Multi-dimensional geological radar data intelligent fusion processing system

Abstract

The invention, which relates to the technical field of geological radar detection, discloses a multi-dimensional geological radar data intelligent fusion processing system comprising a bimodal decoupling module, an adaptive alignment module, a feature fusion module and a dual verification module. The bimodal decoupling module is used for separating input multi-dimensional geological radar data into a spatial topology mode and a spectrum characteristic mode; the adaptive alignment module comprises a lightweight space-time encoder for processing a spatial topology mode to generate a spatial feature vector; and the frequency domain sparse attention network is used for processing a frequency spectrum characteristic mode and embedding a differentiable geometric constraint layer. According to the multi-dimensional geological radar data intelligent fusion processing system, through cooperation of a dynamic physical constraint layer and a thermodynamic diagram driven hierarchical scheduling mechanism, cross-dimensional alignment of geological radar multi-dimensional data under strict real-time requirements and a forced anti-interference alignment strategy of a metal interference area are realized by the system; and the problem of position deviation caused by phase jump is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological radar detection, in particular to a multi-dimensional geological radar data intelligent fusion processing system. BACKGROUND

[0002] The geological radar technology is widely used in the underground detection field, and the data collected by the geological radar technology naturally has multi-dimensional characteristics, including different scanning modes such as B-scan, C-scan, different polarization directions, different frequencies and different time sequences. These multi-dimensional data contain rich and complementary underground information. In order to more comprehensively and accurately analyze the underground structure, it is necessary to intelligently fuse and process the geological radar data from different dimensions.

[0003] In the prior art, a multi-dimensional geological radar data fusion system based on a general electric digital data processing framework faces significant challenges in processing such high-dimensional heterogeneous data streams. Especially in application scenarios that require real-time or near real-time processing, such as tunnel construction advanced prediction or road disease rapid detection, the problem is particularly prominent. The lack of cross-dimensional feature alignment accuracy can significantly amplify the errors in the subsequent fusion processing stage. If the feature points that fail to be accurately corresponded are input into the fusion model, it will lead the model to learn incorrect correlation rules, and the final output of the fusion result, such as a three-dimensional reconstruction image or a target recognition result, may appear target blur, positioning deviation or introduce false abnormalities, which seriously weakens the credibility and practical application value of the fusion result. Therefore, how to effectively improve the accuracy of cross-dimensional feature alignment of multi-dimensional geological radar data under the strict constraint of real-time requirement has become a key bottleneck restricting the performance improvement of the intelligent fusion system. Therefore, the technical problem to be solved at present is: under the premise of meeting the real-time processing requirement, how to significantly improve the accuracy of cross-dimensional feature alignment of multi-dimensional geological radar data to ensure the credibility of the fusion result. SUMMARY

[0004] To achieve the above object, the present application is implemented by the following technical scheme: a multi-dimensional geological 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; The dual-modal decoupling module separates the input multi-dimensional geological radar data into a spatial topology mode and a spectral feature mode; The adaptive alignment module comprises: A lightweight space-time encoder processes the spatial topology mode to generate a spatial feature vector; A frequency domain sparse attention network processes the spectral feature mode and embeds a differentiable geometric constraint layer; A dynamic scheduling unit allocates computing resources according to a feature saliency heat map: high-precision image matching alignment is performed on high-sensitive areas, and low-rank approximate alignment is performed on non-sensitive areas. The feature fusion module receives the aligned bimodal features and outputs a preliminary fusion result; The double verification module performs physical consistency screening and residual error adversarial verification on the fusion result.

[0005] Preferably, the implementation of the differentiable geometric constraint layer is: The wave propagation direction constraint of Snell's law is realized by a differentiable function; The energy constraint of the medium attenuation equation is realized by a differentiable regularization term; The above constraints are used as physical driving optimization targets of the frequency domain sparse attention network.

[0006] Preferably, the method for generating the feature saliency heat map comprises: Extracting the gradient sensitivity of cross-dimensional features in the frequency domain sparse attention network; Identifying the geometric structure mutation area in the spatial feature vector; Fusing the gradient sensitivity and the geometric mutation area to generate a heat map, and marking the area with a sensitivity exceeding a preset threshold as a high sensitivity area.

[0007] Preferably, the specific steps of the high-precision image matching alignment are: Constructing a graph structure node of cross-dimensional feature points; Realizing sub-pixel level alignment of spatial dimensions based on optimal transmission paths; Realizing time dimension alignment based on spectral phase coherence.

[0008] Preferably, the operation of the physical consistency screening comprises: Inputting the aligned features into a cross-domain physical simulator to reconstruct the electromagnetic wave propagation path; When the deviation of the simulated echo from the actual data in terms of reflection energy attenuation rate, multiple reflection time difference, or polarization rotation angle exceeds a threshold, triggering the alignment module to iteratively optimize.

[0009] Preferably, the residual error adversarial verification comprises: The first channel adopts a convolutional neural network to detect boundary misplacement artifacts in the spatial dimension; The second channel adopts a time-frequency analysis network to detect phase jumps in the spectral dimension; When either channel detects a residual error confidence exceeding a threshold, feeding back a reconstructed signal to the feature fusion module.

[0010] Preferably, the electromagnetic wave propagation model of the cross-domain physical simulator comprises: A wave field solver based on Maxwell's equations; An adaptive mapping mechanism for medium dielectric parameters and conductivity.

[0011] Preferably, it also includes: A data preprocessing module performs denoising, gain correction and dimension standardization operations on the original ground penetrating radar data; A result output module converts the fusion data that passes the verification into a three-dimensional geological model.

[0012] A multi-dimensional ground penetrating radar data intelligent fusion processing method adopts a multi-dimensional ground penetrating radar data intelligent fusion processing system, including the following steps: S1: Separate the data into spatial topology modalities and spectral feature modalities through a dual-modality decoupling module; S2: Perform a dynamic resource allocation alignment operation through an adaptive alignment module; S3: Generate a preliminary fusion result through a feature fusion module; S4: Perform a dual inspection of physical rules and data residuals on the result through a dual verification module.

[0013] Preferably, the dynamic resource allocation alignment operation includes: According to the heat map, allocate high sensitivity areas to FPGA computing units to perform graph matching; Allocate non-sensitive areas to GPU computing units to perform low-rank approximation calculation.

[0014] The present application provides a multi-dimensional ground penetrating radar data intelligent fusion processing system. It has the following beneficial effects: The multi-dimensional ground penetrating radar data intelligent fusion processing system, through the cooperation of the dynamic physical constraint layer and the heat map driven hierarchical scheduling mechanism, realizes the cross-dimension alignment of the multi-dimensional ground penetrating radar data under the strict real-time requirement, the forced anti-interference alignment strategy of the metal interference area, and effectively solves the position deviation problem caused by phase jump.

[0015] The multi-dimensional ground penetrating radar data intelligent fusion processing system, relying on the dual closed-loop mechanism of physical screening and residual verification, improves the engineering credibility of the system output fusion result, and the parameter real-time feedback and resource conflict arbitration mechanism driven by the verification result, so that the system has self-optimization ability in complex geological environment, and the output three-dimensional geological model can directly support construction decision. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a module interaction schematic diagram of the multi-dimensional ground penetrating radar data intelligent fusion processing system of the present application; Figure 2 It is a flowchart of the multi-dimensional ground penetrating radar data intelligent fusion processing method of the present application. DETAILED DESCRIPTION

[0017] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0018] Please refer to Figure 1 and Figure 2 The present application provides a technical solution: a multi-dimensional geological radar data intelligent fusion processing system, comprising: a dual-modal decoupling module, an adaptive alignment module, a feature fusion module and a double verification module. The dual-modal decoupling module separates the input multi-dimensional geological radar data into a spatial topology mode and a spectral feature mode. The adaptive alignment module comprises: A lightweight space-time encoder processes the spatial topology mode to generate a spatial feature vector. A frequency domain sparse attention network processes the spectral feature mode and embeds a differentiable geometric constraint layer. A dynamic scheduling unit allocates computing resources according to a feature saliency heat map: high-precision graph matching alignment is performed on high-sensitive areas, and low-rank approximate alignment is performed on non-sensitive areas. The feature fusion module receives the aligned dual-modal features and outputs a preliminary fusion result. The double verification module performs physical consistency screening and residual adversarial verification on the fusion result.

[0019] It needs to be further explained that in the specific implementation process, after the input multi-dimensional geological radar data is preprocessed by denoising and standardization, it is first decoupled into a spatial topology mode and a spectral feature mode. The spatial topology mode contains scan trajectory coordinates and antenna spacing geometric relationships, and the spectral feature mode contains frequency spectrum, phase difference and polarization parameters. The spatial topology mode extracts key geometric features through a lightweight space-time encoder to generate a spatial feature vector. The spectral feature mode is input into a frequency domain sparse attention network, which embeds a differentiable geometric constraint layer. This layer converts the wave propagation angle constraint of Snell's law into a differentiable loss function of direction vector, and converts the energy constraint of medium attenuation equation into a regularization term, forcing the network output to conform to the spectrum feature vector of the physical law.

[0020] Based on two types of feature vectors, a feature saliency heat map is generated: the sensitive area across the dimension data is calculated by the gradient back propagation of the spectrum network, combined with the curvature mutation point detection result in the space vector, and the high sensitive area which has significant influence on the alignment error is marked; the high sensitive area contains the underground cavity boundary and pipeline intersection. The dynamic scheduling unit calls the FPGA computing unit to perform high-precision map matching alignment in the high sensitive area according to the heat map distribution, constructs the feature point map structure, calculates the sub-pixel level displacement compensation through the optimal transmission path, and calibrates the time dimension by using the phase coherence; the GPU computing unit is called to perform low-rank approximation alignment in the non-sensitive area, and the truncated SVD dimension reduction is used to realize fast transformation.

[0021] The aligned bimodal features generate preliminary three-dimensional fusion results through the fusion network, and then enter the double verification stage, including the following two stages: Physical consistency screening stage: the fusion features are input into the cross-domain physical simulator to reconstruct the propagation of electromagnetic waves in the corresponding medium model. If the deviation of the core indicators between the simulation echo and the actual data exceeds the preset threshold, an iteration instruction is sent to the alignment module; the cross-domain physical simulator is constructed based on Maxwell's equations, and the core indicators include the reflection energy attenuation rate and the multiple reflection time difference; Residual adversarial verification stage: the dual-channel discriminator works synchronously, the spatial residual channel detects the boundary misplacement artifacts in the three-dimensional model, and the spectral residual channel identifies the phase jump anomaly. When the residual confidence of any channel output exceeds the standard, the reconstruction training of the fusion network is triggered.

[0022] Finally, the fusion data that passes the verification is output as a three-dimensional geological structure model and a target identification report, and the data that does not pass the verification is returned to the alignment stage for re-optimization.

[0023] The implementation method of the differentiable geometric constraint layer is: The wave propagation direction constraint of Snell's law is realized through a differentiable function; The energy constraint of the medium attenuation equation is realized through a differentiable regularization term; The above constraints are used as the physical driving optimization target of the frequency domain sparse attention network.

[0024] It needs to be further explained that, in the specific implementation process, when the frequency domain sparse attention network processes the spectral feature mode, the embedded differentiable geometric constraint layer dynamically implements the physical rule constraint in the following ways, including: For the electromagnetic wave propagation direction constraint, the incident angle and the refracted angle relationship of Snell's law is converted into a differentiable function of direction vector: when the deviation between the network output wave propagation direction vector and the theoretical direction calculated based on the medium dielectric constant increases, the gradient loss back propagation is generated to force the network to adjust the feature extraction weight to comply with the physical law; if there is strong noise interference in the actual data, such as tunnel steel arch reflection clutter, the constraint automatically reduces the weight to avoid overfitting noise.

[0025] For the energy attenuation constraint, a differentiable regularization term is constructed based on the medium attenuation equation: according to the energy attenuation slope of different frequency components in the spectral feature, the theoretical attenuation curve is dynamically calculated; when the difference between the network output spectral energy distribution and the theoretical curve exceeds the adaptive threshold, the regularization strength is increased to correct the feature vector; in the uniform medium area, such as the soil layer, the constraint uses a loose threshold to improve processing efficiency.

[0026] The implementation of the constraint layer includes a working condition adaptive judgment mechanism, which includes the following: For high noise working conditions: when the signal-to-noise ratio of the input data is lower than the set threshold, the attenuation equation constraint strength is automatically reduced to prioritize feature extraction integrity; For medium mutation working conditions: when the spatial topology mode detects a region with a dramatic change in dielectric constant, such as a rock-soil interface, the wave propagation direction constraint is immediately strengthened to force the network to output a refractive feature that complies with Snell's law in that region; For low frequency dominant working conditions: when the detection depth exceeds the effective range of high frequency signals, switch to low frequency attenuation dominant mode and relax the high frequency energy matching requirement.

[0027] The output of the constraint layer is used as the spectral feature vector to enter the subsequent module, and its physical compliance is checked by the physical consistency screening unit in the double verification module: if the energy attenuation mode of the simulation echo deviates significantly from the spectral feature output by the constraint layer, such as a continuous three-frame attenuation rate exceeding the standard, a parameter reset instruction is sent to the constraint layer.

[0028] The generation method of the feature saliency heat map includes: Extract the gradient sensitivity of cross-dimensional features in the frequency domain sparse attention network; Identify the geometric structure mutation area in the spatial feature vector; Fuse the gradient sensitivity and the geometric mutation area to generate a heat map, and mark the area with sensitivity exceeding the preset threshold as a high sensitivity area.

[0029] It needs to be further explained that in the specific implementation process, the generation of the feature saliency heat map is realized through the dual-channel cooperation of gradient sensitivity analysis and geometric mutation detection, and the process is as follows: Gradient sensitivity analysis: During the forward propagation of the frequency-domain sparse attention network, the gradient amplitude 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 area. If there is a sudden change in spectral energy in this region, such as an energy difference between adjacent frequency points exceeding 30% of the mean value, the sensitivity level is raised to the highest level.

[0030] Geometric discontinuity detection: Calculate the curvature field of the spatial feature vector to identify geometric discontinuities with a second-order derivative of curvature exceeding a threshold, such as underground pipeline edges and rock fracture surfaces. When discontinuities form a closed boundary, they are automatically expanded to sensitive areas.

[0031] The results of the two channels are fused according to logical decision rules, which include the following three working conditions: High sensitivity working condition: When the gradient high-level sensitive area completely overlaps with the geometric boundary, such as pipeline intersection points, it is directly marked as the highest sensitive level, i.e., the red area, triggering subsequent high-precision alignment. Weak sensitivity working condition: Only gradient sensitivity or isolated geometric mutation, such as weak anomalies in homogeneous media, is marked as the secondary sensitive level, i.e., the yellow area, using a simplified alignment strategy. Conflict working condition: When the gradient low-level area overlaps with the geometric strong mutation area, such as false boundaries caused by metal interference, start the spectral coherence verification. If the phase consistency is lower than the threshold, it is determined as a false sensitive area, i.e., not marked.

[0032] The update of the heat map implements a dynamic threshold adjustment mechanism, including: In a strong noise environment, such as an environment with excessive electromagnetic interference intensity, the geometric mutation detection threshold is automatically increased to avoid noise artifacts mislabeling. In deep exploration mode with a depth greater than 10 meters, the gradient sensitivity weight is reduced to prioritize geometric structure integrity. The generated heat map is input into the dynamic scheduling unit in real time to guide the allocation of computing resources. If the same area is repeatedly marked as a sensitive area in five consecutive frames of data but does not pass subsequent verification, the sensitivity analysis module triggers parameter self-correction.

[0033] The specific steps of high-precision map matching alignment are: Constructing graph structure nodes of cross-dimensional feature points; Achieving sub-pixel level alignment in spatial dimensions based on optimal transmission paths; Achieving time dimension alignment based on spectral phase coherence.

[0034] It needs to be further explained that, in the specific implementation process, when the dynamic scheduling unit allocates high-sensitive areas to the high-precision map matching alignment process, the following physical rule-driven operations are performed, including: Cross-dimension graph structure construction: take geometric key points in spatial feature vector as reference nodes; associate corresponding points in frequency spectrum feature vector as cross-dimension edges; wherein, corresponding points: meet phase coherence threshold; when the number of nodes of the same ground object in B-scan and C-scan data does not match, such as the edge of a karst cave, a node fusion mechanism is started: keep strong coherence nodes, and remove isolated points with a signal-to-noise ratio lower than the background level.

[0035] Spatial dimension sub-pixel alignment: calculate the optimal transmission path based on the electromagnetic wave propagation speed model: adopt a straight path constraint in a uniform medium area; automatically switch to a refraction path model in a dielectric constant mutation area, such as a concrete-soil interface; realize sub-pixel level offset compensation by inverting the spatial displacement amount through the path length difference; if the distance between nodes after compensation still exceeds 1 / 4 of the wavelength, trigger iterative optimization until convergence.

[0036] Time dimension phase calibration: perform coherence verification on the spectral phase difference of cross-dimension edges: when the phase difference of homologous signals is less than π / 2, directly align the time stamp; when the phase difference is greater than π / 2, but meets the linear gradient rule, such as the reflection of an inclined interface, use phase unwrapping correction; when the phase jumps caused by metal interference, such as steel mesh reflection, prefer to match the amplitude peak rather than the phase zero point; propagate the calibration results through adjacent nodes.

[0037] The alignment results are fed back to the heat map generation module in real time: if the alignment iteration of a certain area does not converge for three consecutive times, the sensitivity level of the area is downgraded and switched to a low-rank approximate alignment process.

[0038] The operation of the physical consistency screening includes: Input the alignment features into the cross-domain physical simulator to reconstruct the electromagnetic wave propagation path; When the deviation of the simulated echo and the actual data in reflection energy attenuation rate, multiple reflection time difference or polarization rotation angle exceeds the threshold, trigger the iteration optimization of the alignment module.

[0039] It needs to be further explained that, in the specific implementation process, the physical consistency screening unit receives the aligned bimodal features, and reconstructs the electromagnetic wave propagation process through the cross-domain physical simulator. This process implements a hierarchical threshold verification mechanism: Basic index verification: calculate the energy attenuation rate deviation of the simulated echo and the actual data at the target reflection point, and mark the primary anomaly when the deviation exceeds the reference threshold allowed by the medium type; if the multi-path reflection wave arrival time difference simulation value and the measured value of the same target deviate more than half a wavelength corresponding time, mark the timing anomaly; wherein, the reference threshold includes soil layer ≤8%, rock layer ≤5%.

[0040] Key area reinforcement verification: For the highest sensitive area marked in the heat map, increase the consistency detection of the polarization rotation angle. When the difference between the simulation and the measured polarization angle exceeds the angle tolerance, it is directly determined as a physical mismatch. In safety-sensitive areas such as tunnel vaults and bridge pile foundations, three indicators are used in parallel verification, namely: energy + timing + polarization, and any indicator exceeding the standard will trigger an alarm. The highest sensitive area includes underground pipeline intersections, and the angle tolerance includes shallow ≤5° and deep ≤8°.

[0041] Among them, the screening result triggers intelligent iterative decision-making, including: Single-frame mild anomaly: Energy attenuation deviation is within 1-1.5 times of the reference threshold, only record log without triggering iteration; Three consecutive frames of moderate anomaly: Deviation continues to exceed 1.5 times of the reference threshold but does not reach 2 times, reduce the physical constraint weight of the alignment module and reprocess locally; Single-frame severe anomaly or regional joint inspection failure: Deviation exceeds 2 times of the reference threshold or any of the three indicators in the key area fails, freeze the current frame data and start cross-module joint optimization, including: send geometric constraint reinforcement instructions to the alignment module; require the heat map module to recalibrate the sensitive area; perform full-precision map matching on the fault area.

[0042] If joint optimization still cannot pass the verification, mark the data as "physically uninterpretable" and generate a fault report for manual intervention.

[0043] Residual adversarial verification includes: The first channel uses a convolutional neural network to detect boundary misplacement artifacts in the spatial dimension; The second channel uses a time-frequency analysis network to detect phase jumps in the frequency spectrum dimension; When either channel detects residual confidence exceeding the threshold, feedback the reconstructed signal to the feature fusion module.

[0044] It needs to be further explained that in the specific implementation process, the residual adversarial verification unit performs a double-tunnel parallel detection process on the fusion result, including a spatial residual channel and a frequency spectrum residual channel, wherein: Spatial residual channel: Use a convolutional neural network to scan the three-dimensional fusion model to identify the feature patterns of boundary misplacement artifacts, such as jagged edges and local grid distortion; when a suspected artifact is detected, start regional confidence assessment: if the artifact region overlaps with the high sensitive area of the heat map and the curvature changes continuously, it is determined as a real anomaly, i.e. high confidence; if it exists in isolation in a non-sensitive area, perform spectral cross-verification.

[0045] Frequency spectrum residual channel: Use a time-frequency analysis network to extract phase continuity features and detect jump points; classify the causes of jump points, including: Metallic interference: accompanied by sharp amplitude rise: 3 times background and band broadening, marked as physical real; Alignment mismatch: no amplitude anomaly and stable band, marked as processing error; Noise-induced: randomly scattered and SNR below threshold, directly filtered.

[0046] Dual-channel results are processed according to confidence linkage rules, including: High position confidence alarm, that is, any channel confidence greater than 0.8: spatial channel reports real anomaly, freezes output and triggers physical screening unit review; spectral channel reports alignment mismatch, immediately sends reconstruction signal to fusion module; Low position confidence warning, that is, dual-channel confidence is between 0.5-0.8: start cross-channel arbitration: if the spatial distance between the spatial artifact region and the spectral jump point is less than λ / 2, it is jointly marked as potential anomaly; otherwise, it is downgraded to acceptable error; Persistent pending anomaly, that is, three consecutive low position warnings: automatically relax the detection threshold of this region by 20% to avoid excessive interference with real-time processes.

[0047] Verification results drive fusion network fine-tuning: for metallic interference spectrum jump, retain the original phase features; for alignment mismatch, trigger local feature re-fusion.

[0048] The electromagnetic wave propagation model of the cross-domain physical simulator includes: a wave field solver based on Maxwell's equations; an adaptive mapping mechanism for dielectric parameters and conductivity of the medium. It should be further explained that, in the process of implementation, when the cross-domain physical simulator performs electromagnetic wave propagation reconstruction, the adaptive mapping mechanism for medium parameters is used to dynamically construct the calculation model, and the construction process is as follows: Receive geometric structure information in the spatial feature vector and dielectric properties in the spectral feature vector to generate an initial medium grid; wherein the geometric structure information includes layered interfaces and cavity positions, and the dielectric properties include attenuation slope and dispersion relationship; when the actual detection depth exceeds the preset threshold, automatically strengthen the low-frequency dielectric parameter weight and suppress high-frequency parameter noise interference; for the high sensitivity area marked by the thermal map, implement sub-grid encryption mapping: reduce the standard grid size to one fourth of the wavelength, and use dielectric constant gradient smoothing technology in the encryption area to avoid simulation artifacts caused by parameter mutations.

[0049] The wave field solving process implements a multi-scale intelligent switching strategy, including the following: Shallow fast solving mode, that is: depth does not exceed 3 meters: use paraxial approximation wave equation, ignore multiple reflection paths; when a metallic target spectrum amplitude increases by more than 3 times the background, switch to full wave field solution locally; Deep high-precision mode, i.e. depth greater than 3 meters: enable time-domain full-wave solver, automatically truncate the calculation area to the target body extension 1.5 times range; if the signal-to-noise ratio of the deep target is lower than the threshold, activate the waveguide mode enhancement algorithm: preferentially retain the guide wave component propagating along the stratigraphic interface; Safety-sensitive area, i.e. tunnel vault: force full-time domain solution, and increase twice reflection path calculation.

[0050] Boundary condition dynamic compensation mechanism handles metal interference effect: when the frequency spectrum residual channel identifies metal interference type jump, automatically add equivalent dipole boundary at the corresponding spatial position: the dipole moment strength is proportional to the measured echo amplitude; the direction is calculated by polarization angle deviation inversion; after compensation, re-execute simulation, if the energy attenuation rate deviation is reduced to within the threshold, retain the compensation parameter; otherwise, mark as "unmodelable interference".

[0051] Also includes: Data preprocessing module, performs denoising, gain correction and dimension standardization operations on original ground penetrating radar data; Result output module converts the fusion data that passes the verification to a three-dimensional geological model.

[0052] It needs to be further explained that in the specific implementation process, the data preprocessing module implements a working condition adaptive processing chain, including the following: Dynamic denoising strategy: in a strong electromagnetic interference environment, such as around a substation, 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, automatically switch to adaptive notch filtering to eliminate interference at specific frequencies; if the input signal-to-noise ratio is higher than the threshold, only perform basic median filtering to ensure real-time performance.

[0053] Gain correction mechanism: for shallow detection, i.e. detection not exceeding 2 meters, linear gain compensation is used to prevent near-field saturation; for deep detection, i.e. detection greater than 5 meters, switch to exponential gain mode to enhance weak signals; when encountering an abnormally high amplitude area, start local gain suppression to avoid distortion.

[0054] Dimension standardization: uniformly resample different scan mode B / C-scan data to a reference grid; perform time axis registration on time-varying data to eliminate device movement errors.

[0055] The result output module implements verification-driven three-dimensional reconstruction: only receives fusion data that passes the double verification module; adds confidence labels to "low-risk abnormal areas" screened by physical consistency screening in the model; when the residual adversarial verification identifies a metal target, automatically adds an electromagnetic reflection characteristic label; outputs a three-dimensional visual model containing geological horizons, abnormal body spatial distribution and confidence labels.

[0056] A multi-dimensional geological radar data intelligent fusion processing method adopts a multi-dimensional geological radar data intelligent fusion processing system, including the following steps: S1: separating data into a spatial topology mode and a spectral feature mode through a bimodal decoupling module; S2: performing an alignment operation of dynamic resource allocation through an adaptive alignment module; S3: generating a preliminary fusion result through a feature fusion module; S4: performing a double check of physical rules and data residuals on the result through a double verification module.

[0057] It needs to be further explained that in the specific implementation process, when performing multi-dimensional geological radar data processing, intelligent fusion is realized through the following steps: Step a: bimodal decoupling separation: receiving preprocessed multi-source data, implementing decoupling decision based on scanning mode characteristics: B-scan data preferentially extracts spatial topology mode; C-scan data focuses on separating spectral feature mode; when inputting time-varying sequence data, automatically adding time dimension label for alignment module calling.

[0058] Step b: alignment operation of dynamic resource allocation: after generating feature saliency heat map, performing hierarchical scheduling: calling FPGA unit to perform physically guided graph matching in high sensitivity area; enabling GPU parallel low rank alignment in non-sensitive area, if local sensitivity rises during processing, switching to high precision process in real time; forcibly enabling phase anti-interference alignment in metal enrichment area.

[0059] Step c: bimodal fusion and double verification closed loop: preliminary fusion result synchronously triggers physical screening and residual verification: physical screening adopts medium adaptive threshold; residual verification performs cause classification processing; verification result is dynamically fed back to the previous module: physical mismatch triggers geometric constraint strengthening and heat map re-calibration; residual anomaly drives local re-fusion or threshold relaxation.

[0060] Step d: verification-driven result generation: only output data passing double verification, add confidence label to low-risk abnormal area; when physical screening label is "unmodelable interference", keep original spectral feature and add warning label.

[0061] The alignment operation of dynamic resource allocation includes: allocating high sensitivity area to FPGA computing unit to perform graph matching according to heat map; allocating non-sensitive area to GPU computing unit to perform low rank approximation calculation.

[0062] It needs to be further explained that in the specific implementation process, when performing the alignment operation of dynamic resource allocation, the following hierarchical decision mechanism is used to realize the balance between high precision and high efficiency, including: Heat map driven hardware scheduling: For the highest sensitive level region marked in the heat map: red, real-time allocation to FPGA computing unit to perform physically guided graph matching: adopt the optimal transmission path model, automatically enable refractive path calculation when the dielectric constant gradient is greater than 5% per meter region; when the spatial distance of the node is greater than one fourth of λ, start the iterative compensation until the sub-pixel converges or the timeout switching strategy; yellow secondary sensitive area is allocated to GPU unit to perform low rank approximation alignment: adopt truncated SVD dimension reduction acceleration, if the local gradient change rate suddenly increases above the threshold, immediately interrupt and transfer to FPGA processing; non-sensitive area uses batch matrix operation alignment to ensure overall throughput.

[0063] Metal enrichment area forced processing strategy: When the spectrum feature detects a sudden increase in amplitude, that is: more than 3 times the background and the frequency band is widened, it is automatically identified as a metal enrichment area; regardless of the heat map sensitivity level, force to use FPGA to perform anti-interference alignment: use the phase peak value matching instead of zero tracking; propagate the calibration results through adjacent nodes to avoid phase jump traps; if the metal area ratio exceeds the upper limit of the processing unit load, start segmented processing and add boundary cache.

[0064] Real-time resource conflict arbitration, including: FPGA overload working condition: when the number of red areas exceeds the number of FPGA parallel channels, sort by heat map intensity: the highest intensity area is processed first; the next high area is downgraded to GPU high precision mode, enabling full rank SVD; GPU efficiency bottleneck: when the data volume of yellow area reaches the upper limit of video memory: start data block pipelining processing; add overlapping cache to boundary blocks to prevent truncation errors; all interrupted tasks record breakpoint status, and automatically resume after resource release.

[0065] It needs to be further explained that in the specific implementation process, the input multi-dimensional ground penetrating radar data is pre-processed and separated into a spatial topology mode and a spectrum feature mode by a dual-mode decoupling module. The spatial topology mode contains antenna trajectory coordinates and geometric relationships, and extracts key features through a lightweight space-time encoder; the spectrum feature mode contains frequency spectrum and polarization parameters, and inputs a frequency domain sparse attention network embedded with a differentiable geometric constraint layer. The constraint layer dynamically implements physical rules: for the direction of electromagnetic wave propagation, when the deviation between the network output direction vector and the theoretical value increases, a gradient loss correction weight is generated; for energy attenuation, the regularization strength is dynamically adjusted according to the medium type. If the signal-to-noise ratio of the input data is lower than the set standard, the energy constraint weight is automatically reduced to avoid noise interference; when the detection depth exceeds the threshold, the low frequency feature integrity is preferentially guaranteed.

[0066] Based on the dual-mode feature generation feature saliency heat map: through the spectral network gradient analysis to identify sensitive areas, combined with the spatial curvature mutation detection results, the area is divided into high sensitive area and secondary sensitive area. When the gradient high area and the geometric boundary completely coincide, it is marked as the highest sensitive level; if there is only a single sensitive indicator, it is marked as a secondary sensitive level; when the low gradient area overlaps with the strong geometric mutation, the phase coherence verification is started, and if it fails, it is determined as a false sensitive area and is not marked. The heat map generation process implements a noise adaptive mechanism: when the electromagnetic interference exceeds the standard, the geometric mutation judgment threshold is automatically increased.

[0067] The dynamic scheduling unit allocates computing resources according to the heat map: the high sensitive area is allocated to the first computing unit to perform high precision map matching alignment. This process constructs a cross-dimensional feature point map structure, and uses a straight line path model to calculate the spatial displacement compensation in the uniform medium area; when a dielectric constant gradient mutation is detected, it automatically switches to a refractive path model. In the time dimension alignment, the area with a phase difference of less than ninety degrees between homologous signals is directly calibrated; the phase jump area uses a combination strategy of amplitude peak matching and adjacent node propagation calibration. The secondary sensitive area is allocated to the second computing unit to perform low rank approximation alignment, and if the local gradient change rate exceeds the warning value during processing, it is immediately interrupted and handed over to the first computing unit. The non-sensitive area uses batch matrix operation to ensure throughput. The metal-rich area, regardless of the sensitivity level, is forced to enable anti-interference alignment strategy, and when the metal area ratio exceeds the processing unit load capacity, segmented processing is performed and boundary cache is added to prevent truncation error.

[0068] After alignment, the dual-mode features are generated by the fusion network to generate preliminary results and enter the double verification stage: Physical consistency screening: reconstruct the electromagnetic wave propagation through cross-domain physical simulator, the core verification indicators include reflection energy attenuation rate and multiple reflection time difference. Different threshold values are set for different medium types: a loose threshold value is used for soil layer, and a strict threshold value is used for rock layer; energy, time sequence and polarization three indicators are used in parallel verification in safety sensitive area. Screening results trigger hierarchical response: single frame mild abnormality only records log; three consecutive frames of moderate abnormality reduce the physical constraint weight and reprocess locally; single frame serious abnormality or key area verification failure, freeze data and start cross-module joint optimization, strengthen geometric constraint, re-label heat map, and perform full precision map matching.

[0069] Residual adversarial verification: spatial residual channel detects three-dimensional model boundary mispositioning artifacts, and determines as true when the artifact area overlaps with high sensitivity area and the curvature is continuous; if it exists in isolation in a non-sensitive area, it will start spectral cross-verification. The spectral residual channel identifies phase jump points and classifies the causes: metal interference type retains original features, alignment mismatch type triggers local re-fusion, and noise-induced type dynamic filtering. The results of the dual-channel verification are subjected to confidence linkage processing: high position warning freezes output or drives reconstruction immediately; low position warning starts cross-channel arbitration; and persistent undecided abnormalities automatically relax the area detection threshold.

[0070] The final output module generates a three-dimensional geological model with confidence markers. Low-risk abnormal areas marked by physical screening are annotated with tolerance; metal targets are annotated with electromagnetic characteristics; and unmodelable interference areas that fail verification are output with original spectral features and warning information.

[0071] Through the cooperation of dynamic physical constraint layer and thermal map driven hierarchical scheduling mechanism, the system realizes cross-dimension alignment of geological radar multi-dimensional data under strict real-time requirements, forced anti-interference alignment strategy for metal interference areas, and effectively solves the position deviation problem caused by phase jump.

[0072] Relying on the dual closed-loop mechanism of physical screening and residual verification, the engineering credibility of the system output fusion results is improved, and the real-time feedback of parameters and resource conflict arbitration mechanism driven by verification results enable the system to have self-optimization ability in complex geological environment, and the output three-dimensional geological model can directly support construction decision-making.

[0073] It should be noted that, in this text, relationship terms such as first and second are only used to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0074] Although embodiments of the present application 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 thereto without departing from the principles and spirit of the present application, and the scope of the present application 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; 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; 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.

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 method for generating the feature 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; 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.

4. The intelligent fusion processing system for multi-dimensional geological radar data according to claim 1, characterized in that: The specific steps for high-precision map matching and alignment are as follows: Construct a graph structure node for cross-dimensional feature points; Achieve sub-pixel-level alignment in spatial dimensions based on the optimal transmission path; Time dimension alignment is achieved based on spectral phase coherence.

5. The intelligent fusion processing system for multi-dimensional geological radar data according to claim 1, characterized in that: 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.

6. The intelligent fusion processing system for multi-dimensional geological radar data according to claim 1, characterized in that: 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 that the residual confidence level exceeds the threshold, it feeds back a reconstruction signal to the feature fusion module.

7. The intelligent fusion processing system for multi-dimensional ground-penetrating radar data according to claim 5, 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.

8. A multi-dimensional ground-penetrating radar data intelligent fusion processing system according to any one of claims 1-7, 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.

9. A method for intelligent fusion processing of multi-dimensional ground-penetrating radar data, characterized in that, The system described in any one of claims 1-8 includes the following steps: 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.

10. The intelligent fusion processing method for multi-dimensional ground-penetrating radar data according to claim 9, 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