Radar slope deformation monitoring method and system based on dynamic intelligent reference trace point

Abstract

The invention discloses a radar slope deformation monitoring method and system based on a dynamic intelligent reference trace point, and relates to the technical field of geological monitoring, and the method comprises the steps: deploying an FMCW radar to collect slope sparse trace point data; an initial intelligent reference trace point set is established through parameter adaptive clustering, and systematic drift caused by environmental factors is dynamically identified and compensated by using a reference evolution model, so that adaptive updating of the reference is realized; performing double comparative analysis on each frame of new data, and respectively extracting accumulated displacement time sequence characteristics and instantaneous change characteristics; and synthesizing the two types of features through a multi-modal fusion decision model, calculating a deformation risk score and outputting graded early warning. According to the method, the problems of serious reference drift, high false alarm rate and difficulty in giving consideration to slow deformation and sudden instability in long-term monitoring of a traditional method are solved by specially aiming at the sparse trace point characteristic of a low-cost and low-data-rate radar, and intelligent slope deformation monitoring and early warning with high reliability, self-adaption and full-spectrum response are realized.

Description

Technical Field

[0001] This invention relates to the field of geological monitoring technology, and more specifically, to a radar slope deformation monitoring method and system based on dynamic intelligent reference points. Background Technology

[0002] Slope deformation monitoring is a crucial means of preventing geological disasters such as landslides and collapses. Currently, commonly used slope deformation monitoring methods mainly include traditional surveying methods, Global Navigation Satellite System (GNSS) monitoring, spaceborne or ground-based Synthetic Aperture Radar (SAR) technology, and the increasingly applied Frequency Modulated Continuous Wave (FMCW) radar monitoring technology. Traditional surveying methods, such as total stations, while offering high accuracy, rely on manual field operations, resulting in low efficiency, long cycles, and high costs, and are difficult to implement for continuous automated monitoring. While GNSS monitoring technology can achieve continuous monitoring, the sparse deployment of monitoring points and limited spatial resolution make it unable to capture subtle local deformation characteristics of slopes. Although spaceborne or ground-based SAR technology can achieve large-area deformation monitoring, the equipment is expensive, deployment is complex, and revisit cycles are long, making it difficult to meet the needs of cost-effective and flexible engineering monitoring.

[0003] In recent years, low-cost, miniaturized FMCW radars have been explored for slope deformation monitoring due to their ease of deployment and networking capabilities. However, these radars typically have low output data rates, acquiring only tens to hundreds of sparse data points per frame, making it difficult to effectively extract micro-deformation information using traditional dense data point analysis methods. Furthermore, existing monitoring methods often rely on fixed reference points or static thresholds for deformation assessment. During long-term monitoring, these methods are significantly affected by systematic measurement benchmark drift caused by non-deformation factors such as temperature changes and vegetation growth, leading to increased false alarm rates and hindering highly reliable and intelligent deformation monitoring. Therefore, there is an urgent need for a monitoring method that can specifically process sparse radar data, possess adaptive benchmark update capabilities, and achieve highly reliable deformation monitoring and early warning. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a radar slope deformation monitoring method and system based on dynamic intelligent reference points. This method solves the problems of severe reference drift, high false alarm rate, and difficulty in simultaneously addressing slow deformation and sudden instability in traditional methods during long-term monitoring. It achieves highly reliable, adaptive, and full-spectrum-response intelligent slope deformation monitoring and early warning.

[0005] The first aspect of this invention provides a radar slope deformation monitoring method based on dynamic intelligent reference points, comprising the following steps: FMCW radar sensor nodes are deployed at stable benchmark points opposite or to the side of the monitoring area to collect sparse point data of the slope area at a preset cycle and upload them to the intelligent processing module simultaneously. When the intelligent processing module receives historical point data for a preset duration, it triggers the initial reference point establishment process, uses a parameter-adaptive density clustering algorithm to select spatially stable points, and forms and stores the initial reference point set. During continuous monitoring, the systemic drift caused by non-deformation factors is identified by the reference baseline evolution model. When the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated. For each newly acquired point data frame, the new point is matched and the difference is calculated with the current reference point set to generate cumulative displacement time series features. The inter-frame change between the new point and the point in the previous frame or multiple previous frames is calculated in parallel to generate instantaneous change features. The cumulative displacement time-series features and instantaneous change features are input into the pre-trained multimodal fusion decision model to calculate the comprehensive deformation risk score, and the corresponding warning level is output according to the preset threshold range, and the warning information is pushed to the user terminal.

[0006] In this solution, the initial reference point establishment process includes: Within a preset historical time period, radar spot data, environmental monitoring data, and radar equipment status data are collected synchronously; the radar spot data, environmental monitoring data, and radar equipment status data at each moment are fused and aligned through spatiotemporal indexing to form a multimodal spot data unit. Using all historical data units as nodes, an edge structure connecting the nodes is dynamically constructed based on spatiotemporal proximity, generating a graph structure that reflects the spatiotemporal relationship of the data points. The node features include spatial coordinates, echo intensity, environmental parameters, and device status. The graph structure is input into a spatiotemporal graph neural network model, and the neighborhood information of nodes is aggregated through a graph attention mechanism. The stability representation vector of each node is learned, and cluster analysis is performed on the stability representation vector. Based on the clustering results, a set of candidate stable nodes is obtained. Positive and negative sample pairs are constructed within the candidate stable point set. The encoder is trained through contrastive learning so that the positive sample pairs are represented closer and the negative sample pairs are represented further apart. Based on the trained model, a stability confidence score is calculated for each candidate point, and points with a confidence score lower than the preset threshold are removed. For the selected high-confidence points, calculate the statistical characteristics of the historical location distribution, use the distribution mean as the calibration position of the reference points, use the distribution standard deviation as the initial location uncertainty, and output the initial reference point set with confidence weights and uncertainty ranges.

[0007] In this scheme, node neighborhood information is aggregated through a graph attention mechanism to learn the stability representation vector of each node. Cluster analysis is then performed on these stability representation vectors, and a candidate stable node set is obtained based on the clustering results, including: In the spatiotemporal graph neural network model, a multi-head attention mechanism is used to aggregate the information of each node and its spatiotemporal neighbors. After several layers of iteration, a high-dimensional stable representation vector of each node is output and then dimensionality reduction is performed. Calculate the distance between the dimensionality-reduced stability representation vectors, construct a distance matrix, adaptively determine the neighborhood radius and minimum number of points based on the statistical distribution of the distance matrix, and perform density clustering on all stability representation vectors based on the neighborhood radius and minimum number of points. During the clustering process, nodes that contain at least a minimum number of vectors within their neighborhood radius are marked as core points, and clusters are generated by expanding from the core points. Nodes corresponding to vectors that are not assigned to any cluster are marked as noise points. Select the clusters with a size greater than a preset threshold and the smallest average distance between nodes within the cluster from the clustering results, and output the original points corresponding to the nodes within the cluster as a candidate stable point set.

[0008] In this scheme, during continuous monitoring, systematic drift caused by non-deformation factors is identified by referencing a baseline evolution model, including: During continuous monitoring, the real-time input points are matched with the current set of reference points, the inter-frame displacement difference of each reference point is calculated, a displacement difference time sequence is generated, and each displacement difference data point is associated with a multi-dimensional environmental state vector corresponding to the acquisition time. The environmental state vector includes temperature, humidity, time period information and equipment status parameters. The displacement difference time series and associated environmental state vectors are used to construct direct observation terms and environment-driven terms, and feature representations are extracted from the observation vectors of the past several periods to construct historical memory terms. The multivariate system observation vector is constructed using the direct observation term, environment-driven term, and historical memory term. Using a reference baseline evolution model based on a multivariate state estimator, the predicted observation vector of the reference point at the current time is predicted based on the estimated state at the previous time step and the current environmental input change. The predicted observation vector is compared with the multivariate system observation vector to generate a residual vector. The residual vector is then projected onto a preset drift feature direction, which includes: temperature-sensitive drift direction, humidity-sensitive drift direction, seasonal periodic drift direction, equipment aging drift direction, and potential deformation direction. Analyze the projection intensity and persistence of the residual vector onto each drift eigenvector, and identify non-deformable and asymptotic system drift based on the analysis results.

[0009] In this scheme, the projection intensity and persistence of the residual vector onto each drift eigenvector are analyzed. If all preset conditions are met, the current system change is determined to be a non-deformable system drift. The preset conditions include: In the most recent K consecutive monitoring cycles, the absolute values ​​corresponding to the projection value sequence of the residual vector in at least one of the drift feature vector directions are obtained, and the absolute values ​​are all greater than the preset activation threshold in the current direction; The absolute value of the linear correlation coefficient between the projected value sequence and the reference time series representing the changes in the physical driving factors corresponding to the predefined characteristics is greater than the preset correlation threshold. Within the same continuous monitoring period, a sequence of projection values ​​of the residual vector in the potential deformation direction is obtained, and the absolute values ​​of all of them are less than the preset deformation threshold in the current direction.

[0010] In this scheme, when the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated, including: The drift-related environmental drift state components are extracted from the updated estimated state. Each component represents the magnitude of the quantitative drift effect of a specific environmental factor on the monitoring benchmark at the current moment. The components are estimated online based on the projection of the residual vector onto the corresponding drift feature vector. The sensitivity vector is calibrated by considering the slope material, slope aspect, and slant range and incident angle of the radar at the location of the reference point, and then associated with the reference point. The environmental drift state component and the sensitivity vector are linearly synthesized to obtain a three-dimensional coordinate compensation vector for each reference point trace to offset the currently estimated non-deformable system drift. The three-dimensional coordinate compensation vector is then applied to the corresponding reference point trace coordinates to generate an updated reference point trace set.

[0011] In this scheme, a long-term deformation analysis engine is used to match and calculate the differences between the new point traces and the current reference point trace set, generating cumulative displacement time-series features, including: Based on the multidimensional feature descriptor of the point trace, the best matching point is selected in the reference point trace set for the new point trace in the current frame, and the spatial geometric consistency is used for verification to obtain reliable tracking point trace pairs; For each reliable tracking point pair, calculate the displacement difference between the current coordinates and the reference coordinates, and subtract the real-time environmental drift state component provided by the reference baseline evolution model from the displacement difference to generate the net deformation displacement characterizing the deformation trend. The net deformation displacement is added to the displacement history sequence of the corresponding point in chronological order to form and maintain the cumulative displacement temporal feature sequence of each point. The cumulative displacement time-series feature sequence of each point, environmental time-series data and time label are input into the pre-trained time-series prediction model. The time-series prediction model is used to learn the displacement pattern under normal environmental fluctuations and output the predicted value. The standardized cognitive bias between the current actual net deformation displacement and the predicted value is calculated. For each tracking point, extract the standardized cognitive bias value at the current moment, the directional consistency index of the standardized cognitive bias within a preset continuous time window, and the cumulative statistical characteristics of the net deformation displacement of the point within a preset long period, and then fuse them. The fused multi-scale deformation features are input into the classifier, and the output point traces are used as the probability values ​​of abnormal cumulative deformation, which serve as the cumulative displacement time-series features characterizing their long-term deformation risk.

[0012] In this scheme, a real-time mutation detection engine is used to calculate the inter-frame change between the new data point and the data points in the previous frame or multiple previous frames, generating instantaneous change features, including: Based on the point data of the current frame and several previous frames, a multi-channel spatiotemporal feature field is constructed, and the unit-by-unit difference between the feature field at the current moment and the feature field at the previous moment is calculated to generate a multi-channel inter-frame difference feature map that includes distance changes, point number changes and intensity changes. The inter-frame difference feature map is input into a pre-trained convolutional autoencoder. The autoencoder learns to reconstruct the inter-frame change pattern during the historical normal monitoring period. By calculating the overall reconstruction error between the input feature map and the autoencoder output reconstruction map, a transient mutation intensity value is generated. Based on the spatial distribution of the reconstruction error, local abnormal regions are identified. For each local abnormal region, the region centroid displacement and region point cloud distribution change parameters are calculated based on the internal point data. The instantaneous mutation intensity value, the spatial location information of the local abnormal region, and the regional centroid displacement and distribution change parameters are collectively encapsulated and output as the instantaneous change feature.

[0013] In this scheme, a pre-trained multimodal fusion decision model is used to calculate a comprehensive deformation risk score, and the corresponding early warning level is output based on a preset threshold range, including: The long-term deformation characteristics and instantaneous change characteristics from each monitoring area are spliced ​​together and fused into the feature vector of each area. Spatial location encoding and time period encoding are added to construct a feature sequence input fusion decision model. By integrating the multi-head attention mechanism in the decision-making model, the feature sequence is processed to achieve interaction and information fusion between different regions and between long-term and short-term features, and to extract global features that represent the overall risk status. Multiple decoupled risk component scores are calculated using a single decoder, including: slow cumulative deformation risk score, sudden instability risk score, spatial diffusion risk score, and model decision confidence score. The slow cumulative deformation risk score, sudden instability risk score, and spatial diffusion risk score are input into a nonlinear fusion network, and the model decision confidence score is used as a weight adjustment factor to calculate a comprehensive deformation risk score. Based on the spatial diffusion risk score and the model decision confidence score, the preset warning threshold range is dynamically adjusted to determine the final warning level; at the same time, based on the global features, a semantic warning description text corresponding to the current risk status is generated. The output includes the final warning information, which includes the comprehensive deformation risk score, warning level, main risk area identifiers, and semantic warning description text.

[0014] The second aspect of the present invention provides a radar slope deformation monitoring system based on dynamic intelligent reference points, the system comprising: a front-end sensing module, an intelligent processing module, and a user terminal module; The front-end sensing module includes one or more FMCW radar sensor nodes deployed at stable benchmark points, which collect sparse point data of the slope area at a preset cycle and upload it synchronously to the intelligent processing module. The intelligent processing module includes a reference point management unit, a dual comparison analysis unit, and a fusion decision and early warning unit. The reference point management unit uses historical point data and employs a parameter-adaptive density clustering algorithm to select spatially stable points, forming and storing an initial reference point set. During continuous monitoring, it identifies systematic drift caused by non-deformation factors through a reference baseline evolution model. When the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated. The dual comparison analysis unit performs two types of analysis in parallel: using a long-term deformation analysis engine to match and calculate the differences between new points and the current reference point set, generating cumulative displacement time-series features; and using a real-time mutation detection engine to calculate the inter-frame change between new points and points in the previous frame or multiple previous frames, generating instantaneous change features. The fusion decision and early warning unit receives the cumulative displacement time-series features and instantaneous change features, calculates a comprehensive deformation risk score, and outputs the corresponding early warning level based on a preset threshold range. The user terminal module receives the warning information generated by the intelligent processing module through the communication network and displays the warning information visually.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention stably extracts micro-deformation signals from low-data-rate information, thereby achieving high-precision automated monitoring while significantly reducing hardware and deployment costs. It provides a practical technical path for large-scale, universally applicable slope safety monitoring. By referencing a baseline evolution model, the system can not only establish an initial stable baseline but also continuously learn and understand long-term, gradual drift patterns caused by environmental factors, and intelligently update the analysis baseline accordingly. This effectively alleviates the problems of baseline drift and accumulated errors in long-term monitoring, greatly improving the long-term stability and reliability of the system.

[0016] This method achieves comprehensive and sensitive capture of the entire spectrum of slope instability risk through dual comparative analysis and multimodal fusion decision-making. The long-term deformation analysis engine focuses on identifying slowly accumulating deformation trends from noise, while the real-time abrupt change detection engine remains highly sensitive to short-term displacement abrupt changes. Both operate in parallel, and the fusion decision model performs high-level information integration and confidence assessment, enabling the system to provide early warnings for slowly developing potential slippage and to respond quickly to sudden local collapses or instability. This significantly broadens the coverage of monitoring and early warning, enhancing the system's ability to cope with complex risk scenarios.

[0017] Furthermore, this invention can automatically adapt to differences in geological conditions, climate characteristics, and slope structures, exhibiting excellent universality. The multi-layered information package, including a comprehensive deformation risk score, graded early warning levels, location of major risk areas, and interpretable semantic early warning descriptions, closely aligns with the decision-making needs of engineering practice and emergency management. It provides maintenance units with clear risk assessment criteria and precise action guidelines, enabling more scientific and efficient resource allocation and the implementation of preventative or remedial measures, ultimately improving the level of intelligent slope safety management. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0019] Figure 1 A flowchart of a radar slope deformation monitoring method based on dynamic intelligent reference point traces is shown. Figure 2 A flowchart of the initial reference point establishment process is shown; Figure 3 The flowchart illustrates the reference baseline evolution model's identification of drift caused by non-deformation factors. Figure 4 A block diagram of a radar slope deformation monitoring system based on dynamic intelligent reference point traces is shown. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] like Figure 1 As shown, the first embodiment of the present invention provides a radar slope deformation monitoring method based on dynamic intelligent reference point traces, including: FMCW radar sensor nodes are deployed at stable benchmark points opposite or to the side of the monitoring area to collect sparse point data of the slope area at a preset cycle and upload them to the intelligent processing module simultaneously. When the intelligent processing module receives historical point data for a preset duration, it triggers the initial reference point establishment process, uses a parameter-adaptive density clustering algorithm to select spatially stable points, and forms and stores the initial reference point set. During continuous monitoring, the systemic drift caused by non-deformation factors is identified by the reference baseline evolution model. When the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated. For each newly acquired point data frame, the new point is matched and the difference is calculated with the current reference point set to generate cumulative displacement time series features. The inter-frame change between the new point and the point in the previous frame or multiple previous frames is calculated in parallel to generate instantaneous change features. The cumulative displacement time-series features and instantaneous change features are input into the pre-trained multimodal fusion decision model to calculate the comprehensive deformation risk score, and the corresponding warning level is output according to the preset threshold range, and the warning information is pushed to the user terminal.

[0023] It should be noted that after the system starts, the radar sensor continuously collects sparse point data of the slope area at fixed intervals, including information such as distance, angle, and echo intensity. Within a preset time period, in addition to accumulating the original point data collected by the radar, it also simultaneously collects meteorological data and radar equipment status data that are spatiotemporally aligned with the point data. The meteorological data includes temperature, humidity, wind speed, and rainfall data, while the radar equipment status data includes transmit power, receive gain, and system noise figure. Through spatiotemporal indexing, the radar point data, environmental monitoring data, and radar equipment status data at each moment are fused and aligned to form a multimodal point data unit. like Figure 2As shown, using all historical trace data units as nodes, an edge structure connecting the nodes is dynamically constructed based on spatiotemporal proximity, generating a graph structure reflecting the spatiotemporal relationship of the traces. Node features include spatial coordinates, echo intensity, environmental parameters, and device status. The edge structure includes temporal edges and spatial neighborhood edges. Temporal edges connect trace nodes in adjacent time frames within the same spatial region, while spatial neighborhood edges connect trace nodes within the same time frame whose spatial distance is less than a threshold. This graph structure is input into a spatiotemporal graph neural network model, which aggregates node neighborhood information through a graph attention mechanism. Nodes adaptively focus on neighbors with similar features while reducing the weight of neighbors affected by abnormal environments or noise during information aggregation. After multiple layers of graph convolution and attention aggregation, a high-dimensional stability representation vector for each trace node is output.

[0024] Cluster analysis is performed on the dimensionality-reduced stability representation vectors to calculate the distances between them, constructing a distance matrix. The neighborhood radius and minimum number of points are adaptively determined based on the statistical distribution of this distance matrix. The neighborhood radius is... Based on the mean of the effective distance values ​​in the distance matrix with standard deviation Combining the dimension and distribution density of the representation vector, and using a preset adjustment coefficient... The calculation yields the following result, which is expressed as: ;Minimum points Based on the total number of nodes Spatial distribution density estimate of the characterization vector It performs adaptive calculations proportionally and sets a minimum parameter lower bound. , represented as Density clustering is performed on all stability characterization vectors based on the neighborhood radius and minimum number of points. During the clustering process, nodes that contain at least a minimum number of vectors within their neighborhood radius are marked as core points, and clusters are generated by expanding from the core points. Nodes corresponding to vectors that are not assigned to any cluster are marked as noise points. The cluster with a size greater than a preset threshold and the smallest average distance between nodes within the cluster is selected, and the original points corresponding to the nodes within the cluster are output as a set of candidate stable points.

[0025] Within the candidate stable point trace set, positive and negative sample pairs are constructed. Positive sample pairs are two point trace nodes randomly selected from the same candidate cluster, observed under similar time and environmental conditions. Negative sample pairs are candidate stable point trace nodes paired with randomly selected, significantly disturbed point trace nodes. An encoder is trained through contrastive learning using a normalized temperature-scaled cross-entropy loss function. Backpropagation optimizes the encoder parameters to make positive sample pairs converge and negative sample pairs diverge. The encoder is then used to perform forward propagation on all candidate stable point trace nodes to obtain the final contrastive representation vector. For each candidate node, its representation similarity to the cluster center node is calculated, and the similarity is normalized to a stability confidence score. Point traces with confidence scores below a preset threshold are removed. For the selected high-confidence point traces, the statistical characteristics of their historical location distribution are calculated. The distribution mean is used as the calibration position of the reference point trace, and the distribution standard deviation is used as the initial location uncertainty. An initial reference point trace set with confidence weights and uncertainty ranges is output.

[0026] It should be noted that during continuous monitoring, the systematic drift caused by non-deformation factors is identified by referencing the baseline evolution model, which solves the problem of systematic and gradual measurement baseline drift caused by non-deformation environmental factors such as temperature changes, material aging, and vegetation growth in long-term monitoring.

[0027] like Figure 3 As shown, during continuous monitoring, real-time input data points are matched with the current set of reference data points. The three-dimensional displacement difference between each matching pair of reference and real-time data points is calculated, generating a displacement difference time series. Each displacement difference data point is associated with a multi-dimensional environmental state vector corresponding to the acquisition time. The environmental state vector includes temperature, humidity, time period information, and equipment status parameters. The displacement difference time series and the associated environmental state vector are used to construct a direct observation item. Environment-driven items And from the observation vectors of the past several periods Extract feature representations and construct historical memory items. .

[0028] Using the direct observation term Environment-driven items and historical memory items Constructing observation vectors for multivariate systems Based on the observation vector of a multivariate system, the overall dynamic behavior of the reference point track set is modeled as an implicit dynamic system driven by environmental input, and the system state... This characterizes the combined state of the current stable baseline and various systematic drifts. A reference baseline evolution model based on a multivariate state estimator is used, based on the estimated state from the previous time step. and changes in current environmental input Predict the predicted observation vector of the reference point at the current time. ; The predicted observation vector With multivariate system observation vector Compare and generate residual vectors The residual vector is projected onto a preset drift feature direction, including: temperature-sensitive drift direction, humidity-sensitive drift direction, seasonal cycle drift direction, equipment aging drift direction, and potential deformation direction; the drift feature basis vector is learned through historical data, and each drift feature direction corresponds to a specific systematic drift physical mode, including but not limited to: thermal expansion / contraction mode, humidity expansion and contraction mode, seasonal cycle mode, and equipment performance gradual change mode.

[0029] Analyze the projection intensity and persistence of the residual vector on each drift feature vector. If all preset conditions are met, the current system change is determined to be a non-deformable system drift. The preset conditions include: (1) In the most recent K consecutive monitoring periods, obtain the absolute value of the projection value sequence of the residual vector in at least one of the drift feature vector directions. The absolute value is greater than the preset activation threshold of the current direction. (2) The absolute value of the linear correlation coefficient between the projection value sequence and the reference time sequence of the predefined physical driving factor change is greater than the preset correlation threshold. (3) In the same consecutive monitoring period, obtain the projection value sequence of the residual vector in the potential deformation direction. The absolute value is less than the preset deformation threshold of the current direction.

[0030] A baseline update is triggered only when the multivariate state estimator determines that the system change is mainly caused by one or more modeled environmental drift state components, and the estimated values ​​of the potential deformation state components remain within a stable range. The system state estimation vector includes the ideal stable baseline state component, the intensity scalar of the environmental drift mode, and the potential deformation state component. Drift-related environmental drift state components are extracted from the updated estimated state. Each component represents the magnitude of the quantized drift effect of a specific environmental factor on the monitoring baseline at the current moment, estimated online based on the projection of the residual vector onto the corresponding drift feature vector. The sensitivity vector is calibrated using the slope material, aspect, and slant range and incident angle of the radar at the reference point, and then associated with the reference point. The environmental drift state components and the sensitivity vector are linearly synthesized to obtain the three-dimensional coordinate compensation vector for each reference point used to offset the currently estimated non-deformable system drift. , For sensitivity vectors, For environmental drift mode, This represents the total number of environmental drift patterns. For the intensity scalar of the environmental drift state component, Given the unit influence direction vector of the first drift mode in the radar coordinate system, the three-dimensional coordinate compensation vector is applied to the corresponding reference point coordinates to generate an updated reference point set.

[0031] In the dual-comparison analysis unit, a long-term deformation analysis engine is used to match and calculate the differences between new traces and the current reference trace set, generating cumulative displacement temporal features. Based on the multi-dimensional feature descriptor of the traces, the best matching point is selected for the new trace in the current frame from the reference trace set. The descriptor at least includes the coordinates, echo intensity, and distribution statistics of the trace within its local spatial neighborhood. Verification is performed using spatial geometric consistency to obtain reliable tracking trace pairs. The preferred rigid constraint is that the change in spatial distance between matching pairs should be consistent with the overall apparent displacement trend of the region; the topological constraint is that matching should not cause a drastic distortion in the relative topological relationship between the point and its neighboring points.

[0032] For each reliable tracking point pair, the displacement difference between the current coordinates and the reference coordinates is calculated, and the real-time environmental drift state component provided by the reference baseline evolution model is subtracted from the displacement difference to generate the net deformation displacement characterizing the deformation trend. The net deformation displacement is added to the displacement history sequence of the corresponding point in chronological order to form and maintain the cumulative displacement time series feature sequence of each point. The cumulative displacement time series feature sequence of each point, environmental time series data, and time labels are input into a pre-trained time series prediction model. The preferred time series prediction model adopts a recurrent neural network enhanced by an attention mechanism. The time series prediction model is used to learn the displacement pattern under normal environmental fluctuations and output the predicted value to calculate the current actual net deformation displacement. With the predicted value Standardized cognitive bias between , This represents the standard deviation of the predicted values.

[0033] For each tracking point, the standardized cognitive bias value at the current moment, the directional consistency index of the standardized cognitive bias within a preset continuous time window (checking whether the biases of the most recent K observations show directional consistency), and the cumulative statistical characteristics of the point's net deformation displacement within a preset long period (calculating the vector sum of the point's net deformation displacement within the preset long period or the cumulative projected value in a certain main sliding direction) are extracted and fused. The fused multi-scale deformation features are input into a classifier, which outputs the probability value of the point's abnormal cumulative deformation, as the cumulative displacement time-series feature characterizing its long-term deformation risk. Preferably, a grid is defined on the slope space, and the abnormal probability and cumulative displacement vector of all points in the same area are statistically aggregated. The long-term deformation analysis engine outputs the point-level features such as the abnormal probability of each tracking point, and the regional-level features such as the average deformation probability, the resultant displacement vector, and the spatial coordination index.

[0034] In the dual comparison analysis unit, a real-time mutation detection engine is used to capture sudden local instability in a short period of time, calculate the inter-frame change between the new point and the point in the previous frame or multiple previous frames, generate instantaneous change features, and detect abnormal inter-frame changes that do not conform to the recent normal motion pattern.

[0035] Based on the point data of the current frame and several previous frames, a sliding spatiotemporal window is constructed, mapping the points within the window onto a three-dimensional spatiotemporal voxel grid. Two dimensions represent the radar's spatial coordinates, and the third dimension represents time. Statistical features of the points within each voxel are calculated to form a multi-channel spatiotemporal feature field corresponding to the current moment. The unit-by-unit difference between the current and previous feature fields is calculated, generating a multi-channel inter-frame difference feature map containing changes in distance, number of points, and intensity. This inter-frame difference feature map is input into a pre-trained convolutional autoencoder, which learns to reconstruct inter-frame change patterns within historical normal monitoring periods. In real-time detection, the overall reconstruction error between the input feature map and the autoencoder's output reconstruction map is calculated. The error is then spatially weighted and averaged based on the historical stability of each spatial region to obtain an instantaneous abrupt change intensity value characterizing the degree of current inter-frame change anomaly. Regions with more stable historical data have higher weights for the current reconstruction error. Based on the spatial distribution of the reconstruction error, local anomalous regions with significantly higher mutation intensity than the surrounding area are identified. For each local anomalous region, the region centroid displacement and region point cloud distribution change parameters are calculated based on the internal point trace data. The instantaneous mutation intensity value, the spatial location information of the local anomalous region, and the region centroid displacement and distribution change parameters are encapsulated and output as the instantaneous change feature.

[0036] It should be noted that a lightweight Transformer model with spatial-temporal awareness is introduced to construct a multimodal fusion decision model. Using a pre-trained multimodal fusion decision model, a comprehensive deformation risk score is calculated, and the corresponding warning level is output based on a preset threshold range. Each spatial monitoring area is used as a basic unit to generate a regional feature token. Long-term deformation features and instantaneous change features from each monitoring area are concatenated and fused into a feature vector for each area. A global context token is added at the beginning of the entire sequence to aggregate the overall state. Spatial location encoding and temporal period encoding are added to each regional feature token to construct a feature sequence input lightweight Transformer encoder. The feature sequence is processed through the multi-head attention mechanism in the lightweight Transformer encoder corresponding to the fusion decision model, realizing interaction and information fusion between different areas and between long-term and short-term features. Specific regional feature tokens interact to evaluate the spatial synergy or isolation of deformation patterns; long-term deformation features in regional feature tokens interact cross-modally with instantaneous change features from other areas to assess the correlation risk between long-term trends and short-term abrupt changes; the global context token aggregates all regional information through the attention mechanism to form a comprehensive representation of the overall slope risk status.

[0037] Based on the global features of the global context token output by the Transformer encoder, multiple decoupled risk component scores are calculated through a decoding head consisting of four parallel sub-networks, each responsible for calculating an independent risk component score. Each sub-network has the same structure but independent parameters, and is a two-layer feedforward neural network, where the first layer is fully connected and non-linearly activated. The second layer is a fully connected layer with normalized output: , This is the hidden feature vector output by the first activation function. As a global feature, This is the weight matrix of the first fully connected layer. This is the bias vector of the first fully connected layer. The output risk component score, This is the weight matrix of the second fully connected layer. This is the bias vector for the second fully connected layer. For the Sigmoid function, This includes: slow cumulative deformation risk score (mainly based on the time series characteristics of cumulative displacement), sudden instability risk score (mainly based on the instantaneous change characteristics), spatial diffusion risk score (based on the spatial clustering and correlation of risk areas), and model decision confidence score (based on the consistency between features and the uncertainty of model prediction). During the model training phase, regression training is performed using severity labels of historical long-term deformation events confirmed by experts (for each confirmed long-term deformation event period, the final cumulative displacement of key areas of the slope within that period is extracted) to provide a supervisory signal for the slow-accumulating deformation risk score. Mean squared error loss is preferably used for regression training. Regression training is also performed using intensity labels of confirmed sudden deformation events (such as local collapse records) to provide a supervisory signal for the sudden instability risk score. Mean squared error loss is preferably used to penalize prediction errors at the event's peak. Regression training is also performed using the spatial diffusion index (combining historical monitoring data (multi-point displacement gauges, crack gauge data) and on-site investigation reports with expert judgment) of historical events where the impact range of deformation events expands or chain reactions occur in adjacent areas (e.g., cracks at the rear edge of a landslide continuously extend backward) to provide a supervisory signal for the spatial diffusion risk score. During training, noise is randomly added to the input features for the model decision confidence score. The goal of the confidence score is to predict the variance of the model's output risk under noise disturbance; the smaller the variance, the higher the confidence score, enabling the model to learn to assess the certainty of its own judgments.

[0038] The slow-accumulating deformation risk is divided into Sudden instability risk classification and spatial diffusion risk classification The input is a nonlinear fusion network. The three components are concatenated and then processed through a fully connected layer for initial interaction, resulting in a preliminary fused feature vector. The model decision confidence score is used as a weight adjustment factor to adjust the contribution of the fused features. Adjusted features The final fully connected layer is used to calculate the comprehensive deformation risk score using the Sigmoid activation function; where To fuse the fully connected weight matrix of the first layer of the network, To fuse the bias vectors of the first layer of the network, This is a learnable weight adjustment vector that maps scalar confidence scores to adjustment weights corresponding to each fused feature channel. Through training, the model learns which feature channels should be suppressed or enhanced at low confidence levels. When the model's decision confidence score is high, incremental adjustment is provided for each channel of the feature, fine-tuning the feature using the learned pattern. When the model's decision confidence score is low, the adjustment mechanism automatically disables, outputting a fused result without significant correction.

[0039] Four static thresholds T1 < T2 < T3 < T4 are preset, corresponding to four warning levels: stable (S < T1), attention (T1 ≤ S < T2), warning (T2 ≤ S < T3), alarm (S ≥ T3), and T4 is the extreme alarm threshold. According to the spatial diffusion risk score and the model decision confidence score , calculate the threshold offset , apply the offset to the intermediate thresholds T2 and T3, dynamically adjust the preset warning threshold interval, and determine the final warning level, where is the preset sensitivity coefficient; perform threshold dynamic adjustment according to the spatial attributes of the current risk and the self-awareness level of the model, making the warning more intelligent and more in line with the actual situation. At the same time, based on the global features, generate a semantic warning description text corresponding to the current risk situation; output the final warning information including the comprehensive deformation risk score, warning level, main risk area identifier, and semantic warning description text.

[0040] As Figure 4 shown, in the second aspect of the present invention, a radar slope deformation monitoring system based on dynamic intelligent reference point traces is provided, and the system includes: a front-end perception module, an intelligent processing module, and a user terminal module; The front-end perception module includes one or more FMCW radar sensor nodes deployed at stable reference points, collects sparse point trace data of the slope area at a preset period, and synchronously uploads it to the intelligent processing module; The intelligent processing module includes a reference point trace management unit, a dual comparison and analysis unit, and a fusion decision and warning unit; after using the historical point trace data, the reference point trace management unit uses a density clustering algorithm with parameter self-adaptation to screen out the point traces with stable spatial positions, forms and stores the initial reference point trace set; during continuous monitoring, it identifies systematic drifts caused by non-deformation factors through a reference benchmark evolution model, and automatically updates the reference point trace set when it is determined that the drift is progressive and non-deformation caused; the dual comparison and analysis unit executes two types of analysis in parallel, uses a long-term deformation analysis engine to match and calculate the differences between the new point traces and the current reference point trace set, generates cumulative displacement time series features, and uses a real-time mutation detection engine to calculate the inter-frame change amounts between the new point traces and the previous frame or multiple previous frames of point traces, generates instantaneous change features; the fusion decision and warning unit receives the cumulative displacement time series features and instantaneous change features, calculates the comprehensive deformation risk score, and outputs the corresponding warning level according to the preset threshold interval; The user terminal module receives the warning information generated by the intelligent processing module through the communication network, and visualizes the warning information through the slope warning APP.

[0041] The third embodiment of the present invention provides a computer-readable storage medium, which includes a radar slope deformation monitoring method program based on dynamic intelligent reference points. When the radar slope deformation monitoring method program based on dynamic intelligent reference points is executed by a processor, it implements the steps of the radar slope deformation monitoring method based on dynamic intelligent reference points.

[0042] In the several embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms.

[0043] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0044] Alternatively, if the integrated units of the present invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0045] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A radar slope deformation monitoring method based on dynamic intelligent reference point traces, characterized in that, Includes the following steps: FMCW radar sensor nodes are deployed at stable benchmark points opposite or to the side of the monitoring area to collect sparse point data of the slope area at a preset cycle and upload them to the intelligent processing module simultaneously. When the intelligent processing module receives historical point data for a preset duration, it triggers the initial reference point establishment process, uses a parameter-adaptive density clustering algorithm to select spatially stable points, and forms and stores the initial reference point set. During continuous monitoring, the systemic drift caused by non-deformation factors is identified by the reference baseline evolution model. When the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated. For each newly acquired point data frame, the new point is matched and the difference is calculated with the current reference point set to generate cumulative displacement time series features. The inter-frame change between the new point and the point in the previous frame or multiple previous frames is calculated in parallel to generate instantaneous change features. The cumulative displacement time-series features and instantaneous change features are input into the pre-trained multimodal fusion decision model to calculate the comprehensive deformation risk score, and the corresponding warning level is output according to the preset threshold range, and the warning information is pushed to the user terminal.

2. The radar slope deformation monitoring method based on dynamic intelligent reference points according to claim 1, characterized in that, The initial reference point establishment process includes: Within a preset historical time period, radar spot data, environmental monitoring data, and radar equipment status data are collected synchronously; the radar spot data, environmental monitoring data, and radar equipment status data at each moment are fused and aligned through spatiotemporal indexing to form a multimodal spot data unit. Using all historical data units as nodes, an edge structure connecting the nodes is dynamically constructed based on spatiotemporal proximity, generating a graph structure that reflects the spatiotemporal relationship of the data points. The node features include spatial coordinates, echo intensity, environmental parameters, and device status. The graph structure is input into a spatiotemporal graph neural network model, and the neighborhood information of nodes is aggregated through a graph attention mechanism. The stability representation vector of each node is learned, and cluster analysis is performed on the stability representation vector. Based on the clustering results, a set of candidate stable nodes is obtained. Positive and negative sample pairs are constructed within the candidate stable point set. The encoder is trained through contrastive learning so that the positive sample pairs are represented closer and the negative sample pairs are represented further apart. Based on the trained model, a stability confidence score is calculated for each candidate point, and points with a confidence score lower than the preset threshold are removed. For the selected high-confidence points, calculate the statistical characteristics of the historical location distribution, use the distribution mean as the calibration position of the reference points, use the distribution standard deviation as the initial location uncertainty, and output the initial reference point set with confidence weights and uncertainty ranges.

3. The radar slope deformation monitoring method based on dynamic intelligent reference points according to claim 2, characterized in that, The system aggregates node neighborhood information using a graph attention mechanism, learns the stability representation vector for each node, performs cluster analysis on the stability representation vectors, and obtains a candidate stable node set based on the clustering results, including: In the spatiotemporal graph neural network model, a multi-head attention mechanism is used to aggregate the information of each node and its spatiotemporal neighbors. After several layers of iteration, a high-dimensional stable representation vector of each node is output and then dimensionality reduction is performed. Calculate the distance between the dimensionality-reduced stability representation vectors, construct a distance matrix, adaptively determine the neighborhood radius and minimum number of points based on the statistical distribution of the distance matrix, and perform density clustering on all stability representation vectors based on the neighborhood radius and minimum number of points. During the clustering process, nodes that contain at least a minimum number of vectors within their neighborhood radius are marked as core points, and clusters are generated by expanding from the core points. Nodes corresponding to vectors that are not assigned to any cluster are marked as noise points. Select the clusters with a size greater than a preset threshold and the smallest average distance between nodes within the cluster from the clustering results, and output the original points corresponding to the nodes within the cluster as a candidate stable point set.

4. The radar slope deformation monitoring method based on dynamic intelligent reference point traces according to claim 1, characterized in that, During continuous monitoring, systematic drift caused by non-deformation factors is identified by referencing a baseline evolution model, including: During continuous monitoring, the real-time input points are matched with the current set of reference points, the inter-frame displacement difference of each reference point is calculated, a displacement difference time sequence is generated, and each displacement difference data point is associated with a multi-dimensional environmental state vector corresponding to the acquisition time. The environmental state vector includes temperature, humidity, time period information and equipment status parameters. The displacement difference time series and associated environmental state vectors are used to construct direct observation terms and environment-driven terms, and feature representations are extracted from the observation vectors of the past several periods to construct historical memory terms. The multivariate system observation vector is constructed using the direct observation term, environment-driven term, and historical memory term. Using a reference baseline evolution model based on a multivariate state estimator, the predicted observation vector of the reference point at the current time is predicted based on the estimated state at the previous time step and the current environmental input change. The predicted observation vector is compared with the multivariate system observation vector to generate a residual vector. The residual vector is then projected onto a preset drift feature direction, which includes: temperature-sensitive drift direction, humidity-sensitive drift direction, seasonal periodic drift direction, equipment aging drift direction, and potential deformation direction. Analyze the projection intensity and persistence of the residual vector onto each drift eigenvector, and identify non-deformable and asymptotic system drift based on the analysis results.

5. The radar slope deformation monitoring method based on dynamic intelligent reference point traces according to claim 4, characterized in that, Analyze the projection intensity and persistence of the residual vector onto each drift eigenvector. If all preset conditions are met, the current system change is determined to be a non-deformable system drift. The preset conditions include: In the most recent K consecutive monitoring cycles, the absolute values ​​corresponding to the projection value sequence of the residual vector in at least one of the drift feature vector directions are obtained, and the absolute values ​​are all greater than the preset activation threshold in the current direction; The absolute value of the linear correlation coefficient between the projected value sequence and the reference time series representing the changes in the physical driving factors corresponding to the predefined characteristics is greater than the preset correlation threshold. Within the same continuous monitoring period, a sequence of projection values ​​of the residual vector in the potential deformation direction is obtained, and the absolute values ​​of all of them are less than the preset deformation threshold in the current direction.

6. The radar slope deformation monitoring method based on dynamic intelligent reference point traces according to claim 4, characterized in that, When the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated, including: The drift-related environmental drift state components are extracted from the updated estimated state. Each component represents the magnitude of the quantitative drift effect of a specific environmental factor on the monitoring benchmark at the current moment. The components are estimated online based on the projection of the residual vector onto the corresponding drift feature vector. The sensitivity vector is calibrated by considering the slope material, slope aspect, and slant range and incident angle of the radar at the location of the reference point, and then associated with the reference point. The environmental drift state component and the sensitivity vector are linearly synthesized to obtain a three-dimensional coordinate compensation vector for each reference point trace to offset the currently estimated non-deformable system drift. The three-dimensional coordinate compensation vector is then applied to the corresponding reference point trace coordinates to generate an updated reference point trace set.

7. The radar slope deformation monitoring method based on dynamic intelligent reference point traces according to claim 1, characterized in that, The long-term deformation analysis engine is used to match and calculate the differences between the new point traces and the current reference point trace set, generating cumulative displacement time-series features, including: Based on the multidimensional feature descriptor of the point trace, the best matching point is selected in the reference point trace set for the new point trace in the current frame, and the spatial geometric consistency is used for verification to obtain reliable tracking point trace pairs; For each reliable tracking point pair, calculate the displacement difference between the current coordinates and the reference coordinates, and subtract the real-time environmental drift state component provided by the reference baseline evolution model from the displacement difference to generate the net deformation displacement characterizing the deformation trend. The net deformation displacement is added to the displacement history sequence of the corresponding point in chronological order to form and maintain the cumulative displacement temporal feature sequence of each point. The cumulative displacement time-series feature sequence of each point, environmental time-series data and time label are input into the pre-trained time-series prediction model. The time-series prediction model is used to learn the displacement pattern under normal environmental fluctuations and output the predicted value. The standardized cognitive bias between the current actual net deformation displacement and the predicted value is calculated. For each tracking point, the standardized cognitive bias value at the current moment, the directional consistency index of the standardized cognitive bias within a preset continuous time window, and the cumulative statistical characteristics of the net deformation displacement of the point within a preset long period are extracted and fused. The fused multi-scale deformation features are input into the classifier, and the output point traces are used as the probability values ​​of abnormal cumulative deformation, which serve as the cumulative displacement time-series features characterizing their long-term deformation risk.

8. The radar slope deformation monitoring method based on dynamic intelligent reference point traces according to claim 1, characterized in that, A real-time mutation detection engine is used to calculate the inter-frame changes between new data points and data points from the previous frame or multiple previous frames, generating instantaneous change features, including... Based on the point data of the current frame and several previous frames, a multi-channel spatiotemporal feature field is constructed, and the unit-by-unit difference between the feature field at the current moment and the feature field at the previous moment is calculated to generate a multi-channel inter-frame difference feature map that includes distance changes, point number changes and intensity changes. The inter-frame difference feature map is input into a pre-trained convolutional autoencoder, which learns to reconstruct the inter-frame change pattern during the historical normal monitoring period. By calculating the overall reconstruction error between the input feature map and the autoencoder output reconstruction map, a transient mutation intensity value is generated. Based on the spatial distribution of the reconstruction error, local abnormal regions are identified. For each local abnormal region, the region centroid displacement and region point cloud distribution change parameters are calculated based on the internal point data. The instantaneous mutation intensity value, the spatial location information of the local abnormal region, and the regional centroid displacement and distribution change parameters are collectively encapsulated and output as the instantaneous change feature.

9. The radar slope deformation monitoring method based on dynamic intelligent reference point traces according to claim 1, characterized in that, A pre-trained multimodal fusion decision model is used to calculate a comprehensive deformation risk score, and the corresponding warning level is output based on a preset threshold range, including: The long-term deformation characteristics and instantaneous change characteristics from each monitoring area are spliced ​​together and fused into the feature vector of each area. Spatial location encoding and time period encoding are added to construct a feature sequence input fusion decision model. By integrating the multi-head attention mechanism in the decision-making model, the feature sequence is processed to achieve interaction and information fusion between different regions and between long-term and short-term features, and to extract global features that represent the overall risk status. Multiple decoupled risk component scores are calculated using a single decoder, including: slow cumulative deformation risk score, sudden instability risk score, spatial diffusion risk score, and model decision confidence score. The slow cumulative deformation risk score, sudden instability risk score, and spatial diffusion risk score are input into a nonlinear fusion network, and the model decision confidence score is used as a weight adjustment factor to calculate a comprehensive deformation risk score. Based on the spatial diffusion risk score and the model decision confidence score, the preset warning threshold range is dynamically adjusted to determine the final warning level; at the same time, based on the global features, a semantic warning description text corresponding to the current risk status is generated. The output includes the final warning information, which includes the comprehensive deformation risk score, warning level, main risk area identifiers, and semantic warning description text.

10. A radar slope deformation monitoring system based on dynamic intelligent reference point traces, characterized in that, To implement the radar slope deformation monitoring method based on dynamic intelligent reference points as described in any one of claims 1-9, the system includes: a front-end sensing module, an intelligent processing module, and a user terminal module. The front-end sensing module includes one or more FMCW radar sensor nodes deployed at stable benchmark points, which collect sparse point data of the slope area at a preset cycle and upload it synchronously to the intelligent processing module. The intelligent processing module includes a reference point management unit, a dual comparison analysis unit, and a fusion decision and early warning unit. The reference point management unit uses historical point data and employs a parameter-adaptive density clustering algorithm to select spatially stable points, forming and storing an initial reference point set. During continuous monitoring, it identifies systematic drift caused by non-deformation factors through a reference baseline evolution model. When the drift is determined to be gradual and not caused by deformation, the reference point set is automatically updated. The dual comparison analysis unit performs two types of analysis in parallel: using a long-term deformation analysis engine to match and calculate the differences between new points and the current reference point set, generating cumulative displacement time-series features; and using a real-time mutation detection engine to calculate the inter-frame change between new points and points in the previous frame or multiple previous frames, generating instantaneous change features. The fusion decision and early warning unit receives the cumulative displacement time-series features and instantaneous change features, calculates a comprehensive deformation risk score, and outputs the corresponding early warning level based on a preset threshold range. The user terminal module receives the warning information generated by the intelligent processing module through the communication network and displays the warning information visually.

Images