A multi-dimensional ground penetrating radar data intelligent fusion processing system

Abstract

The application discloses a kind of multidimensional geological radar data intelligent fusion processing systems, the present application relates to geological radar detection technical field, including: dual-mode decoupling module, adaptive alignment module, feature fusion module and double verification module;The multidimensional geological radar data of input is separated into spatial topology mode and spectral feature mode by the dual-mode decoupling module;The adaptive alignment module includes: lightweight space-time encoder, generates spatial feature vector by processing spatial topology mode;Frequency domain sparse attention network, processing spectral feature mode and embedding differentiable geometric constraint layer.The multidimensional geological radar data intelligent fusion processing system, through the coordination of dynamic physical constraint layer and thermodynamic diagram driven hierarchical scheduling mechanism, system realizes the cross-dimension alignment of geological radar multidimensional data under the strict real-time requirement, the forced anti-interference alignment strategy of metal interference area, effectively solve the position deviation problem caused by phase jump.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar detection technology, specifically to a multi-dimensional ground-penetrating radar data intelligent fusion processing system. Background Technology

[0002] Ground-penetrating radar (GPR) technology is widely used in underground exploration. The data it collects naturally possesses multi-dimensional characteristics, including different scanning modes such as B-scan and C-scan, different polarization directions, different frequencies, and different time series. This multi-dimensional data contains rich and complementary subsurface information. To more comprehensively and accurately analyze subsurface structures, intelligent fusion processing of GPR data from different dimensions is necessary.

[0003] In existing technologies, multi-dimensional ground-penetrating radar (GPR) data fusion systems built on general-purpose electronic digital data processing frameworks face significant challenges when processing such high-dimensional heterogeneous data streams. This is particularly pronounced in applications requiring real-time or near-real-time processing, such as tunnel construction forecasting or rapid road defect detection. Insufficient accuracy in cross-dimensional feature alignment significantly amplifies errors in subsequent fusion processing stages. Inaccurately corresponding feature points input into the fusion model can lead to the model learning incorrect association rules, resulting in fusion outputs such as 3D reconstructed images or target recognition results that may exhibit target blurring, positioning deviations, or the introduction of false anomalies, severely weakening the credibility and practical application value of the fusion results. Therefore, effectively improving the accuracy of cross-dimensional feature alignment in multi-dimensional GPR data under strict real-time constraints has become a key bottleneck restricting the performance improvement of intelligent fusion systems. Thus, the urgent technical problem to be solved is: how to significantly improve the accuracy of cross-dimensional feature alignment in multi-dimensional GPR data while meeting real-time processing requirements to ensure the credibility of the fusion results. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a multi-dimensional ground-penetrating radar data intelligent fusion processing system, comprising: a dual-modal decoupling module, an adaptive alignment module, a feature fusion module, and a dual verification module;

[0005] The dual-mode decoupling module separates the input multi-dimensional geological radar data into spatial topological modes and spectral feature modes;

[0006] The adaptive alignment module includes:

[0007] A lightweight spatiotemporal encoder that processes spatial topological modes to generate spatial feature vectors;

[0008] A frequency-domain sparse attention network processes spectral feature modes and embeds a differentiable geometric constraint layer;

[0009] The dynamic scheduling unit allocates computing resources based on the feature saliency heatmap: high-precision map matching and alignment are performed on highly sensitive regions, and low-rank approximate alignment is performed on non-sensitive regions;

[0010] The feature fusion module receives the aligned dual-modal features and outputs a preliminary fusion result;

[0011] The dual verification module performs physical consistency screening and residual adversarial verification on the fusion results.

[0012] Preferably, the implementation of the differentiable geometric constraint layer is as follows:

[0013] Wave propagation direction constraint of Snell's law can be implemented using differentiable functions;

[0014] Energy constraints for the medium decay equation are achieved through differentiable regularization terms.

[0015] The above constraints serve as the physical driving optimization objectives for frequency-domain sparse attention networks.

[0016] Preferably, the method for generating the feature saliency heatmap includes:

[0017] Extracting gradient sensitivity of cross-dimensional features in frequency domain sparse attention networks;

[0018] Identify regions of abrupt geometric structural changes in spatial feature vectors;

[0019] Heatmaps are generated by fusing gradient sensitivity with geometric abrupt change regions, and regions with sensitivity exceeding a preset threshold are marked as high-sensitivity areas.

[0020] Preferably, the specific steps for high-precision image matching and alignment are as follows:

[0021] Construct a graph structure node for cross-dimensional feature points;

[0022] Achieve sub-pixel-level alignment in spatial dimensions based on the optimal transmission path;

[0023] Time dimension alignment is achieved based on spectral phase coherence.

[0024] Preferably, the physical consistency screening operation includes:

[0025] The alignment features are input into a cross-domain physics simulator to reconstruct the electromagnetic wave propagation path;

[0026] When the deviation between the simulated echo and the actual data in terms of reflection energy attenuation rate, time difference between multiple reflections, or polarization rotation angle exceeds a threshold, the alignment module is triggered for iterative optimization.

[0027] Preferably, the residual adversarial verification includes:

[0028] The first channel uses a convolutional neural network to detect boundary misalignment artifacts in the spatial dimension;

[0029] The second channel uses a time-frequency analysis network to detect phase jumps in the spectral dimension;

[0030] When any channel detects that the residual confidence level exceeds the threshold, it feeds back a reconstruction signal to the feature fusion module.

[0031] Preferably, the electromagnetic wave propagation model of the cross-domain physics simulator includes:

[0032] Wave field solver based on Maxwell's equations;

[0033] An adaptive mapping mechanism between dielectric parameters and conductivity.

[0034] Preferred options also include:

[0035] The data preprocessing module performs noise reduction, gain correction, and dimension standardization operations on the raw ground-penetrating radar data;

[0036] The results output module converts the validated fused data into a three-dimensional geological model.

[0037] A method for intelligent fusion processing of multi-dimensional ground-penetrating radar data, employing a multi-dimensional ground-penetrating radar data intelligent fusion processing system, includes the following steps:

[0038] S1: The data is separated into spatial topological modes and spectral characteristic modes through a dual-modal decoupling module;

[0039] S2: Perform alignment operations for dynamic resource allocation through the adaptive alignment module;

[0040] S3: Generate preliminary fusion results through the feature fusion module;

[0041] S4: The results are subjected to dual verification using the dual verification module, which checks both the physical rules and the data residuals.

[0042] Preferably, the alignment operation for dynamic resource allocation includes:

[0043] Based on the heatmap, highly sensitive regions are assigned to FPGA computing units for graph matching.

[0044] Non-sensitive regions are allocated to GPU computing units to perform low-rank approximation calculations.

[0045] This invention provides an intelligent fusion processing system for multi-dimensional geological radar data. It has the following beneficial effects:

[0046] This intelligent fusion processing system for multi-dimensional ground-penetrating radar data achieves cross-dimensional alignment of multi-dimensional ground-penetrating radar data under strict real-time requirements through the synergy of a dynamic physical constraint layer and a heat map-driven hierarchical scheduling mechanism. It also features a forced anti-interference alignment strategy for metal interference zones, effectively solving the position deviation problem caused by phase jumps.

[0047] This multi-dimensional ground-penetrating radar data intelligent fusion processing system relies on a dual closed-loop mechanism of physical screening and residual verification to improve the engineering credibility of the system's output fusion results. The verification results-driven real-time parameter feedback and resource conflict arbitration mechanism enable the system to have self-optimization capabilities in complex geological environments, and the output three-dimensional geological model can directly support construction decisions. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the module interaction of a multi-dimensional geological radar data intelligent fusion processing system according to the present invention;

[0049] Figure 2 This is a flowchart illustrating the intelligent fusion processing method for multi-dimensional geological radar data according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 and Figure 2 The present invention provides a technical solution: a multi-dimensional ground-penetrating radar data intelligent fusion processing system, comprising: a dual-modal decoupling module, an adaptive alignment module, a feature fusion module, and a dual verification module;

[0052] The dual-mode decoupling module separates the input multi-dimensional ground-penetrating radar data into spatial topological modes and spectral feature modes;

[0053] The adaptive alignment module includes:

[0054] A lightweight spatiotemporal encoder that processes spatial topological modes to generate spatial feature vectors;

[0055] A frequency-domain sparse attention network processes spectral feature modes and embeds a differentiable geometric constraint layer;

[0056] The dynamic scheduling unit allocates computing resources based on the feature saliency heatmap: high-precision map matching and alignment are performed on highly sensitive regions, and low-rank approximate alignment is performed on non-sensitive regions;

[0057] The feature fusion module receives the aligned dual-modal features and outputs the preliminary fusion result;

[0058] The dual verification module performs physical consistency screening and residual adversarial verification on the fusion results.

[0059] It should be further explained that, in the specific implementation process, the input multi-dimensional ground-penetrating radar data, after denoising and standardization preprocessing, is first decoupled into spatial topological modes and spectral feature modes. The spatial topological modes include the scanning trajectory coordinates and antenna spacing geometry, while the spectral feature modes include the frequency spectrum, phase difference, and polarization parameters. The spatial topological modes extract key geometric features through a lightweight spatiotemporal encoder to generate spatial feature vectors. The spectral feature modes are input into a frequency-domain sparse attention network, which embeds a differentiable geometric constraint layer. This layer transforms the wave propagation angle constraint of Snell's law into a differentiable loss function of the direction vector and converts the energy constraint of the medium attenuation equation into a regularization term, forcing the network to output spectral feature vectors that conform to physical laws.

[0060] A feature saliency heatmap is generated based on two types of feature vectors: Cross-dimensional data-sensitive regions are calculated through gradient backpropagation of a spectral network, and combined with the detection results of curvature abrupt change points in the spatial vectors, highly sensitive regions that significantly affect alignment errors are marked; these highly sensitive regions include underground cavity boundaries and pipeline intersections. The dynamic scheduling unit, based on the heatmap distribution, calls the FPGA computing unit to perform high-precision map matching and alignment for highly sensitive regions, constructing a feature point map structure, calculating sub-pixel-level displacement compensation through the optimal transmission path, and calibrating the time dimension using phase coherence; for non-sensitive regions, the GPU computing unit performs low-rank approximate alignment, employing truncated SVD dimensionality reduction for rapid transformation.

[0061] The aligned bimodal features are processed by a fusion network to generate preliminary 3D fusion results, which then enter a dual verification stage, including the following two phases:

[0062] Physical consistency screening phase: The fused features are input into the cross-domain physical simulator to reconstruct the propagation of electromagnetic waves in the corresponding medium model. If the core indicators of the simulated echo and the actual data deviate from the preset threshold, an iterative command is sent to the alignment module. The cross-domain physical simulator is built based on Maxwell's equations, and the core indicators include the reflection energy attenuation rate and the time difference between multiple reflections.

[0063] Residual adversarial verification phase: The dual-channel discriminators operate synchronously. The spatial residual channel detects boundary misalignment artifacts in the 3D model, while the spectral residual channel identifies phase jump anomalies. When the confidence level of the output residual of either channel exceeds the limit, the reconstruction training of the fusion network is triggered.

[0064] The final output of the fused data that passes the verification is a 3D geological structure model and a target identification report, while the data that fails the verification is returned to the alignment stage for re-optimization.

[0065] The implementation method of the differentiable geometric constraint layer is as follows:

[0066] Wave propagation direction constraint of Snell's law can be implemented using differentiable functions;

[0067] Energy constraints for the medium decay equation are achieved through differentiable regularization terms.

[0068] The above constraints serve as the physical driving optimization objectives for frequency-domain sparse attention networks.

[0069] It should be further explained that, in the specific implementation process, when the frequency domain sparse attention network processes spectral feature modes, the embedded differentiable geometric constraint layer dynamically implements physical rule constraints in the following ways:

[0070] For the electromagnetic wave propagation direction constraint, the relationship between the incident angle and the refraction angle of Snell's law is transformed into a differentiable function of the direction vector: when the deviation between the wave propagation direction vector output by the network and the theoretical direction calculated based on the dielectric constant of the medium increases, gradient loss is generated and backpropagated, forcing the network to adjust the feature extraction weights to conform to physical laws; if there is strong noise interference in the actual data, such as reflection clutter from the steel arch of the tunnel, the constraint automatically reduces the weights to avoid overfitting the noise.

[0071] For energy attenuation constraints, a differentiable regularization term is constructed based on the medium attenuation equation: the theoretical attenuation curve is dynamically calculated according to the energy attenuation slope of different frequency components in the spectral characteristics; when the difference between the spectral energy distribution output by the network and the theoretical curve exceeds the adaptive threshold, the regularization intensity is enhanced to correct the feature vector; in homogeneous medium regions, such as soil layers, this constraint adopts a relaxed threshold to improve processing efficiency.

[0072] The implementation of the constraint layer includes an adaptive judgment mechanism for operating conditions, which includes the following:

[0073] For high-noise operating conditions: when the signal-to-noise ratio of the input data is lower than the set threshold, the constraint strength of the attenuation equation is automatically reduced to prioritize the integrity of feature extraction;

[0074] For abrupt changes in the medium: when the spatial topological mode detects a region of drastic change in dielectric constant, such as the interface between soil and rock, the wave propagation direction constraint is immediately strengthened, and the network is forced to output refraction characteristics that conform to Snell's law in this region;

[0075] For low-frequency dominant operating conditions: when the detection depth exceeds the effective range of the high-frequency signal, switch to low-frequency attenuation dominant mode and relax the high-frequency energy matching requirements.

[0076] The output of the constraint layer is used as a spectral feature vector and enters the subsequent modules. Its physical compliance is verified in a closed loop by the physical consistency screening unit in the dual verification module: if the energy attenuation mode of the simulated echo deviates significantly from the spectral features output by the constraint layer, such as the attenuation rate exceeding the standard for three consecutive frames, a parameter reset command is sent to the constraint layer.

[0077] Methods for generating saliency heatmaps include:

[0078] Extracting gradient sensitivity of cross-dimensional features in frequency domain sparse attention networks;

[0079] Identify regions of abrupt geometric structural changes in spatial feature vectors;

[0080] Heatmaps are generated by fusing gradient sensitivity with geometric abrupt change regions, and regions with sensitivity exceeding a preset threshold are marked as high-sensitivity areas.

[0081] It should be further explained that, in the specific implementation process, the generation of the feature saliency heatmap is achieved through the collaborative implementation of gradient sensitivity analysis and geometric mutation detection via a dual-channel approach, as follows:

[0082] Gradient sensitivity analysis: During the forward propagation of the frequency domain sparse attention network, the gradient magnitude of the cross-dimensional feature map is calculated in real time; when the gradient value of a specific region is continuously higher than three times the standard deviation of the background noise level, it is marked as a potential sensitive region; if there is a sudden change in spectral energy in the region, such as the energy difference between adjacent frequency points exceeding 30% of the mean, the sensitivity level is increased to a high level.

[0083] Geometric abrupt change detection: The curvature field is calculated on the spatial feature vector to identify geometric discontinuities where the second derivative of curvature exceeds a threshold, such as the edge of underground pipelines or the fracture surface of rock strata; when discontinuities cluster together to form a closed boundary, they are automatically expanded into a sensitive area.

[0084] The dual-channel results are fused according to logical decision rules. The fusion process includes the following three operating conditions:

[0085] Highly sensitive operating conditions: When the gradient high-level sensitive area completely overlaps with the geometric boundary, such as the pipeline intersection, it is directly marked as the highest sensitivity level, i.e., the red area, triggering subsequent high-precision alignment;

[0086] Weakly sensitive conditions: When only gradient sensitivity or isolated geometrical abrupt changes occur, such as weak anomalies in a homogeneous medium, they are marked as secondary sensitivity levels, i.e., the yellow area, and a simplified alignment strategy is adopted.

[0087] Conflict condition: When the low gradient region overlaps with the region of strong geometric change, such as a false boundary caused by metal interference, spectral coherence verification is initiated. If the phase consistency is lower than the threshold, it is judged as a false sensitive region, that is, it is not marked.

[0088] The heatmap update implements a dynamic threshold adjustment mechanism, including:

[0089] In high-noise environments, such as those with excessive electromagnetic interference, the geometric change detection threshold is automatically increased to avoid false labeling by noise artifacts. In deep-penetration modes with a depth greater than 10 meters, the gradient sensitivity weight is reduced to prioritize the integrity of the geometric structure. The generated heatmap is input into the dynamic scheduling unit in real time to guide the allocation of computational resources. If the same region is repeatedly marked as a sensitive area in five consecutive frames of data but fails subsequent verification, the parameter self-correction of the sensitivity analysis module is triggered.

[0090] The specific steps for high-precision image matching and alignment are as follows:

[0091] Construct a graph structure node for cross-dimensional feature points;

[0092] Achieve sub-pixel-level alignment in spatial dimensions based on the optimal transmission path;

[0093] Time dimension alignment is achieved based on spectral phase coherence.

[0094] It should be further explained that, in the specific implementation process, when the dynamic scheduling unit allocates highly sensitive areas to the high-precision map matching and alignment process, the following physical rule-driven operations are executed, including:

[0095] Cross-dimensional graph structure construction: Geometric key points in spatial feature vectors are used as reference nodes; corresponding points in spectral feature vectors are associated as cross-dimensional edges; where, corresponding points: satisfy the phase coherence threshold; when the number of nodes of the same ground feature is mismatched in B-scan and C-scan data, such as the edge of karst cavity, the node fusion mechanism is activated: retaining strongly coherent nodes and removing isolated points with a signal-to-noise ratio lower than the background level.

[0096] Sub-pixel alignment in spatial dimension: Calculate the optimal transmission path based on the electromagnetic wave propagation speed model: use straight path constraints in homogeneous medium regions; automatically switch to the refraction path model in regions with abrupt changes in dielectric constant, such as concrete-soil interfaces; realize sub-pixel level offset compensation by inverting spatial displacement through path length difference; if the distance between nodes still exceeds 1 / 4 of the wavelength after compensation, iterative optimization is triggered until convergence.

[0097] Phase calibration in the time dimension: coherence verification of the spectral phase difference across dimensions: when the phase difference of signals from the same source does not exceed π / 2, the timestamp is directly aligned; when the phase difference is greater than π / 2 but meets the linear gradient law, such as reflection from a tilted interface, phase unwrapping correction is used; when encountering phase jumps caused by metal interference, such as reflection from a steel mesh, the amplitude peak is matched first rather than the phase zero; the calibration results are propagated through neighboring nodes.

[0098] The alignment results are fed back to the heatmap generation module in real time: if a region fails to converge after three consecutive alignment iterations, the sensitivity level of that region is downgraded and switched to a low-rank approximate alignment process.

[0099] The physical consistency screening process includes:

[0100] The alignment features are input into a cross-domain physics simulator to reconstruct the electromagnetic wave propagation path;

[0101] When the deviation between the simulated echo and the actual data in terms of reflection energy attenuation rate, time difference between multiple reflections, or polarization rotation angle exceeds a threshold, the alignment module is triggered for iterative optimization.

[0102] It should be further explained that, in the specific implementation process, the physical consistency screening unit receives the aligned dual-modal features and reconstructs the electromagnetic wave propagation process through a cross-domain physical simulator. This process implements a hierarchical threshold verification mechanism:

[0103] Basic index verification: Calculate the energy attenuation rate deviation between the simulated echo and the actual data at the target reflection point. When the deviation exceeds the allowable benchmark threshold for this medium type, mark a primary anomaly. If the simulated value and the measured value of the arrival time difference of the multipath reflected wave of the same target deviate from each other by more than half the wavelength, mark a timing anomaly. The benchmark threshold includes soil ≤8% and rock ≤5%.

[0104] Enhanced verification in key areas: For the most sensitive areas marked on the heat map, a polarization rotation angle consistency check is added: when the difference between the simulated and measured polarization angles exceeds the angle tolerance, it is directly judged as a physical mismatch; in safety-sensitive areas such as tunnel arches and bridge pile foundations, three indicators are used for parallel verification, namely: energy + time series + polarization. An alarm is triggered if any indicator exceeds the standard. The most sensitive areas include underground pipeline intersections, and the angle tolerance includes shallow layers ≤5° and deep layers ≤8°.

[0105] Among them, the screening results trigger intelligent iterative decision-making, including:

[0106] Single-frame minor anomaly: Energy attenuation deviation is within 1-1.5 times the baseline threshold; only log is recorded, no iteration is triggered.

[0107] Three consecutive frames of moderate anomaly: the deviation continuously exceeds the baseline threshold by 1.5 times but does not reach 2 times. After reducing the physical constraint weight of the alignment module, local reprocessing is performed.

[0108] If a single frame is severely abnormal or a region fails joint inspection: the deviation exceeds twice the baseline threshold or any of the three indicators of the key region fails, the current frame data is frozen and cross-module joint optimization is initiated, including: sending a geometric constraint strengthening instruction to the alignment module; requiring the heatmap module to recalibrate the sensitive area; and performing full-precision map matching on the faulty area.

[0109] If the joint optimization still fails to pass the verification, the data will be marked as "physically unexplainable" and a fault report will be generated, prompting manual intervention for verification.

[0110] Residual adversarial verification includes:

[0111] The first channel uses a convolutional neural network to detect boundary misalignment artifacts in the spatial dimension;

[0112] The second channel uses a time-frequency analysis network to detect phase jumps in the spectral dimension;

[0113] When any channel detects that the residual confidence level exceeds the threshold, it feeds back a reconstruction signal to the feature fusion module.

[0114] It should be further explained that, in the specific implementation process, the residual adversarial verification unit performs a dual-channel parallel detection process on the fusion result, including a spatial residual channel and a spectral residual channel, wherein:

[0115] Spatial residual channel: A convolutional neural network is used to scan the 3D fusion model to identify the characteristic patterns of boundary misalignment artifacts, such as jagged edges and local mesh distortion. When a suspected artifact is detected, regional confidence assessment is initiated: if the artifact region overlaps with the high-sensitivity area of ​​the heat map and the curvature change is continuous, it is determined to be a real anomaly, i.e., high confidence level; if it exists in isolation in the non-sensitive area, spectral cross-validation is initiated.

[0116] Spectral residual channel: Phase continuity features are extracted using a time-frequency analysis network to detect transition points; the causes of these transition points are classified, including:

[0117] Metallic interference type: Accompanied by a sharp increase in amplitude: exceeding the background by 3 times and with a widened frequency band, marked as physically real;

[0118] Alignment mismatch type: No amplitude anomaly and stable frequency band, marked as processing error;

[0119] Noise-induced type: randomly scattered occurrences with a signal-to-noise ratio below the threshold, directly filtered.

[0120] The dual-channel results are processed according to the confidence level linkage rule, including:

[0121] High positional signal alarm, that is: when the confidence of any channel is greater than 0.8: when the spatial channel reports a real anomaly, the output is frozen and the physical screening unit is triggered for re-examination; when the spectrum channel reports alignment mismatch, a reconstruction signal is immediately sent to the fusion module.

[0122] Low positional confidence warning, i.e., when the dual-channel confidence level is between 0.5 and 0.8: initiate cross-channel arbitration: if the spatial artifact region and the spectral jump point are spatially less than λ / 2, they are jointly marked as potential anomalies; otherwise, they are downgraded to acceptable errors.

[0123] Continuous unresolved anomalies, i.e., three consecutive low-level warnings: automatically relax the detection threshold for this area by 20% to avoid excessive interference with the real-time process.

[0124] The validation results drive the fine-tuning of the fusion network: for metal interference-type spectral jumps, the original phase features are preserved; for alignment mismatch types, local feature refusion is triggered.

[0125] The electromagnetic wave propagation model of the cross-domain physics simulator includes: a wave field solver based on Maxwell's equations; and an adaptive mapping mechanism between dielectric parameters and conductivity. It should be further noted that, in practical implementation, when the cross-domain physics simulator performs electromagnetic wave propagation reconstruction, it dynamically constructs the computational model through the adaptive mapping mechanism for dielectric parameters. The construction process is as follows:

[0126] The system receives geometric structure information from spatial feature vectors and dielectric properties from spectral feature vectors to generate an initial dielectric mesh. The geometric structure information includes layered interfaces and void locations, while the dielectric properties include attenuation slope and dispersion relationship. When the actual detection depth exceeds a preset threshold, the system automatically strengthens the weight of low-frequency dielectric parameters and suppresses high-frequency parameter noise interference. For highly sensitive areas identified by thermal maps, a sub-mesh densification mapping is implemented: the standard mesh size is reduced to a quarter wavelength, and a dielectric constant gradient smoothing technique is used in the densified area to avoid simulation artifacts caused by parameter abrupt changes.

[0127] The wavefield solution process employs a multi-scale intelligent switching strategy, including the following:

[0128] The shallow, fast solution mode is as follows: for depths not exceeding 3 meters, the paraxial approximation wave equation is used, ignoring multiple reflection paths; when the amplitude of the spectrum of a metal target suddenly increases to more than 3 times the background, the solution mode is locally switched to the full wave field.

[0129] Deep high-precision mode, i.e.: depth greater than 3 meters: enable time-domain full wave field solver, automatically truncate the calculation area to 1.5 times the outer extension of the target body; if the signal-to-noise ratio of the deep target is lower than the threshold, activate the waveguide mode enhancement algorithm: prioritize the preservation of guided wave components propagating along the stratum interface;

[0130] For safety-sensitive areas, such as tunnel arches, a full-time-domain solution is required, and two additional reflection path calculations are added.

[0131] The boundary condition dynamic compensation mechanism handles the metal interference effect: when the spectral residual channel identifies a metal interference-type jump, an equivalent dipole boundary is automatically added at the corresponding spatial location: the dipole moment intensity is proportional to the measured echo amplitude; the direction is calculated by inversion of the polarization angle deviation; after compensation, the simulation is re-executed, and if the energy attenuation rate deviation drops to within the threshold, the compensation parameters are retained; otherwise, it is marked as "unmodelable interference".

[0132] Also includes:

[0133] The data preprocessing module performs noise reduction, gain correction, and dimension standardization operations on the raw ground-penetrating radar data;

[0134] The results output module converts the validated fused data into a three-dimensional geological model.

[0135] It should be further explained that, in the specific implementation process, the data preprocessing module implements an adaptive processing chain based on operating conditions, including the following:

[0136] Dynamic denoising strategy: In environments with strong electromagnetic interference, such as around substations, directional beamforming filtering is enabled to retain effective signals perpendicular to the detection surface; when mechanical vibration noise is detected, i.e., time-domain periodic jitter, it automatically switches to adaptive notch filtering to eliminate interference at specific frequencies; if the input signal-to-noise ratio is higher than the threshold, only basic median filtering is performed to ensure real-time performance.

[0137] Gain correction mechanism: For shallow detection, i.e. detection within 2 meters, linear gain compensation is used to prevent near-field saturation; for deep detection, i.e. detection within 5 meters, exponential gain mode is switched to enhance weak signals; when encountering abnormally high amplitude areas, local gain suppression is activated to avoid distortion.

[0138] Dimensional standardization: Resample B / C-scan data from different scanning modes to a reference grid; perform time axis registration on time-varying data to eliminate equipment movement errors.

[0139] The output module implements verification-driven 3D reconstruction: it only accepts fused data that has passed the dual verification module; it adds confidence labels to the model for "low-risk anomaly areas" marked by physical consistency screening; it automatically adds electromagnetic reflection characteristic labels when residual adversarial verification identifies metallic targets; and it outputs a 3D visualization model that includes geological strata, spatial distribution of anomalies, and confidence labels.

[0140] A method for intelligent fusion processing of multi-dimensional ground-penetrating radar data, employing a multi-dimensional ground-penetrating radar data intelligent fusion processing system, includes the following steps:

[0141] S1: The data is separated into spatial topological modes and spectral characteristic modes through a dual-modal decoupling module;

[0142] S2: Perform alignment operations for dynamic resource allocation through the adaptive alignment module;

[0143] S3: Generate preliminary fusion results through the feature fusion module;

[0144] S4: The results are subjected to dual verification using the dual verification module, which checks both the physical rules and the data residuals.

[0145] It should be further explained that, in the specific implementation process, when performing multi-dimensional ground-penetrating radar data processing, intelligent fusion is achieved through the following steps:

[0146] Step a: Dual-modal decoupling and separation: Receive preprocessed multi-source data and implement decoupling decisions based on scanning mode characteristics: B-scan data prioritizes the extraction of spatial topological modes; C-scan data focuses on separating spectral feature modes; when inputting time-varying sequence data, time dimension labels are automatically added for the alignment module to use.

[0147] Step b: Alignment operation for dynamic resource allocation: After generating a feature saliency heatmap, perform hierarchical scheduling: In highly sensitive areas, call the FPGA unit to perform physically guided graph matching; in non-sensitive areas, enable GPU parallel low-rank alignment. If the local sensitivity increases during processing, switch to high-precision process in real time; in metal-rich areas, force phase anti-interference alignment to be enabled.

[0148] Step c: Dual-modal fusion and dual-verification closed loop: The preliminary fusion results synchronously trigger physical screening and residual verification: Physical screening adopts a media-adaptive threshold; residual verification performs cause classification processing; the verification results are dynamically fed back to the preceding module: physical mismatch triggers geometric constraint enhancement and heat map recalibration; residual anomalies drive local refusion or threshold relaxation.

[0149] Step d: Validation-driven result generation: Only output data that passes dual validation, and add confidence markers to low-risk anomaly areas; when the physical screening is marked as "unmodelable interference", retain the original spectral features and add a warning label.

[0150] Alignment operations for dynamic resource allocation include:

[0151] Based on the heatmap, highly sensitive regions are assigned to FPGA computing units for graph matching.

[0152] Non-sensitive regions are allocated to GPU computing units to perform low-rank approximation calculations.

[0153] It should be further explained that, in the specific implementation process, when performing the alignment operation of dynamic resource allocation, a balance between high precision and high efficiency is achieved through the following hierarchical decision-making mechanism, including:

[0154] Heatmap-driven hardware scheduling: For the highest-sensitivity region marked in the heatmap (red), it is allocated in real time to the FPGA computing unit for physically guided graph matching: using the optimal transmission path model, refraction path calculation is automatically enabled in regions where the dielectric constant gradient is greater than 5% per meter; when the node spatial distance is greater than one-quarter λ, iterative compensation is initiated until sub-pixel convergence or a timeout switching strategy is implemented; the yellow secondary sensitive region is allocated to the GPU unit for low-rank approximate alignment: truncated SVD dimensionality reduction acceleration is adopted, and if the local gradient change rate suddenly increases beyond the threshold, it is immediately interrupted and handed over to the FPGA for processing; non-sensitive regions are aligned using batch matrix operations to ensure overall throughput.

[0155] Forced processing strategy for metal-rich areas: When a sudden increase in amplitude is detected in the spectral characteristics, i.e., exceeding the background by 3 times and the bandwidth is widened, it is automatically identified as a metal-rich area; regardless of the sensitivity level of the heat map, the FPGA is forced to perform anti-interference alignment: the phase peak matching is used to replace zero-point tracking; the calibration results are propagated through neighboring nodes to avoid phase jump traps; if the proportion of the metal area exceeds the upper limit of the processing unit load, segmented processing is started and boundary buffers are added.

[0156] Real-time resource conflict arbitration includes: FPGA overload conditions: when the number of red areas exceeds the number of FPGA parallel channels, sort by heatmap intensity: the highest intensity area is processed first; the second highest intensity area is downgraded to GPU high-precision mode and full-rank SVD is enabled; GPU efficiency bottleneck: when the amount of data in the yellow area reaches the memory limit: data block pipeline processing is started; overlapping cache is added to the boundary blocks to prevent truncation errors; all interrupted tasks record the breakpoint status and automatically resume transmission after resource release.

[0157] It should be further explained that, in the specific implementation process, the input multi-dimensional ground-penetrating radar data, after preprocessing, is separated into spatial topology mode and spectral feature mode by a dual-mode decoupling module. The spatial topology mode includes antenna trajectory coordinates and geometric relationships, and key features are extracted through a lightweight spatiotemporal encoder. The spectral feature mode includes the frequency spectrum and polarization parameters, and is input to a frequency-domain sparse attention network with an embedded differentiable geometric constraint layer. This constraint layer dynamically implements physical rules: for the electromagnetic wave propagation direction, when the deviation between the network output direction vector and the theoretical value increases, gradient loss correction weights are generated; for energy attenuation, the regularization intensity is dynamically adjusted according to the medium type. If the input data signal-to-noise ratio is lower than the set standard, the energy constraint weights are automatically reduced to avoid noise interference; when the detection depth exceeds a threshold, the integrity of low-frequency features is prioritized.

[0158] A heatmap based on dual-modal features is generated: sensitive regions are identified through spectral network gradient analysis, and combined with the results of spatial curvature abrupt change detection, the region is divided into high-sensitivity and secondary-sensitivity regions. When a high-gradient region completely coincides with the geometric boundary, it is marked as the highest sensitivity level; if only a single sensitivity indicator exists, it is marked as the secondary sensitivity level; when a low-gradient region overlaps with a strong geometric abrupt change, phase coherence verification is initiated, and if it fails, it is judged as a false sensitive region and is not marked. A noise adaptive mechanism is implemented in the heatmap generation process: when electromagnetic interference exceeds the standard, the geometric abrupt change judgment threshold is automatically increased.

[0159] The dynamic scheduling unit allocates computing resources based on the heatmap: highly sensitive areas are assigned to the first computing unit for high-precision map matching and alignment. This process constructs a cross-dimensional feature point map structure, using a straight-line path model to calculate spatial displacement compensation in homogeneous medium regions; when a sudden change in dielectric constant gradient is detected, it automatically switches to a refraction path model. In time-dimensional alignment, regions with a phase difference of less than 90 degrees between signals from the same source are directly calibrated; regions with phase jumps employ a combination strategy of amplitude peak matching and neighboring node propagation calibration. Secondary sensitive areas are assigned to the second computing unit for low-rank approximate alignment; if the local gradient change rate exceeds the warning value during processing, the process is immediately interrupted and transferred to the first computing unit. Non-sensitive areas use batch matrix operations to ensure throughput. Metal-rich regions are forcibly enabled with an anti-interference alignment strategy regardless of sensitivity level; when the proportion of metal regions exceeds the processing unit's load capacity, segmented processing is performed and boundary buffers are added to prevent truncation errors.

[0160] The aligned bimodal features are processed by a fusion network to generate preliminary results, which then proceed to the dual verification stage.

[0161] Physical consistency screening: Electromagnetic wave propagation is reconstructed using a cross-domain physical simulator. Core verification metrics include reflection energy attenuation rate and time difference between multiple reflections. Differentiated thresholds are set for different media types: a lenient threshold is used for soil layers, and a strict threshold is used for rock layers; energy, timing, and polarization metrics are verified in parallel for safety-sensitive areas. Screening results trigger tiered responses: minor anomalies in a single frame are only logged; moderate anomalies in three consecutive frames are reduced in physical constraint weights and locally reprocessed; severe anomalies in a single frame or verification failure in critical areas freeze the data and initiate cross-module joint optimization, strengthening geometric constraints, re-labeling heatmaps, and performing full-precision map matching.

[0162] Residual adversarial verification: The spatial residual channel detects boundary misalignment artifacts in the 3D model. When the artifact area overlaps with the high-sensitivity area and the curvature is continuous, it is determined to be a real anomaly; if it exists isolated in the non-sensitive area, spectral cross-verification is initiated. The spectral residual channel identifies phase jump points and classifies their causes: metallic interference types retain original features, alignment mismatch types trigger local re-fusion, and noise-induced types are dynamically filtered. The dual-channel verification results are processed with confidence linkage: high position signal alarms immediately freeze the output or drive reconstruction; low position signal warnings initiate cross-channel arbitration; persistent unresolved anomalies automatically relax the region detection threshold.

[0163] The final output module generates a 3D geological model with confidence level labels. Tolerance labels are added to low-risk anomaly areas marked by physical screening; electromagnetic property labels are added to metallic targets; and raw spectral characteristics and warning information are output for unmodelable interference areas that fail verification.

[0164] Through the synergy of dynamic physical constraint layer and heat map-driven hierarchical scheduling mechanism, the system achieves cross-dimensional alignment of multi-dimensional data from ground-penetrating radar under strict real-time requirements, and a forced anti-interference alignment strategy for metal interference zones, effectively solving the position deviation problem caused by phase jump.

[0165] By relying on the dual closed-loop mechanism of physical screening and residual verification, the engineering credibility of the system output fusion results is improved. The real-time parameter feedback and resource conflict arbitration mechanism driven by the verification results enable the system to have self-optimization capabilities in complex geological environments, and the output three-dimensional geological model can directly support construction decisions.

[0166] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional ground-penetrating radar data intelligent fusion processing system, characterized in that, include: The module includes a dual-modal decoupling module, an adaptive alignment module, a feature fusion module, and a dual verification module. The dual-mode decoupling module separates the input multi-dimensional geological radar data into spatial topological modes and spectral feature modes; The adaptive alignment module includes: A lightweight spatiotemporal encoder that processes spatial topological modes to generate spatial feature vectors; A frequency-domain sparse attention network processes spectral feature modes and embeds a differentiable geometric constraint layer to generate spectral feature vectors. A dynamic scheduling unit allocates computing resources based on the saliency heatmap: performing high-precision map matching and alignment for highly sensitive regions and low-rank approximate alignment for non-sensitive regions; the method for generating the saliency heatmap includes: Extracting gradient sensitivity of cross-dimensional features in frequency domain sparse attention networks; Identify regions of abrupt geometric structural changes in spatial feature vectors; A feature saliency heatmap is generated by fusing gradient sensitivity with regions of abrupt geometric changes, and regions whose feature saliency exceeds a preset threshold are marked as high-sensitivity regions; The specific steps for high-precision map matching and alignment are as follows: Construct a graph structure node for cross-dimensional feature points; cross-dimensional graph structure construction: use geometric key points in spatial feature vectors as reference nodes; associate corresponding points in spectral feature vectors as cross-dimensional edges; where, corresponding points: satisfy the phase coherence threshold; when the number of nodes for the same ground feature is mismatched in B-scan and C-scan data, start the node fusion mechanism, retain strong coherence nodes, and remove isolated points with a signal-to-noise ratio lower than the background level; Achieve sub-pixel-level alignment in spatial dimensions based on the optimal transmission path; Time dimension alignment is achieved based on spectral phase coherence. The feature fusion module receives the aligned dual-modal features and outputs a preliminary fusion result; The dual verification module performs physical consistency screening and residual adversarial verification on the fusion results; The physical consistency screening operation includes: The alignment features are input into a cross-domain physics simulator to reconstruct the electromagnetic wave propagation path; When the deviation between the simulated echo and the actual data in terms of reflection energy attenuation rate, time difference between multiple reflections, or polarization rotation angle exceeds a threshold, the alignment module is triggered for iterative optimization. The residual adversarial verification includes: The first channel uses a convolutional neural network to detect boundary misalignment artifacts in the spatial dimension; The second channel uses a time-frequency analysis network to detect phase jumps in the spectral dimension; When any channel detects a boundary misalignment artifact or a phase jump exceeding a threshold, it feeds back a reconstruction signal to the feature fusion module.

2. The intelligent fusion processing system for multi-dimensional geological radar data according to claim 1, characterized in that: The implementation method of the differentiable geometric constraint layer is as follows: Wave propagation direction constraint of Snell's law can be implemented using differentiable functions; Energy constraints for the medium decay equation are achieved through differentiable regularization terms. The above constraints serve as the physical driving optimization objectives for frequency-domain sparse attention networks.

3. The intelligent fusion processing system for multi-dimensional ground-penetrating radar data according to claim 1, characterized in that: The electromagnetic wave propagation model of the cross-domain physics simulator includes: Wave field solver based on Maxwell's equations; An adaptive mapping mechanism between dielectric parameters and conductivity.

4. A multi-dimensional geological radar data intelligent fusion processing system according to any one of claims 1-3, characterized in that, Also includes: The data preprocessing module performs noise reduction, gain correction, and dimension standardization operations on the raw ground-penetrating radar data; The results output module converts the validated fused data into a three-dimensional geological model.

5. A method for intelligent fusion processing of multi-dimensional ground-penetrating radar data, characterized in that, Using the system described in any one of claims 1-4, the steps include: S1: The data is separated into spatial topological modes and spectral characteristic modes through a dual-modal decoupling module; S2: Perform alignment operations for dynamic resource allocation through the adaptive alignment module; S3: Generate preliminary fusion results through the feature fusion module; S4: The results are subjected to dual verification using the dual verification module, which checks both the physical rules and the data residuals.

6. The intelligent fusion processing method for multi-dimensional ground-penetrating radar data according to claim 5, characterized in that: Alignment operations for dynamic resource allocation include: Based on the heatmap, highly sensitive regions are assigned to FPGA computing units for graph matching. Non-sensitive regions are allocated to GPU computing units to perform low-rank approximation calculations.

Images