Ground penetrating radar data analysis method and device based on causal reasoning

Abstract

The application discloses a ground penetrating radar data analysis method and device based on causal reasoning, comprising: preprocessing and space-time coordinate mapping of ground penetrating radar original data of a target detection area to obtain a preprocessed radar image sequence; inputting the same into a ground penetrating radar target detection model to obtain a preliminary detection result set and a corresponding atomic state probability vector; constructing a causal reasoning graph and estimating structural causal model parameters and conditional probability tables to obtain a causal reasoning graph model; performing time sequence checking, logic evaluation and counterfactual verification on the preliminary detection result set, performing confidence suppression or missing target completion to obtain a modified detection result; and performing weighted fusion and state marking on the modified detection result and the preliminary detection result set to obtain a ground penetrating radar data analysis result. The application effectively suppresses false detection caused by environmental interference, noise influence and target shielding, and significantly improves the detection accuracy and robustness of underground targets, especially disease bodies.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing technology, and in particular to a method and apparatus for analyzing ground-penetrating radar data based on causal reasoning. Background Technology

[0002] Ground-penetrating radar (GPR), a technology that uses high-frequency electromagnetic waves for non-destructive testing of underground media, has been widely applied in fields such as urban road defect detection, engineering quality inspection, and geological surveys. However, existing GPR data interpretation methods still face fundamental limitations: on the one hand, data interpretation mainly relies on expert experience for manual interpretation, and due to the complexity of underground media and the ambiguity of radar images, the false signals generated by environmental interference factors such as metal objects, underground pipelines, and heterogeneous strata interfaces in radar images are similar to the actual defect morphology, making it difficult for traditional methods to effectively distinguish them. In addition, existing deep learning methods only establish the correlation between image features and target categories, lacking the ability to causally infer behavioral patterns and state changes. When target features are sparsified due to noise interference or missing due to occlusion, the recognition performance drops sharply. At the same time, most existing methods are based on independent processing of single-frame static images and fail to utilize the temporal logical consistency inherent in multi-temporal detection data to verify the detection results, resulting in the inability to effectively suppress false detections caused by feature loss or environmental disturbances.

[0003] Therefore, there is an urgent need for a new method that can combine causal reasoning mechanisms to perform in-depth logical analysis of ground-penetrating radar detection data, in order to overcome the false detection problems caused by environmental interference, noise, and target obstruction in existing technologies, and improve the accuracy and robustness of underground target detection. Summary of the Invention

[0004] The main objective of this application is to provide a ground-penetrating radar data analysis method based on causal reasoning, which aims to solve the technical problem of high false detection rate in the analysis of ground-penetrating radar detection data in the prior art due to environmental interference, noise, and target obstruction.

[0005] To achieve the above objectives, this application proposes a ground-penetrating radar (GPR) data parsing method based on causal reasoning. The causal reasoning-based GPR data parsing method includes: The raw ground-penetrating radar data of the target detection area is acquired, and the raw ground-penetrating radar data is preprocessed and subjected to spatiotemporal coordinate mapping to obtain a preprocessed radar image sequence. The preprocessed radar image sequence is input into a ground-penetrating radar target detection model based on a deep learning network for multi-scale feature extraction, temporal fusion, and atomic state prediction to obtain a preliminary detection result set and the corresponding atomic state probability vector. Based on the atomic state probability vector and combined with the prior knowledge of the ground penetrating radar detection scenario, a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer is constructed, and the structural causal model parameters and conditional probability table are estimated to obtain the causal inference graph model. Based on the causal reasoning graph model, the preliminary detection result set is subjected to temporal consistency verification, logical consistency evaluation, and counterfactual reasoning verification. Based on the verification results, confidence suppression or missing target completion is performed to obtain the corrected detection results. The corrected detection results are weighted and fused with the preliminary detection results, and then labeled with a status to obtain the ground penetrating radar data analysis results.

[0006] Furthermore, to achieve the above objectives, this application also proposes a ground-penetrating radar data parsing device based on causal reasoning, which includes: The acquisition module is used to acquire the raw ground-penetrating radar data of the target detection area, and to preprocess and perform spatiotemporal coordinate mapping on the raw ground-penetrating radar data to obtain a preprocessed radar image sequence. The prediction module is used to input the preprocessed radar image sequence into a ground-penetrating radar target detection model based on a deep learning network for multi-scale feature extraction, temporal fusion and atomic state prediction, so as to obtain a preliminary detection result set and the corresponding atomic state probability vector. The construction module is used to construct a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer based on the atomic state probability vector and the prior knowledge of the ground penetrating radar detection scenario, and to estimate the structural causal model parameters and conditional probability table to obtain the causal inference graph model. The correction module is used to perform temporal consistency verification, logical consistency evaluation, and counterfactual reasoning verification on the preliminary detection result set based on the causal reasoning graph model, and to perform confidence suppression or missing target completion based on the verification results to obtain the corrected detection results. The labeling module is used to perform weighted fusion and state labeling on the corrected detection results and the preliminary detection result set to obtain the ground penetrating radar data parsing results.

[0007] Furthermore, to achieve the above objectives, this application also proposes a ground-penetrating radar data parsing device based on causal reasoning. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the ground-penetrating radar data parsing method based on causal reasoning as described above.

[0008] In addition, to achieve the above objectives, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the ground-penetrating radar data parsing method based on causal reasoning as described above.

[0009] This application proposes one or more technical solutions that, by constructing a knowledge fence based on causal reasoning graphs, logically verify and constrain the preliminary detection results of deep learning, breaking through the limitations of traditional deep learning methods that rely solely on correlation analysis. This enables ground-penetrating radar data to transition from "feature-category" correlation recognition to "state-evolution" causal reasoning. Through a counterfactual reasoning mechanism, it uses temporal logic consistency constraints to suppress the confidence of detection results that do not conform to causal evolution laws. This makes it difficult for environmental interference factors to pass causal logic verification even if they present image features similar to real diseases, thus effectively suppressing false detections caused by environmental interference and noise. Through temporal consistency verification and missing target completion based on state transition prediction, when target features are missing due to occlusion or sparsity, it infers the target's state based on the causal logic of previous and subsequent frames and generates virtual targets. This solves the detection failure problem caused by feature sparsity and occlusion, enabling ground-penetrating radar data to transition from correlation recognition to causal reasoning. It effectively suppresses false detections caused by environmental interference, noise, and target occlusion, significantly improving the detection accuracy and robustness of underground targets, especially diseases. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating an embodiment of the ground-penetrating radar data parsing method based on causal reasoning in this application. Figure 2 This is a schematic diagram of the module structure of the ground-penetrating radar data parsing device based on causal reasoning, according to an embodiment of this application.

[0013] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0015] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0016] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a ground-penetrating radar data analysis device based on causal reasoning. The following description uses a ground-penetrating radar data analysis device based on causal reasoning as an example to illustrate this embodiment and the subsequent embodiments.

[0017] Based on this, embodiments of this application provide a ground-penetrating radar data parsing method based on causal reasoning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the ground-penetrating radar data parsing method based on causal reasoning of this application.

[0018] In this embodiment, the ground-penetrating radar data parsing method based on causal reasoning includes steps S10~S50: Step S10: Obtain the raw ground-penetrating radar data of the target detection area, and preprocess and perform spatiotemporal coordinate mapping on the raw ground-penetrating radar data to obtain the preprocessed radar image sequence.

[0019] It should be noted that raw ground-penetrating radar (GPR) data refers to the raw echo data collected by the GPR system in the target detection area according to a preset survey line, including the travel time, amplitude, and phase information of each sampling channel. Preprocessing and spatiotemporal coordinate mapping refer to performing error correction, noise suppression, signal enhancement, and spatial location calibration on the raw data to construct a radar image sequence suitable for subsequent deep learning processing.

[0020] Understandably, raw ground-penetrating radar (GPR) data is affected by antenna coupling effects, system noise, surface reflection, and environmental electromagnetic interference. Directly inputting it into the detection network would introduce a large number of spurious features. Furthermore, the raw data only contains a time-sampling matrix, lacking a correspondence with the actual spatial location. Therefore, a spatiotemporal coordinate mapping must be used to establish the conversion relationship between pixel coordinates and underground spatial coordinates.

[0021] In one feasible implementation, step 10 may include: collecting ground-penetrating radar data in the target detection area according to a preset survey line to obtain raw radar data; preprocessing the raw radar data to obtain preprocessed radar data, wherein the preprocessing includes at least zero-time correction, DC drift removal, background removal, gain adjustment, and bandpass filtering; mapping the preprocessed radar data to spatiotemporal coordinates along the survey line direction and the sampling time direction to establish the correspondence between each sampling channel and the spatial position coordinates, thereby obtaining a preprocessed radar image sequence.

[0022] It should be noted that zero-time correction is a process to eliminate the deviation between the radar recording start time and the actual transmission time. DC drift removal is used to eliminate baseline drift caused by low-frequency offsets in the radar system. Background removal suppresses direct and coupled waves in the horizontal direction through moving averages or singular value decomposition. Gain adjustment uses automatic gain control or an exponential gain function to compensate for energy attenuation of electromagnetic waves propagating underground. Bandpass filtering designs the passband range according to the target's depth and expected size to suppress high-frequency noise and low-frequency interference.

[0023] Understandably, spatiotemporal coordinate mapping converts the time axis into the depth axis and the sampling channel number into the horizontal position based on the survey line trajectory, the sampling point spacing, and the propagation speed of electromagnetic waves in the medium, thereby establishing a spatial coordinate index for the underground two-dimensional profile and obtaining a time-space two-dimensional radar image sequence, which is also the preprocessed radar image sequence.

[0024] In one feasible implementation, to improve imaging accuracy under complex surface conditions, the preprocessing may further include: performing terrain correction on the radar profile based on terrain elevation measurement data to eliminate image distortion caused by surface undulations; performing velocity correction on the depth axis based on dielectric constant inversion results to correct depth calculation errors caused by velocity variations in layered media; and using a wavelet transform-based time-frequency domain denoising algorithm to perform multi-scale noise separation on the radar image to obtain a radar image sequence with enhanced signal-to-noise ratio.

[0025] It should be noted that terrain correction eliminates hyperbolic distortion and depth calculation errors caused by surface undulations by geometrically registering survey elevation data with radar profiles, thus restoring the true geometric shape of targets under sloping surfaces. Velocity correction, based on the dielectric constant inversion results of layered media, uses the Dix formula or layer stripping method to correct the electromagnetic wave propagation velocity of each layer, eliminating nonlinear distortion of the depth scale caused by velocity changes. The time-frequency domain denoising algorithm based on wavelet transform utilizes the differences in the time-frequency distribution of radar signals and noise. By selecting appropriate wavelet bases and decomposition layers, it suppresses random and coherent noise while preserving the target reflection signal. The above alternative preprocessing procedures can be implemented individually or in combination to adapt to complex terrain conditions such as urban roads, mountains, and tunnel faces, thereby improving the positioning accuracy of subsequent target detection.

[0026] Step S20: Input the preprocessed radar image sequence into the ground-penetrating radar target detection model based on deep learning network for multi-scale feature extraction, temporal fusion and atomic state prediction to obtain a preliminary detection result set and the corresponding atomic state probability vector.

[0027] It should be noted that the ground-penetrating radar target detection network is a deep learning network based on the YOLOv11 network architecture and specifically optimized for the characteristics of ground-penetrating radar images.

[0028] Multi-scale feature extraction refers to extracting radar image features from different spatial resolution levels using an improved backbone network to adapt to the scale variations of underground targets, ranging from shallow, small cavities to deep, large, loose areas. Temporal fusion utilizes the temporal dependencies between consecutive radar frames to enhance the detection capability of obscured and small targets. Atomic state prediction quantifies observable low-level features in radar images into probability vectors, providing low-level input for subsequent causal inference graphs.

[0029] Understandably, the regular rectangular receptive field of a standard convolutional neural network is ill-suited to the morphological distortions caused by target tilt and size changes in ground-penetrating radar images. Furthermore, standard detection heads only output target bounding boxes and categories, lacking semantic features for integration with causal inference mechanisms. This implementation, by introducing deformable convolution, coordinate attention mechanisms, and atomic state detection heads, achieves seamless integration between the detection network and the causal inference graph.

[0030] In one feasible implementation, step S20 may include: constructing a ground-penetrating radar target detection network, wherein the ground-penetrating radar target detection network is based on the YOLOv11 network, deformable convolutions are introduced into one or more feature extraction layers of the backbone network of the YOLOv11 network, a coordinate attention module is embedded after each residual block, and dilated convolutional pyramids are constructed at different levels to perform multi-scale receptive field adjustment, the ground-penetrating radar target detection network includes a target detection head and an atomic state prediction head set in parallel; and the preprocessed radar image sequence is input into the ground-penetrating radar target detection network. The backbone network performs multi-scale feature extraction to obtain multi-level feature maps. These multi-level feature maps are then input into a bidirectional long short-term memory network for temporal feature fusion to obtain temporal fusion features and state change vectors. The temporal fusion features are then input into the target detection head and the atomic state prediction head to obtain target bounding box positions, categories, confidence levels, and atomic state probability maps. Region feature extraction is performed on the atomic state probability maps to obtain atomic state probability vectors for each target. Based on the target bounding box positions, categories, confidence levels, atomic state probability vectors, and state change vectors, the preliminary detection result set is generated.

[0031] It should be noted that YOLOv11 is a single-stage object detection network. Its standard structure includes three parts: a backbone network, a neck network, and a decoupled detection head. The backbone network consists of input convolutional layers, multiple cascaded C3k2 feature extraction modules, and a spatial pyramid pooling fast module, used to extract multi-scale features containing edge, texture, and semantic information from the input image layer by layer. The neck network adopts a path aggregation network structure, which bidirectionally fuses deep high-level semantic features with shallow high-resolution spatial features through upsampling, lateral connections, and cascaded fusion to generate three feature maps of P3, P4, and P5 suitable for multi-scale object detection. The decoupled detection head sets up classification and regression branches in parallel on the feature maps of each layer. The classification branch predicts the probability of the target class through convolutional layers, and the regression branch predicts the center coordinates, width, height, and confidence of the target bounding box through convolutional layers. Finally, the detection result is output through non-maximum suppression.

[0032] Understandably, deformable convolution is introduced into one or more feature extraction layers in the backbone network responsible for extracting mid-to-high-level semantic features. By learning additional offsets, the sampling points are adaptively distributed, effectively capturing features of irregular targets such as cavity boundaries and pipeline bends. A coordinate attention module is embedded after one or more residual blocks in the backbone network, embedding positional information into channel attention, enabling the network to more accurately locate sparse reflection signals. A multi-scale receptive field adjustment module employs differentiated receptive field strategies at different levels of the backbone network, maintaining detail resolution in shallow layers, expanding the receptive field in mid-layers, and capturing large-scale structures in deep layers, thereby adapting to the multi-scale changes of underground targets.

[0033] It is worth noting that, based on the YOLOv11 standard structure, this implementation of the ground-penetrating radar target detection network improves upon the characteristics of target morphology distortion, feature sparsity, and multi-scale variations in ground-penetrating radar images. Specifically, deformable convolutions are introduced into the mid-to-high-level feature extraction layers of the backbone network, such as layers 3 to 5, to learn adaptive offsets of sampling points to conform to irregular target morphologies such as hole boundaries and pipeline bends. A coordinate attention module is embedded after one or more residual blocks in the backbone network to embed horizontal and vertical position information into channel attention, enhancing the spatial localization capability for sparse reflection signals. Furthermore, multi-scale receptive field adjustment modules are constructed at different levels of the backbone network. A differentiated hole rate strategy is adopted to preserve details in shallow layers, expand the receptive field in middle layers, and capture large-scale structures in deep layers. The P3-level feature map output by the neck network is input into a bidirectional long short-term memory network for temporal feature fusion, and outputs temporal fusion features and state change vectors. In the detection head part, a standard target detection head and an atomic state prediction head are set in parallel. The target detection head outputs the target box position, category and confidence. The atomic state prediction head predicts the probability maps of six atomic states through six independent convolutional branches, namely local strong amplitude, in-phase axis pull-down, top and bottom reflection separation, multiple reflection, diffraction hyperbola and frequency decay, and obtains the atomic state probability vector of each target through regional feature extraction.

[0034] It is worth noting that after the radar image sequence is input into the backbone network, it passes through the feature extraction layer with deformable convolution and the residual block with embedded coordinate attention module in sequence. Through forward propagation, a multi-level feature map containing shallow detail features and deep semantic features is obtained.

[0035] A bidirectional long short-term memory (LSTM) network is placed after the intermediate feature layer in the backbone network, which balances spatial resolution and semantic information. The input is a sequence of feature maps of T consecutive frames, where T=5. The feature maps of each frame are unfolded along the spatial dimension and then fed into a bidirectional LSTM. The forward and backward outputs are concatenated to obtain the temporal fusion feature. Based on this, a state change prediction branch is added, which outputs the state change vector of the current frame relative to the previous temporal sequence.

[0036] in, and These are the last forward and reverse hidden states of the Bi-LSTM, respectively. The weight matrix is ​​a learnable matrix. For learnable bias terms, The sigmoid activation function outputs... This is the state change vector of the current frame, used to describe the trend of the target state's evolution over time. This state change vector is directly used as input to the causal inference graph, representing the magnitude of changes in atomic states, and is used to trigger counterfactual inference.

[0037] Understandably, the target detection head adopts the standard YOLO detection head structure, outputting the target bounding box position, category, and confidence score. The atomic state prediction head adds six independent convolutional layer branches to the temporal fusion features, predicting the existence probability of six atomic states respectively, including local strong amplitude, in-phase axis pull-down, top and bottom reflection separation, multiple reflections, diffraction hyperbola, and frequency attenuation. Each branch outputs an atomic state probability map of the same size as the feature map, which takes values ​​in the range [0,1] after passing through the Sigmoid activation function.

[0038] For each bounding box output by the target detection head, features of the corresponding region are extracted from the atomic state probability map using the RoI Align algorithm, and the average probability within the region is determined as the atomic state vector of the target. This vector is directly used as the underlying input of the causal inference graph, achieving seamless integration between deep detection networks and causal inference.

[0039] It is worth noting that the preliminary test results set is represented as follows:

[0040] in, This represents the i-th target detected in the t-th frame of the image, containing target category c. it Target location l it Confidence level p it Atomic state probability vector a t and state change vector Δ t Target location l it Including depth and horizontal position, category c it Covering target types such as common underground defects and pipelines, p it The initial detection confidence score output by the deep learning model. This represents the total number of targets initially detected in the t-th frame image.

[0041] It is worth noting that, to further clarify the network structure, one or more feature extraction layers can be specifically set as layers 3 to 5 of the backbone network; the multi-scale receptive field adjustment module can adopt a dilated convolutional pyramid, using a small dilation rate in shallow layers to preserve details, a medium dilation rate in middle layers to expand the receptive field, and a large dilation rate in deep layers to capture large-scale structures. The formula for adjusting the dilation rate is:

[0042] in, As a scale factor, For target depth.

[0043] Optionally, to further enhance the small target detection capability, multi-scale feature extraction may also include: introducing a cross-level feature fusion path in the feature pyramid network of the backbone network, and bidirectionally fusing deep semantic features and shallow positional features through upsampling and lateral connections; and adopting a feature selection module based on an attention mechanism to adaptively weight feature maps of different levels to enhance the feature representation of sparse small targets.

[0044] It should be noted that the cross-layer feature fusion path establishes lateral connections between deep and shallow layers of the network, upsampling the high-level semantic features of the deep layers to the same resolution as the shallow layers, and then adding or concatenating them element-wise with the detailed location features of the shallow layers, thereby preserving both the semantic discriminative power of the deep layers and the spatial localization accuracy of the shallow layers. The attention-based feature selection module adaptively enhances feature channels that strongly respond to small targets and suppresses feature channels dominated by background noise by calculating the attention weights of feature maps at different levels in the channel and spatial dimensions. This allows the network to focus on sparse small target reflection signals even in complex backgrounds. The above feature enhancement strategies can work in conjunction with deformable convolution and coordinate attention modules to form a multi-dimensional small target feature enhancement system.

[0045] Step S30: Based on the atomic state probability vector and combined with the prior knowledge of the ground penetrating radar detection scenario, construct a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer, and estimate the structural causal model parameters and conditional probability table to obtain the causal inference graph model.

[0046] It should be noted that atomic state probability vectors provide data-driven evidence for the activation of underlying nodes in the causal inference graph, while prior knowledge provides physically reasonable node definitions and logical connection constraints. The initial causal inference graph refers to a topological skeleton containing only node types, node levels, and logical connections, lacking executable reasoning capabilities. The causal inference graph model, on the other hand, is a complete probabilistic graphical model formed by adding conditional probability tables and structural causal model parameters to the initial causal inference graph, capable of performing logical reasoning, counterfactual calculations, and probability propagation.

[0047] Understandably, existing deep learning methods only establish correlations between image features and target categories, failing to answer questions like "why is this?" and "what should it be based on the logical relationships between them?" This implementation injects the atomic state probability vectors output by the detection network into the causal reasoning graph skeleton, and further assigns probability parameters and structural equations, achieving a deep fusion of data-driven features and knowledge-driven reasoning, thus making a leap from correlation recognition to causal reasoning.

[0048] Step S40: Based on the causal reasoning graph model, perform temporal consistency verification, logical consistency evaluation, and counterfactual reasoning verification on the preliminary detection result set, and perform confidence suppression or missing target completion according to the verification results to obtain the corrected detection results.

[0049] It should be noted that temporal consistency verification refers to extracting the state sequence of the same target at different times through cross-frame target association and verifying whether its state evolution satisfies the continuity constraint. Logical consistency evaluation refers to inputting the target state change sequence into the causal inference graph model and retrieving whether it conforms to the valid causal path defined in the graph structure. Counterfactual inference verification refers to inferring the expected state based on the structural causal model for detection results that do not conform to logical consistency, and generating a confidence suppression coefficient by comparing the difference between the expected state and the actual state. Missing target completion refers to virtually generating targets missed by the detection network due to occlusion or noise based on the state transition prediction of the causal inference graph model.

[0050] It is understandable that environmental interference factors, such as metallic objects and underground pipelines, often produce spurious signals on radar images that resemble the morphological characteristics of actual defects. However, their temporal evolution patterns differ fundamentally from those of actual defects. For example, the relative positions of underground pipelines remain fixed at different detection periods, which does not conform to the logical pattern of cavity size expansion. Through temporal consistency verification, logical consistency assessment, and counterfactual reasoning verification, this implementation method can effectively distinguish between real targets and interference sources.

[0051] In one feasible implementation, step S40 may include: performing cross-frame target association on the preliminary detection result set to obtain a target state change sequence; inputting the target state change sequence into the causal inference graph model for logical path retrieval to obtain a valid state change subset and an invalid state change subset; for each target in the invalid state change subset, determining the difference between the expected state and the actual detected state based on the structural causal model, and generating a confidence suppression coefficient based on the difference; attenuating the detection confidence of the corresponding target based on the confidence suppression coefficient to obtain a corrected confidence; determining a set of targets that are expected to exist but not detected based on the state transition prediction of the causal inference graph model, and generating virtual targets based on the predicted states of each target in the target set; and generating a corrected detection result based on the valid state change subset, the corrected confidence, and the virtual targets.

[0052] It should be noted that cross-frame target association establishes the correspondence between targets in adjacent frames through a position matching algorithm. Specifically, it is based on the target bounding box position. Using the Euclidean distance threshold and class consistency constraint, the target in frame t-1 is paired with the target in frame t to obtain the target state change sequence {(s)}. t-1 ,st )}, where s t This represents the state of the target at time t. This step implements temporal consistency verification, ensuring the temporal continuity of the same target.

[0053] Understandably, logical path retrieval will consider the target state change pairs (s) t-1 ,s t Input a causal inference graph model and evaluate whether it exists in the causal inference graph G from s t-1 to s t The valid causal path is determined. If a valid path exists, the target state change pair is categorized into the valid state change subset; otherwise, it is categorized into the invalid state change subset. This step achieves the aforementioned logical consistency assessment, eliminating spurious associations that do not conform to the laws of physical evolution through causal logical constraints.

[0054] It is worth noting that for targets within the subset of invalid state changes, a counterfactual reasoning mechanism is activated to predict the desired state using a structural causal model and a causal reasoning graph G:

[0055] in, Represents a causal reasoning diagram. Let be the transition probability of the current state given the previous state and the causal graph.

[0056] Define a structural causal model for counterfactual reasoning Where U is the exogenous variable, V is the endogenous variable, the endogenous variables include atomic state A, event E, state flow S, F is the structure equation, and P(U) is the distribution of exogenous variables.

[0057] All possible counterfactual paths are generated from the causal graph. A backtracking search algorithm is then used to trace back from the desired state to the event layer to determine the supporting event combinations. Then, trace back to the atomic states to determine the required set of atomic states. Comparing actual atomic states and Determine the difference set .

[0058] The degree of difference is calculated using weighted KL divergence:

[0059] in, The number of target state categories, Assigning higher weights to key states, such as holes and expansions, as category weights. Let represent the probability distribution of the i-th type of state in the actual detection state. Let represent the probability distribution of the i-th type of state in the desired state, and KL be the relative entropy calculation function.

[0060] Generate confidence suppression coefficient based on the degree of difference:

[0061] in, The suppression coefficient is set by domain experts based on the allowable error range of disease evolution. The greater the difference between the actual detection state and the expected state, the smaller the suppression coefficient and the lower the corrected detection confidence. Conversely, the smaller the difference, the closer the corrected confidence is to the original detection confidence.

[0062] The confidence level of the original detection is based on the confidence suppression coefficient. Attenuation is performed to obtain the corrected confidence level. :

[0063] For detection results with excessively large discrepancies, the corrected confidence level is directly set to zero and the result is discarded, thereby achieving counterfactual reasoning verification. The detection confidence level is corrected through causal logic to suppress false detections caused by environmental interference.

[0064] It is worth noting that the state transition prediction is based on a causal inference graph model and the context information of the current frame, including the target state of the previous frame, the probability distribution of atomic states in the current frame, and the state change vector. Forward inference determines the expected existence probability of each target in the current frame. If the expected existence probability of a target confirmed in the previous frame exceeds a preset threshold, but the corresponding detection result is not included in the preliminary detection result set, then the target is classified into the target set that is expected to exist but not detected. Based on the predicted state of each target in the target set, including predicted position, predicted category, and predicted atomic state probability vector, virtual targets are generated to maintain the continuity of target tracking and solve the problem of short-term target loss due to occlusion or noise.

[0065] The corrected detection results are generated by summarizing the following three types of targets: targets in the valid state change subset that pass the logical consistency assessment, retaining their original position, category, and confidence level; targets whose confidence level is corrected after verification through counterfactual reasoning, updating their confidence level and retaining other information; and virtual targets generated by missing target completion, assigning them a predicted state and virtual target label. The three types of results are integrated to form a corrected detection result with a format consistent with the initial detection result set, serving as input for subsequent weighted fusion and state labeling.

[0066] Optionally, to further improve the accuracy of counterfactual reasoning, counterfactual reasoning verification may also include: introducing a radar wave propagation simulator based on physical information to perform forward modeling verification of the atomic state intervention in the counterfactual path, calculating the theoretical radar response under the assumed atomic state, and evaluating the matching degree with the actual observed response; and using an uncertainty quantification method based on Monte Carlo sampling to sample the probability distribution of the state change prediction multiple times, statistically analyze the uncertainty range of the expected state, and activate confidence suppression only when the actual detected state falls outside the uncertainty range.

[0067] It should be noted that the radar wave propagation simulator based on physical information calculates the theoretical radar response profile using the finite-difference time-domain method or ray tracing method, based on the dielectric constant distribution, conductivity distribution, and geometric parameters of the underground medium. This transforms the assumed atomic states in the counterfactual path into specific medium parameter models, generating corresponding theoretical radar images. These images are then matched with actual observation images using structural similarity or normalized cross-correlation. Only when the matching degree is below a preset threshold is a significant deviation between the actual observation and the theoretical assumption confirmed, thus triggering confidence suppression. Monte Carlo uncertainty quantification generates a large number of possible state transition trajectories by repeatedly sampling noise terms and exogenous variables in the structural causal model. The probability distribution and confidence interval of the expected state are statistically analyzed. Anomalies are identified when the actual detected state falls outside the 95% confidence interval, effectively avoiding misjudgments caused by model parameter uncertainty. The aforementioned physical information verification and uncertainty quantification mechanisms can be embedded in the main process of counterfactual reasoning as an enhanced verification step for logical consistency assessment.

[0068] Step S50: The corrected detection results are weighted and fused with the preliminary detection results set and then labeled with a status to obtain the ground penetrating radar data analysis results.

[0069] It should be noted that weighted fusion and state labeling refer to integrating the detection results after counterfactual reasoning suppression with the original detection results, assigning differentiated fusion weights and structured semantic labels to each target according to the causal verification state, forming an output with interpretability and traceability; the temporal evolution archive refers to a temporal database stored according to the field structure of target unique identifier, detection time, spatial location, state category, fusion confidence and structured state label, used to record the evolution trajectory of underground targets and feed it back to the causal reasoning graph model for continuous optimization.

[0070] Understandably, existing deep learning methods only output bounding boxes and class probabilities, lacking a hierarchical description of the reliability of detection results and causal logic. This implementation introduces a three-layer differentiated fusion weight mechanism based on causal verification states, structured state labeling, and closed-loop feedback incremental learning, enabling the detection results to not only have clear credibility levels but also continuously improve detection performance through the accumulation of time-series data.

[0071] In one feasible implementation, step S50 may include: dividing the corrected detection results into a set of targets that pass verification, a set of targets with confidence suppression, and a set of targets with virtual completion based on the causal verification status of each target in the corrected detection results; assigning a first fusion weight to the set of targets that pass verification, assigning a second fusion weight to the set of targets with confidence suppression based on the confidence suppression coefficient and generating a suppression cause code, and assigning a third fusion weight to the set of targets with virtual completion based on the predicted state probability and generating a completion basis code, wherein the first fusion weight is greater than the second fusion weight, and the second fusion weight is greater than the third fusion weight; and, based on the first fusion weight, the second fusion weight, and the third fusion weight, dividing the set of targets that pass verification, the set of targets with confidence suppression, and the set of targets with virtual completion based on the causal verification status of each target in the corrected detection results into a set of targets that pass verification, a set of targets with confidence suppression, and a set of targets with virtual completion based on the predicted state probability, wherein the first fusion weight is greater than the second fusion weight, and the second fusion weight is greater than the third fusion weight; and, based on the first fusion weight, the second fusion weight, and the third fusion weight, dividing the set of targets that pass verification, the set of targets with confidence suppression, and the set of targets with virtual completion based on the predicted state probability, and generating a completion basis code. The suppression target set and the virtual completion target set are differentially weighted and fused with the preliminary detection result set to obtain the fused detection result. The fused detection result is then subjected to structured state labeling, which includes at least target state labeling, causal path tracing information, confidence source identifier, and uncertainty quantification index. The state labeling corresponding to the confidence suppression target set embeds the suppression cause code, and the state labeling corresponding to the virtual completion target set embeds the completion basis code. An interpretability analysis report is generated based on the structured state labeling, and a temporal evolution archive is constructed based on the fused detection result and the interpretability analysis report. The temporal evolution archive is fed back to the causal inference graph model to trigger incremental learning updates, thereby obtaining the ground penetrating radar data analysis result.

[0072] It should be noted that the target set for verification includes targets whose existence of a valid causal path has been confirmed through logical consistency assessment and which have not triggered confidence suppression. The target set for confidence suppression includes targets whose discrepancies exceed a preset threshold and whose confidence is attenuated after verification through counterfactual reasoning. The virtual completion target set includes virtual targets generated based on state transition predictions from a causal inference graph model. The division of the three target sets is based on the causal verification state label.

[0073] Understandably, the first, second, and third fusion weights form a decreasing weight sequence. The specific values ​​can be set according to the engineering accuracy requirements. For example, the first fusion weight could be 1.0, indicating complete trust, and the second fusion weight could be based on the corrected confidence level. The weights are mapped proportionally to the 0.3-0.6 range; the deeper the suppression, the lower the weight. The third fusion weight is mapped to the 0.1-0.3 range based on the predicted state probability; the higher the predicted probability, the higher the weight. The suppression cause encoding is generated through a preset encoding table. For example, the encoding "CA-001" represents "does not conform to the hollow evolution logic," and the encoding "CA-002" represents "abnormal temporal state jump." Completion is based on the encoding records of the previous frame target state, predicted state transition probability, and occlusion / noise judgment identifier used to generate the virtual target.

[0074] It is worth noting that the target state labels include three categories: "confirmed target," "target to be verified," and "inference target." The target state label corresponding to the confidence suppression target set is "target to be verified," while the target state label corresponding to the virtual completion target set is "inference target." Causal path tracing information records the sequence of valid causal path nodes traversed by the target when verified through the causal inference graph, such as "local strong amplitude → loosening → void formation." The confidence source identifier distinguishes whether the confidence comes from direct detection by the deep learning network, counterfactual inference correction, or virtual completion prediction. The uncertainty quantification index is determined based on Monte Carlo sampling or Bayesian posterior variance, characterizing the uncertainty range of the target state prediction. The suppression cause encoding and completion basis encoding are embedded in the corresponding target state label fields, serving as direct data sources for interpretability analysis.

[0075] It is worth noting that the interpretability analysis report is output in a structured data format, including the detection box coordinates, category, final confidence score, target state label, causal path tracing information, confidence source identifier, uncertainty quantification index, and suppression reason encoding for the confidence suppression target set and completion basis encoding for the virtual completion target set for each target. The temporal evolution archive aggregates the detection records of each time according to the unique identifier of the target, stores the target state, final confidence score, and structured state label in the fused detection results in a time series, and plots the evolution curve of the target state over time. When the cumulative amount of new data reaches a preset threshold, such as 1000 frames, or when an abnormal pattern that deviates significantly from the historical evolution pattern is detected, the temporal evolution archive is fed back to the causal inference graph model, triggering incremental learning and updating the conditional probability table and structural equation parameters, thereby achieving a closed-loop continuous improvement in detection performance.

[0076] In one feasible implementation, the method further includes: establishing a causal graph template library for typical detection scenarios, wherein the typical detection scenarios include at least road cavity detection scenarios, tunnel lining detection scenarios, metal mine detection scenarios, and archaeological exploration scenarios; acquiring new scenario detection data and extracting scenario feature descriptors from the new scenario detection data, wherein the scenario feature descriptors include at least underground medium type, target geometric feature distribution, and interference source spectral features; determining the semantic similarity between the scenario feature descriptors and the template feature descriptors of each template in the causal graph template library, and sorting the causal graph templates according to the semantic similarity ranking results. The optimal matching template is selected from the template library; based on the optimal matching template, graph structure skeleton transfer and atomic state semantic mapping are performed to obtain the initial causal graph of the new scene. The graph structure skeleton transfer preserves the general causal logic hierarchy, and the atomic state semantic mapping semantically aligns the observable features of the new scene with the atomic states of the optimal matching template, and replaces or prunes mismatched atomic states; labeled samples of the new scene are obtained, and incremental learning and parameter optimization are performed on the initial causal graph of the new scene based on the labeled samples of the new scene, updating the conditional probability table and structural equation parameters to obtain a causal inference graph model adapted to the new scene.

[0077] It should be noted that the causal graph template library is constructed by systematically organizing historical detection data and expert knowledge for various typical scenarios. Each template fully records the causal reasoning graph structure for that scenario. For example, the atomic states of the road cavity detection scenario include local strong amplitude, in-phase axis pull-down, and top-bottom reflection separation, while the events include loosening formation, cavity formation, cavity expansion, and cavity collapse. The atomic states of the tunnel lining detection scenario include steel reinforcement hyperbolic curves, strong reflection of steel arch frames, and lining void characteristics, while the events include void formation, water leakage, and crack development. The atomic states of the metal mine exploration scenario include strong metal reflection, fracture zone characteristics, and aquifer characteristics, while the events include ore body boundary identification, fault detection, and goaf location. The atomic states of the archaeological exploration scenario include tomb cavity, brick and stone structure, and disturbed soil layer, while the events include tomb passage identification, tomb location, and burial pit detection.

[0078] Understandably, the scene feature descriptor is a high-dimensional feature vector extracted from the new scene detection data through statistical and spectral analysis. The underground medium type is obtained by inverting radar wave velocity and reflection coefficient. The geometric feature distribution of the target body is described by the mean, variance, and skewness parameters of the hyperbolic morphology and layered structure in the radar profile through statistical quantification. The spectral features of interference sources are extracted by principal component analysis of the radar data's spectral components, yielding spectral fingerprints of metallic interference, pipeline interference, and ground noise. These features collectively form the basis for semantic matching between the new scene and the template library.

[0079] It is worth noting that semantic similarity is calculated using a weighted cosine similarity metric:

[0080] in, For new scene feature descriptors, For the template feature descriptor of the k-th template, is the weight coefficient of the j-th feature, which is pre-set by experts based on the contribution of each feature to the scene discrimination. D is the dimension of the scene feature descriptor. The calculated result Sim takes values ​​in the range of [0,1]. The larger the value, the higher the scene matching degree.

[0081] The templates with the highest semantic similarity are sorted in descending order and selected as the optimal matching template.

[0082] It's worth noting that the graph structure skeleton migration retains the general causal logic hierarchy of "feature co-occurrence → event triggering → state transition," meaning the three-layer topology of bottom-level atomic state nodes, middle-level AND / OR nodes, and top-level target state flow nodes remains unchanged. Atomic state semantic mapping is achieved by establishing a semantic correspondence table between observable features of the new scene and template atomic states. Directly corresponding atomic states are retained, new atomic states unique to the new scene but not present in the template are added, and states present in the template but without a physical counterpart in the new scene are pruned. For example, when migrating from a road cavity scene to a tunnel lining scene, "diffraction hyperbola" is mapped to "reinforcing steel hyperbola," "frequency attenuation" is pruned, and "strong reflection from steel arch frame" is added.

[0083] It's worth noting that the number of labeled samples for the new scene is approximately 50 to 100, labeled based on radar image features and actual verification results. Incremental learning employs a semi-supervised learning algorithm, using labeled samples to extract state change pairs between adjacent frames as positive examples, and updating the conditional probability table through frequency counting. The gradient descent method was used to optimize the causal effect coefficients in the structural causal model. , After incremental learning, the conditional probability table and structural equation parameters of the initial causal graph of the new scene are updated to match the data distribution of the new scene, and a causal inference graph model adapted to the new scene is output, realizing the rapid migration and adaptation of the detection model to the new scene.

[0084] Optionally, to further improve the automation of scene migration, selecting the optimal matching template may also include: constructing a fuzzy matching matrix between scene feature descriptors and template feature descriptors; using a soft classification method based on fuzzy C-means clustering to determine the membership vector of the new scene to each template; weighting and combining multiple templates according to the membership vector to generate a hybrid template skeleton; and then performing the graph structure skeleton migration and atomic state semantic mapping.

[0085] It should be noted that fuzzy C-means clustering calculates membership by minimizing the objective function:

[0086] in, Let be the membership degree of the new scene i to the template k, and m be the fuzzy weighting index. , where N is the cluster center of the template, N is the total number of scene samples participating in the matching, and K is the total number of existing templates in the causal graph template library.

[0087] The membership vector obtained by minimizing J is directly used as the weight of the weighted combination of multiple templates to generate a hybrid initial skeleton that adapts to the features of the new scene. Compared with single optimal template matching, this method can better adapt to complex detection tasks with multiple scene features, improve the adaptability of the initial skeleton to the new scene, reduce the number of iterations and the amount of labeled samples required for subsequent incremental learning, and reduce the human and time costs of adapting to the new scene.

[0088] When a new scene has a certain similarity to multiple templates, the hybrid template skeleton is generated by weighted averaging of the conditional probability tables and structural equation parameters of each template, avoiding information loss caused by hard classification, and enabling the transferred causal reasoning graph model to have the ability to generalize knowledge from multiple scenes.

[0089] In one feasible implementation, the method further includes: acquiring multimodal detection data, the multimodal detection data including ground-penetrating radar images, geological borehole data, and infrared thermal imaging data; extracting modality-specific features from the multimodal detection data to obtain each modality feature vector; performing cross-modal attention alignment on each modality feature vector to obtain aligned multimodal fusion features; and updating the atomic state probability vector based on the multimodal fusion features.

[0090] It should be noted that the multimodal detection data is obtained through synchronous acquisition by multiple source sensors. Among them, ground-penetrating radar images provide electromagnetic reflection characteristics of underground media, geological borehole data provide information on lithology, water content and density of local strata, and infrared thermal imaging data provide temperature distribution characteristics of the surface or lining surface. The three provide characteristics of underground targets from different physical dimensions.

[0091] Understandably, ground-penetrating radar (GPR) images extract features through the backbone network of the GPR target detection network to obtain radar feature vectors. Geological borehole data extracts borehole feature vectors through a one-dimensional convolutional network or a fully connected network. Infrared thermal imaging data extracts thermal imaging feature vectors through a convolutional neural network. The dimensions of the feature vectors from each modality are uniformly mapped to the same dimensional space to facilitate subsequent cross-modal fusion.

[0092] It is worth noting that the cross-modal attention alignment adopts an attention mechanism based on spatial location indexing, which registers the spatial coordinates of the borehole point with the pixel coordinates of the radar image, and calculates the attention weights between the borehole feature vector and the radar feature vector:

[0093] in, Let i be the feature query vector of the i-th spatial location in the radar image. Let j be the feature key vector of the j-th borehole location. This refers to the attention weight of the borehole feature on the i-th position feature of the radar image.

[0094] By using attention weighting, the geological information at the borehole point is used as an additional supervisory signal for atomic state prediction and integrated into the radar image features to obtain aligned multimodal fusion features.

[0095] It should be noted that the aligned multimodal fusion features are input into the atomic state prediction head to replace or enhance the original single-modal radar features, and the existence probability of each atomic state is recalculated. Alternatively, a decision-level fusion rule can be used to perform Bayesian fusion on the atomic state probabilities predicted separately for each modality.

[0096] in, Let be the predicted probability of the m-th mode for atomic state a. This represents the prior probability.

[0097] The fused atomic state probability vector serves as the underlying input to the causal inference graph model, improving the accuracy and robustness of atomic state recognition in complex scenarios.

[0098] Optionally, to enhance the adaptability of multimodal fusion to occluded regions, cross-modal attention alignment may also include: masking occluded regions in ground-penetrating radar images and reducing the contribution weight of pixels in occluded regions when calculating attention weights; when there is no occlusion at the corresponding location of borehole data or infrared data, increasing the weight ratio of that modal feature in the fused features to achieve adaptive dynamic allocation of modal weights.

[0099] It should be noted that the occlusion mask is automatically identified by analyzing signal continuity breaks in radar images, or obtained through manual annotation. In cross-modal attention calculation, the occlusion mask is introduced as a positional bias term into the attention scoring function:

[0100] in, Let be the occlusion mask value at position j, where 1 represents no occlusion and 0 represents complete occlusion. is the modal compensation coefficient.

[0101] When information is missing in a certain area of ​​a radar image due to occlusion, if the borehole or infrared mode provides effective information in that area, the information can be complemented by increasing the attention weight, which significantly improves the reliability of atomic state prediction under occlusion conditions.

[0102] This embodiment provides a ground-penetrating radar (GPR) data parsing method based on causal reasoning. By constructing a knowledge fence based on a causal reasoning graph, it performs logical verification and secondary constraints on the preliminary detection results of deep learning, breaking through the limitations of traditional deep learning methods that rely solely on correlation analysis. This achieves a leap from "feature-category" correlation identification to "state-evolution" causal reasoning of GPR data. Through a counterfactual reasoning mechanism, it utilizes temporal logical consistency constraints to suppress the confidence of detection results that do not conform to causal evolution laws, ensuring that even if environmental interference factors present image features similar to real diseases, they are still... It is difficult to verify through causal logic, thus effectively suppressing false detections caused by environmental interference and noise. By using temporal consistency verification and missing target completion based on state transition prediction, when target features are missing due to occlusion or sparsity, it infers the target's state based on the causal logic of previous and next frames and generates virtual targets, solving the detection failure problem caused by feature sparsity and occlusion. It realizes the leap from correlation identification to causal reasoning of ground penetrating radar data, effectively suppressing false detections caused by environmental interference, noise, and target occlusion, and significantly improving the detection accuracy and robustness of underground targets, especially diseased objects.

[0103] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S30 includes steps S301 to S306: Step S301: Define an atomic state set and an event set based on prior knowledge of the ground penetrating radar detection scenario. The atomic state set includes local strong amplitude, in-phase axis pull-down, top and bottom reflection separation, multiple reflections, diffraction hyperbola and frequency attenuation. The event set includes loosening formation, void formation, void expansion and void collapse.

[0104] It should be noted that each element in the atomic state set corresponds to a low-level feature pattern that can be directly observed or inferred in the radar image, while each element in the event set is a semantic unit formed by combining atomic states through logical nodes, used to drive changes in the top-level target state flow. For example, the "void formation" event needs to be composed of "top and bottom reflection separation" and "multiple reflections" connected by nodes, while the "void expansion" event can be composed of "diffraction wave hyperbolic shape change" or "top and bottom reflection separation range expansion" connected by OR nodes.

[0105] Step S302: Instantiate the atomic state set into nodes based on the atomic state probability vector to obtain an instantiated atomic state node layer.

[0106] It should be noted that node instantiation refers to assigning the probability values ​​of each dimension of the atomic state probability vector to the corresponding atomic state node, so that each node in the atomic state node layer carries the observation evidence output by the detection network. For example, the atomic state probability vector... The k-th element This value is assigned to the k-th atomic state node as the initial activation probability of that node in the causal inference graph. This instantiation method enables seamless data integration between the deep detection network and the causal inference graph.

[0107] Step S303: Based on the instantiated atomic state node layer and the event set, construct an initial causal reasoning graph using an AND-OR graph. The initial causal reasoning graph includes an atomic state layer, an event layer, and a target state flow layer.

[0108] It should be noted that the AND-OR graph uses instantiated atomic state nodes as the bottom layer and events in the event set as middle-layer nodes, establishing a logical mapping relationship between atomic states and events through AND and OR nodes. The target state flow layer is the top layer, representing the state change process of the target over time, such as the evolutionary path from "safe" to "loose" to "void" to "expand" to "collapse". AND nodes indicate that multiple atomic states must be satisfied simultaneously to trigger an event, and OR nodes indicate that multiple possible paths can lead to the same event. At this point, the initial causal reasoning graph only has a topological structure and does not yet contain the probabilistic parameters required for executable reasoning.

[0109] Step S304: Based on the historical detection dataset, the atomic state probability vector, and the node order constraints defined by the expert, the K2 algorithm is used to learn the graph structure of the initial causal inference graph, and the conditional probability table is estimated based on the atomic state probability vector.

[0110] It should be noted that the K2 algorithm performs a greedy search under expert-defined node order constraints, meaning atomic states precede events, and events precede the state flow, aiming to learn the optimal graph structure using Bayesian information criterion scoring. Historical detection datasets provide statistical regularities of state transitions, and atomic state probability vectors provide real-time observational evidence of node activations; both serve as inputs to the K2 algorithm and conditional probability estimation. (Conditional probability table) The strength of the causal influence of the parent node on the child node is obtained by jointly estimating the frequency count and the statistical distribution of the atomic state probability vector.

[0111] Step S305: Based on the conditional probability table, quantify the causal relationship between the atomic state layer, the event layer and the target state flow layer into a structural causal model, and estimate the parameters of the structural causal model.

[0112] It should be noted that the structural causal model quantifies causal relationships into structural equations, where the event activation probability is a weighted sum of the probabilities of the parent node's atomic states transformed by a sigmoid transform, and the target state probability is a weighted sum of the event activation probabilities of the parent nodes transformed by a softmax transform. The parameters of the structural causal model include causal effect coefficients. , and noise parameters , Its formal expression is: ;

[0113] in, The probability of the atomic state after instantiation. For the event activation probability, For the target state, , The causal effect coefficient is a learnable coefficient. , For noise terms, This refers to the Softmax function.

[0114] Step S306: Assign the conditional probability table and the structural causal model parameters to the initial causal inference graph to obtain the causal inference graph model.

[0115] It should be noted that by injecting the estimated conditional probability table and structural causal model parameters into the corresponding nodes and edges of the initial causal inference graph, the initial causal inference graph is transformed from a static topological structure into a dynamic model capable of probabilistic inference, i.e., a causal inference graph model. At this point, the causal inference graph model can accept atomic state probability vectors as input, calculate event activation probabilities and target state probabilities through probability propagation, and support counterfactual inference queries.

[0116] It is worth noting that the causal inference graph model employs incremental Bayesian updates and a forgetting mechanism for dynamic correction, enabling it to adapt to environmental changes. Incremental Bayesian updates update the conditional probability table and the parameters of the structural causal model by multiplying the likelihood of the new data with the old posterior after each batch of new data (e.g., 100 frames). The forgetting mechanism introduces a time decay factor, assigning decreasing weights to historical data to prevent long-term interference from outdated data on the model's inference accuracy. This ensures the model can gradually optimize its inference performance as geological environments change and detection data accumulates, always adapting to the statistical patterns of the current detection scenario.

[0117] In this embodiment, domain expert knowledge is embedded into the causal graph learning process in the form of node sequence constraints and AND / OR logic. Atomic state nodes are instantiated using atomic state probability vectors output by the detection network. This ensures both the physical rationality and domain interpretability of the causal reasoning graph structure and achieves a deep integration of data-driven features and knowledge-driven reasoning. By employing the K2 algorithm to learn the graph structure under expert constraints and estimating the conditional probability table based on the atomic state probability vectors, the model is quantified into a structural causal model and given an initial causal reasoning graph. This transforms the static topological skeleton into an executable causal reasoning graph model with probability propagation and counterfactual reasoning capabilities. This solves the technical problem that relying solely on deep learning detection cannot leverage domain causal knowledge to improve the interpretability of results, effectively enhancing the interpretability of ground-penetrating radar target detection results.

[0118] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the ground-penetrating radar data analysis method based on causal reasoning in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0119] This application also provides a ground-penetrating radar data analysis device based on causal reasoning; please refer to [reference needed]. Figure 2 The ground-penetrating radar data analysis device based on causal reasoning includes: The acquisition module 10 is used to acquire the raw ground-penetrating radar data of the target detection area, and to preprocess and perform spatiotemporal coordinate mapping on the raw ground-penetrating radar data to obtain a preprocessed radar image sequence.

[0120] The prediction module 20 is used to input the preprocessed radar image sequence into the ground-penetrating radar target detection model based on deep learning network for multi-scale feature extraction, temporal fusion and atomic state prediction, so as to obtain a preliminary detection result set and the corresponding atomic state probability vector.

[0121] The construction module 30 is used to construct a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer based on the atomic state probability vector and the prior knowledge of the ground penetrating radar detection scenario, and to estimate the structural causal model parameters and conditional probability table to obtain the causal inference graph model.

[0122] The correction module 40 is used to perform temporal consistency verification, logical consistency evaluation and counterfactual reasoning verification on the preliminary detection result set based on the causal reasoning graph model, and to perform confidence suppression or missing target completion based on the verification results to obtain the corrected detection results.

[0123] The labeling module 50 is used to perform weighted fusion and state labeling on the corrected detection results and the preliminary detection result set to obtain the ground penetrating radar data parsing results.

[0124] The ground-penetrating radar (GPR) data parsing device based on causal reasoning provided in this application employs the GPR data parsing method based on causal reasoning in the above embodiments, which can solve the technical problem of high false detection rate in the parsing of GPR detection data caused by environmental interference, noise, and target obstruction in the prior art. Compared with the prior art, the beneficial effects of the GPR data parsing device based on causal reasoning provided in this application are the same as those of the GPR data parsing method based on causal reasoning provided in the above embodiments, and other technical features in the GPR data parsing device based on causal reasoning are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0125] This application also proposes a ground-penetrating radar data parsing device based on causal reasoning, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the ground-penetrating radar data parsing method based on causal reasoning as described above.

[0126] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the ground-penetrating radar data parsing method based on causal reasoning as described above.

[0127] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A ground-penetrating radar data analysis method based on causal reasoning, characterized in that, The method includes: The raw ground-penetrating radar data of the target detection area is acquired, and the raw ground-penetrating radar data is preprocessed and subjected to spatiotemporal coordinate mapping to obtain a preprocessed radar image sequence. The preprocessed radar image sequence is input into a ground-penetrating radar target detection model based on a deep learning network for multi-scale feature extraction, temporal fusion, and atomic state prediction to obtain a preliminary detection result set and the corresponding atomic state probability vector. Based on the atomic state probability vector and combined with the prior knowledge of the ground penetrating radar detection scenario, a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer is constructed, and the structural causal model parameters and conditional probability table are estimated to obtain the causal inference graph model. Based on the causal reasoning graph model, the preliminary detection result set is subjected to temporal consistency verification, logical consistency evaluation, and counterfactual reasoning verification. Based on the verification results, confidence suppression or missing target completion is performed to obtain the corrected detection results. The corrected detection results are weighted and fused with the preliminary detection results, and then labeled with a status to obtain the ground penetrating radar data analysis results.

2. The method as described in claim 1, characterized in that, The process involves acquiring raw ground-penetrating radar data of the target detection area, preprocessing the raw ground-penetrating radar data, and performing spatiotemporal coordinate mapping to obtain a preprocessed radar image sequence, including: Ground-penetrating radar data is collected in the target detection area according to the preset survey lines to obtain raw radar data; The original radar data is preprocessed to obtain preprocessed radar data. The preprocessing includes at least zero-time correction, DC drift removal, background removal, gain adjustment, and bandpass filtering. The preprocessed radar data is mapped to spatiotemporal coordinates along the survey line direction and the sampling time direction to establish the correspondence between each sampling channel and the spatial position coordinates, thereby obtaining the preprocessed radar image sequence.

3. The method as described in claim 1, characterized in that, The preprocessed radar image sequence is input into a ground-penetrating radar target detection model based on a deep learning network for multi-scale feature extraction, temporal fusion, and atomic state prediction to obtain a preliminary detection result set and corresponding atomic state probability vectors, including: A ground-penetrating radar target detection network is constructed, wherein the ground-penetrating radar target detection network is based on the YOLOv11 network, deformable convolution is introduced into one or more feature extraction layers of the backbone network of the YOLOv11 network, a coordinate attention module is embedded after each residual block, and a dilated convolution pyramid is constructed at different levels to adjust the receptive field at multiple scales. The ground-penetrating radar target detection network includes a target detection head and an atomic state prediction head set in parallel. The preprocessed radar image sequence is input into the ground-penetrating radar target detection network, and multi-scale feature extraction is performed through the backbone network to obtain multi-level feature maps; The multi-level feature map is input into a bidirectional long short-term memory network for temporal feature fusion to obtain temporal fusion features and state change vectors. The temporal fusion features are input into the target detection head and the atomic state prediction head respectively to obtain the target box position, category, confidence level and atomic state probability map; Regional feature extraction is performed on the atomic state probability map to obtain the atomic state probability vector of each target; The preliminary detection result set is generated based on the target bounding box position, category, confidence level, atomic state probability vector, and state change vector.

4. The method as described in claim 1, characterized in that, Based on the atomic state probability vector and combined with prior knowledge of the ground-penetrating radar detection scenario, a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer is constructed. The structural causal model parameters and conditional probability table are estimated to obtain the causal inference graph model, including: Based on prior knowledge of the ground penetrating radar detection scenario, an atomic state set and an event set are defined. The atomic state set includes local strong amplitude, in-phase axis pull-down, top and bottom reflection separation, multiple reflections, diffraction hyperbola and frequency attenuation. The event set includes loosening formation, void formation, void expansion and void collapse. Based on the atomic state probability vector, the atomic state set is instantiated into nodes to obtain an instantiated atomic state node layer. Based on the instantiated atomic state node layer and the event set, an initial causal reasoning graph is constructed using an AND-OR graph. The initial causal reasoning graph includes an atomic state layer, an event layer, and a target state flow layer. Based on the historical detection dataset, the atomic state probability vector, and the node order constraints defined by the expert, the K2 algorithm is used to learn the graph structure of the initial causal inference graph, and the conditional probability table is estimated based on the atomic state probability vector. Based on the conditional probability table, the causal relationship between the atomic state layer, the event layer, and the target state flow layer is quantified into a structural causal model, and the parameters of the structural causal model are estimated. The conditional probability table and the structural causal model parameters are assigned to the initial causal inference graph to obtain the causal inference graph model.

5. The method as described in claim 1, characterized in that, Based on the causal reasoning graph model, the preliminary detection result set is subjected to temporal consistency verification, logical consistency evaluation, and counterfactual reasoning verification. Based on the verification results, confidence suppression or missing target completion is performed to obtain the corrected detection results, including: Cross-frame target association is performed on the preliminary detection result set to obtain the target state change sequence; The target state change sequence is input into the causal reasoning graph model for logical path retrieval to obtain a subset of valid state changes and a subset of invalid state changes; For each target in the subset of invalid state changes, the degree of difference between the expected state and the actual detected state is determined based on the structural causal model, and a confidence suppression coefficient is generated based on the degree of difference. The detection confidence of the corresponding target is attenuated based on the confidence suppression coefficient to obtain the corrected confidence. Based on the state transition prediction of the causal reasoning graph model, a set of targets that are expected to exist but have not been detected is determined, and virtual targets are generated according to the predicted state of each target in the target set. Based on the effective state change subset, the corrected confidence level, and the virtual target, a corrected detection result is generated.

6. The method as described in claim 1, characterized in that, The step of weightedly fusing and state-labeling the corrected detection results with the preliminary detection results to obtain the ground-penetrating radar data analysis results includes: Based on the causal verification status of each target in the corrected detection results, the corrected detection results are divided into a set of verified targets, a set of targets with suppressed confidence, and a set of targets with virtual completion. The first fusion weight is assigned to the verified target set, the second fusion weight is assigned to the confidence suppression target set based on the confidence suppression coefficient and a suppression cause code is generated, and the third fusion weight is assigned to the virtual completion target set based on the predicted state probability and a completion basis code is generated, wherein the first fusion weight is greater than the second fusion weight, and the second fusion weight is greater than the third fusion weight; Based on the first fusion weight, the second fusion weight, and the third fusion weight, the verification target set, the confidence suppression target set, and the virtual completion target set are differentially weighted and fused with the preliminary detection result set to obtain the fused detection result; The fusion detection results are subjected to structured state labeling. The structured state labeling includes at least target state labeling, causal path tracing information, confidence source identifier, and uncertainty quantification index. The state labeling corresponding to the confidence suppression target set is embedded with the suppression cause code, and the state labeling corresponding to the virtual completion target set is embedded with the completion basis code. An interpretability analysis report is generated based on the structured state markers, and a temporal evolution profile is constructed based on the fusion detection results and the interpretability analysis report. The temporal evolution profile is then fed back to the causal inference graph model to trigger incremental learning updates, thereby obtaining the ground penetrating radar data analysis results.

7. The method as described in claim 1, characterized in that, The method further includes: Establish a cause-effect graph template library for typical detection scenarios, which includes at least road cavity detection scenarios, tunnel lining detection scenarios, metal mine detection scenarios, and archaeological exploration scenarios; Acquire new scene detection data and extract scene feature descriptors from the new scene detection data. The scene feature descriptors include at least the type of underground medium, the geometric feature distribution of the target body, and the spectral features of the interference source. Determine the semantic similarity between the scene feature descriptor and the template feature descriptors of each template in the causal graph template library, and select the optimal matching template from the causal graph template library based on the semantic similarity ranking result; Based on the optimal matching template, graph structure skeleton transfer and atomic state semantic mapping are performed to obtain the initial causal graph of the new scene. The graph structure skeleton transfer preserves the general causal logic hierarchy, and the atomic state semantic mapping semantically aligns the observable features of the new scene with the atomic states of the optimal matching template, and replaces or prunes mismatched atomic states. Obtain labeled samples of the new scene, perform incremental learning and parameter optimization on the initial causal graph of the new scene based on the labeled samples of the new scene, update the conditional probability table and structural equation parameters, and obtain a causal inference graph model adapted to the new scene.

8. A ground-penetrating radar data analysis device based on causal reasoning, characterized in that, The device includes: The acquisition module is used to acquire the raw ground-penetrating radar data of the target detection area, and to preprocess and perform spatiotemporal coordinate mapping on the raw ground-penetrating radar data to obtain a preprocessed radar image sequence. The prediction module is used to input the preprocessed radar image sequence into a ground-penetrating radar target detection model based on a deep learning network for multi-scale feature extraction, temporal fusion and atomic state prediction, so as to obtain a preliminary detection result set and the corresponding atomic state probability vector. The construction module is used to construct a causal inference graph containing an atomic state layer, an event layer, and a target state flow layer based on the atomic state probability vector and the prior knowledge of the ground penetrating radar detection scenario, and to estimate the structural causal model parameters and conditional probability table to obtain the causal inference graph model. The correction module is used to perform temporal consistency verification, logical consistency evaluation, and counterfactual reasoning verification on the preliminary detection result set based on the causal reasoning graph model, and to perform confidence suppression or missing target completion based on the verification results to obtain the corrected detection results. The labeling module is used to perform weighted fusion and state labeling on the corrected detection results and the preliminary detection result set to obtain the ground penetrating radar data parsing results.

9. A ground-penetrating radar data analysis device based on causal reasoning, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the ground-penetrating radar data parsing method based on causal reasoning as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the ground-penetrating radar data parsing method based on causal reasoning as described in any one of claims 1 to 7.

Images