Dam termite nest intelligent detection and control method based on multi-mode AI model

Abstract

The invention discloses a dam termite nest intelligent detection and control method and system based on a multi-mode AI model, and relates to the technical field of hydraulic engineering safety monitoring and termite control. The method comprises the following steps: S1, multi-modal data acquisition: acquiring ground penetrating radar data and surface visual data by using a ground penetrating radar and a video camera, and constructing a multi-modal data set; s2, performing intelligent analysis on the ground penetrating radar data and the ground surface visual data, specifically comprising the steps of S3, performing three-dimensional reconstruction on the termite nest; and step 4, based on a three-dimensional reconstruction result in the step S3, carrying out hidden danger assessment and prevention and treatment decision making, and constructing a'perception-evaluation-decision making-treatment 'whole-process closed-loop management system. According to the invention, ground penetrating radar and visual AI analysis technologies are integrated, and precise detection, positioning and treatment of termite nests are realized through multi-modal data acquisition, intelligent analysis, three-dimensional modeling, hidden danger assessment and prevention and treatment decision.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy project safety monitoring and termite control, and specifically to a method and system for intelligent detection and control of termite nests in dams based on a multimodal AI model. Background Art

[0002] Levees are crucial infrastructure for flood control, irrigation, and water supply projects, and their structural safety directly impacts the safety of people and property. However, termite nest erosion has become a major safety hazard for levees. Termites construct complex tunnel systems within levees, loosening the soil structure and reducing shear strength. This can lead to seepage, piping, landslides, collapses, and even dam failures.

[0003] Currently, termite nest detection relies primarily on manual inspections and single geophysical methods, such as surface sign searches, baiting, ground-penetrating radar, and high-density resistivity. However, these methods suffer from low detection efficiency, reliance on expert experience, significant interference from environmental factors, and high rates of missed detection. These methods fail to meet the demands of modern water conservancy projects for rapid and accurate detection of termite hazards. Furthermore, existing technologies lack the ability to accurately locate the three-dimensional morphology of termite nests and lack an intelligent prevention and control system encompassing the entire "perception-assessment-decision-making-control" process, limiting the scientific nature and effectiveness of prevention and control measures.

[0004] In recent years, artificial intelligence (AI) has been widely applied in engineering safety monitoring. Ground-penetrating radar (GPR) intelligent analysis, computer vision recognition, and multimodal data fusion have provided new approaches for detecting termite nests on embankments. However, in complex embankment environments, using AI to accurately identify ecological signatures of termites, analyze GPR data, construct three-dimensional morphological models of termite nests, and integrate historical data to optimize prevention and control strategies remains a challenge. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a multimodal AI-based intelligent detection and control method and system for termite nests on embankments. By integrating ground-penetrating radar and visual AI analysis technologies, this system uses a deep learning model to automatically identify, 3D reconstruct, and intelligently determine termite ecological signs and nests. This system establishes a full-lifecycle management system for termite detection and control, enabling efficient detection, precise management, and scientific assessment of termite hazards, thereby enhancing the safety monitoring capabilities of water conservancy projects. Compared to existing technologies, this invention overcomes the limitations of traditional detection methods, improves the automation of detection and the intelligence of data processing, and provides a more efficient and accurate termite control solution for water conservancy project safety.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A multimodal AI model-based intelligent detection and prevention method for termite nests on dams includes the following steps:

[0008] Step S1, multimodal data acquisition: using ground penetrating radar and video camera to collect ground penetrating radar data and surface visual data to construct a multimodal dataset;

[0009] Step S2, performing intelligent analysis on the ground penetrating radar data and surface visual data, specifically including:

[0010] Step S3, performing three-dimensional reconstruction of the termite nest;

[0011] Step 4: Based on the three-dimensional reconstruction results of step S3, conduct hidden danger assessment and prevention and control decisions, and build a closed-loop management system for the entire process of "perception-assessment-decision-making-governance".

[0012] Furthermore, the step S2 includes:

[0013] Step S21: Perform AI recognition on the surface visual data:

[0014] Step S22: analyzing the ground penetrating radar data.

[0015] Furthermore, step S3 includes:

[0016] In step S31, the spatial distribution characteristics of the ground penetrating radar data are analyzed through the maximum density projection algorithm and the SegNet target imaging network, and a three-dimensional morphological model of the ant nest is constructed to achieve accurate positioning of the ant nest, and the morphological optimization is performed in combination with the historical dam structure information.

[0017] Step S32, based on the multi-channel two-dimensional ground penetrating radar B-scan data, three-dimensional imaging of underground internal diseases is performed. By deeply mining the spatial feature information of the multi-channel ground penetrating radar B-scan data, the complex three-dimensional morphology of the underground ant nest can be finely depicted and the spatial position can be accurately located.

[0018] Furthermore, it also includes step S5, prevention and control implementation and effect evaluation, executing anthill control according to the prevention and control plan, re-surveying the embankment structure through ground penetrating radar, and combining AI model to analyze the before and after data, quantifying the control effect, realizing long-term tracking monitoring, and building a full life cycle management system for termite control.

[0019] Furthermore, in step S1, multimodal data collection is performed using cameras and ground penetrating radar equipment to collect termite ecological environment signs on the surface of the dam and underground electromagnetic wave signals, thereby constructing a multimodal dataset including surface signs and underground structures.

[0020] Furthermore, in step S21, the ground penetrating radar data and the surface visual data are intelligently analyzed, and a multi-branch convolutional neural network modal fusion algorithm is used, combined with a feature pyramid network and a dual-path lightweight convolutional network, to identify signs of termite activity such as mud sheets, mud lines, cluster holes, and fungi;

[0021] By adopting a deep learning model based on the coupling of ground-penetrating radar electromagnetic wave characteristics and profile features, and integrating incremental learning and transfer learning techniques, high-precision termite nest identification can be achieved under small sample conditions.

[0022] Furthermore, in step S3, the three-dimensional reconstruction of the termite nest is based on the SegNet target imaging network and combined with the spatial distribution characteristics of the ground penetrating radar data to generate a three-dimensional morphological model of the termite nest, thereby achieving accurate positioning of the termite nest.

[0023] Furthermore, in the multimodal data collection, the termite ecological environment sign data include mud lines, mud sheets, flight holes, termite nest indicators and termite-eaten food.

[0024] Furthermore, in step S2, a termite ecological environment sign recognition algorithm based on vision-based multi-branch convolutional neural network modal fusion is adopted, and a feature pyramid network, a dual-path lightweight convolutional neural network with multi-feature fusion, and a parallel multi-branch convolutional neural network technology are introduced to automatically identify termite ecological environment sign information.

[0025] Furthermore, ground-penetrating radar analysis adopts an automatic termite nest identification algorithm based on the coupling of ground-penetrating radar electromagnetic wave characteristics and profile features. It takes the propagation and reflection characteristics of electromagnetic waves in underground media (such as wave velocity attenuation, changes in reflection coefficient caused by differences in dielectric constants, dispersion effects, etc.) as the physical basis, and achieves accurate target identification through algorithm optimization and synergistic combination of feature coupling networks (EDMFEBs), LS-YOLOv5 detection model and YOLACT instance segmentation model.

[0026] The three-dimensional reconstruction of termite nests on embankments adopts a three-dimensional automatic imaging method based on multi-channel ground penetrating radar B-scan data to intelligently extract the nest morphology, location and semantic information, and realize the conversion from two-dimensional radar profile images to three-dimensional stereo modeling.

[0027] Furthermore, the historical treatment database includes process parameters, cost data and treatment effect cases of grouting and plugging, excavation and backfilling, and supports intelligent recommendation of prevention and control plans.

[0028] Furthermore, hidden danger assessment and prevention and control decisions are based on AI intelligent analysis, which enables automatic classification of different levels of termite hidden dangers and recommends optimal prevention and control measures in combination with the historical management database.

[0029] Furthermore, the prevention and control implementation and effect evaluation system adopts multimodal data fusion technology to quantify the treatment effect indicators by comparing the differences in ground penetrating radar data before and after treatment, and establish a full life cycle management system for embankment termite prevention and control.

[0030] Furthermore, the ground penetrating radar data acquisition subsystem has a detection depth of no less than 5 meters, a resolution better than 20 cm, and supports multi-channel data acquisition.

[0031] Furthermore, the ground penetrating radar data analysis unit covers preprocessing operations and intelligent recognition operations on the original electromagnetic wave signals, wherein the data preprocessing operations include denoising, normalization and feature enhancement of the original electromagnetic wave signals;

[0032] The present invention also proposes an intelligent detection and prevention system for termite nests in dams based on the above method, comprising:

[0033] Multimodal data acquisition module: includes a ground penetrating radar data acquisition subsystem and a visual camera data acquisition subsystem, used to acquire multimodal data and support real-time uploading of data to the central processing system;

[0034] The AI ​​analysis module, including a ground-penetrating radar data analysis unit and a visual AI data processing unit, uses deep learning technology to achieve intelligent recognition of termite nests and signs of termite ecological environments, improving recognition accuracy and environmental adaptability;

[0035] The 3D modeling module generates a 3D morphological model of termite nests based on ground-penetrating radar data. Combined with dam geological data and historical termite activity data, it achieves detailed spatial morphology and precise positioning, and has a dynamic update function to adapt to environmental changes.

[0036] Decision-making management module: Based on the full-process system of "perception-evaluation-decision-making-governance", it outputs termite hazard detection results, governance plans and effect evaluation results;

[0037] Furthermore, the decision management platform also includes:

[0038] Dynamic risk assessment module: predicts nest development trends based on historical data and real-time monitoring results;

[0039] Treatment process database: stores grouting material ratios, excavation and backfill process parameters, and cost data, supporting intelligent matching of optimal solutions;

[0040] Visual interactive interface: supports 3D model rotation, potential hazard area marking, and treatment plan simulation, enabling intuitive interaction and intelligent decision-making.

[0041] The beneficial effects of the present invention are as follows:

[0042] This invention uses AI-powered intelligent analysis technology, integrating ground-penetrating radar and visual data, to achieve high-precision detection and intelligent prevention of termite nests. By integrating multimodal data and optimizing decision-making with deep learning algorithms, it overcomes the limitations of traditional manual inspections, which rely on expert experience, have high rates of false positives and missed detections, and exhibit low levels of standardization, thereby improving detection accuracy and stability.

[0043] This invention utilizes deep learning models to automatically identify, accurately locate, and reconstruct termite nests in three dimensions. This allows for automated analysis of ground-penetrating radar signals, reduces data noise, enhances spatial imaging capabilities of nests, and ensures reliable detection in complex environments. Based on a "perception-assessment-decision-control" system, an intelligent termite control and management system is constructed, enabling hierarchical assessment of potential hazards, optimization of control solutions, and long-term monitoring, enhancing the scientific nature and sustainability of control measures.

[0044] Compared with traditional methods, this invention reduces human intervention and improves the standardization and intelligence level of termite nest detection, assessment and management. It can be widely used in water conservancy projects such as dams, reservoirs, and river embankments, improving the accuracy and efficiency of dam safety monitoring, effectively reducing the safety hazards caused by termite infestations, and providing water conservancy facilities with efficient and intelligent termite control solutions.

[0045] The present invention aims to provide a method and system for intelligent detection and prevention of termite nests on dams based on a multimodal AI model, in order to address the problems of low efficiency, high missed detection rate, reliance on expert experience, difficulty in accurate positioning, and lack of intelligent prevention and control decision-making in existing termite nest detection methods. This method combines ground-penetrating radar and visual AI analysis technology to achieve high-precision identification, spatial positioning, and hazard assessment of termite nests through multimodal data acquisition, intelligent analysis, and three-dimensional reconstruction. It also constructs a closed-loop management system for the entire process of "perception-evaluation-decision-making-governance" to achieve intelligent detection, precise prevention, and long-term tracking of termite nests on dams. The present invention can be widely used in the safety monitoring of water conservancy projects such as dams, reservoirs, and riverbanks, to improve the automation, standardization, and intelligence level of termite hazard detection, and to provide scientific and efficient technical support for the long-term stable operation of water conservancy facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a multimodal AI model-based intelligent detection and prevention method for termite nests on embankments provided for the implementation of this application;

[0047] Figure 2 : System block diagram of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] See also Figure 1 The present invention provides a method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model, the method comprising the following steps:

[0050] Step S1, multimodal data acquisition: using ground penetrating radar and video camera to collect ground penetrating radar data and surface visual data to construct a multimodal dataset.

[0051] Surface visual data: Using camera video acquisition equipment, real-time video stream data of termite activity signs such as mud lines, mud covers, cluster holes, fungi, etc. is collected. The specific content is as follows:

[0052] The mud line refers to a tubular or strip-shaped channel constructed by termites using a mixture of saliva, soil, excrement and other materials to maintain the humidity of the nest and avoid natural enemies and external interference. It is characterized by being attached to surfaces such as walls, beams, columns, and pipes, connecting the nest with food sources (such as wood), and having traces of termite activity inside.

[0053] The mud cover refers to the mud protective layer that termites cover the surface of the eroded object to block light and maintain moisture when feeding on wood or plant fibers. It is characterized by being in the form of flakes or blocks, and is commonly found on the surface of wood, under the bark or in soil cracks. After peeling it off, traces of internal decay can be seen.

[0054] The swarming holes are temporary exits through which the reproductive ants (winged adults) fly out of the nest during the swarming season (reproduction and diffusion period) after the termite colony matures. They are characterized by small round holes with a diameter of about 2-5 mm, which are mostly distributed on the surface of walls and wood, and often leave traces of soil sealing after swarming.

[0055] The fungi are: symbiotic fungi cultivated in termite nests (such as Alternaria albuminosa), microbial communities that provide nutrition for termites or decompose wood fibers; characterized by grayish white or light brown spongy fungus beds, or saprophytic fungi that grow on external wood damaged by termites.

[0056] The underground radar data refers to electromagnetic wave data collected by multi-channel ground penetrating radar (detection depth ≥ 5 meters, resolution ≤ 20 cm) along regular survey lines arranged along the surface of the dam.

[0057] Step S2, performing intelligent analysis on the ground penetrating radar data and surface visual data, specifically including:

[0058] Step S21: Perform AI recognition on the surface visual data:

[0059] We use a multi-branch convolutional neural network modal fusion algorithm, combined with a feature pyramid network (FPN) and a dual-path lightweight convolutional network for multi-feature fusion, to extract features from surface visual data. We then combine YOLOv5 object detection with Mask R-CNN instance segmentation to accurately locate termite signs and distribution ranges. The specific process is as follows:

[0060] Step S211, sampling the continuous frames collected by the high-definition camera containing signs of termite activity such as mud lines and mud blankets, and extracting key frames;

[0061] Step S212: performing noise reduction, standardization, and other processing on each key frame to obtain a pre-processed image that meets the model input requirements; the model input requirements refer to image sizes that meet the AI ​​model requirements.

[0062] In step S213, the pre-processed image enters the backbone network. At different stages of the backbone network, the convolutional layer and the pooling layer are used to downsample and extract features and reduce the size (resolution) layer by layer to obtain several layers of feature maps from shallow to deep. Among them, the shallow feature map has a higher resolution (such as 56×56), retaining details such as edges and semantics, but the semantics are weaker; the deep feature map has a lower resolution (such as 7×7), contains global semantics (such as overall concepts such as object categories), and has stronger semantics. Each layer of feature map is no longer an image in the traditional sense, but a high-dimensional tensor (multidimensional array), a mathematical representation that is more abstract and meaningful than the original pixel.

[0063] In step S214, starting from the deepest feature map, upsample the deep feature map to expand its resolution (for example, from 1 / 32 of the original image to 1 / 16), and obtain a feature map of the same size as the next shallow feature map to facilitate feature fusion.

[0064] In step S215, the upsampled feature map is horizontally connected with the shallow features of the corresponding level, and fused by element-by-element addition (for example, the fifth layer features are fused with the shallower fourth layer features).

[0065] Repeat steps S214 and S215 to generate new feature maps layer by layer until the highest resolution layer is reached, so that deep semantic information and shallow spatial information can be expressed at the same time, achieving the complementarity of semantics and details and improving the accuracy of target detection.

[0066] Step S216: Each layer independently runs a region proposal network (RPN) to generate candidate boxes with various aspect ratios (such as 1:1, 1:2, and 2:1). The high-resolution shallow layer generates small-sized anchor boxes to detect small targets, and the low-resolution deep layer generates large-sized anchor boxes to detect large targets.

[0067] Step S217: Dynamically allocate the preset anchor frames. The level selection formula is as follows:

[0068] k=k0+log2(224W×h)

[0069] Where w and h are the normalized detection box width and height (relative to the input image size); k0 is the base level (typically set to 4, corresponding to the fourth level); and k is the feature level number assigned to the object. Dynamic allocation maps continuous object sizes to discrete levels, ensuring that each object is assigned to the most appropriate level and anchor box, improving training stability and detection efficiency.

[0070] In step S218, the anchor frames of each layer output categories (such as clustered holes and fungi) and confidence levels through the classification head (YOLO Head), and then retain high-confidence detection results through non-maximum suppression (NMS).

[0071] Step S219: Map the anchor box containing the category and confidence information back to the original image coordinate system, draw the bounding box, category label, and confidence score on the original image, and detect the target of termite signs in the image. The original image is, for example, the pre-processed image in step S212;

[0072] Step S22, analyzing the ground penetrating radar data, specifically includes:

[0073] In step S221, the electromagnetic wave characteristics and profile features of the ground penetrating radar are coupled to analyze the propagation and reflection characteristics of electromagnetic waves underground, and the algorithm is combined based on the EDMFEBs feature coupling network, the LS-YOLOv5 detection model and the YOLACT instance segmentation model: the feature coupling network serves as the core bridge, and the cross-modal attention mechanism (such as Transformer cross attention) is used to dynamically associate the electromagnetic wave time-frequency features (through wavelet transform, STFT extraction) with the B-scan image texture features (such as GLCM, edge detection). At the same time, LS-YOLOv5 optimizes the positioning and classification accuracy of underground targets through the improved multi-scale feature extraction module (ECSPP, SPD-Conv) and the decoupling detection head, while the mask branch of YOLACT generates instance segmentation contours based on the detection results, and feeds back to the feature fusion layer to enhance morphological analysis (such as soil cracks and layered structures).

[0074] Step S222: performing algorithm optimization, including: adopting a dual-stream network architecture: the electromagnetic wave signal is processed by 1D CNN or Transformer, and the image is subjected to multi-scale feature extraction by an improved LS-YOLOv5 backbone. The two are fused in the neck network through Bi-FPN and the attention mechanism, and jointly optimize detection (GIoU loss), segmentation (Dice loss), and feature consistency loss.

[0075] In step S223, EDMFEBs integrates a convolutional neural network with an attention mechanism, comprising an encoder, a feature extractor, and a decoder. The attention mechanism uses the GCS attention mechanism, which includes two types of attention network blocks: a channel attention network block and a spatial attention network block. The channel attention and spatial attention components of the existing GCS attention mechanism are expanded, and a multi-head attention mechanism is introduced into these modules. By integrating the multi-head attention mechanism into the GCS attention mechanism, the model's feature expression capabilities are significantly enhanced. This is particularly true when dealing with non-uniform and complex underground target inversion tasks, enabling the model to better capture and process feature information at different scales and in different regions.

[0076] The LS-YOLOv5 detection model, targeting the large-scale feature map, uses a deep residual network to better extract disease features. It also uses the deep residual main network feature map to construct feature pyramids at four scales. At a 512×512 input resolution, it reduces the minimum feature map of YOLOv5 from 16×16 to 8×8, thereby giving the network a larger receptive field. It then fuses features with the deep residual network at four scales to form a detection model. The LS-YOLOv5 detection model, targeting the large-scale feature map, uses a deep residual network to better extract disease features. It also uses the deep residual main network feature map to construct feature pyramids at four scales. At a 512×512 input resolution, it reduces the minimum feature map of YOLOv5 from 16×16 to 8×8, thereby giving the network a larger receptive field. It then fuses features with the deep residual network at four scales to form a detection model.

[0077] The YOLACT model is an efficient end-to-end instance segmentation network model that uses ResNet101 as the basic network structure for YOLACT feature extraction. After the feature extraction network, the network is divided into two branches: the mask coefficient network and the fully convolutional neural network. The main task of the mask coefficient network is to predict the mask coefficient and output the classification probability (used to determine the category of the target in the input image), the bounding box probability (used to locate the position of the target in the image), and the mask probability (used to generate the corresponding mask to represent the area of ​​the target). The fully convolutional neural network is responsible for generating the prediction of the prototype mask. The network performs operations through the fully convolutional layer, which can output high-quality mask feature maps while maintaining spatial information.

[0078] Step S3, performing three-dimensional reconstruction of the termite nest, specifically includes:

[0079] In step S31, the spatial distribution characteristics of the ground penetrating radar data are analyzed through the maximum density projection algorithm and the SegNet target imaging network, and a three-dimensional morphological model of the ant nest is constructed to achieve accurate positioning of the ant nest, and the morphological optimization is performed in combination with the historical dam structure information.

[0080] Step S32, based on the multi-channel two-dimensional ground penetrating radar B-scan data, three-dimensional imaging of underground internal diseases is performed. By deeply mining the spatial feature information of the multi-channel ground penetrating radar B-scan data, the complex three-dimensional morphology of the underground ant nest can be finely depicted and the spatial position can be accurately located.

[0081] The method for automatic three-dimensional imaging of internal diseases includes three steps, including projection preprocessing, feature extraction and three-dimensional reconstruction.

[0082] Specifically,

[0083] In step S321, the projection preprocessing module performs multi-angle projection transformation on the original B-scan data, extracts the spatial distribution characteristics of the disease reflection wave through signal enhancement and spatial feature encoding, and generates a two-dimensional projection sequence containing continuous angle information of 0°-180°.

[0084] In step S322, the feature extraction module uses a deep convolutional network to analyze key parameters such as the geometric outline, depth position, and material properties of the defect from projections at various angles, and outputs a multi-perspective two-dimensional defect representation map.

[0085] In step S323, the 3D reconstruction module finally generates a disease model containing 3D geometric topology and spatial coordinates by integrating multi-angle information through spatial registration of feature maps and implicit surface reconstruction algorithm, achieving millimeter-level precision in morphological characterization and spatial positioning.

[0086] Step 4: Based on the 3D reconstruction results of step S3, conduct hazard assessment and prevention and control decisions, and build a closed-loop management system for the entire process of "perception-assessment-decision-making-governance". Specifically, it includes:

[0087] Establish a hazard classification standard based on the "Technical Guidelines for Termite Control in Water Conservancy Projects (Trial)", and make ratings based on data such as signs of the termite ecological environment and nest size;

[0088] Combined with the historical governance database, it intelligently matches the optimal prevention and control plan (including grouting material ratio, excavation and backfilling process parameters and cost data) and intelligently matches the optimal governance plan.

[0089] Furthermore, the method may further include step S5, prevention and control implementation and effect evaluation:

[0090] Ant nest control is carried out according to the prevention and control plan. The embankment structure is re-surveyed through ground penetrating radar, and the data before and after are analyzed in combination with AI models to quantify the control effect, realize long-term tracking and monitoring, and build a full life cycle management system for termite control.

[0091] Based on the above-mentioned method, the present invention also provides a method for intelligent detection and prevention of termite nests in dams. The method also provides a system for intelligent detection and prevention of termite nests in dams, comprising:

[0092] Multimodal data acquisition module: includes ground penetrating radar data acquisition subsystem and visual camera data acquisition subsystem, used to obtain multimodal data and support real-time uploading of data to the central processing system.

[0093] AI analysis module: includes a ground-penetrating radar data analysis unit and a visual AI data processing unit. It uses deep learning technology to achieve intelligent recognition of termite nests and signs of termite ecological environment, improving recognition accuracy and environmental adaptability.

[0094] 3D modeling module: Generates a 3D morphological model of termite nests based on ground-penetrating radar data. Combined with dam geological data and historical termite activity data, it achieves detailed spatial morphology and precise positioning, and has a dynamic update function to adapt to environmental changes.

[0095] Decision-making management module: Based on the full-process system of "perception-evaluation-decision-making-governance", it outputs termite hazard detection results, control plans and effect evaluation results.

[0096] According to one embodiment of the present invention, the dataset construction process is as follows: a multimodal dataset is used, including surface visual data and underground ground penetrating radar data.

[0097] Among them, surface visual data: consists of video camera images and high-resolution images taken by handheld devices, covering signs of termite activity (mud lines, mud blankets, cluster holes, fungi, etc.).

[0098] Ground-penetrating radar data: Collected using multi-channel ground-penetrating radar, the data covers a depth of ≥5 meters and a resolution of ≤20 cm. Radar data undergoes denoising and enhancement processing to optimize signal characteristics.

[0099] Data annotation: Manually annotate termite signs and nest locations, using a semi-automatic annotation method combined with manual correction to improve data quality.

[0100] In one embodiment, the present invention uses three core neural network modules to implement data analysis and detection:

[0101] Visual recognition network: A parallel multi-branch convolutional neural network is used, combined with a feature pyramid network (FPN) and a lightweight two-way convolutional neural network to improve the accuracy of identifying signs of termite ecological environment.

[0102] Radar analysis network: A deep learning model based on the coupling of ground-penetrating radar electromagnetic wave characteristics and profile features is used, combined with the EDMFEBs feature coupling network, LS-YOLOv5 detection model, and YOLACT instance segmentation model to enhance the accuracy of ant nest detection.

[0103] 3D modeling network: Based on the SegNet target imaging network, multi-channel B-scan ground penetrating radar data is integrated, and the maximum intensity projection algorithm and MIP projection preprocessing are used to accurately reconstruct the 3D morphology of termite nests.

[0104] In one embodiment, the training process includes vision model training and ground penetrating radar model training.

[0105] The visual model is trained as follows:

[0106] The dataset is based on 5,000 labeled samples (including mud lines, mud blankets, clustered holes, etc.), and uses transfer learning (pre-trained model is ResNet50) and data augmentation (rotation, noise injection, brightness adjustment) to optimize the model's generalization ability.

[0107] The algorithm deployment process is as follows: using a multi-branch convolutional neural network modal fusion algorithm, a front-end parallel dual-path lightweight convolutional network (channel number 64-256, step size 1-2), and a back-end decision fusion based on DS evidence theory to output termite sign positioning results (accuracy ≥ 95%).

[0108] The radar model training process is as follows:

[0109] First, data preprocessing is performed: the ground penetrating radar B-scan data is denoised (wavelet threshold filtering) and normalized;

[0110] Secondly, the model architecture adopts the EDM FEBs network (encoder-feature extractor-decoder structure), embeds a multi-head GCS attention mechanism (4 heads each for channel and spatial attention branches), and combines it with the LS-YOLOv5 detection model (input resolution 512×512, feature pyramid 4-scale fusion) to achieve ant nest recognition (accuracy ≥ 90%).

[0111] The testing process is as follows:

[0112] Model validation: The mAP (mean average precision) of termite ecological sign detection, ant nest 3D localization error, and IoU (intersection over union) of ground penetrating radar analysis were evaluated on the test set.

[0113] Real-time testing: Deploy the model in a real dam environment to evaluate its adaptability under different weather, lighting, and terrain conditions.

[0114] Decision-making evaluation: Using intelligent decision-making systems, we evaluate termite nests of different severity levels and generate the best prevention and control plan.

[0115] Through the above-mentioned complete implementation scheme, the present invention achieves precise detection, three-dimensional reconstruction, and intelligent control of termite nests, improving the efficiency and reliability of dam safety monitoring. It utilizes deep learning algorithms to analyze ground-penetrating radar electromagnetic wave data and video image data, intelligently identifying signs of termite ecological environments, accurately locating dam termite nests, and constructing three-dimensional morphological models of the nests, thereby improving nest identification accuracy. A "perception-evaluation-decision-governance" system is constructed, combining historical governance databases to intelligently match control plans and optimize governance strategies. After control is implemented, ground-penetrating radar re-measurements and multimodal data comparisons enable quantitative evaluation of control effects and long-term tracking monitoring, establishing a full-lifecycle management system for termite control. This invention overcomes the shortcomings of traditional manual inspection methods, such as low efficiency, high missed detection rates, and reliance on expert experience, while improving detection accuracy and intelligence. It is applicable to water conservancy projects such as reservoirs and riverbanks, and enhances the scientific nature and efficiency of monitoring and control of termite hazards on dams.

[0116] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, and it should be clear that the present invention is not limited to the scope of the specific embodiments, it is obvious to those skilled in the art that as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

Claims

1. A method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model, characterized in that: The following steps are involved: Step S1, multimodal data acquisition: using ground penetrating radar and video camera to collect ground penetrating radar data and surface visual data to construct a multimodal dataset; Step S2, performing intelligent analysis on the ground penetrating radar data and surface visual data, specifically including: Step S3, performing three-dimensional reconstruction of the termite nest; Step 4: Based on the 3D reconstruction results of step S3, conduct hazard assessment and prevention and control decisions, and build a closed-loop management system for the entire process of "perception-assessment-decision-making-governance".

2. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: The step S2 comprises: Step S21: Perform AI recognition on the surface visual data: Step S22: analyzing the ground penetrating radar data.

3. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: The step S3 comprises: In step S31, the spatial distribution characteristics of the ground penetrating radar data are analyzed through the maximum density projection algorithm and the SegNet target imaging network, and a three-dimensional morphological model of the ant nest is constructed to achieve accurate positioning of the ant nest, and the morphological optimization is performed in combination with the historical dam structure information. Step S32, based on the multi-channel two-dimensional ground penetrating radar B-scan data, three-dimensional imaging of underground internal diseases is performed. By deeply mining the spatial feature information of the multi-channel ground penetrating radar B-scan data, the complex three-dimensional morphology of the underground ant nest can be finely depicted and the spatial position can be accurately located.

4. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: It also includes step S5, prevention and control implementation and effect evaluation, executing anthill control according to the prevention and control plan, re-surveying the embankment structure through ground penetrating radar, and combining AI model to analyze the before and after data, quantifying the control effect, realizing long-term tracking and monitoring, and building a full life cycle management system for termite control.

5. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: In step S1, multimodal data collection is performed using cameras and ground penetrating radar equipment to collect termite ecological environment signs on the surface of the dam and underground electromagnetic wave signals, thereby constructing a multimodal dataset including surface signs and underground structures.

6. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: In step S21, the ground penetrating radar data and the surface visual data are intelligently analyzed, and a multi-branch convolutional neural network modal fusion algorithm is used, combined with a feature pyramid network and a dual-path lightweight convolutional network, to identify signs of termite activity such as mud sheets, mud lines, cluster holes, and fungi; By adopting a deep learning model based on the coupling of ground penetrating radar electromagnetic wave characteristics and profile features, and integrating incremental learning and transfer learning techniques, high-precision identification of termite nests can be achieved under small sample conditions.

7. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: In step S3, the three-dimensional reconstruction of the termite nest is based on the SegNet target imaging network and combined with the spatial distribution characteristics of the ground penetrating radar data to generate a three-dimensional morphological model of the termite nest, thereby achieving accurate positioning of the termite nest.

8. The method for intelligent detection and prevention of termite nests on dams based on a multimodal AI model according to claim 1, characterized in that: In multimodal data collection, termite ecological environment sign data include mud lines, mud covers, termite flight holes, termite nest indicators and termite-eaten food.

9. The method according to claim 1, characterized in that In step S2, a termite ecological environment sign recognition algorithm based on a multi-branch convolutional neural network modal fusion based on vision is adopted, which introduces a feature pyramid network, a dual-path lightweight convolutional neural network with multi-feature fusion, and a parallel multi-branch convolutional neural network technology to automatically identify termite ecological environment sign information.

10. An intelligent detection and prevention system for termite nests on dams based on the above method, characterized in that: include: Multimodal data acquisition module: includes a ground penetrating radar data acquisition subsystem and a visual camera data acquisition subsystem, used to acquire multimodal data and support real-time uploading of data to the central processing system; The AI ​​analysis module, including a ground-penetrating radar data analysis unit and a visual AI data processing unit, uses deep learning technology to achieve intelligent recognition of termite nests and signs of termite ecological environments, improving recognition accuracy and environmental adaptability; The 3D modeling module generates a 3D morphological model of termite nests based on ground-penetrating radar data. Combined with dam geological data and historical termite activity data, it achieves detailed spatial morphology and precise positioning, and has a dynamic update function to adapt to environmental changes. Decision-making management module: Based on the full-process system of "perception-evaluation-decision-making-governance", it outputs termite hazard detection results, control plans and effect evaluation results.

Images