal-noise cross-covariance is introduced to accurately identify the coupling relationship. When the signal is weak and the noise is high, the signal's autocovariance (i.e., the traditional power spectrum) does not show any abnormality, but the cross-covariance of the signal and noise can show significant singular peak values, which can highlight the differences between the noise and signal characteristics.

[0061] 2. The technical solution of this invention further introduces uncertainty estimation, uses periodic statistics to quantify point-by-point uncertainty, achieves high-precision signal-to-noise separation in the controllable source electromagnetic method, improves the difficulties in deep learning denoising of the controllable source electromagnetic method, and broadens the application prospects of controllable source electromagnetic method data processing based on deep learning network architecture. Attached Figure Description

[0062] Figure 1 This is a flowchart of a controllable source electromagnetic data noise reduction method based on covariance-driven gating according to an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of a network process based on covariance-driven gating according to an embodiment of the present invention;

[0064] Figure 3This is a schematic diagram of a model based on a covariance-driven temporal convolutional gating network provided in an embodiment of the present invention;

[0065] Figure 4 The diagram shows the effect of processing the analog signal provided by this invention using this method.

[0066] Figure 5 The image shows the recognition effect of the measured data provided by this invention using this method.

[0067] Figure 6 The image shows a comparison of the measured point curves before and after processing, as provided by this invention. Detailed Implementation

[0068] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0070] This invention proposes a hybrid network (Temporal Convolutional Gated Network TCN-GRU) that integrates Temporal Convolutional Network (TCN) and Gated Recurrent Unit Network (GRU). This network is applied to noise reduction of controllable source electromagnetic data. Based on the characteristics of controllable source electromagnetic data, a covariance-driven spatiotemporal gating module is proposed to replace the gating mechanism module in the original GRU. Specifically, the cross-covariance matrix of signal feature representation S and noise feature N is input into the gating module to generate gating coefficients, thereby achieving the update gate function and significantly improving the overall signal noise reduction effect.

[0071] It should be understood that the core improvement of this invention is the introduction of a covariance-driven spatiotemporal gating module. Under the premise of satisfying this core improvement, other adjustments or optimizations to the temporal convolutional network (TCN) and the gated recurrent unit network (GRU), or technical solutions that retain the original network structure, fall within the scope of protection of this invention. Figure 2 As mentioned above, both dimensionality-upgrading and dimensionality-reducing convolutions are embedded in the TCN module. That is, the technical solution of this invention does not impose strict limitations or requirements on the design of other network layers of the Temporal Convolutional Network (TCN) and the Gated Recurrent Unit Network (GRU).

[0072] Therefore, the technical concept of this invention is as follows:

[0073] A covariance-driven spatiotemporal gating module is introduced to replace the gating mechanism module in the original gated recurrent unit network GRU in the temporal convolutional gating network TCN-GRU.

[0074] Noisy controllable source electromagnetic data is input into a temporal convolutional gated network (TCN-GRU) for noise reduction.

[0075] In this process, training and testing sets can be constructed by referring to existing technologies, and then the temporal convolutional gated network TCN-GRU can be trained. Finally, the trained temporal convolutional gated network TCN-GRU can be used to denoise the controllable source electromagnetic method data to be denoised.

[0076] Furthermore, this invention further optimizes the process by introducing uncertainty estimation based on the final features output by the temporal convolutional gated network TCN-GRU. The standard deviation and mean of the uncertainty estimation are used as representations of the denoised data, further enhancing the denoising effect. This invention presents this technical solution as an improved technique, and the following detailed description will be provided with reference to specific embodiments.

[0077] Example 1

[0078] This invention provides a method for denoising controllable source electromagnetic data based on covariance-driven gating, comprising the following steps:

[0079] Step 1: Preprocessing of the input layer. The noisy controllable source electromagnetic data is fitted with the least squares method to obtain the trend term of the controllable source electromagnetic data. Then, the trend term is subtracted from the original controllable source electromagnetic data to obtain the residual time series data. In addition, the residual time series data in Step 1 is cut into time series segments by sliding window with equal periods.

[0080] In practice, the least squares method is used to fit a low-order polynomial to approximate the trend. This technique is existing technology, therefore it will not be described in detail in this invention. The residual data obtained after the detrending process in step 1 consists of high-frequency components and low-frequency and mid-frequency components that are not affected by the trend.

[0081] Step 2: Multi-scale feature extraction. Each time segment after cutting is input into a parallel dilated convolution branch to extract multi-scale feature maps.

[0082] In the technical solution of this invention, the parallel dilated convolution branch has several parallel convolutional layers with different dilation rates to obtain feature maps at different time scales. The feature maps at different time scales are then spliced ​​together in the channel dimension to form a multi-scale feature map.

[0083] This embodiment employs three parallel convolutional layers with different dilation rates of 1, 3, and 7, corresponding to different receptive field sizes to capture short-term, medium-term, and long-term dependencies. It should be understood that the above dilation rate values ​​are illustrative examples of this invention.

[0084] Step 3: Temporal weighting. Perform global average pooling on the multi-scale feature maps of each time segment to obtain the temporal weight vector. Use the Sigmoid activation function to normalize the temporal weights to between 0 and 1. Then multiply the temporal weight vector with the multi-scale feature maps of the corresponding time segments point by point to obtain the temporal weighted feature F of the time segment.

[0085] Step 4: Residual connection. The temporal weighted feature F is input into a one-dimensional convolutional layer, and the output is a noise feature N. Then, the noise feature N is subtracted from the temporal weighted feature F to obtain the signal feature representation S: S = F - N. Here, the convolutional layer is a one-dimensional convolution. In this embodiment, the kernel size of this one-dimensional convolution is 3, the stride is 1, and the output length is equal to the input length. The input is the temporal weighted feature, and the output is the noise feature.

[0086] It should be understood that a Temporal Convolutional Network (TCN) consists of a single TCN residual block or a stack of multiple TCN residual blocks, with the output of the previous TCN residual block serving as the input of the next TCN residual block. Steps 2-4 involve data processing for a single TCN residual block, step 3 is an attention mechanism, and step 4 is a residual-based signal enhancement mechanism. In this embodiment of the invention, the Temporal Convolutional Network (TCN) consists of a single TCN residual block.

[0087] For the technical solution of this invention, the signal feature representation S and noise feature N of the final output of the temporal convolutional network TCN are input into the gated recurrent unit network GRU to obtain the final feature Z.

[0088] Step 5: Input the feature representation S along with the noise feature N or the cross-covariance matrix of the signal feature representation S and the noise feature N into the gated recurrent unit network (GRU) to obtain the final feature Z.

[0089] The technical solution of this invention uses a covariance-driven spatiotemporal gating module to replace the original GRU gating mechanism module. That is, the spatiotemporal gating module replaces the update gate. It should be understood that the gated recurrent unit network (GRU) can consist of multiple GRU layers or a single GRU layer; this invention does not specifically limit this, only constraining the use of a covariance-driven spatiotemporal gating module to replace the update gate of the GRU layer.

[0090] Let X represent the input feature of the GRU layer. For the first GRU layer, X is the input feature of the spatiotemporal gating module, that is, the feature representation of the multi-scale information output by the TCN in the TCN-GRU hybrid network. When there are multiple GRU layers, the input of other GRU layers is the hidden state output by the previous GRU layer. The input feature X corresponds to the feature at time step t; The gating coefficient corresponds to the generative model as follows:

[0091]

[0092] In the formula, Let be the gating coefficient for time step t corresponding to time segment i; It is the hidden state of the previous time step t-1, and the subscript i represents the marker of the time segment; It is the cross covariance matrix The vector obtained by expanding by rows or columns; R represents the set of real numbers, and d represents the dimension; and Learnable weight matrix; This represents the Sigmoid activation function; For bias.

[0093] The hidden state is updated based on gated sparse updates. The gate coefficient is multiplied element-wise by the hidden state or by the candidate state, specifically as follows:

[0094]

[0095] In the formula, Let t be the hidden state and the candidate state. The calculation formula is:

[0096]

[0097] In the formula, This indicates that the door is being reset. Represents the weight matrix. Indicates the bias term. This represents the hyperbolic tangent activation function.

[0098] Among them, the cross covariance matrix satisfy:

[0099]

[0100] In the formula, This represents the number of time sampling points in time segment i. Let S represent the mean vector of the signal feature representation S corresponding to time segment i. T represents the mean vector of the noise feature N corresponding to time segment i, and T is the matrix transpose symbol; , Let S represent the signal feature representation of time segment i and N represent the feature corresponding to time sampling point t, respectively.

[0101] The final feature Z is composed of the hidden states of all time steps in the last GRU layer, specifically: .

[0102] like Figure 2 and Figure 3 The input time-domain sequence fragment is expanded to a high-dimensional space through convolutional layers in the TCN module (dimensionality increase). Then, parallel dilated convolution is used to extract temporal dependencies, compressing the high-dimensional features to the target output dimension, reducing the number of parameters and aggregating information (dimensionality reduction). The Sigmoid activation function is used to generate gating weights, which are multiplied element-wise with the input features to achieve adaptive feature selection. The output of the TCN module is obtained through residual connections. In the GRU module, the hidden state is dynamically updated through cross-covariance calculation and a covariance-driven spatiotemporal gating module, finally obtaining denoised controllable source electromagnetic data.

[0103] Step 6: Introduce uncertainty estimation based on Monte Carlo Dropout to enhance the noise reduction effect.

[0104] For the final feature , This indicates the number of windows, or the number of time-series segments. This indicates the number of time sampling points for each window. This represents the number of feature channels, with a value of 1. The confidence index is calculated point-by-point, meaning the mean and standard deviation are calculated independently for each time sampling point. This embodiment uses a Bayesian estimation method based on Monte Carlo Dropout, defined as follows:

[0105] First, for the final feature Z, M random forward propagations are performed on the final feature Z using Monte Carlo Dropout (random forgetting layer) to obtain M sets of output feature tensors. For the t-th time sampling point within window b, the output scalar value of the m-th forward propagation is defined as: .

[0106] Then, for each of the M feature tensors corresponding to each time sampling point, the mean and standard deviation are calculated respectively. The mean is used as the denoised signal of the corresponding time sampling point, i.e., the signal tensor; the standard deviation is used as the confidence of the denoised signal of the corresponding time sampling point, i.e., the uncertainty tensor. The uncertainty tensor is used to analyze the denoising results. The specific analysis can refer to the existing technology or existing rules. This invention does not impose specific limitations on this.

[0107] The mean estimate (signal tensor) of the denoised signal is:

[0108]

[0109] The uncertainty tensor (standard deviation estimate) is:

[0110]

[0111] In the formula, Let be the mean value corresponding to the t-th time sampling point within window b. Let be the standard deviation corresponding to the t-th time sampling point within window b.

[0112] Finally, a window function and a weighted overlapping summation method are introduced to process the denoised signal, resulting in the final denoised controllable source electromagnetic method data. The specific model is as follows:

[0113]

[0114]

[0115] In the formula, This is the denoised signal at the t-th time sampling point within window b after windowing modulation; Let L be the window function, and L be the sequence length of the window function. This represents the data at time point t in the final controlled-source electromagnetic method data for noise reduction; m is the frame number (considered as further dividing the time-series data into frames), and R is the frame length. Window function For length is The weighted sequence is calculated using the Hanning window:

[0116]

[0117] Through the above steps, denoising of controllable source electromagnetic data based on covariance-driven gating is achieved, which improves the accuracy of electromagnetic data processing using hybrid networks and uncertainty estimation. Hybrid networks can effectively improve the denoising accuracy and effect of controllable source electromagnetic methods, providing reliable data support for electromagnetic inversion interpretation.

[0118] To verify the effectiveness of this embodiment, the method of the present invention was processed in simulated noisy signals and measured data. Figure 4The diagram shows the processing effect of the simulated noisy signal according to the embodiments of this application after being processed by the method of the present invention; analysis Figure 5 As can be seen, the original data contained noisy data. After processing by the method of this invention, high-quality valid signals can be effectively obtained, and the noisy data is completely removed. (Comprehensive analysis) Figure 4 and Figure 5 It can be shown that the method in this embodiment can effectively achieve noise reduction of controllable source electromagnetic data based on covariance-driven gating, and improve the data quality of controllable source electromagnetic methods in the mid and low frequency bands.

[0119] By comparing the electric field curves before and after processing at the measured points, such as... Figure 6 As shown, the original electromagnetic data is affected by noise, which causes fluctuations at some frequency points in the low-frequency band, resulting in unstable amplitude and low curve quality. It is easy to see that after processing by the method in this embodiment, the electric field curve of the measured point shows a stable trend and no abrupt changes at abnormal frequency points, indicating that the processed controllable source electromagnetic data is effectively improved.

[0120] In summary, in the above embodiments, the noisy controllable source electromagnetic data is fitted with least squares and the trend term is subtracted; the preprocessed data is divided into sliding window segments with equal periods, and each segment is subjected to parallel dilated convolution to extract multi-scale features; the multi-scale features are subjected to global average pooling to obtain a temporal weight vector, and the weights are normalized to between 0 and 1 using the Sigmoid activation function; the temporal weight vector is multiplied point by point with the multi-scale features to obtain the temporal weighted feature F; the noise feature N is estimated through a convolutional layer, and the signal feature is obtained by subtracting the noise feature from the temporal weighted feature. A covariance-driven spatiotemporal gating module is constructed to calculate the cross-covariance matrix of signal feature representation S and noise feature N. This matrix is ​​then used to learn and generate gating coefficients through a hybrid network, ultimately outputting finely separated and enhanced features. The mean and standard deviation of each period's data are calculated from these final features as point-by-point confidence indices, outputting signal tensors and uncertainty tensors. Finally, a weighted overlapping summation method is used to smoothly stitch the denoised data segments back into complete continuous time-series data, obtaining effective data for the controlled-source electromagnetic method. Experimental results show that this invention improves the denoising effect of the controlled-source electromagnetic method based on hybrid networks and uncertainty estimation, enhances the quality of controlled-source electromagnetic data, and possesses practical value and innovation.

[0121] In some embodiments, the present invention also provides a controllable source electromagnetic data noise reduction system based on covariance-driven gating, comprising:

[0122] The network construction module is used to construct the temporal convolutional gated network TCN-GRU, that is, to introduce a covariance-driven spatiotemporal gated module to replace the gated mechanism module in the original gated recurrent unit network GRU in the temporal convolutional gated network TCN-GRU. The temporal convolutional gated network TCN-GRU includes a temporal convolutional network TCN and a gated recurrent unit network GRU connected in sequence.

[0123] The noise reduction module is used to input noisy controllable source electromagnetic data into the temporal convolutional gated network TCN-GRU for noise reduction.

[0124] It should also be understood that the specific implementation process of each module is described in the above method. This invention will not repeat it here. The above division of functional modules is only for illustrative purposes. In some embodiments, some functional modules can be combined and some functional modules can be separated. 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-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0125] In some embodiments, the present invention also provides a computer device, including: one or more processors and a memory storing one or more computer programs;

[0126] The processor invokes a computer program to achieve the following:

[0127] The steps of a controllable source electromagnetic data denoising method based on covariance-driven gating are described below. Specific details are as described in the foregoing method embodiments and will not be repeated here.

[0128] The processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0129] The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory and the processor calls the algorithm program to execute the above methods.

[0130] In some embodiments, the present invention also provides a computer-readable storage medium storing a computer program that is invoked by a processor to implement the steps of a method for denoising controllable source electromagnetic data based on covariance-driven gating. Specific details are described in the foregoing method embodiments and will not be repeated here.

[0131] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the hardware and software device described in any of the foregoing embodiments, such as the hard drive or memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both internal storage units and external storage devices 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.

[0132] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this 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 solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also protected by this invention.

Claims

1. A method for covariance-driven gating based controllable source electromagnetic data denoising, characterized in that: The method includes: A covariance-driven spatiotemporal gating module is introduced to replace the gating mechanism module in the original gated recurrent unit network GRU in the temporal convolutional gating network TCN-GRU, which includes a temporal convolutional network TCN and a gated recurrent unit network GRU connected in sequence. The noisy controllable source electromagnetic method data is input into the temporal convolutional gated network TCN-GRU for noise reduction; The processing of the covariance-driven spatiotemporal gating module is as follows: First, obtain the signal feature representation S and noise feature N cross-covariance matrix of the output of the temporal convolutional network TCN; Then, the cross-covariance matrix is used to generate gating coefficients to replace the update gate of the original gated recurrent unit network GRU.

2. The method according to claim 1, characterized in that: The gating coefficients The model is: In the formula, Let be the gating coefficient for time step t corresponding to time segment i; It is the hidden state of the previous time step t-1, and the subscript i represents the marker of the time segment; It is the cross covariance matrix The vector obtained by expanding by rows or columns; R represents the set of real numbers, and d represents the dimension; and Learnable weight matrix; This represents the Sigmoid activation function; For bias; The input features of the current gate at time step t; The hidden state is updated based on gated sparse updates, specifically as follows: In the formula, The hidden state at time step t. These are candidate states for time step t; Cross covariance matrix satisfy: In the formula, This represents the number of time sampling points in time segment i. Let S represent the mean vector of the signal feature representation S corresponding to time segment i. Let T represent the mean vector of the noise feature N corresponding to time segment i, and T be the matrix transpose symbol. , Let S represent the signal feature representation of time segment i and N represent the feature corresponding to time sampling point t in the noise feature representation.

3. The method according to claim 1, characterized in that: The method further includes: for the final feature Z output by the temporal convolutional gated network TCN-GRU, introducing uncertainty estimation based on Monte Carlo Dropout to enhance the noise reduction effect, as follows: First, for the final feature Z, M random forward propagations are performed on the final feature Z using Monte Carlo Dropout to obtain M sets of output feature tensors. ; Then, for each of the M feature tensors corresponding to each time sampling point, the mean and standard deviation are calculated respectively. The mean is used as the denoised signal for the corresponding time sampling point, and the standard deviation is used as the confidence level of the denoised signal for the corresponding time sampling point. Finally, a window function and a weighted overlapping summation method are introduced to process the denoised signal, resulting in the final denoised controllable source electromagnetic method data.

4. The method according to claim 3, characterized in that: The mathematical model for processing the denoised signal by introducing a window function and a weighted overlap summation method is as follows: In the formula, This is the denoised signal at the t-th time sampling point within window b after windowing modulation using a window function; Let L be the window function, and L be the sequence length of the window function. This refers to the data at the t-th time sampling point in the final controlled-source electromagnetic method data for noise reduction; m is the frame number, and R is the frame length. The denoised signal at the t-th time sampling point within window b; In window function Signal at time sampling points.

5. The method according to claim 3, characterized in that: For each of the M feature tensors corresponding to a given time sampling point, the mathematical model for calculating the mean and standard deviation is as follows: In the formula, For characteristic tensors The feature tensor of the m-th group The feature value of the t-th time sampling point within window b. This is the denoised signal at the t-th time sampling point within window b. Let be the standard deviation of the denoised signal at the t-th time sampling point within window b.

6. The method according to claim 1, characterized in that: When noisy, controllable-source electromagnetic method data is input into the temporal convolutional gated network TCN-GRU for noise reduction, the processing procedure of the temporal convolutional gated network TCN-GRU is as follows: S1. Input layer preprocessing: The trend term of the noisy controllable source electromagnetic method data is subtracted to obtain residual time series data, and the residual time series data is cut into equal length segments to obtain time series segments. S2. Data processing of the temporal convolutional network TCN, wherein the TCN residual blocks of the temporal convolutional network TCN are executed sequentially: Multi-scale feature extraction involves inputting each time segment into a parallel dilated convolution branch to extract multi-scale feature maps. Temporal weighting involves generating temporal weight vectors based on the multi-scale feature maps of time segments, and then multiplying the temporal weight vectors with the corresponding multi-scale feature maps point by point to obtain the temporal weighted features of each time segment. The residual connection inputs the time-domain weighted features into a one-dimensional convolutional layer, outputs noise features, and then subtracts the noise features from the time-domain weighted features to obtain the signal feature representation; S3. Data processing of the gated recurrent unit network (GRU), wherein the feature representation S of the final output of the temporal convolutional network (TCN) is input into the gated recurrent unit network (GRU) along with the noise feature N or the cross-covariance matrix of the signal feature representation S and the noise feature N to obtain the final feature Z, which is the output of the temporal convolutional gated network (TCN-GRU).

7. The method according to claim 6, characterized in that: The parallel dilated convolution branch consists of three or more parallel convolutional layers with different dilation rates, used to extract feature maps at different time scales. Finally, the feature maps at different time scales are concatenated in the channel dimension to obtain a multi-scale feature map.

8. A system based on the method of any one of claims 1-6, characterized in that: include: The network construction module is used to construct the temporal convolutional gated network TCN-GRU, that is, to introduce a covariance-driven spatiotemporal gated module to replace the gated mechanism module in the original gated recurrent unit network GRU in the temporal convolutional gated network TCN-GRU. The temporal convolutional gated network TCN-GRU includes a temporal convolutional network TCN and a gated recurrent unit network GRU connected in sequence. The noise reduction module is used to input noisy controllable source electromagnetic data into the temporal convolutional gated network TCN-GRU for noise reduction.

9. A computer device, characterized in that, include: One or more processors; A memory that stores one or more computer programs; The processor invokes a computer program to achieve the following: The steps of the method according to any one of claims 1-7.

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

Images