Drill hole multi-source sensing signal denoising method based on dynamic noise decoupling and self-adaptive mode

Abstract

The invention discloses a drilling multi-source sensing signal denoising method based on dynamic noise decoupling and a self-adaptive mode, and belongs to the technical field of underground processing. The method comprises the following steps of: separating common noise of a noise energy distribution matrix of a sensor and separating specific noise; carrying out adaptive noise set empirical mode decomposition on the separated signals, carrying out variational mode decomposition on residual signals in the signals, and screening effective modes through a kurtosis-entropy joint criterion to obtain signals with effective characteristic components reserved; high-frequency fluctuation of the signal is punished through total variation regularization, an improved alternating direction multiplier method algorithm is used for solving, short-time Fourier transform is carried out on the solved signal, low-frequency and high-frequency features are extracted through a multi-scale convolutional network, and a final denoised signal is output through gating weight fusion. According to the method, the signal denoising precision in a complex noise environment is remarkably improved, the processing time is shortened, and the resource consumption is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of downhole processing technology, and in particular relates to a drilling multi-source sensor signal denoising method based on dynamic noise decoupling and adaptive mode. Background Art

[0002] In underground processing technology, particularly in coal mines, borehole trajectory imaging is a core infrastructure for mine geological exploration. Its accuracy is directly related to resource extraction efficiency and safety costs. However, the raw signals from borehole trajectory sensors (primarily the azimuth / inclination analog voltage signals of the three-dimensional electronic compass and the pulse sequence signals of the depth encoder) exhibit complex characteristics such as multi-source coupling, non-stationarity, and physical constraints. Their noise sources are complex, including common noise coupled across sensors (probe vibration, electromagnetic pulses) and sensor-specific noise (magnetic field baseline drift, pulse counting errors). The signal's time-frequency characteristics exhibit a fusion of low-frequency trends such as azimuth drift and depth accumulation error with high-frequency noise such as vibration, impact, and electromagnetic interference, resulting in a high degree of overlap between valid information and noise in the time-frequency domain. Crucially, the borehole trajectory signal must strictly adhere to the laws of geomechanics (continuous drilling path) to ensure a continuous signal output.

[0003] In view of the above signal characteristics, traditional denoising methods (such as wavelet threshold and empirical mode decomposition) have significant shortcomings in deep and complex scenarios: (1) Weak noise decoupling ability: relying on a single sensor noise model, it is impossible to separate cross-sensor coupled noise; (2) Lack of modal aliasing and physical constraints: the traditional EMD algorithm causes high-frequency noise residue or low-frequency trend distortion due to modal aliasing, and does not embed the borehole trajectory continuity equation, resulting in measured azimuth jump exceeding the limit and depth resolution; (3) Insufficient anti-interference generalization: the frequency domain method is sensitive to transient noise and ignores the geomechanical laws and signal sparsity priors, resulting in insufficient signal-to-noise ratio improvement, making it difficult to meet the needs of deep exploration. Therefore, there is an urgent need for an optimization method that integrates dynamic decoupling, modal coordination and signal continuity constraints to overcome the dual difficulties of noise suppression and trajectory fidelity. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides a drilling multi-source sensor signal denoising method based on dynamic noise decoupling and adaptive mode.

[0005] To implement the above technique, the steps are as follows: S1. Dynamic noise decoupling and coupling interference elimination: Construct a sensor noise energy distribution matrix and perform common noise separation and specific noise separation operations on the constructed matrix to eliminate cross-sensor coupling interference and obtain a preliminary denoised signal. Here are the steps: S1.1. Based on a preset number of sensors, construct a sensor noise energy distribution matrix, which is expressed as follows: In the formula, the matrix elements , indicating the Sensors in the time window The noisy observation signal, where Indicates the Sensors in the time window The real signal, Indicates the Sensors in the time window Noise signal; Indicates the number of sensors; S1.2. Perform eigendecomposition on the sensor noise energy distribution matrix, extract the eigenvector corresponding to the maximum eigenvalue obtained by the decomposition operation, and use the corresponding eigenvector as the common noise component; The expression of the common noise component is as follows: Where, E Represents the sensor noise energy distribution matrix; The eigenvector corresponding to the maximum eigenvalue is used to characterize the weight distribution of common noise on each sensor; Indicates transposition, which is used to convert column vectors into row vectors for easy comparison with matrices E multiplication; S1.3. Subtract the sum of the common noise component and the transposed eigenvector corresponding to the largest eigenvalue from the sensor noise energy distribution matrix to obtain a residual signal. Analyze the resulting residual signal using Fast Independent Component Analysis (FastICA) to separate sensor-specific noise. The expression of the residual signal is as follows: Where, R represents the residual signal; The sensor-specific noise is expressed as: ; S1.4, using the orthogonal projection operator to eliminate cross-sensor coupling interference and generate a preliminary denoised signal; The expression of the orthogonal projection operator is as follows: Where, I represents the identity matrix; The expression for generating the preliminary denoised signal is as follows: Where, represents the preliminary denoised signal.

[0006] S2. Dual-domain modal collaborative decomposition and effective component screening: Adaptive noise ensemble empirical mode decomposition is performed on the preliminary denoised signal, and variational mode decomposition is performed on the residual signal in the adaptive noise ensemble empirical mode decomposition. Effective modes are screened based on the kurtosis-entropy joint criterion to obtain a preliminary denoised signal that retains effective characteristic components. Here are the steps: S2.1. Perform adaptive noise ensemble empirical mode decomposition (CEEMDAN) on the preliminary denoised signal to add adaptive white noise. The expression is as follows: Where, represents the weight coefficient; express The standard deviation of Indicates adding k An adaptive white noise amplitude, used to control the intensity of the added noise; S2.2, performing variational mode decomposition (VMD) on the residual signal of the adaptive ensemble empirical mode decomposition of noise (CEEMDAN) to set the number of modes and dynamically adjust the number of modes; The expression of the residual signal is as follows: Where, represents the residual signal of the ensemble empirical mode decomposition of adaptive noise (CEEMDAN); K represents the decomposition of the adaptive noise ensemble empirical mode decomposition (CEEMDAN) IMF Number of layers; The expression of variational mode decomposition (VMD) is as follows: Where, M express VMD Theoretical value of modal number; Indicates the sampling frequency, used to obtain the refined component ; Modal number The dynamically adjusted expression is as follows: Where, Represents power spectral density The highest frequency corresponding to the significant peak; Indicates the preset frequency band interval; S2.3. Based on the kurtosis-entropy joint criterion, the kurtosis and permutation entropy of each modal component are calculated, and the effective modal component is output to obtain a preliminary denoised signal that retains the effective characteristic components; Through the joint criterion of kurtosis and permutation entropy, IMF and VMD The effective signal components are screened out from the components and a high-fidelity denoised signal is reconstructed, taking into account the screening of pulses and random noise. The high-fidelity reconstructed signal meets the continuity requirements of geomechanics. The kurtosis of each modal component Expressed as: ; Permutation entropy of each modal component E Expressed as: ; By comparing with the preset threshold, the low kurtosis values ​​less than the threshold are retained. and low entropy E Component; Among them, low kurtosis is used to smooth the signal, and low entropy is used to make the signal orderly; the preset threshold is calibrated according to actual experiments; The output effective modal components are expressed as: ; The preliminary denoised signal that retains the effective characteristic components is expressed as: .

[0007] S3, signal smoothness constrained optimization: Take the initial denoised signal that retains the effective characteristic components as input, use total variation regularization to penalize the high-frequency fluctuations of the signal, and use the improved ADMM algorithm to iteratively solve the problem, and output the optimized signal; Here are the steps: S3.1. Construct a total variation optimization objective function and use total variation regularization to penalize high-frequency fluctuations of the signal to eliminate the impulse noise or random jitter remaining after modal decomposition. The expression of the total variation optimization objective function is as follows: Where, represents the reconstructed signal of the effective modal component; represents the time domain signal to be optimized; represents the adaptive weight coefficient; represents the data fidelity item; represents the total variation regularization term, where represents the time length of the signal, t represents the continuous time variable; among them, the adaptive weight coefficient Includes: first adaptive weight coefficient and the second adaptive weight coefficient ,The update rule of the adaptive weight coefficient is as follows: dynamic ,adjustment based on the local signal to noise ratio of the signal; S3.2. Auxiliary variables and Lagrange multipliers are introduced to construct an improved ADMM algorithm to solve the objective function; The update rules of the improved ADMM algorithm include: Where, Represents the iteration index of the ADMM algorithm, indicating the iterations; is the penalty factor; the update rule is as follows ,expression and expressions As shown; is a soft threshold function used to impose sparsity constraints; Represents auxiliary variables; represents the Lagrange multiplier.

[0008] S4, joint optimization in the time-frequency domain: The optimized signal is subjected to a short-time Fourier transform to retain the low-frequency trend components of the signal; low-frequency and high-frequency features are extracted separately through a preset multi-scale convolutional network, and the final denoised signal is output through gated weight fusion; Here are the steps: S4.1. Perform short-time Fourier transform on the optimized signal and use low-frequency protection strategy to mask high-frequency noise bands and avoid excessive signal suppression. S4.2, use three sets of complex convolution kernels to extract low-frequency and high-frequency features respectively, train them through gated weight fusion, and output the final denoised signal; The expression of gated weight fusion is as follows: Where, Represents low-frequency features, which are extracted by the low-frequency convolution kernel in 3 groups of complex convolution kernels; Represents high-frequency features, which are extracted by the high-frequency convolution kernels in 3 groups of complex convolution kernels; represents the gating weight matrix (learnable parameters); Represents the Sigmoid activation function; Represents feature splicing; The training process includes: using Adam as the optimizer and the mean square error between the reconstructed signal and the true signal as the loss function; The expression of the loss function is as follows: Where, represents the reconstructed signal (output of gated weight fusion); Represents the true signal; the stopping condition of the loss function is: when the maximum number of iterations of the loss function is reached, the training stops.

[0009] Beneficial effects of the present invention: The present invention is based on the eigendecomposition and orthogonal projection algorithm of the multi-sensor noise energy matrix to accurately separate cross-channel coupling noise and sensor-specific noise, thereby reducing signal interference at the source. On this basis, the preliminary denoised signal is subjected to a dual-domain collaborative decomposition through the CEEMDAN-VMD hybrid decomposition framework, and the effective modal components are screened by combining the kurtosis-entropy joint criterion, which significantly suppresses the modal aliasing and high-frequency detail loss problems caused by traditional single decomposition methods.

[0010] The present invention combines signal smoothness constraints and sparsity priors to construct the objective function, and uses an improved ADMM algorithm to dynamically optimize core parameters such as the number of decomposition layers and sparse weights to ensure that the denoising results conform to physical laws and have frequency domain sparse characteristics.

[0011] The present invention ultimately performs masking enhancement and feature fusion on the spectrum of the optimized signal through a multi-scale gated convolutional network, thereby maximizing the retention of the denoised signal characteristics while suppressing residual noise.

[0012] The present invention can significantly improve the signal denoising accuracy in complex noise environments, shorten processing time, and reduce resource consumption while ensuring the fidelity of high-frequency weak signals, providing efficient and reliable technical support for deep mineral resource exploration and engineering safety monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the present invention; Figure 2 The original signal diagram collected by the embodiment of the present invention; Figure 3 This is the denoising result diagram of the present invention. DETAILED DESCRIPTION

[0014] In order to better illustrate the purpose, technical solutions and advantages of the present invention, the present invention will be further described below in conjunction with specific embodiments.

[0015] This example uses the ZKXG100 coal mine trajectory measurement instrument, equipped with a three-dimensional electronic compass and depth encoder, including: Three-axis magnetoresistive sensor enables azimuth and inclination measurement; The depth encoder realizes depth measurement and synchronized acquisition of noise signals across frequency bands; Signal source: Drilling depth is 0-100m, covering the main ore body and surrounding rock (granite and phyllite) of the copper mine. The collected signals include geophysical signals and noise interference, mainly involving medium-deep holes at the top of the mining area and large holes downward.

[0016] like Figure 1 As shown in the figure, a drilling multi-source sensor signal denoising method based on dynamic noise decoupling and adaptive mode is proposed. The steps are as follows: S1. Dynamic noise decoupling and coupling interference elimination: Construct a sensor noise energy distribution matrix and perform common noise separation and specific noise separation operations on the constructed matrix to eliminate cross-sensor coupling interference and obtain a preliminary denoised signal. The steps are as follows: S1.1. Based on a preset number of sensors, construct a sensor noise energy distribution matrix, which is expressed as follows: In the formula, the matrix elements , indicating the Sensors in the time window The noisy observation signal, where Indicates the Sensors in the time window The real signal, Indicates the Sensors in the time window Noise signal; Indicates the number of sensors; in this embodiment, the calculation of each sensor in the time window ( t =0.1s) E , =4, the matrix dimension is 4×1000; S1.2. Perform eigendecomposition on the sensor noise energy distribution matrix, extract the eigenvector corresponding to the maximum eigenvalue obtained by the decomposition operation, and use the corresponding eigenvector as the common noise component (such as mechanical vibration); The expression of the common noise component is as follows: Where, E Represents the sensor noise energy distribution matrix; The eigenvector corresponding to the maximum eigenvalue is used to characterize the weight distribution of common noise on each sensor; Indicates transposition, which is used to convert column vectors into row vectors for easy comparison with matrices E multiplication; S1.3. Subtract the sum of the common noise component and the transposed eigenvector corresponding to the largest eigenvalue from the sensor noise energy distribution matrix to obtain a residual signal. Analyze the resulting residual signal using Fast Independent Component Analysis (FastICA) to separate sensor-specific noise. The expression of the residual signal is as follows: Where, R represents the residual signal; The sensor-specific noise is expressed as: ; S1.4, using the orthogonal projection operator to eliminate cross-sensor coupling interference and generate a preliminary denoised signal; After separating common noise and specific noise, the residual signal may still contain interference that has not been completely decoupled (such as cross-coupling between sensors). The orthogonal projection operator is used to further strip away these residual coupling components to ensure that the noise subspaces do not overlap. The expression of the orthogonal projection operator is as follows: Where, I represents the identity matrix; The expression for generating the preliminary denoised signal is as follows: Where, represents the preliminary denoised signal.

[0017] S2. Dual-domain modal collaborative decomposition and effective component screening: Adaptive noise ensemble empirical mode decomposition is performed on the preliminary denoised signal, and variational mode decomposition is performed on the residual signal in the adaptive noise ensemble empirical mode decomposition. Effective modes are screened based on the kurtosis-entropy joint criterion to obtain a preliminary denoised signal that retains effective characteristic components. Here are the steps: S2.1. Perform adaptive noise ensemble empirical mode decomposition (CEEMDAN) on the preliminary denoised signal to add adaptive white noise. The expression is as follows: Where, represents the weight coefficient; express The standard deviation of Indicates adding k An adaptive white noise amplitude, used to control the intensity of the added noise; S2.2, performing variational mode decomposition (VMD) on the residual signal of the adaptive ensemble empirical mode decomposition of noise (CEEMDAN) to set the number of modes and dynamically adjust the number of modes; The expression of the residual signal is as follows: Where, represents the residual signal of the ensemble empirical mode decomposition of adaptive noise (CEEMDAN); K represents the decomposition of the adaptive noise ensemble empirical mode decomposition (CEEMDAN) IMFIn this embodiment, the number of decomposition layers is 6, the number of noise additions is 100, and the generated IMF gather ; The expression of variational mode decomposition (VMD) is as follows: Where, M express VMD Theoretical value of modal number; Indicates the sampling frequency, used to obtain the refined component In this embodiment, set =50Hz, generating VMD gather ; Modal number Dynamic adjustment is performed because the frequency domain characteristics of the residual signal are different at different drilling depths or in noise environments. M Automatically adjust to suit signal characteristics; that is, the number of modes is not a fixed value, but is calculated in real time based on the power spectral density (PSD) of the residual signal; Calculate the residual signal The power spectral density , the expression is as follows: Where, express The highest frequency corresponding to the significant peak in this embodiment =50Hz is set as the preset frequency band interval ; S2.3. Based on the kurtosis-entropy joint criterion, the kurtosis and permutation entropy of each modal component are calculated, and the effective modal component is output to obtain a preliminary denoised signal that retains the effective characteristic components; Through the joint criterion of kurtosis and permutation entropy, IMF and VMD The effective signal components are screened out from the components and a high-fidelity denoised signal is reconstructed, taking into account the screening of pulses and random noise. The high-fidelity reconstructed signal meets the continuity requirements of geomechanics. The kurtosis of each modal component Expressed as: ; Permutation entropy of each modal component E Expressed as: ; By comparing with the preset threshold, the low kurtosis values ​​less than the threshold are retained. and low entropy E Component; Among them, low kurtosis is used to smooth the signal, and low entropy is used to make the signal orderly; the preset threshold is calibrated according to actual experiments; The output effective modal components are expressed as: ; The preliminary denoised signal that retains the effective characteristic components is expressed as: .

[0018] S3, signal smoothness constrained optimization: Take the initial denoised signal that retains the effective characteristic components as input, use total variation regularization to penalize the high-frequency fluctuations of the signal, and use the improved ADMM algorithm to iteratively solve the problem, and output the optimized signal; Here are the steps: S3.1. Construct a total variation optimization objective function and use total variation regularization to penalize high-frequency fluctuations of the signal to eliminate the impulse noise or random jitter remaining after modal decomposition. The expression of the total variation optimization objective function is as follows: Where, represents the reconstructed signal of the effective modal component; represents the time domain signal to be optimized; represents the adaptive weight coefficient; represents the data fidelity item; represents the total variation regularization term, where Indicates the time length of the signal (total time points 1000), t represents a continuous time variable (time domain signal index); where the adaptive weight coefficient Includes: first adaptive weight coefficient and the second adaptive weight coefficient ,The update rule of the adaptive weight coefficient is as follows: dynamic ,adjustment based on the local signal to noise ratio of the signal; The local signal-to-noise ratio is expressed as follows: ; The dynamically adjusted expression is as follows: Where, represents the base coefficient, Take 0.5; S3.2. Auxiliary variables and Lagrange multipliers are introduced to construct an improved ADMM algorithm to solve the objective function; The update rules of the improved ADMM algorithm include: Where, Represents the iteration index of the ADMM algorithm, indicating the iterations; is the penalty factor; the maximum number of iterations of the improved ADMM algorithm is 200; the update rule is as follows: ,expression and expressions As shown; is a soft threshold function used to impose sparsity constraints; Represents auxiliary variables; represents the Lagrange multiplier.

[0019] S4, joint optimization in the time-frequency domain: The optimized signal is subjected to a short-time Fourier transform to retain the low-frequency trend components of the signal; low-frequency and high-frequency features are extracted separately through a preset multi-scale convolutional network, and the final denoised signal is output through gated weight fusion; Here are the steps: S4.1. Perform short-time Fourier transform on the optimized signal and use low-frequency protection strategy to mask high-frequency noise bands and avoid excessive signal suppression. S4.2, use three sets of complex convolution kernels to extract low-frequency and high-frequency features respectively, train them through gated weight fusion, and output the final denoised signal; The expression of gated weight fusion is as follows: Where, Represents low-frequency features, which are extracted by the low-frequency convolution kernel in 3 groups of complex convolution kernels; Represents high-frequency features, which are extracted by the high-frequency convolution kernels in 3 groups of complex convolution kernels; represents the gating weight matrix (learnable parameters); Represents the Sigmoid activation function; Represents feature splicing; The training process includes: using Adam as the optimizer and the mean square error between the reconstructed signal and the true signal as the loss function; The expression of the loss function is as follows: Where, represents the reconstructed signal (output of gated weight fusion); Represents the true signal; the loss function stopping condition is: =1000, stop training.

[0020] In order to verify the present invention, Figure 2 This is the original signal graph collected for the case of this invention. Figure 3 This is the denoising result diagram of the present invention; The original noisy analog voltage signals of the X, Y, and Z axes of the three-dimensional electronic compass contain high-frequency jitter signals from the sensor, resulting in poor smoothness in the results. The depth encoder shows the same result. The denoising results of the present invention are significantly better than the original signals, demonstrating the significant advantages of the patented method. High-frequency noise is completely suppressed, verifying the effectiveness of dynamic noise decoupling (common / specific noise separation). Thanks to the dynamic screening of the CEEMDAN-VMD hybrid decomposition and the kurtosis-entropy criterion in the present invention, the denoised signal has no modal aliasing and the denoising effect is obvious, indicating that the present invention can significantly improve the signal denoising accuracy in complex noise environments, shorten processing time, and reduce resource consumption while ensuring the fidelity of high-frequency weak signals, providing efficient and reliable technical support for deep mineral resource exploration and engineering safety monitoring.

[0021] It should be noted that the above are only preferred embodiments of the present application and do not limit the scope of patent protection of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the scope of patent protection of the present application.

Claims

1. A method for denoising multi-source drilling sensor signals based on dynamic noise decoupling and adaptive mode, characterized in that: The following steps are involved: S1. Dynamic noise decoupling and coupling interference elimination: Construct a sensor noise energy distribution matrix and perform common noise separation and specific noise separation operations on the constructed matrix to eliminate cross-sensor coupling interference and obtain a preliminary denoised signal. S2. Dual-domain modal collaborative decomposition and effective component screening: Adaptive noise ensemble empirical mode decomposition is performed on the preliminary denoised signal, and variational mode decomposition is performed on the residual signal in the adaptive noise ensemble empirical mode decomposition. Effective modes are screened based on the kurtosis-entropy joint criterion to obtain a preliminary denoised signal that retains effective characteristic components. S3, signal smoothness constrained optimization: The initial denoised signal that retains effective characteristic components is used as input, total variation regularization is used to penalize the high-frequency fluctuations of the signal, and the improved alternating direction multiplier method (ADMM) algorithm is used to iteratively solve the problem, and the optimized signal is output; S4, time-frequency domain joint optimization: perform short-time Fourier transform on the optimized signal to retain the low-frequency trend component of the signal; The low-frequency and high-frequency features are extracted separately through the preset multi-scale convolutional network, and the final denoised signal is output through gated weight fusion to complete denoising.

2. The method for denoising multi-source drilling sensor signals based on dynamic noise decoupling and adaptive mode according to claim 1 is characterized in that: The steps of constructing a sensor noise energy distribution matrix and performing a common noise separation operation and a specific noise separation operation on the constructed matrix to eliminate cross-sensor coupling interference and obtain a preliminary denoised signal include: S1.

1. Construct a sensor noise energy distribution matrix based on a preset number of sensors; S1.

2. Perform eigendecomposition on the sensor noise energy distribution matrix, extract the eigenvector corresponding to the maximum eigenvalue obtained by the decomposition operation, and use the corresponding eigenvector as the common noise component; S1.

3. Subtract the sum of the common noise component and the transposed eigenvector corresponding to the maximum eigenvalue from the sensor noise energy distribution matrix to obtain a residual signal; The obtained residual signal was analyzed using fast independent component analysis to separate sensor-specific noise; S1.

4. Eliminate cross-sensor coupling interference through orthogonal projection operators and generate preliminary denoised signals.

3. The method for denoising multi-source drilling sensor signals based on dynamic noise decoupling and adaptive mode according to claim 1 is characterized in that: The steps of performing adaptive noise set empirical mode decomposition on the preliminary denoised signal, performing variational mode decomposition on the residual signal in the adaptive noise set empirical mode decomposition, and screening effective modes based on the kurtosis-entropy joint criterion to obtain the preliminary denoised signal retaining effective characteristic components are as follows: S2.

1. Performing adaptive noise set empirical mode decomposition on the preliminary denoised signal to add adaptive white noise; S2.2, performing variational modal decomposition on the residual signal of the adaptive noise set empirical mode decomposition to set the mode number and dynamically adjust the mode number; The dynamic adjustment method is: real-time calculation based on the power spectrum density of the residual signal; S2.

3. Based on the kurtosis-entropy joint criterion, the kurtosis and permutation entropy of each modal component are calculated, and the effective modal component is output to obtain a preliminary denoised signal that retains the effective characteristic components; The way to output the effective modal component is to compare it with the preset threshold and retain the low kurtosis that is less than the threshold. and low entropy E The weight.

4. The method for denoising multi-source drilling sensor signals based on dynamic noise decoupling and adaptive mode according to claim 1 is characterized in that: The steps of taking the preliminary denoised signal that retains the effective characteristic components as input, using total variation regularization to penalize the high-frequency fluctuations of the signal and using the improved alternating direction multiplier method ADMM algorithm to iteratively solve and output the optimized signal are as follows: S3.

1. Construct a total variation optimization objective function and use total variation regularization to penalize high-frequency fluctuations of the signal to eliminate the impulse noise or random jitter remaining after modal decomposition. The expression of the total variation optimization objective function is as follows: Where, represents the reconstructed signal of the effective modal component; represents the time domain signal to be optimized; represents the adaptive weight coefficient; represents the data fidelity item; represents the total variation regularization term, where Indicates the time length of the signal, t represents a continuous time variable; where the adaptive weight coefficient Including: first adaptive weight coefficient and the second adaptive weight coefficient ,The update rule of the adaptive weight coefficient is as follows: dynamic ,adjustment based on the local signal to noise ratio of the signal; S3.

2. Auxiliary variables and Lagrange multipliers are introduced to construct an improved ADMM algorithm to solve the objective function; The update rules of the improved ADMM algorithm include: Where, Represents the iteration index of the ADMM algorithm, indicating the iterations; is the penalty factor; the update rule is as follows ,expression and expressions As shown; is a soft threshold function used to impose sparsity constraints; Represents auxiliary variables; represents the Lagrange multiplier.

5. The method for denoising multi-source drilling sensor signals based on dynamic noise decoupling and adaptive mode according to claim 1 is characterized in that: The optimized signal is subjected to short-time Fourier transform to retain the low-frequency trend component of the signal; low-frequency and high-frequency features are extracted respectively through a preset multi-scale convolutional network, and the final denoised signal is output through gated weight fusion as follows: S4.

1. Perform short-time Fourier transform on the optimized signal and use low-frequency protection strategy to mask high-frequency noise bands and avoid excessive signal suppression. S4.

2. Use a preset number of complex convolution kernels to extract low-frequency and high-frequency features respectively, train them through gated weight fusion, and output the final denoised signal; The training process includes: using Adam as the optimizer and the mean square error between the reconstructed signal and the true signal as the loss function; The stopping condition of the loss function is: when the maximum number of iterations of the loss function is reached, the training stops.

Images