Density measurement method and system fusing shallow sparse density measurement data and ground penetrating radar data

Abstract

The present application belongs to the technical field of nondestructive testing, in order to solve the problem of low accuracy of existing density, a density measurement method and system are proposed by fusing sparse measurement value and ground penetrating radar data, shallow layer sparse density measurement data is mapped to shallow layer sparse density measurement value aligned with ground penetrating radar scanning data, and measurement position indication mask corresponding to shallow layer sparse density measurement value is generated; the above three channel data are input into the trained inversion model to predict the equivalent dielectric constant and estimate the medium state parameter; wherein the medium state parameter is used to represent the conversion relationship between the equivalent dielectric constant and the density; the density field of the region to be measured is obtained according to the equivalent dielectric constant and the medium state parameter, and then the dense area imaging result is generated. The present application realizes the inversion result with density accuracy and structural details under the condition of sparse measurement value.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing technology, and in particular relates to a method and system for measuring density by fusing shallow sparse density measurement data and ground penetrating radar data. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the advancement of construction and experimental research in filling engineering, the demand for rapid detection and evaluation of the density of layered compacted materials is increasing. Density is a crucial indicator for evaluating the structural stability and load-bearing capacity of filling engineering structures. Currently, on-site density testing mainly relies on point methods such as the sand cone method, ring cutter method, or nucleus-free density gauge. These methods often require sampling, are highly destructive, have low testing efficiency, and the measurement points are discrete and have limited representativeness, making it difficult to achieve large-scale continuous testing. Ground penetrating radar (GPR), as an efficient non-destructive testing method, can quickly obtain the electromagnetic response profile inside the material, which can be used for qualitative identification such as dense interfaces, layered structures, or abnormal areas. However, it essentially detects differences in the electromagnetic properties of the medium (such as the equivalent dielectric constant), and is affected by factors such as the moisture content and other medium state parameters, making it difficult to directly correlate with density values.

[0004] Existing ground-penetrating radar (GPR) density inversion methods are mainly divided into two categories: physical model-driven and data-driven. Physical model methods are typically based on the idealized assumption of homogeneous and isotropic media, using empirical formulas relating electromagnetic parameters to density. However, in actual engineering projects, the medium in the filling material is usually non-uniformly distributed, and GPR signals are highly susceptible to nonlinear interference from medium state parameters such as water content, leading to insufficient stability and significant bias in the inversion results. Data-driven methods utilize deep learning models (such as CNN, U-Net, and Transformer) to directly map the nonlinear relationship between GPR signals and density, possessing stronger feature extraction capabilities. However, their inversion results are highly dependent on the distribution of training data, easily affected by equipment parameter drift, differences in the acquisition environment, and sample bias, and generally suffer from poor field generalization ability and lack of physical meaning.

[0005] Specifically, existing deep learning inversion frameworks face three challenges when processing engineering field data. First, physical measurements in engineering data are typically sparsely distributed and discontinuous, often representing the average density within a specific sampling volume; while ground-penetrating radar profile data is continuously collected along the survey line, reflecting the electromagnetic response characteristics of the filling structure at a higher spatial resolution. Due to the significant differences in sampling methods and spatial resolution between the two types of data, existing network models struggle to effectively fuse them. Directly aligning high-resolution prediction results to the regional mean corresponding to the measured values ​​introduces a scale mismatch problem, causing numerical flattening within the corresponding mapped region and suppressing local differences within that region.

[0006] Secondly, the ground-penetrating radar response essentially reflects the differences in the electromagnetic properties of the medium. These changes can originate from variations in pore structure and component ratios caused by changes in density, or from changes in electromagnetic parameters caused by variations in medium state parameters such as water content. These two types of factors are coupled and superimposed in the radar signal, and their direction and amplitude often change with varying operating conditions. In the absence of effective physical constraints or calibration mechanisms, the model struggles to distinguish the contributions of "medium state changes" and "density changes" to the electromagnetic response, easily misinterpreting state disturbances such as water content as density differences. This leads to increased errors in the quantitative inversion of density and insufficient generalization ability across operating conditions.

[0007] Finally, the internal structure of the filling contains small-scale non-uniform patches, whose spectral energy distribution usually shows a certain pattern; however, conventional networks, under the optimization objective mainly based on average error, tend to generate overly smooth results, thereby suppressing the detailed features of local non-uniform patches in the density field and making it difficult to accurately recover the structural features of the non-uniform distribution of the density field.

[0008] In summary, existing single physical models or conventional deep learning methods are insufficient to achieve inversion results that combine density accuracy and structural detail under conditions of sparse measurements. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, the present invention provides a density measurement method and system that integrates shallow sparse density measurement data and ground penetrating radar data. It effectively integrates sparse measurement values ​​and ground penetrating radar data to achieve quantitative density measurement and dense area imaging, and achieves inversion results with both density accuracy and structural details under the condition of sparse measurement values.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a density measurement method that integrates shallow sparse density measurement data and ground-penetrating radar data, comprising:

[0012] Acquire ground-penetrating radar scanning data and shallow density measurement data of the area to be tested;

[0013] The shallow sparse density measurement data is mapped to shallow sparse density measurement values ​​aligned with the ground penetrating radar scan data, and a measurement position indication mask corresponding to the shallow sparse density measurement values ​​is generated.

[0014] The ground-penetrating radar scan data, the shallow sparse density measurement value, and the measurement location indication mask are input into the trained inversion model to predict the equivalent dielectric constant and estimate the medium state parameters; wherein, the medium state parameters are used to characterize the conversion relationship between the equivalent dielectric constant and the density.

[0015] The density field of the region to be measured is obtained based on the equivalent dielectric constant and the dielectric state parameters, and then the imaging result of the dense region is generated.

[0016] Secondly, the present invention provides a density measurement system that integrates shallow sparse density measurement data and ground-penetrating radar data, comprising:

[0017] The acquisition module is configured to acquire ground-penetrating radar scan data and shallow sparse-dense density measurement data of the area to be measured.

[0018] The alignment module is configured to: map the shallow sparse density measurement data to shallow sparse density measurement values ​​aligned with the ground penetrating radar B-scan data, and generate a measurement position indication mask corresponding to the shallow sparse density measurement values.

[0019] The inversion module is configured to: input the ground-penetrating radar scan data, the shallow sparse density measurement value, and the measurement location indication mask into the trained inversion model, predict the equivalent dielectric constant, and estimate the medium state parameters; wherein, the medium state parameters are used to characterize the conversion relationship between the equivalent dielectric constant and the density;

[0020] The calculation module is configured to: obtain the density field of the region to be measured based on the equivalent dielectric constant and the dielectric state parameters, and then generate the dense region imaging result.

[0021] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0022] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0023] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0024] The above one or more technical solutions have the following beneficial effects:

[0025] In this invention, the equivalent dielectric constant predicted by the inversion network and the estimated dielectric state parameters are used together for quantitative conversion of density, thereby outputting a density field with clear physical dimensions and numerical meaning. Compared with only giving qualitative results of the boundary between dense and non-dense regions, this invention can realize the integrated output of quantitative inversion of density and dense region imaging, and has stronger practicality.

[0026] In this invention, the predicted density within the mapping region is taken as the regional average value, and the error between this average value and the actual measured density at the measuring point is used to construct a mean error constraint term. This term is used to constrain the quantitative consistency of the density field and to provide supervisory information for the estimation and calibration of medium state parameters. By introducing a regional mean constraint, this invention solves the problem of spatial resolution difference caused by the fact that sparse density measurements are average values ​​within a macroscopic sampling volume, while ground-penetrating radar scan data are two-dimensional electromagnetic response data matrices.

[0027] In this invention, to avoid overly smoothing of the density field prediction results and preserve the inhomogeneity of the medium, frequency domain spectral constraints are introduced during the training phase. A two-dimensional Fourier transform is performed on the predicted density field to obtain a two-dimensional power spectrum, and the two-dimensional power spectrum is radially averaged to obtain a radially averaged power spectrum curve as a one-dimensional spectral feature. The difference between the one-dimensional spectral feature and the preset target spectral feature is calculated to construct a frequency domain spectral constraint term, which is then used in training optimization. Frequency domain spectral constraints can suppress the over-smoothing trend of the predicted density field, reduce the weakening of spatial variation information of small-scale inhomogeneous patches in the density field, and thus enhance the ability to preserve the inhomogeneous structural features of the medium.

[0028] In this invention, a measurement location indicator mask is introduced and a mask guidance mechanism is adopted in the inversion network. On the one hand, this enables the network to distinguish between regions with measured density and regions without measured values ​​during feature extraction and fusion, thereby enhancing the constraint effect of sparse measurement point information on network learning. On the other hand, by utilizing the effective calculation area constrained by the mean constraint of the mask channel indicator area, and combining it with the physical mapping area corresponding to the measurement point, the predicted density is averaged regionally and compared with the error in the actual measured density calculation, thereby estimating the medium state parameters or updating them in the field.

[0029] In this invention, by using regional mean constraints and its on-site calibration and update mechanism, a small number of sparse density measurements can be used to adjust the medium state parameters, so that the density conversion relationship can maintain quantitative consistency under different working conditions and different site conditions, thereby improving the engineering applicability and cross-scenario transferability of the method.

[0030] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0032] Figure 1 This is a flowchart of the density measurement method that integrates shallow sparse density measurement data and ground penetrating radar data in Embodiment 1 of the present invention.

[0033] Figure 2 This is a schematic diagram illustrating the construction of the three-dimensional non-uniform random medium model and its multi-source simulation dataset in Embodiment 1 of the present invention;

[0034] Figure 3 This is a schematic diagram of the deep neural network architecture that integrates physical laws in Embodiment 1 of the present invention;

[0035] Figure 4 The images shown are density imaging results in Embodiment 1 of the present invention. (a), (b), and (c) correspond to different morphological conditions of the dense region, respectively. Detailed Implementation

[0036] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0037] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0038] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0039] Example 1

[0040] This embodiment discloses a density measurement method that integrates shallow sparse density measurement data and ground-penetrating radar data, including:

[0041] Acquire ground-penetrating radar scanning data and shallow density measurement data of the area to be tested;

[0042] The shallow sparse density measurement data is mapped to shallow sparse density measurement values ​​aligned with the ground penetrating radar scan data, and a measurement location indication mask corresponding to the shallow sparse density measurement values ​​is generated.

[0043] Ground penetrating radar scan data, shallow sparse density measurements, and measurement location indication masks are input into a trained inversion model to predict the equivalent dielectric constant and estimate the medium state parameters; among which, the medium state parameters are used to characterize the conversion relationship between the equivalent dielectric constant and density.

[0044] The density field of the region under test is obtained based on the equivalent dielectric constant and dielectric state parameters, and then the imaging results of the dense region are generated.

[0045] This embodiment provides a density measurement method that integrates shallow sparse density measurements with ground-penetrating radar data. The specific process is as follows: Figure 1 As shown. In the training phase, this method uses large-scale simulation samples to pre-train the inversion network, enabling it to learn the inversion relationship between the ground-penetrating radar response and the equivalent dielectric constant distribution, as well as the feature representation related to subsequent density conversion. In the practical application phase, the method combines the shallow sparse density measurement values ​​obtained in the field as physical constraints to estimate and calibrate the medium state parameters, thereby adapting the density conversion relationship to the current working conditions and outputting the density field and dense area imaging results.

[0046] The following is a detailed description of the density measurement method proposed in this embodiment, which integrates shallow sparse density measurement data and ground-penetrating radar data, specifically including the following steps:

[0047] Step S1: Construct a simulation training dataset based on a three-dimensional random medium model.

[0048] Step S101: To address the issue of strong heterogeneity in underground soil media, this embodiment uses three-dimensional random media modeling technology to generate a large-scale simulation dataset containing complex geological features.

[0049] First, a three-dimensional non-uniform soil medium model of a predetermined size is constructed based on random medium theory to simulate the complex scenario of coexistence of compaction defects in real soil. The mesh resolution of the soil medium model is set to [value missing]. To simulate the structure formed by the actual roadbed compaction process, a randomly generated semi-ellipsoid is set at the bottom of the soil medium model as a compaction core region, with its geometric parameters randomly sampled within a reasonable range; and the dielectric constant within this region is introduced with a specific variance range (e.g., ±3) centered on a set mean value and following a Gaussian distribution. Random perturbations within the core are used to simulate the random non-uniformity within the actual packing material. Based on this, a transition shell of randomly and dynamically adjusted thickness is constructed around the dense core region. Its dielectric constant decreases continuously from the core region towards the background medium to simulate the physical interface where density gradually changes towards the periphery. Figure 2 As shown on the left.

[0050] Step S102: To further approximate actual working conditions, this embodiment embeds several small ellipsoids in the background area representing insufficiently compacted loose soil in the soil medium model. The number of these ellipsoids is negatively correlated with the volume of the dense core. These randomly distributed anomalies, together with the dense core area and the transition shell, constitute a complex medium environment.

[0051] Step S103: As Figure 2 As shown in survey lines 1, 2, and 3, the above three-dimensional non-uniform model is sliced ​​in two dimensions along the preset survey line direction to extract a series of two-dimensional dielectric constant matrices, and a simulation dataset with strong physical correlation is constructed based on these matrices.

[0052] First, the two-dimensional dielectric constant distribution map is input into the finite-difference time-domain (FDTD) electromagnetic simulation solver. By simulating the propagation and scattering process of radar waves in a non-uniform medium, the corresponding ground-penetrating radar B-scan data is generated. Figure 2 The data shown is from the ground-penetrating radar B-scan.

[0053] This embodiment also includes at least one of the following: zero-time correction, removal of DC component and background noise, and time-varying gain compensation for the original ground-penetrating radar B-scan data.

[0054] Step S2: Construct a three-channel input to simulate the on-site working conditions.

[0055] To enable the network to process sparse density measurements, this step simulates the physical sampling process of field instruments through numerical calculations, constructing a three-channel training input consistent with actual applications, including the ground-penetrating radar B-scan channel, the shallow sparse density measurement value channel, and the measurement position indication mask channel.

[0056] Step S201: Construct the ground-penetrating radar B-scan channel: Normalize the B-Scan data generated by the finite-difference time-domain method while preserving floating-point precision.

[0057] Step S202: Simulate macroscopic volume sampling and generate sparse measurement values.

[0058] To accurately reflect the characteristics of sparse measurements and macroscopic volume average in field testing, this step simulates the sampling process without a nucleus density meter using numerical methods: several virtual measurement points are set up at the horizontal position corresponding to the ground penetrating radar survey line. With each measurement point as the center, a corresponding three-dimensional region is delineated in the simulation model based on the typical sampling volume, and the arithmetic mean of the density in that region is calculated as the measurement value of that measurement point.

[0059] Optional, a typical sampling volume is a cuboid with a length, width, and depth of 20 cm and a depth of 30 cm, as corresponding to a nucleus-free density meter.

[0060] Step S203: Construct shallow sparse density measurement value channel and measurement position indication mask channel.

[0061] Based on the physical sampling depth of the measured value =30cm and equivalent wave velocity of the medium Using the formula: Calculate the effective ground-penetrating radar two-way travel time window corresponding to the measured value, with a sampling interval of . The corresponding number of pixel rows in the image Subsequently, the horizontal coordinate column and vertical direction of the measured density values ​​were [0] Normalized density measurement data is filled into the pixel range of [ ], while the remaining areas remain zero to form a shallow sparse density measurement value channel. Simultaneously, a measurement position indication mask channel is generated, marking the effective area filled with the density measurement values ​​as 1 and the remaining areas as 0. Finally, the ground-penetrating radar B-scan channel, the shallow sparse density measurement value channel, and the measurement position indication mask channel are stacked to form an input of size (3,H,W).

[0062] When randomly cropping, flipping, or scaling the ground-penetrating radar B-scan image, the same geometric transformation is simultaneously performed on the shallow sparse density measurement value channel and the measurement position indication mask channel, while maintaining the spatial alignment of the three.

[0063] Step S3: Construct a density inversion network architecture that integrates physical conversion formulas.

[0064] To achieve quantitative density inversion under the condition of fusing shallow sparse density measurements with ground penetrating radar data, this embodiment constructs a density inversion network structure that integrates physical conversion formulas.

[0065] The inversion network receives three input channels: a ground-penetrating radar B-scan channel, a shallow sparsity and density measurement channel, and a measurement location indication mask channel. The network outputs an equivalent dielectric constant prediction field and the medium state parameter K. During the feature extraction stage, the inversion network generates a gated weight map based on the measurement location indication mask channel. Furthermore, the intermediate features of the network are weighted and modulated to enhance the feature response of the effective region of the measurement value, thereby improving the constraint effect of sparse measurement point information on feature learning.

[0066] In this inversion network, the medium state parameters The conversion coefficient in the compactness conversion formula is set as the quantity to be estimated. During the training phase, equivalent dielectric constant labels can be provided in the simulation samples to supervise the equivalent dielectric constant. At the same time, by utilizing the mean loss of the region corresponding to the sparse compactness measurement points, the difference between the predicted average compactness and the measured compactness in the mapped region is minimized to achieve the estimation of the medium state parameters. Learning estimation; in the field calibration stage, the medium state parameters can also be estimated based on the regional mean loss. Perform a calibration update.

[0067] It should be noted that the density referred to in this embodiment can be calculated according to the definition of compaction.

[0068] The equivalent dielectric constant is obtained by inversion from ground-penetrating radar data. Then, the medium state parameters Substituting the density conversion formula derived from the hybrid dielectric model and the definition of density, the density field is calculated, and the dense region imaging result is output accordingly. During the training phase, frequency domain spectral constraints are further introduced to suppress excessive smoothing of the predicted density field and enhance the ability to preserve non-uniform structures, thus inverting the network structure. Figure 3 As shown.

[0069] Step S301: Mask-guided feature extraction network.

[0070] This embodiment employs a feature extraction network with an encoder-decoder structure.

[0071] The feature extraction network receives three-channel input. It includes a ground-penetrating radar B-scan channel, a shallow sparse measurement value channel, and a measurement position indication mask channel. The encoder consists of five cascaded downsampling stages, each integrating a mask-gated convolutional block for feature extraction and mask-guided weighting in the next stage.

[0072] The calculation logic for the downsampling stage is as follows:

[0073]

[0074]

[0075] In the formula, For the first Layer input features, The corresponding measurement position indicates the mask feature. Represents convolution operation. and For convolution kernel weights, and For bias terms, It is the ReLU activation function. It is the Sigmoid activation function. For element-wise multiplication, This is a gating weight graph.

[0076] Through a mask gating mechanism, the network enhances the feature response of the effective region of the measurement value in the feature space, thereby facilitating the establishment of a correlation between sparse measurement point information and the electromagnetic response features at the corresponding locations. The decoder gradually restores the spatial resolution through transposed convolution, outputting a high-resolution feature map. Based on the high-resolution feature map, the first prediction branch outputs the equivalent dielectric constant prediction field; the second parameter branch performs global feature aggregation on the feature map and outputs the dielectric state parameters. The first prediction branch is used to perform convolutional mapping on the high-resolution feature map output by the feature extraction network to obtain the equivalent dielectric constant prediction field; the second parameter branch is used to perform global feature aggregation on the high-resolution feature map and outputs the dielectric state parameters through the parameter mapping layer, where the parameter mapping layer can specifically be fully connected.

[0077] Step S302: Equivalent dielectric constant inversion and density quantitative conversion.

[0078] This embodiment divides the quantitative calculation of density into two parts: "equivalent dielectric constant acquisition" and "density conversion". First, the electromagnetic propagation characteristics of the medium are estimated based on ground-penetrating radar B-scan data, thereby obtaining the equivalent dielectric constant of the medium through inversion. .

[0079] Then, based on the mixed dielectric model of composition ratio, the correspondence between dielectric components and equivalent dielectric parameters is established, and the density conversion formula is derived by combining the definition of density. To reduce the need for explicit acquisition of intermediate physical quantities such as water content and soil particle specific gravity in the field, this embodiment uses their comprehensive influence on the conversion relationship as dielectric state parameters. The density mapping relationship is then constructed by representing it in the form of:

[0080]

[0081] In the formula, For density; It is the dry density of the soil; The maximum dry density was obtained from a standard laboratory compaction test. The density of water; It is the equivalent dielectric constant; Medium state parameters or conversion factors; Medium state parameters It is the specific gravity of soil particles Moisture content Dielectric constant of water Soil dielectric constant The nonlinear combination term, i.e. .

[0082] During the training data construction phase, the medium state parameters can be calculated using the following formula:

[0083]

[0084] in, This refers to the specific gravity of soil particles. Moisture content (by weight); The dielectric constant of water; is the dielectric constant of the soil.

[0085] During the training phase, the medium state parameters are analyzed using the region mean loss method. Estimate the parameters; during the on-site calibration phase, further adjustments to the medium state parameters can be made based on the regional mean error. Perform calibration and updates. Finally, the equivalent dielectric constant of the dielectric obtained from the inversion is used. With medium state parameters Substituting into the density conversion formula, the overall density value is calculated, providing a basis for subsequent dense area imaging output.

[0086] Step S303: Construction of optimization objects and composite loss during the training phase.

[0087] During the training phase, the inversion network predicts the equivalent dielectric constant field under the condition of a three-channel fused input, and constructs a dielectric parameter loss term under the supervision of the equivalent dielectric constant label provided by the simulation data. To constrain the equivalent dielectric constant of the medium The accuracy of the prediction.

[0088]

[0089] In the formula, N is the number of simulation samples involved in the calculation. To invert the network for the first The equivalent dielectric constant field predicted from each sample. The simulation data provides the first The equivalent dielectric constant field of each sample, It is the square of the L2 norm.

[0090] Furthermore, the equivalent dielectric constant of the dielectric is... With medium state parameters Substitute density The conversion formula yields the compaction field, and a composite loss function is constructed: on the one hand, a regional mean loss term is built based on the measured point mapping area, constraining the difference between the regional average of the predicted compaction and the corresponding measured compaction within the mapping area to decrease, thereby providing a conversion coefficient. The learning estimation or field calibration provides supervision information. On the other hand, frequency domain spectral constraints are also introduced during the training phase to constrain the spatial spectral energy distribution of the compactness field, so that the prediction results avoid excessive smoothing and enhance the ability to preserve the characteristics of non-uniform structures.

[0091] Step S304: Constructing the regional mean-constrained loss.

[0092] Based on the sampling aperture and sampling depth of the measuring tool, and combined with the B-scan channel spacing and time-domain sampling interval, the predicted density field is obtained. The above determines the first The physical mapping area corresponding to each measurement point Calculate the arithmetic mean of all predicted data within the region and compare it with the corresponding measured values. Calculate the error and construct the regional mean loss term. Its calculation formula is:

[0093]

[0094] In the formula, The number of measured values. The number of valid points within the measurement area. This is the density value calculated for the inversion network prediction.

[0095] Step S305: Constructing the frequency domain spectral loss.

[0096] To avoid over-smoothing of the predicted density field and reduce the suppression of local inhomogeneities, this embodiment introduces a frequency domain spectral loss term during the training phase.

[0097] First, specifically, the predicted density field A two-dimensional Fourier transform is performed to obtain the frequency domain representation, and the square of its modulus is calculated to obtain the two-dimensional power spectrum. Further radial averaging of the two-dimensional power spectrum yields the radially averaged power spectrum curve as a one-dimensional spectral feature. The one-dimensional spectral features are compared with the preset target spectral features. Perform difference calculations and construct the frequency domain spectral loss term:

[0098]

[0099] in, Target spectral characteristics Obtained from statistical analysis of training data; () is the normalization operator, and it is preferred to normalize according to the total energy.

[0100] By minimizing the frequency domain spectral loss term, the energy distribution of the predicted density field at different spatial frequencies can be limited, avoiding the suppression of local variations such as small-scale inhomogeneous patches while only retaining the overall variation trend of large-scale dense / non-dense regions, thus enhancing the ability to preserve the non-uniform structural characteristics of the medium.

[0101] It should be noted that the target spectral characteristics The statistical prior used to provide the spatial undulation structure of the compaction field is constrained by the shape of the radial average power spectrum curve, i.e., the relative energy proportions of different spatial frequency components. Frequency domain spectral loss is used to constrain the spectral shape of the predicted compaction field, ensuring that the spatial variation components corresponding to small-scale inhomogeneous patches are not excessively weakened, leveraging statistical data obtained from the training data. Limit the reasonable range of high-frequency components to reduce the risk of misjudging random noise as a valid high-frequency structure.

[0102] Among them, the frequency domain is a two-dimensional spatial frequency domain. The high-frequency components mainly correspond to local changes such as small-scale non-uniform patches in the density field, while the low-frequency components mainly correspond to the overall distribution trend of large-scale dense / non-dense areas. The one-dimensional spectral features are normalized to reduce the impact of amplitude scale differences under different field conditions on the constraint effect.

[0103] Step S4: On-site operating condition calibration.

[0104] Step S401: On-site data acquisition and three-channel input construction.

[0105] Obtain ground-penetrating radar B-scan data for the roadbed section to be tested, and collect shallow sparse density measurements obtained synchronously on-site (such as a small number of measurement point values ​​obtained using a nucleusless density meter). Following the method in step S2, construct a shallow sparse density measurement value channel and a measurement location indication mask channel, and spatially align them with the ground-penetrating radar B-scan channel to form a three-channel input for actual measurement.

[0106] Step S402: Medium state parameters On-site calibration updates.

[0107] The measured three-channel input is fed into a pre-trained inversion network, and the medium state parameter in the compaction conversion formula is adjusted using the region mean loss. Perform calibration updates to adapt it to the current operating conditions.

[0108] Specifically, the conversion coefficient is adjusted with the regional mean error as the update target, so that the regional average of the predicted density within the mapping area of ​​the measurement point is consistent with or the difference between the predicted density and the corresponding measured density is reduced.

[0109] Step S403: Output of density field and dense region imaging.

[0110] Using calibrated medium state parameters The density value is converted to obtain the density field, and the imaging results of the dense area are output based on the density field.

[0111] like Figure 4 As shown in (a)-(c), imaging examples of this method under different dense region morphological conditions are presented. Figure 4 It can be seen that by fusing sparse density measurements with ground-penetrating radar data, this method can achieve quantitative inversion output of the density across the entire field and maintain the non-uniform structural characteristics of the medium in the imaging results.

[0112] In summary, this method can utilize a small number of on-site density measurements to determine the medium's state parameters. The calibration and update process adapts the density conversion relationship to different engineering materials and working conditions, thereby enabling quantitative measurement of density and imaging output of dense areas in various engineering scenarios.

[0113] This embodiment uses ground-penetrating radar B-scan data, a shallow sparse density measurement channel aligned with it, and a measurement position indication mask channel as three-channel inputs to achieve joint characterization of discrete measurement point information and continuous electromagnetic response. The three-channel input data are fed into an inversion network to predict the equivalent dielectric constant of the medium and estimate or update the medium state parameters, which are the conversion coefficients between the equivalent dielectric constant and density. By substituting the equivalent dielectric constant and the medium state parameters into the density conversion formula derived from the hybrid dielectric model based on composition ratio and the definition of density, the full-field density field is calculated, and the dense region imaging results are output. Frequency domain spectral constraints are introduced during the inversion network training phase to limit the spatial spectral energy distribution of the density field, making the energy distribution of the prediction results at different spatial frequencies more consistent with the spatial structural characteristics of the medium. This suppresses excessive smoothing of the prediction results and enhances the ability to preserve the non-uniform structural characteristics of the medium. In addition, a regional mean constraint is introduced during training and field calibration. The conversion coefficient is learned, estimated, or calibrated and updated based on the error between the regional average of the predicted compaction and the average of the measured compaction within the mapping area of ​​the measurement point.

[0114] The method in this embodiment can invert and reconstruct the overall density and achieve imaging of dense areas by fusing shallow sparse measurements and ground-penetrating radar data. It is compatible with existing ground-penetrating radar data acquisition and processing workflows, making it convenient for engineering applications and deployment.

[0115] The method described in this embodiment is applicable to the measurement of compaction density and the assessment of compaction status of similar material filling bodies in geotechnical model tests. It is also applicable to the quantitative measurement of compaction density and imaging of compacted zones in layered filling compacted layers such as roadbeds, site slabs, or embankments.

[0116] Example 2

[0117] The purpose of this embodiment is to provide a density measurement system that integrates shallow sparse density measurement data and ground-penetrating radar data, including:

[0118] The acquisition module is configured to acquire ground-penetrating radar scan data and shallow sparse-dense density measurement data of the area to be measured.

[0119] The alignment module is configured to: map the shallow sparse density measurement data to shallow sparse density measurement values ​​aligned with the ground penetrating radar scan data, and generate a measurement position indication mask corresponding to the shallow sparse density measurement values;

[0120] The inversion module is configured to: stitch together the ground-penetrating radar scan data, the shallow sparse density measurement value, and the measurement location indication mask, and input them into the trained inversion model to predict the equivalent dielectric constant and estimate the medium state parameters; wherein, the medium state parameters are used to characterize the conversion relationship between the equivalent dielectric constant and the density.

[0121] The calculation module is configured to: obtain the density field of the region to be measured based on the equivalent dielectric constant and the dielectric state parameters, and then generate the dense region imaging result.

[0122] In further embodiments, the following is also provided:

[0123] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0124] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0125] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0126] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0127] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0128] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0129] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0130] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0131] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0132] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0133] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A density measurement method integrating shallow sparse density measurement data and ground-penetrating radar data, characterized in that, include: Acquire ground-penetrating radar scanning data and shallow density measurement data of the area to be tested; The shallow sparse density measurement data is mapped to shallow sparse density measurement values ​​aligned with the ground-penetrating radar scan data, and a measurement location indication mask corresponding to the shallow sparse density measurement values ​​is generated; specifically: Based on the physical sampling depth and the equivalent wave velocity of the medium in shallow sparse density measurement data, the corresponding effective ground-penetrating radar two-way travel time window is calculated. The number of pixel rows in the corresponding image is determined based on the effective ground-penetrating radar two-way travel time window. Within the pixel range, the normalized shallow sparse density measurement data is filled into the measured sparse density measurement point positions and corresponding depth ranges, and 0 is filled into the remaining positions to obtain the shallow sparse density measurement values. The effective area of ​​the shallow sparse density measurement value is marked as 1, and the remaining area is marked as 0, thus obtaining the measurement position indication mask; The ground-penetrating radar scan data, the shallow sparse density measurement value, and the measurement location indication mask are input into the trained inversion model to predict the equivalent dielectric constant and estimate the medium state parameters; wherein, the medium state parameters are used to characterize the conversion relationship between the equivalent dielectric constant and the density. The density field of the region to be measured is obtained based on the equivalent dielectric constant and the dielectric state parameters, and then the imaging result of the dense region is generated. The inversion model adopts an encoder-decoder structure. The encoder includes multiple cascaded downsampling stages. The processing of the ground-penetrating radar scan data, the shallow sparse measurement values, and the measurement location indication mask in the first downsampling stage is as follows: Based on the measurement position indication mask, a gating weight map is generated through a mask gating mechanism; The ground-penetrating radar scanning data, the shallow sparse density measurement value, and the measurement location indication mask are downsampled, and the intermediate features of the downsampled data are weighted and modulated based on the gated weight map to enhance the feature response of the effective area of ​​the measurement value. The decoder gradually restores the spatial resolution through transposed convolution and outputs a high-resolution feature map; based on the high-resolution feature map, it outputs the equivalent dielectric constant prediction field through the first prediction branch; and it performs global feature aggregation on the high-resolution feature map through the second parameter branch and outputs the medium state parameters. In the training of the inversion model, the training loss function includes a dielectric parameter loss term, a regional mean loss term, and a frequency domain spectrum loss term; Among them, the dielectric parameter loss term is constructed based on the equivalent dielectric constant obtained from the simulation and the predicted equivalent dielectric constant; For each measuring point, the predicted density within the corresponding mapping area is taken as the regional average value, and the error is calculated by comparing it with the actual measured density of the corresponding measuring point to construct the regional mean loss term. A two-dimensional power spectrum is obtained by performing a two-dimensional Fourier transform on the density field, and a one-dimensional spectral feature is obtained by radial averaging of the two-dimensional power spectrum. A frequency domain spectral loss term is constructed based on the one-dimensional spectral feature and a preset target spectral feature to constrain the spectral shape of the predicted density field. The preset target spectral feature is obtained by statistical analysis of training data, and the constraint object is the shape of the radially averaged power spectrum curve.

2. The density measurement method as described in claim 1, which integrates shallow sparse density measurement data and ground-penetrating radar data, is characterized in that... The process of constructing the training data for the inversion model is as follows: A three-dimensional non-uniform soil medium model was constructed based on the stochastic medium theory. The three-dimensional non-uniform soil medium model was sliced ​​in two dimensions along the preset survey line direction to extract a series of two-dimensional dielectric constants. The two-dimensional dielectric constant distribution map is input into the finite-difference time-domain electromagnetic simulation solver, and the corresponding ground-penetrating radar scanning data is generated by simulating the propagation and scattering process of radar waves in a non-uniform medium. The corresponding medium state parameters are calculated based on the soil particle specific gravity, mass moisture content, water dielectric constant, and soil dielectric constant.

3. The density measurement method as described in claim 1, which integrates shallow sparse density measurement data and ground-penetrating radar data, is characterized in that... It also includes on-site calibration and updating of medium state parameters, specifically: adjusting the medium state parameters with the regional mean error as the update target, so that the regional average value of the predicted density within the measurement point mapping area is consistent with or the difference between the predicted density and the corresponding measured density is reduced.

4. A density measurement system that integrates shallow sparse density measurement data and ground-penetrating radar data, characterized in that, include: The acquisition module is configured to acquire ground-penetrating radar scan data and shallow sparse-dense density measurement data of the area to be measured. The alignment module is configured to: map the shallow sparse density measurement data to shallow sparse density measurement values ​​aligned with the ground-penetrating radar scan data, and generate a measurement position indication mask corresponding to the shallow sparse density measurement values; specifically: Based on the physical sampling depth and the equivalent wave velocity of the medium in shallow sparse density measurement data, the corresponding effective ground-penetrating radar two-way travel time window is calculated. The number of pixel rows in the corresponding image is determined based on the effective ground-penetrating radar two-way travel time window. Within the pixel range, the normalized shallow sparse density measurement data is filled into the measured sparse density measurement point positions and corresponding depth ranges, and 0 is filled into the remaining positions to obtain the shallow sparse density measurement values. The effective area of ​​the shallow sparse density measurement value is marked as 1, and the remaining area is marked as 0, thus obtaining the measurement position indication mask; The inversion module is configured to: input the ground-penetrating radar scan data, the shallow sparse density measurement value, and the measurement location indication mask into a trained inversion model, predict the equivalent dielectric constant, and estimate the medium state parameters; wherein, the medium state parameters are used to characterize the conversion relationship between the equivalent dielectric constant and the density; the inversion model adopts an encoder-decoder structure, the encoder includes multiple cascaded downsampling stages, and the processing procedure of the first downsampling stage for the ground-penetrating radar scan data, the shallow sparse density measurement value, and the measurement location indication mask is as follows: Based on the measurement position indication mask, a gating weight map is generated through a mask gating mechanism; The ground-penetrating radar scan data, the shallow sparse measurement values, and the measurement location indication mask are downsampled, and the intermediate features of the downsampled data are weighted and modulated based on a gated weight map to enhance the feature response of the effective region of the measurement values. The decoder gradually restores the spatial resolution through transposed convolution and outputs a high-resolution feature map; based on the high-resolution feature map, it outputs the equivalent dielectric constant prediction field through the first prediction branch; and it performs global feature aggregation on the high-resolution feature map through the second parameter branch and outputs the medium state parameters. In the training of the inversion model, the training loss function includes a dielectric parameter loss term, a regional mean loss term, and a frequency domain spectrum loss term; Among them, the dielectric parameter loss term is constructed based on the equivalent dielectric constant obtained from the simulation and the predicted equivalent dielectric constant; For each measuring point, the predicted density within the corresponding mapping area is taken as the regional average value, and the error is calculated by comparing it with the actual measured density of the corresponding measuring point to construct the regional mean loss term. A two-dimensional power spectrum is obtained by performing a two-dimensional Fourier transform on the density field, and a one-dimensional spectral feature is obtained by radial averaging of the two-dimensional power spectrum. A frequency domain spectral loss term is constructed based on the one-dimensional spectral feature and a preset target spectral feature to constrain the spectral shape of the predicted density field. The preset target spectral feature is obtained by statistical analysis of training data, and the constraint object is the shape of the radially averaged power spectrum curve. The calculation module is configured to: obtain the density field of the region to be measured based on the equivalent dielectric constant and the dielectric state parameters, and then generate the dense region imaging result.

5. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-3.

7. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-3.

Images