Slope dangerous rock mass identification and prediction method and device based on large model

Abstract

The application discloses a slope dangerous rock mass identification and prediction method and device based on a large model, relates to the technical field of geological disaster response and prevention, and comprises the following steps: constructing a prediction model; acquiring historical data; constructing a first database and a second database; preprocessing the data of the first database and the second database; training the prediction model; acquiring real-time monitoring data; preprocessing the real-time monitoring data; analyzing the preprocessed real-time monitoring data by using the prediction model; dynamically adjusting a risk threshold; analyzing a risk result; determining a response measure according to the risk result and performing visual display; through the method, three-dimensional visualization of a field dangerous rock mass instability probability and a threshold-adjustable risk early warning mechanism are realized, fast and accurate identification of a dangerous rock mass deformation area is realized, the false alarm rate is effectively reduced, the monitoring and early warning credibility is improved, and flexible and reliable auxiliary decision support is provided for emergency rescue and comprehensive management.

Description

Technical Field

[0001] This invention relates to the field of geological disaster response and prevention technology, and in particular to a method and device for identifying and predicting unstable rock masses on slopes based on a large model. Background Technology

[0002] Slope rock mass instability poses a serious threat to people's lives and property, as well as the stability of engineering projects. Slope rock mass instability is characterized by its sudden onset and wide spatial variation, requiring on-site assessment to locate a large area of ​​potential hazards within a short period. Traditional methods rely primarily on manual experience, simple analysis of single-type monitoring data, or conventional numerical simulations. Furthermore, the monitoring data is often point- or linear, lacking the ability to capture global information.

[0003] Empirical judgment is easily limited by subjectivity, and single-source data can only provide a partial view without seeing the whole picture. Numerical simulation has complex parameters and long calculation time, and point and line monitoring lacks three-dimensional continuous information, resulting in frequent missed reports and misjudgments. With the popularization of multi-source non-contact methods such as drones, radar, and cameras, the data dimension has increased dramatically, but existing frameworks are unable to couple deformation fields, geomechanical parameters, and environmental influencing factors in real time within a unified spatiotemporal grid. Summary of the Invention

[0004] The purpose of this invention is to design a method and device for identifying and predicting unstable rock masses on slopes based on a large model in order to solve the above problems.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] A method for identifying and predicting unstable rock masses on slopes based on a large model includes:

[0007] S1. Construct the initial prediction model. The prediction model, from input to output, includes an encoding module, a fusion module, and an instability risk prediction module. The encoding module is used to encode and align the data from the three modalities. The fusion module is used to fuse the encoded feature vectors to obtain a fused feature vector. The instability risk prediction module is used to make predictions based on the fused feature vector to obtain the predicted instability probability. The fusion module, from input to output, includes a feature-level physical association layer, a modality-level reliability weighting layer, a scene-level adaptability calibration layer, and a fusion layer. The feature-level physical association layer is used to inject physical logic into the feature vectors. The modality-level reliability weighting layer is used to compress and splice the spatiotemporal features of each modality and fuse all modalities to obtain a multimodal fusion representation. The scene-level adaptability calibration layer is used to match the multimodal fusion representation with the on-site scene. The fusion layer is used to fuse the matched multimodal fusion representation with the on-site scene to obtain a fused feature vector.

[0008] S2. Acquire historical data, which includes monitoring data on surface deformation of unstable rock mass and geomechanical and environmental data. The three modes are surface deformation of unstable rock mass, geomechanical and environmental data.

[0009] S3. Construct a first database based on surface deformation monitoring data of unstable rock masses, and construct a second database based on geomechanics and environmental data;

[0010] S4. Preprocess the data in the first and second databases;

[0011] S5. Import the preprocessed data into the initial prediction model, train and optimize it to obtain the optimized prediction model.

[0012] S6. Obtain real-time monitoring data;

[0013] S7. Preprocess the real-time monitoring data;

[0014] S8. Analyze the processed real-time monitoring data using the optimized prediction model to obtain the predicted instability probability and slip volume.

[0015] S9. Dynamically adjust the risk threshold according to the volume of the slip body to obtain the adjusted risk threshold.

[0016] S10. Analyze the risk results based on the adjusted risk threshold and instability probability;

[0017] S11. Determine response measures based on the risk results and present them visually.

[0018] A slope unstable rock mass identification and prediction device based on a large model includes:

[0019] Storage; the storage device contains computer programs;

[0020] Actuator; the actuator is used to execute a computer program stored in the memory, which, when executed, implements the slope unstable rock mass identification and prediction method based on a large model as described above.

[0021] The beneficial effects of this invention are as follows: This method realizes a three-dimensional visualization of the probability of unstable rock mass in the field and an adjustable threshold risk early warning mechanism, which enables rapid and accurate identification of the deformation area of ​​the unstable rock mass, effectively reduces the false alarm rate, improves the credibility of monitoring and early warning, and provides flexible and reliable auxiliary decision support for emergency rescue and comprehensive management. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the fusion module of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0026] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0027] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0028] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] A method for identifying and predicting unstable rock masses on slopes based on a large model includes:

[0031] S1. Construct the initial prediction model. The prediction model, from input to output, includes an encoding module, a fusion module, and an instability risk prediction module. The encoding module is used to encode and align the data of the three modalities. The fusion module is used to fuse the encoded feature vectors to obtain a fused feature vector. The instability risk prediction module is used to make predictions based on the fused feature vectors to obtain the predicted instability probability. The fusion module, from input to output, includes a feature-level physical association layer, a modality-level reliability weighting layer, a scene-level adaptability calibration layer, and a fusion layer. The feature-level physical association layer is used to inject physical logic into the feature vectors. The modality-level reliability weighting layer is used to compress and splice the spatiotemporal features of each modality and fuse all modalities to obtain a multimodal fusion representation. The scene-level adaptability calibration layer is used to match the multimodal fusion representation with the on-site scene. The fusion layer is used to fuse the matched multimodal fusion representation with the on-site scene to obtain a fused feature vector.

[0032] The encoding module is used to encode and align data from the three modalities, specifically as follows:

[0033] Dynamic surface deformation tokens (D-Tokens) of unstable rock masses: Surface displacement field obtained from UAV point cloud / photogrammetry Slope radar can obtain LOS displacement. Ground-based cameras can capture images of the crack opening. Unified embedding yields: ;

[0034] Static Geomechanical Tokens (G-Tokens): ;

[0035] Environmental Impact Factor Tokens (E-Tokens): ;

[0036] in, , and These are the feature vectors for surface deformation, geomechanics, and environment of the unstable rock mass, respectively. L is the time step, d is the embedding dimension, M is the length of the original vector after static feature stitching, and N is the time step for the environmental data. The surface deformation data of the unstable rock mass includes the surface displacement field. LOS displacement and crack opening LOS displacement refers to displacement in the radar line-of-sight direction; geomechanical data includes slope. Slope aspect Elevation h, lithology code Jointed birth , Value, type of dangerous rock Rock mass unit weight Cohesion c and internal friction angle Environmental data includes temperature. Wind speed Rainfall and pore water pressure .

[0037] like Figure 1 As shown, the fusion module is used to perform feature fusion on the encoded feature vectors to obtain the fused feature vectors, specifically including:

[0038] 1) The feature-level physical association layer uses a cross-attention mechanism to construct deformation query Q with geological / environmental keys and values ​​K and V. By quantifying attention weights, crack displacement automatically focuses on the most relevant joint surfaces and rainfall events, thereby injecting physical logic into multimodal data fusion; represented as: , , , , and , , Represent the feature dimension; construct the attention weight matrix Attention weight matrix Matrix representation behavior: ,in, The weight of the i-th deformation token in the feature vector of surface deformation of the unstable rock mass to the j-th token in the feature vector of geomechanics or environment is represented. , and These are the feature vectors for surface deformation of the unstable rock mass, geomechanics, and environment, respectively; the attention-weighted feature vector for surface deformation of the unstable rock mass is then calculated. The weighted surface deformation feature vector of the unstable rock mass has incorporated key geological and environmental information;

[0039] 2) The modal-level reliability weighted layer calculates the global mean, compressing the spatiotemporal features of each modality into a single d-dimensional vector for subsequent gating calculations, expressed as: ;

[0040] The compressed vectors are concatenated to generate the gated vector. , represented as: ,in, , and These are the confidence weights for surface deformation, geomechanics, and environmental modes of the unstable rock mass, respectively. Their values ​​are jointly determined by real-time data quality (indicators such as noise and integrity) and geological environmental conditions (parameters such as rainfall and joint angles), and are learned parameters. and The loss function is automatically adjusted continuously during the training of the prediction model. For example, if the training data includes samples from periods of heavy rain, the influence of environmental factors is amplified. It will increase; when samples containing sensor malfunctions reduce the reliability of deformation data. The weights will decrease accordingly, without the need to manually set rules, ensuring that the weight changes conform to the physical laws of the actual scenario;

[0041] The multimodal fusion representation is obtained by performing element-wise scalar multiplications of the gate vector g with the feature vector. , represented as: ;

[0042] 3) The scene-level adaptability calibration layer utilizes a lightweight MetaNet network based on the scene description vector. The representation is generated sequentially through feature mapping, activation, and dimensionality adaptation, and then fused with multimodal representation. Matching scene cue vector Then, the multimodal fusion representation is performed. Scene cue vectors Element-wise multiplication yields the matched multimodal fusion representation. , represented as: This method enables the representation of modal fusion. The secondary weighting process dynamically amplifies the contribution of key features and suppresses the interference of secondary features based on the characteristics of the scene, so that the same large model can be adapted in real time without retraining under different slopes and different monitoring configurations.

[0043] Obtain the scene description vector The approach involves encoding four types of feature attributes for the new scene: inherent geological features, monitoring configuration features, environmental sensitivity features, and dynamic state features. Then, a three-layer fully connected network is used to obtain the scene description vector. ;

[0044] 4) The fusion layer will integrate scene cue vectors Multimodal fusion representation after matching Concatenate to obtain the fused feature vector , represented as: ;

[0045] Fusion feature vectors It also carries dynamic deformation features (D-Tokens weighted results), geomechanical features (G-Tokens weighted results), environmental impact features (E-Tokens weighted results), and scene-specific adaptive features. As a single dense vector, it can be directly fed into the downstream instability risk prediction module to achieve end-to-end prediction of the probability of rockfall instability, and g Both can be visualized online and manually fine-tuned, meeting the needs for interpretability and real-time parameter tuning.

[0046] S2. Acquire historical data, which includes monitoring data on surface deformation of unstable rock mass and geomechanical and environmental data. The three modes are surface deformation of unstable rock mass, geomechanical and environmental data.

[0047] S3. Construct a first database based on surface deformation monitoring data of unstable rock masses, and a second database based on geomechanical and environmental data. The construction of the first database requires the integration of multi-source spatiotemporal data. Specifically, it involves sub-pixel spatial alignment of UAV point clouds, SAR imagery, and ground camera data, preserving the abrupt changes in the deformation acceleration phase, and including data fields covering three-dimensional coordinate changes, crack aperture time series, deformation rate vectors, and microseismic energy spectra, supporting displacement field reconstruction and instability mode inversion. The second database will classify and store topographic data such as slope, aspect, and elevation; geological data such as lithology, joint occurrence, cohesion, and internal friction angle; and dynamic environmental data such as temperature, wind speed, rainfall, and pore water pressure.

[0048] S4. Preprocess the data in the first and second databases;

[0049] S5. Import the preprocessed data into the initial prediction model and use the joint loss function. Train and optimize it to obtain an optimized prediction model; joint loss function Represented as: ,in, For quantile regression loss, For divergence-based physical consistency loss, The loss function is a monotonic constraint. , and Quantile regression loss Physical consistency loss and monotonicity constraint loss function The weighting coefficients satisfy ;

[0050] quantile regression loss use The Pinball loss function applies a 5x weight to samples near the instability event. To improve the sensitivity of predictions during critical periods, as expressed as: ,in, For sample time weights;

[0051] Physical consistency loss Calculation of safety factor based on Hoek-Bray limit equilibrium equation Output probability through KL divergence constraint model and Consistent distribution ensures that the prediction conforms to rock mechanics theory, expressed as: , ;

[0052] Monotonicity constraint loss function Forced cumulative displacement As the value increases, the model predicts the probability of instability. To maintain the non-decreasing property and avoid anti-physics predictions (such as an increase in displacement but a decrease in probability), it is represented as: Among them, the relaxation threshold This allows for small negative fluctuations in the gradient, avoiding excessive constraints that could lead to model rigidity.

[0053] , , These are the weight coefficients of each loss function, satisfying... The weighting coefficients can be chosen to balance the accuracy of data-driven predictions with the consistency with physical laws. For example, if the data quality is high and the sample size is sufficient in the scenario, prediction accuracy can be prioritized, and the initial value can be set as follows: , , If the task has extremely high requirements for physical interpretability, and it is necessary to strictly ensure that physical laws are not violated, the initial value can be set to... , , .

[0054] During actual training, the weight coefficients will be dynamically adjusted adaptively according to the training stage or scenario. In the early stages of training, the model parameters are random and may seriously violate physical laws. In this case, the weights can be adjusted accordingly. , , First, anchor the model to physical constraints to prevent it from going astray. Later in training, once the model has begun to conform to physical laws, its performance can be increased. Reduced to 0.7 respectively , The weighting is adjusted to 0.2 and 0.1, focusing on improving prediction accuracy. The weighting coefficients will dynamically adjust adaptively according to the scene. If environmental noise in the input data suddenly increases, it is necessary to reduce the reliance on low-quality data and strengthen physical constraints to filter noise interference. This can be temporarily achieved by adjusting the weighting coefficients. Decrease by 0.2, Increase by 0.2; if environmental factors such as rainfall dominate instability, then the monotonicity constraint needs to be strengthened, and it can be appropriately adjusted. ,Will Reduce by 0.15; if geological factors dominate instability, then physical consistency constraints need to be strengthened, and the value can be appropriately reduced. Increase by 0.1, Decrease by 0.1.

[0055] S6. Obtain real-time monitoring data;

[0056] S7. Preprocess the real-time monitoring data;

[0057] S8. Analyze the processed real-time monitoring data using the optimized prediction model to obtain the predicted instability probability and the volume of the slip body. The prediction model output includes several core results such as the instability probability, the spatial location of the unstable rock mass, and the volume of the unstable rock mass. The slip body is the unstable rock mass.

[0058] S9. The risk threshold is dynamically adjusted based on the volume of the slip body to obtain the adjusted risk threshold T, which is expressed as: ,in, The baseline threshold is set manually. The dynamic correction term based on multi-source heterogeneous information is represented as: ,in, This represents the volume of the potentially unstable slip body obtained from the inversion of the prediction model. =1×10 4 m³ represents the normalized reference volume; A is the rock mechanics safety factor, ranging from 0 to 1; B is the real-time rainfall intensity. , , The weighting coefficients, determined by regression analysis of historical events, satisfy... + = 1.

[0059] S10. Analyze the risk results based on the adjusted risk threshold and instability probability;

[0060] S11. Determine countermeasures based on the risk results and visualize them. The visualization is as follows: the voxel probability field is rendered into Cesiumjs 3D tiles in real time, supporting profile, isosurface, and time animation playback, which can be loaded by users on the web. The system automatically extracts potential sliding rock masses, calculates their volume, centroid, main sliding direction, and average probability, and generates a PDF report.

[0061] The specific countermeasures determined based on the risk assessment results are as follows: When the predicted instability probability P ≥ T for any spatial unit, the system pushes the information to WeChat mini-programs, SMS messages, and audible and visual alarms via the Kafka message bus, and also triggers road closures and broadcast systems. The system completes the entire process from data collection to risk classification within 30 minutes, with an early warning lead time of ≥ 1 hour and a false alarm rate of < 8%, truly achieving intelligent on-site assessment and rapid decision-making for unstable rock masses on slopes.

[0062] The data preprocessing specifically includes: for static geomechanical data, physical encoding is performed, followed by mapping to the same spatial grid via coordinate projection; for dynamic environmental data and surface deformation data of unstable rock masses, timestamps are resampled and aligned based on permeability coefficients, and denoising, anomaly detection, and normalization are performed; the anomaly detection process specifically involves: setting a steady-state deformation rate threshold, applying strong filtering to stable segments below the steady-state deformation rate threshold to effectively suppress radar speckle noise; utilizing the displacement-rainfall-lithology coupling rule, when the displacement abrupt change in a sudden segment exceeds the set steady-state deformation rate threshold, automatically checking whether the concurrent rainfall exceeds the rock mass's own seepage threshold. If it does not exceed the threshold, a sensor drift correction or geological parameter verification process is triggered to perform anomaly verification and retain the abrupt change segment; the anomaly detection process incorporates relevant physical mechanisms, avoiding excessive smoothing during radar displacement sequence filtering that could lead to the loss of true deformation, thus improving data reliability.

[0063] This prediction model integrates non-contact deformation monitoring data, geomechanical information, and environmental data from multiple sources into a single interpretable multimodal model, achieving high-precision and scene-adaptive prediction of rock instability probability. The core innovation of this model lies in its three-level dynamic weight fusion mechanism, including feature-level physical correlation, modal-level reliability weighting, and scene-level adaptability calibration. This mechanism enables deep fusion and accurate utilization of multimodal data, enhancing the model's perception capabilities, inference accuracy, and scene adaptability in complex scenarios. The entire process is layered and progressive, achieving a complete weight control mechanism from micro-level correlation focusing to meso-level modal balancing, and finally to macro-level scene adaptation. The three-level dynamic weight fusion mechanism solves the problem of heterogeneous data fusion through multi-source spatiotemporal alignment. The adaptive balancing of the three-level dynamic weight fusion mechanism with the data, along with the joint loss function, ensures physical interpretability, forming a closed-loop optimized rock instability early warning model.

[0064] To address the shortcomings of traditional slope rock mass hazard assessment methods, such as reliance on human experience leading to subjective limitations, lack of global information from single-source data, low efficiency of numerical simulation resulting in frequent missed and false alarms, and difficulty in real-time coupling of multi-source data and adaptive weight adjustment, this method constructs a multimodal database, employs mechanistic constraint preprocessing and dynamic weight fusion feature vector technology, and combines joint loss function training to create a large multimodal model. This enables three-dimensional visualization of the probability of unstable rock mass in the field and a risk warning mechanism with adjustable thresholds. This allows for rapid and accurate identification of rock mass deformation areas, effectively reducing false alarm rates and improving the reliability of monitoring and early warning. Furthermore, it provides flexible and reliable auxiliary decision support for emergency rescue and comprehensive management.

[0065] A slope unstable rock mass identification and prediction device based on a large model includes:

[0066] Storage; the storage device contains computer programs;

[0067] Actuator; the actuator is used to execute a computer program stored in the memory, which, when executed, implements the slope unstable rock mass identification and prediction method based on a large model as described above.

[0068] This invention employs techniques such as multi-source spatiotemporal alignment and mechanism-constrained interpolation, mechanism-guided joint denoising and anomaly detection, terrain-adaptive feature fusion, and physical coding to preprocess multi-source data, transforming it into highly reliable features. Surface deformation, geomechanical, and environmental impact features are extracted. A three-level dynamic weight fusion—feature-level physical correlation, modal-level reliability weighting, and scene-level adaptability calibration—achieves an adaptive balance between mechanism and data, forming a fused feature vector. A joint loss function ensures physical interpretability, and dynamically adjustable weight coefficients improve the model's scene adaptability. Finally, a prediction model based on a multimodal large model is constructed. This prediction model can generate rock mass deformation cloud maps and failure probability distribution maps, quickly identifying rock mass deformation areas. Critical thresholds for disaster risk can be set according to actual on-site risk control requirements. The model configuration can be flexibly adjusted based on different rock mass disaster sites to generate rock mass disaster risk predictions, providing auxiliary decision-making for emergency rescue and comprehensive management of high slope disaster engineering. This invention can quickly and accurately identify rock mass deformation areas on slopes, reduce false alarm rates, and improve the reliability of rock mass disaster monitoring and early warning.

[0069] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for identifying and predicting unstable rock masses on slopes based on a large model, characterized in that, include: S1. Construct the initial prediction model. The prediction model includes an encoding module, a fusion module, and an instability risk prediction module from input to output. The encoding module is used to encode and align the data of the three modalities. The fusion module is used to fuse the encoded feature vectors to obtain a fused feature vector. The instability risk prediction module is used to make predictions based on the fused feature vectors to obtain the predicted instability probability. The fusion module, from input to output, consists of a feature-level physical association layer, a modality-level reliability weighting layer, a scene-level adaptability calibration layer, and a fusion layer. The feature-level physical association layer injects physical logic into the feature vector. The modality-level reliability weighting layer compresses and concatenates the spatiotemporal features of each modality and fuses all modalities to obtain a multimodal fusion representation. The scene-level adaptability calibration layer matches the multimodal fusion representation with the actual scene. The fusion layer fuses the matched multimodal fusion representation with the actual scene to obtain a fusion feature vector. The feature-level physical association layer is used to inject physical logic into the feature vector. Specifically, the feature-level physical association layer uses a cross-attention mechanism to construct the deformation query Q with the geological / environment key, and the value K with V, represented as follows: , , , , and , , Represent the feature dimension; construct the attention weight matrix Attention weight matrix Matrix representation behavior: ,in, The weight of the i-th deformation token in the feature vector of surface deformation of the unstable rock mass to the j-th token in the feature vector of geomechanics or environment is represented. , and These are the feature vectors for surface deformation of the unstable rock mass, geomechanics, and environment, respectively; the attention-weighted feature vector for surface deformation of the unstable rock mass is then calculated. ; S2. Acquire historical data, which includes monitoring data on surface deformation of unstable rock mass and geomechanical and environmental data. The three modes are surface deformation of unstable rock mass, geomechanical and environmental data. S3. Construct a first database based on surface deformation monitoring data of unstable rock masses, and construct a second database based on geomechanics and environmental data; S4. Preprocess the data in the first and second databases; S5. Import the preprocessed data into the initial prediction model, train and optimize it to obtain the optimized prediction model. S6. Obtain real-time monitoring data; S7. Preprocess the real-time monitoring data; S8. Analyze the processed real-time monitoring data using the optimized prediction model to obtain the predicted instability probability and slip volume. S9. Dynamically adjust the risk threshold according to the volume of the slip body to obtain the adjusted risk threshold. S10. Analyze the risk results based on the adjusted risk threshold and instability probability; S11. Determine response measures based on the risk results and present them visually.

2. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 1, characterized in that, The encoding module is used to encode and align data from the three modalities, specifically as follows: ； ； ； in, , and These are the feature vectors for surface deformation, geomechanics, and environment of the unstable rock mass, respectively. L is the time step, d is the embedding dimension, M is the length of the original vector after static feature stitching, and N is the time step for the environmental data. The surface deformation data of the unstable rock mass includes the surface displacement field. LOS displacement and crack opening Geomechanical data includes slope Slope aspect Elevation h, lithology code Jointed birth , Value, type of dangerous rock Rock mass unit weight Cohesion c and internal friction angle Environmental data includes temperature. Wind speed Rainfall and pore water pressure .

3. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 1, characterized in that, The fusion module is used to perform feature fusion on the encoded feature vectors to obtain the fused feature vectors, specifically including: 1) The feature-level physical correlation layer uses a cross-attention mechanism to calculate the surface deformation feature vector of the unstable rock mass. ; 2) The modal-level reliability weighted layer calculates the global mean, compressing the spatiotemporal features of each modality into a single d-dimensional vector, represented as: ; The compressed vectors are concatenated to generate the gated vector. , is represented as: ,in, , and These are the confidence weights for surface deformation, geomechanics, and environmental modes of the unstable rock mass, respectively. The multimodal fusion representation is obtained by performing element-wise scalar multiplications of the gate vector g with the feature vector. , is represented as: ; 3) The scene-level adaptability calibration layer utilizes a lightweight MetaNet network based on the scene description vector. Generate scene cue vectors Then, the multimodal fusion representation is performed. Scene cue vectors Element-wise multiplication yields the matched multimodal fusion representation. , is represented as: ; 4) The fusion layer will integrate scene cue vectors Multimodal fusion representation after matching Concatenate to obtain the fused feature vector , is represented as: .

4. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 3, characterized in that, In 3), the lightweight MetaNet network is based on the scene description vector. Generate scene cue vectors Specifically, the lightweight MetaNet network uses scene description vectors... The representation is generated sequentially through feature mapping, activation, and dimensionality adaptation, and then fused with multimodal representation. Matching scene cue vector .

5. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 3, characterized in that, Obtain the scene description vector The approach involves encoding four types of feature attributes for the new scene: inherent geological features, monitoring configuration features, environmental sensitivity features, and dynamic state features. Then, a three-layer fully connected network is used to obtain the scene description vector. .

6. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 1, characterized in that, The data preprocessing specifically includes: for static geomechanical data, physical encoding is performed and then mapped to the same spatial grid through coordinate projection; for dynamic environmental data and surface deformation data of unstable rock masses, timestamps are resampled and aligned based on permeability coefficients, and noise reduction, anomaly detection and normalization are performed.

7. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 6, characterized in that, The anomaly detection and processing are as follows: a steady-state deformation rate threshold is set, and strong filtering is applied to the stable segments below the steady-state deformation rate threshold. When the displacement of the abrupt change segment is greater than the set steady-state deformation rate threshold, the simultaneous rainfall is automatically checked to see if it exceeds the seepage threshold of the rock mass itself. If it does not exceed the threshold, the sensor drift correction or geological parameter verification process is triggered to perform anomaly verification and retain the abrupt change segment.

8. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 1, characterized in that, In S5, the joint loss function is used. The initial prediction model is trained and optimized, as follows: ,in, For quantile regression loss, For divergence-based physical consistency loss, The loss function is a monotonic constraint. , and Quantile regression loss Physical consistency loss and monotonicity constraint loss function The weighting coefficients satisfy ; quantile regression loss use The Pinball loss function applies a 5x weight to samples near the instability event. , is represented as: ,in, For sample time weights; Physical consistency loss Calculation of safety factor based on Hoek-Bray limit equilibrium equation Output probability through KL divergence constraint model and Consistent distribution ensures that the prediction conforms to rock mechanics theory, expressed as: , ; Monotonicity constraint loss function Forced cumulative displacement As the value increases, the model predicts the probability of instability. It always maintains the non-decreasing property, which can be expressed as: Among them, the relaxation threshold .

9. The method for identifying and predicting unstable rock masses on slopes based on a large model according to claim 1, characterized in that, In S9, the adjusted risk threshold T is expressed as: ,in, The baseline threshold is set manually. The dynamic correction term based on multi-source heterogeneous information is represented as: ,in, This represents the volume of the potentially unstable slip body obtained from the inversion of the prediction model. =1×10 4 m³ represents the normalized reference volume; A is the rock mechanics safety factor, ranging from 0 to 1; B is the real-time rainfall intensity. , , The weighting coefficients, determined by regression analysis of historical events, satisfy... + = 1.

10. A slope unstable rock mass identification and prediction device based on a large model, characterized in that, include: Storage; The memory contains computer programs; Actuator; The actuator is used to execute a computer program stored in the memory, which, when executed, implements the slope unstable rock mass identification and prediction method based on a large model as described in any one of claims 1-9.

Images