Machine learning-based complex geological disaster evolution trend deduction method and system

Abstract

The application relates to the technical field of machine learning, in particular to a complex geological disaster evolution trend deduction method and system based on machine learning.The method comprises the following steps: collecting geological survey data, a digital elevation model and multi-source monitoring data of a geological disaster hidden danger point, and establishing a complex geological disaster space-time correlation data set; a disaster evolution state chain representing a continuous disaster evolution process is established based on the complex geological disaster space-time correlation data set; a continuous state segment is generated based on the disaster evolution state chain, the number of monitoring periods and the state duration that each continuous state segment continuously maintains are counted, the continuous maintenance proportion corresponding to each continuous state segment is calculated, and the state stability contribution degree corresponding to each continuous state segment is calculated in combination with the state switching density.The physical continuity constraint is embedded into the machine learning framework, the Transformer model is used to accurately identify the continuous transmission process in the infiltration stagnation period, and the artificial fracture of the disaster evolution state chain is effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a method and system for predicting the evolution of complex geological disasters based on machine learning. Background Technology

[0002] The evolution of geological hazards is highly nonlinear, nonstationary, and involves multiple coupled factors. Its continuous evolution from stable development and gradual progression to accelerated destruction is difficult to accurately characterize using traditional mathematical models. Existing geological hazard monitoring and early warning systems often rely on single or a few monitoring parameters (such as rainfall, tilt angle, and acceleration) and rely on simple threshold discrimination for early warning. This fails to fully utilize the rich evolutionary information contained in multi-source heterogeneous monitoring data and effectively captures the continuous temporal transfer patterns of hazard states. Furthermore, due to spatial differences in the permeability of soil and rock media within slopes, groundwater infiltration exhibits significant transmission lags between different areas. Under continuous rainfall conditions, soil moisture content continuously increases, while slope attitude monitoring parameters such as tilt angle and acceleration fail to respond synchronously or show weakened responses within local monitoring periods. This makes existing methods prone to misjudging interruptions in the hazard evolution process, leading to artificial breaks in the constructed hazard evolution state chain and reducing the accuracy of machine learning models in learning the continuous evolution patterns of hazards. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for predicting the evolution of complex geological disasters based on machine learning.

[0004] The technical solution of this invention: a method for predicting the evolution of complex geological hazards based on machine learning, comprising: S1. Collect geological survey data, digital elevation models and multi-source monitoring data of potential geological disaster sites, and establish a complex spatiotemporal correlation dataset of geological disasters; S2. Establish a disaster evolution state chain characterizing the continuous evolution process of disasters based on a complex geological disaster spatiotemporal correlation dataset; S3. Generate continuous state segments based on the disaster evolution state chain, count the number of monitoring periods in which each continuous state segment is continuously maintained and the duration of the state, calculate the continuous maintenance ratio corresponding to each continuous state segment, and calculate the state stability contribution corresponding to each continuous state segment in combination with the state switching density. Generate a weighted disaster evolution state sequence based on the state stability contribution, and use a pre-trained Transformer model to predict the disaster evolution state corresponding to multiple future monitoring periods based on the weighted disaster evolution state sequence, and generate the future disaster evolution state chain corresponding to multiple future monitoring periods. S4. Calculate the probability of occurrence of each evolution state based on the future disaster evolution state chain, and generate the prediction results of the evolution trend of complex geological disasters.

[0005] As a further improvement to this technical solution, step S1 involves establishing a spatiotemporal correlation dataset for complex geological disasters, including the following steps: Geological survey data, digital elevation models, and multi-source monitoring data collected from monitoring equipment corresponding to various geological hazard hazard points are collected. Based on a unified spatial coordinate benchmark and time benchmark, the geological survey data, digital elevation models, and multi-source monitoring data are standardized in terms of coordinates, time, and data format to form standardized geological hazard monitoring data. The standardized geological hazard monitoring data are linked and organized according to the hazard point number, monitoring time, and data type to establish the correspondence between geological environment information, topographic information, and monitoring response information corresponding to each monitoring time, forming a complex spatiotemporal correlation dataset of geological hazards.

[0006] As a further improvement to this technical solution, in step S2, establishing a disaster evolution state chain characterizing the continuous evolution process of a disaster based on a complex geological disaster spatiotemporal correlation dataset includes the following steps: S2.1. Based on the spatiotemporal correlation dataset of complex geological disasters, extract the geological environment features, topographic features and monitoring response features corresponding to each monitoring time to form a disaster state feature set corresponding to each monitoring time; S2.2 Based on the disaster state feature set corresponding to each monitoring time, analyze the changing relationship of disaster state features between adjacent monitoring times and identify the disaster evolution state corresponding to each monitoring time; S2.3. Based on multi-source monitoring data, establish a groundwater infiltration response chain and identify infiltration response stagnation intervals. Introduce external hydraulic gradient constraints to calculate the effective transmission rate of infiltration driving force. Based on the effective transmission rate of infiltration driving force, add a continuity confidence label to the disaster evolution state within the infiltration response stagnation interval to form an enhanced disaster evolution state. S2.4. Based on the enhanced disaster evolution state and monitoring time sequence, establish the state transition relationship between adjacent disaster evolution states to form a disaster evolution state transition sequence; S2.5. Based on the disaster evolution state transition sequence, connect the disaster evolution states corresponding to each monitoring time to establish a disaster evolution state chain that characterizes the continuous evolution process of the disaster.

[0007] As a further improvement to this technical solution, in step S2.3, a groundwater infiltration response chain is established and infiltration response stagnation intervals are identified. An external hydraulic gradient constraint is introduced to calculate the effective transmission rate of the infiltration driving force. Based on the effective transmission rate of the infiltration driving force, a continuity confidence label is added to the disaster evolution state within the infiltration response stagnation interval to form an enhanced disaster evolution state. This includes the following steps: S2.31. Based on multi-source monitoring data within a continuous monitoring period, establish rainfall change sequences, soil moisture content change sequences, slope angle change sequences, and acceleration change sequences according to a unified monitoring period. Based on the temporal variation patterns corresponding to each change sequence, analyze the temporal response relationship between rainfall changes, soil moisture content changes, and slope attitude changes, identify the order of responses between rainfall, groundwater infiltration, and slope attitude response, and establish a groundwater infiltration response chain. S2.32. Based on the groundwater infiltration response chain, the infiltration-induced deformation delay time and its rate of change corresponding to adjacent monitoring cycles are statistically analyzed according to the continuous monitoring cycle, and the infiltration response stagnation interval is generated. S2.33. Based on the boundary response characteristics before and after the infiltration response stagnation interval, an external hydraulic gradient constraint is introduced. The effective transmission rate is calculated by the ratio of the infiltration front advance rate to the saturated permeability coefficient as a physical verification item. The transmission continuity is analyzed in conjunction with the boundary response characteristics to generate a groundwater infiltration transmission retention coefficient that characterizes the continuous transmission capacity of infiltration. S2.34. Based on the groundwater infiltration transmission retention coefficient, the continuity of the disaster evolution state corresponding to the infiltration response stagnation interval is determined. Based on the continuity determination result, a continuity confidence label is generated, and an enhanced disaster evolution state with a continuity confidence label is formed.

[0008] As a further improvement to this technical solution, in S2.33, based on the boundary response characteristics before and after the infiltration response stagnation interval, an external hydraulic gradient constraint is introduced. The effective transmission rate is calculated as a physical verification item by the ratio of the infiltration front advance rate to the saturated permeability coefficient. The transmission continuity is analyzed in conjunction with the boundary response characteristics to generate a groundwater infiltration transmission retention coefficient that characterizes the continuous transmission capacity of infiltration. This includes the following steps: Soil moisture content, dip angle, and acceleration response characteristics at the start and end times of the infiltration response stagnation interval are extracted as response characteristics of the front and rear boundaries of the stagnation interval. The topographic slope of the corresponding area of ​​the stagnation interval is calculated based on the digital elevation model, and external hydraulic gradient constraints are established in combination with the groundwater flow direction. The effective transmission rate of infiltration driving force is calculated based on the ratio of the groundwater infiltration front advance rate to the saturated permeability coefficient of the corresponding area, and the effective transmission rate is used as a physical verification item. Based on the response characteristics of the boundary before and after the stagnation zone, the consistency of the direction of change, the continuity of the magnitude of change, and the continuity of the rate of change of the groundwater infiltration driving force within the stagnation zone are analyzed. After normalizing the continuity analysis results with the effective transmission rate of the infiltration driving force, a groundwater infiltration transmission retention coefficient, which characterizes the continuous transmission capacity of groundwater infiltration, is generated by weighted fusion.

[0009] As a further improvement to this technical solution, in step S3, a weighted disaster evolution state sequence is generated, and a pre-trained Transformer model is used to predict the disaster evolution state corresponding to multiple future monitoring periods based on the weighted disaster evolution state sequence, thereby generating a future disaster evolution state chain corresponding to multiple future monitoring periods, including the following steps: S3.1 Divide the disaster evolution state chain into multiple continuous state segments according to the monitoring time sequence, count the number of monitoring cycles and the duration of each continuous state segment, calculate the continuous maintenance ratio of each continuous state segment, and calculate the state stability contribution of each continuous state segment in combination with the state switching density. Generate a weighted disaster evolution state sequence based on the state stability contribution. S3.2 Encode the weighted disaster evolution state sequence into the corresponding state sequence and input it into the pre-trained Transformer model. Use the multi-head self-attention mechanism to learn the long-term state transition law between disaster evolution states and establish the disaster evolution state transition mapping relationship. S3.3 Based on the disaster evolution state transition mapping relationship, perform sequence deduction on the weighted disaster evolution state sequence, predict the disaster evolution state corresponding to multiple future monitoring periods, and connect each predicted disaster evolution state according to the monitoring time order to generate the future disaster evolution state chain corresponding to multiple future monitoring periods.

[0010] As a further improvement to this technical solution, in step S3.1, the number of monitoring cycles and the duration of each continuous state segment are statistically analyzed, the continuous maintenance ratio corresponding to each continuous state segment is calculated, and the state stability contribution corresponding to each continuous state segment is calculated in combination with the state switching density. Based on the state stability contribution, a weighted disaster evolution state sequence is generated, including the following steps: S3.11. Divide the disaster evolution state chain into multiple continuous state segments according to the monitoring time sequence. Using the current disaster evolution state as the reference state, backtrack a preset number of continuous disaster evolution states in the historical direction, and count the number of monitoring cycles that each continuous state segment can maintain continuously and the corresponding state duration. S3.12. Calculate the continuous maintenance ratio based on the number of monitoring cycles and the duration of each continuous state segment, and calculate the state stability contribution of each continuous state segment in combination with the state switching density. S3.13. Based on the state stability contribution of each continuous state segment, the weights of each continuous state segment in the disaster evolution state chain are adjusted to generate a weighted disaster evolution state sequence.

[0011] As a further improvement to this technical solution, in S3.12, the continuous maintenance ratio is calculated based on the number of monitoring cycles and the duration of each continuous state segment, and the state stability contribution corresponding to each continuous state segment is calculated in combination with the state switching density, including the following steps: The number of monitoring cycles and the duration of each continuous state segment are normalized respectively. Based on the normalized number of monitoring cycles and the normalized duration of the state, the continuous maintenance ratio corresponding to each continuous state segment is generated by linear weighted fusion. Identify the state transition locations between adjacent continuous state segments, count the number of times the disaster evolution state changes within a unit monitoring period, and calculate the state transition density corresponding to each continuous state segment by combining the monitoring time interval between each state transition location. The system is normalized based on the continuous maintenance ratio and state switching density, and then weighted and fused according to preset weights to generate the state stability contribution of each continuous state segment.

[0012] As a further improvement to this technical solution, in step S4, the probability of occurrence corresponding to each evolution state is calculated based on the future disaster evolution state chain to generate a prediction result of the evolution trend of complex geological disasters, including the following steps: S4.1. Based on the future disaster evolution state chain, statistically analyze the disaster evolution state types corresponding to each future monitoring period and the number of monitoring periods in which each disaster evolution state is continuously maintained, and establish a time series distribution sequence of future disaster evolution states. S4.2 Based on the time-series distribution sequence of future disaster evolution states, the cumulative number of occurrences, the number of consecutive monitoring periods, and the number of state transitions of each disaster evolution state in multiple future monitoring periods are statistically analyzed. The occurrence frequency, state duration ratio, and state transition activity corresponding to each disaster evolution state are calculated, and the occurrence probability corresponding to each disaster evolution state is calculated based on the occurrence frequency, state duration ratio, and state transition activity. S4.3 Sort the disaster evolution states in descending order according to their occurrence probabilities, take the disaster evolution state with the highest occurrence probability as the dominant evolution state for the corresponding future monitoring period, and combine the dominant evolution states for each future monitoring period to form a disaster evolution probability sequence. S4.4. Based on the analysis of the probability sequence of disaster evolution, the changing trend of disaster evolution status in multiple future monitoring periods is analyzed to generate the prediction results of the evolution of complex geological disasters.

[0013] On the other hand, the present invention provides a machine learning-based system for predicting the evolution of complex geological hazards, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the above-mentioned machine learning-based method for predicting the evolution of complex geological hazards.

[0014] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: by embedding physical continuity constraints into the machine learning framework, the Transformer model is used to accurately identify the continuous transmission process during the infiltration stagnation period, effectively avoiding the artificial breakage of the disaster evolution state chain, significantly improving the physical consistency and prediction reliability of the inference results, and providing more effective decision support for early warning and situation assessment of geological disasters. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0016] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please refer to Figure 1 As shown in the figure, this embodiment provides a method for predicting the evolution of complex geological hazards based on machine learning, including the following steps: S1. Collect geological survey data, digital elevation models and multi-source monitoring data of potential geological disaster sites, and establish a complex spatiotemporal correlation dataset of geological disasters; In this embodiment, the establishment of a spatiotemporal correlation dataset for complex geological disasters includes the following steps: Geological survey data (including at least information on soil type, soil layer thickness, and particle composition), digital elevation models (i.e., three-dimensional terrain models that digitally represent the surface undulations using elevation data of regular grid points), and multi-source monitoring data collected by monitoring equipment are collected from various geological hazard hazard points. The multi-source monitoring data includes rainfall, soil moisture content, dip angle, and acceleration monitoring data. Based on a unified spatial coordinate benchmark and time benchmark, the geological survey data, digital elevation models, and multi-source monitoring data are standardized in terms of coordinates, time, and data format to form standardized geological hazard monitoring data. The standardized geological hazard monitoring data are then linked and organized according to hazard point number, monitoring time, and data type to establish the correspondence between geological environment information, terrain information, and monitoring response information corresponding to each monitoring time, forming a complex spatiotemporal correlation dataset of geological hazards.

[0018] S2. Establish a disaster evolution state chain characterizing the continuous evolution process of disasters based on a complex geological disaster spatiotemporal correlation dataset; In this embodiment, establishing a disaster evolution state chain characterizing the continuous evolution process of a disaster includes the following steps: S2.1. Based on the complex geological disaster spatiotemporal correlation dataset, extract the geological environment features (such as lithology, stratigraphic structure, degree of joint and fissure development, landslide boundary and sliding surface burial depth) and topographic features (such as slope, aspect, topographic relief, curvature and elevation) and monitoring response features (such as rainfall response features, groundwater infiltration response features and slope deformation response features) corresponding to each monitoring time. This forms a disaster state feature set corresponding to each monitoring time (normalize various features according to a unified feature encoding method and fuse them according to the monitoring time to form a disaster state feature set corresponding to each monitoring time, i.e., feature vector set). S2.2 Based on the disaster state feature set corresponding to each monitoring time, analyze the changing relationship of disaster state features between adjacent monitoring times, and identify the disaster evolution state corresponding to each monitoring time (calculate the change amplitude and direction of geological environment features, topographic features and monitoring response features between adjacent monitoring times to generate disaster state change features; comprehensively analyze the synergistic change relationship between rainfall infiltration, groundwater response and slope deformation based on the disaster state change features, and identify the corresponding evolution state of the disaster state at each monitoring time, which is in stable development, gradual evolution, accelerated evolution or violent evolution; at the same time, establish a one-to-one correspondence between the identified disaster evolution state and the corresponding monitoring time to generate the disaster evolution state corresponding to each monitoring time). S2.3. Based on multi-source monitoring data, establish a groundwater infiltration response chain and identify infiltration response stagnation intervals. Introduce external hydraulic gradient constraints to calculate the effective transmission rate of infiltration driving force. Based on the effective transmission rate of infiltration driving force, add a continuity confidence label to the disaster evolution state within the infiltration response stagnation interval to form an enhanced disaster evolution state. (During the evolution of complex geological disasters, due to the spatial differences in the permeability of soil and rock media within the slope, groundwater infiltration has a significant transmission time lag between different areas. This leads to a continuous increase in soil moisture content under continuous rainfall conditions, while slope attitude monitoring parameters such as tilt angle and acceleration do not respond synchronously or the response weakens within local monitoring periods. This can easily cause existing methods to misjudge the interruption of the disaster evolution process, thus leading to...) The artificially constructed disaster evolution state chain is disrupted, reducing the accuracy of machine learning models in learning the continuous evolution of disasters. This method establishes a groundwater infiltration response chain and quantitatively identifies stagnant infiltration response intervals. It introduces external hydraulic gradient constraints and the effective transmission rate of infiltration driving force as physical verification terms. A "physical continuity compensation" confidence label is added to the disaster evolution state within the stagnant interval, and an extended feature vector containing the dimension of continuous driving intensity enhancement is generated. This enables the machine learning model to accurately identify physically continuous but lagging evolution stages during training and deduction, thereby effectively preventing state chain breaks and significantly improving the accuracy of the Transformer model in learning the continuous evolution of geological disasters and the physical consistency of the deduction results. The formation of an enhanced disaster evolution state includes the following steps: S2.31. Based on multi-source monitoring data within a continuous monitoring period, establish rainfall variation sequences, soil moisture content variation sequences, slope angle variation sequences, and acceleration variation sequences according to a unified monitoring period. Based on the temporal variation patterns corresponding to each variation sequence, analyze the temporal response relationships between rainfall changes, soil moisture content changes, and slope attitude changes. Identify the sequence of responses among rainfall, groundwater infiltration, and slope attitude responses, and establish a groundwater infiltration response chain (based on the identified sequence of responses among rainfall response, groundwater infiltration response, and slope attitude response, with rainfall response as the starting node of the response chain, soil moisture content response as the groundwater infiltration transmission node, and slope angle and acceleration responses as the slope attitude response nodes). The node to be terminated should be established, and a directed connection relationship between the nodes should be established according to the monitoring time sequence corresponding to each response node. Among them, when the time interval between two adjacent response nodes does not exceed the preset response time window (6-24h) and the change direction of the two is consistent, the response transmission relationship between the corresponding nodes is established; otherwise, the response transmission is judged as interrupted and no node connection is established. Subsequently, all response nodes that meet the response transmission conditions within the continuous monitoring period and their corresponding response transmission relationships are connected in sequence to form a groundwater infiltration response chain that is transmitted step by step in the order of rainfall response, groundwater infiltration response, and slope attitude response. The monitoring time, response type, response duration, and response delay time between adjacent nodes are recorded simultaneously for each response node. S2.32. Based on the groundwater infiltration response chain, the infiltration-induced deformation delay time corresponding to adjacent monitoring periods is statistically analyzed according to the continuous monitoring period. ( For the first The start time of groundwater infiltration response in each monitoring cycle The slope attitude response time and its rate of change are used to generate the infiltration response stagnation interval (specifically, the infiltration-induced deformation delay time is analyzed according to the continuous monitoring cycle). The evolution of its rate of change; then, three judgment conditions are set: one: continuity ( Within each monitoring cycle, Continue to increase or remain above the historical average The level is one time the standard deviation; 2: the corresponding rate of change over the delay time remains continuously positive, or although it is negative, its absolute value is lower than the preset change threshold. Hour / cycle; Three: Soil moisture content continues to rise or remains high (above historical average) during the same period. (multiple standard deviations), while the rate of change of tilt angle and the rate of change of acceleration are both lower than their respective preset response thresholds (the threshold for the rate of change of tilt angle is set to 10 ... / period, acceleration rate of change threshold set to / cycle); when the above three conditions are met simultaneously, it is determined that the driving force of groundwater infiltration is still in effect but has not yet been effectively transmitted to the macroscopic attitude change process of the slope, and the continuous monitoring interval that meets the above conditions is determined as the infiltration response stagnation interval, and the start time, end time, duration and corresponding delay time change characteristics of the stagnation interval are recorded). S2.33. Based on the response characteristics of the infiltration response stagnation zone at the boundaries before and after the stagnation zone, an external hydraulic gradient constraint is introduced. The effective transfer rate is calculated using the ratio of the infiltration front advance rate to the saturated permeability coefficient as a physical verification item. The continuity of transfer is analyzed in conjunction with the boundary response characteristics, generating a groundwater infiltration transfer retention coefficient that characterizes the continuous transfer capacity of infiltration. (This is achieved by extracting the soil moisture content, dip angle, and acceleration response characteristics at the boundaries before and after the stagnation zone, establishing an external hydraulic gradient constraint in conjunction with the terrain slope, and calculating the effective transfer rate using the ratio of the infiltration front advance rate to the saturated permeability coefficient, thereby quantitatively linking the hydraulic driving conditions with the response transfer efficiency; simultaneously combining...) The continuity analysis results of the boundary response in terms of direction, amplitude, and rate were weighted and fused to generate a groundwater infiltration transmission retention coefficient. This coefficient, as a quantitative indicator, objectively reflects whether the infiltration driving force within the stagnant zone remains effectively transmitted. It provides a scientific basis for subsequently adding "physical continuity compensation" confidence markers and expanding feature vectors to the corresponding disaster evolution state, thereby ensuring that the machine learning model can accurately identify physically continuous but lagging monitoring response evolution stages during training and extrapolation. This effectively prevents artificial breaks in the disaster evolution state chain and improves the accuracy of the Transformer model in learning the continuous evolution laws of complex geological disasters. Furthermore, generating the groundwater infiltration transport retention coefficient, which characterizes the sustained transport capacity of infiltration, includes the following steps: Soil moisture content, dip angle, and acceleration response characteristics at the start and end times of the infiltration response stagnation interval are extracted as the response characteristics of the boundaries before and after the stagnation interval. The topographic slope of the area corresponding to the stagnation interval is calculated based on the digital elevation model, and an external hydraulic gradient constraint is established in combination with the groundwater flow direction. ( This represents the elevation difference (m) between the start and end points of the area corresponding to the stationary interval. This represents the path length (m) of groundwater along the main flow direction. (The angle between the mainstream direction of groundwater and the direction of maximum slope gradient); based on the advance rate of the groundwater infiltration front. ( The distance (m) between adjacent soil moisture content monitoring points along the main infiltration direction of groundwater is defined as follows: (First, based on the digital elevation model and groundwater level monitoring data, the main flow direction of groundwater in the corresponding area of ​​the stagnation zone is determined using the steepest slope direction method; then, the spatial line connecting two adjacent soil moisture content monitoring points is orthogonally projected onto the main flow direction of groundwater, and the length of the projected line segment is taken as the distance between adjacent soil moisture content monitoring points along the main infiltration direction of groundwater.) When the angle between the line connecting adjacent monitoring points and the main flow direction of groundwater is greater than 10°C, the groundwater flow direction is more than 10°C. When this happens, the monitoring point pair is discarded, and another monitoring point pair that is closest to the groundwater flow upstream is used for calculation to ensure... It can accurately reflect the propagation distance of the infiltration front along the hydraulic gradient direction. The time difference (s) between the groundwater infiltration front and the next monitoring point, and the corresponding saturated permeability coefficient of the area. (Based on geological survey data, information on soil and rock types, soil layer thickness, and particle composition of the corresponding stagnant area is obtained, and the saturation state of the corresponding area is determined by combining the changes in soil moisture content during the continuous monitoring period; subsequently, a pre-established soil and rock permeability parameter library is matched according to the corresponding soil and rock types (the soil and rock permeability parameter library is established based on the results of variable head permeability tests or constant head permeability tests, covering at least nine common soil and rock types: clay, silty clay, silt, silty sand, fine sand, medium sand, coarse sand, gravelly sand, and gravelly soil; the reference range of saturated permeability coefficients for each type is as follows: clay) m / s, silty clay m / s, silty soil m / s, silt m / s, fine sand m / s, medium sand m / s, coarse sand m / s, gravel m / s, gravelly soil m / s; For special soil and rock types or mixed strata not included in the parameter database, the saturated permeability coefficient of the corresponding area is determined by field double-ring infiltration test or borehole water injection test, and the measured value is used as... Substitute the values ​​into the calculation. When multiple strata exist in the same area, the basic saturated permeability coefficients are weighted and averaged according to the thickness proportion of each stratum to obtain the saturated permeability coefficient of the corresponding area in the stagnation zone. The effective transmission rate of the infiltration driving force is calculated based on the ratio of the basic saturated permeability coefficients of the corresponding area in the stagnation zone to the weighted average of the basic saturated permeability coefficients of each stratum thickness proportion. ( The external hydraulic gradient correction coefficient is used to characterize the enhancing effect of the hydraulic gradient on infiltration transfer efficiency. The specific value is determined by the fitting relationship between the soil moisture content response lag time and rainfall intensity in the historical monitoring period of the region. The effective transfer rate is used as a physical verification item to quantitatively assess whether the infiltration driving force is continuously and effectively transferred in the stagnation interval, and to provide a physical constraint basis for the continuity confidence label. Based on the boundary response characteristics before and after the stagnation interval, and using the changes in soil moisture content, dip angle, and acceleration at the start and end of the stagnation interval as benchmarks, this study analyzes the consistency of the direction of change, the continuity of the amplitude, and the continuity of the rate of change of groundwater infiltration driving force within the stagnation interval. Specifically, the changes in soil moisture content, dip angle, and acceleration at the start and end of the stagnation interval are used as the continuity analysis benchmark; the consistency between the direction of change of the monitoring parameters corresponding to each monitoring cycle within the stagnation interval and the overall direction of change of the boundary response is statistically analyzed to generate the consistency of the direction of change; further analysis is conducted on the fluctuation and dispersion of the response amplitude relative to the boundary response amplitude within each monitoring cycle within the stagnation interval to generate the continuity of the amplitude; and the fluctuation and continuous change characteristics of the rate of change of the response relative to the boundary response rate within each monitoring cycle within the stagnation interval are analyzed to generate the continuity of the rate of change. The continuity analysis results are then normalized with the effective transmission rate of the infiltration driving force (using the Min-Max normalization method), and a weighted fusion method is used to generate the groundwater infiltration transmission retention coefficient, which characterizes the continuous transmission capacity of groundwater infiltration. ( To ensure consistency in the direction of change after normalization, To ensure the continuity of the variation range after normalization, To ensure the continuity of the rate of change after normalization, The normalized infiltration driving force effective transfer rate, Weights for consistency in the direction of change. As a weight for the continuity of the change range, As a weight for the continuity of the rate of change, As the weight of the effective transmission rate of the infiltration driving force, The weights are determined by numerical traversal based on historical monitoring experimental data, with the goal of maximizing the correlation between the retention coefficient and subsequent state transitions. S2.34. Based on the groundwater infiltration transmission retention coefficient, the continuity of the disaster evolution state corresponding to the stagnant infiltration response interval is determined. Based on the continuity determination result, a continuity confidence label is generated. When the groundwater infiltration transmission retention coefficient reaches the preset continuous transmission threshold (0.65~0.85, determined by expert experience), it is determined that the groundwater infiltration in the stagnant interval is still continuously transmitted. A "physical continuity compensation" confidence label is added to the disaster evolution state corresponding to the stagnant interval. The effective transmission rate of the infiltration driving force is used to physically enhance the feature vector of the disaster evolution state corresponding to the interval. Specifically, the original monitoring data values ​​such as rainfall, tilt angle, and acceleration are kept completely unchanged. Only the continuous driving intensity dimension generated by the effective transmission rate is added to the end of the feature vector to form an extended feature vector containing the original physical monitoring features and the continuous driving intensity enhancement dimension. An enhanced disaster evolution state with a continuity confidence label is formed to enable the machine learning model to accurately identify the evolution stage that is physically continuous but has a lagging monitoring response. S2.4. Based on the enhanced disaster evolution state and monitoring time sequence, establish the state transition relationship between adjacent disaster evolution states to form a disaster evolution state transition sequence. Specifically, sort the enhanced disaster evolution states corresponding to each monitoring time according to the monitoring time sequence, take the enhanced disaster evolution state corresponding to the current monitoring time as the preceding state and the enhanced disaster evolution state corresponding to the next monitoring time as the succeeding state, and establish a one-to-one correspondence state transition relationship between adjacent disaster evolution states. For enhanced disaster evolution states with added continuity confidence labels, while keeping the original disaster evolution state type unchanged, the continuous driving intensity enhancement dimension and continuity confidence label are recorded synchronously as accompanying attributes of the corresponding state in the state transition relationship. Then, connect the state transition relationships corresponding to each adjacent monitoring time in sequence to form a disaster evolution state transition sequence arranged continuously according to the monitoring time. S2.5. Based on the disaster evolution state transition sequence, connect the disaster evolution states corresponding to each monitoring time to establish a disaster evolution state chain characterizing the continuous evolution process of the disaster. Specifically, based on the disaster evolution state transition sequence, take the enhanced disaster evolution state corresponding to the first monitoring time as the starting node of the state chain. Read each state transition relationship sequentially according to the monitoring time order, and connect the preceding disaster evolution state and the subsequent disaster evolution state in each state transition relationship in sequence. Simultaneously retain the monitoring time, disaster evolution state type, extended feature vector, and continuity confidence mark corresponding to each monitoring time. When there is a continuity confidence mark between the disaster evolution states corresponding to adjacent monitoring times, mark the corresponding state transition relationship as a physical continuous transition; otherwise, mark it as a regular state transition. Subsequently, connect the disaster evolution states corresponding to each monitoring time continuously according to the time sequence of all state transition relationships to form a complete disaster evolution state chain that records the disaster evolution state, state transition relationship, and continuity information, so as to characterize the continuous evolution process of complex geological disasters from initial evolution to subsequent development.

[0019] S3. Generate continuous state segments based on the disaster evolution state chain, count the number of monitoring periods in which each continuous state segment is continuously maintained and the duration of the state, calculate the continuous maintenance ratio corresponding to each continuous state segment, and calculate the state stability contribution corresponding to each continuous state segment in combination with the state switching density. Generate a weighted disaster evolution state sequence based on the state stability contribution, and use a pre-trained Transformer model to predict the disaster evolution state corresponding to multiple future monitoring periods based on the weighted disaster evolution state sequence, and generate the future disaster evolution state chain corresponding to multiple future monitoring periods. In this embodiment, generating a future disaster evolution state chain corresponding to multiple future monitoring periods includes the following steps: S3.1. Divide the disaster evolution state chain into multiple continuous state segments according to the monitoring time sequence. (For each continuous state segment, firstly, based on the original monitoring data, determine the number of original monitoring cycles and the duration of each segment's continuous state. This calculation strictly uses the original monitoring cycle as the benchmark and does not convert the continuous driving intensity dimension added in S2.34 into an additional number of cycles. At the same time, use the extended feature vector generated in S2.34 (including the original features and the continuous driving intensity dimension) as the accompanying physical features of each state on the time axis into the Transformer model, so that the model can autonomously identify the physical continuity of the compensation interval during the learning process.) Calculate the number of monitoring cycles and the duration of each continuous state segment's continuous state, calculate the continuous maintenance ratio corresponding to each continuous state segment, and calculate the state stability contribution corresponding to each continuous state segment in combination with the state switching density. Generate a weighted disaster evolution state sequence based on the state stability contribution. (In the continuous evolution of complex geological disasters, due to the stage transition zone between different disaster evolution states, adjacent monitoring...) The disaster evolution state corresponding to the measurement cycle simultaneously possesses some evolutionary characteristics of the current state and the next state, resulting in multiple paths for local state transitions in the disaster evolution state chain. Traditional machine learning models treat multiple competing state transitions as effective learning objects simultaneously, making it difficult to distinguish between dominant evolutionary paths and transitional perturbations. This reduces the accuracy of learning the disaster evolution state transition patterns and affects the reliability of predicting future disaster evolution state chains. By statistically analyzing the number of consecutive periods and duration of each continuous state segment to calculate the continuous holding ratio, and combining it with the state switching density to comprehensively evaluate the dominance and persistence of each segment in the overall evolution sequence, a state stability contribution is generated as a sequence weight. This allows high-contribution dominant evolutionary state segments to obtain larger gradient update amplitudes and feature retention weights during Transformer model training, while effectively suppressing the impact of low-contribution transitional perturbation segments on model parameter updates. This guides the model to focus on the dominant evolutionary path of disaster evolution, significantly improving the targetability of state transition pattern learning and the accuracy of prediction results. The generation of a weighted disaster evolution state sequence includes the following steps: S3.11. Divide the disaster evolution state chain into multiple continuous state segments according to the monitoring time sequence. Using the current disaster evolution state as the reference state, backtrack to the historical direction for a preset number (e.g., 3 to 10 monitoring cycles) of continuous disaster evolution states, and count the number of monitoring cycles that each continuous state segment has been maintained continuously and the corresponding state duration (accumulated based on the timestamp of the monitoring cycle). S3.12. Calculate the continuous maintenance ratio based on the number of monitoring cycles and the duration of each continuous state segment, and calculate the state stability contribution of each continuous state segment in combination with the state switching density, which is used to characterize the dominance and persistence of the segment in the overall evolution sequence. Furthermore, the continuous maintenance ratio is calculated based on the number of monitoring cycles and the duration of each continuous state segment, and the state stability contribution of each continuous state segment is calculated in conjunction with the state switching density, including the following steps: The number of monitoring cycles and the duration of each continuous state segment are normalized (using the Min-Max normalization method). Based on the normalized number of monitoring cycles and the normalized duration of the state, a continuous maintenance ratio corresponding to each continuous state segment is generated through a linear weighted fusion method (where the mean of the continuous driving intensity dimension generated by S2.33 within the corresponding time period of the continuous state segment is extracted as the physical continuity modulation coefficient; the continuous maintenance ratio and the physical continuity modulation coefficient are nonlinearly mapped and fused (using the sigmoid function) to obtain the final continuous maintenance ratio), which is used to characterize the degree of continuity of the corresponding disaster evolution state. Identify the state transition locations between adjacent consecutive state segments, count the number of disaster evolution state transitions within a unit monitoring period, and calculate the state transition density corresponding to each consecutive state segment based on the monitoring time interval between each state transition location. (Following the monitoring time sequence corresponding to the disaster evolution state chain, compare the disaster evolution states corresponding to adjacent monitoring times sequentially. When the disaster evolution state type corresponding to two adjacent monitoring times changes, the corresponding monitoring time is determined as the state transition location, and the corresponding monitoring time is recorded. Subsequently, count the number of state transitions within the monitoring time range corresponding to each consecutive state segment, and calculate the number of state transitions within a unit monitoring period based on the number of monitoring periods corresponding to the consecutive state segment. Further extract the state transition density corresponding to adjacent state transition locations.) The monitoring time is calculated by taking into account the monitoring time interval between adjacent state switching positions and calculating the average of all monitoring time intervals within the corresponding continuous state segment. A comprehensive analysis is performed based on the number of state switching times and the average monitoring time interval within a unit monitoring period (i.e., linear weighted fusion after Min-Max normalization) to generate the state switching density corresponding to each continuous state segment. The higher the number of state switching times and the shorter the average monitoring time interval between adjacent state switching positions within a unit monitoring period, the greater the state switching density of the corresponding continuous state segment; conversely, the lower the number of state switching times and the longer the average monitoring time interval within a unit monitoring period, the lower the state switching density of the corresponding continuous state segment. This is used to characterize the frequency of state switching in the corresponding disaster evolution. Normalization is performed based on the continuous maintenance ratio and state switching density (using the Min-Max normalization method), and according to preset weights ( and The weighted fusion is performed to generate the state stability contribution corresponding to each continuous state segment. ( To maintain the proportion after normalization, This represents the normalized state transition density. To maintain the proportion of weights continuously, The weights for state transition density, , The index is a continuous state segment. The weights are determined by numerical traversal based on experimental data of historical disaster evolution state chains, with the goal of maximizing the goodness of fit between the weighted disaster evolution state sequence and the actual disaster evolution state within the historical monitoring period. Among them, the higher the proportion of continuous maintenance and the lower the state switching density, the higher the contribution to state stability; the lower the proportion of continuous maintenance and the higher the state switching density, the lower the contribution to state stability. S3.13. Based on the state stability contribution corresponding to each continuous state segment, the weights of each continuous state segment in the disaster evolution state chain are adjusted to generate a weighted disaster evolution state sequence. Specifically, according to the state stability contribution corresponding to each continuous state segment, all continuous state segments are sorted according to the magnitude of their state stability contribution, and the state stability contribution is used as the sequence weight of the corresponding continuous state segment. Then, based on the state stability contribution corresponding to each continuous state segment, consistent weights are assigned to the disaster evolution state and extended feature vector at each monitoring time within the corresponding continuous state segment in the disaster evolution state chain, so that the same continuous state... All disaster evolution states within a segment have the same segment weight. For continuous state segments that contribute significantly to state stability, their feature retention weight in the weighted disaster evolution state sequence is increased, thereby increasing the contribution of the corresponding state segments to feature learning and state transition learning during model training. For continuous state segments that contribute less to state stability, their corresponding sequence weight is reduced, thereby weakening their impact on model parameter updates. After adjusting the weights of all continuous state segments, the continuous state segments are reconnected according to the original monitoring time order to generate a weighted disaster evolution state sequence that maintains the time order but with different learning weights for each state segment. S3.2. Encode the weighted disaster evolution state sequence into a corresponding state sequence and input it into a pre-trained Transformer model (where the extended feature vectors corresponding to each time step are fused with the state encoding, and the position encoding and state sequence are input into the Transformer model together). Utilize a multi-head self-attention mechanism to learn the long-term state transition patterns between disaster evolution states and establish a disaster evolution state transition mapping relationship. The pre-trained Transformer model includes an input layer, a feature encoding layer, a position encoding layer, a multi-layer Transformer encoder, a multi-layer Transformer decoder, and an output layer. The input layer is used to input the weighted disaster evolution state sequence corresponding to the historical disaster evolution state chain and each monitoring... The extended feature vector corresponding to each time point includes geological environment features, topographic features, monitoring response features, and continuous driving intensity enhancement dimensions. The feature encoding layer maps the disaster evolution state and the extended feature vector to a unified feature space. The location encoding layer introduces monitoring time sequence information and adopts a sine and cosine fixed location encoding method, with the encoding dimension consistent with the output dimension of the feature encoding layer. The multi-layer Transformer encoder uses a multi-head self-attention mechanism to learn the long-term temporal correlation between disaster evolution states. The multi-layer Transformer decoder infers the subsequent disaster evolution state for each monitoring cycle based on the encoding results. The output layer uses a fully connected layer combined with a Softmax function to output the probability distribution of the disaster evolution state and the corresponding disaster evolution state for each future monitoring cycle. The Transformer model uses a historical dataset of complex geological disasters with spatiotemporal correlation as training data. It constructs a corresponding disaster evolution state chain based on continuous historical monitoring processes. The model input consists of a weighted sequence of disaster evolution states and extended feature vectors corresponding to multiple consecutive historical monitoring periods, while the actual disaster evolution states corresponding to subsequent monitoring periods serve as supervision labels. The model parameters are iteratively trained by minimizing the weighted cross-entropy loss function. Specifically, the state stability contribution of each continuous state segment calculated in S3.13 is used as the loss weight coefficient for all monitoring time samples within that segment. The weighted cross-entropy loss function is used to ensure that the prediction error corresponding to high-contribution state segments receives a larger gradient update magnitude during backpropagation. This process continues until the validation set loss converges, resulting in a pre-trained Transformer model. S3.3. Based on the disaster evolution state transition mapping relationship, perform sequence extrapolation on the weighted disaster evolution state sequence to predict the disaster evolution state corresponding to multiple future monitoring periods. Connect each predicted disaster evolution state according to the monitoring time order to generate a future disaster evolution state chain corresponding to multiple future monitoring periods. (During the extrapolation process, an autoregressive iterative method is used to generate subsequent states period by period. After generating the state of each future period, the transition confidence between the predicted state and the state of the previous period is calculated simultaneously. The maximum prediction probability given by the output layer of the Transformer model is extracted and used as the basic measure of the statistical reliability of the prediction results. Secondly, the transition confidence between the previous state and the state in the historical disaster evolution data is statistically analyzed.) The frequency of the shift to the current predicted state is analyzed, and the consistency between the monitoring cycle interval corresponding to the current predicted shift and the average interval of this type of historical shift is compared, serving as measures of historical shift reliability and temporal consistency, respectively. Next, the extended feature vectors from the previous monitoring cycle and the predicted values ​​of the extended feature vectors corresponding to the current predicted state are extracted, with a focus on the physical enhancement dimension reflecting the continuous transmission capability of infiltration. The consistency of the two sets of extended feature vectors in terms of change direction and magnitude is compared to obtain a measure of physical continuity. Finally, the four measures—statistical reliability, historical shift reliability, temporal consistency, and physical continuity—are weighted and fused according to their respective preset weights (each weight coefficient satisfies...). For example, the default values ​​are statistical reliability weights. Historical transfer credibility weight Time consistency weight Physical continuity weight The weight values ​​are optimized and determined using a grid search method based on historical monitoring data to generate the final transition confidence score. When the transition confidence score is lower than the preset confidence threshold (0.55 to 0.70, determined by expert experience), a dynamic truncation strategy is triggered to terminate over-extrapolation and output only the prediction results within the confidence score range. After the dynamic truncation strategy is triggered, all currently generated prediction states with qualified confidence scores (including all periods before the current truncation period) will be retained as components of the final future disaster evolution state chain. At the same time, each iteration feeds back the disaster evolution state predicted in the previous step and its corresponding internal model state representation (including the predicted value of the continuous driving intensity enhancement dimension output by the Transformer decoder) to the model input. In the future-oriented extrapolation process, since it is impossible to obtain measured physical quantities such as the infiltration front advance rate and the actual change in soil moisture content in future monitoring periods in advance, the effective transmission rate of infiltration driving force based on measured data defined in S2.33 is no longer calculated in real time. Instead, the Transformer model has learned the evolution law of the continuous driving intensity enhancement dimension by learning the mapping relationship between historical extended feature vectors and state transitions during the training process. In the extrapolation stage, the model decoder autoregressively predicts and outputs the value of this enhancement dimension as a proxy representation of physical continuity, and feeds it back to subsequent iterations together with the predicted state, thereby avoiding the interruption of the extrapolation chain due to the lack of future measured data.

[0020] S4. Calculate the probability of occurrence of each evolution state based on the future disaster evolution state chain, and generate the prediction results of the evolution trend of complex geological disasters; In this embodiment, generating the evolution trend prediction results of complex geological disasters includes the following steps: S4.1. Based on the future disaster evolution state chain, statistically analyze the disaster evolution state types corresponding to each future monitoring period and the number of monitoring periods in which each disaster evolution state is continuously maintained, and establish a time series distribution sequence of future disaster evolution states. S4.2 Based on the time-series distribution sequence of future disaster evolution states, calculate the cumulative occurrence of each disaster evolution state over multiple monitoring periods (for each disaster evolution state, traverse the future disaster evolution state chain, count the total number of occurrences of the corresponding disaster evolution state, and obtain the cumulative occurrence of each disaster evolution state over multiple monitoring periods), the number of consecutive monitoring periods, and the number of state transitions (traverse the future disaster evolution state chain, and count once if adjacent states change), and calculate the state occurrence frequency corresponding to each disaster evolution state. ( For the first The cumulative number of occurrences of a disaster evolution state over multiple future monitoring periods. (Total number for future monitoring cycles) and the percentage of states remaining unchanged. ( For the first (Number of monitoring periods during which the evolutionary state of a disaster is continuously maintained) and the activity of state transitions ( For the first The number of times a disaster evolution state transitions occurs is used to calculate the probability of occurrence for each disaster evolution state based on the frequency of state occurrence, the proportion of state duration, and the activity of state transitions. ( , The frequency of state occurrence after Min-Max normalization. The proportion of state duration after Min-Max normalization. The activity of state transitions after Min-Max normalization. The weights are the frequency of occurrence of each state. The weights are the proportions of state duration. The weights for the activity of state transitions. , This represents the total number of categories of future disaster evolution states. The index serves as the index for the disaster evolution state, with weights determined through numerical optimization based on historical multi-period extrapolation experimental data, aiming to minimize the difference between the extrapolated probability distribution and the actual occurrence state. S4.3 Sort the disaster evolution states in descending order according to their occurrence probabilities. Take the disaster evolution state with the highest occurrence probability as the dominant evolution state for the corresponding future monitoring period. Combine the dominant evolution states for each future monitoring period to form a disaster evolution probability sequence (encode the dominant evolution states and their occurrence probabilities for each future monitoring period according to the timestamp of the future monitoring period, and connect them in chronological order to generate a disaster evolution probability sequence). S4.4 Analyze the changing trend of disaster evolution status within multiple future monitoring periods based on the disaster evolution probability sequence (based on the disaster evolution probability sequence, construct a sliding analysis window with a preset number of future monitoring periods (this number is 3 to 5 consecutive future monitoring periods), extract the occurrence probability corresponding to the dominant induced state within each sliding analysis window, and use the least squares method to fit the trend and calculate the slope of the change in the corresponding occurrence probability; at the same time, count the number of monitoring periods in which the dominant induced state is continuously maintained and the number of state transitions within each sliding analysis window, and perform time-series correlation analysis in combination with the changes in the dominant induced state between adjacent sliding analysis windows; based on the slope of the change in the occurrence probability, the continuous maintenance of the dominant induced state and the state transition, determine the changing trend of disaster evolution status within multiple future monitoring periods), and generate a complex geological disaster evolution trend prediction result including the disaster evolution status of each future monitoring period, the occurrence probability corresponding to each disaster evolution status, and the disaster evolution trend.

[0021] Example 2: This example provides a machine learning-based system for predicting the evolution of complex geological hazards, including a memory, a processor, and a computer program stored in the memory and run on the processor. The processor executes the computer program to implement the machine learning-based method for predicting the evolution of complex geological hazards described in Example 1.

[0022] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for predicting the evolution of complex geological hazards based on machine learning, characterized in that, include: S1. Collect geological survey data, digital elevation models and multi-source monitoring data of potential geological disaster sites, and establish a complex spatiotemporal correlation dataset of geological disasters; S2. Establish a disaster evolution state chain characterizing the continuous evolution process of disasters based on a complex geological disaster spatiotemporal correlation dataset; S3. Generate continuous state segments based on the disaster evolution state chain, count the number of monitoring periods in which each continuous state segment is continuously maintained and the duration of the state, calculate the continuous maintenance ratio corresponding to each continuous state segment, and calculate the state stability contribution corresponding to each continuous state segment in combination with the state switching density. Generate a weighted disaster evolution state sequence based on the state stability contribution, and use a pre-trained Transformer model to predict the disaster evolution state corresponding to multiple future monitoring periods based on the weighted disaster evolution state sequence, and generate the future disaster evolution state chain corresponding to multiple future monitoring periods. S4. Calculate the probability of occurrence of each evolution state based on the future disaster evolution state chain, and generate the prediction results of the evolution trend of complex geological disasters.

2. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 1, characterized in that, In step S1, establishing a spatiotemporal correlation dataset for complex geological disasters includes the following steps: Geological survey data, digital elevation models, and multi-source monitoring data collected from monitoring equipment corresponding to various geological hazard hazard points are collected. Based on a unified spatial coordinate benchmark and time benchmark, the geological survey data, digital elevation models, and multi-source monitoring data are standardized in terms of coordinates, time, and data format to form standardized geological hazard monitoring data. The standardized geological hazard monitoring data are linked and organized according to the hazard point number, monitoring time, and data type to establish the correspondence between geological environment information, topographic information, and monitoring response information corresponding to each monitoring time, forming a complex spatiotemporal correlation dataset of geological hazards.

3. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 1, characterized in that, In step S2, establishing a disaster evolution state chain characterizing the continuous evolution process of a disaster based on a complex geological disaster spatiotemporal correlation dataset includes the following steps: S2.

1. Based on the spatiotemporal correlation dataset of complex geological disasters, extract the geological environment features, topographic features and monitoring response features corresponding to each monitoring time to form a disaster state feature set corresponding to each monitoring time; S2.2 Based on the disaster state feature set corresponding to each monitoring time, analyze the changing relationship of disaster state features between adjacent monitoring times and identify the disaster evolution state corresponding to each monitoring time; S2.

3. Based on multi-source monitoring data, establish a groundwater infiltration response chain and identify infiltration response stagnation intervals. Introduce external hydraulic gradient constraints to calculate the effective transmission rate of infiltration driving force. Based on the effective transmission rate of infiltration driving force, add a continuity confidence label to the disaster evolution state within the infiltration response stagnation interval to form an enhanced disaster evolution state. S2.

4. Based on the enhanced disaster evolution state and monitoring time sequence, establish the state transition relationship between adjacent disaster evolution states to form a disaster evolution state transition sequence; S2.

5. Based on the disaster evolution state transition sequence, connect the disaster evolution states corresponding to each monitoring time to establish a disaster evolution state chain that characterizes the continuous evolution process of the disaster.

4. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 3, characterized in that, In step S2.3, a groundwater infiltration response chain is established and infiltration response stagnation intervals are identified. An external hydraulic gradient constraint is introduced to calculate the effective transmission rate of the infiltration driving force. Based on the effective transmission rate of the infiltration driving force, a continuity confidence label is added to the disaster evolution state within the infiltration response stagnation interval to form an enhanced disaster evolution state. This includes the following steps: S2.

31. Based on multi-source monitoring data within a continuous monitoring period, establish rainfall change sequences, soil moisture content change sequences, slope angle change sequences, and acceleration change sequences according to a unified monitoring period. Based on the temporal variation patterns corresponding to each change sequence, analyze the temporal response relationship between rainfall changes, soil moisture content changes, and slope attitude changes, identify the order of responses between rainfall, groundwater infiltration, and slope attitude response, and establish a groundwater infiltration response chain. S2.

32. Based on the groundwater infiltration response chain, the infiltration-induced deformation delay time and its rate of change corresponding to adjacent monitoring cycles are statistically analyzed according to the continuous monitoring cycle, and the infiltration response stagnation interval is generated. S2.

33. Based on the boundary response characteristics before and after the infiltration response stagnation interval, an external hydraulic gradient constraint is introduced. The effective transmission rate is calculated by the ratio of the infiltration front advance rate to the saturated permeability coefficient as a physical verification item. The transmission continuity is analyzed in conjunction with the boundary response characteristics to generate a groundwater infiltration transmission retention coefficient that characterizes the continuous transmission capacity of infiltration. S2.

34. Based on the groundwater infiltration transmission retention coefficient, the continuity of the disaster evolution state corresponding to the infiltration response stagnation interval is determined. Based on the continuity determination result, a continuity confidence label is generated, and an enhanced disaster evolution state with a continuity confidence label is formed.

5. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 4, characterized in that, In S2.33, based on the boundary response characteristics before and after the infiltration response stagnation interval, an external hydraulic gradient constraint is introduced. The effective transmission rate is calculated as a physical verification item by the ratio of the infiltration front advance rate to the saturated permeability coefficient. The transmission continuity is analyzed in conjunction with the boundary response characteristics to generate a groundwater infiltration transmission retention coefficient that characterizes the continuous transmission capacity of infiltration. This includes the following steps: Soil moisture content, dip angle, and acceleration response characteristics at the start and end times of the infiltration response stagnation interval are extracted as response characteristics of the front and rear boundaries of the stagnation interval. The topographic slope of the corresponding area of ​​the stagnation interval is calculated based on the digital elevation model, and external hydraulic gradient constraints are established in combination with the groundwater flow direction. The effective transmission rate of infiltration driving force is calculated based on the ratio of the groundwater infiltration front advance rate to the saturated permeability coefficient of the corresponding area, and the effective transmission rate is used as a physical verification item. Based on the response characteristics of the boundary before and after the stagnation zone, the consistency of the direction of change, the continuity of the magnitude of change, and the continuity of the rate of change of the groundwater infiltration driving force within the stagnation zone are analyzed. After normalizing the continuity analysis results with the effective transmission rate of the infiltration driving force, a groundwater infiltration transmission retention coefficient, which characterizes the continuous transmission capacity of groundwater infiltration, is generated by weighted fusion.

6. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 1, characterized in that, In step S3, a weighted disaster evolution state sequence is generated, and a pre-trained Transformer model is used to predict the disaster evolution state corresponding to multiple future monitoring periods based on the weighted disaster evolution state sequence, thereby generating a future disaster evolution state chain corresponding to multiple future monitoring periods. This includes the following steps: S3.1 Divide the disaster evolution state chain into multiple continuous state segments according to the monitoring time sequence, count the number of monitoring cycles and the duration of each continuous state segment, calculate the continuous maintenance ratio of each continuous state segment, and calculate the state stability contribution of each continuous state segment in combination with the state switching density. Generate a weighted disaster evolution state sequence based on the state stability contribution. S3.2 Encode the weighted disaster evolution state sequence into the corresponding state sequence and input it into the pre-trained Transformer model. Use the multi-head self-attention mechanism to learn the long-term state transition law between disaster evolution states and establish the disaster evolution state transition mapping relationship. S3.3 Based on the disaster evolution state transition mapping relationship, perform sequence deduction on the weighted disaster evolution state sequence, predict the disaster evolution state corresponding to multiple future monitoring periods, and connect each predicted disaster evolution state according to the monitoring time order to generate the future disaster evolution state chain corresponding to multiple future monitoring periods.

7. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 6, characterized in that, In step S3.1, the number of monitoring cycles and the duration of each continuous state segment are counted, the continuous maintenance ratio of each continuous state segment is calculated, and the state stability contribution of each continuous state segment is calculated in combination with the state switching density. Based on the state stability contribution, a weighted disaster evolution state sequence is generated, including the following steps: S3.

11. Divide the disaster evolution state chain into multiple continuous state segments according to the monitoring time sequence. Using the current disaster evolution state as the reference state, backtrack a preset number of continuous disaster evolution states in the historical direction, and count the number of monitoring cycles that each continuous state segment can maintain continuously and the corresponding state duration. S3.

12. Calculate the continuous maintenance ratio based on the number of monitoring cycles and the duration of each continuous state segment, and calculate the state stability contribution of each continuous state segment in combination with the state switching density. S3.

13. Based on the state stability contribution of each continuous state segment, the weights of each continuous state segment in the disaster evolution state chain are adjusted to generate a weighted disaster evolution state sequence.

8. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 7, characterized in that, In step S3.12, the continuous maintenance ratio is calculated based on the number of monitoring cycles and the duration of each continuous state segment, and the state stability contribution corresponding to each continuous state segment is calculated in conjunction with the state switching density. This includes the following steps: The number of monitoring cycles and the duration of each continuous state segment are normalized respectively. Based on the normalized number of monitoring cycles and the normalized duration of the state, the continuous maintenance ratio corresponding to each continuous state segment is generated by linear weighted fusion. Identify the state transition locations between adjacent continuous state segments, count the number of times the disaster evolution state changes within a unit monitoring period, and calculate the state transition density corresponding to each continuous state segment by combining the monitoring time interval between each state transition location. The system is normalized based on the continuous maintenance ratio and state switching density, and then weighted and fused according to preset weights to generate the state stability contribution of each continuous state segment.

9. The method for predicting the evolution of complex geological hazards based on machine learning according to claim 1, characterized in that, In step S4, the probability of occurrence corresponding to each evolution state is calculated based on the future disaster evolution state chain to generate a prediction result of the evolution trend of complex geological disasters, including the following steps: S4.

1. Based on the future disaster evolution state chain, statistically analyze the disaster evolution state types corresponding to each future monitoring period and the number of monitoring periods in which each disaster evolution state is continuously maintained, and establish a time series distribution sequence of future disaster evolution states. S4.2 Based on the time-series distribution sequence of future disaster evolution states, the cumulative number of occurrences, the number of consecutive monitoring periods, and the number of state transitions of each disaster evolution state in multiple future monitoring periods are statistically analyzed. The occurrence frequency, state duration ratio, and state transition activity corresponding to each disaster evolution state are calculated, and the occurrence probability corresponding to each disaster evolution state is calculated based on the occurrence frequency, state duration ratio, and state transition activity. S4.3 Sort the disaster evolution states in descending order according to their occurrence probabilities, take the disaster evolution state with the highest occurrence probability as the dominant evolution state for the corresponding future monitoring period, and combine the dominant evolution states for each future monitoring period to form a disaster evolution probability sequence. S4.

4. Based on the analysis of the probability sequence of disaster evolution, the changing trend of disaster evolution status in multiple future monitoring periods is analyzed to generate the prediction results of the evolution of complex geological disasters.

10. A machine learning-based system for predicting the evolution of complex geological hazards, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the machine learning-based method for predicting the evolution of complex geological disasters as described in any one of claims 1-9.

Images