Levee hidden danger identification and minimally invasive verification method based on deep learning

Abstract

The application discloses a kind of embankment hidden danger identification and minimally invasive verification method based on deep learning, and the application relates to the technical field of embankment hidden danger prediction, comprising the following steps: collecting ground penetrating radar image and high-density electrical method image of known hidden danger, labeling hidden danger type, area and size, constructing multi-source atlas database, constructing embankment hidden danger identification model, collecting corresponding image data of the embankment to be measured, and inputting the identification model, selecting the preferred hidden danger pair with high overlap degree through spatial pairing, determining the comprehensive confidence score by comprehensively outputting confidence and image quality weight, evaluating joint uncertainty by combining confidence score and classification probability distribution entropy value, screening potential risk area, carrying out targeted minimally invasive drilling operation to verify the accuracy of identification result, ensure the accuracy of identification result, significantly improve the overall level of embankment safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of dike hazard prediction technology, specifically a deep learning-based method for dike hazard identification and minimally invasive verification. Background Technology

[0002] In the field of levee engineering, hazard identification and monitoring are crucial for ensuring levee safety. Currently, ground-penetrating radar (GPR) and high-density electrical resistivity tomography (EDT) are commonly used techniques for detecting potential hazards and defects within levees. These technologies acquire information about underground structures by emitting electromagnetic waves or electrical currents. However, while these methods can provide a degree of visualization of underground conditions, traditional hazard detection processes still rely on manual analysis and interpretation. Specifically, technicians need to manually annotate and analyze the acquired radar and EDT images to determine the type, location, and size parameters of the hazards. This process is not only time-consuming and labor-intensive but also susceptible to the influence of subjective human judgment, leading to questions about the reliability and consistency of the detection results. In existing technologies, manual annotation and analysis are inefficient and cannot meet the needs of large-scale datasets. With the development of monitoring technology, the amount of data generated in levee hazard monitoring is constantly increasing. Relying solely on manual analysis is not only costly but also makes it difficult to process massive amounts of data within a reasonable timeframe. Secondly, traditional methods have limitations in the accuracy of hazard identification. Because different operators may have different judgments on hazard characteristics, the identification results for the same area may vary significantly, increasing the risk of missed or false detections. These shortcomings mean that existing technologies have failed to effectively meet the dual requirements of efficiency and accuracy for levee safety monitoring, necessitating a new type of automated and intelligent identification method to solve these problems.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for identifying and minimally invasively verifying potential hazards in dikes, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for identifying and minimally invasively verifying potential hazards in dikes, comprising the following steps: Ground-penetrating radar images and high-density electrical resistivity tomography images were collected for dikes with known hidden dangers. The type, region, and size parameters of the hidden dangers in the images were marked by manual labeling to form a multi-source map database. Based on a multi-source map database, a levee hazard identification model was constructed using a YOLOv11 deep learning network. The model was trained using unlabeled original images from the multi-source map database as input and the corresponding labeled images as labels. Ground-penetrating radar (GPR) images and high-density electrical resistivity tomography (HRT) images of the section of the dike to be tested are collected and input into the trained dike hazard identification model. The type, region, and size parameters of the hazards in the GPR images and HRT images of the section of the dike to be tested are obtained respectively. Spatial pairing of potential hazards in ground-penetrating radar images and high-density electrical resistivity tomography images of the embankment section to be tested is performed. Preferred hazard pairs with spatial overlap greater than a preset threshold are selected, and the output confidence of the preferred hazard pairs is determined. The comprehensive confidence score of the preferred hazard pairs is determined based on the output confidence of the preferred hazard pairs and the image quality weight. Based on the comprehensive reliability score of the selected hidden danger pairs, the joint uncertainty measure is determined by combining the classification probability distribution entropy value of the selected hidden danger pairs. Based on the joint uncertainty measure, potential risk areas are screened, targeted minimally invasive drilling operations are carried out, and the accuracy of the identification results is verified by drilling endoscopic images.

[0006] Furthermore, the acquired ground-penetrating radar images and high-density electrical resistivity images were subjected to anomaly screening to remove images with missing regions. The images were then subjected to uniform size cropping and color level normalization preprocessing. The preprocessed ground-penetrating radar images and high-density electrical resistivity images were used as unlabeled original images in the multi-source image database. The specific types of hazards include loose material, voids, and sand-bearing layers; The specific method for annotating the type, region, and size parameters of potential hazards in images is as follows: using the LabelImg annotation tool, according to the YOLO model format, the hazard region is defined frame by frame, and corresponding category labels, region coordinates, and size parameters are assigned. The size parameters specifically refer to the length and depth of the hazard region. Unlabeled raw images are mapped one-to-one with their corresponding labeled images to form a multi-source atlas database.

[0007] Furthermore, the specific method for constructing the levee hazard identification model includes: selecting the YOLOv11 deep learning network as the basic framework of the levee hazard identification model, and initializing the structure of the levee hazard identification model, including setting the input image size, preset the anchor box parameters of the hazard area, and defining the number of category labels; During training, the CIOU-based bounding box regression loss function is used as the loss function of the levee hazard identification model. The Adam or SGD optimizer is used to set the learning rate decay strategy and batch size, and the weight update rate is dynamically adjusted in multiple training iterations. Meanwhile, an image enhancement strategy is introduced during the training process. The image enhancement strategy includes image rotation, flipping, contrast adjustment, random blurring, and color perturbation. The input unlabeled original image is enhanced online through the image enhancement strategy to obtain several enhanced unlabeled original images, which are used as training image data to train the levee hazard identification model. The performance parameters of the levee hazard identification model are monitored in real time during the training process. The performance parameters of the levee hazard identification model include prediction accuracy, recall, and mean precision. If the performance parameters of the levee hazard identification model meet the preset prediction requirements, the training of the levee hazard identification model is completed.

[0008] Furthermore, the specific logic underlying the spatial pairing of potential hazards in the ground-penetrating radar images and high-density electrical resistivity tomography images of the embankment section under test is as follows: All potential hazards identified in the ground-penetrating radar image and high-density electrical resistivity image of the embankment section to be tested are mapped to a unified two-dimensional coordinate system for embankment engineering. The two-dimensional coordinate system for embankment engineering has a fixed station number as the origin, the X-axis along the embankment line direction, and the Y-axis perpendicular to the depth direction of the embankment line. Based on the constructed two-dimensional coordinate system of the dike project, all potential hazard areas were identified; The spatial overlap between pairs of potential hazards in the ground-penetrating radar image and high-density electrical resistivity tomography image of the embankment section under test is calculated based on the potential hazard area. The specific calculation method is as follows: determine the intersection area and union area between pairs of potential hazards, calculate the intersection-union ratio between pairs of potential hazards based on the intersection area and union area, and characterize the spatial overlap by the intersection-union ratio.

[0009] Furthermore, the specific formula used to calculate the union area between any two potential hazards is as follows: In the formula, Let be the union area of ​​the i-th hidden danger in the ground-penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography image. The first one identified in the ground penetrating radar image The area of ​​the potential hazard zone. The first one identified in high-density electrical resistivity images The area of ​​the potential hazard zone. Let be the intersection area of ​​the i-th hidden danger in the ground penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography (EDT) image; i is the index of the hidden danger identified in the ground penetrating radar image, and j is the index of the hidden danger identified in the high-density EDT image. The specific formula used to calculate the intersection-union ratio of each pair of potential hazards based on the intersection area and union area is as follows: In the formula, is the cross-combination ratio of the i-th hidden danger in the ground penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography image.

[0010] Furthermore, the specific method for determining the output confidence of the preferred hidden danger pair is as follows: the output confidence of the preferred hidden danger pair includes the prediction confidence of the hidden danger in the ground penetrating radar image and the prediction confidence of the hidden danger in the high-density electrical resistivity tomography image. Specifically, the output confidence of the preferred hidden danger pair is determined by the dike hidden danger identification model. The method for obtaining the image quality weight is as follows: calculate the local signal-to-noise ratio of the ground penetrating radar image and the inversion residual of the high-density electrical resistivity tomography image, and normalize the local signal-to-noise ratio of the ground penetrating radar image and the inversion residual of the high-density electrical resistivity tomography image through the S-shaped function, and use the normalized value as the image quality weight. The specific formula used to calculate the quality weight of ground-penetrating radar images is as follows: In the formula, For the quality weight of ground-penetrating radar images, For the local signal-to-noise ratio of ground-penetrating radar images, This is an empirical threshold for the signal-to-noise ratio. The steepness control parameter for the S-shaped function of the ground-penetrating radar image; The specific formula used to calculate the quality weight of high-density electrical resistivity images is as follows: In the formula, For high-density electrical resistivity image quality weights, For the inversion residuals of high-density electrical resistivity images, For the empirical threshold of the inverted residual, The steepness control parameter is used to determine the slopingness of the S-shaped function in high-density electrical resistivity images.

[0011] Furthermore, the specific formula used to calculate the overall reliability score of the preferred hidden danger pair is as follows: In the formula, The overall reliability score for the q-th preferred hidden danger pair is... For the q-th preferred hazard pair, the prediction confidence of the hazard in the ground-penetrating radar image is... For the q-th preferred hazard pair, the prediction confidence of the hazard in the high-density electrical resistivity tomography image is... Let be the crossover ratio of the qth preferred hidden danger pair. , and q is the proportionality coefficient, and q is the index of the preferred hidden danger pair; The formula used to determine the joint uncertainty measure of the preferred hidden danger pair is as follows: In the formula, For the q-th preferred hidden danger pair, For balance coefficient, Let be the entropy value of the classification probability distribution of the q-th preferred hazard pair, specifically the average entropy value of the hazard in the ground-penetrating radar image and the high-density electrical resistivity image of the q-th preferred hazard pair; The specific method used to screen potential disease areas based on joint uncertainty measurement is as follows: Based on the selected hidden danger pairs, the union area of ​​the hidden dangers of each selected hidden danger pair is determined as the candidate verification area. According to the joint uncertainty measure of the selected hidden danger pairs, the candidate verification area corresponding to the selected hidden danger pairs whose joint uncertainty measure is not less than the preset verification threshold is recorded as the risk potential disease area. Minimally invasive drilling is performed on the risk potential disease area to verify the accuracy of the identification results.

[0012] Compared with the prior art, the beneficial effects of the present invention are: By constructing a YOLOv11 deep learning network based on a multi-source map database, the automation and accuracy of levee hazard identification have been significantly improved. Compared with traditional manual annotation and analysis methods, this solution can quickly process a large amount of ground-penetrating radar and high-density electrical resistivity tomography image data, automatically extract hazard features and classify them, improve the speed and efficiency of hazard identification, ensure the timely detection of potential risks, and thus provide effective technical support for levee safety. Furthermore, the active learning strategy and joint uncertainty measurement introduced in the solution enable the identification model to self-optimize through continuous iteration, dynamically adjusting the screening and classification criteria for hazards. This not only improves the accuracy of hazard identification and reduces the risk of false positives and false negatives, but also gives the hazard identification results higher reliability by comprehensively analyzing the comprehensive reliability scores of the optimal hazard pairs. Targeted minimally invasive drilling operations and endoscopic image verification further ensure the accuracy of the identification results, significantly improving the overall level of dike safety monitoring and contributing to ensuring public safety and the reliability of infrastructure. Attached Figure Description

[0013] Fig. 1 This is a schematic diagram of the overall method flow of the present invention; Fig. 2 The graph shows the fitting curve of local signal-to-noise ratio of ground-penetrating radar image versus quality weight of ground-penetrating radar image. Fig. 3 A dotted line plot showing the distribution of quality weights for each image. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0015] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0016] Example: Please see Figs. 1-3 The present invention provides a technical solution: A deep learning-based method for identifying and minimally invasively verifying potential hazards in dikes, comprising the following steps: Step 1: Collect ground-penetrating radar images and high-density electrical resistivity tomography images of dikes with known hidden dangers. Use manual labeling to annotate the type, region, and size parameters of the hidden dangers in the images to form a multi-source map database.

[0017] The acquired ground-penetrating radar images and high-density electrical resistivity images were subjected to anomaly screening to remove images with missing regions. The images were then preprocessed by uniform size cropping and color level normalization. The preprocessed ground-penetrating radar images and high-density electrical resistivity images were used as unlabeled original images in the multi-source image database. The goal of image anomaly screening is to remove images of poor quality or incompleteness, such as images with missing regions. Ground penetrating radar images and high-density electrical resistivity tomography images may contain noise, missing regions, or signal anomalies, such as black blocks, blank spots, and stripes. These anomalies may interfere with subsequent analysis and modeling. Different equipment or acquisition conditions may result in ground-penetrating radar images and high-density electrical resistivity tomography (EDT) images having different resolutions and sizes. To facilitate subsequent analysis, it is necessary to crop the images to a uniform size. The principle of cropping is to retain the effective information in the image while removing invalid or abnormal data at the edges. Generally, a fixed target size is selected, such as 256×256 or 512×512 pixels, and the image size is adjusted by center cropping or proportional cropping. If the original image size does not match the target size, interpolation methods, such as bilinear interpolation, can be used to adjust the cropped image.

[0018] The grayscale values ​​or color level distributions of ground-penetrating radar images and high-density electrical resistivity tomography (EDT) images may not be consistent. For example, different equipment or acquisition conditions may lead to differences in parameters such as image brightness and contrast. In order to eliminate color level differences between different images and facilitate subsequent comparative analysis, it is necessary to normalize the color levels of the images. Histogram equalization is used to adjust the grayscale distribution of the images, making the pixel value distribution of the images more uniform and enhancing the contrast. This can be achieved using the equalizeHist function in OpenCV or MATLAB.

[0019] The process involves retaining clear, representative original image data with distinct target features, and performing preprocessing operations such as uniform size cropping, resolution adjustment, and color level normalization to improve the standardization and readability of the images when inputting them into the model. Unlabeled original images require uniform preprocessing to ensure consistent quality across multiple data sources. This facilitates feature extraction and fusion analysis by subsequent deep learning models.

[0020] The specific types of hazards include loose material, voids, and sand-bearing layers; The specific method for annotating the type, region, and size parameters of potential hazards in images is as follows: Using the LabelImg annotation tool, following the YOLO model format, the hazard region is defined frame by frame in each image, and corresponding category labels, region coordinates, and size parameters are assigned. The size parameters specifically refer to the length and depth of the hazard region. The specific annotation steps include: opening the LabelImg tool, selecting "Open Dir" to open the folder containing the images to be annotated, selecting "Change Save Dir" to set the folder for saving the annotation results, selecting the YOLO format in LabelImg to generate an annotation file that conforms to the YOLO model requirements, selecting the hazard region in the image, defining the hazard region and its type, such as loose material, voids, and sand-bearing layers, entering or selecting the corresponding category label in the pop-up dialog box, and recording the length and depth of the hazard region. This information can be added to the LabelImg property box during annotation, or these parameters can be recorded through subsequent processing after annotation is completed. Repeat the above steps until all images are annotated.

[0021] Unlabeled raw images are mapped one-to-one with their corresponding labeled images to form a multi-source atlas database.

[0022] Step 2: Based on the multi-source map database, a dike hazard identification model is constructed using the YOLOv11 deep learning network. The model is trained using unlabeled original images from the multi-source map database as input and the corresponding labeled images as labels.

[0023] The specific method for constructing a levee hazard identification model includes: selecting the YOLOv11 deep learning network as the basic framework of the levee hazard identification model, and initializing the structure of the levee hazard identification model, including setting the input image size, preset the anchor box parameters of the hazard area, and defining the number of category labels; During training, the CIOU-based bounding box regression loss function is used as the loss function of the levee hazard identification model. The Adam or SGD optimizer is used to set the learning rate decay strategy and batch size, and the weight update rate is dynamically adjusted in multiple training iterations. Meanwhile, an image enhancement strategy is introduced during the training process. The image enhancement strategy includes image rotation, flipping, contrast adjustment, random blurring, and color perturbation. The input unlabeled original image is enhanced online through the image enhancement strategy to obtain several enhanced unlabeled original images, which are used as training image data to train the levee hazard identification model. Organize the unlabeled raw images and corresponding labeled images from the multi-source image database into a designated folder; ensure that the labeled image format conforms to the YOLO format and that the category labels match the model requirements; in the YOLOv11 configuration file, set the size of the input image, for example: 640x640 or 416x416, to ensure that the image size is uniform in order to improve training efficiency. In the YOLOv11 configuration file, define the number of category labels for dike hazards, such as loose material, voids, and sandy layers. Run the training script to start model training. Monitor the loss value and accuracy during training. Use the Adam or SGD optimizer to adjust hyperparameters to optimize model performance. The performance parameters of the levee hazard identification model are monitored in real time during the training process. The performance parameters of the levee hazard identification model include prediction accuracy, recall, and mean precision. If the performance parameters of the levee hazard identification model meet the preset prediction requirements, the training of the levee hazard identification model is completed.

[0024] Step 3: Collect ground-penetrating radar images and high-density electrical resistivity tomography (HRT) images of the section of the dike to be tested within the dike to be tested, and input them into the trained dike hazard identification model to obtain the type, area, and size parameters of the hazards in the ground-penetrating radar images and high-density HRT images of the section of the dike to be tested.

[0025] The specific method for acquiring ground-penetrating radar (GPR) and high-density electrical resistivity tomography (HERS) images of the dike section under test is as follows: High-quality GPR and HERS equipment are used to provide clear and high-resolution image data. The GPR equipment is moved along a predetermined path on the dike section to record radar reflection signals, generating GPR images. These images show different features of the underground structure and can indicate potential hazard areas. HERS equipment is then used to measure along the same or similar paths to obtain the dike's resistivity data. This data can be converted into images reflecting the electrical characteristics of different soil layers, which helps in identifying hazard areas. All images are adjusted to a uniform size and format to meet the input requirements of the trained dike hazard identification model. The acquired images are preprocessed, and the preprocessed GPR and HERS images are input into the trained dike hazard identification model. The model analyzes the input images, identifies potential hazard areas, and outputs the type, region, and size parameters of the hazard, marking them in the corresponding areas of the images. Table 1 shows the hazard identification results for each dike station. Table 1. Hazard Identification Data for Each Lot Number in the Dike Model Step 4: Spatially pair the hidden dangers in the ground-penetrating radar image and high-density electrical resistivity tomography image of the embankment section to be tested, select the preferred hidden danger pairs with a spatial overlap greater than a preset threshold, determine the output confidence of the preferred hidden danger pairs, and determine the comprehensive confidence score of the preferred hidden danger pairs based on the output confidence of the preferred hidden danger pairs and the image quality weight.

[0026] The logic behind spatially pairing potential hazards in ground-penetrating radar images and high-density electrical resistivity tomography (EDT) images of the embankment section under test is as follows: All potential hazards identified in the ground-penetrating radar image and high-density electrical resistivity image of the embankment section to be tested are mapped to a unified two-dimensional coordinate system for embankment engineering. The two-dimensional coordinate system for embankment engineering has a fixed station number as the origin, the X-axis along the embankment line direction, and the Y-axis perpendicular to the depth direction of the embankment line. Based on the constructed two-dimensional coordinate system of the dike project, all potential hazard areas were identified; The spatial overlap between pairs of potential hazards in the ground-penetrating radar image and high-density electrical resistivity tomography image of the embankment section under test is calculated based on the potential hazard area. The specific calculation method is as follows: determine the intersection area and union area between pairs of potential hazards, calculate the intersection-union ratio between pairs of potential hazards based on the intersection area and union area, and characterize the spatial overlap by the intersection-union ratio.

[0027] The specific formula used to calculate the union area between any two potential hazards is as follows: In the formula, Let be the union area of ​​the i-th hidden danger in the ground-penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography image. The first one identified in the ground penetrating radar image The area of ​​the potential hazard zone. The first one identified in high-density electrical resistivity images The area of ​​the potential hazard zone. Let be the intersection area of ​​the i-th hidden danger in the ground penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography (EDT) image; i is the index of the hidden danger identified in the ground penetrating radar image, and j is the index of the hidden danger identified in the high-density EDT image. It should be noted that the union of two sets refers to the set of all elements in two sets. Geometrically, the union of two rectangles is the smallest region that contains both rectangles. The logic for calculating the area of ​​the union is: add the areas of the two rectangles. However, this will double-count the area of ​​the intersection. Therefore, it is necessary to subtract the area of ​​the intersection to ensure that the area of ​​the union is calculated only once for the overlapping area. The formula is based on the basic principles of set theory. When calculating the union of two sets, the problem of double-counting the intersection needs to be considered. From a geometric point of view, the union of two rectangles can be obtained by adding the areas of each rectangle and then subtracting the area of ​​the overlapping part to ensure the accuracy of the result. The specific formula used to calculate the intersection-union ratio of each pair of potential hazards based on the intersection area and union area is as follows: In the formula, is the cross-combination ratio of the i-th hidden danger in the ground penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography image.

[0028] It should be noted that the intersection area can be zero or a positive value. When two potential hazards overlap, the intersection area is greater than zero; if there is no overlap, the intersection area is zero. The calculation of the union area depends on the existence and size of the intersection area to ensure that the overlapping area is not counted repeatedly. The intersection-union ratio provides a clear proportion to represent the degree of overlap between two regions; when A value of 1 indicates that the two regions completely overlap; when A value of 0 indicates no overlap, while other values ​​indicate partial overlap, which can intuitively reflect the spatial relationship between the two. The intersection-union ratio is a value between 0 and 1, which can eliminate the measurement bias caused by different area sizes, making it possible to compare the degree of overlap between different areas. At the same time, regardless of the type of image (such as ground-penetrating radar image and high-density electrical resistivity image) or different hidden dangers, Both methods can effectively assess the degree of spatial overlap and have strong versatility. By combining the intersection and union areas, the intersection-union ratio provides a comprehensive analysis of regional overlap, effectively identifying true overlapping areas rather than relying solely on simple area comparisons. It can clearly represent the degree of overlap and also has advantages such as simplicity, ease of understanding, and strong adaptability, providing a reliable basis for decision-making in the process of hazard identification and analysis.

[0029] Preferred potential hazard pairs are selected based on their spatial overlap exceeding a preset threshold. The preset threshold is specifically set based on expert experience and is generally set between 0.7 and 0.9.

[0030] The specific method for determining the output confidence of the preferred hidden danger pair is as follows: the output confidence of the preferred hidden danger pair includes the prediction confidence of the hidden danger in the ground penetrating radar image and the prediction confidence of the hidden danger in the high-density electrical resistivity tomography image. Specifically, the output confidence of the preferred hidden danger pair is determined by the dike hidden danger identification model. The method for obtaining the image quality weight is as follows: calculate the local signal-to-noise ratio of the ground penetrating radar image and the inversion residual of the high-density electrical resistivity tomography image, and normalize the local signal-to-noise ratio of the ground penetrating radar image and the inversion residual of the high-density electrical resistivity tomography image through the S-shaped function, and use the normalized value as the image quality weight. The specific formula used to calculate the quality weight of ground-penetrating radar images is as follows: In the formula, For the quality weight of ground-penetrating radar images, For the local signal-to-noise ratio of ground-penetrating radar images, This is an empirical threshold for the signal-to-noise ratio. The steepness control parameter for the S-shaped function of the ground-penetrating radar image; It should be noted that this formula uses a sigmoid function (logistic function), which has a smooth transition characteristic and can map the local signal-to-noise ratio to between 0 and 1, thus making the image quality weights more balanced. It can effectively represent the quality of an image; This represents the quality weight of a ground-penetrating radar image, with a value ranging from 0 to 1. The closer the value is to 1, the higher the image quality, and therefore the higher the weight. Greater than hour, The value will gradually approach 1, indicating high image quality; when Less than hour, A value that gradually approaches 0 indicates low image quality; Local signal-to-noise ratio of ground penetrating radar images The signal-to-noise ratio (SNR) reflects the ratio of signal strength to noise strength. A higher SNR indicates a more prominent effective signal and better image quality. By using the SNR as input, the quality of ground-penetrating radar (GPR) images can be directly reflected. A higher SNR indicates more effective information and less noise in the image, thus resulting in higher quality. The use of the sigmoid function allows for a smooth transition in the quality weight when the local SNR value changes, avoiding instability caused by abrupt changes. Among them, the empirical threshold of signal-to-noise ratio Used to determine The elevation of this value can be determined based on experience or statistical data, and serves as a reference level; the steepness of the S-curve function in a ground-penetrating radar image is controlled by parameters. This determines the slope of the function; a larger slope... A value that makes the function change more steeply near the threshold, and more sensitive; a smaller value... The value makes the changes smoother and more adaptable; the specific setting can be based on historical experience. The specific formula used to calculate the quality weight of high-density electrical resistivity images is as follows: In the formula, For high-density electrical resistivity image quality weights, For the inversion residuals of high-density electrical resistivity images, For the empirical threshold of the inverted residual, The steepness control parameter for the S-shaped function of high-density electrical resistivity tomography (EDT) images is set using the same method as that for the S-shaped function of ground-penetrating radar (GPR) images. The inversion residual of a high-density electrical resistivity tomography (EDT) image refers to the difference between the actual observed value and the image observed value during the inversion process of high-density EDT data. Specific methods for obtaining this residual include: acquiring image observed values ​​through high-density EDT measurements; establishing an inversion model based on geological models, physical properties, and other information, such as using a certain algorithm for resistivity inversion to obtain a preliminary resistivity distribution map; calculating the difference between the actual observed value and the image observed value for each measurement point; and obtaining the overall inversion residual value by statistically analyzing the residuals of all measurement points, which is usually expressed using indicators such as root mean square error. The empirical threshold for inversion residuals is set based on expert experience.

[0031] The specific formula used to calculate the overall reliability score of the preferred hidden danger pair is as follows: In the formula, The overall reliability score for the q-th preferred hidden danger pair is... For the q-th preferred hazard pair, the prediction confidence of the hazard in the ground-penetrating radar image is... For the q-th preferred hazard pair, the prediction confidence of the hazard in the high-density electrical resistivity tomography image is... Let be the crossover ratio of the qth preferred hidden danger pair. , and q is the proportionality coefficient, and q is the index of the preferred hidden danger pair; It should be noted that the first The overall reliability score of the preferred hidden danger pair The larger the value, the more significant the first... The higher the credibility of the preferred potential hazard pair, the more reliable the hazard identification result will be; This formula unifies the confidence level of hazard prediction using ground-penetrating radar and high-density electrical resistivity tomography with the influence of image quality and overlap into a single score, reflecting the need for comprehensive judgment; through multi-dimensional evaluation, it can more comprehensively reflect the credibility of hazards. and The product of these values ​​reflects the confidence contribution of ground-penetrating radar images to hazard identification; when image quality is high and prediction confidence is high, this part of the score will be significantly improved. and The product of these two factors also reflects the reliability of high-density electrical resistivity tomography (EDT) images in identifying potential hazards. The relationship between them is the same as above: high-quality images and high confidence levels together improve the score. The introduction of this feature compensates for the spatial overlap between the two image recognition results, providing additional validation reliability. A high degree of overlap indicates that the two potential issues are more likely to be the same problem, thus improving the overall score. Introducing a proportionality coefficient This allows for flexible adjustment of the influence of each factor in the overall score, among which... And the proportionality coefficient All are greater than 0; the intersection-union ratio is an important indicator for measuring the degree of overlap between two potential hazards. A high degree of overlap usually means that there is a stronger correlation between the two. Therefore, it is given a higher weight in the comprehensive score. Through such weight setting, it can be ensured that the influence of different factors on the comprehensive score is relatively balanced; high-quality images and high-confidence predictions will play a positive role in promoting the score, while the influence of low-quality or low-confidence predictions will be limited to a certain extent to avoid adverse effects on the comprehensive evaluation.

[0032] Step 5: Based on the comprehensive reliability score of the selected hidden danger pairs, and combined with the entropy value of the classification probability distribution of the selected hidden danger pairs, determine their joint uncertainty measure. Based on the joint uncertainty measure, screen potential risk areas, carry out targeted minimally invasive drilling operations, and verify the accuracy of the identification results through borehole endoscopy images.

[0033] The formula used to determine the joint uncertainty measure of the preferred hidden danger pair is as follows: In the formula, For the q-th preferred hidden danger pair, For balance coefficient, Let be the entropy value of the classification probability distribution of the q-th preferred hazard pair, specifically the average entropy value of the hazard in the ground-penetrating radar image and the high-density electrical resistivity image of the q-th preferred hazard pair; It should be noted that the overall reliability score is calculated... Entropy of classification probability distribution Combining these methods can effectively quantify the overall uncertainty of each potential risk pair. Higher confidence and lower entropy should result in a smaller uncertainty measure, and vice versa; when... The larger the value, the more significant the th . The higher the joint uncertainty of a preferred hazard pair, the less reliable the identification result of that hazard is, and the greater the potential risk of the hazard. Therefore, a high joint uncertainty measure may indicate the need for further verification steps, such as conducting targeted minimally invasive drilling operations to confirm the accuracy of the identification results using endoscopic images. Introducing the balance coefficient This is to adjust the degree of influence of the reliability score on the joint uncertainty measure. In different scenarios, this coefficient may need to be adjusted according to the actual situation in order to better reflect the reliability of hazard identification; The entropy value of the classification probability distribution of the qth preferred hidden danger pair It measures the uncertainty of hazard classification results. A higher entropy value indicates greater uncertainty in the classification result, implying a greater risk. Therefore, using it as the denominator in the formula effectively highlights the uncertainty under high entropy conditions. The increase in joint uncertainty This will increase accordingly; this indicates that the more uncertain the hazard classification results are, the higher the risk assessment of the entire hazard pair will be. The entropy value of the classification probability distribution of the q-th preferred hidden danger pair Specifically, the entropy values ​​of the hazards in the ground-penetrating radar (GPR) image and the high-density electrical resistivity tomography (HRT) image of the preferred hazard pair are determined separately, and the average value of the entropy values ​​of the hazards in the GPR image and the HRT image is calculated as the joint uncertainty measure classification probability distribution entropy value of the q-th preferred hazard pair; the specific entropy calculation formula is a conventional existing technology and will not be elaborated here.

[0034] The specific method used to screen potential disease areas based on joint uncertainty measurement is as follows: Based on the preferred hazard pairs, the union area of ​​the hazards of each preferred hazard pair is determined as the candidate verification area. According to the joint uncertainty measure of the preferred hazard pairs, the candidate verification area corresponding to the preferred hazard pairs whose joint uncertainty measure is not less than the preset verification threshold is recorded as the risk potential disease area. Minimally invasive drilling is performed on the risk potential disease area to verify the accuracy of the identification results. The preset verification threshold is specifically set by expert experience combined with historical data.

[0035] By analyzing the optimal pairs of potential hazards and the area of ​​their clusters, a more comprehensive understanding of the risks in areas with potential defects can be achieved. This method can effectively identify potential hazards, reduce the risk of missed detections, and verify the identification results through minimally invasive drilling operations. It allows for the acquisition of accurate geological and hazard information without affecting the overall structure, effectively reducing costs and risks.

[0036] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0037] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0038] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for identifying and minimally invasively verifying potential hazards in dikes based on deep learning, characterized in that, The specific steps include: Ground-penetrating radar images and high-density electrical resistivity tomography images were collected for dikes with known hidden dangers. The type, region, and size parameters of the hidden dangers in the images were marked by manual labeling to form a multi-source map database. Based on a multi-source map database, a levee hazard identification model was constructed using a YOLOv11 deep learning network. The model was trained using unlabeled original images from the multi-source map database as input and the corresponding labeled images as labels. Ground-penetrating radar (GPR) images and high-density electrical resistivity tomography (HRT) images of the section of the dike to be tested are collected and input into the trained dike hazard identification model. The type, region, and size parameters of the hazards in the GPR images and HRT images of the section of the dike to be tested are obtained respectively. Spatially pairing potential hazards in ground-penetrating radar images and high-density electrical resistivity tomography images of the embankment section under test, selecting preferred hazard pairs with spatial overlap greater than a preset threshold, determining the output confidence of the preferred hazard pairs, and comprehensively determining the overall confidence score of the preferred hazard pairs based on the output confidence of the preferred hazard pairs and the image quality weight. Based on the comprehensive reliability score of the selected hidden danger pairs, the joint uncertainty measure is determined by combining the classification probability distribution entropy value of the selected hidden danger pairs. Based on the joint uncertainty measure, potential risk areas are screened, targeted minimally invasive drilling operations are carried out, and the accuracy of the identification results is verified by drilling endoscopy images. The entropy value of the classification probability distribution is specifically the average entropy value of the hidden danger in the ground-penetrating radar image and the high-density electrical resistivity tomography image of each preferred hidden danger pair.

2. The method for identifying and minimally invasively verifying levee hazards based on deep learning according to claim 1, characterized in that: The acquired ground-penetrating radar images and high-density electrical resistivity images were subjected to anomaly screening to remove images with missing regions. The images were then preprocessed by uniform size cropping and color level normalization. The preprocessed ground-penetrating radar images and high-density electrical resistivity images were used as unlabeled original images in the multi-source image database. The specific types of hazards include loose material, voids, and sand-bearing layers; The specific method for annotating the type, region, and size parameters of potential hazards in images is as follows: using the LabelImg annotation tool, according to the YOLO model format, the hazard region is defined frame by frame, and corresponding category labels, region coordinates, and size parameters are assigned. The size parameters specifically refer to the length and depth of the hazard region. Unlabeled raw images are mapped one-to-one with their corresponding labeled images to form a multi-source atlas database.

3. The method for identifying and minimally invasively verifying levee hazards based on deep learning according to claim 2, characterized in that: The specific method for constructing a levee hazard identification model includes: selecting the YOLOv11 deep learning network as the basic framework of the levee hazard identification model, and initializing the structure of the levee hazard identification model, including setting the input image size, preset the anchor box parameters of the hazard area, and defining the number of category labels; During training, the CIOU-based bounding box regression loss function is used as the loss function of the levee hazard identification model. The Adam or SGD optimizer is used, and the learning rate decay strategy and batch size are set to dynamically adjust the weight update rate in multiple training iterations. Meanwhile, an image enhancement strategy is introduced during the training process. The image enhancement strategy includes image rotation, flipping, contrast adjustment, random blurring, and color perturbation. The input unlabeled original image is enhanced online through the image enhancement strategy to obtain several enhanced unlabeled original images, which are used as training image data to train the levee hazard identification model. The performance parameters of the levee hazard identification model are monitored in real time during the training process. The performance parameters of the levee hazard identification model include prediction accuracy, recall, and mean precision. If the performance parameters of the levee hazard identification model meet the preset prediction requirements, the training of the levee hazard identification model is completed.

4. The method for identifying and minimally invasively verifying levee hazards based on deep learning according to claim 2, characterized in that: The logic behind spatially pairing potential hazards in ground-penetrating radar images and high-density electrical resistivity tomography (EDT) images of the embankment section under test is as follows: All potential hazards identified in the ground-penetrating radar image and high-density electrical resistivity image of the embankment section to be tested are mapped to a unified two-dimensional coordinate system for embankment engineering. The two-dimensional coordinate system for embankment engineering has a fixed station number as the origin, the X-axis along the embankment line direction, and the Y-axis perpendicular to the depth direction of the embankment line. Based on the constructed two-dimensional coordinate system of the dike project, all potential hazard areas are identified. Based on the potential hazard areas, the spatial overlap between each pair of potential hazards in the ground-penetrating radar image and high-density electrical resistivity tomography image of the dike section to be tested is calculated. The specific calculation method is as follows: determine the intersection area and union area between each pair of potential hazards, calculate the intersection-union ratio between each pair of potential hazards based on the intersection area and union area, and characterize the spatial overlap by the intersection-union ratio.

5. The method for identifying and minimally invasively verifying levee hazards based on deep learning according to claim 4, characterized in that: The specific formula used to calculate the union area between any two potential hazards is as follows: In the formula, Let be the union area of ​​the i-th hidden danger in the ground-penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography image. The first one identified in the ground penetrating radar image The area of ​​the potential hazard zone. The first one identified in high-density electrical resistivity images The area of ​​the potential hazard zone. Let be the intersection area of ​​the i-th hidden danger in the ground penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography (EDT) image; i is the index of the hidden danger identified in the ground penetrating radar image, and j is the index of the hidden danger identified in the high-density EDT image. The specific formula used to calculate the intersection-union ratio of each pair of potential hazards based on the intersection area and union area is as follows: In the formula, is the cross-combination ratio of the i-th hidden danger in the ground penetrating radar image and the j-th hidden danger in the high-density electrical resistivity tomography image.

6. The method for identifying and minimally invasively verifying levee hazards based on deep learning according to claim 5, characterized in that: The specific method for determining the output confidence of the preferred hazard pair is as follows: the output confidence of the preferred hazard pair includes the prediction confidence of the hazard in the ground penetrating radar image and the prediction confidence of the hazard in the high-density electrical resistivity tomography image, both of which are specifically determined by the levee hazard identification model; The method for obtaining the image quality weight is as follows: calculate the local signal-to-noise ratio of the ground penetrating radar image and the inversion residual of the high-density electrical resistivity tomography image, and normalize the local signal-to-noise ratio of the ground penetrating radar image and the inversion residual of the high-density electrical resistivity tomography image through the S-shaped function, and use the normalized value as the image quality weight. The specific formula used to calculate the quality weight of ground-penetrating radar images is as follows: In the formula, For the quality weight of ground-penetrating radar images, For the local signal-to-noise ratio of ground-penetrating radar images, This is an empirical threshold for the signal-to-noise ratio. The steepness control parameter for the S-shaped function of the ground-penetrating radar image; The specific formula used to calculate the quality weight of high-density electrical resistivity images is as follows: In the formula, For high-density electrical resistivity image quality weights, For the inversion residuals of high-density electrical resistivity images, For the empirical threshold of the inverted residual, The steepness control parameter is the S-shaped function of the high-density electrical resistivity image.

7. The method for identifying and minimally invasively verifying levee hazards based on deep learning according to claim 6, characterized in that: The specific formula used to calculate the overall reliability score of the preferred hidden danger pair is as follows: In the formula, The overall reliability score for the q-th preferred hidden danger pair is... For the q-th preferred hazard pair, the prediction confidence of the hazard in the ground-penetrating radar image is... For the q-th preferred hazard pair, the prediction confidence of the hazard in the high-density electrical resistivity tomography image is... Let be the crossover ratio of the qth preferred hidden danger pair. , and q is the proportionality coefficient, and q is the index of the preferred hidden danger pair; The formula used to determine the joint uncertainty measure of the preferred hidden danger pair is as follows: In the formula, For the q-th preferred hidden danger pair, For balance coefficient, Let be the entropy value of the classification probability distribution of the q-th preferred hidden danger pair; The specific method used to screen potential disease areas based on joint uncertainty measurement is as follows: Based on the selected hidden danger pairs, the union area of ​​the hidden dangers of each selected hidden danger pair is determined as the candidate verification area. According to the joint uncertainty measure of the selected hidden danger pairs, the candidate verification area corresponding to the selected hidden danger pairs whose joint uncertainty measure is not less than the preset verification threshold is recorded as the risk potential disease area. Minimally invasive drilling is performed on the risk potential disease area to verify the accuracy of the identification results.

Images