Stratum exploration multi-sensor signal noise reduction reconstruction method based on singular value decomposition

Abstract

The invention discloses a stratigraphic exploration multi-sensor signal noise reduction reconstruction method based on singular value decomposition. The method comprises the following steps: step 1, signal acquisition and preprocessing; step 2, singular value decomposition SVD adaptive layered noise reduction is carried out; step 3, performing wavelet transformation and adaptive threshold processing; 4, reconstructing time-space domain signals and extracting stratum features; and 5, self-adaptive parameter adjustment is carried out. According to the method, a noise filtering algorithm combining singular value decomposition with wavelet transform is constructed by fusing displacement, drilling speed, pressure, torque and other multi-source sensor signals. The SVD is adopted to extract the principal component of the signal, and the wavelet transform is combined for detail enhancement, so that the efficient noise reduction and feature retention of the multi-sensor signal are realized, the accuracy of signal processing is effectively improved, the noise interference is reduced, and the efficiency and accuracy of stratum exploration are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geophysical exploration, and particularly relates to a processing and noise suppression technology for multi-source sensor signals in engineering geological exploration. Especially for multi-dimensional sensor signals such as displacement, drilling speed, pressure, and torque, a joint denoising model in time and space domain is constructed to realize effective signal extraction and reconstruction in a complex noise environment, and to provide high-precision data support for formation structure analysis. BACKGROUND

[0002] In geophysical exploration, the underground rock physical parameters collected by multi-sensor arrays (displacement, drilling speed, pressure, and torque sensors) are often affected by environmental noise (electromagnetic interference, mechanical vibration), instrument noise (background noise, transmission distortion), and signal coupling noise. The traditional method has defects such as insufficient noise suppression, signal distortion, and waste of multi-source information due to independent processing of single signals, fixed threshold strategy, and low computational efficiency. SUMMARY

[0003] In order to overcome the deficiencies of insufficient noise suppression, low utilization rate of multi-source information, and the contradiction between noise reduction and fidelity in existing multi-sensor signal processing, the present application provides a multi-sensor signal denoising and reconstruction method for stratum exploration based on singular value decomposition, which realizes high-precision reconstruction of effective signals in a complex noise environment by constructing a joint decomposition model of multi-dimensional observation matrix and a dynamic threshold optimization mechanism.

[0004] The technical solution adopted by the present application to solve its technical problems is: A multi-sensor signal denoising and reconstruction method for stratum exploration based on singular value decomposition, comprising the following steps: Step 1, signal acquisition and preprocessing: by deploying displacement, drilling speed, pressure, and torque sensor arrays on the exploration drilling rig, synchronously collecting stratum response signals under a unified time reference, ensuring the spatio-temporal consistency of each signal, and through standardization processing of multi-sensor signals, eliminating the dimensional differences of different sensor signals, and making all signals in the same order of magnitude; Step 2, singular value decomposition (SVD) adaptive hierarchical denoising: the normalized multi-sensor signal matrix is divided into blocks according to a fixed size, singular value decomposition (SVD) is performed on each block, the singular values and singular vectors reflecting the principal components of the signal are extracted, the effective singular values are screened through the "energy proportion" and "singular value difference" two standards, the signal matrix is reconstructed using the screened effective singular values and corresponding singular vectors, the matrix rank is reduced by discarding small singular values dominated by noise, and the goal of "preserving effective signals and filtering random noise" is achieved, obtaining a preliminarily denoised signal matrix; Step 3, wavelet transform and adaptive threshold processing: selecting a wavelet base suitable for non-stationary signals, and determining the decomposition layer number according to the number of sampling points, the signal is decomposed into two parts of "low frequency approximation" (overall trend) and "high frequency detail" (mutation characteristics), the statistical characteristics (median absolute deviation) of high frequency detail coefficients are used to estimate the noise intensity, and the interference of abnormal values is avoided; according to the noise intensity, the number of sampling points and the number of sensors, the wavelet threshold is dynamically calculated to ensure that the threshold can effectively filter noise and not damage the signal details; at the same time, the shrinkage factor is introduced, so that the threshold is adaptively adjusted with the signal energy, realizing "strong signal less shrinkage, weak signal more protection"; the smooth threshold function is applied to the high frequency detail coefficients, and the coefficients exceeding the threshold are proportionally shrunk, and the coefficients below the threshold are zeroed; Step 4, space-time signal reconstruction and formation feature extraction: the processed low frequency approximation coefficients and high frequency detail coefficients are combined through inverse wavelet transform to restore the complete waveform of the time domain signal; combining the global features (multi-sensor spatial correlation) of SVD and the local features (time domain details) of wavelet transform, the reconstructed signal matrix is optimized; the time-frequency distribution of the reconstructed signal is analyzed by short-time Fourier transform (STFT), the energy peak position of the formation interface reflection wave is identified, and the reflection wave travel time is determined; the energy attenuation degree of the reflection wave in the propagation process is quantified, and the formation lithology difference is inverted combined with geological prior knowledge; Step 5, adaptive parameter adjustment: the proportion of noise in the signal is evaluated by calculating the local singular entropy - the smaller the singular entropy, the purer the signal; the larger the singular entropy, the more serious the noise; by checking the error between the reconstructed signal and the original signal, the wavelet decomposition layer number, threshold size and other parameters are dynamically adjusted to ensure that the denoised signal has no obvious noise residue and no effective feature loss.

[0005] The technical concept of the application is: a multi-sensor signal processing framework for cross-domain feature fusion is constructed, a space-time correlation matrix containing multi-sensor signals such as displacement, drilling speed, pressure and torque is constructed, the feature complementation between different physical quantities is realized, the main components of the signal are extracted by standardization processing and singular value decomposition, and double-layer noise reduction is realized by wavelet transform and adaptive half-soft threshold function, wherein the threshold formula innovatively incorporates the number of sensors to reduce the noise estimation variance, and the threshold is adaptively adjusted according to the energy ratio and singular value difference to optimize the truncation decision; in the signal reconstruction stage, the spatial correlation and time details are fused by the space-time domain collaborative algorithm, and the reflection wave features are accurately extracted by combining the short-time Fourier transform, and finally the signal-to-noise ratio is improved.

[0006] The beneficial effects of the application mainly include: improving the accuracy of multi-source sensor signal processing, reducing the interference of noise on effective signals, reducing the distortion of effective signals in the signal processing process, ensuring the reliability of the data, improving the efficiency of formation exploration, and helping to more accurately obtain formation information. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 This is a flowchart of a multi-sensor signal noise reduction and reconstruction method for stratigraphic exploration based on singular value decomposition. Detailed Implementation

[0008] The present invention will now be further described with reference to the accompanying drawings.

[0009] Reference Figure 1 A method for denoising and reconstructing multi-sensor signals in stratigraphic exploration based on singular value decomposition includes the following steps: Step 1, Signal Acquisition and Preprocessing: By deploying sensor arrays such as displacement, drilling speed, pressure, and torque on the exploration drilling rig, formation response signals are synchronously acquired under a unified time reference to ensure the spatiotemporal consistency of each signal. By standardizing the multi-sensor signals, the dimensional differences between different sensor signals are eliminated, so that all signals are at the same order of magnitude. The process of step 1 is as follows: A four-channel high-precision sensor array is installed on the exploration drilling rig, including: (1.1) Displacement sensor: displacement gauge (accuracy 0.01mm, sampling frequency 100Hz); (1.2) Drilling speed sensor: magnetoelectric speed sensor (resolution 0.1 m / min, sampling frequency 100 Hz). (1.3) Pressure sensor: Piezoresistive pressure transmitter (range 0~100MPa, accuracy 0.1MPa, sampling frequency 100Hz). (1.4) Torque sensor: strain gauge torque sensor (range 0~5000N·m, accuracy 1N·m, sampling frequency 100Hz). (1.5) GNSS clock synchronization technology is adopted, and multi-channel synchronous sampling is achieved through distributed data acquisition units, with a time error of ≤0.5ms, to ensure spatiotemporal consistency.

[0010] n =4-channel sensors m The signal from each sampling point is constructed as follows: m × n When matrix X continuously collects data for 30 minutes, m =30×60×100=180000, forming an 180000×4 matrix. Each column corresponds to the time series of one sensor: ; in, x ij Indicates the first j The sensor at the first i Observations at each sampling point ( i=1,2,…, m ; j =1,2,…, n ).

[0011] Perform standardization preprocessing on matrix X: calculate the mean of each sensor signal. and standard deviation Then standardize the matrix elements. Standardized matrix Each row has a mean of 0 and a variance of 1, ensuring that subsequent singular value decompositions are unaffected by dimensions.

[0012] Step 2, Singular Value Decomposition (SVD) Adaptive Hierarchical Noise Reduction: The standardized multi-sensor signal matrix is ​​divided into blocks of fixed size. Singular Value Decomposition (SVD) is performed on each block to extract singular values ​​and singular vectors that reflect the principal components of the signal. Valid singular values ​​are selected using two criteria: "energy percentage" and "singular value difference". The signal matrix is ​​reconstructed using the selected valid singular values ​​and corresponding singular vectors. By discarding small singular values ​​dominated by noise, the rank of the matrix is ​​reduced, achieving the goal of "preserving valid signals and filtering random noise", and obtaining the signal matrix after preliminary noise reduction. The process of step 2 is as follows: The singular value decomposition is performed on the preprocessed standardized matrix X′. The core of this process is to decompose the matrix into the product of three orthogonal matrices and a diagonal matrix. The sub-steps are as follows: (2.1) Singular Value Decomposition Expression like (Assuming) If ), then there exists a unique decomposition. ; in, Let u1, u2, ..., u be a left singular matrix with column vectors u1, u2, ..., u3. n For matrix The orthogonal unit eigenvectors represent the "principal component directions" of the standardized multi-sensor signals, reflecting the correlation between signals from different sensors; V∈R m×m Let v1, v2, ..., v be a right singular matrix. m For matrix The orthogonal unit eigenvectors represent the characteristic patterns of the signal in the time domain, corresponding to components with different frequencies or time-series variations.

[0013] ∑∈R n×m This is a diagonal matrix, representing the quantization of signal energy, with diagonal elements... Singular values ​​( r For matrix The rank of the singular value is 0, and the rest of the elements are 0. The energy level of the corresponding principal component was measured; (2.2) The process of singular value decomposition is as follows: 2.2.1 Calculate the eigenvalues ​​and eigenvectors of the covariance matrix: Calculate eigenvalues and the corresponding orthogonal unit eigenvectors u1, u2, ..., u n This forms a left singular matrix U; calculate eigenvalues and the corresponding orthogonal unit eigenvectors v1, v2, ..., v m This forms a left singular matrix V; 2.2.2 Determining Singular Values: Singular Values Arranged in descending order, they form a diagonal matrix Σ, where All other elements are 0; 2.2.3 Matrix Decomposition Verification: Verification This ensures the correctness of the decomposition.

[0014] Through the above decomposition, the standardized matrix X′ is transformed into singular values ​​and eigenvectors that reflect the signal energy distribution, providing a mathematical basis for subsequent noise reduction and reconstruction based on threshold decision.

[0015] (2.3) Adaptive truncation decision: Calculate the cumulative energy ratio E( r ): ; Select the smallest r Make E( r )≥ α ( α For the energy retention threshold, ∈[0.9,0.98]), construct the truncation matrix Σ r : ; Obtain the noise reduction matrix Among them, U r and V r The front of U and V respectively r A matrix composed of columns.

[0016] Step 3: Wavelet Transform and Adaptive Thresholding: Select a suitable wavelet basis for the non-stationary signal and determine the number of decomposition levels based on the number of sampling points. Decompose the signal into two parts: "low-frequency approximation" (overall trend) and "high-frequency details" (abrupt features). Utilize the statistical characteristics (median absolute deviation) of the high-frequency detail coefficients to estimate the noise intensity and avoid interference from outliers. Dynamically calculate the wavelet threshold based on the noise intensity, number of sampling points, and number of sensors to ensure that the threshold effectively filters noise without damaging signal details. Simultaneously, introduce a shrinkage factor to adaptively adjust the threshold according to changes in signal energy, achieving "less shrinkage for strong signals and more protection for weak signals." Apply a smooth threshold function to the high-frequency detail coefficients, shrinking coefficients exceeding the threshold proportionally and setting coefficients below the threshold to zero. The process of step 3 is as follows: (3.1) To X r Each column of signals x r,j conduct J Layer wavelet decomposition: ; in, For the first J Low-frequency approximation coefficients of the layer; For the first j High-frequency detail factor of the layer.

[0017] (3.2) Adaptive thresholding: The first threshold is based on Stein's unbiased likelihood estimation (SURE). j Layer threshold The calculation formula is as follows: ; in, For the first j The standard deviation of layer noise is estimated using the median absolute deviation (MAD): ; Apply a semi-soft thresholding function to detail coefficients: ; in, These are the wavelet detail coefficients after thresholding. For the original wavelet detail coefficients (the first one) j Layer i (coefficients); sgn(·) is the sign function, which returns the sign of the input value; This is an indicator function that takes the value 1 when the condition is met and 0 otherwise. Adaptive parameters To ensure the adaptability of threshold processing, where Let L2 norm be the square of the detail coefficients of the j-th layer.

[0018] Step 4, Spatiotemporal Signal Reconstruction and Stratigraphic Feature Extraction: The processed low-frequency approximation coefficients and high-frequency detail coefficients are merged through inverse wavelet transform to restore the complete waveform of the time-domain signal; the reconstructed signal matrix is ​​optimized by combining the global features of SVD (multi-sensor spatial correlation) and the local features of wavelet transform (time-domain details); the time-frequency distribution of the reconstructed signal is analyzed by short-time Fourier transform (STFT) to identify the energy peak position of the reflected wave at the stratigraphic interface and determine the travel time of the reflected wave; the energy attenuation of the reflected wave during propagation is quantified, and stratigraphic lithological differences are inverted by combining prior geological knowledge; The process of step 4 is as follows: By reconstructing each column of signals using inverse wavelet transform, the final denoised signal matrix Y is obtained, whose elements... Indicates the first j The sensor at the first i The denoised value of each sampling point; An inverse transform is performed on the wavelet coefficients after semi-soft thresholding to recover the detailed features of the time-domain signal. The low-frequency approximation coefficients are then... c A J High-frequency coefficients after thresholding Perform inverse wavelet transform (Mallat algorithm): ; Where IDWT represents the inverse discrete wavelet transform; y j ( t ) is the first j Time-domain reconstructed signal of road sensor ( j =1,2,…, n J represents the wavelet decomposition level, calculated according to... Dynamic determination ensures effective analysis of high-frequency noise and low-frequency signals; The wavelet basis chosen is Daubechies-4 (db4), whose compact support and symmetry reduce reconstruction distortion and are suitable for abrupt changes in formation signals (such as interface reflection spikes). The signals after inverse wavelet transform are reconstructed into matrices and combined with the left singular matrix U of SVD. r and the right singular matrix V r To achieve spatiotemporal domain collaborative optimization, the matrix reconstruction formula is as follows: ; Among them, U r For the left singular matrix before r Column (capturing spatial correlation of multi-sensor signals); Σ r To retain the previous r A diagonal matrix with singular values; V r For the right singular matrix before rColumn (representing the principal components of the signal in the time domain); This is the detail error matrix after the inverse wavelet transform (its magnitude is much smaller than the effective signal).

[0019] By using the inverse SVD transform, the reconstructed matrix Y not only preserves the spatial correlation characteristics of multi-sensor signals (such as the synchronous change of displacement-torque), but also repairs the temporal details (such as the phase characteristics of reflected waves) through the inverse wavelet transform, thus achieving the dual preservation of "spatial correlation + temporal details".

[0020] Time-frequency analysis was performed on the reconstructed signal matrix Y to extract key parameters reflecting the formation structure. travel time of reflected waves t r calculate: ; in, For the first j The signal at the center frequency f c The short-time Fourier transform is performed using the Hann window as the window function; f c =50Hz is the characteristic frequency of the interaction between the drill bit and the formation (which can be adaptively adjusted according to the drilling machinery parameters).

[0021] Amplitude attenuation coefficient calculation: ; Where t0 is the reference time point (such as the start time of drilling); α is the attenuation factor (value 0.5~1.0, determined by the damping characteristics of the formation medium), which reflects the degree of energy attenuation of the signal when it propagates underground.

[0022] Step 5, Adaptive parameter adjustment: The proportion of noise in the signal is evaluated by calculating the local singular entropy—the smaller the singular entropy, the purer the signal; the larger the singular entropy, the more severe the noise; by verifying the error between the reconstructed signal and the original signal, parameters such as the number of wavelet decomposition layers and the threshold size are dynamically adjusted to ensure that the denoised signal has neither obvious noise residue nor loss of effective features. The process of step 5 is as follows: By calculating local singular entropy Identify the noise-dominant region.

[0023] in, p i For the first i The energy percentage of each singular value ; k The number of singular values ​​is equal to the number of sensors in this invention. k =4 (displacement, drilling speed, pressure, and torque 4-channel sensors); The i-th singular value is obtained through SVD decomposition and satisfies... ; When H sing When the noise level is less than 0.5, it is considered a low-noise environment, and standard noise reduction parameters are used. When H sing When the value is greater than 0.8, it is determined to be a strong noise environment, and the noise reduction intensity is automatically increased (e.g., by increasing the energy threshold α). This invention utilizes H sing Dynamically adjust the SVD energy threshold α and wavelet threshold λ j It enables intelligent noise reduction that adapts to the noise environment.

[0024] In this embodiment, for a sandstone-mudstone interbedded strata exploration project in a certain area, there are three distinct stratigraphic interfaces (one sandstone-mudstone interface and two mudstone-shale interfaces). Due to the influence of drilling machinery vibration (10~50Hz low-frequency interference) and electromagnetic radiation from nearby high-voltage lines (100~500Hz high-frequency interference), the original signal-to-noise ratio (SNR) is only 8.5~10.2dB. Therefore, the method described in this patent is required to achieve noise reduction and feature extraction.

[0025] The hardware and sensor configurations are shown in Table 1 below: The implementation details and results of each step are as follows: Step 1: Signal Acquisition and Preprocessing; Original matrix construction: After continuous data acquisition for 30 minutes, m = 30 × 60 × 100 = 180,000 sampling points were generated, constructing an original signal matrix X of 180,000 × 4. Some of the original data is shown in Table 2 below (first 5 sampling points): Data standardization processing: Calculate the average value of each signal channel He Standard Deviation : Displacement: =23.78mm, =5.23; Drilling speed: =4.5m / min, =1.8m / min; Pressure: =9.2MPa, =2.1MPa; Torque: =1500 N·m, =300 N·m. After standardization The average value of each row is 0, and the difference is 1, thus eliminating the difference in dimensions.

[0026] Step 2: Singular Value Decomposition (SVD) Adaptive Hierarchical Noise Reduction, the process is as follows; (2.1) Block SVD decomposition Will Divide the data into 512×4 blocks (352 sub-blocks in total), and calculate the global singular values ​​using the LAPACK library: =132.6, =41.5, =15.2, =4.8.

[0027] (2.2) Double cutoff decision Energy percentage calculation criteria: ; Singular value difference calculation: Calculating the difference spectrum =132.6-41.5=91.1; =41.5 - 15.2 = 26.3; =15.2-4.8=10.4, median median( ) = 26.3, which satisfies =26.3=1.0*26.3 (take 26.3 as the base value) =1.0, adapted for weak noise scenarios), determine the cutoff point r=2.

[0028] (2.3) Low-rank matrix reconstruction Reconstructed matrix ,in, After reconstruction, the signal SNR increased from the original 10.2dB to 17.1dB, and drilling mechanical vibration noise was significantly suppressed.

[0029] Step 3, wavelet transform and adaptive thresholding, the process is as follows: (3.1) Wavelet decomposition parameters Select the Daubechies-4 (db4) wavelet basis, and press... In practice, J=5 levels of decomposition were used to obtain the low-frequency coefficients. and high frequency coefficient .

[0030] (3.2) Noise and threshold calculation (using the level 3 detail coefficient) (For example) Noise standard deviation: ; Adaptive threshold calculation: ; Shrinkage factor calculation: ,but .

[0031] (3.3) Semi-soft threshold processing right Applying the threshold function: After processing, high-frequency electromagnetic noise is suppressed by 65%, and the reflection peaks at the sandstone-mudstone interface are preserved intact (peak value attenuation ≤3%).

[0032] Step 4, Spatiotemporal Reconstruction and Stratigraphic Feature Extraction, the process is as follows: (4.1) Spatiotemporal domain reconstruction By using Mallat inverse transform fusion coefficients and combining them with SVD singular matrix optimization to reconstruct the matrix Y, the reconstructed signal fidelity (correlation coefficient with the noiseless reference signal) reaches 0.981, with a mean square error of [missing value]. .

[0033] (4.2) Feature extraction Reflected wave travel time: STFT analysis of the torque signal (Hann window length 1024 points, ), identify 3 time-frequency peaks: =8.2s, =15.6s, =22.3s.

[0034] Corresponding formation depth (P-wave velocity 2800 m / s): , , ; Amplitude attenuation coefficient: Take , =0.7, calculated as follows: =0.75, , =0.39, corresponding to lithology of sandstone (low attenuation) → mudstone (medium attenuation) → shale (high attenuation), consistent with the verification of drilling core samples.

[0035] Step 5, adaptive parameter adjustment, the process is as follows: (5.1) Singular entropy calculation: Energy percentage =0.986, =0.012, =0.0015, =0.0005, Singular entropy: H sing =-(0.986ln0.986+0.012ln0.012+0.0015ln0.0015+0.0005ln0.0005)≈0.048<0.5, therefore it is determined to be a weak noise environment, and the maintenance energy threshold is set. =0.95.

[0036] (5.2) Accuracy closed-loop verification Reconstructed MSE = 0.032 ≤ 0.05, no parameter adjustment required; if H in subsequent acquired signals... sing =0.85 (strong noise), automatically Increased to 0.98, threshold Increase by 15% to ensure noise reduction effect.

[0037] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for denoising and reconstructing multi-sensor signals for formation prospecting based on singular value decomposition, characterized in that, The method comprises the following steps: Step 1, signal acquisition and pretreatment: by deploying displacement, drilling speed, pressure, torque sensor array on the exploration drilling rig, the formation response signal is synchronously collected under a unified time reference, the space-time consistency of each signal is ensured, the dimension difference of different sensor signals is eliminated by standardizing the multi-sensor signal, and all signals are in the same order of magnitude Step 2, singular value decomposition SVD adaptive hierarchical denoising: the normalized multi-sensor signal matrix is blocked according to a fixed size, singular value decomposition SVD is performed on each block, singular values and singular vectors reflecting the principal components of the signal are extracted, effective singular values are screened through two standards of "energy proportion” and "singular value difference”, the signal matrix is reconstructed by using the screened effective singular values and corresponding singular vectors, the matrix rank number is reduced by discarding small singular values dominated by noise, the goal of "retaining effective signals and filtering random noise” is achieved, and a signal matrix after preliminary denoising is obtained; Step 3, wavelet transform and adaptive threshold processing: a wavelet basis suitable for non-stationary signals is selected, the decomposition layer number is determined according to the sampling point number, the signal is decomposed into "low-frequency approximation” and "high-frequency detail” two parts, the noise intensity is estimated by using the statistical characteristics of high-frequency detail coefficients, the wavelet threshold is dynamically calculated according to the noise intensity, the sampling point number and the number of sensors, a shrinkage factor is introduced, a smooth threshold function is applied to the high-frequency detail coefficients, the coefficients exceeding the threshold are proportionally shrunk, and the coefficients below the threshold are zeroed; Step 4, time-space domain signal reconstruction and formation feature extraction: the processed low-frequency approximation coefficients and high-frequency detail coefficients are combined through inverse wavelet transform to restore the complete waveform of the time domain signal; the reconstructed signal matrix is optimized by combining the global features of SVD and the local features of wavelet transform; the time-frequency distribution of the reconstructed signal is analyzed by short-time Fourier transform, the energy peak position of the formation interface reflection wave is identified, and the reflection wave travel time is determined; the energy attenuation degree of the reflection wave in the propagation process is quantified, and the formation lithology difference is inverted combined with the geological prior knowledge; Step 5, adaptive parameter adjustment: the proportion of noise in the signal is evaluated by calculating the local singular entropy; the error between the reconstructed signal and the original signal is checked, the wavelet decomposition layer number and the threshold size parameter are dynamically adjusted, and it is ensured that the denoised signal has neither obvious noise residue nor effective feature loss.

2. The singular value decomposition based formation prospecting multi-sensor signal denoising reconstruction method according to claim 1, characterized in that, In step 1, the displacement, drilling speed, pressure, torque multi-source sensor signals are synchronously collected to construct m x n The observation matrix X is observed, and the multi-source signal space-time domain fusion model is inputted for zero-mean and standardization preprocessing.

3. The singular value decomposition based formation prospecting multi-sensor signal denoising reconstruction method according to claim 2, characterized in that, In step 2, singular value decomposition is performed on the pretreated matrix X' to calculate the singular value cumulative energy ratio E(r) and the difference spectrum , and the optimal cutoff point is determined based on double criteria r ; In step 2, the determination of the cut-off point r satisfies the following conditions: ； wherein, ∈ [0.9, 0.98], γ ∈ [1.2, 2.0] are tunable parameters.

4. The singular value decomposition based formation prospecting multi-sensor signal denoising reconstruction method according to claim 3, characterized in that, In step 3, the signal matrix Xr reconstructed by singular value decomposition SVD is subjected to wavelet decomposition, and each layer of high-frequency detail coefficients is subjected to semi-soft threshold processing based on SURE estimation.

5. The singular value decomposition based formation prospecting multi-sensor signal denoising reconstruction method according to claim 4, characterized in that, In step 3, the wavelet transform adopts Daubechies wavelet basis, and the decomposition layer number J is dynamically determined: ； wherein m is the signal length; The semi-soft threshold function is: ； in, These are the wavelet detail coefficients after thresholding. λ represents the original wavelet detail coefficients, i.e., the i-th coefficient at the j-th level; sgn() is the sign function, returning the sign of the input value; λ j For the first j The threshold of the layer is calculated using the following formula: ,in, Here, m is the noise standard deviation estimate, m is the signal length, and n is the number of sensors. β j For adaptive adjustment parameters, they are defined as follows: ,in For the first j Layer detail factor Norm square; This is an indicator function that takes the value 1 when the condition is met and 0 otherwise.

6. The singular value decomposition based formation prospecting multi-sensor signal denoising reconstruction method according to claim 5, characterized in that, In step 4, the formation features are extracted from the restructured signal, and the travel time t of the reflected wave is calculated by time-frequency analysis r and the amplitude attenuation coefficient A t : ； ； where STFT is a short-time Fourier transform, is a center frequency, is a reference time point, and a is a decay factor.

Images