Three-dimensional geological radar AI processing and interpretation platform

Abstract

The invention relates to the technical field of geological radar signal processing, and discloses a three-dimensional geological radar AI processing interpretation platform, which comprises a geological causal structure chart construction module for obtaining a geological causal structure chart representing a causal dependency relationship between geological features and radar signals; the causal-driven adaptive filtering module is used for acquiring an adaptive filter capable of selectively retaining geological feature information; the signal geological bidirectional mapping module is used for realizing mutual conversion between signal features and geological features; the double-loop iterative optimization module is used for obtaining an optimized clean geological radar signal; a causal physical double-circulation framework is constructed, and through a geological causal graph, adaptive filtering, bidirectional mapping and iterative optimization, the geological feature recovery rate is improved in a high-noise environment, and interpretability and consistency are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of ground penetrating radar signal processing, and more specifically, to a three-dimensional ground penetrating radar AI processing and interpretation platform. Background Art

[0002] When processing ground penetrating radar signals in complex geological environments, multiple technical challenges are faced. There is a contradiction between signal fidelity and noise removal: traditional filtering methods are difficult to effectively remove noise while retaining key geological information. Especially in weak signal areas, valuable geological features are often misjudged as noise; the interpretation of ground penetrating radar signals highly depends on experts' experience, lacking objective and unified interpretation criteria, resulting in inconsistent interpretation results, and there may be significant differences in the interpretation results between different experts; Existing methods have difficulties in effectively integrating physical laws and geological knowledge into the signal processing process, resulting in the separation of the two links of signal processing and geological interpretation, and unable to form effective coordination; in extreme noise environments (signal-to-noise ratio less than 5dB), conventional methods are difficult to extract effective geological features from severely noise-polluted data; automated signal processing methods usually lack interpretability of processing decisions, making it difficult for experts to evaluate and trust the processing results. Summary of the Invention

[0003] The present invention provides a three-dimensional ground penetrating radar AI processing and interpretation platform to solve the technical problem of difficult extraction of ground penetrating radar signals in complex geological environments in related technologies.

[0004] The present invention provides a three-dimensional ground penetrating radar AI processing and interpretation platform, which is characterized by including: A geological causal structure diagram construction module that constructs a geological causal structure diagram and obtains a geological causal structure diagram representing the causal dependence relationship between geological features and radar signals based on a constraint-based causal discovery algorithm; A causality-driven adaptive filtering module that constructs a causality-driven adaptive filter and obtains an adaptive filter capable of selectively retaining geological feature information by using a causal weighted adaptive threshold algorithm; A signal-geology bidirectional mapping module that constructs a signal-geology bidirectional mapping model, establishes a bidirectional mapping relationship between the signal domain and the geology domain, and realizes the mutual conversion between signal features and geological features; A double-loop iterative optimization module that performs iterative optimization between the signal domain and the geology domain based on double-loop iterative optimization through a causality-driven dual-domain alternating optimization algorithm to obtain an optimized clean ground penetrating radar signal; A causality interpretation module that receives the geological causal structure diagram and the record of the iterative optimization process as inputs and outputs a causality interpretation report including text explanations, visualization charts, and interactive exploration components; The feedback learning module takes the above - constructed functional modules as inputs and outputs an integrated system with continuous learning ability and an ever - updated causal knowledge base.

[0005] Furthermore, the construction of the geological causal structure diagram includes: Standardize, de - trend, and detect outliers in historical ground - penetrating radar data to obtain a pre - processed data set; Extract physical constraint conditions from electromagnetic wave propagation theory and geological knowledge to form a constraint set; Apply the constraint - based PC algorithm to learn the causal graph skeleton. The PC algorithm determines the conditional independence between nodes through the conditional independence test statistic: ; where is the conditional independence test statistic, are any two nodes in the node set , is the conditional independence test statistic, is the physical constraint set, is the basic conditional independence test value, are the -th and -th nodes respectively, is the condition set, is the influence factor of the -th physical constraint on this conditional independence judgment, represents the product operation from to , is the total number of physical constraint conditions; Determine the direction of the edges in the graph based on the V - structure, acyclicity constraint, and physical prior knowledge to obtain a directed acyclic graph.

[0006] Furthermore, the construction of the causal - driven adaptive filter includes: Use wavelet transform to decompose the original ground - penetrating radar signal into coefficients of multiple frequency scales; Establish a mapping relationship between the wavelet coefficients and the signal feature nodes in the causal structure diagram; Calculate the causal protection coefficient of the signal feature nodes based on the causal structure diagram: ; where is the causal protection coefficient of the signal feature node , and this coefficient characterizes the importance of the signal feature for geological feature inference; is a subset of geological feature nodes; is a subset of signal feature nodes; is from node Causal strength to the node ; Indicates summation over all geological feature nodes ; is a set of geological feature nodes, and is the node index, is from the node to the node Causal strength; Construct an adaptive threshold function for fusion causal protection in the wavelet domain: ; where is the threshold at position of the wavelet detail coefficients of the th layer, is the base threshold of the th layer, is the causal protection strength parameter, is the causal protection coefficient corresponding to the signal feature node, is the local geological feature significance function, quantifying the possible geological feature strength at position ; Process the wavelet coefficients using an improved soft threshold function and reconstruct the filtered signal.

[0007] Furthermore, the construction of the signal-geology bidirectional mapping model includes: Construct a forward mapping model to achieve the conversion from signal features to geological features; Construct a reverse mapping model to achieve the conversion from geological features to signal features; Construct a bidirectional consistency metric function to ensure the self-consistency of the bidirectional mapping; Set a joint optimization objective to integrate the forward mapping loss, reverse mapping loss, bidirectional consistency loss, physical consistency loss, and causal consistency regularization term; Train the bidirectional mapping model through an alternating optimization strategy to achieve the mutual conversion between the signal domain and the geological domain.

[0008] Furthermore, the double-loop iterative optimization includes: Use a causality-driven adaptive filter to perform initial processing on the original signal; Use the forward mapping model to infer geological features; Use the reverse mapping model to generate the corresponding theoretical signal; Perform adaptive fusion of the filtered signal and the theoretical signal; Adjust the parameters of the adaptive filter according to the geological features and signal quality of the current iteration; Filter the original signal again using the updated filter parameters; Calculate the changes in the signal and geological features for two adjacent rounds of iteration, and stop the iteration when the convergence condition is met.

[0009] Furthermore, the causal interpretation module includes: Compare the original signal and the processed signal to identify the regions where significant changes occur; Analyze the reasons for the changes in each key change region to determine whether it is noise removal or signal enhancement; For each key change region, trace the causal path that affects the processing decision; Map the set of interpretation paths to the geophysical laws; Quantitatively evaluate the uncertainty of the interpretation results; Generate multi-level interpretation information according to user needs; Generate an intuitive visual interpretation component; Form a causal interpretation report that includes text explanations, visual charts, and interactive exploration functions.

[0010] Furthermore, the feedback learning module includes: Integrate each functional module into a unified processing platform; Build an expert feedback collection component to obtain evaluations and correction suggestions for the processing results; Structurally process and classify the expert feedback; Update the causal structure diagram based on the expert feedback; Optimize the algorithm parameters of each module based on the expert parameter feedback; Improve the interpretation criteria based on the interpretation feedback; Continuously optimize the system in an incremental learning manner; Implement a continuous monitoring and optimization mechanism for the system performance.

[0011] Furthermore, the tracing of the causal path that affects the processing decision includes: Identify the corresponding nodes in the signal feature space; Identify the relevant nodes in the geological feature space; Find all paths from the signal feature nodes to the geological feature nodes in the causal diagram; Calculate the importance score of each path: ; where represents the importance score of each path, is the path, is the edge on the path, is the corresponding causal intensity value; Select several paths with the highest scores as the set of interpretation paths.

[0012] Further, the updating of the causal structure diagram based on expert feedback includes: Updating the weight of the causal edge for the causal relationship feedback: ; wherein and are the weights of the causal edge before and after updating respectively, is the smoothing factor, is the weight recommended value of the expert for the edge ; Adding, removing or modifying causal edges to conform to expert knowledge: ; wherein represents the set of causal edges after updating, represents the set of causal edges before updating, is the set of edges to be added, is the set of edges to be removed.

[0013] A computer-readable storage medium is used to store computer-readable instructions, which can run the 3D geological radar AI processing and interpretation platform as described above when read by a computer.

[0014] The beneficial effects of the present invention are as follows: By constructing a causal physical double-loop framework, most of the key geological features can be restored under extreme conditions where the signal-to-noise ratio is less than 8 dB; Through causal relationship guidance, a balanced optimization of signal fidelity and noise removal is achieved, solving the contradiction between signal fidelity and noise removal in traditional methods; By introducing a geological causal structure and a bidirectional mapping model, the objectivity and consistency of geological interpretation are improved, reducing the dependence of the interpretation result on personal experience; By tracing the geological basis through the causal path, the interpretability of the processing result is enhanced; By continuously learning from expert feedback, adaptive learning and knowledge accumulation are achieved, improving the system performance and processing accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a module diagram of the 3D geological radar AI processing and interpretation platform in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Embodiment 1 In at least one embodiment of the present invention, a 3D geological radar AI processing and interpretation platform is disclosed, as Figure 1 shown, including: Geological causal structure diagram construction module: constructing a geological causal structure diagram; In at least one embodiment of the present invention, a geological causal structure diagram is obtained through a constraint-based causal discovery algorithm, which receives a historical ground penetrating radar dataset (including ground penetrating radar data and geological feature data), a geological expert knowledge base, and a set of physical constraint conditions as inputs, and outputs a geological causal structure diagram (the geological causal structure diagram characterizes the causal dependence relationship between geological features and radar signals).

[0017] Data preprocessing: Standardize, detrend, and detect outliers in the historical ground penetrating radar data to obtain a preprocessed dataset.

[0018] Physical constraint extraction: Extract physical constraint conditions from the electromagnetic wave propagation theory and geological knowledge to form a constraint set.

[0019] Causal skeleton learning: Apply the constraint-based PC algorithm (PCPhyK) to learn the causal graph skeleton, and the PC algorithm is executed as follows: Initialize a fully connected undirected graph: ; where is an undirected graph, is a set of nodes, including all geological features and signal features; is a set of edges, initially an empty set; For each pair of nodes and the condition set in the node set , calculate the conditional independence test statistic: ; where is the conditional independence test statistic, is a set of nodes any two nodes in, is the conditional independence test statistic, is a set of physical constraints, is the basic conditional independence test value, are respectively the th and the th nodes, is a set of conditions, is the influence factor of the th physical constraint on this conditional independence judgment, represents the product operation from to , is the total number of physical constraint conditions; When ( is the significance threshold), remove the edge between and .

[0020] Edge direction determination: The direction of the edges in the graph is determined based on the V-shaped structure, acyclic constraints and physical prior knowledge to obtain a directed acyclic graph.

[0021] Causal strength quantification: Calculate the strength of each causal edge and use the causal effect quantification indicator: ; in Represents a slave node To Node The causal strength of the node Changes to Nodes The average magnitude of the impact; Representation Node The amount of change, Representation Node The amount of change; Express expectations.

[0022] Expert knowledge fusion: Integrate the causal relationship knowledge provided by experts with the results of data-driven learning in the following ways: Determine the confidence level in causal relationships derived from data-driven methods; Assess the degree to which expert knowledge supports or opposes each causal relationship; The causal relationship obtained by data-driven learning is corrected through the weight adjustment mechanism. When there is a conflict between expert knowledge and data-driven results, the final causal relationship is determined based on the degree of certainty of expert knowledge and the strength of data support.

[0023] The geological causal structure diagram obtained through the above: ; in is a weighted directed graph. is a node set (including geological feature nodes and signal feature nodes), is a set of directed edges representing causal relationships, is the corresponding set of causal strength weights (i.e., the set of all edge weights).

[0024] The geological causal structure diagram clearly characterizes the causal dependency between geological features and radar signal characteristics, providing a knowledge basis for subsequent signal processing and interpretation.

[0025] Causal-driven adaptive filtering module: constructs a causal-driven adaptive filter; In at least one embodiment of the present invention, a method for constructing an adaptive filter for geological radar signals through a causal weighted adaptive threshold algorithm is provided; the original geological radar signal and the obtained geological causal structure map are used as input, and an adaptive filter that can selectively retain geological feature information according to geological causal relationships is output.

[0026] Multi-scale signal decomposition: Wavelet transform is used to decompose the original geological radar signal into coefficients of multiple frequency scales: ; in is the original geological radar signal, Respectively Layer approximation coefficient, Layer detail coefficient, Layer detail factor, is the number of wavelet transform layers.

[0027] Signal feature and causal node mapping: Establish the mapping relationship between wavelet coefficients and signal feature nodes in the causal structure diagram: ; in is the mapping function, is a subset of nodes in the causal structure graph that represents signal features.

[0028] Causal protection coefficient calculation: Calculate the causal protection coefficient of the signal feature node based on the causal structure diagram: ; in Signal feature node The causal protection coefficient, which represents the importance of signal characteristics to the inference of geological characteristics; is a subset of geological feature nodes; is a subset of signal feature nodes; For slave nodes To Node the causal strength of Represents all geological feature nodes sum; is a set of geological feature nodes, and is the node index, For slave nodes To Node causal strength.

[0029] Construction of causal adaptive threshold function: Construct an adaptive threshold function for fusion causal protection in the wavelet domain: ; in For the The layer wavelet detail coefficients are at positions The threshold value at is the basic threshold, is the causal protection strength parameter, is the causal protection coefficient of the corresponding signal feature node, is the local geological feature significance function, quantifying the position where the intensity of possible geological features may exist.

[0030] Adaptive soft-threshold filtering implementation: The wavelet coefficients are processed using an improved soft-threshold function: ; where is the processed wavelet coefficient, is the sign function, is the detail coefficient of the j-th layer after wavelet transform, is the threshold of the j-th layer wavelet detail coefficient at position .

[0031] Introduction of spatial correlation constraint: The geological spatial correlation constraint is introduced during the filtering process to ensure the spatial continuity of geological features: ; where is the neighborhood set of position , is the correlation coefficient between position and , is the spatial constraint intensity parameter, is the processed wavelet coefficient of the j-th layer wavelet detail coefficient at position .

[0032] Signal reconstruction: The filtered signal is reconstructed using the processed wavelet coefficients: ; where is the filtered ground penetrating radar signal, represents the inverse wavelet transform operation, , are the processed detail coefficients of the J-th layer, the processed detail coefficients of the J1-th layer, and the processed detail coefficients of the 1st layer, respectively.

[0033] The obtained causal-driven adaptive filter incorporates geological causal relationship knowledge into the filtering process, can distinguish noise and weak geological signals, selectively protects important geological features, avoids the loss of geological information caused by over-filtering, and is applicable to the processing of ground penetrating radar signals in complex geological environments with low signal-to-noise ratio.

[0034] Signal-geology bidirectional mapping module: Construct a signal-geology bidirectional mapping model; In at least one embodiment of the present invention, a signal-geology bidirectional mapping model is constructed, where the model receives a training data set (data pairs including signal features and geological features), the obtained geological causal structure diagram, and the obtained causality-driven adaptive filter as inputs, and generates a mapping model capable of bidirectional conversion between signal features and geological features as an output; Among them, the signal-geology bidirectional mapping model includes a forward mapping model and a reverse mapping model: Construct a forward mapping model (signal to geology): Model architecture definition: Construct a causality-aware deep neural network , whose structure includes: an input layer (processing filtered signal features), a feature extraction layer, a causality layer, and a geological feature prediction layer; Causal structure mapping: Map the relationships in the geological causal structure diagram to the network connections, and set the initial network weights according to the causal strength: ; Where is the initial weight of the connection nodes and in the neural network, is the causal strength from the node to the node , is the mapping function; Causal regularization term definition: Introduce a causal consistency regularization term to ensure that the model learning result is consistent with the causal structure: ; Where is the causal consistency regularization term, and are regularization strength parameters, is the edge set in the causal graph, represents the sum over all existing edges in the causal graph, represents the sum over all non-existing edges in the causal graph.

[0035] Construction of the reverse mapping model (geology to signal): Model architecture definition: Construct a neural network based on physical constraints , whose structure includes: an input layer (receiving geological features), a physical parameter inference layer, a waveform simulation layer, and a signal generation layer; Integration of physical parameters and constraints: Embed the physical model of electromagnetic wave propagation in the network to ensure that the generated signal conforms to physical laws: ; Where is the physical consistency loss metric, is the signal feature, is the geological feature, is the signal generation function based on the physical model; Geological continuity constraint: Introduce the spatial continuity constraint of the geological feature: ; where is the spatial continuity constraint of the geological feature, is the position of the set of neighborhood points, is the weight coefficient based on the spatial distance, is the position of the geological feature, is the position of the geological feature.

[0036] Definition of bidirectional consistency metric: Construct a bidirectional consistency metric function between the signal domain and the geological domain: ; where is the geological feature; is the signal feature; is the forward mapping model; is the inverse mapping model; is the bidirectional consistency metric, which ensures the self-consistency of the bidirectional mapping, that is, the signal and the geological feature should be as close as possible to the original data after the bidirectional conversion.

[0037] Setting of joint optimization objective: Integrate various indicators and set the joint optimization objective of the bidirectional mapping model: ; where represents the total loss, and are the prediction losses of the forward and inverse mappings respectively, , , , are the coefficients for weighing various items.

[0038] Model training and verification: Use the training dataset to train the model, adopting an alternating optimization strategy: Fix the parameters of and optimize the parameters of ; Fix the parameters of and optimize the parameters of ; Jointly optimize the parameters of the two models to minimize the consistency loss; Evaluate the model performance through the validation dataset to ensure that the model has sufficient generalization ability.

[0039] The constructed signal-geology bidirectional mapping model can achieve the mutual conversion between the signal domain and the geology domain. Among them, the forward mapping ( ) converts the filtered radar signal features into corresponding geological features (including geological structures, physical property parameters, etc.), and the reverse mapping ( ) converts the geological features into theoretical radar signals (including waveforms, amplitudes, phases, etc.). This model ensures the rationality and reliability of the conversion by introducing causal constraints and physical constraints, providing basic support for the subsequent double-loop iterative optimization.

[0040] Double-loop iterative optimization module: Implement the double-loop iterative optimization system; In at least one embodiment of the present invention, a method for implementing a double-loop iterative optimization system is provided. This system uses a causality-driven dual-domain alternating optimization algorithm, receives the original ground penetrating radar signal, the geological causality structure diagram, the causality-driven adaptive filter, and the signal-geology bidirectional mapping model as inputs, and outputs the optimized clean ground penetrating radar signal.

[0041] Initialization: Use the causality-driven adaptive filter to perform initial processing on the original signal: ; where is the initial processing signal obtained in the 0th round of iteration, is the original ground penetrating radar signal, is the causality-driven adaptive filter.

[0042] Forward loop (signal → geology): For the iteration round , use the forward mapping model to infer geological features: ; where is the geological feature inferred in the th round of iteration, is the forward mapping model in the signal-geology bidirectional mapping model, is the processing signal obtained in the th round of iteration.

[0043] Reverse loop (geology → signal): Use the reverse mapping model to generate the corresponding theoretical signal: ; where is the theoretical signal generated based on the geological feature , is the reverse mapping model in the signal-geology bidirectional mapping model.

[0044] Signal fusion and update: Adaptively fuse the filtered signal with the theoretical signal to obtain an updated signal: ; in is a position-dependent fusion weight function that is dynamically adjusted according to signal reliability and geological feature certainty. is the processed signal obtained in the round iteration.

[0045] Causal feedback adjustment: Adjust the parameters of the adaptive filter based on the geological characteristics and signal quality of the current iteration: ; in For the The filter threshold used in round iterations, is the change in protection factor calculated based on the current geological interpretation results, To adjust the rate parameter, For the The filter threshold used in the round iteration.

[0046] Re-filtering: Use the updated filter parameters to re-filter the original signal: ; in To refilter the result, is the original geological radar signal, For the The filter threshold used in round iterations, It is a causally driven adaptive filter.

[0047] Further integrate the re-filtering result with the fusion result: ; in For the final clean geological radar signal, For the The processed signal obtained by round iteration, For the re-filtering result, To balance the parameters, control the relative importance of the two signals.

[0048] Convergence determination and early stopping: Calculate the changes in signals and geological characteristics between two adjacent iterations, and stop the iteration when the convergence condition is met or the maximum number of iterations is reached.

[0049] Counterfactual verification: The reliability of the processing results is evaluated by making small disturbances to the geological features and observing the corresponding signal changes. When the disturbance response ratio exceeds the preset threshold (the preset threshold is the confidence level of the geological feature prediction in the signal-geology bidirectional mapping model), the corresponding area is marked as an uncertain area, indicating that further verification by experts is required.

[0050] The clean geological radar signal obtained through the above iterative optimization has a higher signal-to-noise ratio and geological interpretability, and can effectively remove noise interference while retaining key geological features. This double-loop iterative optimization system integrates signal processing and geological interpretation into a unified closed-loop optimization process, and achieves continuous improvement in processing quality through mutual verification and correction in the signal domain and geological domain. It is particularly suitable for high-noise radar data processing in complex geological environments.

[0051] Causal explanation module: build a causal explanation system; In at least one embodiment of the present invention, a method for constructing an interpretable system for processing results using a causal path tracing algorithm is provided, the system receiving original signals, optimized clean signals, geological causal structure diagrams, and iterative optimization process records as inputs, and outputting a causal interpretation report containing textual explanations, visual charts, and interactive exploration components.

[0052] Identification of key change areas: Compare the original signal and the processed signal to identify areas where significant changes have occurred: ; in, For location The signal change amplitude at For the optimized clean signal at position The value at is the original signal at position The value at position ,when ( When is the threshold, the position is marked as a key change area.

[0053] Change attribution analysis: Analyze the cause of change in each key change area to determine whether it is noise removal or signal enhancement, and classify it based on the local signal-to-noise ratio and signal change amplitude.

[0054] Causal path tracing: For each key change area, the corresponding nodes are identified in the signal feature space and the geological feature space, and the important paths connecting them are found in the causal graph, and the paths with the highest scores are selected as the explanation path set; Calculate the importance score for each path: ; in For path The importance score of For the path, is the edge on the path, is the causal intensity value corresponding to the geological causal structure diagram; Physical law mapping: Map the set of explanation paths to geophysical laws, and transform abstract causal relationships into specific geophysical explanations.

[0055] Uncertainty Quantification: Quantitative assessment of the uncertainty in the interpretation of the results: ; in Lower values ​​indicate higher certainty in the explanation. represents the summation symbol, For all possible causal paths, To explain the set of paths.

[0056] Multi-level explanation generation: Three different depths of explanations are generated based on user needs: foundational explanations (concise and concise descriptions of processing decisions), technical explanations (containing processing parameters and algorithm details), and professional explanations (containing detailed geological causal relationships and physical principles).

[0057] Visualization: Generates intuitive visualization explanation components, including signal comparison charts before and after processing, causal path highlighting, decision uncertainty heat map, and A / B comparison view.

[0058] Interactive Exploration Interface: Provides a functional interface for experts to interactively explore explanations, allowing experts to click on areas of interest to obtain detailed explanations, adjust the level of explanation detail, explore hypothetical questions, and provide feedback.

[0059] The generated causal interpretation report contains textual explanations, visualization charts, and interactive exploration components, transforming the "black box" process of signal processing into an interpretable and transparent process. By tracing the causal path, the system reveals the geophysical principles behind the processing decisions, enabling experts to understand, evaluate, and trust the processing results, thereby improving the reliability and acceptance of geological radar data interpretation.

[0060] Feedback learning module: system integration and feedback learning mechanism implementation; In at least one embodiment of the present invention, a method for implementing a system integration and feedback learning mechanism is provided. The method adopts a modular integration architecture and an incremental causal learning algorithm, takes the functional modules constructed above as input, and outputs an integrated system with continuous learning capabilities and a continuously updated causal knowledge base.

[0061] Modular system integration: Integrate various functional modules (geological causal structure diagram, causal-driven adaptive filter, signal-geology bidirectional mapping model, double-loop iterative optimization system, causal interpretation system) into a unified processing platform, and adopt a loosely coupled architecture to ensure that the modules are independent and upgradeable, including defining a unified data interface standard, building a central scheduler, implementing a message passing mechanism between modules, and establishing a unified parameter management system.

[0062] Expert feedback collection: Build an expert feedback collection component to obtain evaluations and correction suggestions for processing results, including annotation interface support, structured feedback forms, interactive parameter adjustment functions, and expert explanation and evaluation collection mechanisms.

[0063] Feedback data structuring and classification: Expert feedback (including algorithm parameter feedback, interpretation feedback, and causal relationship feedback) is structured and classified: ; in For expert feedback, For feedback about causality, For feedback on algorithm parameters, For feedback on interpretation results.

[0064] Causal knowledge update: Update the causal structure diagram based on expert feedback: Feedback on causality , update the weights of causal edges: ; in is the smoothing factor (value range 0-1), For experts The recommended weight value of and are the causal edge weights before and after the update, respectively.

[0065] Add, remove, or modify causal edges based on expert knowledge: ; in and are the edge sets before and after the update, respectively. is the edge set to be added, is the set of edges that need to be removed.

[0066] Adaptive adjustment of parameters and interpretation criteria: Based on algorithm parameter feedback and interpretation feedback, optimize the algorithm parameters and interpretation criteria of each module, and use methods such as Bayesian optimization to integrate expert experience and historical data to achieve continuous optimization of system parameters and continuous improvement of interpretation criteria.

[0067] Incremental learning and knowledge transfer: Use incremental learning to continuously optimize the system, perform incremental training on newly collected data, and achieve knowledge transfer between different geological environments, so that the system can adapt to a wider range of application scenarios.

[0068] System performance monitoring and optimization: Establish a complete performance evaluation system, including processing accuracy, efficiency, and consistency indicators, and guide the iterative upgrade of the system through regular evaluation, resource optimization, and performance analysis.

[0069] The integrated system realized through the above is a complete geological radar signal processing platform, which has the causal physics double loop processing capability and expert feedback learning capability. The system realizes the closed-loop mechanism of "processing feedback learning optimization", which continuously improves the processing capability and can adapt to the signal processing requirements in different geological environments.

[0070] A computer-readable storage medium, which is used to store computer-readable instructions. When the computer-readable instructions are read by a computer, the three-dimensional geological radar AI processing and interpretation platform as mentioned above can be run.

[0071] Here, the present invention provides an implementation example: This embodiment is applied in the construction process of a high-speed railway tunnel. The engineering geological survey team needs to accurately detect the geological conditions in front of the tunnel, especially to identify possible unfavorable geological bodies such as karst caves, faults and karst fissures, so as to prevent safety accidents such as water inrush and collapse during the construction process; The survey area is located in a complex karst development area with complex stratum structure, and is greatly disturbed by surrounding construction machinery, resulting in serious geological radar signal noise; This application scenario has typical characteristics of "high noise, multiple interferences, and complex geological structures", which poses severe challenges to geological radar signal processing. Traditional processing methods are difficult to obtain sufficiently clear geological structure images in this environment, resulting in insufficient reliability of survey results and posing potential risks to tunnel construction safety. In this tunnel engineering project, we first constructed a geological causal structure map to reveal the cause-effect relationship between radar signals and geological features in the karst area; Data collection and preprocessing: 142 sets of historical geological radar data in this area and similar geological conditions were collected, including 36 verification borehole data with known karst characteristics. These data were standardized and used for causal structure learning.

[0072] Extraction of physical constraints: Based on the electromagnetic wave propagation characteristics in the karst area, the following were extracted: the relationship between the electromagnetic wave velocity and the dielectric constant of the medium, the relationship between the interface reflection coefficient and the dielectric constant difference, the relationship between the water content in the karst area and the signal amplitude attenuation, the relationship between the fracture density and the scattering attenuation, and the relationship between the cave space and the hyperbolic reflection characteristics; Causal structure learning: Applying the PC algorithm based on physical constraints (PCPhyK) to conditional independence significance level = 0.05 for causal skeleton learning, and integrate expert knowledge to determine the edge direction. The types and numbers of nodes in the constructed geological causal structure graph are shown in Table 1: Table 1: Types and numbers of nodes in the constructed geological causal structure diagram

[0073] The constructed causal structure diagram clearly reflects the causal dependency between geological features and radar signal features in the karst area. For example, groundwater level changes affect electromagnetic wave attenuation by changing the water content of the medium; the cave space affects the hyperbolic characteristics of the signal through boundary reflection; the crack density affects the spectral characteristics of the signal through the scattering effect, etc. These causal relationships provide knowledge guidance based on physical meaning for subsequent signal processing.

[0074] Based on the constructed geological causal structure map, we developed a causal-driven adaptive filter for georadar signals in karst areas.

[0075] Wavelet decomposition and causal node mapping: The original radar signal is decomposed into 5 layers using the db5 wavelet basis to obtain 5 detail coefficient layers and 1 approximate coefficient. The mapping relationship between the wavelet coefficients and the causal graph signal nodes is shown in Table 2: Table 2: Mapping table of wavelet coefficients and causal graph signal feature nodes

[0076] Calculation of causal protection coefficient: According to the connection strength between each signal feature node and the geological feature node in the causal graph, the causal protection coefficients of different frequency bands are calculated as shown in Table 3: Table 3: Causal protection coefficients for signals in different frequency bands

[0077] Adaptive threshold function implementation: The parameters of the adaptive threshold function designed based on the causal protection coefficient are shown in Table 4: Table 4: Adaptive threshold function parameter settings where σ is the estimated noise level, obtained by estimating in the non-target area.

[0078] 2.3 Signal-geology bidirectional mapping model and double-loop iterative optimization example In the geological radar data processing of the area ahead of the tunnel, we constructed a signal-geology bidirectional mapping model for the karst environment and implemented a double-loop iterative optimization.

[0079] Bidirectional mapping model construction and training: Using the geological radar data and drilling verification data obtained in the previous tunnel project as the training set, the forward mapping (signal → geology) and reverse mapping (geology → signal) models were constructed, as shown in Table 5: Table 5: Bidirectional mapping model architecture and parameters

[0080] The training process adopts an alternating optimization strategy to minimize the following comprehensive loss function, as shown in Table 6: Table 6: Loss function weight configuration for bidirectional mapping model training

[0081] The model training adopted an early stopping strategy, and stopped training when the comprehensive performance on the validation set reached the best. The final model achieved a forward mapping geological feature prediction accuracy of 83.6% on the test set, and the average error of reverse mapping signal reconstruction was 0.14.

[0082] Double loop iterative optimization implementation: The double loop iterative optimization system is applied to the high noise radar profile in front of the tunnel. The key parameters of the iterative process are shown in Table 7: Table 7: Double loop iterative optimization process parameters and convergence records

[0083] After the sixth iteration, the changes in the signal and geological features were less than the preset convergence threshold, and the system automatically stopped iterating. The signal quality of the optimized radar profile was significantly improved, with the signal-to-noise ratio increased from the initial 8.2dB to +4.3dB, and the recognizability of key geological features was greatly improved.

[0084] Counterfactual verification: To verify the reliability of the processing results, we conducted counterfactual intervention tests on key geological features. For example, the predicted cave area size was perturbed by ±20%, and the theoretical signal was generated through reverse mapping to calculate the perturbation response ratio R. The results showed that in the predicted main cave area, the R value was 2.3 on average (much higher than the threshold of 1.0), confirming the high reliability of the processing results for the prediction of this geological feature; while in some weak reflection areas, the R value was only 0.7, which was marked as an uncertain area and required drilling verification.

[0085] To support tunnel engineering geology experts in understanding and trusting the processed results, we built a causal interpretation system for the processed radar data.

[0086] Key change areas and change attribution: The system automatically identifies and marks the key change areas in the profile and analyzes the causes of the changes as shown in Table 8: Table 8: Attribution statistics of key change areas

[0087] Example of causal path analysis: For important geological feature areas (such as the predicted main cave), the system generates a causal path analysis. Part of the causal path analysis of the cave feature at the 170m position of profile line 25 is shown in Table 9: Table 9: Key causal path analysis of cave characteristics (ranking by importance)

[0088] Multi-level interpretation generation: For each key change area, the system generates different levels of interpretation information to meet the needs of different users. An example of multi-level interpretation is shown in Table 10: Table 10: Example of multi-level interpretation (for cave characteristics at 170m on Line 25)

[0089] Visualization and interactive interface: The system provides geological experts with a variety of visual interpretation components, including before-and-after comparison diagrams, causal path visualization diagrams, uncertainty heat maps, etc., and supports interactive exploration. Through these functions, engineering geological experts can fully understand the basis for treatment decisions and enhance their trust in the treatment results.

[0090] In the actual application of this tunnel project, we verified the technical effect of this method in many aspects, focusing on verifying the following two key technical effects: In order to verify the geological feature recognition ability of this method in a high noise environment, we conducted drilling verification on the radar profiles before and after processing. The comparison of the accuracy of geological feature recognition under different noise environments is shown in Table 11: Table 11: Comparison of geological feature recognition accuracy under different noise environments

[0091] In the actual tunnel excavation process, this method successfully predicted two previously undiscovered karst caves (sizes of 4.2×3.1m and 2.8×2.3m respectively). These two karst caves were masked by noise in the traditional processing method, but were clearly visible after processing by this method and verified by subsequent drilling. This result avoids potential construction safety risks and demonstrates the practical value of this method in complex geological environments.

[0092] In order to verify the interpretability of this method, we conducted a geological expert evaluation experiment. The evaluation comparison of geological experts on the results of different processing methods is shown in Table 12: Table 12: Comparison of geological experts’ evaluation of the results of different treatment methods

[0093] The results show that compared with traditional filtering methods and pure deep learning methods, this method not only scored higher in result reliability, but also had significant advantages in explanation satisfaction and credibility of decision-making basis. In particular, the willingness of experts to make engineering decisions based on the processing results of this method reached 89.3%, which is much higher than other methods.

[0094] Expert feedback indicated that being able to understand the geophysical basis behind processing decisions was a key factor in building trust. Through causal path explanations, experts were able to relate signal processing results to their familiar geological knowledge, which greatly improved their acceptance of automated processing results.

[0095] In actual engineering applications, the interpretation system of this method provides a reliable basis for tunnel design and construction plan adjustment, helping the engineering team to make reasonable decisions quickly after discovering caves and faults, avoiding potential engineering risks, and optimizing construction progress and costs.

[0096] Embodiment 2: At least one embodiment of the present invention discloses a 3D geological radar AI processing and interpretation platform. Based on the technical solution of embodiment 1, while retaining the basic processing flow, a hierarchical collaborative scheduling method of asynchronous decision-making is introduced, including: Construct regionalized geological cause-effect structure map; Input data types: 3D geological radar original data set, geological prior knowledge base data, regional segmentation parameters; Implementation method: The 3D geological radar data is divided into regions according to the complexity of geological features; Construct a geological causal structure diagram for each region and mark the strength of causal relationships between nodes; Identify and label causal connections between regions; Specific output results: A collection of regionalized geological causal structure maps, including digital representations of causal relationships within each region and causal connections between regions.

[0097] Generate region-level causal-driven adaptive filters; Input data types: Regionalized geological radar raw data, regionalized geological causal structure map collection; Implementation method: Perform wavelet decomposition on radar data in each region to obtain multi-scale coefficients; Determine the causal protection coefficient matrix based on the causal structure diagram of each region; Construct a region-adaptive threshold function: ; in is the basic threshold, is the causal protection coefficient, is the spatial constraint coefficient, and To adjust the parameters; Specific output results: regional-level causal-driven adaptive filter parameter set, including the filtering parameters of each region and boundary filtering coordination parameters.

[0098] Construct a regional adaptive signal-geology bidirectional mapping model; Input data types: regional filtered radar dataset, regionalized geological causal structure map collection, regional characteristic description data.

[0099] Implementation method: A parameter-adaptive forward mapping model is constructed for each region to map the filtered signal into a geological feature representation; A parameter-adaptive inverse mapping model is constructed for each region to map the geological feature representation into an ideal signal; Adjust model parameters according to regional characteristics to ensure mapping consistency between regions; Specific output results: a set of regional adaptive signal geological bidirectional mapping models, including forward mapping model parameters and reverse mapping model parameters for each region.

[0100] Perform asynchronous hierarchical double loop iterative optimization; Input data types: regionalized geological radar raw data, regional level causal driven adaptive filter parameter set, regional adaptive signal geological bidirectional mapping model set; Implementation method: Data area division and task value assessment Divide the data into multiple processing areas based on the geological characteristics of the 3D radar data ,in , , They are the 1st, 2nd, and nth treatment areas respectively; Calculate the mission value index for each area: ; in For Region The task value index, For Region The causal importance index of is the feature significance index, To handle the uncertainty index, , , is the weight coefficient; Priority stratification and resource allocation: Divide the areas into three levels of priority according to the task value: ; ; ; in, is a collection of high priority areas. is a set of medium priority areas. is a collection of low priority areas. and Assign thresholds for priority; Allocate computing resources by priority: ; in , , Respectively represent the amount of computing resources allocated to the high priority area set, the medium priority area set, and the low priority area set; Asynchronous iteration and boundary state management: Determine the iteration update frequency based on priority: ; That , , Respectively represent the iterative update frequencies assigned to the high priority area set, the medium priority area set, and the low priority area set; Implement state coordination on adjacent region boundaries: ; in and The adjacent areas and The processing status, is the boundary state after coordination, is the boundary coordination factor; Regional level adaptive convergence judgment: Calculate the regional characteristic adaptive convergence threshold: ; in For Region The convergence threshold of is the benchmark convergence threshold, and To adjust the parameters, is the causal importance, is the characteristic significance; Termination condition of regional iteration: when the state change between two consecutive iterations is less than the adaptability threshold or the maximum number of iterations is reached, it is terminated; Task dependency-aware scheduling execution: Build inter-region task dependency graph and determine execution order based on topological sorting, optimizing parallel execution by considering resource availability and dependencies; Specific output results: regionalized geological radar signal data set after asynchronous optimization processing, including optimized signal data and boundary coordination data of each region.

[0101] Generate multi-level causal explanation reports; Input data types: Regionalized geological radar raw data, regionalized geological radar signal dataset after asynchronous optimization processing, regionalized geological causal structure map collection, regional priority allocation data Implementation method: Generate a separate interpretation document for each treatment area, including a description of the main geological features and treatment parameters; Generate priority decision explanation content to explain the basis for allocating priorities to each area; Generate multi-scale interpretation views, including region-level processing views, feature-level interpretation views, and signal-level interpretation views; Generate processing efficiency and quality analysis reports to compare the effects of asynchronous processing and synchronous processing; Specific output results: multi-level causal explanation report data package, including regional level explanation documents, priority decision explanation documents, multi-scale explanation view sets and processing efficiency analysis reports.

[0102] Implement a dynamic scheduler for computing resources; Input data types: task priority data, system resource status data, processing node load data, regional convergence status data; Implementation method: Collect resource usage indicators of each computing node, including CPU, memory, GPU and other resource usage; Calculate load balancing indicators to ensure that the load between nodes is relatively balanced; Dynamically allocate resources based on region priority and convergence status: ; in For Region exist The amount of resources allocated at a given time, As the basic resource allocation, For Region exist The convergence of time, Adjust parameters for resource allocation; Realize adaptive batch processing and dynamic task migration to optimize resource utilization efficiency; Collect and process performance indicator data to provide a basis for subsequent resource allocation; Specific output results: computing resource dynamic scheduling control data, including resource allocation strategy parameters, load balancing control parameters, batch optimization parameters and task scheduling sequence.

[0103] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. 3D geological radar AI processing and interpretation platform, characterized by: include: A geological causal structure diagram construction module is used to construct a geological causal structure diagram, and based on the constraint-based causal discovery algorithm, a geological causal structure diagram is obtained that represents the causal dependency relationship between geological features and radar signals; The causal-driven adaptive filtering module constructs a causal-driven adaptive filter and uses a causal-weighted adaptive threshold algorithm to obtain an adaptive filter that can selectively retain geological feature information; The signal-geology bidirectional mapping module builds a signal-geology bidirectional mapping model, establishes a bidirectional mapping relationship between the signal domain and the geological domain, and realizes the mutual conversion between signal features and geological features; The dual-loop iterative optimization module is based on dual-loop iterative optimization. It performs iterative optimization between the signal domain and the geological domain through a causal-driven dual-domain alternating optimization algorithm to obtain the optimized clean geological radar signal. The causal interpretation module receives the geological causal structure diagram and the iterative optimization process record as input, and outputs a causal interpretation report containing textual explanations, visual charts, and interactive exploration components; The feedback learning module takes the functional modules constructed above as input and outputs an integrated system with continuous learning capabilities and a continuously updated causal knowledge base.

2. The 3D geological radar AI processing and interpretation platform according to claim 1 is characterized in that: The construction of the geological causal structure diagram comprises: Standardize, detrend and detect outliers on historical geological radar data to obtain preprocessed data sets; Extract physical constraints from electromagnetic wave propagation theory and geological knowledge to form a constraint set; The causal graph skeleton is learned using a constraint-based PC algorithm that determines the conditional independence between nodes through the conditional independence test statistic: ； in is the conditional independence test statistic, For node set For any two nodes in is the conditional independence test statistic, is the set of physical constraints, is the basic conditional independence test value, Respectively and nodes, is the condition set, It is The influence factor of the physical constraint on the independence judgment of the condition, Indicates from arrive The multiplication operation of is the total number of physical constraints; Based on the V-shaped structure, acyclic constraints and physical prior knowledge, the directions of the edges in the graph are determined to obtain a directed acyclic graph.

3. The 3D geological radar AI processing and interpretation platform according to claim 1 is characterized in that: The construction of the causal driven adaptive filter comprises: Wavelet transform is used to decompose the original geological radar signal into coefficients of multiple frequency scales; Establish the mapping relationship between wavelet coefficients and signal feature nodes in the causal structure diagram; Calculate the causal protection coefficient of the signal feature node based on the causal structure graph: ； in Signal feature node The causal protection coefficient, which represents the importance of signal characteristics to the inference of geological characteristics; is a subset of geological feature nodes; is a subset of signal feature nodes; Slave nodes for computing To Node the causal strength of Represents all geological feature nodes sum; is a set of geological feature nodes, and is the node index, For slave nodes To Node the causal strength of Construct an adaptive threshold function for fusion causal protection in wavelet domain: ； in For the The layer wavelet detail coefficients are at positions The threshold value at For the The base threshold of the layer, is the causal protection strength parameter, is the causal protection coefficient of the corresponding signal characteristic node, is a local geological feature significance function, quantifying the location the strength of geological features that may be present at the location; The improved soft threshold function is used to process the wavelet coefficients and reconstruct the filtered signal.

4. The 3D geological radar AI processing and interpretation platform according to claim 1, characterized in that: The construction of the signal-geology bidirectional mapping model comprises: Construct a forward mapping model to achieve the conversion from signal features to geological features; Construct a reverse mapping model to achieve the conversion from geological features to signal features; Construct a bidirectional consistency measurement function to ensure the self-consistency of the bidirectional mapping; Set a joint optimization objective that integrates forward mapping loss, reverse mapping loss, bidirectional consistency loss, physical consistency loss, and causal consistency regularization term; The bidirectional mapping model is trained through an alternating optimization strategy to achieve mutual conversion between the signal domain and the geological domain.

5. The 3D geological radar AI processing and interpretation platform according to claim 1, characterized in that: The double loop iterative optimization includes: Initial processing of the raw signal using a causally driven adaptive filter; Inferring geological features using forward mapping models; Use the reverse mapping model to generate the corresponding theoretical signal; Adaptively fuse the filtered signal with the theoretical signal; Adjusting the parameters of the adaptive filter according to the geological characteristics and signal quality of the current iteration; Re-filter the original signal using the updated filter parameters; Calculate the changes in signals and geological characteristics between two adjacent iterations, and stop the iteration when the convergence condition is met.

6. The 3D geological radar AI processing and interpretation platform according to claim 1, characterized in that: The causal explanation module includes: Compare the original signal with the processed signal to identify areas where significant changes have occurred; Analyze the cause of change for each key change area to determine whether it is noise removal or signal enhancement; For each key area of ​​change, trace the causal paths that affect treatment decisions; Mapping the set of interpretation pathways to geophysical laws; Quantitative assessment of uncertainty in interpretation of results; Generate multi-level explanatory information based on user needs; Generate intuitive visual explanation components; Generate causal explanation reports that include textual explanations, visual charts, and interactive exploration capabilities.

7. The 3D geological radar AI processing and interpretation platform according to claim 1, characterized in that: The feedback learning module includes: Integrate various functional modules into a unified processing platform; Build an expert feedback collection component to obtain evaluations and correction suggestions for processing results; Structure and categorize expert feedback; Update the causal structure diagram based on expert feedback; Optimize algorithm parameters of each module based on expert parameter feedback; Improve interpretation criteria based on interpretation feedback; Use incremental learning to continuously optimize the system; Implement continuous monitoring and optimization mechanism of system performance.

8. The 3D geological radar AI processing and interpretation platform according to claim 6, characterized in that: The causal paths that trace back to affect the processing decision include: Identify corresponding nodes in the signal feature space; Identify relevant nodes in geological feature space; Find all paths from signal feature nodes to geological feature nodes in the causal graph; Calculate the importance score for each path: ； in represents the importance score of each path, For the path, is the edge on the path, is the corresponding causal strength value; Select several paths with the highest scores as the explanation path set.

9. The 3D geological radar AI processing and interpretation platform according to claim 7, characterized in that: The updating of the causal structure diagram based on expert feedback includes: For causal feedback, update the weight of the causal edge: ； in and are the causal edge weights before and after the update, is the smoothing factor, For experts The weight recommended value of Add, remove, or modify causal edges to match expert knowledge: ； in represents the updated causal edge set, represents the causal edge set before updating, is the edge set to be added, is the set of edges that need to be removed.

10. A computer-readable storage medium, characterized in that: It is used to store computer-readable instructions, which, when read by a computer, can run the three-dimensional geological radar AI processing and interpretation platform as described in any one of claims 1 to 9.

Images