A geological disaster monitoring and early warning method and system based on big data

Abstract

The present application relates to the technical field of monitoring and early warning, in particular to a geological disaster monitoring and early warning method and system based on big data, which comprises the following steps: collecting real-time monitoring data and transmitting the data to an edge computing node to generate a landslide multi-source time sequence monitoring data set; based on edge computing, calculating the landslide displacement time sequence characteristics of each monitoring point using the landslide multi-source time sequence monitoring data set, extracting crack propagation characteristics, slope body inclination characteristics and rainfall process characteristics, and using the deformation acceleration and slope body inclination angle change rate corresponding to each monitoring time, analyzing the cooperative change relationship, generating the landslide dynamic response coupling factor corresponding to each monitoring point, and generating the landslide deformation and time response relationship. The present application greatly reduces the data volume uploaded to the cloud platform by sinking the tasks of cleaning, synchronization, feature extraction and landslide dynamic response coupling factor calculation of multi-source monitoring data to the edge computing node for on-site processing.

Description

Technical Field

[0001] This invention relates to the field of monitoring and early warning technology, specifically to a geological disaster monitoring and early warning method and system based on big data. Background Technology

[0002] In the field of landslide geological disaster monitoring and early warning, there are significant differences in the collection frequency, dimensions, and physical response time of multi-source heterogeneous monitoring data (including displacement, cracks, dip angle, rainfall, etc.). Traditional early warning methods that rely on a single displacement threshold or comprehensive judgment based on human experience are no longer able to meet the real-time and accuracy requirements of large-scale, high-frequency monitoring scenarios. At the same time, although historical landslide monitoring big data has accumulated a wealth of evolution cases, the huge differences in geological conditions, triggering mechanisms, and deformation patterns among different landslides mean that directly using them as mixed samples for model training will produce serious sample noise interference. Existing data-driven methods lack effective means to deeply integrate prior knowledge of physics and mechanics with statistical learning frameworks. In addition, with the continuous increase in the number of monitoring points and sampling frequency, the network transmission pressure and computing latency brought about by the convergence of multi-source data to the cloud are becoming increasingly prominent. Especially in scenarios where communication conditions are limited in remote mountainous areas, the centralized cloud processing mode cannot guarantee the real-time and continuous early warning response. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a geological disaster monitoring and early warning method and system based on big data.

[0004] The technical solution of this invention: A geological disaster monitoring and early warning method based on big data, comprising: S1. Collect real-time monitoring data and transmit it to the edge computing node to generate a multi-source time-series monitoring dataset for landslides; S2. Based on edge computing, the landslide displacement time series characteristics of each monitoring point are calculated using the multi-source time series monitoring dataset of landslides. Crack propagation characteristics, slope tilt characteristics and rainfall process characteristics are extracted. The deformation acceleration and slope tilt angle change rate corresponding to each monitoring time are used to analyze the cooperative change relationship, generate the landslide dynamic response coupling factor corresponding to each monitoring point, and generate the landslide deformation and time response relationship. S3. Construct a dynamic evolution constraint domain for landslides based on the relationship between landslide deformation and time response, generate a constraint evolution sample set, and output the membership degree of the landslide deformation stage based on the constraint evolution sample set using a long short-term memory network. S4. Determine the landslide deformation stage based on the membership degree, and conduct landslide trend early warning analysis and displacement threshold alarm analysis in combination with the cumulative displacement of monitoring points, and generate comprehensive landslide monitoring and early warning results.

[0005] As a further improvement to this technical solution, in step S1, real-time monitoring data is collected and transmitted to the edge computing node to generate a landslide multi-source time-series monitoring dataset, including the following steps: Real-time data collection is performed on landslide surface displacement, crack displacement, slope tilt, and rainfall data according to a preset sampling period. The data is then transmitted to edge computing nodes via a wireless communication network. The edge computing nodes perform time synchronization processing on various monitoring data based on a unified time reference, and correlate and fuse various monitoring data according to monitoring time and spatial location of monitoring points to generate a multi-source time-series monitoring dataset for landslides.

[0006] As a further improvement to this technical solution, in step S2, the landslide displacement time-series characteristics of each monitoring point are calculated, crack propagation characteristics, slope tilt characteristics, and rainfall process characteristics are extracted, and the deformation acceleration and slope tilt angle change rate corresponding to each monitoring time are used to analyze the coordinated change relationship, generate the landslide dynamic response coupling factor corresponding to each monitoring point, and generate the landslide deformation and time response relationship, including the following steps: S2.1 Based on edge computing nodes, read the landslide surface displacement data from the multi-source time-series monitoring dataset of landslides according to the monitoring points, and generate landslide displacement time-series features; S2.2. Generate crack propagation features based on crack displacement data, generate slope tilt features based on slope tilt data, and generate rainfall process features based on rainfall data; S2.3. Using the deformation acceleration and slope tilt angle change rate corresponding to each monitoring time, analyze the synergistic change relationship between the two in the continuous monitoring period, and generate the landslide dynamic response coupling factor corresponding to each monitoring point. S2.4. The landslide displacement time series characteristics, crack propagation characteristics, slope tilt characteristics, rainfall process characteristics, and landslide dynamic response coupling factors are fused according to a unified monitoring time axis to construct the landslide dynamic evolution characteristic matrix corresponding to each monitoring point. S2.5. Based on the unified monitoring time axis, perform time-series correlation analysis on the landslide dynamic evolution characteristic matrix to generate the relationship between landslide deformation and time response.

[0007] As a further improvement to this technical solution, in step S2.3, the deformation acceleration and slope tilt angle change rate corresponding to each monitoring time are used to analyze the coordinated change relationship between the two within the continuous monitoring period, and to generate the landslide dynamic response coupling factor corresponding to each monitoring point, including the following steps: S2.31. Using the current monitoring time as the analysis endpoint, search in reverse along the time axis for the starting time when the deformation acceleration first shows a continuous increasing trend, and take the time interval between the starting time and the current monitoring time as the effective collaborative analysis period. Simultaneously extract the corresponding slope tilt change rate time series within the effective collaborative analysis period to construct a response-driven sliding window. S2.32. Based on the variation characteristics of the deformation acceleration time series within the response-driven sliding time window, the evolution process of deformation acceleration is identified, and the effective collaborative analysis period is divided into acceleration segment, constant speed segment and deceleration segment in sequence. S2.33. Based on the time synchronization relationship, the time series of slope inclination change rate is segmented accordingly to form the deformation acceleration subsequence and slope inclination change rate subsequence corresponding to each dynamic stage. S2.34. Based on the effective collaborative analysis period division results, combined with the deformation acceleration subsequence and the slope tilt change rate subsequence, a weighted trend collaborative index is generated. The weighted trend collaborative index is then normalized to generate the landslide dynamic response coupling factor.

[0008] As a further improvement to this technical solution, step S2.34, generating the weighted trend synergy index, includes the following steps: The direction of change of the deformation acceleration subsequence and the slope tilt change rate subsequence between adjacent monitoring times in the acceleration, constant speed, and deceleration phases are calculated respectively. When the direction of change is the same, it is determined to be a state of coordinated change. The time interval of continuous coordinated change in each phase is counted, and the proportion of coordinated change time to the total duration of the corresponding phase is calculated. The coordination degree of acceleration phase, constant speed phase, and deceleration phase are obtained respectively. The weighted trend coordination index corresponding to each monitoring point is generated by weighted fusion.

[0009] As a further improvement to this technical solution, in step S3, a dynamic evolution constraint domain for landslides is constructed based on the relationship between landslide deformation and time response, a constraint evolution sample set is generated, and the membership degree of the landslide deformation stage is output using a long short-term memory network based on the constraint evolution sample set, including the following steps: S3.1 Receive the relationship between landslide deformation and time response, retrieve the historical monitoring records of the corresponding area in the historical landslide monitoring big data, perform time registration and feature standardization processing on the current landslide deformation and time response relationship and historical monitoring data according to the unified time axis, and generate a landslide evolution feature sequence between the current monitoring object and historical landslide samples; S3.2 Construct a dynamic evolution constraint domain for landslides based on the relationship between landslide deformation and time response, and perform evolution constraint screening on historical landslide monitoring big data based on the dynamic evolution constraint domain to generate a constraint evolution sample set; S3.3 Using the constrained evolution sample set as the training reference sample, and combining it with the landslide dynamic evolution feature matrix corresponding to the current landslide, construct the stage evolution feature vector corresponding to the current landslide. S3.4 Input the stage evolution feature vector into the pre-trained long short-term memory network, and use the evolution law of the corresponding landslide deformation stage in the constrained evolution sample set to constrain and correct the current landslide stage identification result, and generate the membership degree of the current landslide belonging to the initial deformation stage, uniform deformation stage and accelerated deformation stage.

[0010] As a further improvement to this technical solution, in step S3.2, a dynamic evolution constraint domain for landslides is constructed based on the relationship between landslide deformation and time response. Based on this dynamic evolution constraint domain, historical landslide monitoring big data is used to filter evolutionary constraints and generate a constraint evolution sample set. This includes the following steps: S3.21. Based on the relationship between landslide deformation and time response, and construct a multidimensional monitoring feature sequence according to a unified time axis; use a nonlinear multidimensional projection method to map each monitoring feature sequence to a unified evolution space, and introduce a local neighborhood preservation metric to generate state nodes and state transition trajectories corresponding to each monitoring time, and construct the dynamic evolution state space corresponding to the current landslide. S3.22. Based on the temporal connection relationship and trajectory continuous change law between each state node in the dynamic evolution state space, analyze the continuous reachability relationship between any adjacent state nodes and identify the state boundary in the state transition process. S3.23. Using a time-sharing recursive construction mode, calculate the evolution constraint strength corresponding to each state boundary to generate a landslide evolution constraint domain that describes the normal evolution range of the landslide. S3.24. Based on the landslide evolution constraint domain, perform point-by-point positive progressive verification on the current state nodes to form a constraint evolution sample set.

[0011] As a further improvement to this technical solution, in S3.23, a time-sharing recursive construction mode is adopted to calculate the evolution constraint strength corresponding to each state boundary, and generate a landslide evolution constraint domain describing the normal evolution range of the landslide, including the following steps: Based on the evolution states in the dynamic evolution feature matrix of landslides, they are mapped as state nodes, and adjacent state nodes are connected in chronological order to form local state evolution paths. The connection area between adjacent state nodes is defined as the state boundary, and incremental adaptive density clustering is used to statistically analyze the frequency of discrete partitioning. The changes in Mahalanobis distance between the state nodes on both sides of each state boundary in the original high-dimensional feature space are calculated respectively; the changes in state transition probability between adjacent discrete states are calculated based on the discrete partition transition frequency; and the degree of continuous change of the dynamic response coupling factor of the landslide on both sides of the state boundary is calculated by combining the relationship between landslide deformation and time response, and a dynamic response continuity index is generated. The statistical soft constraint strength corresponding to each state boundary is obtained by normalizing the changes in Mahalanobis distance, the changes in state transition probability, and the continuity index of dynamic response, and then performing weighted fusion calculations to generate the conditional evolution constraint domain.

[0012] As a further improvement to this technical solution, in step S4, the landslide deformation stage is determined based on the membership degree, and landslide trend early warning analysis and displacement threshold alarm analysis are carried out in combination with the cumulative displacement of monitoring points, and a comprehensive landslide monitoring and early warning result is generated, including the following steps: S4.1 Extract the maximum membership degree based on the membership degree and perform stage confidence determination to determine the deformation stage to which the current landslide belongs; S4.2 Based on the landslide deformation stage and the corresponding benchmark deformation rate, the cumulative displacement at each monitoring time is normalized to the time dimension to obtain standardized time parameters with the same dimension as the monitoring time, and a standardized deformation-time curve is generated. S4.3 Based on the standardized deformation and time curves, calculate the improved tangent angle corresponding to each monitoring time, and combine the trend of the improved tangent angle to refine the identification of the accelerated deformation stage of the landslide. Divide the accelerated deformation stage into the initial acceleration stage, the intermediate acceleration stage and the imminent landslide stage, and generate landslide trend early warning results. S4.4. Based on the cumulative displacement of the monitoring points, calculate the current cumulative displacement value of each monitoring point and the number of monitoring points that exceed the preset displacement threshold, as well as the proportion of the total number of monitoring points. Then, determine the threshold according to the preset alarm rules and generate a displacement threshold alarm result. S4.5. Based on the comprehensive landslide trend warning results and displacement threshold alarm results, a comprehensive assessment of the current landslide hazard level is made, and a comprehensive monitoring and early warning result is generated.

[0013] On the other hand, the present invention provides a geological disaster monitoring and early warning system based on big data, 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 geological disaster monitoring and early warning method based on big data.

[0014] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: by offloading tasks such as cleaning, synchronizing, feature extraction and landslide dynamic response coupling factor calculation of multi-source monitoring data to edge computing nodes for local processing, the amount of data uploaded to the cloud platform is greatly reduced, effectively alleviating the network bandwidth pressure and cloud storage computing burden caused by concurrent transmission of multi-source heterogeneous data in geological disaster monitoring scenarios; at the same time, the edge computing nodes have the ability to independently complete all feature condensation and coupling analysis when the network is interrupted or congested, ensuring the continuity and low-latency response of the early warning process, and improving the real-time performance, reliability and scalability of the overall system. 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 geological disaster monitoring and early warning method based on big data, including the following steps: S1. Collect real-time monitoring data and transmit it to the edge computing node to generate a multi-source time-series monitoring dataset for landslides; In this embodiment, real-time monitoring data is collected and transmitted to edge computing nodes to generate a multi-source time-series monitoring dataset for landslides, including the following steps: GNSS monitoring equipment, crack monitoring equipment, tilt monitoring equipment, and rainfall monitoring equipment are deployed in the landslide monitoring area. Data on landslide surface displacement, crack displacement, slope tilt, and rainfall are collected in real time according to a preset sampling cycle and transmitted to edge computing nodes via wired or wireless communication networks (such as LoRa wireless communication networks). The edge computing nodes perform time synchronization processing on various monitoring data based on a unified time reference, sequentially completing outlier removal, missing data compensation, noise filtering, and standardization preprocessing. Subsequently, the various monitoring data are correlated and fused according to monitoring time and spatial location of monitoring points to generate a multi-source time-series monitoring dataset for landslides, providing a unified data foundation for subsequent landslide stability analysis.

[0018] S2. Based on edge computing, the landslide displacement time series characteristics of each monitoring point are calculated using the multi-source time series monitoring dataset of landslides. Crack propagation characteristics, slope tilt characteristics and rainfall process characteristics are extracted. The deformation acceleration and slope tilt angle change rate corresponding to each monitoring time are used to analyze the cooperative change relationship, generate the landslide dynamic response coupling factor corresponding to each monitoring point, and generate the landslide deformation and time response relationship. In this embodiment, generating the relationship between landslide deformation and time response includes the following steps: S2.1 Based on edge computing nodes, the surface displacement data of the landslide body in the multi-source time-series monitoring dataset of the landslide is read according to the monitoring points. Taking the initial monitoring time as the displacement benchmark, the cumulative displacement corresponding to the current time of each monitoring point is calculated, and the stage displacement increment is calculated according to the displacement difference between adjacent monitoring times. Furthermore, based on the cumulative displacement and stage displacement increment, the deformation rate and deformation acceleration of each monitoring point are calculated using a sliding time window to generate the landslide displacement time-series characteristics. S2.2. Generate crack propagation characteristics based on crack displacement data (based on crack displacement data, statistically analyze the crack width, crack length, and crack propagation direction corresponding to each monitoring time, calculate the crack width change rate, crack length growth rate, and crack propagation direction change, and generate crack propagation characteristics based on the changing trends of crack parameters at each monitoring time), generate slope tilt characteristics based on slope tilt data (based on slope tilt data, calculate the slope tilt angle change rate, tilt direction change, and cumulative tilt angle change corresponding to each monitoring point, and analyze the consistency of slope tilt changes in conjunction with the spatial location of each monitoring point to generate slope tilt characteristics), and generate rainfall process characteristics based on rainfall data (based on rainfall data, statistically analyze rainfall intensity, cumulative rainfall, and continuous rainfall time within each monitoring period, and generate rainfall process characteristics by combining the start time, peak time, and end time of rainfall). S2.3. Using the deformation acceleration and slope tilt angle change rate corresponding to each monitoring time, the synergistic relationship between the two in the continuous monitoring period is analyzed, and a landslide dynamic response coupling factor corresponding to each monitoring point is generated to characterize the degree of synergy between deformation acceleration and slope tilt change. The larger the landslide dynamic response coupling factor, the more obvious the synergistic evolution between the two and the stronger the landslide dynamic response. In this embodiment, during the landslide deformation process, displacement acceleration and slope tilt angle change rate reflect the landslide dynamic behavior from the two dimensions of force change and attitude response, respectively. However, the two are not synchronized in time: when rainfall infiltration causes an increase in thrust, deformation acceleration responds first, while the slope tilt angle change rate is affected by soil stiffness. The lag often delays the appearance of significant changes by several hours to several days. This invention constructs a response-driven sliding time window to statistically analyze the consistency of the changing directions of deformation acceleration and tilt angle change rate in stages within an effective collaborative analysis period. It also introduces a weighted trend synergy index to upgrade the comparison between the two from physical magnitudes to a quantitative assessment of the degree of evolutionary trend synergy. This can effectively distinguish between mechanically driven active deformation (high degree of synergy between the two) and environmental disturbance-driven passive swaying (low degree of synergy between the two), thereby eliminating false acceleration signals caused by short-term disturbances such as rainfall and construction. At the same time, when the coupling factor is consistently high, it serves as a quantitative basis for the precursor of landslides entering the accelerated creep stage, thus upgrading the early warning from post-event response to pre-event perception. The process of generating the landslide dynamic response coupling factor for each monitoring point includes the following steps: S2.31. Using the current monitoring time as the analysis endpoint, search backward along the time axis for the starting time when the deformation acceleration first shows a continuous increasing trend. (First, read the original deformation acceleration sequence within the preset historical analysis window before the current monitoring time, and adaptively adjust the median filter window size according to the local fluctuation amplitude to perform adaptive median filtering and noise reduction on the original deformation acceleration sequence to eliminate the influence of high-frequency noise and abnormal sampling values ​​on trend identification; then, use a fixed-length sliding window to calculate the least squares fitted line of the deformation acceleration sequence within the window window window by window along the time axis, and calculate the corresponding fitting slope and its significance level. When the fitting slope is greater than zero and the significance test meets the preset conditions (p < 0.05, p is the approximate value of the deformation acceleration), the curve is considered to be in the range of 0.05.) When the value of the significance test corresponding to the slope is obtained, the deformation acceleration within the window is determined to have a significant increasing trend. Further, the number of monitoring points in each window where the deformation acceleration is continuously monotonically increasing is counted. When the number of continuously monotonically increasing monitoring points is no less than 3, the window is determined to meet the local continuous increasing condition. Only when the determination result of a significantly increasing fitted slope and the determination result of continuous monotony are simultaneously met is the corresponding window identified as a continuously increasing trend window. Subsequently, starting from the current monitoring time, a reverse search is performed along the time axis window by window until the earliest window that meets the continuous increasing trend condition is located, and the starting monitoring time corresponding to this window is determined as the starting time when the deformation acceleration first shows a continuous increasing trend. If the continuous increasing trend occurs during the reverse search process... If the increasing trend is interrupted, the backtracking stops, and the determined starting monitoring time is taken as the starting point of the effective collaborative analysis period. The time interval between the starting time and the current monitoring time is taken as the effective collaborative analysis period. Within the effective collaborative analysis period, the corresponding slope inclination change rate time series is extracted synchronously, and a response-driven sliding time window is constructed. (After determining the effective collaborative analysis period, the starting time and the current monitoring time of this period are used as time boundaries. In the landslide multi-source time series monitoring dataset, slope inclination change rate data of each monitoring point within the corresponding time period are extracted synchronously according to a unified time axis, and a one-to-one correspondence registration is performed based on the timestamp and deformation acceleration sequence. When there is an inconsistency in sampling time, time interpolation or the most recent time is used.) The adjacent time matching method completes the time synchronization processing, so that the deformation acceleration sequence and the slope tilt change rate sequence have the same sampling time and sampling length. Then, the effective collaborative analysis period is used as the boundary of the sliding time window. The synchronously registered deformation acceleration time series and slope tilt change rate time series within this time interval are jointly encapsulated into a response-driven sliding time window. As the monitoring time is updated, the effective collaborative analysis period is re-determined in reverse with the current monitoring time as the endpoint, and the response-driven sliding time window is updated synchronously. This provides a unified time series analysis window for subsequent collaborative change analysis of deformation acceleration and slope tilt change rate, so as to ensure that each collaborative analysis fully covers the dynamic evolution process of the landslide from stress adjustment to attitude response. S2.32. Based on the variation characteristics of the deformation acceleration time series within the response-driven sliding time window, identify the evolution process of deformation acceleration from continuous increase to basic stability to continuous decrease, and divide the effective collaborative analysis period into acceleration segment, constant velocity segment, and deceleration segment in sequence; specifically: based on the deformation acceleration time series within the response-driven sliding time window, calculate the change in deformation acceleration between adjacent monitoring times in chronological order, and use the sliding window to calculate the local change slope and slope change rate to identify the trend of the deformation acceleration sequence; when the local change slope corresponding to several consecutive monitoring times is greater than the preset positive threshold q (e.g., 0), .1) When the slope of the local change is less than the preset stable threshold b (e.g., 0.02) and the preset monitoring duration is maintained, the corresponding time interval is determined to be the continuous increasing stage; when the slope of the local change is less than the preset negative threshold c (e.g., -0.1) and the slope of the local change is maintained to be ... S2.33. Based on the time synchronization relationship, the slope inclination change rate time series is segmented accordingly to form the deformation acceleration subsequence and slope inclination change rate subsequence corresponding to each dynamic stage (based on the start and end times of the acceleration, constant speed and deceleration stages, the slope inclination change rate time series is time-synchronized and mapped according to the unified monitoring time axis, and the slope inclination change rate time series is synchronously segmented according to the time boundaries of each dynamic stage; at the same time, the deformation acceleration time series is correspondingly truncated according to the same time boundaries, so that the deformation acceleration data and slope inclination change rate data in each stage are in one-to-one correspondence in time; when the sampling times of the two types of data are inconsistent, time interpolation or nearest neighbor time matching is used to complete time alignment, and finally the deformation acceleration subsequence and slope inclination change rate subsequence corresponding to the acceleration stage, constant speed stage and deceleration stage are formed respectively). S2.34. Based on the effective synergistic analysis period division results, combined with the deformation acceleration subsequence and the slope tilt change rate subsequence, a weighted trend synergistic index is generated. The weighted trend synergistic index is normalized (using the Min-Max normalization method) to generate the landslide dynamic response coupling factor. Furthermore, the weighted trend correlation index is generated, including the following steps: The direction of change of the deformation acceleration subsequence and the slope tilt change rate subsequence between adjacent monitoring times in the acceleration, constant speed, and deceleration phases are calculated respectively. When the directions of change of the two are the same, it is determined to be a state of coordinated change. The time intervals in which the state of coordinated change is continuous in each phase are counted, and the proportion of the coordinated change time to the total duration of the corresponding phase is calculated to obtain the coordination degree of the acceleration phase, the coordination degree of the constant speed phase, and the coordination degree of the deceleration phase. The weighted trend coordination index corresponding to each monitoring point is generated by weighted fusion. (First, the actual phase category (acceleration phase, constant speed phase, or deceleration phase) that exists in the effective coordination analysis period is detected. If all three phases exist completely, the coordination degree of the acceleration phase, the coordination degree of the deceleration phase, and the coordination degree of the constant speed phase are respectively set to 0.6, ... Fixed weights of 0.3 and 0.1 are used for weighted fusion. If only the acceleration and constant speed segments are detected but the deceleration segment is not detected, the deceleration segment coordination degree is set to 0.5 (neutral value) by default, and the weights of the acceleration and constant speed segments are normalized and redistributed proportionally, i.e., their weights are adjusted to 0.667 and 0.333 respectively. If only the acceleration segment is detected, the acceleration segment weight is set to 1.0. If neither is detected, the landslide dynamic response coupling factor at the current moment of the monitoring point is marked as "unavailable," and the benchmark acceleration marker corresponding to that point is simultaneously set to "unavailable" in subsequent S4.1. Through the above dynamic normalization process, it is ensured that the total weight is always 1 under any combination of stages, avoiding the invalidation of the coupling factor calculation benchmark due to stage absence. S2.4. The landslide displacement time series characteristics, crack propagation characteristics, slope tilt characteristics, rainfall process characteristics, and landslide dynamic response coupling factors are fused according to a unified monitoring time axis to construct a landslide dynamic evolution characteristic matrix corresponding to each monitoring point (a time series characteristic matrix constructed with the unified monitoring time axis as the row index and various monitoring characteristics as the column index, where each row corresponds to a monitoring time and each column corresponds to a landslide evolution characteristic, including cumulative displacement, stage displacement increment, deformation rate, deformation acceleration, crack propagation characteristics, slope tilt characteristics, rainfall process characteristics, and landslide dynamic response coupling factors), which is used to characterize the co-evolution relationship between landslide dynamic evolution characteristics and rainfall-induced response characteristics; S2.5. Based on the unified monitoring time axis, perform time-series correlation analysis on the landslide dynamic evolution characteristic matrix to generate the relationship between landslide deformation and time response, and upload the relationship between landslide deformation and time response to the cloud platform to provide an analytical basis for subsequent landslide deformation stage identification and landslide stability evaluation.

[0019] S3. Construct a dynamic evolution constraint domain for landslides based on the relationship between landslide deformation and time response, generate a constraint evolution sample set, and output the membership degree of the landslide deformation stage based on the constraint evolution sample set using a long short-term memory network. In this embodiment, a dynamic evolution constraint domain for landslides is constructed, a constraint evolution sample set is generated, and the membership degree of the landslide deformation stage is output using a long short-term memory network based on the constraint evolution sample set, including the following steps: S3.1 Receive the landslide deformation and time response relationship uploaded by the edge computing node, and retrieve the historical monitoring records of the corresponding area or similar landslides in the historical landslide monitoring big data (at least including the cumulative displacement, stage displacement increment, and deformation rate of the historical landslides throughout the entire evolution process). Based on the unified time axis, perform time registration and feature standardization processing on the current landslide deformation and time response relationship and historical monitoring data to generate a landslide evolution feature sequence that can be compared between the current monitoring object and historical landslide samples. S3.2. Construct a dynamic evolution constraint domain for landslides based on the relationship between landslide deformation and time response. Then, based on this constraint domain, perform evolution constraint screening on historical landslide monitoring big data to generate a constraint evolution sample set. This removes historical samples inconsistent with the current landslide evolution pattern and retains historical evolution samples matching the current landslide dynamic evolution characteristics, improving the accuracy and reliability of subsequent landslide deformation stage identification. In this embodiment, the historical landslide monitoring big data comes from multiple types of landslide samples with different geological conditions, triggering mechanisms, and evolution patterns. These landslides exhibit significant differences in deformation rate magnitude, temporal morphology, and acceleration response characteristics. If these samples are directly mixed for model training, the decision boundary learned by the model will be affected by the mutual interference of samples from different mechanisms, leading to distorted stage discrimination results for the current specific landslide. For example, landslide samples with deformation rate magnitudes differing by tens of times interfere with each other after normalization, misjudging normal fluctuations of high-rate landslides as abnormal acceleration. The true acceleration signals of low-rate landslides are submerged in background noise. At the same time, key parameters such as the uniform velocity stage rate and peak acceleration time of landslides with different mechanisms lack a unified physical reference, making it impossible for the model to accurately distinguish whether the current landslide is in the uniform deformation stage or the early stage of accelerated deformation, ultimately leading to stage misjudgment and frequent false alarms. This invention constructs a constraint domain that reflects the real-time dynamic evolution characteristics of the current landslide, and selects samples similar to the current landslide evolution path from historical big data as needed to generate a constraint evolution sample set. This ensures that the training reference samples and the current monitoring object maintain local topological consistency in the feature space. It not only eliminates historical interference samples that are irrelevant to the mechanism, but also retains the common characteristics of the same type of landslide in different evolution stages. Thus, while ensuring the purity of the model training samples, it deeply integrates the prior knowledge of physics and mechanics with the data-driven method, effectively solving the industry pain points of stage misjudgment caused by sample mixing and the disconnect between physical mechanisms and statistical laws in traditional data-driven methods. The process of constructing a dynamic evolution constraint domain for landslides and generating a constraint evolution sample set includes the following steps: S3.21. Based on the relationship between landslide deformation and time response, the temporal characteristics of landslide displacement, crack propagation, slope tilt, rainfall response, and landslide dynamic response coupling factors within a continuous monitoring period are processed for time synchronization, and a multidimensional monitoring feature sequence is constructed according to a unified time axis. A nonlinear multidimensional projection method (e.g., t-SNE) is used to map each monitoring feature sequence to a unified low-dimensional evolution space, maintaining the temporal adjacency and dynamic evolution relationships between each monitoring feature. A local neighborhood preservation metric is introduced to retain the local geodesic structure of the feature space during the projection process, providing a consistent geometric benchmark for subsequent distance measurements. State nodes and state transition trajectories corresponding to each monitoring time are generated, and the dynamic evolution state space corresponding to the current landslide is constructed. (Specifically, after mapping the multidimensional monitoring feature sequence to a unified low-dimensional evolution space, a local neighborhood preservation metric is introduced. Using the feature vector corresponding to each monitoring time as the center, the Euclidean distance between it and the feature vectors of adjacent monitoring times within a preset neighborhood is calculated (the preset neighborhood uses the K-nearest neighbor method, with K value taken from the current monitoring point). 5% to 15% of the total (minimum 5, maximum 20) are selected, and the three points with the smallest distance are chosen as their nearest neighbors. When mapping high-dimensional feature vectors to low-dimensional space, the optimization objective is to keep the relative distance between each point and its nearest neighbor unchanged before and after mapping. The low-dimensional coordinates are adjusted iteratively (the iteration termination condition is: the total number of iterations reaches a preset upper limit (e.g., the upper limit is 100 times), so that the sum of the squares of the difference between the ratio of the distance between any point and its nearest neighbor after mapping and the ratio before mapping is minimized (this criterion ensures that the state nodes in the low-dimensional space can truly reflect the local topological relationships in the original feature space). Then, the low-dimensional feature vectors corresponding to each monitoring time are defined as state nodes, and adjacent state nodes are connected sequentially according to the monitoring time to form a state transition trajectory. Finally, all state nodes and their corresponding state transition trajectories are used to construct the dynamic evolution state space corresponding to the current landslide, which is used to characterize the temporal evolution relationship and dynamic evolution path between the various evolution states of the landslide in the continuous monitoring period, providing a basis for subsequent state boundary identification and landslide evolution constraint domain construction). S3.22. Based on the temporal connectivity and trajectory continuous change patterns between state nodes in the dynamic evolution state space, analyze the continuous reachability between any adjacent state nodes and identify the state boundaries during state transitions. Specifically, each state node in the dynamic evolution state space is used as an analysis unit. The Euclidean distance, Mahalanobis distance, and motion trend angle between adjacent nodes in the unified low-dimensional evolution space are calculated sequentially along the time axis. The transition rate and direction change rate between adjacent states are derived by combining the corresponding time intervals. If the transition rate suddenly jumps to more than three times the standard deviation of the local historical mean or the motion direction angle changes abruptly by more than 90 degrees, the transition rate is considered a critical factor. If the continuous reachability between adjacent nodes is significantly weakened, then the local rate of change of multi-source monitoring features such as deformation acceleration, slope tilt angle change rate, crack width expansion rate, and rainfall response intensity within the transfer interval is further extracted. By calculating the sliding Pearson correlation coefficient between deformation acceleration and tilt change rate, if the correlation coefficient changes from positive to negative and the absolute value is consistently lower than the preset threshold of 0.3, it is determined to be a dynamic response instability boundary. At the same time, if the time lag between the moment of sudden change in crack expansion direction and the moment of peak rainfall response exceeds the preset response lag threshold (e.g., 6 hours), it is marked as a multi-source monitoring feature mismatch boundary. S3.23. Using a time-sharing recursive construction mode, the evolution constraint strength corresponding to each state boundary is calculated to generate a landslide evolution constraint domain that describes the normal evolution range of the landslide, so as to dynamically characterize the boundary range of the normal evolution state of the landslide and provide a constraint basis for subsequent screening of historical samples that conform to the current landslide evolution law and identification of abnormal evolution states. Furthermore, generating a landslide evolution constraint domain describing the normal evolution range of the landslide includes the following steps: In each recursive step, based on the landslide dynamic evolution feature matrix... to Evolutionary state within the interval ( The index of the monitoring time is used to map the state nodes, and adjacent state nodes are connected in chronological order to form a local state evolution path. The connection area between adjacent state nodes is defined as the state boundary. Incremental adaptive density clustering is used to count the frequency of discrete partition transitions. (Incremental adaptive density clustering (such as online mean drift) is used to discretize the continuous evolution state space within the current prefix interval, transforming the prefix trajectory into a discrete state sequence. The clustering bandwidth is set to 30% of the median Euclidean distance between all pairs of state nodes. Starting from the first node, nodes are processed sequentially. If the minimum distance between the current node and the existing cluster center is less than the bandwidth, the node is assigned to that cluster and the center coordinates are updated. Otherwise, a new cluster is created. Every ten nodes are processed, a merging check is performed on the cluster centers. If the distance between two centers is less than half the bandwidth, they are merged. The forgetting factor is 0.95. After clustering, each state node is replaced with the category number of its cluster, thus transforming the continuous state space into a discrete state sequence.) Only the frequency of discrete partition transitions between discrete partitions within this prefix interval is counted, thus providing a pure historical statistical sample space for probability calculation. The changes in Mahalanobis distance between state nodes on both sides of each state boundary in the original high-dimensional feature space are calculated to characterize the true magnitude of changes in the evolutionary state. Based on the frequency of discrete partition transitions, the changes in state transition probabilities between adjacent discrete states are calculated (the number of state transitions between adjacent discrete states is counted in the order of monitoring time to construct a discrete state transition statistics matrix; for any discrete state, the total number of state transitions of that state as the current state is used as the denominator, and the number of transitions from that state to adjacent discrete states is used as the numerator to calculate the state transition probability from the current state to each adjacent discrete state; then, the difference in state transition probabilities corresponding to the discrete states on both sides of the current state boundary is calculated, and combined with the magnitude of changes in state transition probabilities at adjacent monitoring times, the change in state transition probability is generated) to characterize the landslide evolution state crossing the current state boundary. The statistical difficulty is assessed; simultaneously, combining the relationship between landslide deformation and time response, the degree of continuous change of the landslide dynamic response coupling factor on both sides of the state boundary is calculated to generate a dynamic response continuity index (based on the landslide dynamic response coupling factor at the corresponding monitoring time on both sides of the state boundary, the time series of the landslide dynamic response coupling factor within the preset time window before and after the state boundary is extracted, and time synchronization is completed according to the relationship between landslide deformation and time response; then, the average value and rate of change of the landslide dynamic response coupling factor on both sides of the state boundary are calculated, and the difference in the average value and rate of change of the landslide dynamic response coupling factor before and after the state boundary are further calculated, and normalization is performed respectively (using the Min-Max normalization method); the landslide dynamic response continuity index is generated by weighted fusion based on the difference in the normalized average value and the difference in the rate of change). The parameters are normalized based on the changes in Mahalanobis distance, state transition probability, and dynamic response continuity (using the Min-Max normalization method), and then weighted and fused to obtain the statistical soft constraint strength corresponding to each state boundary, thus generating the conditional evolution constraint domain. (Using the fixed physical hard constraint boundary marked in step S3.22 as the insurmountable outer envelope, within this envelope, the statistical soft constraint strength within this prefix interval is less than the dynamic adaptive threshold) The region consisting of interconnected state nodes (either the 95th quantile of the distribution of soft constraint intensity in the neighborhood of the current state node within this prefix, or twice the standard deviation of the mean within the sliding window of this prefix) is defined as the normal evolution region of the landslide under this prefix condition, and is used to generate a region for judgment. Conditional evolution constraint domain of state at time step ); S3.24. Based on the landslide evolution constraint domain, perform point-by-point positive progressive verification of the current state nodes to form a constraint evolution sample set (for the current monitoring time to be verified). Extract only the current state node This is compared with the pure prefix data in step S3.23. to Constructed condition constraint domain Spatial location comparison is performed first, followed by hard boundary violation detection. If If a point is located outside the physical hard constraint boundary (i.e., beyond the fixed outer envelope), it is immediately identified as an anomaly indicating irreversible instability, directly removed, and an early warning is triggered; if If a point is located inside the physical hard boundary, it is further determined whether it falls into the statistical soft constraint region: if it does, the point is retained as a normal evolution sample; if it is located outside the soft constraint region but still inside the hard boundary, a local temporal consistency backtracking check (rather than a global trajectory average angle check) is initiated: the last point within this prefix interval is extracted. At any moment (e.g.) ) local trajectory direction vector, calculate its relationship with the current point The angle between the local motion directions; simultaneously, a corrected local dynamic time-warped distance is introduced, limited only to this prefix. to The nearest neighbor trajectory segment within the current local trajectory segment is used as a reference template for secondary similarity verification. This corrected distance is achieved through three constraints based on standard dynamic time warping: the endpoint of the warped path is forced to align with the current monitoring time to anchor the reference; the maximum offset of the time axis curvature is limited to maintain local temporal continuity; and an additional distance cost is added to the time index difference between matching point pairs to penalize severely misaligned matching methods. During verification, the minimum cumulative distance with penalty is first calculated under the three constraints, and then this distance and the average time offset of the optimal path are compared with the preset threshold and the allowable offset, respectively. If the cumulative distance with penalty does not exceed the threshold and the average time offset is not greater than one monitoring interval, the current deviation is determined to be a temporary external disturbance (such as a single drop). If the deviation is caused by peak rainfall or instantaneous construction load, and the deviation is within the range of physically recoverable energy potential well (i.e., the deviation of the current deformation rate from the reference rate of its uniform deformation stage does not exceed 1.5 times, and the relative deviation of the cumulative displacement from the historical average value does not exceed ±15%, and the deviation lasts for no more than 3 consecutive monitoring cycles and shows a decaying trend during this period (e.g., the rate decreases in adjacent cycles), and exceeds this range, it is judged as an irreversible deviation), then it is temporarily retained and marked as "critical concern" (not included in the normal evolution sample, but placed in the observation buffer); if the local deviation shows a continuous diffusion trend (deviation from the soft constraint area for 3 or more consecutive moments) or the sudden change in the local motion direction violates the constitutive relationship of landslide creep mechanics, then it is removed; After the time is determined, Incorporate into historical databases, Repeat steps S3.23 to S3.24 above to realize the constraint domain is updated and evolved over time, and finally outputs a constraint evolution sample set that satisfies strict time causality. S3.3 Using the constrained evolution sample set as training reference samples, and combining it with the landslide dynamic evolution feature matrix corresponding to the current landslide, construct the stage evolution feature vector corresponding to the current landslide (using the constrained evolution sample set as training reference samples, extracting the landslide dynamic evolution feature matrix corresponding to each historical sample according to the unified monitoring time axis, and statistically analyzing the feature evolution patterns corresponding to different landslide deformation stages; combining the landslide dynamic evolution feature matrix corresponding to the current landslide, extracting the temporal variation features of cumulative displacement, deformation rate, deformation acceleration, crack propagation features, slope tilt features, rainfall process features, and landslide dynamic response coupling factors within the current monitoring time and preset historical time window, and performing feature mapping and unified encoding according to the stage evolution patterns in the constrained evolution sample set, constructing a stage evolution feature vector representing the current comprehensive evolution state of the landslide, which is used for subsequent landslide deformation stage identification). S3.4. Input the stage evolution feature vector into a pre-trained Long Short-Term Memory (LSTM) network, and use the evolution law of the corresponding landslide deformation stages in the constraint evolution sample set to constrain and correct the current landslide stage identification result, generating the membership degree of the current landslide belonging to the initial deformation stage, uniform deformation stage, and accelerated deformation stage. The pre-trained LSM network includes an input layer, a hidden layer, a fully connected layer, and an output layer. The input layer receives the stage evolution feature vector, and the hidden layer consists of two stacked LSM units: the first layer contains 128 memory units, and the second layer contains 64 memory units. Each layer uses a hyperbolic tangent activation function, and the initial bias of the forget gate is set to 1.0 to enhance the retention of long-term dependencies. A random deactivation layer with a dropout rate of 20% is added between the two layers to prevent overfitting. The fully connected layer contains 32 neurons and uses a linear rectified activation function. The output layer consists of three neurons and employs a flexible maximum value classifier to output the probability values ​​of belonging to the initial deformation stage, the uniform deformation stage, and the accelerated deformation stage, respectively. Network training uses a constrained evolution sample set as training data. Each sample is divided into a sliding window of one year in chronological order (with a step size of one month). The feature sequence within each window serves as a training sequence, and the corresponding true stage label is taken from the stage marked at the end of the window. The loss function is cross-entropy, and the optimization algorithm uses adaptive moment estimation. During training, the learning rate is reduced to one-tenth of its original value when the validation set loss no longer decreases for five consecutive epochs. The maximum number of training epochs is 200, and the early termination condition is that the validation set loss does not improve for ten consecutive epochs. The ratio of training set to validation set is 8:2.

[0020] S4. Determine the landslide deformation stage based on the membership degree, and conduct landslide trend early warning analysis and displacement threshold alarm analysis respectively in combination with the cumulative displacement of monitoring points, and generate landslide monitoring and early warning results in a comprehensive manner. In this embodiment, the comprehensive generation of landslide monitoring and early warning results includes the following steps: S4.1 Extract the maximum membership degree based on the membership degree and determine the stage confidence level to identify the deformation stage of the current landslide. Select the subsequent analysis path based on the stage identification confidence level. When the maximum membership degree reaches the preset confidence threshold a (e.g., 0.65, dynamically adjusted according to the quality of monitoring data), the corresponding deformation stage is taken as the current landslide deformation stage. When the maximum membership degree is lower than the preset confidence threshold a, the stage corresponding to the maximum membership degree is taken as the current landslide deformation stage for analysis, and a low confidence level label is added. At the same time, extract the average deformation rate of the corresponding uniform deformation stage as the standardization analysis benchmark based on the determined landslide deformation stage. If there is no complete uniform deformation stage in the historical monitoring data, the average deformation rate of the uniform deformation stage corresponding to the same type of landslide in the historical landslide monitoring big data is used as the benchmark parameter. S4.2. Based on the landslide deformation stage and the corresponding benchmark deformation rate, the cumulative displacement at each monitoring time is normalized to a time dimension to obtain standardized time parameters with the same dimension as the monitoring time. This generates a standardized deformation-time curve (using the benchmark deformation rate determined in S4.1 as the time dimension conversion benchmark, dividing the cumulative displacement at each monitoring time by the benchmark deformation rate yields the corresponding standardized time parameters). To standardize time parameters Compared with actual monitoring time Having the same time dimension; subsequently, based on the actual monitoring time. For the x-axis, standardized time parameters Using the vertical axis, plot the data points corresponding to each monitoring time and connect them to form a standardized deformation relative to time. The curve represents the standardized time increment between adjacent monitoring moments during the uniform deformation phase. Equal to the corresponding actual time interval (Right now Standardized time increment between adjacent monitoring times during the initial deformation phase Less than the corresponding actual time interval (Right now Standardized time increment between adjacent monitoring times during the accelerated deformation phase Greater than the corresponding actual time interval (Right now This eliminates the problem of non-unique tangent angles in traditional cumulative displacement-time curves due to different coordinate dimensions, providing a standardized deformation curve with unified dimensions, uniqueness, and comparability for subsequent improvements in tangent angle analysis. S4.3. Based on the standardized deformation and time curve, calculate the improved tangent angle corresponding to each monitoring time. , To standardize the increment of the time parameter, (This refers to the actual monitoring time interval), and combined with the improved tangent angle change trend, the accelerated deformation stage of the landslide is further identified, and the accelerated deformation stage is divided into the initial acceleration stage ( ), mid-acceleration phase ( ) and the skid (acceleration) phase ( The system generates landslide trend warning results, with yellow warnings for the initial acceleration stage, orange warnings for the intermediate acceleration stage, and red warnings for the imminent (accelerating) landslide stage. For the uniform deformation stage, the improved tangent angle is approximately 45°, and a green normal monitoring status is output. This green status is uniformly mapped to level 0 (normal monitoring, no alarm triggered) in subsequent comprehensive judgments. If the judgment result is the initial deformation stage, the improved tangent angle calculation and refined identification in this step are skipped, and the trend warning level is internally marked as "no level (initial, stable stage)". This mark only represents that the current deformation is in the initial stage and the rate is lower than the uniform speed benchmark, and it is mapped to the basic state of level 0 in subsequent comprehensive judgments. S4.4. Based on the cumulative displacement of monitoring points, calculate the current cumulative displacement value of each monitoring point and the number of monitoring points exceeding the preset displacement threshold, as well as the proportion of the total number of monitoring points. Then, determine the threshold according to the preset alarm rules and generate a displacement threshold alarm result: Prioritize Level 4 alarms: If, within a defined short period, at least one-third of the monitoring points have a cumulative displacement exceeding 100mm (the short period is a fixed physical time window of 24 hours prior to the current monitoring time; the sampling period is only used to determine whether the window contains at least 12 valid monitoring sampling points. If the monitoring frequency is less than once per 2 hours, the window is equivalently extended to 48 hours to ensure a statistical sample size of at least 6, but the displacement rate benchmark is still converted and compared based on 24 hours to ensure the statistical meaning is consistent). (i) Directly classified as a Level 4 alarm; where "not less than one-third" is calculated by multiplying the total number of monitoring points by 1 / 3 and rounding up; if the Level 4 alarm condition is not met, it is classified as a Level 3 alarm: if the cumulative displacement of not less than one-third of the monitoring points exceeds 50mm, it is classified as a Level 3 alarm; if the Level 3 condition is not met, it is classified as a Level 2 alarm: if the cumulative displacement of not less than one-third of the monitoring points exceeds 30mm, it is classified as a Level 2 alarm; if none of the above higher-level conditions (Level 4, Level 3, Level 2) are met, and there is any monitoring point (i.e., the number of points exceeding 20mm is greater than 0) with a cumulative displacement exceeding 20mm, it is classified as a Level 1 alarm; if the cumulative displacement of all monitoring points does not exceed 20mm (i.e., the number of points exceeding 20mm is 0), it is classified as Level 0 (no alarm); S4.5. Based on the comprehensive landslide trend warning results and displacement threshold alarm results, a comprehensive judgment is made on the current landslide hazard level, and a comprehensive monitoring and early warning result is generated. Specifically, a mapping table between the trend warning and displacement alarm levels is first established for quantitative comparison (green (uniform speed) is mapped to level 0; "no level (initial-stable stage)" is mapped to level 0 (basic stable state)). Blue alert level is mapped to level 0.5 (internal reference only, no official warning issued); yellow alert level is mapped to level 2; orange alert level is mapped to level 3; red alert level is mapped to level 4), unifying the two to the same dimension; if S4.3 outputs "no level (initial, stable stage)" due to the initial deformation stage and without triggering tangent angle analysis, then the final comprehensive level is determined by the displacement threshold alarm level (i.e., max(displacement threshold alarm level, 0)). In this case, level 0 physically represents "initial slow deformation stage, no external intervention required"; when S4.3 outputs a valid official level (green (level 0), yellow, orange, or red), max(trend warning mapping level, displacement threshold alarm level) is taken as the comprehensive level; if the highest level ultimately adopted comes from the trend warning, then the trend warning level is used as the primary warning. Based on this, a "level deviation prompt" is added to the early warning report (e.g., a low displacement alarm but a prominent trend, indicating a potential acceleration risk). When the final level originates from the displacement threshold alarm level, the displacement threshold alarm level is used as the primary response basis, and the corresponding alarm measures are immediately activated. When the highest levels of the two are the same, the comprehensive monitoring and early warning results are directly issued according to the corresponding level. When the deformation stage is determined to be "initial deformation stage" but the displacement threshold alarm level reaches level 2 or above, it indicates that the cumulative displacement has accumulated to a significant level over a long period. At this time, the system automatically corrects the stage identification to "mid-to-late stage of uniform deformation" and records the "stage review and correction" log in the early warning report. At the same time, the comprehensive level is set to max(displacement level, 0.5), and a "stage correction prompt" is issued to the public to avoid confusion in on-site decision-making due to a serious disconnect between the stage name and the alarm level. The final output comprehensive monitoring and early warning results include at least: landslide deformation stage, trend warning level, displacement threshold alarm level, and corresponding handling suggestions. Based on the comprehensive level, the results are issued through SMS notifications, platform alarms, on-site audible and visual alarms, and emergency responses.

[0021] Example 2: This example provides a geological disaster monitoring and early warning system based on big data, 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 geological disaster monitoring and early warning method based on big data 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 geological disaster monitoring and early warning method based on big data, characterized in that, include: S1. Collect real-time monitoring data and transmit it to the edge computing node to generate a multi-source time-series monitoring dataset for landslides; S2. Based on edge computing, the landslide displacement time series characteristics of each monitoring point are calculated using the multi-source time series monitoring dataset of landslides. Crack propagation characteristics, slope tilt characteristics and rainfall process characteristics are extracted. The deformation acceleration and slope tilt angle change rate corresponding to each monitoring time are used to analyze the cooperative change relationship, generate the landslide dynamic response coupling factor corresponding to each monitoring point, and generate the landslide deformation and time response relationship. S3. Construct a dynamic evolution constraint domain for landslides based on the relationship between landslide deformation and time response, generate a constraint evolution sample set, and output the membership degree of the landslide deformation stage based on the constraint evolution sample set using a long short-term memory network. S4. Determine the landslide deformation stage based on the membership degree, and conduct landslide trend early warning analysis and displacement threshold alarm analysis in combination with the cumulative displacement of monitoring points, and generate comprehensive landslide monitoring and early warning results.

2. The geological disaster monitoring and early warning method based on big data according to claim 1, characterized in that, In step S1, real-time monitoring data is collected and transmitted to edge computing nodes to generate a multi-source time-series monitoring dataset for landslides, including the following steps: Real-time data collection is performed on landslide surface displacement, crack displacement, slope tilt, and rainfall data according to a preset sampling period. The data is then transmitted to edge computing nodes via a wireless communication network. The edge computing nodes perform time synchronization processing on various monitoring data based on a unified time reference, and correlate and fuse various monitoring data according to monitoring time and spatial location of monitoring points to generate a multi-source time-series monitoring dataset for landslides.

3. The geological disaster monitoring and early warning method based on big data according to claim 1, characterized in that, In step S2, the landslide displacement time-series characteristics of each monitoring point are calculated, crack propagation characteristics, slope tilt characteristics, and rainfall process characteristics are extracted, and the deformation acceleration and slope tilt angle change rate corresponding to each monitoring time are used to analyze the coordinated change relationship, generate the landslide dynamic response coupling factor corresponding to each monitoring point, and generate the landslide deformation and time response relationship, including the following steps: S2.1 Based on edge computing nodes, read the landslide surface displacement data from the multi-source time-series monitoring dataset of landslides according to the monitoring points, and generate landslide displacement time-series features; S2.

2. Generate crack propagation features based on crack displacement data, generate slope tilt features based on slope tilt data, and generate rainfall process features based on rainfall data; S2.

3. Using the deformation acceleration and slope tilt angle change rate corresponding to each monitoring time, analyze the synergistic change relationship between the two in the continuous monitoring period, and generate the landslide dynamic response coupling factor corresponding to each monitoring point. S2.

4. The landslide displacement time series characteristics, crack propagation characteristics, slope tilt characteristics, rainfall process characteristics, and landslide dynamic response coupling factors are fused according to a unified monitoring time axis to construct the landslide dynamic evolution characteristic matrix corresponding to each monitoring point. S2.

5. Based on the unified monitoring time axis, perform time-series correlation analysis on the landslide dynamic evolution characteristic matrix to generate the relationship between landslide deformation and time response.

4. The geological disaster monitoring and early warning method based on big data according to claim 3, characterized in that, In step S2.3, the deformation acceleration and slope tilt angle change rate at each monitoring time are used to analyze their coordinated change relationship within the continuous monitoring period, generating the landslide dynamic response coupling factor for each monitoring point. This includes the following steps: S2.

31. Using the current monitoring time as the analysis endpoint, search in reverse along the time axis for the starting time when the deformation acceleration first shows a continuous increasing trend, and take the time interval between the starting time and the current monitoring time as the effective collaborative analysis period. Simultaneously extract the corresponding slope tilt change rate time series within the effective collaborative analysis period to construct a response-driven sliding window. S2.

32. Based on the variation characteristics of the deformation acceleration time series within the response-driven sliding time window, the evolution process of deformation acceleration is identified, and the effective collaborative analysis period is divided into acceleration segment, constant speed segment and deceleration segment in sequence. S2.

33. Based on the time synchronization relationship, the time series of slope inclination change rate is segmented accordingly to form the deformation acceleration subsequence and slope inclination change rate subsequence corresponding to each dynamic stage. S2.

34. Based on the effective collaborative analysis period division results, combined with the deformation acceleration subsequence and the slope tilt change rate subsequence, a weighted trend collaborative index is generated. The weighted trend collaborative index is then normalized to generate the landslide dynamic response coupling factor.

5. The geological disaster monitoring and early warning method based on big data according to claim 4, characterized in that, In step S2.34, generating the weighted trend coordination index includes the following steps: The direction of change of the deformation acceleration subsequence and the slope tilt change rate subsequence between adjacent monitoring times in the acceleration, constant speed, and deceleration phases are calculated respectively. When the direction of change is the same, it is determined to be a state of coordinated change. The time interval of continuous coordinated change in each phase is counted, and the proportion of coordinated change time to the total duration of the corresponding phase is calculated. The coordination degree of acceleration phase, constant speed phase, and deceleration phase are obtained respectively. The weighted trend coordination index corresponding to each monitoring point is generated by weighted fusion.

6. The geological disaster monitoring and early warning method based on big data according to claim 1, characterized in that, In step S3, a dynamic evolution constraint domain for landslides is constructed based on the relationship between landslide deformation and time response, a constraint evolution sample set is generated, and the membership degree of the landslide deformation stage is output using a long short-term memory network based on the constraint evolution sample set, including the following steps: S3.1 Receive the relationship between landslide deformation and time response, retrieve the historical monitoring records of the corresponding area in the historical landslide monitoring big data, perform time registration and feature standardization processing on the current landslide deformation and time response relationship and historical monitoring data according to the unified time axis, and generate a landslide evolution feature sequence between the current monitoring object and historical landslide samples; S3.2 Construct a dynamic evolution constraint domain for landslides based on the relationship between landslide deformation and time response, and perform evolution constraint screening on historical landslide monitoring big data based on the dynamic evolution constraint domain to generate a constraint evolution sample set; S3.3 Using the constrained evolution sample set as the training reference sample, and combining it with the landslide dynamic evolution feature matrix corresponding to the current landslide, construct the stage evolution feature vector corresponding to the current landslide. S3.4 Input the stage evolution feature vector into the pre-trained long short-term memory network, and use the evolution law of the corresponding landslide deformation stage in the constrained evolution sample set to constrain and correct the current landslide stage identification result, and generate the membership degree of the current landslide belonging to the initial deformation stage, uniform deformation stage and accelerated deformation stage.

7. The geological disaster monitoring and early warning method based on big data according to claim 6, characterized in that, In step S3.2, a dynamic evolution constraint domain for landslides is constructed based on the relationship between landslide deformation and time response. Then, historical landslide monitoring big data is used to filter evolution constraints based on this domain to generate a constraint evolution sample set. This includes the following steps: S3.

21. Based on the relationship between landslide deformation and time response, and construct a multidimensional monitoring feature sequence according to a unified time axis; use a nonlinear multidimensional projection method to map each monitoring feature sequence to a unified evolution space, and introduce a local neighborhood preservation metric to generate state nodes and state transition trajectories corresponding to each monitoring time, and construct the dynamic evolution state space corresponding to the current landslide. S3.

22. Based on the temporal connection relationship and trajectory continuous change law between each state node in the dynamic evolution state space, analyze the continuous reachability relationship between any adjacent state nodes and identify the state boundary in the state transition process. S3.

23. Using a time-sharing recursive construction mode, calculate the evolution constraint strength corresponding to each state boundary to generate a landslide evolution constraint domain that describes the normal evolution range of the landslide. S3.

24. Based on the landslide evolution constraint domain, perform point-by-point positive progressive verification on the current state nodes to form a constraint evolution sample set.

8. The geological disaster monitoring and early warning method based on big data according to claim 7, characterized in that, In step S3.23, a time-sharing recursive construction mode is adopted to calculate the evolution constraint strength corresponding to each state boundary, and generate a landslide evolution constraint domain describing the normal evolution range of the landslide, including the following steps: Based on the evolution states in the dynamic evolution feature matrix of landslides, they are mapped as state nodes, and adjacent state nodes are connected in chronological order to form local state evolution paths. The connection area between adjacent state nodes is defined as the state boundary, and incremental adaptive density clustering is used to statistically analyze the frequency of discrete partitioning. The changes in Mahalanobis distance between the state nodes on both sides of each state boundary in the original high-dimensional feature space are calculated respectively; the changes in state transition probability between adjacent discrete states are calculated based on the discrete partition transition frequency; and the degree of continuous change of the dynamic response coupling factor of the landslide on both sides of the state boundary is calculated by combining the relationship between landslide deformation and time response, and a dynamic response continuity index is generated. The statistical soft constraint strength corresponding to each state boundary is obtained by normalizing the changes in Mahalanobis distance, the changes in state transition probability, and the continuity index of dynamic response, and then performing weighted fusion calculations to generate the conditional evolution constraint domain.

9. The geological disaster monitoring and early warning method based on big data according to claim 1, characterized in that, In step S4, the landslide deformation stage is determined based on the membership degree, and landslide trend early warning analysis and displacement threshold alarm analysis are carried out in combination with the cumulative displacement of monitoring points. The landslide monitoring and early warning results are then generated comprehensively, including the following steps: S4.1 Extract the maximum membership degree based on the membership degree and perform stage confidence determination to determine the deformation stage to which the current landslide belongs; S4.2 Based on the landslide deformation stage and the corresponding benchmark deformation rate, the cumulative displacement at each monitoring time is normalized to the time dimension to obtain standardized time parameters with the same dimension as the monitoring time, and a standardized deformation-time curve is generated. S4.3 Based on the standardized deformation and time curves, calculate the improved tangent angle corresponding to each monitoring time, and combine the trend of the improved tangent angle to refine the identification of the accelerated deformation stage of the landslide. Divide the accelerated deformation stage into the initial acceleration stage, the intermediate acceleration stage and the imminent landslide stage, and generate landslide trend early warning results. S4.

4. Based on the cumulative displacement of the monitoring points, calculate the current cumulative displacement value of each monitoring point and the number of monitoring points that exceed the preset displacement threshold, as well as the proportion of the total number of monitoring points. Then, determine the threshold according to the preset alarm rules and generate a displacement threshold alarm result. S4.

5. Based on the comprehensive landslide trend warning results and displacement threshold alarm results, a comprehensive assessment of the current landslide hazard level is made, and a comprehensive monitoring and early warning result is generated.

10. A geological disaster monitoring and early warning system based on big data, 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 geological disaster monitoring and early warning method based on big data as described in any one of claims 1-9.

Images