An ai-based mountain disaster scenario library and intelligent protection recommendation method and system

Abstract

The application discloses a mountain disaster scene library and intelligent protection recommendation method and system based on AI, and relates to the technical field of mountain geological disaster risk assessment.The application integrates terrain, remote sensing, monitoring, engineering and hazard-affected body multi-source data, and forms an integrated data base through unified coordinate and coding annotation standardization processing; adopts an improved U-Net with attention and a YOLOv8n lightweight model to automatically extract complete features of disasters and hazard-affected bodies, distinguish spatial relationships and build standardized risk scene units, and build a hierarchical mountain disaster risk scene library; after new monitoring data drives the generation of new scenes, similar cases are retrieved, and multiple candidate schemes are generated in combination with a protection device parameter library; adaptive weight-based energy dissipation and protection adaptation quantitative indexes are introduced to comprehensively score, and the optimal parameter layout scheme is output according to a five-level protection chain; and the scene library and the protection parameter library are continuously iteratively updated after running feedback.The application realizes intelligent decision-making of the whole process of mountain disaster protection, and has wide application value.

Description

Technical Field

[0001] This invention relates to the field of mountain geological disaster risk assessment technology, specifically to an AI-based mountain disaster scenario database and intelligent protection recommendation method and system. Background Technology

[0002] With the continuous expansion of construction scale in plateau mountainous areas, canyon transportation corridors, and major field engineering construction areas in my country, mountain disasters such as high-altitude collapses, landslides, debris flows, and avalanches are occurring frequently, directly threatening the safety of various disaster-bearing structures such as construction camps, highways, tunnels, and energy and communication facilities. Currently, the industry has conducted extensive technical research on disaster identification, risk classification, and optimization of protective devices. Air-ground integrated observation technology is rapidly becoming widespread, and the stock of multi-source geographic data, including DEM, UAV 3D models, InSAR deformation data, multi-phase remote sensing, and on-site surveys, continues to grow. Artificial intelligence image segmentation and target detection algorithms are gradually being applied to the automatic identification of geological hazards, and various risk assessment, disaster early warning, and protective structure optimization technologies are being implemented extensively, enabling the extraction of localized hazards and the recommendation of mitigation measures to a certain extent.

[0003] Chinese patent (publication number CN114022765A) discloses an intelligent monitoring method and system for landslides, collapses and falling rocks using image recognition. This method identifies dangerous rocks and falling rocks on slopes through an optical image target detection model and achieves early warning of danger by delineating the scope of the danger. However, this solution only identifies the disaster itself and does not simultaneously identify disaster-bearing objects such as roads and camps. It cannot quantify the coupled risk relationship between the disaster and the disaster-bearing objects, and only outputs early warning signals without having the function of intelligent matching of protection schemes.

[0004] Chinese patent (publication number CN120087851A) discloses a cascaded learning method for assessing the susceptibility of flash floods and debris flows. This method selects topographic and hydrological indicators to construct a machine learning assessment model and outputs the results of watershed disaster zoning. However, this method only completes the static assessment of hazard and cannot automatically generate multi-level barrier, diversion, and energy dissipation combination deployment parameters based on risk intensity. The protection selection relies entirely on manual experience.

[0005] Chinese patent (publication number CN119377769A) discloses a method and device for predicting geological disasters in plateau and mountainous areas based on millimeter-wave SAR. This method integrates SAR imagery and DEM topographic data to extract disaster hazard features and classifies disaster risk levels through feature thresholds. However, the data source coverage of this method is limited, and it lacks a unified and standardized processing flow for multiple types of data such as point clouds, GNSS, snow and ice, and engineering cases. It also lacks a standardized risk scenario storage system and cannot achieve the retrieval and reuse of historical risk cases.

[0006] In addition, existing technologies generally have shortcomings in data fusion, object coupling, scenario reuse, and collaborative optimization of protection parameters, making it difficult to form a complete technical link from multi-source data to the automatic generation of multi-level protection schemes, and failing to meet the actual needs of refined and automated disaster prevention design for major high-altitude canyon projects. Summary of the Invention

[0007] To address the aforementioned technical issues, this application discloses an AI-based method and system for intelligent recommendation of mountain disaster scenarios and protection measures. The method specifically comprises:

[0008] Acquire multi-source data of the target area, and perform coordinate unification, spatial registration, scale normalization, data cleaning, object encoding and semantic annotation on the data to construct a multi-source object data foundation;

[0009] Based on the multi-source object data base, AI disaster object identification and AI disaster-bearing body object identification are carried out respectively to obtain the characteristics of disaster objects and disaster-bearing body objects.

[0010] Based on the characteristics of the disaster objects and the disaster-bearing bodies, the spatial relationship between the disaster objects and the disaster-bearing bodies is identified, and corresponding risk scenario units are generated. Multiple risk scenario units are organized according to disaster type, disaster-bearing body type, spatial relationship type and risk level to construct a mountain disaster risk scenario library.

[0011] When new monitoring data is acquired, corresponding updated risk scenario units are generated. Based on the updated risk scenario units, similar risk scenarios are retrieved from the mountain disaster risk scenario database, and candidate protective device parameter combinations are generated in combination with the protective device parameter database.

[0012] Based on the parameter combinations of each candidate protective device, the corresponding energy dissipation adaptability and protection adaptability are calculated. Based on the energy dissipation adaptability and protection adaptability, the target protective device parameter combination is selected.

[0013] Based on the parameter combination of the target protection device, the protection deployment is implemented, and the mountain disaster risk scenario database and protection device parameter database are updated according to expert verification, on-site review, engineering implementation effect and post-disaster review results.

[0014] Preferably, the construction of the multi-source object data base specifically involves: after standardizing the multi-source data through coordinate unification, spatial registration, scale normalization, data cleaning, object encoding, and semantic annotation, constructing an integrated data set that combines mountain disaster-related data and disaster-bearing body-related data to obtain the multi-source object data base.

[0015] Preferably, the disaster object features are specifically defined as follows: the disaster-related data in the multi-source object data base is cropped into samples of a fixed size, and data augmentation processing is performed. The augmentation methods include random rotation, flipping, scaling, and brightness adjustment to generate training sets, validation sets, and test sets.

[0016] An improved U-Net model is adopted, and an attention mechanism module is added between the encoder and decoder of the U-Net architecture. The feature map extracted by the encoder is multiplied with the attention weight to obtain a weighted feature map, which is then input into the decoder for upsampling.

[0017] The preprocessed disaster data of the target area is input into the trained improved U-Net model, which outputs the segmentation results of the disaster objects. The features of the disaster objects are extracted by combining spatial coordinate information. The features of the disaster objects include disaster type, source area location, geometric range, potential movement direction and disaster energy level.

[0018] Preferably, the AI ​​disaster-bearing object identification specifically involves: using the YOLOv8n lightweight target detection model to annotate the disaster-bearing object data in the multi-source object data base, with the annotation content including the disaster-bearing object type and spatial location coordinates, generating a VOC format annotation file, and constructing a standardized dataset;

[0019] The labeled disaster-bearing body data is divided into training, validation, and test sets, which are then input into the YOLOv8n lightweight target detection model for training. The trained YOLOv8n lightweight target detection model outputs disaster-bearing body type, spatial location, geometric range, personnel exposure characteristics, structural vulnerability, functional importance, portability, usage period, emergency accessibility, and recognition reliability.

[0020] Preferably, the construction of the mountain disaster risk scenario database specifically involves: assigning a unique scenario code to each risk scenario unit, with the scenario coding rule being disaster type code - disaster-bearing body type code - spatial relationship code - risk level code; constructing a B+ tree index for the scenario code, disaster type, and disaster-bearing body type to achieve fast multi-condition retrieval and complete the construction of the mountain disaster risk scenario database.

[0021] Preferably, the step of retrieving similar risk scenarios from the mountain disaster risk scenario database specifically involves using a retrieval method that combines preliminary screening, precise matching, and knowledge graph assistance.

[0022] The preliminary screening filters the risk scenario units in the mountain disaster risk scenario database according to disaster type, disaster-bearing body type, risk level, and spatial relationship type, and selects candidate scenario units that match the basic characteristics of the updated risk scenario units.

[0023] The updated risk scenario units' disaster object characteristics, disaster-bearing body characteristics, spatial relationships, risk levels, and disaster energy levels are normalized, generating dimensions as follows: eigenvectors The corresponding features of the selected candidate scene units are also normalized to generate feature vectors. ;

[0024] Calculate the cosine similarity between the updated risk scenario unit feature vector X and the feature vector of each candidate scenario unit. The calculation formula is as follows:

[0025]

[0026] in, For the updated risk scenario unit feature vector With the scene library Feature vectors of candidate scene units Cosine similarity; The dimension of the feature vector; For the updated risk scenario unit feature vector The Middle Normalized values ​​of each feature; For the first Feature vectors of candidate scene units The Middle Normalized values ​​of each feature; , They are the feature vectors The modulus length;

[0027] A preset similarity threshold is used to filter out risky scenario units with a cosine similarity greater than or equal to the threshold, classifying them as similar risky scenarios; when the number of similar risky scenarios exceeds the preset number of risky scenarios... When, take the top of the similarity ranking. Each scenario unit is selected as the final similar risk scenario; at the same time, knowledge graph reasoning is used to assist in judging the rationality of matching similar scenarios and eliminate scenario units with large matching deviations.

[0028] Preferably, the calculation of the corresponding energy dissipation adaptability and protection adaptability specifically involves: based on the combination of protective device parameters in similar risk scenarios, evaluating the comprehensive ability of the protective device parameter combination to intercept, decelerate, divert, disperse, buffer, control the accumulation of disaster energy, and reduce residual energy through the energy dissipation adaptability. The calculation formula is as follows:

[0029]

[0030] in, Energy dissipation adaptability (value range 0-100%, the higher the value, the better the energy dissipation effect). The initial total energy of the disaster. This refers to the remaining energy after a disaster has passed through multiple levels of protective devices. The total energy dissipation capacity of the entire protective device, For the first The energy dissipation capacity of a single protective device For the first The number of protective devices deployed is calculated by cumulatively adding up the energy dissipation capacity of multiple levels of protective devices.

[0031] Protection adaptability is used to evaluate the protective effect of the combination of protective device parameters on the safety of the disaster-bearing body, as well as the degree of matching between engineering accessibility, construction difficulty, implementation period, maintenance conditions, cost constraints, and emergency response capabilities. The formula is:

[0032]

[0033] in, To ensure compatibility, The geographical adaptability coefficient for the deployment location of the protective device. The structural adaptation coefficient of the protective device. The feasibility coefficient for the construction of the protective device. This is the engineering cost coefficient. For later operation and maintenance coefficients, As the emergency rescue accessibility coefficient, , , , , , The weights are assigned based on the magnitude of each item's impact.

[0034] Preferably, the candidate protective device parameter combination is specifically configured as follows: based on energy dissipation adaptability and protection adaptability, a weighted comprehensive scoring method is used to score and screen each candidate protective device parameter combination. If the disaster-bearing body corresponding to the updated risk scenario unit has the conditions for relocation, the protective combination that meets the requirements of relocation site selection, restricted area delineation, monitoring and early warning, and evacuation routes is output first. If the relocation conditions are not met, the candidate protective device parameter combination with the highest comprehensive score is selected as the target protective device parameter combination. If there are multiple combinations with the same and highest comprehensive score, the combination with the lowest overall engineering cost is selected first.

[0035] Preferably, the updating of the mountain disaster risk scenario database and the protective device parameter database specifically includes updating the mountain disaster risk scenario database and the protective device parameter database.

[0036] Among them, the update of the mountain disaster risk scenario database is achieved by collecting expert verification opinions, on-site review data, engineering implementation effects and post-disaster review results. The database is supplemented with the risk scenario units involved by correcting information and adding new risk scenario units, while invalid and outdated risk scenario units are deleted. The risk level, evolution information and implementation effect of protection plans of the scenario units are updated simultaneously.

[0037] The update of the protective device parameter library will supplement the actual implementation parameters of the target protective device parameter combination, the feedback information on the engineering implementation effect, and the energy dissipation capacity and protection effect data measured after the disaster review into the parameter library. It will correct the energy dissipation capacity, adaptability coefficient, cost coefficient, operation and maintenance coefficient and other parameters of various protective devices, optimize the recommendation rules for protective device parameter combinations, and supplement the optimal parameter range for the deployment of protective devices in different scenarios.

[0038] The system includes, specifically:

[0039] It includes a multi-source data governance module, a dual-target AI recognition module, a risk scenario library construction module, a protection combination generation module, a protection solution intelligent optimization module, and a scenario library;

[0040] The multi-source data governance module stores DEM, DSM, remote sensing, UAV 3D, deformation monitoring, meteorological and hydrological data, disaster cases, and disaster-bearing body spatial multi-source datasets. It has built-in standardized processing operator set, unified coding rule library and semantic annotation label library, and outputs multi-source object data base.

[0041] The dual-target AI recognition module includes an improved U-Net disaster recognition submodule and a YOLOv8n disaster-bearing body recognition submodule. The two submodules are equipped with a preset hierarchical convolutional network, a corresponding loss function and training parameters, and are connected in parallel to a multi-source object data base to output disaster object features and disaster-bearing body object features.

[0042] The risk scenario library construction module has a built-in spatial relationship determination operator, an octet risk scenario unit generator, a B+ tree index engine, and a hierarchical sub-library storage medium to store and structurally manage all risk scenario units.

[0043] The protection combination generation module has built-in feature vector normalization operator, cosine similarity calculation unit, knowledge graph reasoning engine, and seven categories of protection device parameter library storage. It reads new monitoring data to generate and update scene unit and outputs multiple sets of candidate protection device parameter combinations.

[0044] The intelligent optimization module for the protection scheme has a built-in energy dissipation adaptability solving unit, an adaptive weight protection adaptability solving unit, a weighted comprehensive scoring calculation unit, and a five-level protection chain layout scheduler, which outputs parameterized energy dissipation combinations of the target protection device;

[0045] The scenario library has a built-in multi-source feedback data access interface, parameter correction operator, and scenario unit addition, deletion, modification and storage controller, which communicate bidirectionally with the risk scenario library construction module and the protection device parameter library storage.

[0046] Compared with the prior art, the technical solution of this application has the following technical effects:

[0047] This invention establishes a standardized multi-source object data foundation, which uniformly completes coordinate registration, scale normalization, unified coding and semantic annotation of multi-source air-ground observation, engineering ledger and monitoring time series data, and realizes structured and reusable storage of all types of mountain disaster-related data.

[0048] This invention employs two independent AI models, improved U-Net and YOLOv8n, to simultaneously identify disaster objects and disaster-bearing objects. It fully extracts the spatial, energy, and vulnerability multidimensional structured features of the two types of objects, enabling the synchronous output of the coupled features of the disaster-causing body and the disaster-affected body, thus supporting refined coupled risk analysis.

[0049] This invention constructs a standardized octet risk scenario unit and builds a mountain disaster risk scenario library containing eleven hierarchical sub-libraries. It achieves rapid matching of similar historical scenarios through a joint retrieval mechanism combining B+ tree indexing, cosine similarity, and knowledge graph, and improves the rationality of protection recommendations by combining historical engineering experience.

[0050] This invention sets up dual quantitative evaluation indicators of energy dissipation adaptability and protection adaptability with adaptive weights. It quantifies the advantages and disadvantages of various protection combinations from two dimensions: disaster energy dissipation and engineering implementation adaptability. The quantitative indicators can objectively distinguish the protection effects of different multi-level protection chains.

[0051] This invention establishes a full-link iterative update mechanism, which uses feedback from multiple channels such as expert verification, on-site review, and post-disaster debriefing to synchronously update the scenario library, protection parameter library, AI model, and adaptation calculation rules, so as to achieve continuous iterative optimization of the system with engineering practice and improve the accuracy of protection recommendations in the long term.

[0052] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0053] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0055] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0056] Figure 1 This is a flowchart of an AI-based mountain disaster scenario database and intelligent protection recommendation method.

[0057] Figure 2 This is an architecture diagram of an AI-based mountain disaster scenario database and intelligent protection recommendation method;

[0058] Figure 3 This is an architecture diagram of the improved U-Net model in this application;

[0059] Figure 4 This is an architecture diagram of the YOLOv8n lightweight model in this application;

[0060] Figure 5 This is an architecture diagram of an AI-based mountain disaster scenario database and intelligent protection recommendation system.

[0061] Figure 6 This is a cross-sectional schematic diagram of the high-altitude rockfall risk scenario and the deployment of protective devices in the embodiments of this application. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0063] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0064] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0065] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0066] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0067] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0068] Example 1 mainly describes an AI-based mountain disaster scenario database and intelligent protection recommendation method, such as... Figure 1 As shown, it specifically includes:

[0069] Acquire multi-source data of the target area, and perform coordinate unification, spatial registration, scale normalization, data cleaning, object encoding and semantic annotation on the data to construct a multi-source object data foundation;

[0070] Based on the multi-source object data base, AI disaster object identification and AI disaster-bearing body object identification are carried out respectively to obtain the characteristics of disaster objects and disaster-bearing body objects.

[0071] Based on the characteristics of the disaster objects and the disaster-bearing bodies, the spatial relationship between the disaster objects and the disaster-bearing bodies is identified, and corresponding risk scenario units are generated. Multiple risk scenario units are organized according to disaster type, disaster-bearing body type, spatial relationship type and risk level to construct a mountain disaster risk scenario library.

[0072] When new monitoring data is acquired, corresponding updated risk scenario units are generated. Based on the updated risk scenario units, similar risk scenarios are retrieved from the mountain disaster risk scenario database, and candidate protective device parameter combinations are generated in combination with the protective device parameter database.

[0073] Based on the parameter combinations of each candidate protective device, the corresponding energy dissipation adaptability and protection adaptability are calculated. Based on the energy dissipation adaptability and protection adaptability, the target protective device parameter combination is selected.

[0074] Based on the parameter combination of the target protection device, the protection deployment is implemented, and the mountain disaster risk scenario database and protection device parameter database are updated according to expert verification, on-site review, engineering implementation effect and post-disaster review results.

[0075] Furthermore, the multi-source data includes at least one of the following: DEM data, DSM data, remote sensing imagery, UAV orthophotos, UAV oblique photogrammetry 3D models, point cloud data, InSAR surface deformation monitoring data, GNSS surface deformation monitoring data, meteorological and hydrological data, snow and ice cover data, field survey data, historical disaster cases, protective engineering cases, and spatial data of disaster-bearing bodies.

[0076] Furthermore, a multi-source object data foundation is constructed. Specifically, after standardizing the multi-source data through coordinate unification, spatial registration, scale normalization, data cleaning, object encoding, and semantic annotation, an integrated data set that combines mountain disaster-related data and disaster-bearing body-related data is constructed, thus obtaining the multi-source object data foundation.

[0077] Furthermore, the multi-source object data foundation features data standardization, object structuring, and semantic association, which can provide unified, standardized, and callable data support for AI disaster object identification, AI disaster-bearing object identification, and subsequent risk scenario construction.

[0078] The coordinates are uniformly based on the WGS-84 geographic coordinate system, with spatial registration errors controlled within 0.5 meters. The scale normalization adopts a 1:5000 scale standard. Data cleaning removes outliers, missing values, and duplicate data. Object encoding adopts a 16-bit encoding rule of data type encoding - region encoding - time encoding - sequence number encoding. Semantic annotation combines the industry standard for mountain disasters and the classification standard for disaster-bearing bodies to annotate the disaster objects, disaster-bearing body objects, and related attributes in the data, thereby realizing the semantic expression of the data.

[0079] Furthermore, the disaster object characteristics are specifically defined as follows: disaster-related data in the multi-source object data base are cropped into samples of fixed size, and data augmentation processing is performed. Augmentation methods include random rotation, flipping, scaling, and brightness adjustment to generate training sets, validation sets, and test sets.

[0080] An improved U-Net model is adopted, and an attention mechanism module is added between the encoder and decoder of the U-Net architecture. The feature map extracted by the encoder is multiplied with the attention weight to obtain a weighted feature map, which is then input into the decoder for upsampling.

[0081] The preprocessed disaster data of the target area is input into the trained improved U-Net model, which outputs the segmentation results of the disaster objects. The features of the disaster objects are extracted by combining spatial coordinate information. The features of the disaster objects include disaster type, source area location, geometric range, potential movement direction and disaster energy level.

[0082] Furthermore, such as Figure 3 The diagram shows the architecture of the improved U-Net model. The improved U-Net model adopts a symmetrical encoder-decoder hierarchical structure. Considering the feature extraction requirements of mountain disaster remote sensing images and 3D point cloud data, it sets up 5 encoder layers and 5 decoder layers. Feature fusion is achieved between the encoders and decoders through skip connections. An attention mechanism module is added between each encoder and decoder layer. The specific hierarchical structure parameters are as follows:

[0083] The first layer encoder consists of two 3×3 convolutional layers (64 kernels, stride 1, padding 1), one batch normalization (BN) layer, and one ReLU activation function layer, followed by a 2×2 max pooling layer (stride 2, padding 0) to achieve downsampling.

[0084] The second encoder layer consists of two 3×3 convolutional layers (128 kernels, stride 1, padding 1), one BN layer, and one ReLU activation function layer, followed by a 2×2 max pooling layer.

[0085] The third encoder layer consists of two 3×3 convolutional layers (256 kernels, stride 1, padding 1), one BN layer, and one ReLU activation function layer, followed by a 2×2 max pooling layer.

[0086] The fourth encoder layer contains two 3×3 convolutional layers (512 kernels, stride 1, padding 1), one BN layer, and one ReLU activation function layer, followed by a 2×2 max pooling layer.

[0087] The fifth encoder layer is the deepest layer of the model, containing two 3×3 convolutional layers (1024 kernels, stride 1, padding 1), one BN layer, and one ReLU activation function layer, with no pooling layer.

[0088] The first decoder layer contains a 2×2 transposed convolutional layer (512 kernels, stride 2, padding 0) for upsampling. After fusing with the feature maps from the corresponding encoder layer, it is connected to two 3×3 convolutional layers (512 kernels, stride 1, padding 1), one batch normalization (BN) layer, and one ReLU activation function layer. The second decoder layer contains a 2×2 transposed convolutional layer (256 kernels, stride 2, padding 0). After feature fusion, it is connected to two 3×3 convolutional layers (256 kernels, stride 1, padding 1), one batch normalization (BN) layer, and one ReLU activation function layer. The third decoder layer contains a 2×2 transposed convolutional layer (128 kernels). The first layer is a 2×2 transposed convolutional layer (64 kernels, stride 2, padding 0). After feature fusion, it is connected to two 3×3 convolutional layers (128 kernels, stride 1, padding 1), one BN layer, and one ReLU activation function layer. The second layer is a 2×2 transposed convolutional layer (64 kernels, stride 2, padding 0). After feature fusion, it is connected to two 3×3 convolutional layers (64 kernels, stride 1, padding 1), one BN layer, and one ReLU activation function layer. The third layer is the output layer, which contains a 1×1 convolutional layer (the number of kernels is the same as the number of disaster object categories, stride 1, padding 0). It uses the Sigmoid activation function to output the disaster object segmentation results.

[0089] Furthermore, the improved U-Net model uses the cross-entropy loss function as the loss function during training, with the following formula:

[0090]

[0091] in, The loss value for the improved U-Net model. The total number of samples, For the true labels of the samples, The model predicts labels; the Adam optimizer is used, the learning rate is set to 0.001, the training epochs are 100, and an early stopping strategy is adopted, which stops training when the accuracy on the validation set does not improve for 10 consecutive epochs.

[0092] Furthermore, the AI-based disaster-bearing object identification specifically involves: using the YOLOv8n lightweight target detection model to annotate the disaster-bearing object data in the multi-source object data base. The annotation content includes the disaster-bearing object type and spatial location coordinates, generating VOC format annotation files, and constructing a standardized dataset.

[0093] The labeled disaster-bearing body data is divided into training, validation, and test sets, which are then input into the YOLOv8n lightweight target detection model for training. The trained YOLOv8n lightweight target detection model outputs disaster-bearing body type, spatial location, geometric range, personnel exposure characteristics, structural vulnerability, functional importance, portability, usage period, emergency accessibility, and recognition reliability.

[0094] Furthermore, such as Figure 4 The diagram shows the architecture of the YOLOv8n lightweight model, which adopts a three-segment architecture consisting of a backbone network, a neck network, and a detection head.

[0095] The backbone network uses a lightweight improved version of CSPDarknet53, which contains 5 convolutional modules and 4 C2f modules. The specific parameters are as follows: the first layer is a Conv module, which contains 64 3×3 convolutional kernels, stride 2, padding 1, and connects the BN layer and SiLU activation function to achieve downsampling (the output feature map is 1 / 2 of the input).

[0096] The second layer is a Conv module, which contains 128 3×3 convolutional kernels with a stride of 2 and padding of 1. It connects the BN layer and the SiLU activation function and downsamples to 1 / 4 of the input.

[0097] The third layer is the C2f module, which contains 128 convolutional kernels and has two branches. Each branch contains a 3×3 convolutional layer, a BN layer and a SiLU activation function to achieve feature extraction and fusion.

[0098] The fourth layer is a Conv module, which contains 256 3×3 convolutional kernels with a stride of 2 and padding of 1. It connects the BN layer and the SiLU activation function and downsamples to 1 / 8 of the input.

[0099] The fifth layer is the C2f module, which contains 256 convolutional kernels and has the same internal structure as the third layer.

[0100] The 6th layer is a Conv module, containing 512 3×3 convolutional kernels with a stride of 2 and padding of 1. It connects the BN layer and the SiLU activation function and downsamples to 1 / 16 of the input.

[0101] The 7th layer is the C2f module, which contains 512 convolutional kernels and has the same internal structure as the 3rd layer.

[0102] The 8th layer is a Conv module, containing 1024 3×3 convolutional kernels with a stride of 2 and padding of 1. It connects the BN layer and the SiLU activation function and downsamples to 1 / 32 of the input.

[0103] The 9th layer is the C2f module, which contains 1024 convolutional kernels. Its internal structure is the same as the 3rd layer, and it serves as the output layer of the backbone network.

[0104] The Neck Network adopts the PANet structure to achieve multi-scale feature fusion. It includes 3 upsampling modules, 3 downsampling modules, and 3 C2f modules. The specific parameters are as follows: upsampling uses 2×2 transposed convolution (stride 2, padding 0), downsampling uses 3×3 convolution (stride 2, padding 1), and the number of convolution kernels in each C2f module is 512, 256, and 128, respectively, to achieve the fusion of feature maps at three scales: P3, P4, and P5, and adapt to the identification of disaster-bearing bodies of different sizes.

[0105] The head employs an anchorless decoupled head, separating the classification branch from the regression branch. The classification branch contains a 3×3 convolutional layer, a BN layer, a SiLU activation function, and a 1×1 convolutional layer (the number of convolutional kernels is consistent with the number of disaster-bearing body categories). The regression branch contains a 3×3 convolutional layer, a BN layer, a SiLU activation function, and a 1×1 convolutional layer (the number of convolutional kernels is 4, corresponding to the target bounding box coordinates). It outputs the category probability and location coordinates of the disaster-bearing body.

[0106] Furthermore, the YOLOv8n lightweight model uses the CIoU loss function as the bounding box regression loss function, and the calculation formula is as follows:

[0107]

[0108] in, The loss value for the YOLOv8n lightweight model. For prediction boxes With real frame The intersection and union ratio, The squared Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box is given. Let be the diagonal length of the smallest bounding rectangle that encloses the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. To measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box, the learning rate was set to 0.0005, the training batch size to 32, and the image input size to 640×640. The training continued until the model converged.

[0109] Furthermore, identifying the spatial relationship between the disaster object and the disaster-bearing body specifically involves identifying eight types of engineering-specific spatial relationships: direct impact, valley diversion, slope toe diffusion, road cutting, accumulation and burial, long-range impact, indirect functional interruption, and chain transmission. These relationships are identified using spatial topology analysis methods, and the formula for determining spatial topology relationships is as follows:

[0110]

[0111] in, For the spatial geometric set of disaster objects, For the spatial geometric set of disaster-bearing objects, This is a function for determining topological relationships. This is a result of spatial relationships;

[0112] By combining spatial relationship results, disaster energy levels, and the vulnerability of disaster-bearing bodies, standardized risk scenario units are generated. These risk scenario units are represented using an octet structure. ,in As a characteristic of the disaster itself, Characteristics of the disaster-bearing body Features of spatial relationships As a characteristic of the risk boundary, As a disaster energy characteristic, To expose vulnerability, For the combination of parameters of candidate protective devices, To identify credibility.

[0113] Furthermore, the construction of the mountain disaster risk scenario database involves: assigning a unique scenario code to each risk scenario unit, with the scenario coding rule being disaster type code - disaster-bearing body type code - spatial relationship code - risk level code; constructing a B+ tree index for scenario codes, disaster types, and disaster-bearing body types to achieve fast multi-condition retrieval, thus completing the construction of the mountain disaster risk scenario database.

[0114] Furthermore, the risk scenario library contains 11 hierarchical sub-libraries, namely the disaster ontology sub-library, the disaster-bearing body sub-library, the spatial relationship sub-library, the risk boundary sub-library, the disaster energy sub-library, the exposure vulnerability sub-library, the protective device parameter sub-library, the protective device combination sub-library, the similar case sub-library, the recommendation rule sub-library, and the feedback update sub-library.

[0115] Each sub-database independently stores the characteristics, parameters, historical cases, and expert feedback data of the corresponding object. Among them, the disaster ontology sub-database stores information on various types of mountain disaster objects, the disaster-bearing body sub-database stores information on various types of disaster-bearing bodies, the protective device parameter sub-database stores information on the type and parameters of protective devices, and the feedback update sub-database stores feedback data such as expert verification and engineering implementation effects.

[0116] Furthermore, the search for similar risk scenarios from the mountain disaster risk scenario database specifically involves: using a preliminary screening, precise matching, and knowledge graph-assisted search method;

[0117] The preliminary screening filters the risk scenario units in the mountain disaster risk scenario database according to disaster type, disaster-bearing body type, risk level, and spatial relationship type, and selects candidate scenario units that match the basic characteristics of the updated risk scenario units.

[0118] The updated risk scenario units' disaster object characteristics, disaster-bearing body characteristics, spatial relationships, risk levels, and disaster energy levels are normalized, generating dimensions as follows: eigenvectors The corresponding features of the selected candidate scene units are also normalized to generate feature vectors. ;

[0119] Calculate the cosine similarity between the updated risk scenario unit feature vector X and the feature vector of each candidate scenario unit. The calculation formula is as follows:

[0120]

[0121] in, For the updated risk scenario unit feature vector With the scene library Feature vectors of candidate scene units Cosine similarity; The dimension of the feature vector; For the updated risk scenario unit feature vector Middle Normalized values ​​of each feature; For the first Feature vectors of candidate scene units Middle Normalized values ​​of each feature; , respectively, feature vectors The modulus length;

[0122] A preset similarity threshold is used to filter out risky scenario units with a cosine similarity greater than or equal to the threshold, classifying them as similar risky scenarios; when the number of similar risky scenarios exceeds the preset number of risky scenarios... When, take the top of the similarity ranking. Each scenario unit is selected as the final similar risk scenario; at the same time, knowledge graph reasoning is used to assist in judging the rationality of matching similar scenarios and eliminate scenario units with large matching deviations.

[0123] Furthermore, the calculation of the corresponding energy dissipation adaptability and protection adaptability is as follows: Based on the combination of protective device parameters in similar risk scenarios, the comprehensive ability of the protective device parameter combination to intercept, decelerate, divert, disperse, buffer, control the accumulation of disaster energy, and reduce residual energy is evaluated through the energy dissipation adaptability. The calculation formula is as follows:

[0124]

[0125] in, Energy dissipation adaptability (value range 0-100%, the higher the value, the better the energy dissipation effect). The initial total energy of the disaster. This refers to the remaining energy after a disaster has passed through multiple levels of protective devices. The total energy dissipation capacity of the entire protective device, For the first The energy dissipation capacity of a single protective device For the first The number of protective devices deployed is calculated by cumulatively adding up the energy dissipation capacity of multiple levels of protective devices.

[0126] Protection adaptability is used to evaluate the protective effect of the combination of protective device parameters on the safety of the disaster-bearing body, as well as the degree of matching between engineering accessibility, construction difficulty, implementation period, maintenance conditions, cost constraints, and emergency response capabilities. The formula is:

[0127]

[0128] in, To ensure compatibility, The geographical adaptability coefficient for the deployment location of the protective device. The structural adaptation coefficient of the protective device. The feasibility coefficient for the construction of the protective device. This is the engineering cost coefficient. For later operation and maintenance coefficients, As the emergency rescue accessibility coefficient, , , , , , The weights are assigned based on the magnitude of each item's impact.

[0129] Furthermore, all coefficients range from 0 to 1. Among them, the geographical adaptation coefficient is determined based on the terrain slope and soil properties of the deployment area; the structural adaptation coefficient is determined based on the matching degree between the type of protective device and the type of disaster; the construction feasibility coefficient is determined based on the on-site construction conditions; the engineering cost coefficient is determined based on the construction cost of the protective project; the post-operation and maintenance coefficient is determined based on the operation and maintenance difficulty and cost of the protective device, with a higher coefficient for more convenient operation and maintenance; and the emergency rescue accessibility coefficient is determined based on the emergency rescue accessibility of the protected area, with a higher coefficient for better accessibility.

[0130] Furthermore, the specific candidate protective device parameter combinations are as follows: based on energy dissipation adaptability and protection adaptability, a weighted comprehensive scoring method is used to score and screen each candidate protective device parameter combination. If the disaster-bearing body corresponding to the updated risk scenario unit has the conditions for relocation and avoidance, the protective combination that meets the requirements of avoidance site selection, restricted area delineation, monitoring and early warning, and evacuation routes is given priority. If the avoidance conditions are not met, the candidate protective device parameter combination with the highest comprehensive score is selected as the target protective device parameter combination. If there are multiple combinations with the same and highest comprehensive score, the combination with the lowest overall engineering cost is selected first.

[0131] Furthermore, the parameter library for protective devices is divided into at least seven categories: active control devices in the source area, interception and energy dissipation devices in the flow area, ditches for diversion and drainage, slope toe accumulation control devices, near-end protection devices for disaster-bearing bodies, temporary modular protection devices, and monitoring and early warning auxiliary devices. The complete combination of target protective device parameters includes the following parameters: deployment location, number of deployments, deployment density, deployment spacing, number of rows, row spacing, device energy level, device height, device width, device length, diversion angle, interception capacity, energy dissipation capacity, and combination sequence.

[0132] The formula for calculating the layout density is as follows: , For deployment density, This refers to the total number of protective devices deployed. The total area of ​​the protected zone; the formula for calculating the spacing between the installations is: , To set the spacing, The effective protection radius of the protective device.

[0133] Furthermore, following the five-level protection chain sequence of source area control—path interception—diversion and energy dissipation—end protection—emergency support, various protective devices are deployed sequentially. Among them, active control devices in the source area are deployed first, while interception and energy dissipation devices in the flow area and diversion and drainage devices in the ditch are deployed simultaneously. Slope toe accumulation control devices and near-end protection devices for disaster-bearing bodies are key deployment links, and temporary modular protective devices are deployed flexibly according to the construction progress and on-site emergency needs.

[0134] Furthermore, the update of the mountain disaster risk scenario database and the protective device parameter database specifically includes updating the mountain disaster risk scenario database and the protective device parameter database.

[0135] Among them, the update of the mountain disaster risk scenario database is achieved by collecting expert verification opinions, on-site review data, engineering implementation effects and post-disaster review results. The database is supplemented with the risk scenario units involved by correcting information and adding new risk scenario units, while invalid and outdated risk scenario units are deleted. The risk level, evolution information and implementation effect of protection plans of the scenario units are updated simultaneously.

[0136] The update of the protective device parameter library will supplement the actual implementation parameters of the target protective device parameter combination, the feedback information on the engineering implementation effect, and the energy dissipation capacity and protection effect data measured after the disaster review into the parameter library. It will correct the energy dissipation capacity, adaptability coefficient, cost coefficient, operation and maintenance coefficient and other parameters of various protective devices, optimize the recommendation rules for protective device parameter combinations, and supplement the optimal parameter range for the deployment of protective devices in different scenarios.

[0137] This embodiment details an AI-based mountain disaster scenario database and intelligent protection recommendation method. It integrates multi-source data from topography, remote sensing, monitoring, engineering, and disaster-bearing bodies, forming an integrated data foundation through standardized processing with unified coordinates and coding annotations. It automatically extracts complete features of disasters and disaster-bearing bodies using attention-enhanced U-Net and YOLOv8n lightweight models, distinguishes spatial relationships, constructs standardized risk scenario units, and builds a hierarchical mountain disaster risk scenario database. After new monitoring data drives the generation of new scenarios, similar cases are retrieved, and multiple candidate solutions are generated by combining them with a protective device parameter database. An adaptive weighted energy dissipation and protective adaptation quantification index is introduced for comprehensive scoring, and the optimal parameter deployment scheme is output according to a five-level protection chain. Feedback is continuously used to iteratively update the scenario database and protective parameter database after operation.

[0138] Example 2, based on Example 1, details an AI-based mountain disaster scenario database and intelligent protection recommendation system, such as... Figure 5 As shown, it includes a multi-source data governance module, a dual-target AI recognition module, a risk scenario library construction module, a protection combination generation module, a protection scheme intelligent selection module, and a scenario library;

[0139] The multi-source data governance module stores DEM, DSM, remote sensing, UAV 3D, deformation monitoring, meteorological and hydrological data, disaster cases, and disaster-bearing body spatial multi-source datasets. It has built-in standardized processing operator set, unified coding rule library and semantic annotation tag library for outputting multi-source object data base.

[0140] The dual-target AI recognition module includes an improved U-Net disaster recognition submodule and a YOLOv8n disaster-bearing body recognition submodule. The two submodules are equipped with a preset hierarchical convolutional network, a corresponding loss function and training parameters, and are connected in parallel to a multi-source object data base to output disaster object features and disaster-bearing body object features.

[0141] The risk scenario library construction module has a built-in spatial relationship determination operator, an octet risk scenario unit generator, a B+ tree index engine, and a hierarchical sub-library storage medium, which are used to store and structure all risk scenario units.

[0142] The protection combination generation module has built-in feature vector normalization operator, cosine similarity calculation unit, knowledge graph reasoning engine, and seven categories of protection device parameter library storage, which are used to read new monitoring data to generate and update scene units and output multiple sets of candidate protection device parameter combinations.

[0143] The intelligent optimization module for the protection scheme has a built-in energy dissipation adaptability solving unit, an adaptive weight protection adaptability solving unit, a weighted comprehensive scoring calculation unit, and a five-level protection chain layout scheduler, which are used to output parameterized energy dissipation combinations of the target protection devices;

[0144] The scenario library has a built-in multi-source feedback data access interface, parameter correction operator, and scenario unit addition, deletion, modification and storage controller, which communicate bidirectionally with the risk scenario library construction module and the protection device parameter library storage.

[0145] Furthermore, the multi-source data governance module incorporates coordinate transformation operators, spatial registration operators, scale normalization operators, data cleaning operators, object encoding operators, and semantic annotation operators. The coordinate transformation operators incorporate WGS-84 coordinate system transformation logic, the spatial registration operator has a registration error threshold set to 0.5 meters, and the scale normalization operator uniformly adopts a 1:5000 scale standard. The object encoding operator incorporates a 16-bit encoding generator, with encoding rules of data type encoding - region encoding - time encoding - sequence number encoding. The multi-source data governance module is equipped with an independent high-speed storage partition for persistent storage of the standardized multi-source object data base.

[0146] Furthermore, the improved U-Net disaster recognition submodule incorporates a 5-layer encoder, a 5-layer decoder, a cross-layer skip connection channel, and an attention mechanism embedding unit; the encoder and decoder each layer have fixed convolutional kernel numbers, convolutional sizes, and pooling / transposed convolutional parameters; the improved U-Net disaster recognition submodule incorporates a cross-entropy loss calculation unit and an Adam optimizer scheduling unit, and pre-stores a fixed learning rate of 0.001, a maximum of 100 training epochs, and early stopping logic for 10 consecutive epochs where accuracy does not improve; the improved U-Net disaster recognition submodule is equipped with an independent GPU inference computing unit, used to output disaster object features in real time, consisting of disaster type, source area location, geometric range, potential movement direction, and disaster energy level.

[0147] Furthermore, the YOLOv8n disaster-bearing body identification submodule is equipped with a lightweight CSPDarknet53 backbone network, a PANet neck multi-scale fusion unit, and an anchor-free decoupled detection head; the backbone network, neck network, and detection head have fixed convolution channel numbers and downsampling / upsampling parameters; the YOLOv8n disaster-bearing body identification submodule has a built-in CIoU loss calculation unit, pre-configured with a learning rate of 0.0005, batch size of 32, and input image size of 640×640; the dual-target AI recognition module has a data interaction bus that synchronously transmits disaster object features and disaster-bearing body object features to the risk scenario library construction module.

[0148] This embodiment details an AI-based mountain disaster scenario database and intelligent protection recommendation system, including a multi-source data governance module, a dual-target AI recognition module, a risk scenario database construction module, a protection combination generation module, a protection scheme intelligent optimization module, and a scenario database. The multi-source data governance module integrates multiple types of mountain-related spatiotemporal datasets and a complete set of standardized processing operators, generating a structured multi-source object data base according to unified coordinates, scale, and coding rules. The dual-target AI recognition module uses two independent convolutional inference sub-modules—a five-layer symmetric improved U-Net with attention mechanism and a lightweight YOLOv8n—to extract complete features of disasters and disaster-bearing bodies, which are then transmitted via an internal bus. The risk scenario database construction module uses spatial discrimination operators, an octet unit generator, and a B+ tree index to achieve hierarchical storage of risk scenarios. The similar scenario retrieval module combines vector normalization, cosine similarity, and knowledge graph matching to generate candidate protection schemes. The protection optimization module is equipped with a dual-adaptability solving unit and a comprehensive scoring calculator to output tiered protection deployment schemes. The iterative update module connects to multi-channel feedback data to maintain the scenario database and protection parameter database bidirectionally.

[0149] Example 3, based on Example 1 or 2, describes in detail the implementation process of intelligent protection recommendations using this method in a construction camp in a high mountain canyon valley, as follows:

[0150] The construction camp is located at the foot of a river valley slope, with high-level unstable rock masses on the upper part of the mountain. These rocks are prone to collapse due to unloading, freeze-thaw cycles, and rainfall, and the potential collapse path points directly towards the camp area. The area near the camp includes accommodation, office, material storage, construction access roads, emergency assembly points, and temporary communication facilities. The high-level unstable rock masses are classified as high-energy mountain hazards and require a multi-level protection and energy dissipation system.

[0151] Collect DEM, DSM, UAV orthophotos, oblique photogrammetry 3D model and point cloud of the study area, InSAR / GNSS deformation monitoring data, rainfall, freeze-thaw and snow monitoring data, on-site crack investigation records, historical rockfall traces, past protective engineering drawings, and vector disaster bearing data of camps and supporting facilities.

[0152] All data undergoes unified standardization processing: uniform conversion to the WGS-84 coordinate system, spatial registration error controlled within 0.5m, and unified 1:5000 scale normalization; missing, distorted, and duplicate data are removed through data cleaning; multi-dimensional object coding is set, including disaster, disaster-bearing body, spatial relationship, risk boundary, disaster energy, protective device, protective combination, and data credibility coding; semantic annotation is completed according to the mountain disaster industry standard, forming a standardized multi-source object data base, stored in an independent high-speed storage partition, providing a unified data source for AI recognition.

[0153] The AI ​​disaster identification uses an improved U-Net model with a 5-layer encoder and a 5-layer decoder, with attention modules embedded between layers. Disaster samples within the base are rotated, flipped, scaled, and have their brightness enhanced. The model is divided into training, validation, and test sets. A cross-entropy loss function and the Adam optimizer are used, with a learning rate of 0.001 and a maximum training epoch of 100. Training stops if the validation set accuracy does not improve after 10 consecutive epochs. The model outputs the following disaster object characteristics: disaster type is high-altitude rockfall, specific latitude and longitude coordinates of the source area, geometric range of approximately 800 square meters, relative height difference of 320 meters, potential movement direction is vertically downwards towards the camp, disaster energy level is high, and identification confidence level.

[0154] The AI-based disaster-bearing body identification uses a lightweight YOLOv8n model with a CSPDarknet53 backbone, a PANet multi-scale fusion structure for the neck, and an anchorless decoupled detection head. CIoU is used as the boundary loss function, the input image size is set to 640×640, and the training batch size is 32. The model outputs disaster-bearing body features: construction camp type, specific spatial coordinates, lightweight prefabricated house structure, high nighttime exposure characteristics, functional importance level 1, medium transferability, moderate emergency accessibility, and recognition reliability.

[0155] The output features of the two models are synchronously transmitted to the risk scenario unit generation module via the internal data bus of the module.

[0156] Using spatial topology analysis, the spatial relationship between the disaster and the affected body was determined to be a combination of a direct impact type and a slope-toe diffusion type. Four risk boundaries were defined: the core impact zone completely covers the camp accommodation area; the potential impact zone covers the camp office area and construction access road; the verification buffer zone is a 50-meter radius around the core impact zone; and the uncertainty envelope is a 30-meter radius around the verification buffer zone. Based on the disaster energy level and the vulnerability of the affected body, the comprehensive risk level was calculated to be level 5 (highest risk), with a calculated intervention index of 0.26. In this case, the camp is not ready for complete relocation in the short term, indicating low intervention capability.

[0157] A standardized octet risk scenario unit is constructed, specifically: the disaster body characteristics are parameters related to high-altitude unstable rockfall, the disaster-bearing body characteristics are parameters related to construction camps, the spatial relationship characteristics are positive direct impact type and slope toe diffusion type, the risk boundary characteristics are the above four layers of risk boundary range, the disaster energy characteristics are high energy level, the exposure vulnerability characteristics are high personnel exposure and light structure vulnerability, the combination of protection parameters to be matched is reserved blank, and the comprehensive identification confidence is 0.92;

[0158] A unique code is assigned to this scenario, with the coding rule being disaster type code - disaster-bearing body type code - spatial relationship code - risk level code. Specifically, the code is collapse - camp - direct impact spread - level 5, and a B+ tree index is used for storage. This scenario is stored in a mountain disaster risk scenario library containing 11 hierarchical sub-libraries. The 11 sub-libraries are: disaster ontology sub-library, disaster-bearing body sub-library, spatial relationship sub-library, risk boundary sub-library, disaster energy sub-library, exposure vulnerability sub-library, protective device parameter sub-library, protective device combination sub-library, similar case sub-library, recommendation rule sub-library, and feedback update sub-library.

[0159] Receive real-time slope deformation monitoring data and update all characteristic parameters of the current risk scenario unit;

[0160] A preliminary screening was conducted, matching basic features such as landslide disasters, construction camp disaster-bearing bodies, direct-impact spatial relationships, and level 5 high risk to select 12 candidate historical scenarios. The features of the selected candidate scenarios and the current scenario were normalized. Finally, the cosine similarity was calculated, and a similarity threshold of 0.8 was set. The top 5 historical similar scenarios with the highest similarity were selected, and a knowledge graph was introduced to remove 2 scenarios with large matching deviations, thus determining 3 valid similar scenarios.

[0161] The parameter database of protective devices is retrieved. This database is divided into seven categories: active control in the source area, interception and energy dissipation in the flow area, diversion and drainage in the gully, slope toe accumulation control, near-end protection of the disaster-bearing body, temporary modular protection, and monitoring and early warning auxiliary protection. According to the five-level protection chain of source area control, path interception, diversion and energy dissipation, end protection and emergency support, and combined with historical protection schemes for effective similar scenarios, six sets of candidate protective device parameter combinations are generated. Each combination includes device type, deployment location, number of devices, deployment density, deployment spacing, number of rows, device energy level, diversion angle, and energy dissipation capacity parameters.

[0162] The total energy dissipation capacity of each protective device group is accumulated step by step. The initial total energy of the disaster is compared with the remaining energy after multiple levels of protection. The energy dissipation adaptability of 6 candidate combinations is calculated. The adaptability ranges from 62% to 91%, with the highest energy dissipation adaptability being 91% and the lowest being 62%.

[0163] The protection adaptability is calculated using adaptive weights. Considering the high risk and low interventionability of this case, the weights of the geographical adaptability coefficient, construction feasibility coefficient, and emergency rescue accessibility coefficient are automatically increased, while the weights of the engineering cost coefficient and post-operation and maintenance coefficient are decreased. The protection adaptability of 6 candidate combinations is calculated, with an adaptability range of 0.65 to 0.92.

[0164] The six candidate schemes were ranked using a weighted comprehensive scoring method, prioritizing the assessment of whether the disaster-bearing body had the conditions for avoidance. In this case, the camp could not be relocated as a whole, so the scheme with the highest comprehensive score was selected, which had a comprehensive score of 0.90. If multiple schemes had the same score, the combination with the lowest overall engineering cost was selected.

[0165] Output the complete parameters of the parameterized energy dissipation combination of the target protection device, such as Figure 6 As shown, an active protective netting was installed on the outer side of the high-level rock mass, covering an area of ​​800 square meters in the source zone of the unstable rock mass, to constrain, reinforce, and provide initial protection for the source rock mass. Three source zone monitoring and early warning devices were installed near the high-level rock mass to monitor the expansion of rock fissures, block deformation, or local instability. Along the direction of collapse, two passive protective netting systems with an energy level of 500 kJ were installed in the middle of the slope to intercept and decelerate falling rocks or debris. Five slope monitoring and early warning devices were installed at key locations on the slope to monitor slope roll-off, impact, vibration, or local deformation. A multi-level retaining pile group was installed at the lower part of the slope or the leading edge of the slope toe, with a total of [number missing] piles. Three rows of retaining piles are arranged with a pile spacing of 2 meters and a row spacing of 8 meters, in a tiered manner along the direction of the collapse movement. This is used to intercept the collapse in stages, disperse the impact, and dissipate energy step by step. A retaining wall with a height of 3 meters is set below the multi-level retaining pile group or at the front edge of the slope toe accumulation area to limit the continued spread of the collapse accumulation towards the construction camp. A near-end protective wall with a length of 120 meters and a height of 2.5 meters is set on the mountainside of the construction camp to provide secondary protection for the disaster-bearing body near the camp. Two evacuation routes and one emergency assembly point are planned simultaneously. The density of the protective devices is 0.05 units / square meter, with a spacing of 4 meters, calculated by the module's built-in computing unit.

[0166] Construction will proceed in stages according to the five-level protection chain. First, active protection netting and three source area monitoring and early warning devices will be deployed on the outer side of the high-level rock mass. Next, two passive protection netting systems will be constructed in the middle of the slope, and five slope monitoring and early warning devices will be deployed at key locations on the slope. Three rows of multi-level retaining piles will be constructed at the lower part of the slope or the leading edge of the slope toe. A 3-meter-high retaining wall will be constructed below the multi-level retaining piles or at the leading edge of the slope toe accumulation area. A 120-meter-long, 2.5-meter-high near-end protection wall will be built on the mountainside of the construction camp. Two evacuation routes and one emergency assembly point will be planned simultaneously. After construction is completed, the deployment parameters of all protection devices will be verified, and the installation stability and energy dissipation performance of the devices will be tested. The devices will be put into use after passing acceptance inspection.

[0167] The system collects four types of feedback data: expert review opinions, on-site inspection data, flood season project operation results, and post-disaster review data, and performs comprehensive updates:

[0168] The Mountain Disaster Risk Scenario Database has added a complete scenario unit for this case, revised the risk classification standards and evolution information for high-energy direct-impact scenarios, and deleted two outdated scenarios.

[0169] The protective device parameter library is updated with the measured energy dissipation capacity and adaptation coefficient of various protective devices, and the optimal deployment parameter range of protective devices in high-energy collapse scenarios is supplemented.

[0170] The inference parameters of the AI ​​recognition model, the adaptive weight allocation rules for protection adaptability, the similarity threshold for similar scene retrieval, and the protection recommendation ranking logic are updated synchronously.

[0171] This embodiment details the implementation process of intelligent energy dissipation recommendation using this method at a construction camp in a high mountain canyon. The process outputs a multi-level rigid-flexible coupled energy dissipation protection combination, including an active protection net, source area monitoring and early warning device, passive protection net, slope monitoring and early warning device, multi-level retaining pile group, retaining wall, near-end protection wall, and an integrated emergency evacuation system. This comprehensively covers the entire path of energy dissipation from high-altitude rockfalls to the construction camp, with each device's function and deployment location perfectly matching the scenario requirements. The output results include complete recommendation reasons, applicable conditions, and a checklist of verification parameters, which can be directly used in engineering design. After multiple rounds of on-site review and iteration, the matching accuracy between the scenario library and the protection parameter library has continuously improved, achieving long-term stable protection effects and verifying the efficiency and accuracy of the proposed solution recommendation.

[0172] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. An AI-based mountain disaster scenario database and intelligent protection recommendation method, characterized in that, include: Acquire multi-source data of the target area, and perform coordinate unification, spatial registration, scale normalization, data cleaning, object encoding and semantic annotation on the data to construct a multi-source object data foundation; Based on the multi-source object data base, AI disaster object identification and AI disaster-bearing body object identification are carried out respectively to obtain the characteristics of disaster objects and disaster-bearing body objects. Based on the characteristics of the disaster objects and the disaster-bearing bodies, the spatial relationship between the disaster objects and the disaster-bearing bodies is identified, and corresponding risk scenario units are generated. Multiple risk scenario units are organized according to disaster type, disaster-bearing body type, spatial relationship type and risk level to construct a mountain disaster risk scenario library. When new monitoring data is acquired, corresponding updated risk scenario units are generated. Based on the updated risk scenario units, similar risk scenarios are retrieved from the mountain disaster risk scenario database, and candidate protective device parameter combinations are generated in combination with the protective device parameter database. Based on the parameter combinations of each candidate protective device, the corresponding energy dissipation adaptability and protection adaptability are calculated. Based on the energy dissipation adaptability and protection adaptability, the target protective device parameter combination is selected. Based on the parameter combination of the target protection device, the protection deployment is implemented, and the mountain disaster risk scenario database and protection device parameter database are updated according to expert verification, on-site review, engineering implementation effect and post-disaster review results.

2. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 1, characterized in that, The construction of the multi-source object data base specifically involves: after standardizing the multi-source data through coordinate unification, spatial registration, scale normalization, data cleaning, object encoding, and semantic annotation, constructing an integrated data set that combines mountain disaster-related data and disaster-bearing body-related data, thus obtaining the multi-source object data base.

3. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 1, characterized in that, The disaster object characteristics are specifically defined as follows: the disaster-related data in the multi-source object data base is cropped into samples of a fixed size, and data augmentation processing is performed. The augmentation methods include random rotation, flipping, scaling, and brightness adjustment to generate training sets, validation sets, and test sets. An improved U-Net model is adopted, and an attention mechanism module is added between the encoder and decoder of the U-Net architecture. The feature map extracted by the encoder is multiplied with the attention weight to obtain a weighted feature map, which is then input into the decoder for upsampling. The preprocessed disaster data of the target area is input into the trained improved U-Net model, which outputs the segmentation results of the disaster objects. The features of the disaster objects are extracted by combining spatial coordinate information. The features of the disaster objects include disaster type, source area location, geometric range, potential movement direction and disaster energy level.

4. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 3, characterized in that, The AI-based disaster-bearing object identification specifically involves: using the YOLOv8n lightweight target detection model to annotate the disaster-bearing object data in the multi-source object data base. The annotation content includes the disaster-bearing object type and spatial location coordinates, generating VOC format annotation files, and constructing a standardized dataset. The labeled disaster-bearing body data is divided into training, validation, and test sets, which are then input into the YOLOv8n lightweight target detection model for training. The trained YOLOv8n lightweight target detection model outputs disaster-bearing body type, spatial location, geometric range, personnel exposure characteristics, structural vulnerability, functional importance, portability, usage period, emergency accessibility, and recognition reliability.

5. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 1, characterized in that, The construction of the mountain disaster risk scenario database specifically involves: assigning a unique scenario code to each risk scenario unit, with the scenario coding rule being disaster type code - disaster-bearing body type code - spatial relationship code - risk level code; A B+ tree index was constructed for scene coding, disaster type, and disaster-bearing body type to enable fast retrieval based on multiple conditions, thus completing the construction of a mountain disaster risk scene database.

6. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 1, characterized in that, The process of retrieving similar risk scenarios from the mountain disaster risk scenario database specifically involves using a retrieval method that combines preliminary screening, precise matching, and knowledge graph assistance. The preliminary screening filters the risk scenario units in the mountain disaster risk scenario database according to disaster type, disaster-bearing body type, risk level, and spatial relationship type, and selects candidate scenario units that match the basic characteristics of the updated risk scenario units. The updated risk scenario units' disaster object characteristics, disaster-bearing body characteristics, spatial relationships, risk levels, and disaster energy levels are normalized, generating dimensions as follows: eigenvectors The corresponding features of the selected candidate scene units are also normalized to generate feature vectors. ; Calculate the cosine similarity between the updated risk scenario unit feature vector X and the feature vector of each candidate scenario unit; A preset similarity threshold is used to filter out risky scenario units with a cosine similarity greater than or equal to the similarity threshold, which are then identified as similar risky scenarios. When the number of similar risk scenarios exceeds the preset number of risk scenarios When, take the top of the similarity ranking. Each scenario unit serves as the final similar risk scenario; At the same time, knowledge graph reasoning is used to assist in judging the rationality of matching similar scenes and eliminate scene units with large matching deviations.

7. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 6, characterized in that, The calculation of the corresponding energy dissipation adaptability and protection adaptability is specifically as follows: based on the combination of protective device parameters in similar risk scenarios, the comprehensive ability of the protective device parameter combination to intercept, decelerate, divert, disperse, buffer, control the accumulation of disaster energy and reduce residual energy is evaluated through the energy dissipation adaptability. Protection adaptability is used to evaluate the protective effect of the combination of protective device parameters on the safety of the disaster-bearing body, as well as the degree of matching between engineering accessibility, construction difficulty, implementation period, maintenance conditions, cost constraints and emergency response capabilities.

8. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 1, characterized in that, The candidate protective device parameter combinations are specifically defined as follows: based on energy dissipation adaptability and protection adaptability, a weighted comprehensive scoring method is used to score and screen each candidate protective device parameter combination. If the disaster-bearing body corresponding to the updated risk scenario unit has the conditions for relocation and avoidance, the protective combination that meets the requirements of relocation site selection, restricted area delineation, monitoring and early warning, and evacuation routes is given priority. If the conditions for relocation and avoidance are not met, the candidate protective device parameter combination with the highest comprehensive score is selected as the target protective device parameter combination. If there are multiple combinations with the same and highest comprehensive score, the combination with the lowest overall engineering cost is selected first.

9. The AI-based mountain disaster scenario database and intelligent protection recommendation method according to claim 8, characterized in that, The update of the mountain disaster risk scenario database and the protective device parameter database specifically includes updating the mountain disaster risk scenario database and the protective device parameter database. Among them, the update of the mountain disaster risk scenario database is achieved by collecting expert verification opinions, on-site review data, engineering implementation effects and post-disaster review results. The database is supplemented with the risk scenario units involved by correcting information and adding new risk scenario units, while invalid and outdated risk scenario units are deleted. The risk level, evolution information and implementation effect of protection plans of the scenario units are updated simultaneously. The update of the protective device parameter library will supplement the actual implementation parameters of the target protective device parameter combination, the feedback information on the engineering implementation effect, and the energy dissipation capacity and protection effect data measured after the disaster review into the parameter library. It will correct the energy dissipation capacity, adaptability coefficient, cost coefficient, operation and maintenance coefficient and other parameters of various protective devices, optimize the recommendation rules for protective device parameter combinations, and supplement the optimal parameter range for the deployment of protective devices in different scenarios.

10. An AI-based mountain disaster scenario database and intelligent protection recommendation system, used to implement the AI-based mountain disaster scenario database and intelligent protection recommendation method according to any one of claims 1-9, comprising a multi-source data governance module, a dual-target AI recognition module, a risk scenario database construction module, a protection combination generation module, an intelligent protection scheme selection module, and a scenario database; The multi-source data governance module stores DEM, DSM, remote sensing, UAV 3D, deformation monitoring, meteorological and hydrological data, disaster cases, and disaster-bearing body spatial multi-source datasets. It has built-in standardized processing operator set, unified coding rule library and semantic annotation label library, and outputs multi-source object data base. The dual-target AI recognition module includes an improved U-Net disaster recognition submodule and a YOLOv8n disaster-bearing body recognition submodule. The two submodules are equipped with a preset hierarchical convolutional network, a corresponding loss function and training parameters, and are connected in parallel to a multi-source object data base to output disaster object features and disaster-bearing body object features. The risk scenario library construction module has a built-in spatial relationship determination operator, an octet risk scenario unit generator, a B+ tree index engine, and a hierarchical sub-library storage medium to store and structurally manage all risk scenario units. The protection combination generation module has built-in feature vector normalization operator, cosine similarity calculation unit, knowledge graph reasoning engine, and seven categories of protection device parameter library storage. It reads new monitoring data to generate and update scene unit and outputs multiple sets of candidate protection device parameter combinations. The intelligent optimization module for the protection scheme has a built-in energy dissipation adaptability solving unit, an adaptive weight protection adaptability solving unit, a weighted comprehensive scoring calculation unit, and a five-level protection chain layout scheduler, which outputs parameterized energy dissipation combinations of the target protection device; The scenario library has a built-in multi-source feedback data access interface, parameter correction operator, and scenario unit addition, deletion, modification and storage controller, which communicate bidirectionally with the risk scenario library construction module and the protection device parameter library storage.

Images