A method and system for intelligent identification in highway engineering geological survey

Abstract

This invention relates to the technical field of highway engineering, and more particularly to an intelligent identification method for geological exploration in highway engineering. The method involves fusing borehole and radar imagery data within a fixed-width spatial window, identifying local stratigraphic structures through a structural perception network, and dividing reliable and anomalous regions based on radar reflection characteristics and boundary abrupt change intensity. For identifying unstable regions, a directional consistency guidance and structural complexity control regularization mechanism is introduced to generate reinforcing structures, solving the problem of poor structural continuity in complex strata with soft interlayers or fault zones. Finally, based on window-level structural splicing, a confidence-weighted overlapping fusion strategy is designed to output a continuous structural map of the entire segment. An engineering suggestion set is generated based on structural anomalies and regions with insufficient confidence, achieving a closed-loop system from perception to response. This is a novel geological exploration identification method with deployment value, high structural stability, and strong anomaly handling capabilities.

Description

Technical Field

[0001] This invention relates to the technical field of highway engineering, and more particularly to an intelligent identification method and system for highway engineering geological survey. Background Technology

[0002] Geological surveys for highway engineering are a core prerequisite for ensuring route site selection, design optimization, and construction safety. The main task is to identify the geological structure and potential adverse geological units along the route, including weak interlayers, slip zones, fault structures, and groundwater anomalies.

[0003] Current conventional exploration methods primarily rely on drilling and geophysical techniques, typically combining point borehole information with linear radar and other detection data indirectly to form engineering profiles. However, due to the typical characteristics of highway alignments—long distances, traversing multiple landform types, and frequent changes in geological structure—traditional "point-area inference" methods based on sparse sampling suffer from problems in practical applications, such as discontinuous structural judgments and insufficient spatial interpretability.

[0004] Meanwhile, highway surveys are often constrained by site conditions, terrain undulations, and accessibility, limiting the placement of sampling points. This leads to reliance on expert experience for structural interpretation, lacking real-time and intelligent judgment mechanisms. When dealing with complex structural areas, traditional methods often lag behind the construction progress in identifying anomalous structures such as soft interlayers, failing to provide early warnings.

[0005] Furthermore, existing methods generally lack a systematic structure identification process, cannot form a closed loop between data collection, identification, anomaly localization, and structure suggestions, and do not have the ability to spatially express and weight the credibility of the identification results.

[0006] With the development of digital exploration, there is an urgent need for a highway engineering geological identification method that has window-level spatial organization capabilities, controllable identification credibility, reinforcement of abnormal areas, and operable results, so as to support the high requirements of linear engineering for continuity, stability, and dynamic early warning capabilities. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent identification method for highway engineering geological surveys to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] On the one hand, an intelligent identification method for highway engineering geological survey is provided, including:

[0010] Using a spatial sliding window as the basic unit, the highway path is divided into multiple sliding windows at fixed intervals. The collected borehole data and radar waveform data are combined into a unified data unit. A continuous spatial interval from the starting point to the preset window width is extracted from each sliding window. The borehole points within the spatial interval are extracted to form spatial borehole data. At the same time, the radar waveform data is standardized to ensure structural consistency and energy contrast consistency, resulting in a normalized radar map.

[0011] The radar image is input into a two-dimensional convolutional network, which outputs a radar image feature map. The two-dimensional convolutional network is used to capture the interlayer texture, weak reflection structure and tilting stratigraphic trend in the radar image. The spatial borehole data is processed by structural embedding operation and mapped into a virtual label map. The borehole location is spatially interpolated and filled with lithology number, and non-blowhole areas are filled with -1. Then, the virtual label map is embedded into a single-channel tensor and spliced ​​with the radar image feature map to obtain a spliced ​​map, thus obtaining a preliminary stratigraphic identification map.

[0012] The radar atlas feature map and the preliminary stratigraphic identification map are unified to the same resolution through interpolation and downsampling, and pixel-level credibility assessment and window-level structural stability assessment are performed to obtain a set of credible regions and a set of abnormal regions.

[0013] The credible part of the preliminary stratigraphic identification map is formalized into a guide map. The guide map and radar atlas are input into the structure generation module, which outputs the reinforcement structure and confidence map generated in the anomalous area. The confidence map represents the channel value with the highest probability of model output at each point in the generated structure. The confidence map is used to evaluate the stability of the results.

[0014] Confirm the final window structure diagram, overlap and stitch all window structure diagrams according to their spatial coordinates, and introduce a confidence weighted fusion mechanism: when each pixel is predicted by multiple windows, its final output structure is the confidence weighted average result of the structure prediction values ​​of each window at that point. Based on the confidence weighted average result, extract abnormal structure areas and confidence abnormal areas as the basis for generating engineering suggestions, and output geological identification results based on the preset geological engineering knowledge base.

[0015] Preferably, the borehole data includes borehole number, coordinates, top and bottom depth of each layer, lithology code, and permeability grade, wherein 1 in the lithology code represents clay and 2 represents sand.

[0016] Preferably, the radar waveform data is collected in real time along the highway by a towed ground-penetrating radar. The signal is stored in the form of a two-dimensional map, with each scan track recording a reflected waveform signal. After collection, the radar waveform data is converted into a matrix. Where m = 512, and n is the number of channels in the sliding window.

[0017] Preferably, the two-dimensional convolutional network structure is as follows:

[0018] Layer 1: Convolutional kernel size is 5×5, number of channels is 32, activation is ReLU;

[0019] Layer 2: Convolutional kernel size is 3×3, number of channels is 64, with batch normalization;

[0020] Layer 3: The size of the dilated convolution kernel is 3×3, the dilation rate is 2, and the number of channels is 64.

[0021] Preferably, the preliminary stratigraphic identification map is a two-dimensional discrete image matrix of size m′×n′, where the value at each position in the map is a lithology code, the same as that in the borehole data.

[0022] Preferably, the pixel-level credibility assessment is used to detect the prediction credibility of local regions. For each pixel, a boundary-signal inconsistency score function is constructed. When the value of the constructed boundary-signal inconsistency score function is greater than a preset threshold, the pixel is a low-credibility boundary point.

[0023] The window-level structural stability assessment is used to identify whether the entire window is a structurally complex or suspicious region. A window-level structural complexity constraint is introduced, and a window uncertainty index based on a two-dimensional structural oscillation regularization term is adopted. The larger the index, the more violent the predicted oscillation of the window is and the inconsistent with the signal, and the lower the recognition reliability. When the index is greater than a preset threshold, the entire window is marked as an abnormal window.

[0024] Each pixel is marked as a trusted region or an abnormal region based on the boundary low-confidence points and the abnormal window.

[0025] Preferably, the structure generation module adopts a U-Net variant network, which consists of a symmetrical encoder-decoder structure. The encoder contains three convolutional blocks, each containing 3×3 convolution and ReLU normalization, and has 32-64-128 output channels. The decoder structure is symmetrical and contains a skip-connection mechanism to ensure that the spatial structure is preserved.

[0026] As a preferred approach, in order to ensure that the generated results of the reinforcement region are consistent with the structural direction of the reliable region, a contextual structure guidance term is introduced to extract the dominant stratigraphic boundary direction from the reliable region. A direction vector field is constructed through the gradient direction map to prevent structural fracture or direction reversal in the anomalous region.

[0027] Preferably, the anomalous structural zone includes soft interlayers and gravelly zones in the anomalous lithology coding.

[0028] In a second aspect of the invention, a smart identification system for highway engineering geological survey is also provided, comprising interconnected components:

[0029] The geological data standardization and integration module is used to divide the highway route into multiple sliding windows at fixed intervals, using spatial sliding windows as the basic unit. It combines the collected borehole data and radar waveform data into a unified data unit, extracts a continuous spatial interval from the starting point to the preset window width from each sliding window, extracts the borehole points within the spatial interval to form spatial borehole data, and performs standardization processing on the radar waveform data to make it have structural consistency and energy contrast consistency, thus obtaining a normalized radar map.

[0030] The preliminary stratigraphic identification map generation module is used to input radar images into a two-dimensional convolutional network and output radar image feature maps. The two-dimensional convolutional network is used to capture interlayer textures, weak reflection structures, and dipping stratigraphic trends in the radar images. Spatial borehole data is mapped into virtual label maps through structural embedding operations. The borehole locations are spatially interpolated and filled with lithology numbers, and non-blowhole areas are filled with -1. Then, the virtual label map is embedded into a single-channel tensor and spliced ​​with the radar image feature map to obtain a spliced ​​map, thus obtaining the preliminary stratigraphic identification map.

[0031] The regional assessment and division module is used to unify the radar atlas feature map and the preliminary stratigraphic identification map to the same resolution through interpolation and downsampling, and to perform pixel-level credibility assessment and window-level structural stability assessment to obtain a set of credible regions and a set of abnormal regions.

[0032] The anomaly area structure reinforcement identification module is used to formalize the credible part in the preliminary stratigraphic identification map into a guide map. The guide map and radar map are input into the structure generation module, which outputs the reinforcement structure and confidence map generated in the anomaly area. The confidence map represents the maximum probability channel value of the model output at each point in the generated structure. The confidence map is used to evaluate the stability of the results.

[0033] The full-segment structure recognition and engineering suggestion generation module is used to confirm the final window structure map. All window structure maps are overlapped and stitched according to their spatial coordinates. A confidence weighted fusion mechanism is introduced: when each pixel is predicted by multiple windows, its final output structure is the confidence weighted average result of the structure prediction values ​​of each window at that point. Based on the confidence weighted average result, abnormal structure areas and confidence abnormal areas are extracted as the basis for engineering suggestion generation. Based on the preset geological engineering knowledge base, the geological recognition result is matched and output.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This invention addresses the challenges of complex geological structures, high continuity requirements, and difficulty in identifying anomalous units in highway engineering geological exploration. It proposes a sliding window-based data organization method and a hierarchical structure identification mechanism, constructing a closed-loop intelligent identification system encompassing data aggregation, preliminary identification, confidence assessment, structural reinforcement, and suggestion generation. The invention integrates borehole and radar image data within a fixed-width spatial window, identifies local stratigraphic structures through a structural perception network, and delineates reliable and anomalous regions based on radar reflection characteristics and boundary abrupt change intensity. For identifying unstable regions, a directional consistency guidance and structural complexity control regularization mechanism are introduced to generate reinforcement structures, solving the problem of poor structural continuity in complex strata with soft interlayers or fault zones. Finally, based on window-level structural splicing, a confidence-weighted overlapping fusion strategy is designed to output a continuous structural map of the entire segment. An engineering suggestion set is generated based on structural anomalies and regions with insufficient confidence, achieving a closed-loop system from perception to response. The process proposed in this invention focuses on the operability of engineering implementation and adapts to the dual constraints of the variable geology and high reliability requirements of linear engineering. It is a new type of geological exploration and identification system with deployment value, high structural stability and strong anomaly handling capabilities. Attached Figure Description

[0036] Figure 1 This is a flowchart of a smart identification method for highway engineering geological survey in a specific embodiment of the present invention.

[0037] Figure 2 This is a framework diagram of an intelligent identification system for highway engineering geological exploration in a specific embodiment of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] On one hand, a smart identification method for highway engineering geological survey is provided, the method comprising:

[0040] S1. Using a spatial sliding window as the basic unit, the highway path is divided into multiple sliding windows at fixed intervals. The collected borehole data and radar waveform data are combined into a unified data unit. A continuous spatial interval from the starting point to the preset window width is extracted from each sliding window. The borehole points within the spatial interval are extracted to form spatial borehole data. At the same time, the radar waveform data is standardized to ensure structural consistency and energy contrast consistency, resulting in a normalized radar spectrum.

[0041] The goal of this step is to establish a standardized spatial data input structure during highway surveying, ensuring that subsequent identification models can receive continuous, uniform, and normalized geological data. Considering the inherent spatial variability of geological information, and the structural inconsistencies between traditional acquisition methods (often presenting "hole-point" and "continuous radar" data), this step proposes using a spatial sliding window as the basic unit to combine and organize borehole data and radar waveform data into a unified data unit T. i =(D i ,R i In addition, the radar data is standardized to ensure structural consistency and energy contrast consistency.

[0042] Borehole data D(x): Obtained from engineering drilling points along the highway. Borehole data typically includes XY coordinates, elevation, layer information (layer number, lithology, thickness), soil sample records, compaction values, and moisture content. The sampling depth is generally 10–30 meters, depending on the geological section. Data is exported in Excel or WITSML format, with fields including: borehole number, coordinates (UTM projection), top and bottom depths of each layer, lithology code (e.g., 1 for clay, 2 for sand), and permeability grade. Data collection frequency is low, with an average of 3–6 boreholes per kilometer.

[0043] Radar waveform data R(x): Acquired in real-time along the highway by a towed ground-penetrating radar (such as the MALAProExGPR system, frequency 500MHz). The signal is stored in the form of a two-dimensional map, with one reflected waveform signal recorded for each scan track. The sampling interval is 25cm, and the number of sampling points in each depth direction is 512 points, with a depth detection range of approximately 8–12 meters. The data format is SEG-Y or a custom binary format, and after acquisition, it is converted into a matrix. Where m = 512, and n is the number of channels in the sliding window, which is generally between 200 and 300.

[0044] This step first divides the highway path S into multiple sliding windows W at fixed intervals. i Assume the total path length is...

[0045] L total =10km, window width set to L=50m, step size ΔL=25m, then approximately 398 overlapping windows are generated, each window W i =[x i ,x i +L] indicates starting from the starting point x i The initial continuous spatial interval.

[0046] In window W i Extract: All located in [x i ,xi The borehole points within the +L] interval constitute spatial borehole data D i All radar channels j (scan point position x) j ∈W i ) constitutes R i .

[0047] Because radar has a high acquisition speed and large coverage density, but relatively few boreholes, the number of boreholes in each window is typically 0–3, while the number of radar channels can reach 200–300. To ensure data structure alignment, boreholes are spatially bound by coordinates during data preprocessing. This involves inserting their lithology labels into the corresponding positions in the radar atlas as auxiliary channels, and only boreholes that fall within a window participate in the window's identification.

[0048] The original sampling amplitude of the radar spectrum fluctuates significantly due to differences in device gain and self-attenuation. To ensure that the subsequent model can more stably identify structural boundaries in each window, it is necessary to adjust the radar spectrum R within each window. i Standardization process:

[0049]

[0050] Where: R i : Represents the radar map matrix extracted within the i-th window, with a size of m×n, where m is the number of sampling points in the depth direction and n is the number of channels within the window; μ i :R i The mean of all pixel values; σ i :R i The standard deviation of all pixel values ​​in the dataset; Φ(·): represents the standard deviation normalization operation, which transforms each pixel into a value with zero mean and unit variance; R i ′: The normalized radar spectrum, used as input for subsequent identification models.

[0051] For example, a certain road segment W 17 The CCP includes two drilling points (located at x=425m and x=468m respectively), and this section of the radar map is as follows: After normalization to a mean of 0.38 and a standard deviation of 0.12, the result is transformed into R0. 17 ′, and compared with the borehole data D of that section 17 Together they form T 17 =(D 17 ,R 17 If there is no borehole data in a certain window, only radar channel R is retained. i This window will be compensated for by adjacent structures in subsequent steps.

[0052] The normalization process is performed only within the window to avoid the disappearance of weak signals caused by overall standardization, while ensuring the consistency of the model in capturing local structural boundary features.

[0053] This step involves obtaining the window data structure T. i =(D i ,R i ′), D i Window W i All borehole data (including spatial location, lithology number, and layer thickness); R i ′: The radar map matrix after standardization.

[0054] S2. Input the radar map into a two-dimensional convolutional network and output the radar map feature map. The two-dimensional convolutional network is used to capture the interlayer texture, weak reflection structure and tilting stratigraphic trend in the radar map. The spatial borehole data is mapped into a virtual label map through a structure embedding operation. The borehole location is spatially interpolated and filled with lithology number, and non-blowhole areas are filled with -1. Then, the virtual label map is embedded into a single-channel tensor and spliced ​​with the radar map feature map to obtain a spliced ​​map, thus obtaining a preliminary stratigraphic identification map.

[0055] This step plays a crucial role in the "first structural identification" in this scheme, directly extracting data from the window-level fused data T. i =(D i ,R i Identifying spatially continuous and depth-oriented stratigraphic categories from the data is the core geological information extraction step of the entire intelligent identification system. Considering the special characteristics of highway engineering geological exploration, this step not only faces engineering challenges such as sparse data distribution (few borehole points), heterogeneous data types (map + structured data), and drastic spatial variations (nonlinear stratigraphic jumps), but also needs to ensure good interpretability and local stability of the output. Therefore, a stratigraphic identification model architecture with spatial guidance mechanism, low-density label assistance, and adaptive context reinforcement capability is proposed. It integrates the spatial perception advantages of CNN structure with the regional classification stability based on structural priors to complete the learning of the mapping relationship between radar map – borehole information – stratigraphic labels.

[0056] The standardized window data T from step one i =(D i ,R i ′). Where R i ′ represents the radar map, which is normalized to standard deviation to form a two-dimensional matrix of size m×n, where m represents the number of depth sampling points and n is the number of lateral channels, typically 512×240;

[0057] radar image R i Input a three-layer 2D convolutional network (CNN)radar The structure is as follows:

[0058] Layer 1: Convolutional kernel size is 5×5, number of channels is 32, activation is ReLU;

[0059] Layer 2: Convolutional kernel size is 3×3, number of channels is 64, with batch normalization;

[0060] Layer 3: Dilated convolution with a kernel size of 3×3, a dilation rate of 2, and 64 channels.

[0061] Output radar spectrum feature map Where m′ and n′ represent the depth and number of transverse channels after convolution, respectively. This structure is used to capture interlayer textures, weak reflection structures, and dipping stratigraphic trends in radar images.

[0062] Space Drilling Data D i After structural embedding operations, it is mapped to a "virtual tag graph". The borehole locations are spatially interpolated using lithology numbers, with non-borehole areas filled with -1. Then, L... i Embed a single-channel tensor with the radar spectral feature map F r spliced ​​into F i .

[0063] Because of D i The labels are very sparse. To prevent the model from overfitting to borehole points, a "sparse stability regularization term" is designed to penalize the stratigraphic continuity in the model output. The final prediction layer is a combination of convolution, pointwise softmax, and argmax layers, outputting the lithology prediction result for each pixel within the output window, resulting in a preliminary stratigraphic identification map.

[0064]

[0065] To enhance the model's sensitivity to discontinuous formations and soft interlayers, a specially designed regularization term is introduced during training. The formula used to penalize excessive smoothing in local areas of the prediction graph is as follows:

[0066]

[0067] in: This is a preliminary stratigraphic type map output by the model, with a size of m′×n′, and its values ​​are discrete lithology numbers; Conv 1×1 It is a 1×1 convolutional layer, and the number of output channels is equal to the number of lithology categories (usually 5 to 7); This represents the second-order Laplace gradient at the point (x,y), which measures the local rate of change in the prediction plot. The regularization term in the loss function is used to penalize "oversmoothing," i.e., over-predicting a single formation; λ is an adjustment parameter, empirically set to 0.01 or 0.05, depending on the complexity of the data.

[0068] The design of this regularization term takes into account the "small-scale rapid abrupt changes" characteristic of geological strata changes in highway scenarios (such as soft interlayers only tens of centimeters thick). Without a change perception mechanism, the model is prone to ignoring important structural boundaries, resulting in missed detections. Therefore, this paper innovatively incorporates geological "change sensitivity" into the identification model.

[0069] During training, cross-entropy loss and regularization term are used for joint optimization. During inference, only the backbone network is used for prediction, which does not increase the runtime burden.

[0070] This step yielded the preliminary stratigraphic identification map. The output is a two-dimensional discrete image matrix of size m′×n′. The value of each position in the image is a lithology code (e.g., 0 represents silt, 1 represents clay, 2 represents sand, 3 represents gravel, 4 represents sand and gravel mixed layer, etc.), which is the same as in the borehole data. This output will be passed to the next step for credibility judgment and anomaly identification.

[0071] S3. The radar map feature map and the preliminary stratigraphic identification map are unified to the same resolution through interpolation and downsampling, and pixel-level credibility assessment and window-level structural stability assessment are performed to obtain the set of credible regions and the set of abnormal regions.

[0072] This step, within the "self-assessment of the reliability of identification results" stage of the entire intelligent identification system for highway engineering geological exploration, is a crucial logical node that connects the previous step's preliminary stratigraphic structure identification results. And in combination with the original radar imagery R i Under the premise of ′, the spatial reliability of the recognition results is analyzed and judged, and the recognition area is finally divided into reliable sub-regions C. i and the anomalous sub-region A that requires subsequent structural reinforcement i This provides a precise spatial positioning basis for subsequent "structure generation and reinforcement".

[0073] In highway engineering, geological changes are often characterized by "narrow scale and high frequency." Anomalous geological units such as localized soft interlayers and fault fracture zones are spatially irregularly distributed, making them highly susceptible to misjudgment by initial identification models. Furthermore, they exhibit a dual characteristic of "weak data support + unstable identification" in data distribution. Therefore, relying solely on the model output is insufficient to fully identify these weak areas in the identification process. Based on the experience in engineering practice that "raw radar signal responses are often better than model predictions," this step constructs an anomaly identification scheme that combines signal interpretation and structural complexity assessment capabilities. It also proposes an innovative regularization term strongly bound to highway scene characteristics to identify areas with "low signal support + drastic structural changes" as anomalous regions, significantly improving the overall system's anomaly identification capability and engineering practicality.

[0074] To ensure scale alignment, this step involves adjusting the radar imagery R during operation. i ′ and preliminary structure identification results The samples are then interpolated / downsampled to achieve the same resolution.

[0075] The credibility analysis is divided into two stages: pixel-level credibility assessment and window-level structural stability assessment. The former is used to detect the prediction credibility of local regions, while the latter is used to identify whether the entire window is a structurally complex or suspicious region.

[0076] First, from Extract the boundary map B i This figure shows the locations where predicted lithological categories change, calculated using the Sobel operator or first-order difference method. For each boundary point, its reflection gradient in the radar image is extracted. Constructing the boundary-signal inconsistency score function δ i (x,y) indicates whether the boundary lacks radar signal support:

[0077]

[0078] Among them: B i (x,y) is the predicted boundary mask, and =1 indicates that the pixel is the boundary of lithological variation; This represents the radar signal reflection intensity gradient at that point, reflecting structural changes. The maximum gradient value in the graph is used for normalization; ∈ represents a very small positive number to prevent the denominator from being zero; δ i (x,y) approaching 1 indicates that the structural prediction lacks physical signal support at this point, representing a potential boundary for misjudgment.

[0079] Secondly, to prevent misidentification of geologically volatile regions as reliable areas, a window-level structural complexity constraint is introduced, employing the following window uncertainty index based on a two-dimensional structural oscillation regularization term:

[0080]

[0081] in: The Laplace gradient (second difference) at point (x,y) represents the rate of change of the structure; ω(x,y) is the "radar-prediction mismatch weight map" with a value of δ. i The weighting function of (x,y) enhances the weight of inconsistent regions; Z is a normalization factor (such as the total number of pixels or the maximum value); The larger the value, the more volatile the predicted fluctuations are and the less consistent they are with the signal, resulting in lower reliability.

[0082] The innovation of this regularization term lies in integrating structural complexity and radar support into a joint evaluation index, which is superior to the traditional judgment method that only uses image edges or entropy values. It can simultaneously capture weak recognition areas with "weak signal support + drastic local jumps".

[0083] The system sets two thresholds:

[0084] If δ i If (x,y)>θ1, then this pixel is a "low-confidence boundary point"; if The entire window will then be marked as an abnormal window.

[0085] Finally, a binary mask image M is generated. i (x,y) is used to label each point as a reliable region C. i Or abnormal area A i ,Right now:

[0086] M i (x,y) = 1 represents the reliable region; M i (x,y) = 0 indicates an outlier region;

[0087] Get C i ={(x,y)|M i (x,y)=1},A i ={(x,y)|M i (x,y)=0}.

[0088] For example, a certain window W 66 Its radar reflection spectrum shows a clear interlayer reflection interface, but the prediction results... Structural oscillation value (Higher than the engineering threshold θ2 = 0.25), while local δ 66 Pixels with (x,y)>0.85 account for more than 15% of the total area, and the entire window is classified as an abnormal region A. 66 It needs to proceed to the subsequent reinforcement and identification process.

[0089] This solution is highly practical in engineering deployment. All indicators can be automatically calculated from existing data and structural diagrams without the need for external human intervention. It also has a strong ability to identify small-scale soft interlayers and complex faults in highway engineering geological scenarios, and is a key bridge step in the whole solution from preliminary identification to stable and interpretable identification.

[0090] This step yields C. i : A set of reliable regions, representing sub-regions with high confidence in structural prediction; and A i The set of abnormal regions, which are complementary regions, represent regions where the prediction is unstable or inconsistent with the signal, and are used for structural reinforcement in the next step S4.

[0091] S4. The credible part in the preliminary stratigraphic identification map is formalized into a guide map. The guide map and radar map are input into the structure generation module, and the reinforcement structure and confidence map generated in the abnormal area are output. The confidence map represents the maximum channel value of the model output probability at each point in the generated structure. The confidence map is used to evaluate the stability of the results.

[0092] This step is the most crucial and practically applicable part of the patented solution, designed to address the anomaly region A output in the third step. i Structural reinforcement identification is performed, and finally, the continuous structural prediction result S of the "discontinuity section" in the structural diagram is generated. i Unlike ordinary image completion or classification, anomaly completion of geological structures has extremely strong spatial dependence, geological regularity constraints, and contextual consistency requirements. Especially in highway engineering scenarios, since roads linearly traverse multiple geomorphic units, anomalies such as faults, soft interlayers, and sand interlayers often occur locally and have poor spatial continuity. Under the premise of weak support from the original data and blurred boundaries, it is very easy to cause "breaks" or "abrupt changes" in the structure prediction results. To solve the above difficulties, this step proposes a regional anomaly reinforcement generation method based on reliable structure guidance, and introduces two innovative mechanisms that are highly compatible with the geological scenario: one is a contextual structure guidance term, and the other is an anomaly complexity control regularization term, so that the generated structure not only meets the data fitting requirements but also conforms to the geological evolution laws.

[0093] This step employs the "weak supervision – strong structural constraint" approach, transforming the structural identification problem of anomaly regions into a condition generation process. The reinforcement network constructs a generative completion structure with structural boundary continuity and radar response rationality as constraint objectives. The process is as follows:

[0094] First, a preliminary structural diagram will be identified. The credible part in the graph is taken as input as the "known structural region" and formalized into a guiding graph G. i ,in:

[0095] If (x,y)∈C i ,but If (x,y)∈A i Then G i (x,y)=-1 indicates that the part to be completed is missing.

[0096] G i With radar map R i The common input structure generation module employs a variant of the U-Net network and consists of a symmetric encoder-decoder structure. The encoder contains three convolutional layers, each with 3×3 convolutions, ReLU, and batch normalization, and outputs 32-64-128 channels. The decoder structure is symmetric and includes a skip-connection mechanism to ensure spatial structure preservation.

[0097] To guide the generation result S of the reinforcement region i To align with the structure orientation of the trusted region, a contextual structure guidance term is introduced. This term extracts the dominant stratigraphic boundary orientation from the trusted region, using the gradient direction map D. i Construct a direction vector field:

[0098]

[0099] in:

[0100] This indicates that in generating the structure diagram S i The local boundary gradient direction on the plane (obtained using the Sobel or Scharr operator);

[0101] The dominant orientation pattern is constructed by extrapolating the direction of the credible region boundary.

[0102] This encourages the generated structure to be aligned with the direction of the trusted region in the spatial direction, preventing structural breakage or reversal in the abnormal region.

[0103] α is a weighting term, with an empirical value between 0.1 and 0.3.

[0104] Secondly, to prevent the model from overfitting to complex textures or local pseudo-structures, an abnormal complexity regulation term is introduced during the training phase. Encourage the completion of anomalous regions to maintain interlayer smoothness in the projection perpendicular to the boundary direction:

[0105]

[0106] in: This represents the second derivative along the boundary normal direction, reflecting the rate of change of interlayer thickness; this term controls the structure in the anomalous zone from exhibiting unnatural oscillations, such as excessively thin interlayers or high-frequency pseudolayers; β is an adjustment coefficient, with typical values ​​of 0.01–0.05.

[0107] The overall training objective function is:

[0108]

[0109] in The cross-entropy main loss term is used to train and generate the result S. i It approximates manual or simulated annotations. During the inference phase, only the main network and orientation field guidance terms are retained, without increasing the computational burden.

[0110] For example, in a certain mountainous section of road, window W 81 The middle part A 81 There is a suspected fault in the area, C 81 The region structure consists of two alternating soft and hard layers with approximately horizontal boundary directions. The radar image shows R... 81 ′ in A 81 The location presents a low-reflectivity area. This method utilizes C 81 interlayer direction Guided by the map, a continuous but slightly concave reinforcement structure S was generated. 81 This result accurately located the mudstone interlayer during actual drilling verification.

[0111] This step yields S i ′: In abnormal region A i The internally generated reinforcement structure identification results, and with Consistent structure; confidence plot Λ i : Represents the channel value with the highest probability of model output at each point in the generated structure, which can be used to evaluate the stability of the results.

[0112] S5. Confirm the final window structure diagram, overlap and stitch all window structure diagrams according to their spatial coordinates, and introduce a confidence-weighted fusion mechanism: when each pixel is predicted by multiple windows, its final output structure is the confidence-weighted average result of the predicted structure values ​​of each window at that point. Based on the confidence-weighted average result, extract abnormal structure areas and confidence-abnormal areas as the basis for generating engineering suggestions, and match and output geological identification results based on the preset geological engineering knowledge base.

[0113] This step, as the final stage of the invention, has the core objective of transforming the aforementioned sliding windows W... i The completed structural identification results (including preliminary structure) Abnormal reinforcement result S i ′ and confidence plot Λ i The entire segment is stitched together to output a complete recognition map S that is spatially continuous, has smooth boundaries, and has a reliable structure. Based on areas with structural anomalies and confidence anomalies, and combined with the construction risk logic of the highway engineering scenario, an engineering suggestion set R is automatically generated, such as suggestions for densifying drilling point locations, pre-reinforcement area ranges, and construction path adjustment suggestions.

[0114] In actual engineering surveys, data acquisition exhibits significant non-uniformity, with micro-scale misalignment between windows, particularly at the edges of abnormal reinforcement areas where continuity breaks occur. Without systematic fusion, this directly leads to discontinuous structural diagrams, reduced recognition accuracy, and inability to be used for subsequent design. Therefore, this step focuses on addressing the following three issues: (1) the boundary fusion problem of window structural diagrams; (2) the spatial integration of recognition confidence; and (3) the design of a reasoning mechanism to map recognition results to engineering recommendations.

[0115] Final window structure output It is defined as: in the abnormal region A i Take S from the inside i ′, the remaining areas are taken Right now:

[0116]

[0117] Where A i (x, y) is the anomalous region mask from step three, with a value of 0 or 1, indicating whether the point enters the structural reinforcement. This stitching method ensures that each window outputs a uniform structural image, without requiring the model to participate in additional inference, and all anomalous regions are completed by the model.

[0118] Then, all window structure diagrams According to its spatial coordinates {x i Overlapping and stitching are performed. To address the potential structural inconsistencies in overlapping areas of sliding windows, this scheme proposes a confidence-weighted fusion mechanism: when each pixel (x,y) is predicted by multiple windows, its final output structure is the confidence-weighted average of the predicted structure values ​​for that point from each window.

[0119]

[0120] in: Indicates whether the pixel is in the i-th window; This represents the softmax confidence of the window at this pixel for class k; the softmax response values ​​of all windows that appear at this point are summed by class, and the class corresponding to the maximum value is taken as the final structure class;

[0121] This mechanism significantly improves the consistency of overlapping areas, reduces splicing boundary errors, and prioritizes the use of structural judgment results in windows with high confidence.

[0122] After the structural diagrams are assembled, the system automatically extracts the following two types of sensitive engineering areas as the basis for generating engineering suggestions:

[0123] Abnormal structural regions It belongs to the abnormal lithology code, such as soft interlayers, gravel zones, etc.;

[0124] Confidence outlier regions This indicates the region of uncertainty in structural judgment.

[0125] Combining these two regions, a suggestion set R is generated through rule-based reasoning, such as:

[0126] like The continuous area exceeds 20m 2 It is recommended to use encrypted drilling; if If the location coincides with the slope, it indicates a risk of slope instability, and it is recommended to strengthen the support structure; if the continuous distribution of abnormal lithology is longer than 100m, it is recommended to adjust the route or change the construction process.

[0127] These rules and logics are provided by the geological engineering knowledge base, and the system uses a template-based rule matching module to implement them, without the need for additional model reasoning.

[0128] In a second aspect of the invention, a smart identification system for highway engineering geological survey is also provided, comprising interconnected components:

[0129] The geological data standardization and integration module is used to divide the highway route into multiple sliding windows at fixed intervals, using spatial sliding windows as the basic unit. It combines the collected borehole data and radar waveform data into a unified data unit, extracts a continuous spatial interval from the starting point to the preset window width from each sliding window, extracts the borehole points within the spatial interval to form spatial borehole data, and performs standardization processing on the radar waveform data to make it have structural consistency and energy contrast consistency, thus obtaining a normalized radar map.

[0130] The preliminary stratigraphic identification map generation module is used to input radar images into a two-dimensional convolutional network and output radar image feature maps. The two-dimensional convolutional network is used to capture interlayer textures, weak reflection structures, and dipping stratigraphic trends in the radar images. Spatial borehole data is mapped into virtual label maps through structural embedding operations. The borehole locations are spatially interpolated and filled with lithology numbers, and non-blowhole areas are filled with -1. Then, the virtual label map is embedded into a single-channel tensor and spliced ​​with the radar image feature map to obtain a spliced ​​map, thus obtaining the preliminary stratigraphic identification map.

[0131] The regional assessment and division module is used to unify the radar atlas feature map and the preliminary stratigraphic identification map to the same resolution through interpolation and downsampling, and to perform pixel-level credibility assessment and window-level structural stability assessment to obtain a set of credible regions and a set of abnormal regions.

[0132] The anomaly area structure reinforcement identification module is used to formalize the credible part in the preliminary stratigraphic identification map into a guide map. The guide map and radar map are input into the structure generation module, which outputs the reinforcement structure and confidence map generated in the anomaly area. The confidence map represents the maximum probability channel value of the model output at each point in the generated structure. The confidence map is used to evaluate the stability of the results.

[0133] The full-segment structure recognition and engineering suggestion generation module is used to confirm the final window structure map. All window structure maps are overlapped and stitched according to their spatial coordinates. A confidence weighted fusion mechanism is introduced: when each pixel is predicted by multiple windows, its final output structure is the confidence weighted average result of the structure prediction values ​​of each window at that point. Based on the confidence weighted average result, abnormal structure areas and confidence abnormal areas are extracted as the basis for engineering suggestion generation. Based on the preset geological engineering knowledge base, the geological recognition result is matched and output.

[0134] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent identification in highway engineering geological survey, characterized in that, include: Using a spatial sliding window as the basic unit, the highway path is divided into multiple sliding windows at fixed intervals. The collected borehole data and radar waveform data are combined into a unified data unit. A continuous spatial interval from the starting point to the preset window width is extracted from each sliding window. The borehole points within the spatial interval are extracted to form spatial borehole data. At the same time, the radar waveform data is standardized to ensure structural consistency and energy contrast consistency, resulting in a normalized radar map. The radar image is input into a two-dimensional convolutional network, which outputs a radar image feature map. The two-dimensional convolutional network is used to capture the interlayer texture, weak reflection structure and tilting stratigraphic trend in the radar image. The spatial borehole data is processed by structural embedding operation and mapped into a virtual label map. The borehole location is spatially interpolated and filled with lithology number, and non-blowhole areas are filled with -1. Then, the virtual label map is embedded into a single-channel tensor and spliced ​​with the radar image feature map to obtain a spliced ​​map, thus obtaining a preliminary stratigraphic identification map. The radar atlas feature map and the preliminary stratigraphic identification map are unified to the same resolution through interpolation and downsampling, and pixel-level credibility assessment and window-level structural stability assessment are performed to obtain a set of credible regions and a set of abnormal regions. The pixel-level reliability assessment is used to detect the prediction reliability of local regions. For each pixel, a boundary-signal inconsistency score function is constructed. When the value of the constructed boundary-signal inconsistency score function is greater than a preset threshold, the pixel is a low-reliability boundary point. The constructed boundary-signal inconsistency score function is expressed as follows: ； in, For boundary-signal inconsistency scoring function, To predict the boundary mask, This indicates that the pixel represents a boundary of lithological variation. This represents the radar signal reflection intensity gradient at that point, reflecting structural changes. Normalize the maximum gradient value in the graph; The value should be a very small positive number to prevent the denominator from being zero. A value close to 1 indicates that the structural prediction lacks physical signal support at this point, representing a potential boundary for misjudgment. The window-level structural stability assessment is used to identify whether the entire window is a structurally complex or suspicious region. It introduces window-level structural complexity constraints and adopts a window uncertainty index based on a two-dimensional structural oscillation regularization term. The larger the window uncertainty index, the more violent the window prediction oscillation is and the inconsistent with the signal, resulting in lower recognition reliability. When the window uncertainty index is greater than a preset threshold, the entire window is marked as an abnormal window. The window uncertainty index based on the two-dimensional structural oscillation regularization term is expressed as follows: ； in: Indicates in The Laplace gradient at a point reflects the rate of change of the structure. For radar – prediction mismatch weight map, the value is... The weighting function enhances the weighting of inconsistent regions; Normalization factor; This is an indicator of window uncertainty. Each pixel is marked as a trusted region or an abnormal region based on the boundary low-confidence points and the abnormal window. The credible part of the preliminary stratigraphic identification map is formalized into a guide map. The guide map and radar atlas are input into the structure generation module, which outputs the reinforcement structure and confidence map generated in the anomalous area. The confidence map represents the channel value with the highest probability of model output at each point in the generated structure. The confidence map is used to evaluate the stability of the results. The final window structure diagram is confirmed, and all window structure diagrams are overlapped and stitched together according to their spatial coordinates. A confidence-weighted fusion mechanism is introduced: when each pixel is predicted by multiple windows, its final output structure is the confidence-weighted average result of the predicted structure values ​​of each window pixel. Based on the confidence-weighted average result, abnormal structure areas and confidence-abnormal areas are extracted as the basis for generating engineering suggestions. Based on the preset geological engineering knowledge base, the geological identification results are matched and output.

2. The intelligent identification method for highway engineering geological exploration according to claim 1, characterized in that, The borehole data includes borehole number, coordinates, top and bottom depth of each layer, lithology code, and permeability grade, where 1 in the lithology code represents clay and 2 represents sand.

3. The intelligent identification method for highway engineering geological exploration according to claim 1, characterized in that, The radar waveform data was collected in real time by a towed ground-penetrating radar along the highway. The signals were stored in the form of a two-dimensional map, with each scan track recording a reflected waveform signal. After collection, the radar waveform data was converted into a matrix. ,in , This refers to the number of channels within the sliding window.

4. The intelligent identification method for highway engineering geological exploration according to claim 1, characterized in that, The structure of the two-dimensional convolutional network is as follows: Layer 1: Kernel size is The number of channels is 32, and it is activated as ReLU; Layer 2: Kernel size is The number of channels is 64, with batch normalization included; Layer 3: The size of the dilated convolution kernel is... The void ratio is 2, and the number of channels is 64.

5. The intelligent identification method for highway engineering geological exploration according to claim 1, characterized in that, The structure generation module employs a variant of the U-Net network, consisting of a symmetrical encoder-decoder structure. The encoder contains three convolutional layers, each layer containing... Convolution and ReLU normalization are used, with 32-64-128 output channels. The decoder structure is symmetrical and includes a skip-connection mechanism to ensure that the spatial structure is preserved.

6. The intelligent identification method for highway engineering geological exploration according to claim 1, characterized in that, To ensure that the generated reinforcement region is consistent with the structural orientation of the reliable region, a contextual structure guide term is introduced to extract the dominant stratigraphic boundary orientation from the reliable region. A direction vector field is constructed through a gradient direction map to prevent structural fracture or orientation reversal in the anomalous region.

7. The intelligent identification method for highway engineering geological exploration according to claim 1, characterized in that, The anomalous structural zone includes soft interlayers and gravelly zones in the anomalous lithology coding.

8. A smart identification system for highway engineering geological survey, characterized in that, Including interconnected: The geological data standardization and integration module is used to divide the highway route into multiple sliding windows at fixed intervals, using spatial sliding windows as the basic unit. It combines the collected borehole data and radar waveform data into a unified data unit, extracts a continuous spatial interval from the starting point to the preset window width from each sliding window, extracts the borehole points within the spatial interval to form spatial borehole data, and performs standardization processing on the radar waveform data to make it have structural consistency and energy contrast consistency, thus obtaining a normalized radar map. The preliminary stratigraphic identification map generation module is used to input radar images into a two-dimensional convolutional network and output radar image feature maps. The two-dimensional convolutional network is used to capture interlayer textures, weak reflection structures, and dipping stratigraphic trends in the radar images. Spatial borehole data is mapped into virtual label maps through structural embedding operations. The borehole locations are spatially interpolated and filled with lithology numbers, and non-blowhole areas are filled with -1. Then, the virtual label map is embedded into a single-channel tensor and spliced ​​with the radar image feature map to obtain a spliced ​​map, thus obtaining the preliminary stratigraphic identification map. The regional assessment and division module is used to unify the radar atlas feature map and the preliminary stratigraphic identification map to the same resolution through interpolation and downsampling, and to perform pixel-level credibility assessment and window-level structural stability assessment to obtain a set of credible regions and a set of abnormal regions. The pixel-level reliability assessment is used to detect the prediction reliability of local regions. For each pixel, a boundary-signal inconsistency score function is constructed. When the value of the constructed boundary-signal inconsistency score function is greater than a preset threshold, the pixel is a low-reliability boundary point. The constructed boundary-signal inconsistency score function is expressed as follows: ； in, For boundary-signal inconsistency scoring function, To predict the boundary mask, This indicates that the pixel represents a boundary of lithological variation. This represents the radar signal reflection intensity gradient at that point, reflecting structural changes. Normalize the maximum gradient value in the graph; The value should be a very small positive number to prevent the denominator from being zero. A value close to 1 indicates that the structural prediction lacks physical signal support at this point, representing a potential boundary for misjudgment. The window-level structural stability assessment is used to identify whether the entire window is a structurally complex or suspicious region. It introduces window-level structural complexity constraints and adopts a window uncertainty index based on a two-dimensional structural oscillation regularization term. The larger the window uncertainty index, the more violent the window prediction oscillation is and the inconsistent with the signal, resulting in lower recognition reliability. When the window uncertainty index is greater than a preset threshold, the entire window is marked as an abnormal window. The window uncertainty index based on the two-dimensional structural oscillation regularization term is expressed as follows: ； in: Indicates in The Laplace gradient at a point reflects the rate of change of the structure. For radar – prediction mismatch weight map, the value is... The weighting function enhances the weighting of inconsistent regions; Normalization factor; This is an indicator of window uncertainty. Each pixel is marked as a trusted region or an abnormal region based on the boundary low-confidence points and the abnormal window. The anomaly area structure reinforcement identification module is used to formalize the credible part in the preliminary stratigraphic identification map into a guide map. The guide map and radar map are input into the structure generation module, which outputs the reinforcement structure and confidence map generated in the anomaly area. The confidence map represents the maximum probability channel value of the model output at each point in the generated structure. The confidence map is used to evaluate the stability of the results. The full-segment structure recognition and engineering suggestion generation module is used to confirm the final window structure map. All window structure maps are overlapped and stitched according to their spatial coordinates. A confidence-weighted fusion mechanism is introduced: when each pixel is predicted by multiple windows, its final output structure is the confidence-weighted average result of the structure prediction values ​​of each window pixel. Based on the confidence-weighted average result, abnormal structure areas and confidence-abnormal areas are extracted as the basis for engineering suggestion generation. Based on the preset geological engineering knowledge base, the geological recognition result is matched and output.

Images