Depth regression-based multi-frequency radar thickness compaction degree inversion system and method

Abstract

The application relates to the field of road construction detection, and discloses a multi-frequency radar thickness compaction degree inversion system and method based on depth regression, which comprises the following modules: a multi-frequency data acquisition and positioning module, which is used for collecting ground penetrating radar echo data of at least two different frequency bands or center frequencies and synchronously acquiring sampling position information; a cross-frequency alignment module, which is used for time zero point and phase reference alignment of echo of different frequency bands; a preprocessing and quality control module, which is used for preprocessing of the aligned echo; an interface parameter extraction module, which is used for picking up interlayer interfaces to obtain propagation time parameters and estimate equivalent dielectric parameters; a cross-frequency feature construction and fusion module, which is used for constructing a cross-frequency feature tensor; and a depth regression inversion module, which is used for inputting the cross-frequency feature tensor into a multi-task depth regression model to output thickness prediction values and compaction degree prediction values. The application can improve the comparability of data, guarantee traceability, and realize closed-loop control on site.

Description

Technical Field

[0001] This invention relates to the field of road construction inspection, specifically to a multi-frequency radar thickness compaction inversion system and method based on depth regression. Background Technology

[0002] Quality control during asphalt pavement construction typically revolves around whether the structural layer thickness meets design requirements and whether compaction quality meets specified standards. Thickness deviations directly affect structural load-bearing capacity and lifespan; insufficient compaction often manifests as high porosity, water permeability sensitivity, and decreased resistance to rutting and fatigue. Therefore, the industry has long pursued the creation of more continuous and timely quality information at construction sites to support adjustments to construction parameters and quality traceability. Simultaneously, with the advancement of intelligent compaction and digital construction, the evolution of on-site quality control from "spot checks at a small number of points" to "process monitoring and spatial distribution evaluation" has become a clear trend.

[0003] Current methods for testing thickness and compaction quality can be broadly categorized into several types. The first type is traditional sampling-based testing, such as core drilling for thickness measurement, and laboratory or field methods for determining density and porosity, which provides relatively direct "point-to-point true values." The second type is rapid on-site density testing, among which nuclear density meters are widely used in asphalt mixture construction quality control. They are rapid, relatively non-destructive density measurement devices, and corresponding test methods exist in many local regulations / standards (such as the nuclear method standard for in-situ density of asphalt mixtures). The third type is alternatives to the nuclear method, including non-nuclear density meters or electrical density meters based on electromagnetic / impedance principles. Some research and engineering evaluation work revolves around their applications. The applicability, influencing factors, and alternative feasibility of hot-mix asphalt and other materials are investigated; the fourth category is construction process monitoring technology, such as intelligent compaction (IC), which integrates a measurement system and an on-board computing / recording system on a vibratory roller, and combines GNSS positioning to form a spatial distribution map of "compaction measurement value + pass trajectory" for process control and quality management; the fifth category is electromagnetic non-destructive testing and continuous evaluation technology, such as ground penetrating radar (GPR), which can estimate the thickness of structural layers through electromagnetic wave response, and based on the relationship between dielectric properties and material volume fraction, it can be used to infer compaction-related indicators under certain conditions; related studies also indicate that thickness inversion usually requires the estimation or determination of dielectric constant.

[0004] The aforementioned technical approaches still face several common limitations in engineering implementation. While sampling methods (such as core drilling) can provide relatively reliable location information, they are destructive or disruptive to construction, and their coverage is limited, making it difficult to reflect continuous spatial distribution. Nuclear density meters, although fast and relatively non-destructive, involve radiation safety, regulatory, and operational management requirements, and their engineering applications still rely on standardized calibration and on-site condition control. Non-nuclear density equipment has advantages in terms of "no regulation required / easy deployment," but multiple assessments indicate that their measurement results are affected by factors such as material composition, temperature, moisture content, surface condition, and on-site operating conditions, often requiring project-level calibration and control of influencing factors. Stability and substitutability remain key concerns in engineering projects. Intelligent compaction technology can provide process data with a wider coverage, but its core output is usually "compaction measurement values ​​(IC-MV, etc.)" related to material stiffness / response rather than intrinsic indicators such as density or porosity. In engineering, it is still necessary to establish the correlation between these values ​​and the target quality indicators and to handle differences in working conditions. When ground-penetrating radar is used for thickness or compaction-related inference, the inversion results are highly sensitive to the estimation of dielectric constant, which is significantly affected by the volume fraction of material components. At the same time, factors such as moisture can change the electromagnetic response and introduce uncertainties. Therefore, achieving stable and traceable numerical inversion under different material sources, moisture content, and temperature conditions remains challenging. To address these issues, this invention provides a multi-frequency radar thickness and compaction degree inversion system and method based on depth regression. Summary of the Invention

[0005] Based on the above description, the present invention provides a multi-frequency radar thickness and compaction inversion system and method based on depth regression to solve the problem that existing methods for detecting and evaluating road thickness and compaction are difficult to achieve continuous, stable and traceable full-domain numerical inversion and real-time closed-loop control under complex working conditions.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a multi-frequency radar thickness compaction inversion system based on depth regression, comprising: The multi-frequency data acquisition and positioning module is used to acquire ground-penetrating radar echo data of at least two different frequency bands or center frequencies and simultaneously obtain sampling location information. The cross-frequency alignment module is used to align the time zero point and phase reference of echoes from different frequency bands. The preprocessing and quality control module is used to preprocess the aligned echoes and uses the cross-frequency coherence index as a necessary criterion for anomaly detection to remove or reduce the weight of abnormal echoes, and outputs removal marks or weight information; wherein, the cross-frequency coherence index is determined by the time window corresponding to the inter-layer interface. The interface parameter extraction module is used to pick up the propagation time parameters from the interlayer interface and estimate the equivalent dielectric parameters. The cross-frequency feature construction and fusion module is used to construct cross-frequency feature tensors based on multi-frequency effective echoes and propagation time parameters or equivalent dielectric parameters. The deep regression inversion module is used to input the cross-frequency feature tensor into the multi-task deep regression model to output the thickness prediction value and the compaction degree prediction value. The thickness prediction value is subject to consistency constraints or consistency checks with the physical thickness obtained from the propagation time parameter and the equivalent dielectric parameter, and the compaction degree prediction value is limited to the density estimate obtained by the equivalent dielectric parameter through field calibration mapping and the reference density. The uncertainty assessment and spatial mapping output module is used to generate the spatial continuous distribution of thickness, compaction degree and their confidence or uncertainty, and output low confidence area markers and retest or supplementary test prompts based on the confidence or uncertainty for use in on-site quality closed-loop control.

[0007] Through the above technical solutions, this system uses multi-frequency radar to collect and locate the spatially continuous distribution of thickness and compaction, and performs cross-frequency alignment and preprocessing on echoes from different frequency bands to improve data comparability. It uses the cross-frequency coherence of the interlayer interface time window as a necessary basis for anomaly detection, eliminating or downweighting abnormal echoes to suppress coupling anomalies and interference from operating condition fluctuations. Based on the propagation time and equivalent dielectric constant of the interface, a cross-screen feature tensor is constructed, and multi-task deep regression is used to simultaneously invert thickness and compaction. Thickness is verified or constrained by physical consistency to ensure its rationality, while compaction is determined by the density estimate from dielectric calibration mapping combined with a reference density to ensure traceability. The system outputs confidence or uncertainty, providing prompts for low confidence and retesting, thus achieving closed-loop control on-site.

[0008] Based on the above technical solution, the present invention can be further improved as follows.

[0009] Furthermore, the multi-frequency data acquisition and positioning module adopts at least one of step frequency, linear frequency modulation or multi-channel multi-frequency combination to achieve multi-frequency acquisition, and at least two different frequency bands or center frequencies meet the preset frequency band difference configuration to simultaneously take into account penetration depth and resolution. The cross-frequency alignment module controls the time zero-point alignment error of echoes from different frequency bands within a preset upper limit, and uses the aligned co-position echoes as a prerequisite for calculating the cross-frequency coherence index. The preprocessing and quality control module includes background removal, bandpass filtering, time zero-point correction, and gain compensation.

[0010] Furthermore, the preprocessing and quality control module includes an abnormal echo identification unit. The abnormal echo identification unit identifies anomalies based on at least one of the following criteria: abnormal antenna coupling, echo saturation, sudden noise energy change, and excessive time zero drift. It must also identify anomalies based on the cross-frequency coherence criterion based on the time window. The time window includes at least an interface time window determined by the interface picking result of the interface parameter extraction module, and a reference time window set separately from the interface time window; the cross-frequency coherence criterion is to calculate the cross-frequency coherence index and form a coherence comparison value in the interface time window and the reference time window respectively; when the coherence comparison value is lower than the threshold, the corresponding echo must be removed or given a weight lower than the preset upper limit to participate in the construction of the cross-frequency feature tensor, and the removal mark or weight information is output to the cross-frequency feature construction and fusion module and the deep regression inversion module, so that low coherence echoes cannot participate in the inversion in an unweighted state.

[0011] Furthermore, the deep regression inversion module includes a cross-frequency attention fusion layer and a multi-scale convolutional feature extraction layer, which are used to perform weighted fusion of information from different frequency bands; The cross-frequency attention fusion layer and the multi-scale convolutional feature extraction layer use coherence contrast value as a gate factor and weight information as a weighting factor to suppress the feature channels corresponding to low coherence echoes. The cross-frequency feature tensor includes at least time-domain energy features, frequency-domain amplitude features, cross-frequency amplitude ratio features, cross-frequency phase difference features, and cross-frequency coherence features and coherence comparison features corresponding to the interface time window; and a joint loss function is used to simultaneously constrain the thickness prediction error and the compaction degree prediction error, and includes a physical consistency term to constrain the consistency between the thickness prediction value and the physical thickness.

[0012] Furthermore, the interface parameter extraction module further includes a field calibration submodule. The field calibration submodule obtains the true value of the core thickness or density based on a preset number of calibration points, establishes or updates the field calibration mapping relationship between the equivalent dielectric parameter and the thickness and density, and determines or updates the coherence comparison value threshold and weighting rules based on the calibration point data, so that the abnormal echo rejection or weighting processing under different projects, different materials and different working conditions has a traceable threshold source.

[0013] Furthermore, the uncertainty assessment and spatial mapping output module obtains the confidence or uncertainty of the thickness prediction and compaction degree prediction values ​​using at least one of model integration, Monte Carlo random deactivation, or predictive distribution estimation. It also uses the cross-frequency coherence index, coherence comparison value, and elimination markers or weight information as mandatory inputs for confidence generation or threshold determination. When the cross-frequency coherence index is below the threshold, the coherence comparison value is below the threshold, or the elimination marker is true, it directly outputs a low-confidence region marker and a retest or supplementary test prompt. The retest or supplementary test prompt and the low-confidence region marker are then superimposed onto the spatially continuous distribution results.

[0014] This invention also provides a method for inverting the thickness and compaction of multi-frequency radar based on depth regression, including the aforementioned multi-frequency radar thickness and compaction inversion system based on depth regression, comprising the following steps: S1. Collect at least two different frequency bands or center frequencies of ground-penetrating radar echo data in the road section to be detected, and simultaneously collect the location information corresponding to each echo sampling location. S2. Cross-frequency alignment and preprocessing are performed on echoes from different frequency bands to obtain multi-frequency effective echoes; S3. Based on the multi-frequency effective echo, interlayer interface picking is performed to obtain propagation time parameters and estimate equivalent dielectric parameters; S4. Determine the interface time window based on the inter-layer interface picking results and set a separate reference time window. Calculate the cross-frequency coherence within the interface time window and the reference time window and form a coherence comparison value. When the coherence comparison value is lower than the threshold, perform elimination or weight reduction processing on the corresponding echo data and generate elimination marks or weight information. S5. Construct a multi-frequency feature tensor containing coherence comparison information and input it into a multi-task deep regression model. Output the structural layer thickness and compaction values ​​corresponding to the sampling positions. Perform physical consistency verification or introduce physical consistency constraints on the thickness values. The compaction values ​​are limited to the density estimate obtained by on-site calibration mapping of the equivalent dielectric parameters and the reference density. S6. Output the confidence level or uncertainty of the thickness and compaction values, and use the removal mark or weight information as one of the criteria for low confidence. Combine the location information to generate spatially continuous distribution results and prompts for retesting or supplementing low confidence areas.

[0015] Furthermore, the cross-frequency alignment in S2 includes at least time zero-point cross-frequency alignment and phase reference alignment; the preprocessing in S2 includes at least one of background removal, bandpass filtering, time zero-point correction, and gain compensation. The coherence comparison value in S4 is obtained in the following way: the cross-frequency coherence index between echoes in the same position of different frequency bands is calculated within the interface time window, and the corresponding cross-frequency coherence index is calculated within the reference time window. The ratio or difference between the two is then used as the coherence comparison value. When the coherence comparison value is lower than the threshold, the corresponding echo data must be removed or assigned a weight lower than the preset upper limit to participate in the subsequent feature tensor construction. The weight information in S4 is used at least to constrain the feature fusion contribution at that position.

[0016] Furthermore, the multi-frequency feature tensor in S5 includes at least one of the following: time-domain energy feature, frequency-domain amplitude feature, cross-frequency amplitude ratio feature, cross-frequency phase difference feature, and cross-frequency coherence feature and coherence contrast feature corresponding to the interface time window; The multi-task deep regression model in S5 performs gating or weighting on the weight information, so that the feature channels corresponding to the downweighted echoes are suppressed in cross-frequency fusion. The physical consistency check or constraint in S5 is used to ensure that the thickness value is consistent with the physical thickness derived from the propagation time parameter and the equivalent dielectric parameter. The reference density in S5 is the baseline density or standard density corresponding to the mixture to be tested. The density estimate is obtained through the on-site calibration mapping relationship between the equivalent dielectric parameter and the density.

[0017] Furthermore, the multi-task deep regression model in S5 is obtained through supervised learning training, and the training sample labels are obtained by ground truth detection. Ground truth detection includes one of core thickness measurement, density detection, and porosity detection. Based on ground truth detection, the equivalent dielectric parameter and the field calibration mapping relationship between thickness and density are established or updated, and the coherence comparison value threshold and weighting rules are determined or updated based on ground truth detection. For different projects, material ratios, or working conditions, a small number of calibration samples with truth labels are collected to perform small-sample calibration or migration updates on the normalized parameters of the model parameters or feature tensors of the multi-task deep regression model, and the project-level calibration parameters are saved. Based on the spatially continuous distribution results and their confidence or uncertainty, on-site closed-loop control output is performed. The closed-loop control output includes the location and graded alarm of thickness anomaly area, the location and graded alarm of compaction anomaly area, the prompt for retesting or supplementing low confidence area, and the quality result map or report with spatial coordinates.

[0018] Compared with the prior art, the technical solution of this application has the following beneficial technical effects: 1. This invention achieves a one-to-one correspondence between thickness and compaction results and spatial location by binding multi-frequency ground-penetrating radar echo acquisition and positioning. It also performs time zero-point and phase reference alignment on echoes of different frequency bands, and performs preprocessing such as background removal, bandpass filtering, zero-point correction and gain compensation to make cross-frequency data comparable and consistent. This results in outputting continuous distribution results at the road segment level and improving the inversion stability and repeatability under complex working conditions such as different materials, temperatures and water content. 2. This invention calculates cross-frequency coherence using the time window corresponding to the interlayer interface, and uses the coherence comparison between the interface time window and the reference time window as the essential basis for anomaly judgment. Low coherence echoes are forcibly removed or downweighted, and the removal mark or weight is integrated into the feature construction, fusion gating and deep regression inference process to avoid abnormal echoes from participating in the inversion without downweighting. This mechanism suppresses the contamination of the results by coupling anomalies, saturation, noise mutations and zero-point drift. At the same time, the thickness output is introduced with physical consistency verification and constraints of thickness-time-dielectric, and the compaction degree is limited to the density estimate obtained by the equivalent dielectric parameter through on-site calibration and mapping combined with the reference density, so that the inversion results are more in line with physical laws and can be traced and corresponded with the engineering quality indicators. 3. This invention outputs the uncertainty or confidence level of thickness and compaction, and uses the coherence index, coherence comparison value, and rejection mark or weight as the basis for low confidence judgment. It directly generates low confidence area marks and retest prompts and superimposes them on the spatial results, thereby realizing closed-loop control of on-site quality. At the same time, it establishes or updates the calibration mapping relationship between dielectric and thickness and density by combining the true values ​​of core thickness and density, and simultaneously determines the threshold and weight rules. It also supports project-level calibration or migration update of a small number of calibration samples to reduce the risk of cross-project drift and improve deployment efficiency and engineering promotion. Attached Figure Description

[0019] Figure 1 An overall system block diagram of a multi-frequency radar thickness compaction inversion system and method based on depth regression provided in an embodiment of the present invention; Figure 2 This is a system schematic diagram of the preprocessing and quality control module in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of ground-penetrating radar thickness measurement in an embodiment of the present invention; Figure 4 This is a schematic diagram of the scanning and result denoising processing data of the ground penetrating radar in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the interface between asphalt and the base layer represented by a band signal in an embodiment of the present invention; Figure 6 This is a system block diagram of the uncertainty assessment and spatial mapping output module according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the system flow for anomaly identification by the preprocessing and quality control module in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the cooperation between the interface parameter extraction module and the uncertainty assessment and spatial mapping output module in an embodiment of the present invention. Figure 9 This is a schematic diagram of the system flow of Embodiment 2 of the present invention; Figure 10This is a flowchart illustrating the multi-frequency radar thickness compaction inversion method based on depth regression in an embodiment of the present invention. Detailed Implementation

[0020] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0021] Unless otherwise defined, 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 application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. Example 1:

[0022] refer to Figure 1 A multi-frequency radar thickness compaction inversion system based on depth regression includes: a multi-frequency data acquisition and positioning module for acquiring ground-penetrating radar echo data from at least two different frequency bands or center frequencies and simultaneously obtaining sampling location information; a cross-frequency alignment module for aligning the echoes of different frequency bands with time zeros and phase references; a preprocessing and quality control module for preprocessing the aligned echoes and using cross-frequency coherence indices as a necessary criterion for anomaly detection to remove or reduce the weight of abnormal echoes, and outputting removal marks or weight information; wherein, the cross-frequency coherence indices are determined by a time window corresponding to the interlayer interface; an interface parameter extraction module for picking up the propagation time parameters from the interlayer interface and estimating the equivalent dielectric parameters; and a cross-frequency feature construction and fusion module. The module is used to construct a cross-frequency characteristic tensor based on multi-frequency effective echoes and propagation time parameters or equivalent dielectric parameters; the deep regression inversion module is used to input the cross-frequency characteristic tensor into a multi-task deep regression model to output thickness prediction values ​​and compaction degree prediction values. The thickness prediction value is subject to consistency constraints or consistency checks with the physical thickness obtained from the propagation time parameters and equivalent dielectric parameters, and the compaction degree prediction value is limited to the density estimate obtained by on-site calibration mapping of the equivalent dielectric parameters and the reference density; the uncertainty assessment and spatial mapping output module is used to generate the spatial continuous distribution of thickness, compaction degree and their confidence or uncertainty, and output low confidence area markers and retest or supplementary test prompts based on the confidence or uncertainty for on-site quality closed-loop control.

[0023] In this embodiment, the multi-frequency data acquisition and positioning module can be implemented using a vehicle-mounted air-coupled horn antenna ground-penetrating radar system. For example, the GSSI RoadScan 30 can be used as the platform, with antenna options including 1GHz and 2GHz horn antennas to generate echoes at least two different center frequencies. The system can also include a wheel-mounted distance measuring instrument to generate mileage information synchronously with the echo frames. For synchronously acquiring sampling location information, the Trimble R12i GNSS system from Trimble can be used as the RTK positioning receiver to output coordinates and time references, so as to bind the radar echo frames with mileage coordinates and absolute coordinates. Furthermore, the mileage pulses can be provided by an incremental encoder, such as the SICK DFS60 series incremental encoder installed on the ranging wheel or wheel axle to output distance pulses, forming a synchronous sampling link of "echo frame-mileage-coordinates".

[0024] In this embodiment, the cross-frequency alignment module can be deployed in an onboard industrial computer or edge computing unit to perform time zero-point alignment and phase reference unification on the echoes of the 1GHz and 2GHz frequency bands. To ensure the physical meaning of the cross-frequency coherence calculation, this embodiment uses the "aligned co-position echo" as a prerequisite for subsequent coherence calculations and records the alignment residual as a traceable field in the metadata of each sampling point so as to locate the source of the anomaly during retesting, supplementary testing, or quality traceability.

[0025] refer to Figure 2 The preprocessing and quality control module is preferably deployed on an edge-side industrial control computer or embedded computing unit, for each sampling point. Multi-frequency echoes are processed in real time and marked for removal. With weight Two types of fields output quality conclusions, among which , Indicates removal. Let represent the weighting coefficient, and satisfy the constraint condition: when There will always be a time for it. The module's internal functions are divided into a preprocessing unit, a time window generation unit, a cross-frequency coherence calculation unit, an anomaly detection and weight generation unit, and a work condition recording unit. The work condition recording unit is optional and is used to form a traceable chain of operational evidence. To enhance the stability of project implementation, this embodiment can be configured with a road surface temperature sensor as the input for work condition recording. For example, the Optris CS LT infrared temperature sensor can be used to collect road surface temperature and store it synchronously with the echo frame according to the timestamp or mileage number. This is used to interpret the differences in electromagnetic response and anomaly detection caused by temperature changes.

[0026] Furthermore, the preprocessing unit performs a consistent preprocessing procedure on the echoes of each frequency band after cross-frequency alignment, assuming the first... Each frequency band at the sampling point The original echo is The echoes after preprocessing and cross-frequency alignment are uniformly denoted as Preprocessing includes at least background removal, bandpass filtering, time zero-point correction, and gain compensation, as detailed below: Background removal: Calculate the background component for multiple echoes within adjacent mileage ranges in the same frequency band and subtract it from the current echo to suppress system coupling, slow-varying noise floor, and slow platform drift; the background window length is set according to vehicle speed, sampling interval, and road structure change scale. Bandpass filtering: The passband is set based on the center frequency of each frequency band to filter out out-of-band noise and low-frequency drift, ensuring that the cross-frequency echo can participate in subsequent analysis within the comparable frequency band; Zero-point correction: The arrival time of the first wave of the direct wave or coupled wave is used as the zero-point reference. The zero-point position is determined by threshold triggering, energy mutation detection or correlation matching. The waveform is then time-shifted to make the zero-point definition consistent with different sampling points and different frequency bands, thereby ensuring the effectiveness of time window positioning and coherence calculation. Gain compensation: Time-varying gain or segmented automatic gain control is used to improve the visibility of deep reflections. The gain parameter is kept consistent within the same frequency band to avoid additional disturbances to cross-frequency consistency.

[0027] Furthermore, the time window generation unit determines the time window corresponding to the inter-layer interface based on the inter-layer interface position, which is derived from the interface two-way propagation time output by the interface parameter extraction module. The time window generation unit generates at least two types of time windows: one is the interface time window. The first is used to cover the main energy range of reflections at the interlayer interface; the second is a reference time window. This is used to select a baseline within a relatively stable and weak reflection event range outside the interface time window; the widths of the interface time window and the reference time window are set based on the sampling rate, waveform main lobe width, and the thickness range of the layer to be measured, and are adaptively updated as the construction layer structure changes; if interface picking fails or This is unreliable; in this embodiment, the sampling point is directly set to a discard state, and the output is... and This is to avoid misjudgments caused by incorrect time windows.

[0028] Furthermore, the cross-frequency coherence calculation unit in the interface time window The cross-frequency coherence is calculated internally and used as an essential criterion for anomaly detection. For ease of engineering implementation, this embodiment uses a normalized cross-correlation form of coherence index for any two frequency bands. and Within a given time window Internal definition: ;in The value ranges from 0 to 1. A larger value indicates that the waveforms of the two frequency bands are more consistent within the time window. For cases with more than two frequency bands, this embodiment aggregates the coherence results of all frequency band pairs to obtain the interface window coherence aggregate value. Coherence convergence value with reference window The convergence method is fixed at one level within the same project: selecting the minimum value for more conservative anomaly control, and selecting the average value for smoother spatial output; based on this, to enhance robustness to slow changes in operating conditions, this embodiment calculates a coherence comparison value. As a reference for judgment: ;in To prevent the use of tiny constants with a denominator of zero; coherence comparison values. It is used to characterize the degree of change in interface consistency relative to baseline consistency, thereby reducing the impact of overall drift such as temperature, water content, and coupling conditions on the judgment threshold.

[0029] Furthermore, the anomaly detection and weight generation unit uses and As a necessary basis for final processing, it outputs a removal flag. With weight The rules can be implemented in the following way: Hard rejection rule: when Below the threshold ,or Below the threshold When this happens, the sampling point is identified as an abnormal echo, and the output is... Simultaneously output This prevents the point from participating in subsequent inversion calculations; weight reduction rule: when When the sample point is close to the threshold but does not meet the hard rejection condition, it is determined to be usable but lacks confidence, and the output is... At the same time Set to a continuous or graded value between 0 and 1, and Follow The number of levels and the threshold interval are determined by on-site calibration results or project experience and recorded in the system configuration for traceability. Engineering anomaly and coherence linkage rules: When engineering anomaly clues such as saturation, shearing, extremely low signal-to-noise ratio, or alignment residual exceeding limits are detected, the point is first listed as a candidate anomaly, and then calculated within the interface time window. and And based on this, decide on the output. and This ensures that any anomaly handling is confirmed through cross-frequency coherence, thereby meeting the requirement that cross-frequency coherence is a necessary basis for anomaly determination.

[0030] Finally, the output of the preprocessing and quality control module includes: the preprocessed multi-frequency echo sequence. Interface time window parameters Reference time window parameters Interface window coherence convergence value Reference window coherence convergence value Coherence comparison value Remove markers Weight and operating condition record fields (such as road surface temperature); among which and The data is output along with the sample point data as an adjunct field and serves as a mandatory input for subsequent cross-frequency feature construction and fusion modules, deep regression inversion modules, and uncertainty assessment modules. It is used for sample weighting, gating suppression, and output credibility interpretation to prevent low-quality echoes from affecting the inversion results in an unweighted state, while providing a clear, quantitative, and traceable basis for anomaly handling.

[0031] In this embodiment, the interface parameter extraction module is used to pick up the interlayer interface (such as the top-to-bottom boundary or interlayer interface) from the echo and output the interface two-way propagation time. Further estimate the equivalent relative permittivity Equivalent relative permittivity This is used to map electromagnetic propagation characteristics to material states and geometric parameters, and serves as an important intermediate quantity for thickness and compaction inversion; when interface picking fails or the interface response confidence is insufficient, this embodiment submits the point to the preprocessing and quality control module for hard rejection, and outputs... , This is reflected as a low-confidence point in subsequent uncertainty assessments and spatial mapping outputs.

[0032] refer to Figure 3 and Figure 4 To further supplement the principle of ground-penetrating radar (GPR) thickness measurement, the following is provided: First, GPR measures the incident and reflected waves; second, the dielectric constant of asphalt is calculated using the incident and reflected waves, i.e., high-frequency electromagnetic waves are emitted to the ground, the reflected wave signal from the road surface is acquired through a receiving antenna, the amplitudes of the incident and reflected waves are measured, and the relative dielectric constant is derived based on the reflection mechanism; third, the asphalt compaction degree and paving thickness are calculated using the relative dielectric constant. When compaction degree increases, porosity decreases while the dielectric constant increases; conversely, when compaction degree decreases, porosity increases while the dielectric constant decreases. By measuring the dielectric constant and establishing the relationship between compaction degree and dielectric constant, rapid, reliable, safe, and non-destructive testing is achieved; fourth, GIS and IoT technologies are integrated for real-time measurement of asphalt compaction degree and thickness. Regarding thickness calculation, this embodiment uses sampling points... Using a physical thickness model: ;in The speed of light in a vacuum. For the two-way propagation time of the interface, It is the equivalent relative permittivity.

[0033] It should be noted that, Figure 4 This is a schematic diagram of the ground-penetrating radar (GPR) scanning process and the denoising data of the results. Figure 5 This is a schematic diagram of the interface between asphalt and the base layer represented by a band signal. By measuring the amplitude ratio of the incident wave and the reflected wave from the interface, the equivalent dielectric constant of the layer can be calculated. Under near-normal incident conditions, the reflection coefficient is determined by the dielectric constant. In actual calculations, the short time window energy ratio is often used instead of the single-point amplitude ratio to reduce the impact of noise. Specifically, after completing the raw data acquisition of the ground penetrating radar, the signal first needs to be preprocessed. After suppression processing by background removal and bandpass filtering, noise reduction processing is performed using methods such as curve transform denoising or MLP denoising. Both methods can be used simultaneously for noise reduction.

[0034] In this embodiment, the cross-frequency feature construction and fusion module will integrate multi-frequency effective echoes. Interface two-way propagation time Equivalent relative permittivity The organization is a cross-frequency feature tensor, and the quality control output field is used. and As a gating factor and sample weight, it is input together to achieve system-level consistency in the transmission of quality information along with the data stream; the organization of the cross-frequency feature tensor is composed of frequency band dimension, time sampling dimension, and feature channel dimension, enabling the model to learn cross-frequency complementary information and suppress single-frequency instability factors; where when or At that time, the sampling point is gated as an invalid input on the feature side to prevent it from entering the inversion main link.

[0035] In this embodiment, the deep regression inversion module is deployed on the vehicle-mounted edge computing unit and uses sampling points To handle granular operations, the input is a cross-frequency feature tensor and the same point. , , , Its output includes at least the predicted thickness value. With compaction degree It also simultaneously outputs intermediate and verification quantities for traceability, including physical thickness. Thickness consistency deviation Density estimates Reference density Source identifier.

[0036] It should be noted that the model input is organized as follows: the cross-frequency feature tensors are organized according to the frequency band dimension, the time sampling dimension, and the feature channel dimension, and each sampling point corresponds to a set of tensors; to ensure that the quality information has a rigid constraint on the inversion, this embodiment uses weights Used for multiplicative gating, multiplying the overall features of the corresponding sample points by . ,when If the point is not included in the inference process, then the confidence level of the output at that point will be determined if the engineering implementation requires a mandatory output. Set it to 0 and mark it as a low-confidence point in the spatial mapping.

[0037] To further explain, the multi-task deep regression model structure is as follows: The model adopts a multi-task structure with a shared backbone network and dual output heads. The shared backbone network is used to extract the joint representation of the cross-frequency and time dimensions, and the thickness output head is used to regress the thickness prediction value. The compaction output head is used to output the intermediate predicted value of compaction. The learning process incorporates working condition compensation and residual features. The shared backbone network is implemented as follows: The first stage is a temporal feature extraction layer, which performs one-dimensional convolution stacking on the time series features of each frequency band to extract multi-scale temporal features; the second stage is a cross-frequency fusion layer, which performs cross-frequency attention fusion on the temporal representations of each frequency band; the third stage is a global aggregation layer, which performs pooling or weighted aggregation on the fused representations in the time dimension to obtain sampling point-level feature vectors.

[0038] Further explanation: Physical thickness calculation and consistency processing: This embodiment uses data-driven thickness prediction. With physical thickness Simultaneous calculations are performed in parallel to suppress drift under complex operating conditions. The physical thickness is calculated using the above formula. This embodiment also employs consistency constraints and consistency checks: the consistency constraint is reflected in the loss function during the training phase, making... Close in a statistical sense Consistency verification is reflected in the operational logic of the inference stage, calculating the thickness consistency deviation. ;when Exceeding the preset threshold At that time, output the value at that point. Reduce and record the reasons for exceeding the limit and related intermediate quantities in the log to facilitate retesting and traceability.

[0039] Further explanation regarding the compaction degree limitation path and calculation method: In this embodiment, the compaction degree finally released by the system is uniformly defined as... And calculate strictly according to the defined path: First, the interface parameter extraction module outputs... Secondly, the field calibration submodule provides a calibration mapping from dielectric to density, yielding a density estimate. ,in Write the configuration and include the version number; the reference density will then be provided by the project configuration. Record its source identifier and version number; the final compaction degree is determined by the density ratio: When percentage output is required, the system will... Multiply by 100 to get the intermediate predicted value of the compaction output head. This is used for internal learning compensation, but the system uniformly releases the compaction level externally. For reference, and record in the log. , Calibration mapping version number and reference density source.

[0040] To further clarify, traceable records are maintained: the deep regression inversion module records at least the model version number, training data version number, calibration mapping version number, reference density source identifier, and threshold for each task. The mileage range in which the limit was exceeded and the corresponding , , , , It is used for project acceptance and accountability.

[0041] refer to Figure 6The uncertainty assessment and spatial mapping output module receives the thickness and compaction values ​​output by the depth regression inversion module, and simultaneously receives the rejection flag, weights, and coherence comparison values ​​output by the preprocessing and quality control module as "confidence constraint inputs." Furthermore, the confidence level or uncertainty is preferably characterized by the dispersion of repeated inference results, the dispersion of predictions in model ensembles, or the width of the prediction distribution. Repeated inference can be achieved through random deactivation or light perturbations to meet real-time requirements at the edge. Furthermore, the generation of confidence levels follows a hard rule: when the rejection flag is true, the confidence level is high. The confidence level is set to 0 directly. When the weight is less than 1, the upper limit of the confidence level decreases as the weight decreases. When the uncertainty increases or the coherence comparison value decreases, the confidence level decreases further, thus ensuring that low-quality data is explicitly exposed in the result as "low confidence" and providing a direct basis for retesting decisions. Furthermore, spatially continuous distribution results are rasterized or segmented interpolated based on mileage coordinates or geographic coordinates. The interpolation or raster weighting is mainly based on the confidence level, so that high-confidence sampling points contribute more to the layer. A conservative interpolation strategy is adopted for areas with missing data, and the interpolation confidence level is reduced simultaneously to avoid high-confidence display in areas without data support. Furthermore, low-confidence area markers are obtained by spatial clustering of sampling points with confidence levels below the confidence threshold and by removing sampling points marked as true. The retest or supplementary test prompts include at least the start and end mileage or boundary coordinates, the suggested retest method, the suggested number of retests and priorities, and are superimposed on the thickness layer, compaction layer and confidence layer, so that the retest loop can be directly organized on-site according to priority, forming a closed-loop management link of "spatial results - low-confidence prompts - retest execution - result backfilling".

[0042] It should be noted that the multi-frequency data acquisition and positioning module uses at least one of the following methods to achieve multi-frequency acquisition: step frequency, linear frequency modulation, or multi-channel multi-frequency combination. At least two different frequency bands or center frequencies must meet a preset frequency band difference configuration to simultaneously consider penetration depth and resolution. Specifically, in this embodiment, multi-frequency acquisition preferably adopts a frequency band difference configuration of "low frequency band for penetration, high frequency band for resolution": the low frequency band is used to enhance the visibility and anti-attenuation capability of deep interfaces, while the high frequency band is used to enhance the resolution and thickness sensitivity of shallow interfaces. Furthermore, the preset frequency band difference configuration can be fixed in the system configuration in the form of center frequency difference, center frequency ratio, or effective bandwidth difference, and the system automatically verifies the configuration validity before the operation begins to prevent cross-frequency complementarity caused by excessively close frequency bands. The system suffers from insufficient performance. Furthermore, when using a step frequency method, the system acquires multiple sub-frequency responses within the same routing path and reconstructs them according to preset frequency bands, ensuring that echoes from different frequency bands have a unified sampling beat and timestamp. When using a linear frequency modulation method, the system normalizes and demodulates echoes corresponding to different frequency modulation parameters to obtain comparable echo sequences. When using a multi-channel, multi-frequency combination, the system uniformly manages multi-channel triggering and synchronization clocks to ensure the co-location matching of multi-frequency data is established. Furthermore, positioning employs a dual-channel binding of RTK coordinates and odometer pulses: RTK provides spatial coordinates for GIS mapping, and odometer pulses provide uniform sampling intervals for co-location matching and retest alignment, writing "frame number-timestamp-odometer-coordinates" into the accompanying data to meet traceability requirements.

[0043] The cross-frequency alignment module controls the zero-point alignment error of echoes from different frequency bands within a preset upper limit and uses the aligned co-occurrence echoes as a prerequisite for calculating cross-frequency coherence indices. Specifically, zero-point alignment uses the arrival time of the first wave of the direct wave or coupled wave as the zero-point reference. The zero-point position is determined through correlation matching or energy mutation detection, and time shift is performed to ensure that the zero-point definition is consistent for different frequency bands at the same sampling point. Furthermore, the zero-point alignment error is measured by "zero-point residual" or "reference window cross-correlation peak difference" and controlled within a preset upper limit. When the alignment error exceeds the limit, the sampling point first enters the candidate anomaly set, but is not directly eliminated. It must still be subject to final confirmation based on the cross-frequency coherence criterion based on the time window to ensure the consistency of the evidence chain for anomaly handling. Furthermore, the definition of co-occurrence echoes is preferably bound based on mileage numbers: when the sampling point density of different frequency bands is inconsistent, nearest neighbor mileage matching is used and interpolation alignment is performed within a small range to ensure that the "co-occurrence samples" required for cross-frequency coherence calculation are valid. The matching residuals and interpolation markers are written into the accompanying field to explain the sources of difference during retesting. The preprocessing and quality control module's preprocessing includes background removal, bandpass filtering, time zero-point correction, and gain compensation. Specifically, background removal is used to suppress system coupling, slow-varying background noise, and slow platform drift. The background window length is set according to vehicle speed, sampling interval, and structural change scale, and the window parameters are recorded in the job configuration for reproduction. Bandpass filtering sets the passband based on the center frequency and effective bandwidth of each frequency band to ensure that cross-frequency echoes participate in subsequent analysis within comparable frequency bands. Time zero-point correction is used to further eliminate inter-point zero-point micro-drift after cross-frequency alignment to avoid the accumulation of time window positioning errors. Gain compensation uses time-varying gain or piecewise gain to improve the visibility of deep reflections, and maintains a uniform gain rule within the same frequency band to avoid artificially introducing cross-frequency inconsistencies. Furthermore, the preprocessed echoes are entered into the interface picking and coherence calculation process as "multi-frequency effective echoes," while retaining the preprocessing parameter version number to support post-audit.

[0044] refer to Figure 7The preprocessing and quality control module includes an abnormal echo identification unit. This unit identifies anomalies based on at least one criterion among antenna coupling anomalies, echo saturation, noise energy mutations, and time zero-point drift exceeding limits. It must also simultaneously identify anomalies based on a time-window-based cross-frequency coherence criterion. The abnormal echo identification unit employs a two-level judgment mechanism: the first level lists sampling points as candidate anomalies based on engineering anomaly clues; the second level must confirm these anomalies using a time-window-based cross-frequency coherence criterion before outputting removal markers or weight information. Furthermore, antenna coupling anomalies are determined by a combination of direct wave amplitude, coupled wave morphology, and arrival stability; echo saturation is determined by the waveform clipping ratio or the number of continuous saturation samples; noise energy mutations are determined by the rate of energy change within the window and the frequency band energy distribution mutations; and time zero-point drift exceeding limits is determined by the zero-point estimation residual exceeding a preset upper limit. Moreover, the aforementioned engineering criteria only determine whether to proceed to the coherence confirmation stage; the final processing must be based on the coherence criteria, thus ensuring that anomaly identification does not rely on a single sensor state or a single threshold to avoid misjudgment.

[0045] Furthermore, the time window includes at least the interface time window determined by the interface picking results of the interface parameter extraction module, and a reference time window set separately from the interface time window; the cross-frequency coherence criterion is to calculate the cross-frequency coherence index and form a coherence comparison value within the interface time window and the reference time window respectively; when the coherence comparison value is lower than the threshold, the corresponding echo must be removed or assigned a weight lower than the preset upper limit to participate in the construction of the cross-frequency feature tensor, and the removal mark or weight information is output to the cross-frequency feature construction and fusion module and the deep regression inversion module, so that low coherence echoes cannot participate in the inversion in an unweighted state; the interface time window is based on the interface two-way propagation time. The system is centrally configured to cover the main lobe energy range of the interface reflection. The reference time window and interface time window are set separately, preferably located outside the interface window in a region where reflection events are weak and relatively stable, to ensure baseline consistency. Furthermore, the cross-frequency coherence index preferably adopts a normalized cross-correlation form, and the coherence results for all frequency band pairs in multi-band scenarios are aggregated to obtain the interface window coherence and reference window coherence, which are then used to form a coherence comparison value. In engineering implementation, a ratio form is preferred to enhance robustness against overall drift. When the coherence comparison value is below a threshold, the corresponding echo must be discarded or assigned a weight below a preset upper limit; a discard flag is output during discarding. Weight Output during weight reduction , ;in Strictly less than 1 to ensure that low-coherence echoes cannot participate in the inversion in an unweighted state; further, markers are removed. With weight As an accompanying field, it is transmitted along with the sampling point data to the cross-frequency feature construction and fusion module and the deep regression inversion module. On the feature construction side, it is used to constrain the feature fusion contribution of the point, and on the model inference side, it is used for gating suppression and weighting, so that low coherence echoes are consistently weakened or eliminated in the end-to-end link. At the same time, the threshold version number and the weight classification rule version number are written to the log to ensure that the source of the threshold is traceable.

[0046] In addition, the deep regression inversion module includes a cross-frequency attention fusion layer and a multi-scale convolutional feature extraction layer, used to weight and fuse information from different frequency bands. The multi-scale convolutional feature extraction layer is used to extract reflection patterns and interface-sensitive features at different scales in the time dimension. The cross-frequency attention fusion layer is used to adaptively weight information from different frequency bands in the frequency band dimension, so that the penetration advantage of the low-frequency band and the resolution advantage of the high-frequency band complement each other at the same sampling point. Furthermore, the multi-scale convolution can be implemented in parallel or cascaded manner using different convolutional kernel lengths to take into account both the main lobe and secondary reflection structures of the interface. The cross-frequency attention fusion preferably adopts a query-key-value-based weight calculation to make the frequency band The weights can be adaptively adjusted according to changes in materials and working conditions, and the attention weights are written to the debugging log to explain the model behavior. The cross-frequency attention fusion layer and the multi-scale convolutional feature extraction layer use coherence contrast value as a gating factor and weight information as a weighting factor to suppress feature channels corresponding to low coherence echoes. The coherence contrast value is used to provide a direct quantitative basis for "the strength of cross-frequency consistency at this point", and serves as a gating factor to control whether the feature channel enters the high-weight fusion. The weight information serves as a weighting factor to control the contribution ratio of this point in fusion and regression. Furthermore, the gating mechanism can adopt segmented gating: when the coherence contrast value is lower than the threshold, the gating is 0 and it is synchronized with the model. Consistent; when the coherence contrast value is near the threshold, the gating coefficient increases with the increase of the coherence contrast value, while using The feature amplitude or attention weight is scaled so that the point participates in the inference process with low impact, avoiding spike disturbances in the thickness and compaction output caused by low coherence echoes. Furthermore, the above suppression results are consistent with the back-end confidence generation rules, so that "input-side weight reduction - model-side suppression - output-side low confidence" form a closed loop.

[0047] The cross-frequency feature tensor includes at least time-domain energy features, frequency-domain amplitude features, cross-frequency amplitude ratio features, cross-frequency phase difference features, and cross-frequency coherence features and coherence contrast features corresponding to the interface time window. A joint loss function is used to simultaneously constrain the errors in thickness prediction and compaction degree prediction, and a physical consistency term is included to constrain the consistency between the thickness prediction and the physical thickness. In this embodiment, the cross-frequency feature tensor includes at least the following features, which are written into the feature channels in a fixed order to ensure consistency of model input across different items: Time-domain energy features: Interface time window Echo energy, peak amplitude, and energy concentration within the interface time window; frequency domain amplitude characteristics: dominant frequency amplitude and bandwidth energy obtained from spectral analysis of echoes within the interface time window; cross-frequency amplitude ratio characteristics: amplitude ratio and energy ratio of different frequency bands within the interface time window; cross-frequency phase difference characteristics: phase difference statistics of different frequency bands within the interface time window; cross-frequency coherence characteristics: coherence convergence value within the interface window. As a quality-sensitive feature, it is written into the feature channel; coherence contrast feature: coherence contrast value. Write to the feature channel to reduce the impact of overall operating condition drift on threshold determination; by... and By explicitly incorporating feature tensors, the model can perceive the reliability of data while learning thickness and compaction, thereby reducing the pulling effect of abnormal echoes on the model output. Furthermore, the model training employs a joint loss function to simultaneously optimize both thickness and compaction tasks, and introduces a physical consistency term. To control the number of formulas, this embodiment only presents the core form of the joint loss: ;in This is the thickness prediction error term. This is the compaction error term. This is a physical consistency term used for constraints. Close ; The weighting coefficients are determined by parameter tuning on the validation set and written into the model configuration file; the true thickness values ​​in the training data come from the core thickness or known calibration thickness, and the true compaction value is calculated by comparing the true density value with... The calculations show that the loss curve, threshold parameters, and data version number are recorded during the training process to ensure that the results are reproducible and traceable.

[0048] refer to Figure 8 The interface parameter extraction module further includes a field calibration submodule. The field calibration submodule obtains the true value of the core thickness or density based on a preset number of calibration points, establishes or updates the field calibration mapping relationship between the equivalent dielectric parameters and the thickness and density, and determines or updates the coherence comparison value threshold and weighting rules based on the calibration point data, so that the abnormal echo rejection or weighting processing under different projects, different materials and different working conditions has a traceable threshold source.

[0049] In this embodiment, the on-site calibration submodule is used to perform calibration establishment or calibration update under circumstances such as the initial stage of project commencement, material source switching, mix ratio changes, significant changes in construction temperature, and changes in moisture content. This ensures that the project-level mapping relationship of dielectric-thickness and dielectric-density, along with the coherence comparison value threshold and weighting classification rules, remain consistent with the current project, and forms an auditable versioned record. Its specific execution process includes at least the following steps: 1. Calibration point layout: Select a preset number of calibration points on the road section to be tested. The calibration points are evenly distributed along the mileage direction and cover typical working condition sections. The system locks the sampling position at each calibration point and saves the data. , , , , , Metadata such as alignment residuals, operating speed, and road surface temperature; 2. Core sampling and true value determination: Core sampling is performed at each calibration point. The core sampling hole position and radar sampling point are matched one-to-one with the mileage number and coordinates to obtain the true value of thickness and density, and the test method identification, test personnel and test time are recorded. 3. Establishing or updating mapping relationships: based on the calibration point. The results were obtained by fitting the true value of the core density. Used for output The fitting coefficients and version numbers are fixed, and thickness-related mapping parameters are established or updated to assist in thickness verification and drift suppression. After fitting, the fitting residual statistics are recorded to judge the calibration effectiveness. 4. Determination or update of coherence comparison threshold: The set of sampling points near the calibration point that have "alignment residuals not exceeding the limit and good consistency of core data" is used as the reliable sample set, and its statistical analysis is performed. Distribution, and threshold Set as the conservative lower quantile of this distribution, update the threshold, write it to the configuration, and generate the threshold version number; 5. Determining or updating weighted classification rules: (Based on...) Divide the input into several levels and assign corresponding weights. For example, three or five levels, the grading boundary is determined by the credible sample set. The distribution is determined together with the calibration point inversion error distribution, and the weighting rules are saved as independent configuration items and updated synchronously with the threshold version. 6. Traceable Records and Audit Fields: The system will record the number and number of calibration points, coordinates and mileage, true values ​​of core thickness and core density, test method identifier, mapping function type, fitting coefficient, fitting residual statistics, and threshold. The weighted grading boundary, threshold version number, calibration mapping version number, material batch, construction temperature label, moisture content label, and model version number are written into the database or configuration file and saved in an overwriteable manner to form a traceable link.

[0050] It should be noted that the uncertainty assessment and spatial mapping output module obtains the confidence or uncertainty of the thickness prediction and compaction degree prediction values ​​using at least one of the following methods: model integration, Monte Carlo random deactivation, or predictive distribution estimation. The module also uses the cross-frequency coherence index, coherence comparison value, and elimination mark or weight information as mandatory inputs for confidence generation or threshold determination. When the cross-frequency coherence index is below the threshold, the coherence comparison value is below the threshold, or the elimination mark is true, the module directly outputs a low-confidence area mark and a retest or supplementary test prompt. The retest or supplementary test prompt and the low-confidence area mark are then superimposed on the spatial continuous distribution results.

[0051] In this embodiment, the uncertainty assessment and spatial mapping output module completes point-level uncertainty calculation, point-level confidence generation, low-confidence area generation, spatial continuous distribution generation, GIS publishing, and IoT uploading on the edge side. Its output is used for on-site closed-loop scheduling, specifically as follows: 1. Uncertainty acquisition: The system selects one of the following methods to obtain point-level uncertainty: model integration, Monte Carlo random deactivation, or prediction distribution estimation. For ease of real-time deployment, repeated inference is preferred: for the same sampling point, the thickness prediction value and compaction degree output by the depth regression inversion module are repeatedly inferred a preset number of times under the same input conditions to obtain a set of repeated prediction results. The uncertainty is characterized by the dispersion of this set of results. The number of repeated inferences is written as a system parameter into the configuration file and recorded in the work log; 2. Point-level uncertainty calculation and confidence generation: The system Thickness uncertainty and compaction uncertainty are calculated separately. Thickness uncertainty is represented by the dispersion of "multiple thickness prediction results obtained through repeated inference," preferably using variance or standard deviation as the dispersion index. Compaction uncertainty is represented by the dispersion of "multiple compaction results obtained through repeated inference," with the calculation caliber consistent with thickness uncertainty. The point-level confidence level (conf) is jointly determined by thickness uncertainty, compaction uncertainty, weight w, coherence comparison value R, and rejection flag f, and the following hard rules are applied: when rejection flag f is 1, confidence level Conf is directly set to 0; when weight w decreases, the upper limit of confidence level Conf decreases accordingly; when coherence comparison value R decreases, the upper limit of confidence level Conf decreases accordingly; when thickness uncertainty or compaction uncertainty increases, confidence level Conf decreases. The above mapping relationship is fixed in the configuration file, so that the built-in confidence caliber of the same project is consistent and traceable; 3. Low confidence area marking and retesting prompt generation: When the confidence Conf is lower than the confidence threshold T_Conf, or the removal mark f is 1, the sampling point is marked as a low confidence point; the system will cluster the low confidence points according to the mileage continuity or spatial adjacency to form low confidence areas, and the low confidence areas must contain at least the start and end mileage or boundary coordinates.The retest / supplementary test prompts should include at least the suggested retest start and end mileage, suggested retest method, suggested retest number, and suggested priority. Priority is determined by the confidence level (Conf), region length, and whether it is located in a critical control zone, and is written into the retest task list. 4. Spatial Continuous Distribution Generation and Overlay Output: The system projects the predicted thickness, compaction degree, and confidence level (Conf) onto mileage coordinates or geographic coordinates, and generates a spatial continuous distribution using a rasterization method. During rasterization, a weighted average is performed using the confidence level (Conf) as a weighting coefficient, ensuring that high-confidence points contribute more to the raster values. For missing raster values, neighborhood interpolation is used, and the interpolation confidence level is simultaneously reduced; low-confidence areas are marked. The retest and supplementary test prompts are output as overlay layers, allowing the spatially continuous distribution results and low-confidence prompts to be presented simultaneously in the same view; 5. GIS layer generation and IoT publishing: The system generates GIS layer data from the spatial results. The GIS layers include at least a thickness layer, a compaction layer, a confidence layer, a low-confidence area layer, and a retest task layer; The system uploads the layer data and task list to the cloud platform or project server via 4G, 5G, or a private network, and includes the work batch number, timestamp, equipment number, model version number, calibration mapping version number, and threshold version number in each upload, supporting real-time viewing, quality traceability, and closed-loop scheduling on multiple terminals. Example 2:

[0052] refer to Figure 1 and Figure 9 This embodiment provides a hand-push or lightweight towed modular multi-frequency ground-penetrating radar inversion system suitable for narrow areas such as ramps, bridge decks, urban roads, and tunnel entrances and exits, or scenarios requiring fine retesting. Compared with Embodiment 1, this embodiment focuses on lightweight equipment, rapid deployment, repeatable route retesting, and maintaining consistency between echo and location binding under low-speed or variable-speed conditions. Unless otherwise stated, the inversion principle, cross-frequency alignment, cross-frequency coherent quality control, equivalent relative permittivity estimation, thickness and compaction inversion, and GIS and IoT fusion output of this embodiment are consistent with those of Embodiment 1, thereby ensuring that the output caliber of the two deployment forms is consistent and the parameters are traceable in the same project.

[0053] The difference lies in the fact that, in this embodiment, the multi-frequency data acquisition and positioning module consists of a radar controller, a replaceable horn antenna, a positioning sensor, a mileage sensor, and a synchronization triggering and power supply unit. The radar controller is used for echo acquisition, timestamp marking, data storage, and peripheral interface management. For example, the GSSI SIR 30 data acquisition unit is selected as the echo acquisition and storage platform. The multi-frequency antenna is used to form at least two frequency band echo acquisitions. For example, the GSSI horn antennas Model 4108 (1GHz) and Model 4105 (2GHz) are selected. The two antennas are installed on the same trolley crossbeam or towing bar using a quick-release mechanism. The quick-release mechanism preferably has mechanical limits to make the antenna centerline coaxial with the longitudinal baseline of the trolley, reducing geometric errors caused by repeated installation. The trolley structure preferably has an antenna ground clearance adjustment mechanism and a height locking mechanism. The ground clearance remains constant during the same retest and is recorded as a working condition record field in the log.

[0054] The positioning sensor provides the absolute coordinates and time reference of the sampling point. For example, the u-blox ZED-F9P high-precision GNSS module is selected as the core of RTK positioning calculation to output centimeter-level coordinates and synchronize with the radar echo frame timestamp. The odometer provides relative displacement and sampling interval constraints. For example, a combination of a ranging wheel and an incremental rotary encoder is used. The encoder can be the OMRON E6B2-C series, specifically the E6B2-CWZ6C model, which outputs A, B, and Z phase pulses. The circumference of the ranging wheel is pre-calibrated as a system parameter and written into the configuration. Before the operation begins, the system performs a "ranging wheel idle check" to eliminate tire slippage or loose installation. To improve the attitude and heading stability under conditions of curves, slopes, or speed fluctuations, this embodiment can optionally configure the Xsens MTi-680 RTK GNSS / INS module to output attitude, heading, and fused position. This is used to write attitude disturbances such as trolley turning and pitch changes into the working condition record field and to be used for subsequent abnormal echo interpretation and retest path matching.

[0055] Regarding the synchronization of sampling location information, this embodiment adopts a "dual-channel binding" strategy: the echo frame number output by the radar controller is used as the main index, while the RTK coordinates, ranging wheel mileage, and operation timestamp of the same frame are recorded simultaneously; when the RTK signal is briefly lost, the system prioritizes using the ranging wheel mileage to maintain the consistency of the sampling point spacing, and performs drift correction through mileage-coordinate fusion after the RTK is restored, thereby ensuring the continuous sampling capability in narrow and obstructed areas.

[0056] In this embodiment, the cross-frequency alignment module is integrated into the embedded computing unit or lightweight industrial control computer of the handheld terminal. Cross-frequency alignment includes time zero-point alignment and phase reference alignment. To adapt to two acquisition methods, this embodiment provides the implementation paths respectively: Synchronous dual-antenna acquisition: 1GHz and 2GHz antennas acquire data along the same path, and echo frames share the same timestamp system; the cross-frequency alignment module uses the arrival time of the first wave of each frame as the time zero-point reference, calculates the time zero-point difference between the two frequency bands and performs compensation, so that the echoes of the two frequency bands are aligned to a unified zero point within the same sampling point. Phase reference alignment uses a phase reference within a fixed reference window for unification, and the reference window is selected in the direct wave stable region or the system coupling stable region; Acquisition by changing antennas in stages: two acquisitions are completed on the same path using 1GHz and 2GHz antennas respectively; the cross-frequency alignment module first performs position matching based on the distance measuring wheel mileage, using the same mileage position or the nearest neighbor mileage position as the matching point, and supplements it with RTK trajectory projection to ensure correct matching in the turning area; after completing the position matching, time zero-point alignment and phase reference alignment are then performed. To avoid inconsistent sampling point density caused by differences in operation speeds across different trips, this embodiment preferably sets the sampling trigger to "mileage trigger" at the acquisition end, that is, triggering an echo acquisition once every fixed distance traveled, so that the data from the two trips can be naturally aligned in the mileage dimension; the cross-frequency alignment module uses the "aligned co-position echo" as a prerequisite for subsequent cross-frequency coherence calculation and writes the alignment residual into the log field; when the alignment residual exceeds the preset upper limit, the sampling point is marked as a candidate anomaly point and enters the coherence confirmation process of the subsequent quality control unit.

[0057] In this embodiment, the manual pushing operation is characterized by speed fluctuations, slight changes in antenna height above the ground, changes in trolley attitude, and human-induced disturbances. Therefore, the preprocessing and quality control module, based on the processing flow of Embodiment 1, adds constraints and recording items adapted to "low-speed variable-speed retest": Mileage triggering is preferred for sampling to ensure constant spatial spacing between sampling points and reduce the impact of speed changes on feature statistics. If time triggering is used, speed is written to the log as a forced operating condition field, and speed segmented thresholds are used in subsequent coherence threshold determination. Trolley attitude and antenna height above the ground changes are recorded as operating condition fields in the log. When the attitude change exceeds a preset threshold or the antenna height above the ground changes... When the degree change exceeds the preset threshold, the sampling point is first marked as a candidate anomaly, but the final elimination or weighting must still be confirmed by the cross-frequency coherence index; the quality control output fields are still uniformly the elimination flag f and weight w, and are forcibly passed to the cross-frequency feature construction and fusion module and the deep regression inversion module for gating suppression and sample weighting; when the elimination flag f is 1, the weight w is fixed at 0 to ensure that the elimination logic is consistent throughout the entire link; in this embodiment, in order to facilitate the retesting of the location anomaly cause, in addition to saving the cross-frequency coherence comparison value R, the "repeated route number, retesting trip number, trolley operator number or equipment number" are also saved, so that the same section can be traced and compared between multiple retests.

[0058] In this embodiment, the processing flow of the interface parameter extraction module is the same as in Embodiment 1. Its output includes the interface two-way propagation time Δt and the equivalent relative permittivity εr, and the interface picking confidence is written to the log field. To adapt to retesting in narrow areas, this embodiment adds an "interface picking stability check": when the interface arrival time jump in the same segment exceeds a preset threshold within adjacent sampling points, the system marks the segment as a low-confidence candidate segment and highlights it in the spatial output. The final low-confidence determination is still based on f, w, and the subsequent confidence Conf output by the quality control module. The thickness calculation still uses the physical thickness model of Embodiment 1 as the verification quantity. Where c is the speed of light in vacuum, Δt is the two-way propagation time at the interface, and εr is the equivalent relative permittivity; this formula is only used for physical consistency verification and calibration parameter interpretation, and does not replace the deep regression model's ability to compensate for complex working conditions.

[0059] In this embodiment, the cross-frequency feature tensor construction method is consistent with that in Embodiment 1. The input includes multi-frequency effective echo, Δt, and εr, and the quality control output fields f and w are used as mandatory input fields for gating. The difference from Embodiment 1 is that this embodiment places greater emphasis on the rapid location and retesting strategy for "low-confidence areas": in the feature construction stage, the weight w is directly used as a gating factor to automatically reduce the feature contribution of unstable sections of the trolley operation, avoiding peak disturbances in the thickness and compaction output of this section; in the feature fusion stage, a "retesting consistency channel" is added: when there are multiple retesting data in the same section, the system can align the features of different trips according to mileage and calculate the inter-trip difference statistics as an additional quality field and write it to the log; this field does not change the basic structure of the claims, but is used in engineering to quickly locate key retesting sections with "multiple trip inconsistencies".

[0060] In this embodiment, the deep regression inversion module is deployed in a portable edge computing terminal or a vehicle-mounted small industrial control computer. It outputs the predicted thickness and compaction degree according to the sampling points and performs physical consistency verification. The thickness verification uses the predicted thickness value and the physical thickness. The deviation threshold is determined; when the deviation exceeds the threshold, the confidence level of that point is reduced at the output and a retest prompt is triggered. The compaction degree output follows a defined path: the density estimate is obtained by on-site calibration mapping from the equivalent relative permittivity εr, and then the compaction degree is determined by combining it with the reference density; the calibration mapping version number and the reference density source identifier are written to the log with each output; to adapt to on-site retesting, this embodiment provides a "parameter package switching mechanism": the model file, calibration mapping parameters, threshold parameters, and weighting classification rules are stored in the same configuration package, which includes the project number, material source batch, date, and version number; the system selects and locks the corresponding configuration package before starting the retest to avoid untraceability caused by mixing parameters from different projects.

[0061] In this embodiment, the uncertainty assessment and spatial mapping output module preferably organizes its output with retesting efficiency as the goal: low-confidence areas are directly highlighted on the GIS interface and a retesting list is generated according to priority. Simultaneously, the retesting tasks and results are transmitted back to the cloud via an IoT link, forming a closed-loop record and project-level quality archive. Specifically, this is achieved by repeatedly inferring from the same sampling point to obtain repeated prediction sequences for thickness and compaction, with the degree of dispersion characterizing the uncertainty. The confidence level (Conf) is determined by uncertainty, weight (w), coherence comparison value (R), and elimination flag (f), and a hard rule is executed: when f is 1, Conf is 0; when Conf is lower than a certain value, the confidence level is reduced to 0. The sampling points of the confidence threshold are clustered to generate low-confidence regions, and a retest list is output. The retest list includes at least the start and end mileage, region length, suggested retest method, suggested number of retests, and suggested priority. The priority is determined by the Conf size and region length. The spatially continuous distribution is output in the form of three layers: thickness, compaction degree, and Conf. The low-confidence region layer and the retest task layer are displayed overlaid. The layers and task list are uploaded to the cloud platform or project server via 4G, 5G or private network, and include the device number, work batch number, model version number, calibration mapping version number, and threshold version number. It supports real-time viewing and closed-loop scheduling on multiple terminals. Example 3:

[0062] refer to Figure 10 This embodiment uses the depth regression-based multi-frequency radar thickness compaction inversion system from Embodiment 1 to provide a depth regression-based multi-frequency radar thickness compaction inversion method. For ease of description and based on the scheme of Embodiment 1, this embodiment numbers the sampling points in mileage order as follows: The different frequency bands are numbered as The two-way propagation time of the interface is denoted as The equivalent relative permittivity is denoted as The removal mark is recorded as And take the value 0 or 1, and record the weight as And take values ​​from 0 to 1, and record the coherence comparison value as The point-level confidence level is denoted as Conf, and it is agreed that when... There will always be a time for it. This includes the following steps: S1. Collect ground-penetrating radar echo data of at least two different frequency bands or center frequencies on the road section to be detected, and simultaneously collect the location information corresponding to each echo sampling location; the ground-penetrating radar echo data of at least two different frequency bands or center frequencies are acquired using a synchronous acquisition method along the same path, or by changing antennas multiple times along the same path; further, to ensure the feasibility and traceability of subsequent cross-frequency alignment, the sampling process simultaneously binds timestamp, mileage information and coordinate information to each frame of echo; among which, the mileage information is output by the ranging wheel or wheel encoder, and the coordinate information is output by RTK positioning, and is written into the header of the original data file or the accompanying metadata file in the format of "frame number-timestamp-mileage-coordinate"; in addition, this embodiment preferably uses mileage-triggered sampling to make the distance between adjacent sampling points approximately constant, thereby reducing the impact of speed fluctuations on subsequent feature statistics and threshold determination; when the site conditions require time-triggered sampling, speed is written into the metadata as a mandatory operating condition field, and a speed segmented threshold or speed normalization strategy is used in the subsequent threshold determination to maintain the consistency of the determination caliber under different operating conditions.

[0063] S2. Cross-frequency alignment and preprocessing are performed on echoes from different frequency bands to obtain multi-frequency effective echoes. Specifically, cross-frequency alignment includes at least time zero-point cross-frequency alignment and phase reference alignment. Specifically, time zero-point cross-frequency alignment uses the arrival time of the first wave of the direct wave or coupled wave as the zero-point reference. The zero-point position is first determined in each frequency band, and then the echoes from different frequency bands are uniformly shifted to the same zero-point definition. Phase reference alignment establishes a phase reference within a preset reference window so that echoes from different frequency bands have a consistent phase reference at the same sampling point. In addition, preprocessing includes at least one of background removal, bandpass filtering, time zero-point correction, and gain compensation. In this embodiment, background removal is preferably performed sequentially. Background removal, bandpass filtering, time zero-point correction, and gain compensation: Background removal is used to suppress system coupling and slow-varying background noise; bandpass filtering is used to filter out out-of-band noise and low-frequency drift; time zero-point correction is used to ensure zero-point consistency after cross-frequency alignment; gain compensation is used to improve the visibility of deep interface reflections and maintain cross-frequency comparability. It should be noted that, in order to form "multi-frequency effective echoes", this embodiment first marks the echo frames with obvious saturation, shearing, or alignment residual exceeding the limit as candidate anomalies at the end of S2, but does not perform final rejection in the S2 stage. The final rejection or weighting is completed by the coherence comparison value judgment closed loop in S4, thereby ensuring the consistency of quality control criteria and the integrity of the evidence chain.

[0064] S3. Interlayer interface picking is performed based on multi-frequency effective echoes to obtain propagation time parameters and estimate equivalent dielectric parameters; interlayer interface picking takes multi-frequency effective echoes as input and outputs the two-way propagation time of the interlayer interface. Furthermore, the interface picking is implemented using one of the following methods: envelope peak detection, energy mutation detection, gradient extremum detection, and multi-scale correlation matching. An interface picking confidence field is output for each sampling point. When the interface picking confidence is lower than a preset threshold, the point is marked as an "untrustworthy candidate point," and a more stringent quality control strategy is adopted in S4. Additionally, this embodiment further estimates the equivalent dielectric parameters. The equivalent dielectric parameter is estimated based on the relationship between incident and reflected amplitudes or on project-level calibration relationships, and the estimation method identifier and parameter version number are recorded in the metadata to support traceability; among which, and It is used not only for subsequent thickness and compaction inversion, but also for the calculation of physical consistency verification quantity in S5 and interface time window positioning in S4.

[0065] S4. Determine the interface time window based on the inter-layer interface picking results and set a separate reference time window. Calculate the cross-frequency coherence within the interface time window and the reference time window and form a coherence comparison value. When the coherence comparison value is lower than the threshold, perform elimination or weight reduction processing on the corresponding echo data and generate elimination marks or weight information. In this embodiment, based on A defined interface time window is established, covering the main energy range of interlayer interface reflections. Simultaneously, a reference time window, separate from the interface time window, is selected from an area outside the interface time window where reflection events are weak and relatively stable, to establish baseline consistency. Further, coherence contrast values ​​are... The cross-frequency coherence index is obtained as follows: Calculate the cross-frequency coherence index between echoes in different frequency bands within the interface time window, and calculate the corresponding cross-frequency coherence index within the reference time window. Then, the ratio of the two is used as the coherence comparison value. The preferred cross-frequency coherence metric is normalized cross-correlation coherence. For cases involving more than two frequency bands, the coherence results for all frequency band pairs are aggregated to obtain the interface window coherence and reference window coherence, and then the comparison value is calculated. Additionally, when the coherence comparison value... When the data is below a threshold, the corresponding echo data must be discarded or assigned a weight below a preset upper limit to participate in the subsequent feature tensor construction; this embodiment uses a "discarding flag". +weight The combined output method sets a hard rejection threshold. With the weighting and grading rules: when Time output , And remove that point; when Output when the value is near the threshold but does not meet the hard rejection condition. , And the weight follows Increase and increase; further, weight At least to constrain the feature fusion contribution at that location, therefore the weights are used in S5. Used as a gating factor or sample weighting factor, the contribution of the deweighted point in cross-frequency fusion is suppressed, and the threshold version number and weight classification rule version number are recorded in the log to meet the requirement of traceability of threshold source.

[0066] S5. Construct a multi-frequency feature tensor containing coherence comparison information and input it into a multi-task deep regression model. Output the structural layer thickness and compaction values ​​corresponding to the sampling positions. Perform physical consistency verification or introduce physical consistency constraints on the thickness values. The compaction values ​​are limited to the density estimate obtained by on-site calibration mapping of the equivalent dielectric parameters and the reference density. Specifically, the constructed multi-frequency feature tensor explicitly includes coherence contrast information, at least including coherence contrast values. The corresponding feature channels, and the quality control output fields , As a forced adjoint field input to the model-side processing logic; further, the multi-frequency feature tensor includes at least one of time-domain energy features, frequency-domain amplitude features, cross-frequency amplitude ratio features, cross-frequency phase difference features, cross-frequency coherence features corresponding to the interface time window, and coherence contrast features; preferably, this embodiment simultaneously includes time-domain energy features, frequency-domain amplitude features, cross-frequency amplitude ratio features, cross-frequency phase difference features, cross-frequency coherence features, and coherence contrast features, thereby taking into account both physical interpretability and cross-frequency complementarity; in addition, the multi-task deep regression model applies weights Gating or weighting is performed to suppress the feature channels corresponding to the downweighted echoes during cross-frequency fusion; in this embodiment, the feature tensor of the sampling point is weighted at the model input. Perform gating, so that The elimination point does not enter the valid inference path. The weighting point contributes proportionally to the fusion; furthermore, the thickness value is subject to physical consistency verification or physical consistency constraints are introduced, and the physical consistency is used to constrain the thickness value to be consistent with the physical thickness derived from the propagation time parameter and the equivalent dielectric parameter. In this embodiment, the physical thickness is calculated as a verification quantity during the inference phase: ;in The speed of light in a vacuum. For two-way travel time, The equivalent relative permittivity; when the model output thickness is... When the deviation exceeds the threshold, the confidence level of that point is reduced and a retest prompt is triggered. During the training phase, the deviation is further used as a physical consistency constraint in the loss function to suppress thickness drift caused by working condition drift. In addition, the compaction value is limited to the density estimate obtained by on-site calibration mapping of the equivalent dielectric parameter and the reference density. In this embodiment, the reference density is set as the benchmark density or standard density corresponding to the mixture to be tested, and is written into the configuration as a project-level fixed parameter and the source identifier and version number are recorded. The density estimate is obtained from the on-site calibration mapping relationship between the equivalent dielectric parameter and the density and the mapping version number is recorded. Finally, the compaction is calculated according to the density ratio and converted into a percentage output as needed. Furthermore, the multi-task deep regression model is obtained by supervised learning training, and the training sample labels are obtained by ground truth detection. Ground truth detection includes at least one of core thickness measurement, density detection, and porosity detection. The on-site calibration mapping relationship between the equivalent dielectric parameter and the thickness and density is established or updated based on ground truth detection, and the coherence comparison value threshold is determined or updated based on ground truth detection. And the weighting rules; specifically, in the early stages of the project, a small number of calibration points are selected for core sampling and thickness and density testing, which are used as thickness and compaction label samples for supervised training or supervised fine-tuning. At the same time, this batch of calibration points is used to update The mapping function parameters to density are used, and the threshold is updated based on the distribution of coherence contrast values ​​at the calibration points. The weighted grading rules provide a traceable truth basis for the threshold and deweighting rules; In addition, for different projects, different material ratios, or different working conditions, this embodiment collects a small number of calibration samples with truth labels to perform small-sample calibration or migration update of the normalized parameters of the model parameters or feature tensors of the multi-task deep regression model, and saves the project-level calibration parameters; among them, the project-level calibration parameters are bound to the model version number, threshold version number, and mapping version number for storage to avoid cross-project mixing and causing output drift.

[0067] S6. Output the confidence level or uncertainty of the thickness and compaction values, and use the rejection marker or weight information as one of the criteria for low confidence. Combine the location information to generate spatially continuous distribution results and prompts for retesting or supplementing low-confidence areas for on-site quality closed-loop control. In this embodiment, the output terminal generates the thickness value, compaction value, and its point-level confidence level (Conf) or point-level uncertainty for each sampling point, and includes the quality control field. , As one of the criteria for determining low reliability; furthermore, this embodiment executes a hard rule: when When the confidence level (Conf) of that point is set to 0, it is directly included in the low-confidence point set; when The confidence level for a given point decreases as the weight decreases, and the confidence level is further adjusted based on the uncertainty. Furthermore, when generating spatially continuous distribution results using location information, this embodiment projects the sampling point output onto mileage coordinates or geographic coordinates, and uses the confidence level as a weighting coefficient to generate a spatially continuous distribution layer of thickness and compaction, while simultaneously generating a low-confidence region layer. Low-confidence regions are formed by clustering low-confidence points according to mileage continuity or spatial adjacency, and a retest or supplementary test suggestion is output. This suggestion includes at least the start and end mileage, suggested retest method, suggested number of retests, and priority. Furthermore, based on the spatially continuous distribution results and their confidence level or uncertainty, on-site closed-loop control output is performed. The closed-loop control output includes the location and graded alarm of thickness anomaly area, the location and graded alarm of compaction anomaly area, the prompt for retesting or supplementing low confidence area, and the quality result map or report with spatial coordinates. The priority of graded alarm is determined by the anomaly amplitude, the length of the anomaly area, and the confidence level, and is pushed to the on-site terminal or project server through GIS and IoT links to realize real-time closed-loop control and traceable recording of construction quality.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-frequency radar thickness compaction inversion system based on depth regression, characterized in that, include: The multi-frequency data acquisition and positioning module is used to acquire ground-penetrating radar echo data of at least two different frequency bands or center frequencies and simultaneously obtain sampling location information. The cross-frequency alignment module is used to align the time zero point and phase reference of echoes from different frequency bands. The preprocessing and quality control module is used to preprocess the aligned echoes and uses the cross-frequency coherence index as a necessary criterion for anomaly detection to remove or reduce the weight of abnormal echoes, and outputs removal marks or weight information; wherein, the cross-frequency coherence index is determined by the time window corresponding to the inter-layer interface. The interface parameter extraction module is used to pick up the propagation time parameters from the interlayer interface and estimate the equivalent dielectric parameters. The cross-frequency feature construction and fusion module is used to construct cross-frequency feature tensors based on multi-frequency effective echoes and propagation time parameters or equivalent dielectric parameters. The deep regression inversion module is used to input the cross-frequency feature tensor into the multi-task deep regression model to output the thickness prediction value and the compaction degree prediction value. The thickness prediction value is subject to consistency constraints or consistency checks with the physical thickness obtained from the propagation time parameter and the equivalent dielectric parameter, and the compaction degree prediction value is limited to the density estimate obtained by the equivalent dielectric parameter through field calibration mapping and the reference density. The uncertainty assessment and spatial mapping output module is used to generate the spatial continuous distribution of thickness, compaction degree and their confidence or uncertainty, and output low confidence area markers and retest or supplementary test prompts based on the confidence or uncertainty for use in on-site quality closed-loop control.

2. The multi-frequency radar thickness compaction inversion system based on depth regression according to claim 1, characterized in that, The multi-frequency data acquisition and positioning module adopts at least one of the following methods to achieve multi-frequency acquisition: step frequency, linear frequency modulation, or multi-channel multi-frequency combination. At least two different frequency bands or center frequencies meet the preset frequency band difference configuration to simultaneously take into account the penetration depth and resolution. The cross-frequency alignment module controls the time zero-point alignment error of echoes from different frequency bands within a preset upper limit, and uses the aligned co-position echoes as a prerequisite for calculating the cross-frequency coherence index. The preprocessing and quality control module includes background removal, bandpass filtering, time zero-point correction, and gain compensation.

3. The multi-frequency radar thickness compaction inversion system based on depth regression according to claim 2, characterized in that, The preprocessing and quality control module includes an abnormal echo identification unit. The abnormal echo identification unit identifies anomalies based on at least one of the following criteria: abnormal antenna coupling, echo saturation, sudden noise energy change, and excessive time zero drift. It must also identify anomalies based on the cross-frequency coherence criterion based on the time window. The time window includes at least an interface time window determined by the interface picking result of the interface parameter extraction module, and a reference time window set separately from the interface time window; the cross-frequency coherence criterion is to calculate the cross-frequency coherence index and form a coherence comparison value in the interface time window and the reference time window respectively; when the coherence comparison value is lower than the threshold, the corresponding echo must be removed or given a weight lower than the preset upper limit to participate in the construction of the cross-frequency feature tensor, and the removal mark or weight information is output to the cross-frequency feature construction and fusion module and the deep regression inversion module, so that low coherence echoes cannot participate in the inversion in an unweighted state.

4. The multi-frequency radar thickness compaction inversion system based on depth regression according to claim 3, characterized in that, The deep regression inversion module includes a cross-frequency attention fusion layer and a multi-scale convolutional feature extraction layer, which are used for weighted fusion of information from different frequency bands; The cross-frequency attention fusion layer and the multi-scale convolutional feature extraction layer use coherence contrast value as a gate factor and weight information as a weighting factor to suppress the feature channels corresponding to low coherence echoes. The cross-frequency feature tensor includes at least time-domain energy features, frequency-domain amplitude features, cross-frequency amplitude ratio features, cross-frequency phase difference features, and cross-frequency coherence features and coherence comparison features corresponding to the interface time window; and a joint loss function is used to simultaneously constrain the thickness prediction error and the compaction degree prediction error, and includes a physical consistency term to constrain the consistency between the thickness prediction value and the physical thickness.

5. The multi-frequency radar thickness compaction inversion system based on depth regression according to claim 4, characterized in that, The interface parameter extraction module further includes a field calibration submodule. The field calibration submodule obtains the true value of the core thickness or the true value of the density based on a preset number of calibration points, and establishes or updates the field calibration mapping relationship between the equivalent dielectric parameters and the thickness and density. Based on calibration point data, the coherence comparison value threshold and weighting rules are determined or updated, so that the abnormal echo rejection or weighting processing under different projects, materials and working conditions has a traceable threshold source.

6. The multi-frequency radar thickness compaction inversion system based on depth regression according to claim 5, characterized in that, The uncertainty assessment and spatial mapping output module obtains the confidence or uncertainty of the thickness prediction and compaction degree prediction values ​​using at least one of model integration, Monte Carlo random deactivation, or predictive distribution estimation. It uses the cross-frequency coherence index, coherence comparison value, and elimination mark or weight information as mandatory inputs for confidence generation or threshold determination. When the cross-frequency coherence index is below the threshold, the coherence comparison value is below the threshold, or the elimination mark is true, it directly outputs the low confidence area mark and the retest or supplementary test prompt. The retest or supplementary test prompt and the low confidence area mark are superimposed on the spatial continuous distribution results.

7. A multi-frequency radar thickness compaction inversion method based on depth regression, comprising the multi-frequency radar thickness compaction inversion system based on depth regression as described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Collect at least two different frequency bands or center frequencies of ground-penetrating radar echo data in the road section to be detected, and simultaneously collect the location information corresponding to each echo sampling location. S2. Cross-frequency alignment and preprocessing are performed on echoes from different frequency bands to obtain multi-frequency effective echoes; S3. Based on the multi-frequency effective echo, interlayer interface picking is performed to obtain propagation time parameters and estimate equivalent dielectric parameters; S4. Determine the interface time window based on the inter-layer interface picking results and set a separate reference time window. Calculate the cross-frequency coherence within the interface time window and the reference time window and form a coherence comparison value. When the coherence comparison value is lower than the threshold, perform elimination or weight reduction processing on the corresponding echo data and generate elimination marks or weight information. S5. Construct a multi-frequency feature tensor containing coherence comparison information and input it into a multi-task deep regression model. Output the structural layer thickness and compaction values ​​corresponding to the sampling positions. Perform physical consistency verification or introduce physical consistency constraints on the thickness values. The compaction values ​​are limited to the density estimate obtained by on-site calibration mapping of the equivalent dielectric parameters and the reference density. S6. Output the confidence level or uncertainty of the thickness and compaction values, and use the removal mark or weight information as one of the criteria for low confidence. Combine the location information to generate spatially continuous distribution results and prompts for retesting or supplementing low confidence areas.

8. The multi-frequency radar thickness compaction inversion method based on depth regression according to claim 7, characterized in that, The cross-frequency alignment in S2 includes at least time zero-point cross-frequency alignment and phase reference alignment; The preprocessing in S2 includes at least one of background removal, bandpass filtering, time zero-point correction, and gain compensation. The coherence comparison value in S4 is obtained in the following way: the cross-frequency coherence index between echoes in the same position of different frequency bands is calculated within the interface time window, and the corresponding cross-frequency coherence index is calculated within the reference time window. The ratio or difference between the two is then used as the coherence comparison value. When the coherence comparison value is lower than the threshold, the corresponding echo data must be removed or assigned a weight lower than the preset upper limit to participate in the subsequent feature tensor construction. The weight information in S4 is used at least to constrain the feature fusion contribution at that position.

9. The multi-frequency radar thickness compaction inversion method based on depth regression according to claim 8, characterized in that, The multi-frequency feature tensor in S5 includes at least one of the following: time-domain energy feature, frequency-domain amplitude feature, cross-frequency amplitude ratio feature, cross-frequency phase difference feature, and cross-frequency coherence feature and coherence contrast feature corresponding to the interface time window; The multi-task deep regression model in S5 performs gating or weighting on the weight information, so that the feature channels corresponding to the downweighted echoes are suppressed in cross-frequency fusion. The physical consistency check or constraint in S5 is used to ensure that the thickness value is consistent with the physical thickness derived from the propagation time parameter and the equivalent dielectric parameter. The reference density in S5 is the baseline density or standard density corresponding to the mixture to be tested. The density estimate is obtained through the on-site calibration mapping relationship between the equivalent dielectric parameter and the density.

10. The multi-frequency radar thickness compaction inversion method based on depth regression according to claim 9, characterized in that, The multi-task deep regression model in S5 is obtained through supervised learning training. The training sample labels are obtained by ground truth detection, which includes one of core thickness measurement, density detection, and porosity detection. Based on the ground truth detection, the model establishes or updates the field calibration mapping relationship between the equivalent dielectric parameter and the thickness and density, and determines or updates the coherence comparison value threshold and weighting rules based on the ground truth detection. For different projects, material ratios, or working conditions, a small number of calibration samples with truth labels are collected to perform small-sample calibration or migration updates on the normalized parameters of the model parameters or feature tensors of the multi-task deep regression model, and the project-level calibration parameters are saved. Based on the spatially continuous distribution results and their confidence or uncertainty, on-site closed-loop control output is performed. The closed-loop control output includes the location and graded alarm of thickness anomaly area, the location and graded alarm of compaction anomaly area, the prompt for retesting or supplementing low confidence area, and the quality result map or report with spatial coordinates.

Images