Intelligent monitoring and early warning method and system for dangerous rock falling of high and steep slope

Abstract

The invention discloses a high and steep slope dangerous rock falling intelligent monitoring and early warning method and system, and the method comprises the following steps: S1, collecting and preprocessing real-time monitoring data, and generating a standardized input sample set; s2, constructing a long-term prediction network model, and outputting a state prediction sequence of multiple time steps in the future; s3, optimizing structure parameters and training parameters of the prediction network model based on a bald eagle search algorithm; s4, executing multi-step state prediction by using the optimized model, and outputting a crack trend, a displacement trend and an abnormal probability value; s5, constructing an early warning risk scoring function, and fusing multiple prediction indexes to generate a risk scoring value; and S6, setting a multi-level early warning threshold value, outputting an early warning level, and issuing the early warning level in linkage with voice, a terminal and a platform. According to the invention, through constructing the intelligent monitoring and early warning method fusing the long-term prediction network model and the bald eagle search algorithm, accurate prediction and multi-stage linkage early warning of the high and steep slope dangerous rock falling risk are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring and early warning of high and steep slope disasters, and in particular to an intelligent monitoring and early warning method and system for dangerous rockfall on high and steep slopes. Background Art

[0002] In the field of geological disaster prevention and control and mountain engineering safety assurance, monitoring the stability of high and steep slopes and providing early warnings for dangerous rockfalls are critical tasks. In recent years, with the rapid development of transportation, mining, hydropower projects, and other projects in mountainous areas, the number of large slopes formed by artificial excavation has increased rapidly. High and steep slopes have complex geological structures and are significantly affected by factors such as rainfall, weathering, and earthquakes. These slopes are prone to frequent potential collapses, slides, and rockfalls, posing a serious threat to human safety and the stability of infrastructure operations. Therefore, developing efficient and accurate intelligent monitoring and early warning systems for high and steep slopes has become a hot topic in current research and engineering applications.

[0003] Existing high-steep slope monitoring and early warning methods mostly rely on sensor deployment and manual threshold judgment strategies. By deploying a variety of monitoring instruments, such as crack meters, displacement meters, accelerometers, rain gauges, and microseismic monitoring equipment, they acquire data on changes in physical quantities on and within the slope surface. These methods typically employ sliding averages, trend comparisons, and single-indicator critical value alarms for risk assessment. Some systems incorporate fuzzy reasoning, expert knowledge bases, or logical rules to construct scoring models. While these systems can provide early warnings of abnormal slope changes to a certain extent, their reliance on static rules and low-dimensional analysis often makes it difficult to fully exploit the nonlinear correlations and evolutionary patterns among multi-source monitoring data. This results in limited early warning accuracy and a significant response lag.

[0004] In recent years, with the development of artificial intelligence, particularly deep learning, the capabilities of neural network models in time series prediction and state recognition have garnered widespread attention. Some studies have attempted to incorporate models such as long short-term memory (LSTM) networks and convolutional neural networks (CNN) into slope data analysis for trend prediction or anomaly identification. However, these approaches generally face the following challenges: First, the model structures lack specificity and fail to incorporate the temporal-spatial coupling characteristics of geological hazard scenarios; second, model parameter configuration relies on empirical settings, making it difficult to achieve personalized optimal configuration for specific engineering scenarios; and third, risk identification often relies on independent classification methods, failing to integrate them deeply with trend prediction, resulting in poor system connectivity and weak risk identification capabilities.

[0005] In terms of prediction methods, the currently commonly used models mostly use fixed sliding window inputs for single-point trend prediction, ignoring the spatial differences between sensors and the driving effect of inducing factors on state evolution. The prediction results are not accurate enough in the mutation stage. In terms of model optimization, traditional training strategies mostly rely on fixed hyperparameter settings and lack intelligent tuning of model structure and training configuration, making it difficult to achieve the generalization and adaptability of the model in different slope geological environments. In terms of risk scoring, existing methods often use a single indicator drive or linear weighted combination, ignoring the dynamic coupling relationship between various state variables and the fusion of future multi-time step information, making it difficult to accurately reflect the evolutionary risk of potential collapse or rockfall events.

[0006] In addition, the existing systems generally have the following deficiencies in their response mechanisms: the anomaly identification and trend prediction modules are deployed separately, lacking information sharing and feedback mechanisms, resulting in the inability of risk identification results to assist in the adaptive correction of the prediction model; the output results are often presented as single-variable curves or static warning levels, failing to build a closed-loop warning process of multivariable joint prediction-identification-scoring, making it difficult to achieve high-precision, real-time dynamic disaster risk assessment; at the same time, the model training process often lacks the special perception ability of high-risk points, and cannot strengthen model learning and resource allocation for sensitive areas, resulting in the risk of false alarms and missed alarms in local mutation environments.

[0007] Therefore, how to provide an intelligent monitoring and early warning method and system for dangerous rockfall on steep slopes is a problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose an intelligent monitoring and early warning method and system for dangerous rockfalls on steep slopes. The present invention fully integrates deep learning time series modeling technology and swarm intelligence optimization algorithm. By constructing a long-term prediction network model and introducing a vulture search algorithm to jointly optimize structural parameters and training parameters, it describes in detail the entire process of preprocessing, trend prediction, abnormal probability identification, risk scoring and early warning response of multi-source monitoring data. It has the advantages of high prediction accuracy, timely response, strong risk identification ability and good adaptability, and is suitable for dynamic early warning applications of slope disasters under complex geological conditions.

[0009] According to an embodiment of the present invention, a method for intelligent monitoring and early warning of dangerous rockfall on a steep slope includes the following steps:

[0010] S1. Collect real-time monitoring data from multiple types of sensors installed in high and steep slope areas, synchronize the real-time monitoring data by timestamp, construct a multi-source time series dataset, perform data preprocessing on the multi-source time series dataset, and generate a standardized input sample set;

[0011] S2, build a long-term prediction network model, receive a standardized input sample set and output a sequence of prediction results for each monitoring point in multiple time steps in the future;

[0012] S3. Jointly optimize the structural parameters and training parameters of the long-term prediction network model based on the vulture search algorithm to obtain the optimal long-term prediction network model parameters;

[0013] S4. Obtain the optimized long-term prediction network model parameters and apply them to perform multi-step state prediction of real-time monitoring data, outputting the crack development trend, displacement change trend, and abnormal response probability value corresponding to each future time step;

[0014] S5. Construct an early warning risk scoring function based on the multi-step state prediction results, input the predicted crack value, displacement value, acceleration value, rainfall index and microseismic amplitude into the scoring function, and generate a risk score value for each monitoring point;

[0015] S6. Set multi-level warning risk thresholds and perform threshold comparison on the risk scores of each monitoring point. When the risk score exceeds the preset threshold, the corresponding warning level signal is output and the information is released through voice broadcast, mobile terminals and control platforms.

[0016] Optionally, the real-time monitoring data specifically includes multi-source geological and environmental time series data of crack width, rock displacement, tilt angle, acceleration, rainfall data and microseismic signals.

[0017] Optionally, the data preprocessing of the multi-source time series data set specifically includes missing value filling, outlier removal, normalization processing and sliding window sample construction.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Represent the standardized input sample set as a three-dimensional tensor Among them, T represents the number of time steps, N represents the number of monitoring points, and D represents the feature dimension of each monitoring point. is the set of real numbers;

[0020] S22. Introduce spatial position embedding vector for three-dimensional tensor X The latitude, longitude and altitude of each monitoring point are embedded into a vector representation and nonlinearly transformed through the position embedding network, where d p Indicates the dimension of the spatial position embedding vector;

[0021] S23, introduce the inducement context vector for each time step The inducement context vector includes factors affecting the rainfall rate, surface temperature difference, wind speed and humidity environment, and generates an inducement context embedding matrix through a context embedding network;

[0022] S24, inputting the spatial position embedding vector, the inducement context vector, and the three-dimensional tensor into an enhanced input layer, wherein the enhanced input layer includes a feature splicing unit and a linear projection network, and the spatial position embedding vector, the inducement context vector, and the three-dimensional tensor are fused through the enhanced input layer to form an enhanced multi-channel input sequence Z;

[0023] S25. Input the enhanced multi-channel input sequence Z to the multimodal decomposition module, which includes a trend extraction submodule, a period extraction submodule and a disturbance response submodule. The trend extraction submodule uses sliding mean filtering and gated convolution to extract long-term change trends. The period extraction submodule extracts repetitive change structures based on a sparse self-attention mechanism. The disturbance response submodule uses the inducement context feature as input to construct a dynamic weight control mechanism and utilizes a trainable disturbance perception function f. θ (C t ) guides the residual extraction process, decomposing the enhanced multi-channel input sequence Z into trend components Z (tr) , periodic component Z (se) and the disturbance component Z (no) The disturbance perception function includes a two-layer fully connected network and a dynamic normalization activation module. The dynamic normalization activation module calculates the normalized adjustment value of the current disturbance weight by inputting the inducement context features, and introduces trainable scaling and offset parameters to generate a disturbance adjustment coefficient to dynamically control the influence of the disturbance component on the prediction result.

[0024] S26. Constructing a long-term prediction network model, wherein the long-term prediction network model includes an enhanced input layer, a multimodal decomposition module, a period modeling encoder, a disturbance modeling module, a trend modeling state space module, a multi-step prediction decoder module, and an abnormal risk output module;

[0025] S27, the periodic component Z (se) Input to the periodic modeling encoder, which is built based on a multi-head sparse attention mechanism to extract periodic internal features and output a periodic feature vector sequence H e ;

[0026] S28, the disturbance component Z (no) Input to the disturbance modeling module, use the gated residual block to extract the non-stationary change characteristics dominated by external inducements, and output the disturbance feature vector sequence H n ;

[0027] S29, the trend component Z (tr) Input to the trend modeling state space module, use one-dimensional convolution and gating structure to model its long-term trend characteristics, and output trend feature vector sequence H t ;

[0028] S210, the periodic feature vector sequence H e , perturbation feature vector sequence H n , trend feature vector sequence H t The three types of feature vectors are concatenated and fused in the time dimension to generate a unified decoding input representation H d ;

[0029] S211, the unified decoding input is represented as H d Input to the multi-step prediction decoder module, which uses an autoregressive structure to simultaneously predict the state change trend of multiple time steps in the future and generate a prediction sequence

[0030] S212, input the unified decoded input representation to the abnormal risk output module, the abnormal risk output module includes an abnormal probability classification output head set in parallel with the trend prediction decoder, constructs an abnormality detection substructure, uses the abnormal probability classification output head based on the unified decoded input representation H d Output the abnormal probability vector corresponding to the future prediction sequence Form a predicted abnormal probability sequence.

[0031] Optionally, the S3 specifically includes:

[0032] S31. Initialize the vulture search algorithm population and jointly construct the structural parameters and training parameters of the long-term prediction network model into an optimization vector:

[0033] θ=[l,w,h a ,d e ,η,b];

[0034] Among them, l represents the number of coding layers, w represents the sliding window length, h a represents the number of attention heads, d e represents the embedding dimension, η represents the learning rate, and b represents the training batch size; a set of initial parameter vectors θ are generated for each vulture individual i , construct the population set P = {θ1,θ2,…,θ N}, i∈{1,2,…,N}, where N represents the population size;

[0035] S32. Based on the long-term prediction network model, the state prediction results of each monitoring point in multiple time steps in the future and the corresponding abnormal probability output are used to divide the monitoring points into risk clusters, and multiple regional risk-related sub-populations are constructed. Each sub-population independently maintains the parameter individuals in the region and runs the optimization process in parallel;

[0036] S33. Establish a co-evolutionary memory vector for each vulture individual Record several rounds of historical parameter update trajectories and generate a collaborative preference vector M based on a sliding window approach(t) , guiding the evolutionary direction of individuals in the current generation;

[0037] S34, extract the prediction residual sequence from the output of the long-term prediction network model, the prediction residual sequence refers to the time-step difference sequence between the future multi-time-step prediction value of each monitoring point output by the long-term prediction network model and the corresponding actual observation value, and calculate the residual information entropy H t and the first-order derivative ΔH t , construct the entropy-driven stage gating function G(H t ,ΔH t ), output the weight ratio of the three behavior stages [λ1,λ2,λ3], which correspond to the three behavior stages of circling search, tentative descent and dive attack respectively;

[0038] S35. Execute the vulture individual parameter update operation based on the behavior weight and the collaborative preference vector:

[0039]

[0040] Among them, λ s ∈{λ1,λ2,λ3},λ s represents the control weight corresponding to the current behavior stage, β1 and β2 are the collaborative direction guidance coefficient and the global optimal aggregation coefficient respectively, represents the current global optimal parameter combination, is the current parameter vector combination of the i-th vulture individual in the t+1th iteration, is the current parameter vector combination of the i-th vulture individual in the t-th iteration;

[0041] S36, mark the monitoring points in the predicted abnormal probability sequence that are greater than the set threshold δ as high-risk points, count the distribution density of high-risk points in the spatial area and construct the risk perception state vector S t ;

[0042] S37, build a reinforcement learning agent module to perceive the risk state vector S t As the state input, the individual fitness improvement of the vulture is used as the reward function, and the output behavior recommendation action a t , as a dynamic regulation signal of the weight ratio of the behavior stage;

[0043] S38. Construct a fitness function f(θ) to measure the prediction accuracy, risk response consistency, and structural complexity of each parameter combination:

[0044]

[0045] in, is the predicted result, Y is the true value, is the subset of monitoring points with abnormal probability higher than the threshold, ‖θ‖2 is the L2 norm of the parameter vector, dim(θ) represents the number of parameter dimensions, represents the mean square error between the predicted value and the actual value, The variance of the subset of monitoring points whose anomaly probability is higher than the threshold;

[0046] S39, every set round T s Perform an individual migration operation between subpopulations. Based on the risk level and fitness score of each area, individuals with the best fitness are migrated from high-risk areas to low-risk areas, optimizing the population search diversity and global convergence ability.

[0047] S310, when the maximum number of iterations is reached or the fitness change for multiple rounds is lower than the set convergence threshold, the parameter combination θ with the minimum fitness function value is output g , as the final optimized parameter configuration of the long-term prediction network model.

[0048] Optionally, the S4 specifically includes:

[0049] S41, obtaining structural parameters and training parameters of the optimized long-term prediction network model, constructing the final optimized long-term prediction network model as a state prediction module to receive real-time monitoring data and perform multi-step prediction tasks;

[0050] S42. Using the standardized input sample set as the main input, the sensor activity level, measurement point status, and risk score of the previous cycle at the current time step as auxiliary dynamic guidance information are input into the optimized long-term prediction network model input interface to construct a multi-source enhanced input structure integrating dynamic control factors.

[0051] S43. Within the state prediction module, a dual-channel state prediction mechanism based on temporal structure decoupling is executed, wherein the trend prediction channel outputs the crack development trend, and the displacement prediction channel outputs the displacement change trend. The two channels share encoder features and enhance the interrelated features through a cross-trend attention mechanism.

[0052] S44, based on the abnormal risk parallel classification branch output by the optimized long-term prediction network model, combined with the trend change amplitude, sensor feature distribution and dynamic feedback information corresponding to each time step, to generate the abnormal response probability value of each monitoring point in each future time step;

[0053] S45. Output the crack development trend, displacement change trend, and abnormal response probability value in a structured manner, and automatically index and organize them based on the monitoring point type and prediction dimension;

[0054] S46. Construct a consistency evaluation mechanism for prediction results, conduct dynamic consistency judgment on the joint change pattern of the crack development trend sequence and the displacement change trend sequence within the prediction time window, and adjust the update strategy of the long-term prediction network model based on the consistency evaluation results.

[0055] Optionally, the S5 specifically includes:

[0056] S51, receiving multi-step state prediction results output by the long-term prediction network model, including crack development trend, displacement change trend and abnormal response probability value;

[0057] S52, classify and summarize the multi-step state prediction results according to the monitoring point number, construct a multi-dimensional prediction information set based on the monitoring point, and complete and mark incomplete data;

[0058] S53. Construct a risk score input structure based on the predicted data of each monitoring point, wherein the risk score input structure includes a prediction sequence of crack values, displacement values, acceleration values, rainfall indexes, and microseismic amplitudes in each future time step;

[0059] S54, constructing an early warning risk scoring function for comprehensively evaluating the potential risk level of each monitoring point, wherein the early warning risk scoring function is calculated based on the weight relationship between the crack development trend, the displacement change trend, and the abnormal response probability value and the time step characteristics;

[0060] S55, inputting the crack development trend, displacement change trend and abnormal response probability value of each monitoring point into the early warning risk scoring function, and calculating the risk score value of each monitoring point in the entire prediction time window;

[0061] S56. The risk score value of each monitoring point is used as the input basis for the subsequent warning level determination and information linkage release mechanism to realize the intelligent warning process of dangerous rock falls on steep slopes.

[0062] An intelligent monitoring and early warning system for dangerous rockfall on steep slopes according to an embodiment of the present invention includes the following modules:

[0063] Data acquisition and preprocessing module, used to collect real-time monitoring data for preprocessing and generate standardized input sample sets;

[0064] Model construction and optimization module, which is used to build a long-term prediction network model and jointly optimize the structural parameters and training parameters using the vulture search algorithm;

[0065] The state prediction module is used to apply the optimized long-term prediction network model to the standardized input sample set to perform multi-step state prediction;

[0066] The anomaly recognition module is used to output the abnormal response probability value sequence of each monitoring point in the future time step based on the parallel classification branch of the optimized long-term prediction network model;

[0067] Result output module, used to organize and structure the multi-step state prediction results;

[0068] The risk scoring module is used to input multiple types of prediction indicators into the scoring function to generate the risk score value of each monitoring point;

[0069] The early warning response module is used to divide the early warning level according to the risk score value, and release it through voice broadcast, mobile terminal and control platform when the threshold is exceeded.

[0070] The beneficial effects of the present invention are:

[0071] The intelligent monitoring and early warning method for dangerous rockfalls on steep slopes provided by the present invention is based on the deep integration of a long-term prediction network model and a vulture search algorithm. It breaks through key bottlenecks in existing technologies, such as weak multi-source monitoring data modeling capabilities, poor model generalization adaptability, and delayed early warning results, and significantly improves the dynamic perception and intelligent response capabilities of slope disaster risks. By performing time synchronization and standardization on key physical quantities such as cracks, displacements, accelerations, rainfall, and microseismic data collected by sensors in real time, a standardized input sample set is constructed, and the long-term prediction network model predicts the state evolution trend of each monitoring point in multiple time steps in the future, thereby achieving early prediction of crack development and displacement changes. The vulture search algorithm is further introduced to jointly optimize the model structure parameters and training parameters, so that the model has better adaptability and prediction accuracy, and meets the dynamic feature extraction requirements of diverse monitoring data in complex geological scenarios.

[0072] The present invention sets parallel classification branches to output the abnormal response probability value of each monitoring point in each future time step while predicting the trend, thereby realizing the collaborative modeling of anomaly detection and trend prediction, and solving the problem of separation and difficulty in linkage between abnormal response and state prediction in the prior art. The constructed early warning risk scoring function comprehensively considers multi-dimensional prediction information such as crack value, displacement value, acceleration value, rainfall index and microseismic amplitude, and can accurately assess the risk level of the monitoring point, and accordingly sets multi-level early warning thresholds to output early warning level signals, forming a complete prediction-scoring-early warning closed-loop mechanism. In addition, the system also introduces mechanisms such as co-evolutionary memory, entropy-driven behavior control, and reinforcement learning strategy guidance, which strengthens the model's ability to focus on high-risk areas and effectively improves the recognition rate of local mutation risks and the overall stability of the system.

[0073] Compared with existing methods, the present invention realizes data-driven high-precision state prediction, model adaptive adjustment under structural optimization, deep integration of prediction results and risk judgment, and real-time linkage output of warning results. It has the beneficial effects of high prediction accuracy, fast warning response, strong adaptability and high deployment efficiency. It is suitable for high and steep slope disaster monitoring and intelligent warning tasks in various complex engineering scenarios such as transportation, water conservancy, and mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0075] Figure 1 This is a flow chart of an intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes proposed by the present invention;

[0076] Figure 2 This is a structural diagram of an intelligent monitoring and early warning system for dangerous rockfall on steep slopes proposed by the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0078] refer to Figure 1 , an intelligent monitoring and early warning method for dangerous rockfall on steep slopes, comprising the following steps:

[0079] S1. Collect real-time monitoring data from multiple types of sensors installed in high and steep slope areas, synchronize the real-time monitoring data by timestamp, construct a multi-source time series dataset, perform data preprocessing on the multi-source time series dataset, and generate a standardized input sample set;

[0080] S2, build a long-term prediction network model, receive a standardized input sample set and output a sequence of prediction results for each monitoring point in multiple time steps in the future;

[0081] S3. Jointly optimize the structural parameters and training parameters of the long-term prediction network model based on the vulture search algorithm to obtain the optimal long-term prediction network model parameters;

[0082] S4. Obtain the optimized long-term prediction network model parameters and apply them to perform multi-step state prediction of real-time monitoring data, outputting the crack development trend, displacement change trend, and abnormal response probability value corresponding to each future time step;

[0083] S5. Construct an early warning risk scoring function based on the multi-step state prediction results, input the predicted crack value, displacement value, acceleration value, rainfall index and microseismic amplitude into the scoring function, and generate a risk score value for each monitoring point;

[0084] S6. Set multi-level warning risk thresholds and perform threshold comparison on the risk scores of each monitoring point. When the risk score exceeds the preset threshold, the corresponding warning level signal is output and the information is released through voice broadcast, mobile terminals and control platforms.

[0085] This invention achieves efficient prediction and dynamic response to dangerous rockfall risks in high-steep slope areas by constructing a multi-step intelligent monitoring and early warning process. By collecting monitoring data in real time through multiple types of sensors and constructing a multi-source time-series dataset, the comprehensiveness and timeliness of data acquisition are improved. A long-term prediction network model is used to perform multi-step predictions on standardized input samples, and the model's prediction accuracy and generalization capabilities are significantly improved through the optimization of the vulture search algorithm. The model outputs, including crack development trends, displacement change trends, and abnormal response probabilities, comprehensively reflect the future evolution of the slope structure. Based on this, a warning risk scoring function is constructed that comprehensively considers multiple key indicators to quantify the potential risk of each monitoring point. Multi-level warning thresholds are set based on the risk score to ensure the timeliness and classification accuracy of warning signals. Finally, through voice broadcasting, mobile terminals, and control platforms, this system is linked and released, establishing an integrated closed-loop system of information collection, state prediction, risk scoring, and early warning response, effectively enhancing the intelligent level and practical value of high-steep slope disaster prevention and control.

[0086] In this embodiment, the real-time monitoring data specifically includes multi-source geological and environmental time series data of crack width, rock displacement, tilt angle, acceleration, rainfall data and microseismic signals.

[0087] In this embodiment, the data preprocessing of the multi-source time series data set specifically includes missing value filling, outlier removal, normalization processing and sliding window sample construction.

[0088] In this embodiment, S2 specifically includes:

[0089] S21. Represent the standardized input sample set as a three-dimensional tensor Among them, T represents the number of time steps, N represents the number of monitoring points, and D represents the feature dimension of each monitoring point. is the set of real numbers;

[0090] S22. Introduce spatial position embedding vector for three-dimensional tensor X The latitude, longitude and altitude of each monitoring point are embedded into a vector representation and nonlinearly transformed through the position embedding network, where d pIndicates the dimension of the spatial position embedding vector;

[0091] S23, introduce the inducement context vector for each time step The inducement context vector includes factors affecting the rainfall rate, surface temperature difference, wind speed and humidity environment, and generates an inducement context embedding matrix through a context embedding network;

[0092] S24, inputting the spatial position embedding vector, the inducement context vector, and the three-dimensional tensor into an enhanced input layer, wherein the enhanced input layer includes a feature splicing unit and a linear projection network, and the spatial position embedding vector, the inducement context vector, and the three-dimensional tensor are fused through the enhanced input layer to form an enhanced multi-channel input sequence Z;

[0093] S25. Input the enhanced multi-channel input sequence Z to the multimodal decomposition module, which includes a trend extraction submodule, a period extraction submodule and a disturbance response submodule. The trend extraction submodule uses sliding mean filtering and gated convolution to extract long-term change trends. The period extraction submodule extracts repetitive change structures based on a sparse self-attention mechanism. The disturbance response submodule uses the inducement context feature as input to construct a dynamic weight control mechanism and utilizes a trainable disturbance perception function f. θ (C t ) guides the residual extraction process, decomposing the enhanced multi-channel input sequence Z into trend components Z (tr) , periodic component Z (se) and the disturbance component Z (no) The disturbance perception function includes a two-layer fully connected network and a dynamic normalization activation module. The dynamic normalization activation module calculates the normalized adjustment value of the current disturbance weight by inputting the inducement context features, and introduces trainable scaling and offset parameters to generate a disturbance adjustment coefficient to dynamically control the influence of the disturbance component on the prediction result.

[0094] S26. Constructing a long-term prediction network model, wherein the long-term prediction network model includes an enhanced input layer, a multimodal decomposition module, a period modeling encoder, a disturbance modeling module, a trend modeling state space module, a multi-step prediction decoder module, and an abnormal risk output module;

[0095] S27, the periodic component Z (se) Input to the periodic modeling encoder, which is built based on a multi-head sparse attention mechanism to extract periodic internal features and output a periodic feature vector sequence H e ;

[0096] S28, the disturbance component Z (no) Input to the disturbance modeling module, use the gated residual block to extract the non-stationary change characteristics dominated by external inducements, and output the disturbance feature vector sequence Hn ;

[0097] S29, the trend component Z (tr) Input to the trend modeling state space module, use one-dimensional convolution and gating structure to model its long-term trend characteristics, and output trend feature vector sequence H t ;

[0098] S210, the periodic feature vector sequence H e , perturbation feature vector sequence H n , trend feature vector sequence H t The three types of feature vectors are concatenated and fused in the time dimension to generate a unified decoding input representation H d ;

[0099] S211, the unified decoding input is represented as H d Input to the multi-step prediction decoder module, which uses an autoregressive structure to simultaneously predict the state change trend of multiple time steps in the future and generate a prediction sequence

[0100] S212, input the unified decoded input representation to the abnormal risk output module, the abnormal risk output module includes an abnormal probability classification output head set in parallel with the trend prediction decoder, constructs an abnormality detection substructure, uses the abnormal probability classification output head based on the unified decoded input representation H d Output the abnormal probability vector corresponding to the future prediction sequence Form a predicted abnormal probability sequence.

[0101] The long-term prediction network model construction method described in this invention effectively integrates the geographic attributes of monitoring points with external environmental factors by introducing spatial location embedding and causal context embedding, enhancing the input representation's ability to express multi-source heterogeneous slope information. By enhancing the input layer and multimodal decomposition module, the input data is decomposed into trend, periodic, and perturbation components, corresponding to long-term evolution, repetitive fluctuations, and non-stationary perturbation characteristics, respectively, enabling structured modeling of the slope evolution mechanism. The periodic modeling encoder, perturbation modeling module, and trend modeling state-space module extract features for different component types, respectively, enhancing the model's ability to decouple complex dynamic states. The decoder module generates future state trends through multi-step prediction, and the anomaly risk output module outputs future anomaly probabilities in parallel, forming a joint modeling path for state prediction and anomaly identification. This method possesses the ability to model multiple influencing factors, decompose complex structural components, and discriminate future risks, improving the model's prediction accuracy, interpretability, and anomaly response sensitivity under the dynamic changes of steep slopes.

[0102] In this embodiment, S3 specifically includes:

[0103] S31. Initialize the vulture search algorithm population and jointly construct the structural parameters and training parameters of the long-term prediction network model into an optimization vector:

[0104] θ=[l,w,h a ,d e ,η,b];

[0105] Among them, l represents the number of coding layers, w represents the sliding window length, h a represents the number of attention heads, d e represents the embedding dimension, η represents the learning rate, and b represents the training batch size; a set of initial parameter vectors θ are generated for each vulture individual i , construct the population set P = {θ1,θ2,…,θ N}, i∈{1,2,…,N}, where N represents the population size;

[0106] S32. Based on the long-term prediction network model, the state prediction results of each monitoring point in multiple time steps in the future and the corresponding abnormal probability output are used to divide the monitoring points into risk clusters, and multiple regional risk-related sub-populations are constructed. Each sub-population independently maintains the parameter individuals in the region and runs the optimization process in parallel;

[0107] S33. Establish a co-evolutionary memory vector for each vulture individual Record several rounds of historical parameter update trajectories and generate a collaborative preference vector M based on a sliding window approach (t) , guiding the evolutionary direction of individuals in the current generation;

[0108] S34, extract the prediction residual sequence from the output of the long-term prediction network model, the prediction residual sequence refers to the time-step difference sequence between the future multi-time-step prediction value of each monitoring point output by the long-term prediction network model and the corresponding actual observation value, and calculate the residual information entropy H t and the first-order derivative ΔH t , construct the entropy-driven stage gating function G(H t ,ΔH t ), output the weight ratio of the three behavior stages [λ1,λ2,λ3], which correspond to the three behavior stages of circling search, tentative descent and dive attack respectively;

[0109] S35. Execute the vulture individual parameter update operation based on the behavior weight and the collaborative preference vector:

[0110]

[0111] Among them, λ s ∈{λ1,λ2,λ3},λ s represents the control weight corresponding to the current behavior stage, β1 and β2 are the collaborative direction guidance coefficient and the global optimal aggregation coefficient respectively, represents the current global optimal parameter combination, is the current parameter vector combination of the i-th vulture individual in the t+1th iteration, is the current parameter vector combination of the i-th vulture individual in the t-th iteration;

[0112] Execute condor individual parameters The update operation formula represents the updating process of individual parameter vectors in the improved vulture search algorithm, comprehensively considering the influence of the current individual position, the global optimal solution, and the historical co-evolutionary direction. Its practical significance lies in achieving efficient search and dynamic adjustment of parameters in the long-term prediction network model. This formula controls the evolutionary behavior of individuals by introducing three key factors: the collaborative preference vector, which guides individuals along historically favorable directions, enhancing the stability and memory of the search; the global optimal parameter combination, which guides the population toward the current optimal solution and accelerates convergence; and the behavioral stage weight, which is dynamically adjusted based on the information entropy of the prediction residual to determine whether the current phase is exploration, trial, or convergence, thereby improving the algorithm's adaptability to different optimization stages. Through this combination of control variables, the algorithm maintains diversity while enhancing convergence, effectively avoiding local optima and ensuring that the search results for the optimal configuration of long-term prediction network structural parameters and training parameters are more accurate and robust in dynamic slope monitoring tasks, thereby improving the overall system's prediction performance and risk response capabilities.

[0113] S36, mark the monitoring points in the predicted abnormal probability sequence that are greater than the set threshold δ as high-risk points, count the distribution density of high-risk points in the spatial area and construct the risk perception state vector S t ;

[0114] S37, build a reinforcement learning agent module to perceive the risk state vector S t As the state input, the individual fitness improvement of the vulture is used as the reward function, and the output behavior recommendation action a t , as a dynamic regulation signal of the weight ratio of the behavior stage;

[0115] S38. Construct a fitness function f(θ) to measure the prediction accuracy, risk response consistency, and structural complexity of each parameter combination:

[0116]

[0117] in, is the predicted result, Y is the true value, is the subset of monitoring points with abnormal probability higher than the threshold, ‖θ‖2 is the L2 norm of the parameter vector, dim(θ) represents the number of parameter dimensions, represents the mean square error between the predicted value and the actual value, The variance of the subset of monitoring points whose anomaly probability is higher than the threshold;

[0118] The fitness function f(θ) formula is practically significant in that it constructs a multi-objective evaluation criterion to guide the optimization process toward a goal that simultaneously achieves high prediction accuracy, consistent anomaly responses, and low structural complexity. The first term in the formula is the mean squared error, which measures the deviation between the model's predicted values ​​and the actual observed values, reflecting the model's basic predictive capability. The second term is the anomaly probability variance, which assesses the consistency of the model's response to anomalies at high-risk monitoring points. Large fluctuations in anomaly probability can lead to unstable warnings. The third term is the normalized L2 norm of the parameter, which controls model complexity to prevent overfitting and excessive resource consumption. Through the weighted combination of these three components, the fitness function achieves a comprehensive balance between prediction accuracy, risk identification stability, and model simplicity. It guides the vulture search algorithm to find the optimal parameter combination within the multi-objective optimization space, ensuring that the resulting prediction model is both engineering-ready and capable of accurately detecting the evolutionary trends and sudden anomalies of high-steep slope hazards.

[0119] S39, every set round T s Perform an individual migration operation between subpopulations. Based on the risk level and fitness score of each area, individuals with the best fitness are migrated from high-risk areas to low-risk areas, optimizing the population search diversity and global convergence ability.

[0120] S310, when the maximum number of iterations is reached or the fitness change for multiple rounds is lower than the set convergence threshold, the parameter combination θ with the minimum fitness function value is output g , as the final optimized parameter configuration of the long-term prediction network model.

[0121] By optimizing the structure of the vulture search algorithm, this paper proposes an improved mechanism for joint optimization of parameters in a long-term prediction network model, enhancing the model's adaptive tuning capabilities and search convergence efficiency in complex slope monitoring scenarios. By uniformly encoding structural and training parameters to construct an optimization vector, risk clustering is performed based on prediction results and anomaly probabilities, forming regional risk-associated subpopulations and implementing a differentiated optimization strategy. A co-evolutionary memory mechanism enhances the utilization of historical update trajectories, improving the continuity and stability of individual evolutionary directions. A behavioral gating function is introduced during the residual information entropy construction phase to achieve dynamic weight control for the circling search, tentative descent, and dive attack phases. The integration of a co-preference vector with the global optimal solution guides individual evolution, improving optimization directionality and convergence speed. Furthermore, by identifying high-risk points and modeling risk perception states, combined with a reinforcement learning agent module to adjust behavioral strategies, the search capability for local anomaly areas is enhanced. A multi-objective fitness function comprehensively evaluates individual performance based on prediction accuracy, anomaly consistency, and structural complexity. Global convergence is achieved through subpopulation migration and an optimal solution output mechanism. The overall optimization process improves the robustness, accuracy and adaptability of the prediction model, and effectively enhances the system's ability to model the complex dynamic evolution process of the slope.

[0122] In this embodiment, the S4 specifically includes:

[0123] S41, obtaining structural parameters and training parameters of the optimized long-term prediction network model, constructing the final optimized long-term prediction network model as a state prediction module to receive real-time monitoring data and perform multi-step prediction tasks;

[0124] S42. Using the standardized input sample set as the main input, the sensor activity level, measurement point status, and risk score of the previous cycle at the current time step as auxiliary dynamic guidance information are input into the optimized long-term prediction network model input interface to construct a multi-source enhanced input structure integrating dynamic control factors.

[0125] S43. Within the state prediction module, a dual-channel state prediction mechanism based on temporal structure decoupling is executed, wherein the trend prediction channel outputs the crack development trend, and the displacement prediction channel outputs the displacement change trend. The two channels share encoder features and enhance the interrelated features through a cross-trend attention mechanism.

[0126] S44, based on the abnormal risk parallel classification branch output by the optimized long-term prediction network model, combined with the trend change amplitude, sensor feature distribution and dynamic feedback information corresponding to each time step, to generate the abnormal response probability value of each monitoring point in each future time step;

[0127] S45. Output the crack development trend, displacement change trend, and abnormal response probability value in a structured manner, and automatically index and organize them based on the monitoring point type and prediction dimension;

[0128] S46. Construct a consistency evaluation mechanism for prediction results, conduct dynamic consistency judgment on the joint change pattern of the crack development trend sequence and the displacement change trend sequence within the prediction time window, and adjust the update strategy of the long-term prediction network model based on the consistency evaluation results.

[0129] By constructing a state prediction module with dynamic input guidance, dual-channel decoupled modeling, and joint anomaly identification capabilities, this invention significantly improves the modeling accuracy and response efficiency of long-term prediction network models for the complex state evolution of steep slopes. Dynamic guidance factors such as sensor activity, measurement point status, and previous cycle risk scores are introduced to construct a multi-source enhanced input structure, enabling the model to perceive changes in the monitoring environment in real time and enhancing the ability of input information to regulate prediction results. A structure in which trend prediction channels and displacement prediction channels operate in parallel, combined with a cross-trend attention mechanism, enables information sharing between trends and enhances associated features, improving the model's ability to express the dynamic coupling relationships of multivariate states. The anomaly risk classification branch jointly outputs anomaly probability values ​​based on trend amplitude, feature distribution, and feedback information, addressing the issues of lagging and isolated anomaly detection in traditional methods. The final output results are structured and automatically indexed, improving the efficiency of system data interaction and result analysis. A consistency assessment mechanism dynamically determines the joint change pattern of crack and displacement prediction results and optimizes the model update strategy accordingly, achieving closed-loop control of prediction and feedback. Overall, this method has higher prediction accuracy, anomaly recognition sensitivity and system adaptability, significantly enhancing the intelligence and real-time level of disaster risk identification.

[0130] In this embodiment, the S5 specifically includes:

[0131] S51, receiving multi-step state prediction results output by the long-term prediction network model, including crack development trend, displacement change trend and abnormal response probability value;

[0132] S52, classify and summarize the multi-step state prediction results according to the monitoring point number, construct a multi-dimensional prediction information set based on the monitoring point, and complete and mark incomplete data;

[0133] S53. Construct a risk score input structure based on the predicted data of each monitoring point, wherein the risk score input structure includes a prediction sequence of crack values, displacement values, acceleration values, rainfall indexes, and microseismic amplitudes in each future time step;

[0134] S54, constructing an early warning risk scoring function for comprehensively evaluating the potential risk level of each monitoring point, wherein the early warning risk scoring function is calculated based on the weight relationship between the crack development trend, the displacement change trend, and the abnormal response probability value and the time step characteristics;

[0135] S55, inputting the crack development trend, displacement change trend and abnormal response probability value of each monitoring point into the early warning risk scoring function, and calculating the risk score value of each monitoring point in the entire prediction time window;

[0136] S56. The risk score value of each monitoring point is used as the input basis for the subsequent warning level determination and information linkage release mechanism to realize the intelligent warning process of dangerous rock falls on steep slopes.

[0137] This invention achieves quantitative assessment and intelligent grading of future state risks at monitoring points on steep slopes by constructing a multivariate-driven early warning risk scoring mechanism. By receiving crack development trends, displacement change trends, and abnormal response probability values ​​output by a long-term prediction network model, the system comprehensively understands the evolutionary trends and potential abnormalities at monitoring points over the future time period. Prediction results are categorized and aggregated by monitoring point number to form a clearly structured multidimensional prediction information set. Missing or incomplete data is supplemented and marked to ensure the integrity and accuracy of the scoring data. Furthermore, the constructed early warning risk scoring input structure integrates key indicators such as crack values, displacement values, acceleration, rainfall, and microseismic amplitude. The scoring function dynamically combines the weights of different prediction indicators and their temporal characteristics to form a highly targeted risk measurement method. The scoring result for each monitoring point reflects its overall risk level over multiple future time steps and serves as the core basis for determining warning levels and disseminating information, forming a continuous and efficient risk conversion output mechanism. This method improves the sensitivity, accuracy, and automation of slope hazard early warnings, enables refined characterization of monitoring point-level risks and early warning responses, and provides reliable technical support for the prevention and control of major geological disasters.

[0138] refer to Figure 2 , an intelligent monitoring and early warning system for dangerous rockfall on steep slopes, including the following modules:

[0139] Data acquisition and preprocessing module, used to collect real-time monitoring data for preprocessing and generate standardized input sample sets;

[0140] Model construction and optimization module, which is used to build a long-term prediction network model and jointly optimize the structural parameters and training parameters using the vulture search algorithm;

[0141] The state prediction module is used to apply the optimized long-term prediction network model to the standardized input sample set to perform multi-step state prediction;

[0142] The anomaly recognition module is used to output the abnormal response probability value sequence of each monitoring point in the future time step based on the parallel classification branch of the optimized long-term prediction network model;

[0143] Result output module, used to organize and structure the multi-step state prediction results;

[0144] The risk scoring module is used to input multiple types of prediction indicators into the scoring function to generate the risk score value of each monitoring point;

[0145] The early warning response module is used to divide the early warning level according to the risk score value, and release it through voice broadcast, mobile terminal and control platform when the threshold is exceeded.

[0146] Example 1:

[0147] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the entrance and exit section of a certain mountain tunnel. To ensure road safety during the construction period and the subsequent traffic opening period, the local geological disaster monitoring department introduced the "intelligent monitoring and early warning system for dangerous rockfalls on steep slopes" proposed in the present invention, and conducted continuous monitoring and early warning pilot projects on a typical steep rock slope in the surrounding area.

[0148] The rock mass in this area is fractured, with slopes exceeding 70° and a height of approximately 95 meters. Rainfall is frequent year-round, and there are localized pockets of loose rock and developed joints. Historically, this has resulted in numerous small-scale rockfalls, posing a serious threat to the safety of the passages below. Traditional manual inspections are characterized by long inspection cycles, numerous blind spots, and delayed response to sudden disasters. There is an urgent need for a new intelligent, all-weather, and low-power monitoring and early warning system.

[0149] First, a drone equipped with an oblique camera captured high-precision terrain texture and point cloud data of the slope. Combined with ground-based laser scanning and total station measurements, the slope was modeled in 3D using CAD software. This data was then used for numerical simulation analysis using FLAC3D and Rockfall software to comprehensively assess slope stability and potential rockfall paths. The simulation identified five potential slip or collapse points. Based on the simulation results and the impact range of rockfall, ten sets of the rockfall monitoring devices described in this invention were deployed from the toe to the top of the slope.

[0150] Each device is equipped with a seismic wave detector, a high-definition wide-angle camera, a solar panel, a signal transmission module, and a rock-embedded barb fixing structure. It provides all-weather power supply and embedded stability, ensuring continuous operation in environments such as mountain disturbances, rain, fog, and high winds. The device also features combined image and seismic wave monitoring. Seismic signal monitoring is enabled by default, and the camera is automatically activated to capture images when abnormal amplitudes are detected.

[0151] All sensor data is uploaded in real time to a central signal receiving and processing system via an encrypted wireless protocol. The system is equipped with an intelligent analysis module based on deep learning and anomaly recognition. It uses a long-term prediction network model to fuse sensor data for trend analysis, and combines it with a vulture search algorithm to continuously optimize the model structure and training parameters. The model outputs crack trends, displacement trends, and abnormal response probabilities for each monitoring point over multiple future time steps, while quantifying the risk level using a predefined risk scoring function. This scoring function integrates multiple indicators, such as acceleration rate, microseismic amplitude, rainfall intensity, and crack evolution rate, to automatically calculate the risk score for each monitoring point and trigger a graded early warning mechanism.

[0152] At 17:32 on April 23, 2024, strong seismic wave signals (maximum acceleration of 0.36m / s) were continuously captured at the P4 monitoring point on the slope. 2 , microseismic amplitude reached 155μV), the system automatically triggered image acquisition, and the image recognition module identified the obvious displacement of rock blocks at the top of the slope and quickly located the abnormal area. The model prediction results showed that the crack growth trend at this point would exceed 1.6mm / h in the next hour, the abnormal probability value increased to 0.91, and the risk score reached 89.6 points (the high-risk threshold is 80 points). The system completed the processing within 30 seconds after the abnormality occurred and issued a three-level early warning instruction to engineering personnel, inspection vehicles, and management APP within a 5-kilometer range, successfully closing the channel. Subsequently, a local slope collapse actually occurred, but fortunately no casualties were caused.

[0153] The following are the prediction and response data of 8 typical monitoring points in the system operation during this test period, which verify the system's early identification and early warning response effects for high-risk conditions.

[0154] Table 1 Typical monitoring point prediction and risk response effect table

[0155]

[0156] As can be seen from the data in Table 1, the system combines the output results of the long-term prediction network model with the scoring function built based on multiple indicators to comprehensively evaluate the potential risks of each monitoring point in the future time period, and triggers early warning responses according to the risk score value.

[0157] Judging from the two monitoring indicators of acceleration and microseismic amplitude, the peak acceleration of monitoring point P4 reached 0.36m / s 2 The microseismic amplitude reached 155.2μV, the highest in the table. Combined with the crack development trend of 1.64mm / h and the abnormal probability value of 0.91, the system identified this area as a high-risk area, with a corresponding risk score of 89.6, triggering a Level 3 warning. In actual application, a local collapse event did occur at this location after the warning, demonstrating the system's high accuracy and proactiveness in identifying and responding to high-risk trends.

[0158] For example, P3 and P6 were also identified by the system as high-risk areas when the crack trend value exceeded 1.2 mm / h and the microseismic amplitude exceeded 130 μV, triggering level 2 and level 3 warnings, respectively. This demonstrates that the model not only responds to static structural changes but also has a strong ability to capture dynamic risks caused by external disturbances. While P1 and P8 had relatively moderate indicators, they still triggered level 1 and level 2 warnings due to their comprehensive scores exceeding 70 points, demonstrating that the system can maintain reasonable risk assessment sensitivity under boundary conditions.

[0159] It is worth noting that despite slight fluctuations in some individual indicators, P5 and P7 achieved comprehensive scores of 24.8 and 19.3, respectively, which did not reach the warning threshold, and the system did not falsely issue warnings. This demonstrates that the scoring mechanism is stable in avoiding false alarms and can effectively constrain over-responses in non-abnormal situations, demonstrating the superiority of the present invention in terms of model robustness and practical tolerance design.

[0160] Comprehensive analysis results show that this system, by integrating multi-source data (acceleration, microseismicity, cracks, rainfall, etc.), a multi-step prediction model, and an adaptive optimization algorithm, achieves precise slope disaster prediction, real-time response, and intelligent early warning. The system can not only identify high-risk locations in advance and output warning levels promptly, but also maintain high discrimination accuracy in non-risk areas. This effectively addresses the shortcomings of traditional methods in terms of weak disaster prediction capabilities, delayed response, and frequent false alarms, fully demonstrating the application advantages and engineering value of this invention in actual slope engineering scenarios.

[0161] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent monitoring and early warning method for dangerous rockfall on steep slopes, characterized in that: The steps include: S1. Collect real-time monitoring data from multiple types of sensors installed in high and steep slope areas, synchronize the real-time monitoring data by timestamp, construct a multi-source time series dataset, perform data preprocessing on the multi-source time series dataset, and generate a standardized input sample set; S2, build a long-term prediction network model, receive a standardized input sample set and output a sequence of prediction results for each monitoring point in multiple time steps in the future; S3. Jointly optimize the structural parameters and training parameters of the long-term prediction network model based on the vulture search algorithm to obtain the optimal long-term prediction network model parameters; S4. Obtain the optimized long-term prediction network model parameters and apply them to perform multi-step state prediction of real-time monitoring data, outputting the crack development trend, displacement change trend, and abnormal response probability value corresponding to each future time step; S5. Construct an early warning risk scoring function based on the multi-step state prediction results, input the predicted crack value, displacement value, acceleration value, rainfall index and microseismic amplitude into the scoring function, and generate a risk score value for each monitoring point; S6. Set multi-level warning risk thresholds and perform threshold comparison on the risk scores of each monitoring point. When the risk score exceeds the preset threshold, the corresponding warning level signal is output and the information is released through voice broadcast, mobile terminals and control platforms.

2. The intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes according to claim 1 is characterized in that: The real-time monitoring data specifically includes multi-source geological and environmental time series data of crack width, rock displacement, tilt angle, acceleration, rainfall data and microseismic signals.

3. The intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes according to claim 1 is characterized in that: The data preprocessing of the multi-source time series data set specifically includes missing value filling, outlier removal, normalization and sliding window sample construction.

4. The intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes according to claim 1 is characterized in that: The S2 specifically includes: S21. Represent the standardized input sample set as a three-dimensional tensor Among them, T represents the number of time steps, N represents the number of monitoring points, and D represents the feature dimension of each monitoring point. is the set of real numbers; S22. Introduce spatial position embedding vector for three-dimensional tensor X The latitude, longitude and altitude of each monitoring point are embedded into a vector representation and nonlinearly transformed through the position embedding network, where d p Indicates the dimension of the spatial position embedding vector; S23, introduce the inducement context vector for each time step The inducement context vector includes factors affecting the rainfall rate, surface temperature difference, wind speed and humidity environment, and generates an inducement context embedding matrix through a context embedding network; S24, inputting the spatial position embedding vector, the inducement context vector, and the three-dimensional tensor into an enhanced input layer, wherein the enhanced input layer includes a feature splicing unit and a linear projection network, and the spatial position embedding vector, the inducement context vector, and the three-dimensional tensor are fused through the enhanced input layer to form an enhanced multi-channel input sequence Z; S25. Input the enhanced multi-channel input sequence Z to the multimodal decomposition module, which includes a trend extraction submodule, a period extraction submodule and a disturbance response submodule. The trend extraction submodule uses sliding mean filtering and gated convolution to extract long-term change trends. The period extraction submodule extracts repetitive change structures based on a sparse self-attention mechanism. The disturbance response submodule uses the inducement context feature as input to construct a dynamic weight control mechanism and utilizes a trainable disturbance perception function f. θ (C t ) guides the residual extraction process, decomposing the enhanced multi-channel input sequence Z into trend components Z (tr) , periodic component Z (se) and the disturbance component Z (no) The disturbance perception function includes a two-layer fully connected network and a dynamic normalization activation module. The dynamic normalization activation module calculates the normalized adjustment value of the current disturbance weight by inputting the inducement context features, and introduces trainable scaling and offset parameters to generate a disturbance adjustment coefficient to dynamically control the influence of the disturbance component on the prediction result. S26. Constructing a long-term prediction network model, wherein the long-term prediction network model includes an enhanced input layer, a multimodal decomposition module, a period modeling encoder, a disturbance modeling module, a trend modeling state space module, a multi-step prediction decoder module, and an abnormal risk output module; S27, the periodic component Z (se) Input to the periodic modeling encoder, which is built based on a multi-head sparse attention mechanism to extract periodic internal features and output a periodic feature vector sequence H e ; S28, the disturbance component Z (no) Input to the disturbance modeling module, use the gated residual block to extract the non-stationary change characteristics dominated by external inducements, and output the disturbance feature vector sequence H n ; S29, the trend component Z (tr) Input to the trend modeling state space module, use one-dimensional convolution and gating structure to model its long-term trend characteristics, and output trend feature vector sequence H t ; S210, the periodic feature vector sequence H e , perturbation feature vector sequence H n , trend feature vector sequence H t The three types of feature vectors are concatenated and fused in the time dimension to generate a unified decoding input representation H d ; S211, the unified decoding input is represented as H d Input to the multi-step prediction decoder module, which uses an autoregressive structure to simultaneously predict the state change trend of multiple time steps in the future and generate a prediction sequence S212, input the unified decoded input representation to the abnormal risk output module, the abnormal risk output module includes an abnormal probability classification output head set in parallel with the trend prediction decoder, constructs an abnormality detection substructure, uses the abnormal probability classification output head based on the unified decoded input representation H d Output the abnormal probability vector corresponding to the future prediction sequence Form a predicted abnormal probability sequence.

5. The intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes according to claim 1 is characterized in that: The S3 specifically includes: S31. Initialize the vulture search algorithm population and jointly construct the structural parameters and training parameters of the long-term prediction network model into an optimization vector: θ=[l,w,h a ,d e ,η,b]; Among them, l represents the number of coding layers, w represents the sliding window length, h a represents the number of attention heads, d e represents the embedding dimension, η represents the learning rate, and b represents the training batch size; a set of initial parameter vectors θ are generated for each vulture individual i , construct the population set P = {θ1,θ2,…,θ N }, i∈{1,2,…,N}, where N represents the population size; S32. Based on the long-term prediction network model, the state prediction results of each monitoring point in multiple time steps in the future and the corresponding abnormal probability output are used to perform risk clustering on the monitoring points, and multiple regional risk-related subpopulations are constructed. Each subpopulation independently maintains the parameter individuals within the region and runs the optimization process in parallel. S33. Establish a co-evolutionary memory vector for each vulture individual Record several rounds of historical parameter update trajectories and generate a collaborative preference vector M based on a sliding window approach (t) , guiding the evolutionary direction of individuals in the current generation; S34, extract the prediction residual sequence from the output of the long-term prediction network model, the prediction residual sequence refers to the time-step difference sequence between the future multi-time-step prediction value of each monitoring point output by the long-term prediction network model and the corresponding actual observation value, and calculate the residual information entropy H t and the first-order derivative ΔH t , construct the entropy-driven stage gating function G(H t ,ΔH t ), output the weight ratio of the three behavior stages [λ1,λ2,λ3], which correspond to the three behavior stages of circling search, tentative descent and dive attack respectively; S35. Execute the vulture individual parameter update operation based on the behavior weight and the collaborative preference vector: Among them, λ s ∈{λ1,λ2,λ3},λ s represents the control weight corresponding to the current behavior stage, β1 and β2 are the collaborative direction guidance coefficient and the global optimal aggregation coefficient respectively, represents the current global optimal parameter combination, is the current parameter vector combination of the i-th vulture individual in the t+1th iteration, is the current parameter vector combination of the i-th vulture individual in the t-th iteration; S36, mark the monitoring points in the predicted abnormal probability sequence that are greater than the set threshold δ as high-risk points, count the distribution density of high-risk points in the spatial area and construct the risk perception state vector S t ; S37, build a reinforcement learning agent module to perceive the risk state vector S t As the state input, the individual fitness improvement of the vulture is used as the reward function, and the output behavior recommendation action a t , as a dynamic regulation signal of the weight ratio of the behavior stage; S38. Construct a fitness function f(θ) to measure the prediction accuracy, risk response consistency, and structural complexity of each parameter combination: in, is the predicted result, Y is the true value, is the subset of monitoring points with abnormal probability higher than the threshold, ‖θ‖2 is the L2 norm of the parameter vector, dim(θ) represents the number of parameter dimensions, represents the mean square error between the predicted value and the actual value, The variance of the subset of monitoring points whose anomaly probability is higher than the threshold; S39, every set round T s Perform an individual migration operation between subpopulations. Based on the risk level and fitness score of each area, individuals with the best fitness are migrated from high-risk areas to low-risk areas, optimizing the population search diversity and global convergence ability. S310, when the maximum number of iterations is reached or the fitness change for multiple rounds is lower than the set convergence threshold, the parameter combination θ with the minimum fitness function value is output g , as the final optimized parameter configuration of the long-term prediction network model.

6. The intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes according to claim 1 is characterized in that: The S4 specifically includes: S41, obtaining structural parameters and training parameters of the optimized long-term prediction network model, constructing the final optimized long-term prediction network model as a state prediction module to receive real-time monitoring data and perform multi-step prediction tasks; S42. Use the standardized input sample set as the main input, and the sensor activity, measurement point status, and risk score of the previous cycle at the current time step as auxiliary dynamic guidance information to input the optimized long-term prediction network model input interface to construct a multi-source enhanced input structure that integrates dynamic control factors; S43. Within the state prediction module, a dual-channel state prediction mechanism based on temporal structure decoupling is executed, wherein the trend prediction channel outputs the crack development trend, and the displacement prediction channel outputs the displacement change trend. The two channels share encoder features and enhance the interrelated features through a cross-trend attention mechanism. S44, based on the abnormal risk parallel classification branch output by the optimized long-term prediction network model, combined with the trend change amplitude, sensor feature distribution and dynamic feedback information corresponding to each time step, to generate the abnormal response probability value of each monitoring point in each future time step; S45. Output the crack development trend, displacement change trend, and abnormal response probability value in a structured manner, and automatically index and organize them based on the monitoring point type and prediction dimension; S46. Construct a consistency evaluation mechanism for prediction results, conduct dynamic consistency judgment on the joint change pattern of the crack development trend sequence and the displacement change trend sequence within the prediction time window, and adjust the update strategy of the long-term prediction network model based on the consistency evaluation results.

7. The intelligent monitoring and early warning method for dangerous rockfall on high and steep slopes according to claim 1 is characterized in that: The S5 specifically includes: S51, receiving multi-step state prediction results output by the long-term prediction network model, including crack development trend, displacement change trend and abnormal response probability value; S52, classify and summarize the multi-step state prediction results according to the monitoring point number, construct a multi-dimensional prediction information set based on the monitoring point, and complete and mark incomplete data; S53. Construct a risk score input structure based on the predicted data of each monitoring point, wherein the risk score input structure includes a prediction sequence of crack values, displacement values, acceleration values, rainfall indexes, and microseismic amplitudes in each future time step; S54, constructing an early warning risk scoring function for comprehensively evaluating the potential risk level of each monitoring point, wherein the early warning risk scoring function is calculated based on the weight relationship between the crack development trend, the displacement change trend, and the abnormal response probability value and the time step characteristics; S55, inputting the crack development trend, displacement change trend and abnormal response probability value of each monitoring point into the early warning risk scoring function, and calculating the risk score value of each monitoring point in the entire prediction time window; S56. The risk score value of each monitoring point is used as the input basis for the subsequent warning level determination and information linkage release mechanism to realize the intelligent warning process of dangerous rock falls on steep slopes.

8. An intelligent monitoring and early warning system for dangerous rockfall on steep slopes, an intelligent monitoring and early warning method for dangerous rockfall on steep slopes according to any one of claims 1 to 7, characterized in that: Includes the following modules: Data acquisition and preprocessing module, used to collect real-time monitoring data for preprocessing and generate standardized input sample sets; Model construction and optimization module, which is used to build a long-term prediction network model and jointly optimize the structural parameters and training parameters using the vulture search algorithm; The state prediction module is used to apply the optimized long-term prediction network model to the standardized input sample set to perform multi-step state prediction; The anomaly recognition module is used to output the abnormal response probability value sequence of each monitoring point in the future time step based on the parallel classification branch of the optimized long-term prediction network model; Result output module, used to organize and structure the multi-step state prediction results; The risk scoring module is used to input multiple types of prediction indicators into the scoring function to generate the risk score value of each monitoring point; The early warning response module is used to divide the early warning level according to the risk score value, and release it through voice broadcast, mobile terminal and control platform when the threshold is exceeded.

Images