A method and system for detecting intelligent attenuation compensation of a signal

Abstract

This application relates to the field of signal processing technology and provides a method and system for intelligent attenuation compensation of detection signals. The method includes: performing multi-scale frequency band decomposition on the original electromagnetic detection signal to obtain low-frequency sub-signals, mid-low-frequency sub-signals, mid-high-frequency sub-signals, and high-frequency sub-signals; constructing a comprehensive attenuation weight matrix for the deep medium; generating an attention constraint mask matrix for the deep medium based on the comprehensive attenuation weight matrix; extracting features from the low-frequency sub-signals, mid-low-frequency sub-signals, mid-high-frequency sub-signals, and high-frequency sub-signals respectively, and fusing the extracted features to obtain fused features; and compensating the original electromagnetic detection signal based on the fused features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal. This application can improve the accuracy of electromagnetic detection signal attenuation compensation.

Description

Technical Field

[0001] This application belongs to the field of signal processing technology, and in particular relates to a method and system for intelligent attenuation compensation of detection signals. Background Technology

[0002] Transient electromagnetic detection technologies, such as electromagnetic wave detection, are core techniques for non-destructive testing of deep underground targets, geological structure exploration, and municipal infrastructure inspection. When electromagnetic waves propagate through lossy underground media, the signal energy attenuates exponentially with increasing depth due to absorption and scattering effects, especially attenuating high-frequency components. This results in weak signal energy, poor phase axis continuity, and a narrowed effective bandwidth for deep targets, severely limiting the resolution and imaging accuracy of deep detection. Signal attenuation compensation is a crucial technology for improving the quality of electromagnetic detection data and enhancing deep detection capabilities.

[0003] Existing electromagnetic detection signal attenuation compensation techniques are mainly divided into two categories: traditional signal processing methods and deep learning methods. Specific technical solutions and limitations are as follows: (I) Traditional Attenuation Compensation Methods Traditional electromagnetic detection signal attenuation compensation methods often draw on mature technologies from seismic signal processing. This is because electromagnetic waves and seismic waves have similar propagation characteristics in the underground environment. Mainstream methods include spectral whitening, deconvolution, and inverse Q filtering.

[0004] 1. The spectral whitening method achieves high-frequency attenuation compensation by broadening the signal amplitude spectrum. This method enhances the resolution of non-stationary signals through amplitude-frequency partitioning. It is simple to operate and has high computational efficiency. However, it ignores the relative distribution relationship between phase information and amplitude, which can easily cause signal phase distortion. The geological interpretation reliability of the compensated signal is insufficient. 2. Deconvolution methods improve signal resolution through wavelet compression and are a classic method for improving resolution in seismic and radar signal processing. Among them, robust wavelet estimation and blind deconvolution methods have been widely used in seismic and radar signal processing. However, their performance is highly dependent on accurate wavelet estimation and must meet strict theoretical assumptions. In subsurface non-homogeneous media scenarios, the accuracy of wavelet estimation drops significantly, and the method has poor adaptability. 3. The inverse Q-filtering method improves the continuity of the phase axis by correcting amplitude and phase distortion, and is currently the most widely used traditional method. The stable Q-analysis method based on vertical seismic profile data provides the core theoretical support for inverse Q-filtering. The subsequent development of inversion-based time-domain inverse Q-filtering method has further improved the effect of seismic resolution enhancement. However, this type of method is extremely sensitive to noise and requires accurate estimation of Q-values ​​of the subsurface medium. In deep and complex medium scenarios, the Q-values ​​have strong spatial heterogeneity and are difficult to obtain, which directly leads to unstable compensation effect and even noise amplification. Existing improved methods have not yet completely solved the stability problem in low signal-to-noise ratio scenarios.

[0005] The aforementioned traditional methods are generally limited by theoretical assumptions and reliance on prior parameters. In complex scenarios such as deep non-uniform media and temperature-pressure coupling, it is difficult to balance the accuracy and stability of attenuation compensation, thus limiting their practical application effectiveness.

[0006] (II) Deep Learning Attenuation Compensation Method In recent years, deep learning technology, with its powerful nonlinear mapping capabilities, has been widely applied in the field of geophysical signal processing, achieving superior results compared to traditional methods in scenarios such as ground-penetrating radar (GPR) subsurface target identification, seismic signal noise attenuation, geophysical data reconstruction, and GPR data inversion. Regarding the problem of electromagnetic signal attenuation compensation, scholars both domestically and internationally have also conducted research on related deep learning methods. 1. The fusion of deep neural networks (FDNN) enables the simultaneous processing of Q-value estimation and seismic signal attenuation compensation, breaking away from the strong dependence of traditional methods on prior Q-values ​​and achieving end-to-end attenuation compensation; 2. A spectral broadening method based on the U-Net architecture achieves synchronous correction of seismic signal amplitude and phase distortion through an encoder-decoder structure, with compensation accuracy and generalization performance significantly better than the traditional inverse Q filtering method; 3. For GPR signal attenuation compensation, the Spatiotemporal Neural Network (STNN) achieves attenuation compensation by capturing the spatiotemporal characteristics of the GPR signal. It breaks away from the theoretical assumptions of traditional methods and has better noise resistance and compensation effect than traditional methods. It is currently the mainstream deep learning method in the field of GPR attenuation compensation.

[0007] However, existing deep learning attenuation compensation methods still face three major technical bottlenecks: 1. Lacking strict physical constraints, existing methods mostly adopt pure data-driven end-to-end mapping, without incorporating physical priors such as the skin effect of electromagnetic wave propagation and medium attenuation, which easily produces false compensation results that do not conform to physical laws, and the geological interpretation of the compensation results is not reliable enough. 2. It is not well adapted to the complex attenuation characteristics of deep media. It does not take into account the deep temperature and pressure coupling effect and the heterogeneity of attenuation law caused by the non-uniform scattering of the medium. It cannot adapt to the nonlinear attenuation characteristics of deep multi-band signals and has a serious deficiency in the ability to recover high-frequency components of weak deep signals. 3. Lack of differentiated feature enhancement mechanisms: Existing networks mostly adopt a globally unified feature processing paradigm, which cannot perform differentiated compensation for signal regions with different depths and attenuation levels. This easily leads to overcompensation of shallow signals and undercompensation of deep signals, resulting in insufficient robustness in low signal-to-noise ratio scenarios.

[0008] In summary, current electromagnetic detection signal attenuation compensation methods suffer from low accuracy. Summary of the Invention

[0009] This application provides a method and system for intelligent attenuation compensation of detection signals, which can solve the problem of low accuracy in electromagnetic detection signal attenuation compensation.

[0010] In a first aspect, embodiments of this application provide a method for intelligent attenuation compensation of detection signals, including: The original electromagnetic detection signal is decomposed into low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band sub-signals, and high-frequency band sub-signals by multi-scale decomposition. Construct a comprehensive attenuation weight matrix for deep media; the comprehensive attenuation weight matrix is ​​used to describe the full-dimensional attenuation distribution information of electromagnetic wave propagation in deep media; An attention constraint mask matrix for deep media is generated based on a comprehensive attenuation weight matrix. The attention constraint mask matrix is ​​used to describe the spatial distribution of attenuation regions and compensation priorities of electromagnetic signals in deep media. Features of low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band sub-signals, and high-frequency band sub-signals are extracted separately, and the extracted features are fused to obtain fused features; The original electromagnetic detection signal is compensated based on the fusion features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal.

[0011] Secondly, embodiments of this application provide an intelligent signal attenuation compensation system, comprising: The decomposition module is used to perform multi-scale frequency band decomposition on the original electromagnetic detection signal to obtain low-frequency band sub-signals, mid-low frequency band sub-signals, mid-high frequency band signals, and high frequency band signals; The module is used to construct the comprehensive attenuation weight matrix for deep media; the comprehensive attenuation weight matrix is ​​used to describe the full-dimensional attenuation distribution information of electromagnetic wave propagation in deep media; The generation module is used to generate an attention constraint mask matrix for deep media based on the comprehensive attenuation weight matrix; the attention constraint mask is used to describe the spatial distribution of attenuation regions and compensation priorities of electromagnetic signals in deep media. The extraction module is used to extract features from low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band signals, and high-frequency band signals respectively, and to fuse the extracted features to obtain fused features; The compensation module is used to compensate the original electromagnetic detection signal based on the fused features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal.

[0012] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, the original electromagnetic detection signal is first decomposed into low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band sub-signals, and high-frequency band sub-signals. Then, features of the low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band sub-signals, and high-frequency band sub-signals are extracted using a deep network. The extracted features are then fused. Finally, the original electromagnetic detection signal is compensated based on the constructed attention constraint mask matrix and the fused features to obtain the compensated electromagnetic detection signal. The deep network in this application is a multi-scale deep network architecture, which can achieve differentiated extraction and accurate adaptation of multi-band attenuation features, improving the recovery capability of high-frequency components of weak signals in deep media. In addition, since the attention constraint mask matrix can describe the spatial distribution of attenuation regions of electromagnetic signals in deep media, it can provide intrinsic physical constraints for the deep learning compensation process, ensuring the physical rationality of the compensation results and the reliability of geological interpretation. At the same time, the attention constraint mask matrix can also guide the deep network to perform focused and differentiated compensation for signal regions with different depths and attenuation levels. Therefore, compared with existing electromagnetic detection signal attenuation compensation methods, the method in this application can effectively improve the accuracy of electromagnetic detection signal attenuation compensation.

[0013] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a method for intelligent attenuation compensation of detection signals provided in an embodiment of this application. Detailed Implementation

[0016] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0017] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0018] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0019] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0020] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0022] To address the problem of low accuracy in electromagnetic detection signal attenuation compensation, this application provides an intelligent attenuation compensation method for detection signals. This method first decomposes the original electromagnetic detection signal into low-frequency, mid-low-frequency, mid-high-frequency, and high-frequency sub-signals. Then, it uses a deep network to extract features from these sub-signals, fuses the extracted features, and finally compensates the original electromagnetic detection signal based on the constructed attention-constrained mask matrix and the fused features to obtain the compensated electromagnetic detection signal. The deep network in this application is a multi-scale deep network architecture, which can achieve differentiated extraction and accurate adaptation of multi-band attenuation features, improving the recovery capability of high-frequency components of weak signals in deep media. In addition, since the attention constraint mask matrix can describe the spatial distribution of attenuation regions of electromagnetic signals in deep media, it can provide intrinsic physical constraints for the deep learning compensation process, ensuring the physical rationality of the compensation results and the reliability of geological interpretation. At the same time, the attention constraint mask matrix can also guide the deep network to perform focused and differentiated compensation for signal regions with different depths and attenuation levels. Therefore, compared with existing electromagnetic detection signal attenuation compensation methods, the method in this application can effectively improve the accuracy of electromagnetic detection signal attenuation compensation.

[0023] The intelligent attenuation compensation method for detection signals provided in this application will be illustrated below with reference to specific embodiments.

[0024] like Figure 1 As shown in the embodiments of this application, the intelligent attenuation compensation method for detection signals includes the following steps: Step 11: Perform multi-scale frequency band decomposition on the original electromagnetic detection signal to obtain low-frequency band sub-signals, mid-low frequency band sub-signals, mid-high frequency band sub-signals, and high-frequency band signals.

[0025] The aforementioned raw electromagnetic detection signal is the raw deep electromagnetic detection signal, which can be ground-penetrating radar profile data or single-channel A-scan time-domain signal. Understandably, to facilitate data processing, after acquiring the raw electromagnetic detection signal, it can first undergo background removal, DC component correction, and gain normalization preprocessing to eliminate system errors and background clutter interference, resulting in a preprocessed standardized signal matrix. (M represents the number of sampling channels, and N represents the number of sampling points per channel); then, wavelet packet transform is used to perform multi-scale frequency band decomposition on the preprocessed signal, decomposing it into four different frequency bands: low-frequency band, mid-low-frequency band, mid-high-frequency band, and high-frequency band (i.e., low-frequency band sub-signal, mid-low-frequency band sub-signal, mid-high-frequency band sub-signal, and high-frequency band sub-signal). This captures different frequency components of the deep signal, adapts to the differentiated attenuation patterns of different frequency bands, and provides a foundation for subsequent multi-scale network feature extraction. The low-frequency band sub-signal refers to signals with frequencies of 0–200 MHz, the mid-low-frequency band sub-signal refers to signals with frequencies of 200–500 MHz, the mid-high-frequency band sub-signal refers to signals with frequencies of 500–800 MHz, and the high-frequency band sub-signal refers to signals with frequencies of 800–1000 MHz.

[0026] Step 12: Construct a comprehensive attenuation weight matrix for the deep medium; this comprehensive attenuation weight matrix is ​​used to describe the full-dimensional attenuation distribution information of electromagnetic wave propagation in the deep medium.

[0027] The aforementioned comprehensive attenuation weight matrix provides a physical basis for subsequent attenuation level classification, attention mask construction, and physical consistency loss calculation. This matrix integrates the triple attenuation laws of depth dimension, spatial channel order dimension, and frequency dimension, characterizing the comprehensive attenuation degree of each spatial sampling point, each depth layer, and each sub-band. It integrates the three major physical components of basic skin attenuation, thermo-baric coupling attenuation, and dielectric scattering attenuation, realizing point-by-point, frequency-by-frequency, and layer-by-layer quantization of attenuation information.

[0028] The aforementioned deep media specifically refers to underground rock and soil media in the depth range of 30m to 200m below the surface under shallow surface high-frequency electromagnetic detection scenarios.

[0029] The above comprehensive attenuation weight matrix It is a three-dimensional tensor with dimension . ;in, This indicates the number of spatial sampling channels for the data, corresponding to the underground space sampling dimension. The value is determined by the total length of the survey lines and the spacing between channels. Each index corresponds to a spatial sampling location. This represents the number of time-domain sampling points per channel, corresponding to the longitudinal depth (i.e., the depth dimension), which determines the sampling accuracy and longitudinal resolution in the depth direction. This indicates the number of sub-bands in the wavelet packet decomposition, corresponding to the signal frequency dimension, which is perfectly aligned with the input signal dimension. (This is from an embodiment of the application.) Corresponding to the four sub-bands of low, medium-low, medium-high, and high frequency, it is the core dimension for realizing multi-band differentiated attenuation compensation.

[0030] In some embodiments of this application, a comprehensive attenuation weight matrix can be obtained by fusing prior modeling of attenuation physics based on thermo-pressure coupling and non-uniform scattering. Specifically, the comprehensive attenuation weight matrix... Matrix elements Indicates the first Dao, Di The depth sampling point, the first The cumulative attenuation weight of electromagnetic waves in each sub-band , , The matrix elements for: ; Among them, the depth of dissemination , This represents the average velocity of electromagnetic waves in the Earth's crust. Indicating two-way travel, Indicates depth Place, No. The comprehensive attenuation constant corresponds to each sub-band. It's worth noting that path integrals are used to achieve precise quantization of the entire attenuation range, replacing the traditional single-point approximation algorithm. The value ranges from 0 to 1, and the smaller the value, the stronger the signal attenuation.

[0031] ; = ; = ; ; = ; ; in, Indicates depth Place, No. The fundamental attenuation constant corresponding to each sub-band Indicates depth Place, No. The temperature-pressure coupling attenuation factor corresponding to each sub-band. Indicates depth Place, No. The scattering attenuation constant corresponding to each sub-band Angular frequency represents the oscillation rate of an electromagnetic wave signal. Permeability represents the ability of a deep medium to conduct magnetic fields. Indicates as depth Changing dielectric parameters, Indicates as depth Sub-band Changing dynamic conductivity, Represents the relative permittivity of the Earth's surface matrix. This represents the depth gradient correction coefficient. This is used to quantify the rate of change of the electromagnetic parameters of the medium with depth, enabling point-by-point dynamic updates of the attenuation constant with depth and frequency, and providing accurate physical weights for subsequent attention-constrained masks. Indicates dynamic conductivity. This represents the baseline conductivity of the Earth's surface at normal temperature and pressure. This represents the dispersion matching coefficient that adaptively adjusts with the signal frequency. This represents the activation energy of thermo-baric coupling. Indicates formation temperature. Indicates formation pressure. This represents the ideal gas constant, enabling normalized conversion between energy units and temperature units. Indicates the formation porosity compressibility. Represents the dielectric contrast scattering coefficient. Represents the base scattering coefficient of the formation. The dielectric constant of a porous fluid. The dielectric constant of the soil matrix is ​​represented by the dielectric constant of the soil matrix. This represents the depth gradient correction coefficient.

[0032] in, Indicates as depth Changing dielectric parameters: ; Represents the relative permittivity of the Earth's surface matrix. Represents the dielectric constant and depth attenuation coefficient. Represents the relative permittivity of pore fluid. Indicates depth The porosity of the formation changes dynamically with temperature and pressure gradients.

[0033] Indicates as depth Sub-band Changing dynamic conductivity: ; This represents the baseline conductivity of the Earth's surface at normal temperature and pressure. This represents the activation energy of thermo-baric coupling. Represents the gas constant. , Representing depth Temperature and pressure, Indicates the formation porosity compressibility. Indicates the current sub-band center frequency. Indicates the reference frequency. The depth-dependent dispersion index represents the strength of the change in conductivity with frequency.

[0034] Indicates the depth gradient correction coefficient: ; This represents the empirical proportionality coefficient. This is the baseline value for the electrical conductivity of the Earth's surface at normal temperature and pressure. This formula is used to reflect the gradient change of electromagnetic parameters with increasing depth.

[0035] This represents the dispersion matching coefficient that adaptively adjusts with the signal frequency. ; Indicates the reference dispersion matching coefficient; Indicates the current sub-band center frequency; This represents the reference frequency for the entire signal. This formula provides stronger temperature and pressure attenuation correction for high-frequency components, conforming to the deep electromagnetic dispersion characteristics.

[0036] Indicates the activation energy of thermo-baric coupling: ; This represents the baseline value of the activation energy at room temperature and pressure. This represents the temperature correction factor; Indicates the pressure correction factor; Indicates formation temperature. This represents the formation pressure. This formula reflects the coupled relationship between activation energy and increasing temperature and pressure.

[0037] Indicates the depth gradient correction coefficient: ; Indicates the benchmark coefficient; Indicates as depth The changing dielectric parameter is the depth. The relative permittivity; It serves as a reference value for the dielectric constant of the Earth's surface. It is used to describe the gradient variation of scattering and attenuation with increasing depth.

[0038] It should be noted that, Indicates depth Place, No. The fundamental attenuation constant corresponding to each sub-band can be understood as the fundamental attenuation model of the skin effect. This fundamental attenuation constant is derived based on the propagation theory of electromagnetic waves in lossy media, considering the fusion of thermo-baric coupling and non-uniform scattering in deep media. The specific calculation process is as follows: First, starting from Maxwell's curl equation for lossy media, a depth-dependent dynamic complex permittivity is introduced. ,in, To determine the dynamic conductivity that varies with depth and frequency, the dispersion characteristics of the medium parameters caused by the deep temperature and pressure gradient are coupled. Then, a depth-direction Fourier transform is performed on the wave equation to decouple the propagation and attenuation terms, deriving the fundamental attenuation constant in both depth and frequency dimensions. .

[0039] :depth ,frequency The complex permittivity at a given point describes the electromagnetic energy storage and loss characteristics of a lossy medium. :depth The relative permittivity at a given point reflects the energy storage capacity of the medium.

[0040] Vacuum permittivity. Imaginary unit, representing the imaginary part of the loss in the complex permittivity. :depth ,frequency The dynamic conductivity at a given point characterizes the conductive loss of the dielectric. Angular frequency , describes the oscillation rate of electromagnetic waves.

[0041] The above-mentioned basic attenuation constant Compared to the classical fundamental decay constant, the following significant differences exist: Introducing the deep dynamics of dielectric parameters: the fixed dielectric constant in classical formulas... Electrical conductivity Improved to a dynamic parameter that changes continuously with depth. , This aligns with the spatial heterogeneity of medium parameters caused by temperature and pressure gradients and changes in pore structure in deep formations; Achieving two-dimensional coupling of frequency and depth: breaking through the limitations of classical formulas that only calculate in the single frequency dimension, allowing conductivity... Simultaneously coupling depth and frequency characteristics, and introducing a depth gradient correction coefficient, the following derivation is obtained. This enables dynamic calculation of the attenuation constant, which adapts to depth and frequency point by point, rather than the globally fixed value of the classical method.

[0042] Generally, the attenuation constant includes: the complex permittivity of the lossy medium, the propagation constant, and the fundamental attenuation constant.

[0043] The aforementioned temperature-pressure coupling attenuation factor is obtained by correcting the temperature-pressure coupling effect attenuation factor. Specifically, it addresses the changes in electromagnetic parameters of the medium caused by the coupling effect of temperature and pressure in deep formations, introducing a temperature-pressure coupling attenuation factor to obtain the corrected attenuation constant. The calculation process of this temperature-pressure coupling attenuation factor is as follows: First, a dynamic conductivity model under temperature-pressure coupling is established: the conductivity of deep formations is dominated by the conductivity of pore fluids; temperature-pressure coupling changes the viscosity and ion activity of pore water, leading to the derivation of... ,in The activation energy for thermo-baric coupling, The formation porosity compressibility coefficient, The baseline conductivity at ambient temperature and pressure is defined; then, a dispersion-adaptive temperature-pressure coupling attenuation factor is defined. = .

[0044] It is worth mentioning that, from The calculation formula shows that... It possesses the characteristics of nonlinear coupling due to both temperature and pressure factors, integration with the structural properties of porous media, and deep dynamic correlation. The nonlinear coupling due to both temperature and pressure factors refers to the activation energy of the temperature-pressure coupling. Replacing the traditional fixed activation energy, it accurately characterizes the changes in ion activity under the synergistic effect of temperature and pressure, rather than a simple linear superposition of "temperature + pressure"; incorporating the structural characteristics of porous media means: introducing the formation pore compressibility coefficient, binding the influence of pressure on conductivity with the deformation of deep pore structure, so that the change in conductivity fits the actual physical properties of the rock and soil, rather than the traditional abstract linear correction; deep dynamic correlation means: all parameters in the model ( , All of these can be adaptively adjusted with depth, enabling three-dimensional dynamic calculation of conductivity with depth, temperature, and pressure, rather than the traditional global fixed value calculation, thus perfectly matching the gradient change law of deep exploration.

[0045] As can be seen from the calculation formula of the temperature-pressure coupling attenuation factor, the temperature-pressure coupling attenuation factor binds the attenuation factor to the signal frequency. The temperature-pressure coupling attenuation factor increases non-linearly with frequency, allowing high-frequency signals to receive higher temperature-pressure correction weights while maintaining moderate correction for low-frequency signals, thus achieving precise attenuation compensation based on frequency differences. Simultaneously, the temperature-pressure coupling attenuation factor utilizes the dynamic conductivity of temperature-pressure coupling. As the core variable, it achieves three-dimensional linkage correction of "temperature, pressure, conductivity, and frequency," rather than a single factor unrelated to frequency as in traditional methods; all parameters of the factor have a clear physical correspondence. Corresponding frequency sensitivity, (Corresponding to the dielectric conductivity characteristics under temperature and pressure), avoiding the arbitrariness of traditional purely empirical correction factors, and ensuring that the attenuation correction fully conforms to the physical laws of electromagnetic wave propagation.

[0046] The above scattering attenuation constant yes This is obtained through correction using a scattering attenuation factor for media inhomogeneity. Specifically, it addresses the electromagnetic wave scattering attenuation caused by inhomogeneous underground media, introducing a scattering attenuation constant based on Rayleigh scattering theory, where the attenuation constant is proportional to the fourth power of the frequency. The calculation process for this scattering attenuation constant is as follows: First, a depth-dependent equivalent scattering unit is defined: based on the solid-liquid two-phase structure of deep porous media, an equivalent scattering particle size that varies with depth is introduced. The porosity compressibility is corrected by the temperature-pressure coupling to satisfy the following conditions. (Electromagnetic wave wavelength), strictly conforming to the applicable boundaries of Rayleigh scattering; then, the dielectric contrast scattering coefficient is introduced. Finally, the scattering attenuation constant is derived: = It retains the core physical law that Rayleigh scattering is proportional to the fourth power of the frequency, while introducing a depth gradient correction coefficient. It adapts to the increasing characteristics of scattering intensity in deep strata with depth, and realizes point-by-point dynamic calculation of scattering attenuation.

[0047] The formula for calculating the dielectric contrast scattering coefficient reveals that it precisely quantifies the physical meaning of the medium: it directly links the scattering coefficient to the dielectric contrast between the solid and liquid phases. Precise quantification of the dielectric constant difference between pore fluids and soil matrices transforms the scattering coefficient from an "empirical value" into a quantifiable parameter with clear physical meaning, perfectly aligning with the scattering nature of deep porous media. Simultaneously, the dielectric contrast scattering coefficient can achieve dynamic depth adaptation: the scattering coefficient is designed as a dynamic parameter that varies with depth. , All of them can adaptively adjust according to the deep temperature and pressure gradient and pore structure deformation, replacing the traditional global fixed coefficient, realizing the depth-by-depth point-by-point calculation of the scattering coefficient, and matching the real law that the deep scattering attenuation increases with depth.

[0048] From the formula for calculating the scattering attenuation constant, it can be seen that the scattering coefficient is a dynamic scattering coefficient: replacing the globally fixed coefficient in the classical formula with the depth-dynamic dielectric contrast scattering coefficient. This approach binds the scattering attenuation constant to the depth of the dielectric inhomogeneity of the deep medium, enabling a physical quantitative calculation of scattering attenuation rather than empirical fitting. Simultaneously, a depth gradient correction term is introduced into the scattering attenuation constant: a new depth gradient correction coefficient is added. ,pass The quantitative analysis of the increasing scattering attenuation with depth (increased formation pressure, denser pore structure, and more significant dielectric inhomogeneity in deeper formations lead to stronger scattering attenuation) allows for dynamic, point-by-point updates of the scattering attenuation constant, overcoming the limitation of the classical formula's "globally fixed value." Furthermore, the scattering attenuation constant retains physical laws and expands its applicability: while strictly preserving the core law of Rayleigh scattering being "proportional to the fourth power of the frequency," dynamic depth correction allows the Rayleigh scattering model, originally applicable only to shallow homogeneous media, to be adapted to real-world detection scenarios in deep, non-homogeneous porous media, achieving an engineering extension of classical theory.

[0049] Step 13: Generate an attention constraint mask matrix for the deep medium based on the comprehensive attenuation weight matrix; the attention constraint mask is used to describe the spatial distribution of the attenuation region and the compensation priority of the electromagnetic signal in the deep medium.

[0050] The aforementioned attention constraint mask matrix precisely divides the effective compensation region and the ineffective background region in binary form: a value of 1 represents a high-attenuation target region that requires focused compensation; a value of 0 represents a background region with no or weak attenuation, which does not require focused optimization by the network. After optimization by opening operations of the ellipsoidal three-dimensional structuring element, the mask matrix filters out false noise and retains continuous attenuation characteristics. It transforms the quantized attenuation information of the comprehensive attenuation weight matrix into a hard constraint that can be embedded in the cross-attention module of the decoder, guiding the network to focus on high-attenuation regions for precise compensation, improving compensation accuracy and spatial continuity, and ensuring that network training conforms to physical laws.

[0051] The above attention constraint mask matrix Matrix elements This represents the mask value optimized by three-dimensional morphological opening operations. This matrix element... for: ; ; ; in, Represents an ellipsoidal three-dimensional structural element. , express The local neighborhood offset, Fixed value It fully covers the 3×3×3 neighborhood space. The third morphological erosion operation represents the result of the three-dimensional erosion operation. Dao, Di The depth sampling point, the first Sub-band mask values, Representing the ellipsoidal three-dimensional structural element in The weighting coefficient of location, This represents the mask value after the erosion operation. This indicates that after the erosion operation, the first... Dao, Di The depth sampling point, the first Sub-band mask values, Indicates the original mask in the neighborhood The value at that location, For binarizing the attention-constrained mask tensor, The dimension is perfectly aligned with the input signal. , For spatial sampling channel number, Number of depth sampling points per channel The number of sub-bands for wavelet packet decomposition is determined by elements taking only 0 (uncompensated region) / 1 (compensated region). Indicates the first Dao, Di The depth sampling point, the first The original binary attention mask value of the sub-band. The attenuation weighting benchmark matrix Matrix elements Indicates the first Dao, Di The overall attenuation level of each depth sampling point; , Indicates the first The proportion of signal energy in each sub-band This represents the energy-weighted adaptive threshold. , Indicates the baseline threshold. Represents the total energy of the signal. The condition within the brackets is Iverson. The output is 1 if the condition inside the brackets is true, and 0 otherwise. The output is strictly guaranteed to be a binary tensor.

[0052] The above-mentioned ellipsoidal three-dimensional structural element This is a morphological neighborhood constraint kernel, a morphological computational neighborhood kernel that conforms to the physical laws of electromagnetic detection. Its basic size is 3×3×3, and its three axes correspond one-to-one with the three dimensions of the mask tensor. The dimensions are defined as follows: First dimension: Corresponding mask sequence ( The first dimension matches the continuity of the transverse spatial sampling of the survey line; the second dimension corresponds to the mask depth. The third dimension matches the continuous attenuation pattern of electromagnetic waves in the depth direction; the fourth dimension corresponds to the mask frequency band. ( ) dimension, matching the attenuation correlation of adjacent sub-bands.

[0053] In some embodiments of this application, after obtaining the above-mentioned attenuation weight matrix... Then, by using a weighted average of the main energy frequency bands, the three-dimensional tensor is mapped into a two-dimensional space-depth attenuation weighted reference matrix that is perfectly aligned with the dimensions of the input signal profile. Each element in the baseline matrix Representing the Dao, Di The overall attenuation level of each depth sampling point. From the attenuation weight baseline matrix. As can be seen from the calculation formula, the attenuation weighting benchmark matrix can accurately match the multi-band attenuation characteristics of electromagnetic detection: the weighting factor is set to the signal energy proportion of each sub-band. This allows the main frequency band, which has a higher energy content, to have a higher weight in the attenuation weighting fusion process, resulting in a fused frequency band. It accurately reflects the attenuation level of a signal during actual energy propagation, rather than a simple arithmetic average without physical meaning, thus avoiding interference from the attenuation weights of secondary frequency bands in determining the true attenuation level. Simultaneously, it achieves a precise mapping from a three-dimensional tensor to a two-dimensional physical region: the formula is specific to the system constructed in this invention. The design of the three-dimensional attenuation weight matrix is ​​achieved by adjusting the frequency band dimension ( The weighted summation of the attenuation weights achieves dimensionality reduction from the three-dimensional tensor of "depth-channel order-frequency band" to the two-dimensional spatial-depth reference matrix of "depth-channel order". Furthermore, the dimensionality-reduced reference matrix is ​​perfectly aligned with the physical profile dimensions of the GPR probe signal, solving the core problem that multi-band attenuation weights cannot be directly used for signal physical region classification. This is a customized requirement that traditional energy weighting formulas have not considered and cannot fulfill. In addition, the spatial-depth point-by-point characteristics of the attenuation weights are preserved: the formula applies weights to each spatial-depth point (…). Independently perform weighted fusion along the frequency band dimension, and in the final baseline matrix, each element... Each point corresponds uniquely to a physical sampling point of the signal, perfectly preserving the point-by-point attenuation characteristics of the depth-channel order in the three-dimensional attenuation weight matrix. This allows subsequent attenuation level division to achieve pixel-by-pixel accurate determination of the physical region of the signal, breaking through the limitation of traditional weighted fusion which easily loses local details.

[0054] In some embodiments of this application, based on the attenuation weight benchmark matrix The element-by-element threshold determination divides the spatial-depth region of the signal into four levels according to the degree of attenuation. Each region corresponds to a different depth distribution and compensation priority. The division rules are as follows: Slight attenuation region: meets the requirements For shallow, low-attenuation strata, the signal integrity is good, and the compensation priority is the lowest. Moderate attenuation region: meets the requirements Corresponding to the shallow to medium attenuation strata, the signal exhibits slight distortion, and the compensation priority is medium. Severe attenuation region: meets the requirements For medium-deep high-attenuation strata, the high-frequency components of the signal are severely lost, so compensation is of high priority. Extremely high attenuation region: meets the requirements For deep, highly attenuated strata, where effective signals are drowned out by noise, compensation has the highest priority.

[0055] Based on the attenuation level division results, a binarized attention constraint mask tensor that is perfectly aligned with the input signal size is generated. The four channels correspond to four attenuation level regions, and the mask weight setting rule is: the higher the degree of attenuation, the higher the mask weight value. That is, the weight of the extremely attenuated region is 1.0, the attenuated region is 0.8, the attenuated region is 0.5, and the attenuated region is 0.2.

[0056] In some embodiments of this application, the binarized attention constraint mask tensor This ensures a one-to-one correspondence between tensor dimensions and pixel coordinates, achieving complete alignment between physical dimensions and sampling intervals.

[0057] To facilitate the determination of the binarized attention constraint mask tensor Set an energy-weighted adaptive threshold for each sub-band. , , Indicates the baseline threshold. Represents the total energy of the signal. Binarized attention-constrained mask tensor. Matrix elements The area marked with 1 represents the attenuation region that requires significant compensation. The normal area is marked with 0, enabling the generation of multi-band differentiated masks.

[0058] After obtaining the binarized attention constraint mask tensor Subsequently, 3D morphological opening operations and connected component analysis are used to smooth and optimize the mask, eliminating isolated mask noise, ensuring the spatial continuity and depth-direction monotonicity of the mask, and avoiding discontinuities in compensation results caused by abrupt mask changes. The final optimized attention-constrained mask is then obtained to guide the differential feature enhancement and compensation of deep networks. Specifically, the attention-constrained mask matrix... The calculation process is as follows: First, adaptive 3D structuring element design for matching signal characteristics: An ellipsoidal 3D structuring element is designed to suit the 3D spatial characteristics of the mask tensor. The depth direction (N-axis) step size matches the depth sampling interval of the GPR signal, the frequency direction (F-axis) step size matches the attenuation correlation of adjacent sub-bands, and the spatial channel direction (M-axis) step size matches the inter-channel sampling interval. The structuring element weights are adaptively adjusted according to the depth-frequency correlation to avoid the over-smoothing problem of fixed structuring elements. Then, the core calculation is performed in steps: following the process of erosion followed by dilation, the erosion operation filters out isolated mask noise points through neighborhood minimum value calculation, and the calculation formula is: ; During the erosion operation, with As input, for each pixel in the core region, traverse the 3×3×3 neighborhood covered by the structuring element, calculate the minimum product of the mask value and the structuring element weight, and output 1 only when the mask of all valid weight positions in the neighborhood is 1, otherwise output 0, thus accurately filtering out isolated false mask noise.

[0059] The dilation operation restores the spatial continuity of the main mask region through neighborhood maximization. The calculation formula is as follows: ; During the expansion operation, the output is erosion. As input, a completely consistent structuring element is used to traverse the neighborhood, calculate the maximum product, and output 1 if there is a mask of 1 at an effective weight position in the neighborhood. This restores the three-dimensional spatial continuity of the effective attenuation region, while retaining the original values ​​in the edge region to avoid losing effective information. Finally, the optimized mask is output. That is, the final output is the attention constraint mask matrix. While denoising, the depth-frequency spatial continuity of the attenuation region is fully preserved.

[0060] It is worth mentioning that the embodiments of this application adopt a three-dimensional tensor step-by-step calculation formula, which adapts to the three-dimensional spatial characteristics of the mask and abandons the traditional two-dimensional coordinate calculation method. This is specifically designed for three-dimensional attention-constrained mask tensors. The order of the Tao ,depth ,frequency band The three-dimensional coordinates reconstruct the step-by-step calculation formulas for erosion and dilation, allowing the opening operation to complete neighborhood operations in three-dimensional space. This not only filters out isolated noise points in the three-dimensional mask (such as false mask points with single frequency band and single depth), but also fully preserves the spatial correlation between depth, frequency band and channel order of the three-dimensional mask. This is a core improvement that traditional two-dimensional opening operation formulas cannot achieve.

[0061] Step 14: Extract features from low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band sub-signals, and high-frequency band sub-signals respectively, and fuse the extracted features to obtain fused features.

[0062] In some embodiments of this application, a four-level parallel deep encoder architecture can be constructed to extract features of low-frequency band sub-signals, mid-to-low-frequency band sub-signals, mid-to-high-frequency band sub-signals, and high-frequency band sub-signals in parallel.

[0063] Specifically, the low-frequency band signal, the mid-low-frequency band signal, the mid-high-frequency band signal, and the high-frequency band signal can be input into four encoders in a one-to-one correspondence for feature extraction, so that each encoder can extract the features of its input data to obtain the features of the low-frequency band signal, the mid-low-frequency band signal, the mid-high-frequency band signal, and the high-frequency band signal.

[0064] The four encoders described above share the same structure. The backbone of this encoder employs a one-dimensional residual convolutional module. Each residual module contains two 1×3 convolutional layers, a batch normalization layer, and a ReLU activation layer. Downsampling is achieved through a max-pooling layer with a stride of 2. The number of convolutional kernels across the four layers gradually increases from 64 to 512, extracting low-dimensional to high-dimensional features from different frequency bands. A cross-channel attention mechanism is introduced into the encoder to achieve information interaction and fusion of features from different frequency bands, capturing the correlation of attenuation features across different frequency bands. The output is a high-dimensional feature map that integrates global information from multiple frequency bands with local details from a single frequency band, providing a feature foundation for subsequent compensation decoding.

[0065] In some embodiments of this application, the encoder performs feature extraction as follows: ; ; ; in, Representing residual characteristics, This represents a one-dimensional convolution operation. This indicates batch normalization operations. Represents the ReLU activation function. This indicates one of the following: low-frequency band signal, mid-low-frequency band signal, mid-high-frequency band signal, and high-frequency band signal. For the encoder input, , For batch size, Number of channels (corresponding to multi-band feature dimensions) The length of the signal depth sequence. express One-dimensional convolution operation, To achieve dimension-matched convolution, ensuring dimension alignment of residual connections, this addresses the gradient vanishing problem in deep encoders and adapts to the increasing channel count across multiple scale levels. Indicates channel attention weights. This represents the Sigmoid activation function. This indicates the use of a global average pooling layer for residual features. The result after processing , Represents the attenuation weight baseline matrix. This represents the characteristics of the encoder output. This is for element-wise multiplication.

[0066] In the encoder, the attenuation physical prior is endogenously incorporated into the residual coding process to achieve cross-band information attention and attenuation-sensitive feature adaptive enhancement of the encoder's multi-level features. The calculation process includes residual branch basic feature extraction, cross-channel attention weight generation of physical prior modulation, and final output encoder level features.

[0067] Residual branch basic feature extraction: encoder single-level input features are ,in For batch size, Number of channels (corresponding to multi-band feature dimensions) Let be the length of the signal depth sequence. A dual-convolutional residual coding branch is designed to simultaneously achieve local feature extraction and identity mapping, as shown in the formula: .

[0068] Cross-channel attention weight generation based on physical prior modulation: for residual features Global context features of channels are extracted by global averaging-max pooling fusion. Introducing a decay weight benchmark matrix (The average attenuation level corresponding to each frequency band) is physically modulated, and then channel attention weights are generated through shared convolution and sigmoid activation. It is positively correlated with the attenuation level of the corresponding frequency band, realizing cross-channel weight allocation driven by physical meaning.

[0069] The final output encoder-level features This enables adaptive enhancement of attenuation-sensitive channels while retaining the gradient propagation advantages of the residual structure, achieving deep coupling between feature extraction and physical constraints.

[0070] Each encoder outputs hierarchical features. Then, feature fusion can be achieved by splicing these four levels of features to obtain fused features.

[0071] It should be noted that the encoder described above can be trained using common deep learning training methods, and the embodiments of this application do not limit the training method.

[0072] Step 15: Compensate the original electromagnetic detection signal based on the fusion features and attention constraint mask matrix to obtain the compensated electromagnetic detection signal.

[0073] In some embodiments of this application, the fused feature input decoding module can be processed to obtain the compensated electromagnetic detection signal.

[0074] The aforementioned decoding module comprises multiple decoders connected in sequence. The input of the first decoder is the fused feature, and the output of the last decoder is the compensated electromagnetic detection signal. As an optional example, the number of decoders can be four.

[0075] The decoder described above includes an upsampling layer, a cross-attention unit, and a residual convolution unit. The upsampling layer is used to upsample the input of the decoder, the cross-attention unit is used to perform attention calculation on the upsampling result based on the attention constraint mask matrix, and the residual convolution unit is used to perform residual calculation on the attention calculation result and the fused features.

[0076] In some optional embodiments, the aforementioned cross-attention unit can be a cross-attention module in deep learning, and the residual convolution unit can be a residual convolution module in deep learning.

[0077] It is worth mentioning that the decoding module achieves end-to-end accurate compensation for signals attenuated at different depths and frequency bands by upsampling and feature fusion at each stage.

[0078] The cross-attention module is embedded after bilinear upsampling and before residual convolution. It also serves as a fusion node for the encoder's same-level skip connection features and the decoder's upsampled features. This not only enables the mask to guide the decoding process throughout, but also completes the precise alignment of deep and shallow features across levels, avoiding the problem of deep attenuation features being submerged by shallow features due to the indiscriminate fusion of traditional decoders.

[0079] It should be noted that the above decoding module is trained, and the loss function during the decoding module training process is: ; in, This represents the value of the loss function. Indicates the first Dao, Di Electromagnetic detection signal after compensation at each depth sampling point Indicates the first Dao, Di The original electromagnetic detection signal of each depth sampling point.

[0080] The decoding module can be trained using common deep learning training methods; however, this embodiment does not limit the training method. During training, a large number of electromagnetic detection signals can be used, and the aforementioned loss function is employed to adjust the model parameters. It is worth noting that this loss function is a physical consistency loss function, constraining the network's compensation results to conform to the attenuation law of electromagnetic wave propagation.

[0081] In some embodiments of this application, the aforementioned four encoders and decoders can be trained as a single deep learning network. Specifically, the training dataset is constructed as follows: a simulation dataset integrating different geological structures, different medium conductivity, different temperature and pressure conditions, and different degrees of non-uniformity is constructed. The dataset contains paired unattenuated standard signals and attenuated noisy signals, while adding 5-20dB Gaussian white noise to simulate noise interference in the actual scene. The dataset covers the full depth range from shallow to deep and includes multi-band attenuation modes. The training optimization strategy is as follows: the Adam optimizer is used to optimize network parameters, the initial learning rate is set to 0.001, and a warm start + cosine annealing strategy is used for learning rate scheduling. The first 5 epochs are the warm start period, during which the learning rate is linearly increased from 1e-6 to 0.001, and then gradually decreased using a cosine annealing strategy. The training batch size is 64, and the total training epochs are 80. Gradient clipping is used to limit the gradient norm to avoid gradient explosion and improve training stability.

[0082] In practical applications, physical consistency verification and output of the compensation results are also required. Specifically, physical consistency verification involves verifying the frequency band energy distribution and depth-direction attenuation trend of the compensated signal output by the network. This verifies that the high-frequency energy enhancement trend of the compensated signal conforms to the comprehensive attenuation physical model constructed in step two, and eliminates abnormal compensation results that do not conform to physical laws. The output results are as follows: for the compensated signal that passes the physical consistency verification, the final compensated signal matrix and the corresponding imaging profile are output. At the same time, quantitative indicators such as the signal amplitude spectrum, signal-to-noise ratio, and structural similarity before and after compensation are also output, completing the entire process of intelligent attenuation compensation for the detected signal.

[0083] The following example illustrates the effectiveness of the intelligent attenuation compensation method for detection signals proposed in this application, using simulation data.

[0084] The intelligent attenuation compensation method for the detection signal proposed in this application has been systematically verified through simulation and field measurement data. The comparison methods include the traditional inverse Q filtering method, the 1D U-Net method, and the STNN method. Peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and high-frequency energy recovery rate are used as the core evaluation indicators. The verification results are shown in Tables 1, 2, and 3, as follows: Simulation data validation: A multi-layer geological model including asphalt, cement, and soil layers was used, incorporating typical targets such as underground cavities and pipelines. Simulations were conducted for low, medium, and high attenuation levels, with 5dB Gaussian white noise added to simulate actual engineering interference. Results show that in medium and high attenuation scenarios, the proposed method achieves an average PSNR of 37.4dB, a 2.9dB improvement over the STNN method and a 2.0dB improvement over the 1D U-Net method; the average SSIM reaches 0.56, significantly outperforming the comparative methods; and the energy recovery rate for high-frequency components above 1GHz reaches 82%, more than 1.6 times that of the comparative methods, effectively recovering weak signals and high-frequency details from deep targets, with significantly improved phase axis continuity.

[0085] Validation using measured data: Field-measured ground-penetrating radar (GPR) data were used for validation. The measured data showed severe signal attenuation with increasing depth, with targets below 2m becoming almost undetectable. Results indicate that, compared to other methods, STNN can only compensate for shallow signals and cannot recover effective deep information; while 1D U-Net can achieve some depth compensation, it suffers from insufficient detail recovery and is prone to phase distortion. The method proposed in this application can effectively compensate for weak signals below 2m, clearly recover deep strata interfaces and target reflection characteristics, while completely preserving the waveform structure of the original signal without overcompensation. The effective bandwidth of the compensated signal is expanded by more than two times, significantly improving the resolution and interpretation accuracy of deep imaging.

[0086] Robustness verification: Full-scene testing was conducted under different signal-to-noise ratios (SNRs) of 5-20dB. The proposed method significantly outperformed the comparative method in both PSNR and SSIM at all SNR levels. Even in the extremely low SNR scenario of 5dB, the PSNR of the proposed method remained above 30dB, which is more than 3dB higher than the comparative method. This demonstrates that the proposed method still possesses stable compensation performance under strong noise interference scenarios and exhibits excellent robustness.

[0087] Overall experimental results consistently demonstrate that the method proposed in this application is significantly superior to existing mainstream methods in terms of attenuation compensation accuracy, high-frequency detail recovery capability, physical rationality, and noise resistance robustness. It can effectively solve the attenuation compensation problem of electromagnetic detection signals in deep and complex scenarios and has extremely strong practical application value.

[0088] Table 1. Performance Comparison of Various Methods under Simulation Scenarios with Different Attenuation Levels

[0089] Table 2 Noise robustness verification

[0090] Table 3 Verification of Field Measurement Data

[0091] In summary, the intelligent attenuation compensation method for detection signals proposed in this application has the following advantages: I. A comprehensive physical prior constraint system was constructed to ensure the physical rationality of the compensation results. This application overcomes the shortcomings of existing pure data-driven methods that lack physical constraints. For the first time, it integrates the skin effect basic model, the deep temperature and pressure coupling effect, and the scattering attenuation factor of medium inhomogeneity to construct a comprehensive attenuation physical model. Furthermore, physical priors are endogenously integrated into the entire process of network training and inference through loss function constraints and mask guidance, avoiding false compensation results that do not conform to the laws of electromagnetic wave propagation and significantly improving the reliability of the geological interpretation of the compensation results.

[0092] Second, it achieves precise adaptation to the heterogeneity of nonlinear attenuation in deep multi-band structures, improving the recovery capability of weak signals in deep environments. This application addresses the heterogeneity of attenuation patterns for signals at different depths and frequency bands in deep environments by designing a multi-band decomposition and multi-scale parallel encoder-decoder architecture. This enables differentiated extraction and adaptation of attenuation characteristics in different frequency bands, solving the problem of insufficient recovery capability of existing methods for high-frequency weak signals in deep environments, and effectively expanding the effective detection depth and resolution of electromagnetic detection.

[0093] Third, a differentiated compensation mechanism guided by a physical mask is proposed, balancing compensation accuracy and noise robustness. This application constructs an attention-constrained mask based on theoretical attenuation weights, guiding the network to prioritize learning capacity in deeply attenuated signal regions while protecting shallow, slightly attenuated regions. This breaks through the paradigm of globally uniform processing in existing methods, avoiding overcompensation in shallow regions and undercompensation in deep regions. It achieves stable and accurate compensation even in low signal-to-noise ratio scenarios, significantly improving noise robustness.

[0094] Fourth, it possesses strong scene adaptability and engineering practicality. The physical model of this application can be adapted to the medium characteristics of different geological scenes, and the multi-scale network can be compatible with electromagnetic detection signals of different frequency bands and sampling rates. The training dataset covers complex deep scenes such as temperature-pressure coupling and non-uniform scattering. Compared with existing methods that rely on predefined models, it has stronger generalization ability and can be directly applied to signal processing in various underground non-destructive testing scenarios such as ground penetrating radar and transient electromagnetic detection, with high engineering application value.

[0095] This application also provides an intelligent signal attenuation compensation system, including: The decomposition module is used to perform multi-scale frequency band decomposition on the original electromagnetic detection signal to obtain low-frequency band sub-signals, mid-low frequency band sub-signals, mid-high frequency band signals, and high frequency band signals; The module is used to construct the comprehensive attenuation weight matrix for deep media; the comprehensive attenuation weight matrix is ​​used to describe the full-dimensional attenuation distribution information of electromagnetic wave propagation in deep media; The generation module is used to generate an attention constraint mask matrix for deep media based on the comprehensive attenuation weight matrix; the attention constraint mask is used to describe the spatial distribution of attenuation regions and compensation priorities of electromagnetic signals in deep media. The extraction module is used to extract features from low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band signals, and high-frequency band signals respectively, and to fuse the extracted features to obtain fused features; The compensation module is used to compensate the original electromagnetic detection signal based on the fused features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal.

[0096] In the embodiments of this application, the intelligent attenuation compensation system for the detection signal is a system corresponding to the intelligent attenuation compensation method for the detection signal described above, which can solve the problem of low accuracy of electromagnetic detection signal attenuation compensation.

[0097] It should be noted that the information interaction and execution process between the above modules / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0098] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0099] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for intelligent attenuation compensation of detection signals, characterized in that, include: The original electromagnetic detection signal is decomposed into low-frequency band sub-signals, mid-low-frequency band sub-signals, mid-high-frequency band sub-signals, and high-frequency band sub-signals by multi-scale decomposition. A comprehensive attenuation weight matrix is ​​constructed for deep media; the comprehensive attenuation weight matrix is ​​used to describe the full-dimensional attenuation distribution information of electromagnetic wave propagation in deep media; An attention constraint mask matrix for the deep medium is generated based on the comprehensive attenuation weight matrix; the attention constraint mask matrix is ​​used to describe the spatial distribution of the attenuation region and the compensation priority of the electromagnetic signal in the deep medium. Features of the low-frequency band sub-signal, mid-low-frequency band sub-signal, mid-high-frequency band signal and high-frequency band signal are extracted respectively, and the extracted features are fused to obtain fused features; The original electromagnetic detection signal is compensated based on the fusion features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal.

2. The intelligent attenuation compensation method for detection signals according to claim 1, characterized in that, Comprehensive attenuation weight matrix It is a three-dimensional tensor with dimension . ; in, This indicates the number of spatial sampling channels for the data, corresponding to the underground space sampling dimension. The value is determined by the total length of the survey lines and the spacing between channels. Each index corresponds to a spatial sampling location. This indicates the number of time-domain sampling points per channel, corresponding to the longitudinal depth, and determines the sampling accuracy and longitudinal resolution in the depth direction; This indicates the number of subbands in the wavelet packet decomposition. .

3. The intelligent attenuation compensation method for detection signals according to claim 2, characterized in that, Comprehensive attenuation weight matrix Matrix elements for: ； Among them, matrix elements Indicates the first Dao, Di The depth sampling point, the first Electromagnetic wave cumulative attenuation weight in each sub-band, propagation depth , This represents the average velocity of electromagnetic waves in the Earth's crust. Indicating two-way travel, Indicates depth Place, No. The comprehensive attenuation constant corresponding to each sub-band; ； = ； = ； ； = ； ； in, Indicates depth Place, No. The fundamental attenuation constant corresponding to each sub-band Indicates depth Place, No. The temperature-pressure coupling attenuation factor corresponding to each sub-band. Indicates depth Place, No. The scattering attenuation constant corresponding to each sub-band Angular frequency represents the oscillation rate of an electromagnetic wave signal. Permeability represents the ability of a deep medium to conduct magnetic fields. Indicates as depth Changing dielectric parameters, Indicates as depth Sub-band Changing dynamic conductivity, Represents the relative permittivity of the Earth's surface matrix. This represents the depth gradient correction coefficient. Indicates dynamic conductivity. This represents the baseline conductivity of the Earth's surface at normal temperature and pressure. This represents the dispersion matching coefficient that adaptively adjusts with the signal frequency. This represents the activation energy of thermo-baric coupling. Indicates formation temperature. Indicates formation pressure. This represents the ideal gas constant, enabling normalized conversion between energy units and temperature units. Indicates the formation porosity compressibility. Represents the dielectric contrast scattering coefficient. Represents the base scattering coefficient of the formation. The dielectric constant of a porous fluid. The dielectric constant of the soil matrix is ​​represented by the dielectric constant of the soil matrix. This represents the depth gradient correction coefficient.

4. The intelligent attenuation compensation method for detection signals according to claim 3, characterized in that, Attention Constraint Mask Matrix Matrix elements for: ； ； ； in, This represents the mask value after optimization by three-dimensional morphological opening operations. Represents an ellipsoidal three-dimensional structural element. , express The local neighborhood offset, Fixed value It fully covers the 3×3×3 neighborhood space. The third morphological erosion operation represents the result of the three-dimensional erosion operation. Dao, Di The depth sampling point, the first Sub-band mask values, Representing the ellipsoidal three-dimensional structural element in The weighting coefficient of location, This indicates that after the erosion operation, the first... Dao, Di The depth sampling point, the first Sub-band mask values, Indicates the original mask in the neighborhood The value at that location, Indicates the first Dao, Di The depth sampling point, the first The original binary attention mask value of the sub-band. The attenuation weighting benchmark matrix Matrix elements Indicates the first Dao, Di The overall attenuation level of each depth sampling point; , Indicates the first The proportion of signal energy in each sub-band This represents the energy-weighted adaptive threshold. , Indicates the baseline threshold. This represents the total energy of the signal.

5. The intelligent attenuation compensation method for detection signals according to claim 4, characterized in that, The extraction of features from the low-frequency sub-signal, mid-low-frequency sub-signal, mid-high-frequency sub-signal, and high-frequency sub-signal respectively includes: The low-frequency band signal, the mid-low-frequency band signal, the mid-high-frequency band signal, and the high-frequency band signal are input one-to-one into four encoders for feature extraction, thereby obtaining the features of the low-frequency band signal, the mid-low-frequency band signal, the mid-high-frequency band signal, and the high-frequency band signal.

6. The intelligent attenuation compensation method for detection signals according to claim 5, characterized in that, The encoder performs feature extraction as follows: ； ； ； in, Representing residual characteristics, This represents a one-dimensional convolution operation. This indicates batch normalization operations. Represents the ReLU activation function. This refers to one of the low-frequency band signal, mid-low frequency band signal, mid-high frequency band signal, and high frequency band signal. express One-dimensional convolution operation, Indicates channel attention weights. This represents the Sigmoid activation function. This indicates the use of a global average pooling layer for residual features. The result after processing Represents the attenuation weight baseline matrix. This indicates the characteristics of the encoder output.

7. The intelligent attenuation compensation method for detection signals according to claim 6, characterized in that, The compensation of the original electromagnetic detection signal based on the fused features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal includes: The fused features are input into the decoding module for processing to obtain the compensated electromagnetic detection signal; The decoding module includes multiple decoders connected in sequence. The input of the first decoder is the fused feature, and the output of the last decoder is the compensated electromagnetic detection signal. The decoder includes an upsampling layer, a cross-attention unit, and a residual convolution unit. The upsampling layer is used to upsample the input of the decoder. The cross-attention unit is used to perform attention calculation on the upsampling result based on the attention constraint mask matrix. The residual convolution unit is used to perform residual calculation on the attention calculation result and the fused feature.

8. The intelligent attenuation compensation method for detection signals according to claim 7, characterized in that, The loss function for the training process of the decoding module is: ； in, This represents the value of the loss function. Indicates the first Dao, Di Electromagnetic detection signal after compensation at each depth sampling point Indicates the first Dao, Di The original electromagnetic detection signal of each depth sampling point.

9. A smart signal attenuation compensation system, characterized in that, include: The decomposition module is used to perform multi-scale frequency band decomposition on the original electromagnetic detection signal to obtain low-frequency band sub-signals, mid-low frequency band sub-signals, mid-high frequency band signals, and high frequency band signals; A construction module is used to construct a comprehensive attenuation weight matrix for deep media; the comprehensive attenuation weight matrix is ​​used to describe the full-dimensional attenuation distribution information of electromagnetic wave propagation in deep media; The generation module is used to generate an attention constraint mask matrix for the deep medium based on the comprehensive attenuation weight matrix; the attention constraint mask is used to describe the spatial distribution of the attenuation region and the compensation priority of the electromagnetic signal in the deep medium. The extraction module is used to extract features from the low-frequency band sub-signal, the mid-low-frequency band sub-signal, the mid-high-frequency band signal, and the high-frequency band signal respectively, and to fuse the extracted features to obtain fused features; The compensation module is used to compensate the original electromagnetic detection signal based on the fusion features and the attention constraint mask matrix to obtain the compensated electromagnetic detection signal.

Images