Debris flow geological disaster early warning and monitoring system

Abstract

The invention relates to the technical field of geological disaster monitoring and early warning, and discloses a debris flow geological disaster early warning and monitoring system, which comprises a data preprocessing module for collecting and preprocessing original monitoring data; extracting space information of the monitoring points, and calculating a space distance between the monitoring points; the weight distribution module is used for carrying out feature learning and weight distribution of multi-modal features; the feature fusion module is used for extracting statistical features and time sequence features and carrying out feature fusion; the risk assessment module performs disaster situation intelligent perception and risk quantitative assessment; the propagation analysis module is used for analyzing a space cooperation relationship between the monitoring points and a disaster propagation rule; the early warning processing module generates an early warning decision and processes the early warning decision in real time; the feedback optimization module is used for performing verification feedback, identifying problem monitoring points and performing optimization; according to the invention, intelligent fusion and adaptive weight distribution of multi-source heterogeneous data are realized by fusing a multi-modal deep learning network architecture of a debris flow physical evolution mechanism.

Description

A debris flow geological disaster early warning and monitoring system Technical field

[0001] The present invention relates to the technical field of geological disaster monitoring and early warning, and more specifically, it relates to a debris flow geological disaster early warning and monitoring system. Background technique

[0002] As a common sudden geological disaster in mountainous areas, debris flow geological disasters are characterized by high occurrence frequency, strong destructive power, and great difficulty in early warning, posing a serious threat to the traffic infrastructure in mountainous areas and the safety of people's lives and property. With the continuous advancement of mountainous area development and construction, the importance of debris flow disaster monitoring and early warning technology has become increasingly prominent. Traditional debris flow monitoring and early warning technologies mainly collect data based on single-type sensors, such as relying solely on rainfall monitoring or surface displacement monitoring. This single-index monitoring method is difficult to comprehensively reflect the complexity of the debris flow gestation process and the characteristics of multi-factor coupling. In recent years, multi-sensor fusion monitoring technology has gradually emerged. By deploying multiple types of devices such as rainfall sensors, soil moisture sensors, surface displacement sensors, and groundwater pressure sensors, it attempts to monitor the occurrence and development process of debris flow disasters from multiple dimensions.

[0003] However, the existing multi-sensor fusion technology faces many technical bottlenecks in practical applications. There are certain differences in the data formats, sampling frequencies, and spatio-temporal benchmarks of different types of sensors, and there is a lack of unified data quality assessment standards and intelligent repair mechanisms, resulting in serious information loss during the data fusion process and difficult to guarantee the fusion accuracy; the spatial collaborative early warning ability is insufficient. Existing technologies mostly adopt a single-point early warning mode for independent monitoring points, lacking in-depth analysis of the spatial correlation relationships between monitoring points, unable to effectively identify the spatial propagation path and influence range of disasters, and difficult to achieve regional collaborative early warning and multi-point linkage response; the early warning threshold setting mostly adopts fixed thresholds or simple statistical thresholds, lacking consideration of the geological environment differences of different monitoring points and the personalized characteristics of disaster evolution laws, which easily leads to insufficient adaptability of early warning.

[0004] There are obvious deficiencies in the existing technology in terms of real-time guarantee. The traditional centralized data processing architecture collects all monitoring data to the central server for unified analysis and processing. In cases such as network transmission delay, peak data processing load, and communication link failure, it is difficult to meet the strict timeliness requirements of debris flow disaster second-level early warning response. At the same time, the existing early warning system lacks an effective early warning effect evaluation and feedback optimization mechanism, and is unable to dynamically adjust and continuously optimize system parameters according to the actual early warning effect, resulting in the difficulty of continuously improving the early warning performance of the system with the accumulation of operation experience.

[0005] These technical issues directly affect the accuracy, timeliness, and reliability of debris flow disaster early warning. Under complex and ever-changing geological conditions, missed or false reports are prone to occur, and the advance warning time is insufficient, which seriously restricts the actual effect of debris flow disaster prevention and control. There is an urgent need to develop more intelligent, collaborative, and real-time debris flow geological disaster early warning and monitoring technologies. Summary of the Invention

[0006] This invention provides a debris flow geological disaster early warning and monitoring system, which solves the technical problems of low accuracy of debris flow early warning, difficulty in multi-source data fusion, and insufficient spatial collaborative analysis capability in related technologies.

[0007] This invention provides a debris flow geological disaster early warning and monitoring system, comprising:

[0008] The data preprocessing module collects raw monitoring data, performs preprocessing, and obtains a high-quality multi-source monitoring dataset; it extracts spatial information of monitoring points, calculates the spatial distance between monitoring points, and obtains the spatial topology network structure of monitoring points.

[0009] The weight allocation module, based on a high-quality multi-source monitoring dataset, performs feature learning and multi-modal feature weight allocation to obtain an intelligent indicator weight set.

[0010] The feature fusion module extracts statistical features and time-series features based on an intelligent indicator weight set and a high-quality multi-source monitoring dataset, and then performs feature fusion to obtain a deep time-series feature vector.

[0011] The risk assessment module, based on deep temporal feature vectors and the spatial topology network structure of monitoring points, performs intelligent perception of disaster situation and quantitative risk assessment, and obtains the risk assessment results of each monitoring point;

[0012] The propagation analysis module analyzes the spatial coordination relationship and disaster propagation patterns among monitoring points based on the risk assessment results of each monitoring point, and obtains the regional propagation assessment results;

[0013] The early warning processing module generates early warning decisions and processes them in real time based on the risk assessment results of each monitoring point and the regional propagation assessment results, thus obtaining intelligent early warning processing results.

[0014] The feedback optimization module verifies and provides feedback on the intelligent early warning processing results, identifies problem monitoring points and optimizes them to obtain an optimized configuration scheme.

[0015] In a preferred embodiment, the step of collecting and preprocessing the raw monitoring data includes:

[0016] Perform integrity and rationality quality checks on the monitoring data, extract the timestamp field, data value field, sensor identification field and data type field from the received monitoring data records, and perform a data integrity verification process;

[0017] For numerical data fields, theoretical upper and lower limits are obtained based on the physical measurement principle and measurement range constraints of the sensor. A normal numerical distribution range is established using an outlier detection method based on statistical distribution. The current data value is compared with the theoretical boundary and the statistical distribution range. When the value exceeds the theoretical boundary or deviates from the statistical distribution range by more than a preset deviation threshold, the data is marked as an outlier.

[0018] In a preferred embodiment, the step of extracting spatial information of monitoring points and calculating the spatial distance between monitoring points includes:

[0019] The unified transformation of the spatiotemporal reference is performed, the spatial coordinate records and coordinate system type information of each monitoring point are extracted, and the coordinates under different coordinate systems are transformed to a unified spatial reference coordinate system using a coordinate system transformation algorithm;

[0020] Based on the location information of monitoring points under a unified coordinate system, the spatial distance between any two monitoring points is calculated, the digital elevation model data of the monitoring area is extracted to obtain the elevation information of each monitoring point, and the relative elevation difference between monitoring points is calculated.

[0021] By combining watershed distribution data and surface runoff path data of the monitoring area, the hydrological connectivity between monitoring points is analyzed, and a spatial topology network of monitoring points is constructed based on spatial distance, elevation difference and hydrological connectivity information.

[0022] In a preferred embodiment, the steps of performing feature learning and multimodal feature weight allocation include:

[0023] A multimodal deep learning network architecture integrating the physical evolution mechanism of debris flow is constructed. A multi-branch structure is adopted, with each branch corresponding to the physical evolution process of a type of monitoring index. The activation function and loss function of the branch network are designed to incorporate the constraints of the corresponding physical process.

[0024] Deep feature learning and adaptive weight optimization of multimodal monitoring data are performed. The intrinsic relationship between different monitoring indicators is learned through a self-attention mechanism. The feature representations of each monitoring indicator are linearly combined through learnable weight coefficients.

[0025] In a preferred embodiment, the steps of extracting statistical features and time-series features include:

[0026] Construct multi-scale time windows and extract basic statistical features. Set time window scales of short-term, medium-term and long-term. Calculate basic statistical feature parameters for the time series data of each monitoring indicator within each time window.

[0027] Specific time-series features of various monitoring indicators are extracted. For rainfall monitoring indicators, instantaneous rainfall intensity peak features and rainfall duration features are extracted. For surface displacement monitoring indicators, instantaneous displacement velocity features and displacement increment sequence features are extracted. For soil moisture monitoring indicators, moisture growth rate features and saturation change features are extracted.

[0028] In a preferred embodiment, the step of performing intelligent disaster situation perception includes:

[0029] A spatiotemporal graph neural network architecture and risk assessment model are constructed. Each monitoring point in the monitoring area is constructed as a graph structure, with the monitoring points as nodes of the graph. The node features are depth-series feature vectors. The edge connection relationships of the graph are determined based on the spatial topology network structure of the monitoring points.

[0030] A spatiotemporal graph neural network architecture is constructed, which includes a graph convolutional layer, a temporal attention layer, and a risk assessment layer. The graph convolutional layer aggregates neighbor node information through a message passing mechanism, and the temporal attention layer learns the importance weights of features at different time steps using a multi-head attention mechanism.

[0031] In a preferred embodiment, the risk quantification assessment step includes:

[0032] Adaptive threshold dynamic determination based on graph attention mechanism: The attention weight information in graph neural network is used to adaptively adjust the warning threshold.

[0033] The system can determine the current disaster risk level of each monitoring point in real time, compare the current risk probability value with the threshold boundary of the monitoring point, and determine the status as normal, attention, warning and danger respectively.

[0034] In a preferred embodiment, the step of analyzing the spatial coordination relationship between monitoring points and the disaster propagation pattern includes:

[0035] Perform spatiotemporal collaborative anomaly identification of upstream and downstream monitoring points, identify monitoring points with current risk status at warning or dangerous level as anomaly source nodes, find the downstream node set for each anomaly source node, and track the change trajectory of risk probability value of each downstream node in the time period after the anomaly trigger time.

[0036] Predict the disaster's propagation path, impact range, and arrival time. Calculate the estimated propagation time from the anomaly source node to each downstream node based on the expected propagation speed and total propagation path length.

[0037] In a preferred embodiment, the step of generating and processing early warning decisions in real time includes:

[0038] Establish multi-level early warning decision rules and early warning information generation, and construct an early warning decision rule library, including local early warning rules based on single-point risk status and regional early warning rules based on multi-point collaborative status. When the risk status level of a single monitoring point reaches the early warning state, a local early warning decision is triggered. When a preset number of spatially adjacent monitoring points simultaneously reach the early warning state and there is a disaster collaborative relationship or propagation path, a regional joint early warning decision is triggered.

[0039] Edge computing node devices are deployed at each monitoring point. The edge nodes perform data quality assessment, intelligent repair and spatiotemporal benchmark unification processing, call the simplified risk assessment model stored locally to calculate the short-term risk probability value, and activate the local audible and visual early warning device when the value exceeds the rapid response threshold.

[0040] In a preferred embodiment, the steps of verifying and feeding back the intelligent early warning processing results, identifying problem monitoring points, and optimizing them include:

[0041] Establish a database for recording and verifying early warning events, record detailed information of every early warning event triggered during system operation, set an early warning verification time window, continuously monitor the occurrence of actual disaster events within the early warning verification time window, and verify and label early warning events based on the comparison between early warning results and actual conditions;

[0042] Calculate the early warning accuracy rate, false alarm rate, and false alarm rate. When the early warning accuracy rate of a monitoring point is lower than the preset lower threshold, the false alarm rate is higher than the preset upper threshold, or the false alarm rate is higher than the preset upper threshold, the monitoring point is marked as a problem monitoring point.

[0043] Perform reverse optimization of model parameters for the problem monitoring points.

[0044] The beneficial effects of this invention are as follows:

[0045] By integrating a multimodal deep learning network architecture that incorporates the physical evolution mechanism of debris flows, intelligent fusion and adaptive weight allocation of multi-source heterogeneous data are achieved. Compared with traditional methods, the early warning accuracy is improved, effectively solving the problem of low data fusion accuracy.

[0046] A spatiotemporal graph neural network architecture is adopted for intelligent perception of disaster development. By learning the spatial correlation patterns and temporal evolution laws between monitoring points through graph convolutional layers and temporal attention mechanisms, adaptive threshold dynamic adjustment based on graph attention mechanism is realized, thereby reducing the false alarm rate.

[0047] This invention establishes an edge-cloud collaborative early warning processing architecture, enabling edge nodes to achieve rapid local response within seconds, while the cloud performs in-depth global analysis. This reduces the early warning response time from minutes to seconds, meeting the real-time requirements for rapid early warning of debris flow disasters.

[0048] This invention constructs a complete early warning effect evaluation feedback and adaptive optimization mechanism. Through continuous performance evaluation and parameter optimization, the system's early warning performance continuously improves over time, realizing the self-evolution capability of the early warning system.

[0049] This invention achieves a technological leap from single-point monitoring to regional collaboration, and from passive response to proactive prediction, providing reliable technical support for debris flow disaster prevention and control of important infrastructure such as mountain highways. Attached Figure Description

[0050] Figure 1 is a block diagram of the present invention;

[0051] Figure 2 is a detailed flowchart of the present invention. Detailed Implementation

[0052] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0053] At least one embodiment of the present invention discloses a debris flow geological disaster early warning and monitoring system, as shown in Figures 1 and 2, comprising the following steps:

[0054] The data preprocessing module collects raw monitoring data, performs preprocessing, and obtains a high-quality multi-source monitoring dataset; it extracts spatial information of monitoring points, calculates the spatial distance between monitoring points, and obtains the spatial topology network structure of monitoring points.

[0055] Step 1.1: Perform a quality check on the completeness and reasonableness of the monitoring data;

[0056] Based on various sensor devices deployed in debris flow-prone areas, including rainfall sensors, soil moisture sensors, surface displacement sensors, groundwater pressure sensors, and tilt sensors, the system receives raw monitoring data in real time from on-site sensors at each monitoring point. For each received monitoring data record, the system extracts its timestamp, data value, sensor identifier, and data type fields, and performs a data integrity verification process. By checking each field, the system determines if the timestamp is missing or formatted incorrectly, if the data value is null or invalid, if the sensor identifier matches the system's pre-defined sensor registration list, and if the data type identifier conforms to the specified definition. For the timestamp field, the system further checks the continuity of the time series by calculating the time interval between two adjacent data records to identify any time series breaks or time jumps, thus identifying discontinuous data segments. For the data value field, based on the physical measurement principles and measurement range constraints of this type of sensor, the theoretical upper and lower limits of the sensor's values ​​are obtained. An outlier detection method based on statistical distribution is used to extract historical normal data from this type of sensor at the monitoring point, calculate the mean and standard deviation of the historical data, and establish a normal value distribution interval. The current data value is compared with the theoretical boundary and statistical distribution range. When the value exceeds the theoretical boundary or deviates from the statistical distribution range by more than a set deviation threshold multiple, the data is marked as an outlier. Through the above multi-dimensional detection process, a quality status label is generated for each monitoring data record. The label types include five quality statuses: normal data, missing time series, missing values, abnormal values, and incomplete fields, resulting in the quality status labeling results.

[0057] In some embodiments, due to the possibility of transient sensor malfunctions under extreme environmental conditions leading to abnormal data fluctuations, relying solely on statistical distribution detection may misclassify normal extreme data as outliers. An anomaly detection method based on a joint judgment of physical constraints and temporal continuity can be employed to reduce the false positive rate and retain true extreme event data. Specifically, when performing statistical distribution-based anomaly detection, for data points marked as anomalous, the data change trend and rate of change before and after the data point are further analyzed. Normal data values ​​are extracted from the three sampling times before and after the anomalous data point, and the gradient and direction of change of these six times are calculated. It is determined whether the value of the anomalous data point lies on a reasonable extension path of the data change trend before and after. By fitting the change trend line of the data from the three previous times, the expected value range of the anomalous time is predicted, and the anomalous data point value is compared with the expected range. When the anomalous data point value deviates from the historical statistical distribution but conforms to the predicted change trend at the current time, the data is determined to be true extreme data rather than an anomalous value, its original value is retained, and the quality label is modified to normal data. By employing a time-series continuity verification mechanism, the system effectively distinguishes between sensor malfunctions and abnormalities and real-world extreme environmental changes, thereby improving the accuracy of anomaly detection.

[0058] Step 1.2, Intelligent repair of missing and abnormal data based on spatiotemporal correlation;

[0059] Based on the quality status labeling results generated in step 1.1, all data records labeled as time-series missing, numerically missing, or numerically abnormal are extracted, and an intelligent repair process is executed. For isolated single-point missing data in time series, i.e., cases where normal data exists before and after the missing data, a time series interpolation repair method is used. Normal data values ​​for this type of monitoring indicator at the monitoring point are extracted for the five sampling times before and after the missing time, constructing a data sequence for the preceding and following time periods. The temporal variation characteristics of this data sequence are analyzed, calculating the first-order difference sequence to represent the rate of change and the second-order difference sequence to represent the acceleration of change. Based on the rate and direction of change of the preceding and following data, it is determined whether the data evolution trend is linear, accelerating, or decelerating. For linear change patterns, linear interpolation is used to estimate the value at the missing time based on the uniform rate of change of the preceding and following data. For nonlinear change patterns, cubic spline interpolation is used to construct a cubic spline curve based on the preceding and following data points, and the function value of the spline curve at the missing time is used as the repaired value. For batch data loss in continuous periods, i.e., when the length of the missing period exceeds a set threshold, relying solely on temporal interpolation methods is insufficient to guarantee repair accuracy. In such cases, a spatial correlation repair method is employed. This involves extracting a set of neighboring monitoring points in the spatial topology network. Neighboring monitoring points are defined as those located within the same or adjacent watersheds, with a spatial distance less than a set spatial proximity threshold. Normal monitoring data from each neighboring monitoring point during the missing period are acquired, and the data association patterns between the monitoring point and its neighbors over historical periods are analyzed. Historical data from the past 30 days for both the monitoring point and its neighbors are extracted, and the Pearson correlation coefficient and time lag cross-correlation coefficient are calculated to quantify the spatial correlation strength and time response delay. For neighboring monitoring points with correlation coefficients greater than the set correlation threshold, their data are used as the repair reference source. Based on a data association model established over historical periods, which characterizes the numerical proportion and time lag relationship between the data from two monitoring points, the data from neighboring monitoring points during the missing period are mapped and transformed using this association model to generate an estimated data sequence for the monitoring point during the missing period. When multiple highly correlated neighboring monitoring points exist, estimated sequences are generated based on the data of each neighboring point, and then weighted and averaged using the correlation coefficient as the weighting coefficient. Through the above-mentioned repair strategy combining temporal interpolation and spatial mapping, intelligent repair processing for different missing patterns is achieved, resulting in a repaired integrity monitoring dataset.

[0060] In some embodiments, because some monitoring points may be located in terrain-isolated areas, the number of their spatially adjacent monitoring points may be insufficient or their correlation may be low, limiting the effectiveness of spatial correlation restoration methods. A restoration method based on historical contemporaneous data patterns can be adopted, aiming to estimate missing data using the historical patterns of the monitoring point itself. Specifically, for data with missing periods at specific time nodes, monitoring data from the same time nodes in historical years are extracted. For example, if the current missing data occurs between 10:00 and 12:00 on June 15th, historical data from the same time period on June 15th in the past three years are extracted. The numerical distribution characteristics and change patterns of the historical contemporaneous data are analyzed, and the mean, median, and quantile distributions of the historical contemporaneous data are calculated. Based on the data numerical levels before and after the current missing period, the degree of deviation of the overall data level in the current period relative to the historical contemporaneous period is determined, and a deviation ratio coefficient is calculated. The typical change patterns of the historical contemporaneous data are scaled and adjusted according to the deviation ratio coefficient to generate an estimated sequence adapted to the current data level, which serves as the restoration data for the missing period. This historical contemporaneous pattern reference method provides an effective data restoration approach for spatially isolated monitoring points.

[0061] Step 1.3, Unified conversion of spatiotemporal reference and construction of monitoring point topology network;

[0062] For the integrity monitoring dataset obtained in step 1.2, a unified processing procedure is performed on the time reference, sampling frequency, and spatial coordinates. Regarding time reference unification, the timestamp field of each monitoring data record is extracted. A timestamp format parsing algorithm is used to identify the time representation format. This algorithm identifies the date format, time precision level, and time zone identifier by matching the structural features of the timestamp string. For timestamps containing time zone information, their time zone offset is extracted. For timestamps without an explicitly marked time zone, their time zone is queried based on the geographical location of the monitoring point to which the sensor belongs. A time zone conversion algorithm is used to convert all timestamps to Coordinated Universal Time (UTC). During the conversion process, the time values ​​are adjusted by adding or subtracting based on the time zone offset. Furthermore, according to the standard time zone setting of the monitoring area, the UTC time is converted to the local standard time of the monitoring area, establishing a globally unified time reference system and eliminating time reference differences between different data sources. Regarding the standardization of sampling frequencies, the original sampling frequencies of various sensor types were statistically analyzed: rainfall sensors sampled once every 5 minutes, soil moisture sensors once every 10 minutes, surface displacement sensors once every 1 minute, groundwater pressure sensors once every 15 minutes, and tilt sensors once every 2 minutes. There were some differences in the sampling frequencies among these sensors. A unified target sampling frequency of once every 5 minutes was set, which meets the temporal resolution requirements for debris flow disaster monitoring while also considering data processing efficiency. For sensor data with original sampling frequencies higher than the target frequency, a sliding window statistical downsampling method was adopted. The time window length was set to 5 minutes, and each window contained multiple original sampling points. The statistical characteristic values ​​of all sampling points within the window were calculated as the downsampling result for that window. The statistical characteristic values ​​include the mean (representing the average level within the window), the maximum value (representing the peak value within the window), and the standard deviation (representing the degree of fluctuation within the window). For sensor data with an original sampling frequency lower than the target frequency, a cubic spline interpolation upsampling method is used. A cubic spline interpolation function is constructed based on existing low-frequency sampling points. This interpolation function generates a smooth, continuous curve between adjacent sampling points. The function value at the target sampling time is calculated as the interpolation result. Through downsampling and upsampling processing, all monitoring data are unified to the target sampling frequency, resulting in a time-series data set with consistent sampling frequencies.

[0063] Regarding spatial coordinate unification, the spatial coordinate records and coordinate system type information of each monitoring point are extracted. Monitoring point coordinates may be represented using different coordinate systems, including latitude and longitude representation in geographic coordinate systems and planar coordinate representation in projected coordinate systems. A coordinate system transformation algorithm is employed to identify the coordinate system type and reference datum of each coordinate record. Coordinates under different coordinate systems are transformed to a unified spatial reference coordinate system through a coordinate transformation model. This invention uses the WGS84 geographic coordinate system as the unified reference datum. The coordinate transformation process considers the ellipsoidal parameters, projection parameters, and datum transformation parameters of the coordinate system, achieving accurate transformation through mathematical transformation formulas. Based on the monitoring point location information under the unified coordinate system, the spatial distance between any two monitoring points is calculated, using a spherical distance calculation formula based on latitude and longitude coordinates to calculate the surface distance. Digital elevation model data of the monitoring area is extracted to obtain the elevation information of each monitoring point, and the relative elevation difference between monitoring points is calculated. Combining the watershed distribution data and surface runoff path data of the monitoring area, the hydrological connectivity between monitoring points is analyzed, identifying monitoring point pairs located in the same watershed or with upstream and downstream confluence relationships.

[0064] Based on spatial distance, elevation difference, and hydrological connectivity information, a spatial topology network of monitoring points is constructed. The network uses monitoring points as nodes and pairs of monitoring points with potential disaster propagation paths as directed edges. The direction of the edges points from upstream nodes with higher elevations to downstream nodes with lower elevations, and the weight of each edge is the spatial distance and elevation difference between the two nodes. Through the unified processing of the above three dimensions (time, frequency, and space), a high-quality multi-source monitoring dataset and a spatial topology network structure of monitoring points are obtained, featuring a unified spatiotemporal benchmark, consistent sampling frequency, and clearly defined spatial relationships.

[0065] The weight allocation module, based on a high-quality multi-source monitoring dataset, performs feature learning and multi-modal feature weight allocation to obtain an intelligent indicator weight set.

[0066] Step 2.1: Construct a multimodal deep learning network architecture that integrates the physical evolution mechanism of debris flows;

[0067] Based on the physical evolution mechanism of debris flow disasters, including key physical processes such as rainfall infiltration, soil saturation, slope instability, material initiation, and fluid movement, a physical constraint-guided deep neural network architecture is constructed. The network architecture adopts a multi-branch structure, with each branch corresponding to the physical evolution process of a class of monitoring indicators. The rainfall branch models the nonlinear response relationship between rainfall intensity and soil moisture content; the soil moisture branch models the soil saturation process and pore water pressure change mechanism; the surface displacement branch models the slope deformation and critical instability transition process; the groundwater pressure branch models the groundwater dynamic response and seepage field evolution process; and the tilt angle branch models the slope overturning moment and stability coefficient variation law. The activation function and loss function of each branch network incorporate the constraints of the corresponding physical process, ensuring that the network learning process conforms to the physical evolution law. Through the physical constraint loss term, the network must not only minimize the prediction error during training but also satisfy the physical consistency constraint, resulting in a deep learning network architecture that conforms to the physical mechanism of debris flow.

[0068] Step 2.2: Perform deep feature learning and adaptive weight optimization of the multimodal monitoring data;

[0069] The high-quality multi-source monitoring dataset output from step 1 is input into the multimodal deep learning network constructed in step 2.1 to perform a multimodal feature learning process. The network learns the intrinsic correlation between different monitoring indicators through a self-attention mechanism, and the self-attention weight matrix reflects the differences in the contribution of each indicator to the disaster gestation process.

[0070] For each monitoring point, historical disaster case data and corresponding monitoring data sequences are extracted to construct training sample pairs of monitoring data and disaster labels. These training samples are input into a deep network for end-to-end learning. The network automatically learns the optimal weight allocation for each monitoring indicator through backpropagation. The network's output layer adopts a weighted fusion structure, where the feature representations of each monitoring indicator are linearly combined using learnable weight coefficients. These weight coefficients are automatically optimized during training using gradient descent. A physical consistency regularization term is introduced to ensure that the learned weight allocation conforms to the physical evolution mechanism of debris flows. For example, rainfall indicators have higher weights in the early stages of rainfall, while displacement and tilt indicators have increased weights during the slope instability stage. Through multiple rounds of iterative training, the network converges to obtain an intelligent indicator weight allocation scheme for each monitoring point. This weight scheme integrates the advantages of data-driven learning and physical mechanism constraints.

[0071] Step 2.3, Global optimization and verification of weight parameters based on physical constraints;

[0072] Based on the initial weight allocation scheme output in step 2.2, a global optimization and cross-validation process is performed. A global optimization algorithm is used to jointly optimize the weight parameters of all monitoring points, ensuring that the weight allocation of the entire monitoring network reaches a globally optimal state. The optimization objective function comprehensively considers three dimensions: single-point early warning accuracy, multi-point collaborative consistency, and the rationality of the physical mechanism. The generalization ability of the weight parameters is evaluated using a cross-validation method. Historical data is divided into training and validation sets according to time series. Weight parameters are optimized on the training set, and early warning performance is evaluated on the validation set. For monitoring points whose validation performance is substandard, their weight parameters are readjusted and iteratively optimized. A physical constraint verification mechanism is introduced to verify whether the weight allocation conforms to the physical laws of debris flow evolution. Weight combinations that violate physical constraints are corrected. Finally, an intelligent indicator weight set that has undergone global optimization and physical constraint verification is output, ensuring that the weight allocation has both data-driven accuracy and conforms to the rationality of physical evolution.

[0073] The feature fusion module extracts statistical features and time-series features based on an intelligent indicator weight set and a high-quality multi-source monitoring dataset, and then performs feature fusion to obtain a deep time-series feature vector.

[0074] Step 3.1: Construct a multi-scale time window and extract basic statistical features;

[0075] Three time window scales were set: short-term, medium-term, and long-term. The short-term time window was set to 2 hours, used to capture short-term abrupt anomalies and rapid changes in the monitoring data. The medium-term time window was set to 24 hours, used to analyze the intraday evolution trend and periodic change characteristics of the monitoring data. The long-term time window was set to 7 days, used to assess the cumulative effect and long-term background state shift characteristics of the monitoring data. For the current analysis time, the monitoring data of the most recent 2 hours covered by the short-term window, the most recent 24 hours covered by the medium-term window, and the most recent 7 days covered by the long-term window were extracted to construct time-series data matrices for the three time scales. For each monitoring indicator, the basic statistical characteristic parameters were calculated for the time-series data within each time window. The mean was calculated as the average value of all sampling points within the window, representing the average level of the indicator within that time scale. The standard deviation was calculated as the degree of deviation of the sampling point values ​​within the window from the mean, representing the fluctuation range of the indicator within that time scale. The maximum and minimum values ​​were calculated as the extreme values ​​of the sampling point values ​​within the window, and the range was calculated as the difference between the maximum and minimum values, representing the range of change of the indicator within that time scale. The rate of change is calculated as the ratio of the numerical change within a window to the time span. By fitting a linear trend line to the time-series data, the slope of the trend line represents the average rate of change, characterizing how fast the indicator changes within that time scale. The acceleration of change is calculated as the time derivative of the rate of change. By fitting a curve representing the rate of change over time, the slope of the curve represents the acceleration of change, characterizing the trend of the indicator's rate of change accelerating or slowing down within that time scale. These statistical characteristics are calculated for each monitoring indicator at the three time scales, resulting in a set of basic statistical characteristics across multiple time scales, including mean, standard deviation, range, rate of change, and acceleration of change.

[0076] Step 3.2: Extract the specific time-series characteristics of various monitoring indicators;

[0077] To address the differences in the physical meaning and disaster indication mechanisms of different types of monitoring indicators, specific features are further extracted based on a multi-timescale set of basic statistical features. For rainfall monitoring indicators, peak instantaneous rainfall intensity features are extracted within short-term time windows to identify the maximum rainfall intensity and its occurrence time within the window. The duration of heavy rainfall is also extracted, and the cumulative duration of rainfall exceeding a set heavy rainfall threshold within the window is statistically analyzed. Within medium-term time windows, cumulative rainfall features are extracted, calculating the sum of rainfall at all sampling times within the window. The frequency of rainfall events is extracted, identifying the turning point where rainfall intensity rises from zero to a non-zero value, and statistically analyzing the number of rainfall events within the window. Within long-term time windows, accumulated rainfall features are extracted, calculating the total rainfall within the window to characterize the long-term cumulative effect of rainfall. The background soil saturation state is extracted, and the initial soil saturation level is determined by analyzing the baseline soil moisture level within the long-term window. For surface displacement monitoring indicators, instantaneous displacement velocity features are extracted within short-term time windows, and the ratio of displacement increment to time within the short-term window is calculated. Within a medium-term time window, displacement increment sequence features are extracted. The medium-term window is divided into several sub-periods, and the displacement increment of each sub-period is calculated. The change pattern of the displacement increment sequence is analyzed, and the acceleration phase and the uniform phase are identified. The acceleration phase is defined as the period in which the displacement increment shows an increasing trend, and the uniform phase is defined as the period in which the displacement increment remains stable. The duration of each phase is statistically analyzed. Within a long-term time window, cumulative displacement features are extracted, the total change in displacement within the long-term window is calculated, displacement benchmark drift features are extracted, and the systematic shift trend of the displacement benchmark value within the long-term window is analyzed. For soil moisture monitoring indicators, within a short-term time window, moisture growth rate features are extracted, and the rate of change of moisture values ​​within the short-term window is calculated. Within a medium-term time window, saturation change features are extracted, soil moisture values ​​are converted into saturation percentages, the process of saturation rising from the initial value to near saturation is analyzed, the duration of saturation is extracted, and the duration of saturation exceeding the set saturation threshold is statistically analyzed. Within a long-term time window, moisture benchmark level features are extracted, and the average level of moisture within the long-term window is calculated as the current period's moisture background benchmark. For groundwater pressure monitoring indicators, this study extracts pressure abrupt change characteristics within a short-term time window. By detecting abrupt changes in pressure time-series data, it identifies the timing and magnitude of rapid pressure increases or decreases. Within a medium-term time window, it extracts pressure response delay characteristics, analyzing the time lag of groundwater pressure response to rainfall events. By calculating the time difference between the start of rainfall and the start of groundwater pressure increases, it quantifies the groundwater response delay to rainfall. Within a long-term time window, it extracts the baseline pressure change trend, analyzing the long-term upward or downward trend of groundwater pressure.For the tilt angle monitoring index, the tilt angle change rate feature is extracted in the short-term time window, the cumulative tilt angle change feature is extracted in the medium-term time window, and the long-term tilt angle drift trend feature is extracted in the long-term time window. Through the extraction of specific features of the above-mentioned monitoring indicators, a set of specific time-series features is obtained.

[0078] Step 3.3: Perform deep fusion processing of multi-scale temporal features;

[0079] Based on the multi-timescale basic statistical feature set output in step 3.1 and the special time-series feature set output in step 3.2, a multi-level feature fusion architecture is established. The first layer of fusion is the horizontal fusion of features of different monitoring indicators within the same time scale. For short-term time windows, all feature parameters of five types of monitoring indicators—rainfall, soil moisture, surface displacement, groundwater pressure, and dip angle—are extracted at that time scale to construct a short-term feature parameter matrix. The dynamic weighting system based on the output in step 2 is called to obtain the weight coefficients of each monitoring indicator at the current time. For each feature parameter in the feature parameter matrix, the feature value is weighted according to the weight coefficient of the monitoring indicator to which the feature belongs. A weighted summation method is used to accumulate and fuse the weighted feature values ​​of each monitoring indicator within the same time scale to generate a comprehensive feature vector for that time scale. The short-term time scale comprehensive feature vector mainly reflects the abrupt changes and rapid changes in the monitoring data. The above process is repeated for the medium-term and long-term time windows, generating medium-term and long-term comprehensive feature vectors respectively. The medium-term comprehensive feature vector mainly reflects the trend evolution information and periodic change characteristics of the monitoring data, while the long-term comprehensive feature vector mainly reflects the cumulative effect information and background state change characteristics of the monitoring data. The second layer of fusion is a cross-scale vertical fusion of comprehensive feature vectors at different time scales. The comprehensive feature vectors of the short-term, medium-term, and long-term time scales are used as fusion inputs. According to the temporal evolution law of debris flow disasters, differentiated fusion weights are assigned to features at different time scales. Short-term features reflect immediate abnormal states and have an important indicative role for short-term early warning, so they are assigned a sudden change weight coefficient. Medium-term features reflect evolution trends and have key value for trend prediction, so they are assigned a trend weight coefficient. Long-term features reflect background benchmarks and provide a benchmark reference for state assessment, so they are assigned a background weight coefficient. The sudden change weight coefficient is set to a relatively high value to strengthen the response capability to sudden anomalies, the trend weight coefficient is set to a medium value to balance short-term and long-term information, and the background weight coefficient is set to a relatively low value to provide auxiliary reference. The comprehensive feature vectors of the three time scales are multiplied by their corresponding fusion weight coefficients, and then the vectors are concatenated or weighted summed to generate a deep temporal feature vector that integrates information from multiple time scales.

[0080] In some embodiments, since the time span of disaster gestation processes may differ across monitoring points, a fixed time window scale may not optimally suit the characteristics of all monitoring points. An adaptive time window setting method based on historical disaster gestation cycles can be adopted, aiming to customize personalized time window parameters for different monitoring points. Specifically, for monitoring points with sufficient historical disaster cases, monitoring data prior to the historical disasters are extracted, and the time span from the appearance of anomalies in the monitoring data to the actual occurrence of the disaster is analyzed and defined as the typical disaster gestation cycle for that monitoring point. The gestation cycles of multiple historical cases are statistically analyzed, and the average gestation cycle and its distribution range are calculated. Based on the average gestation cycle, a personalized time window scale is set for that monitoring point: a short-term window is set to one-tenth of the average gestation cycle, a medium-term window to half the average gestation cycle, and a long-term window to twice the average gestation cycle. Through adaptive time window setting, the time window scale matches the disaster evolution pattern of the monitoring point, improving the targeting of feature extraction and the accuracy of the early warning model.

[0081] The risk assessment module, based on deep temporal feature vectors and the spatial topology network structure of monitoring points, performs intelligent perception of disaster situation and quantitative risk assessment, and obtains the risk assessment results of each monitoring point;

[0082] Step 4.1: Construct the spatiotemporal graph neural network architecture and risk assessment model;

[0083] A spatiotemporal graph structure of monitoring points is constructed to support subsequent graph neural network analysis. Each monitoring point within the monitoring area is constructed as a graph structure, with the monitoring points acting as nodes. The node features are the depth-time series feature vectors generated above. Based on the spatial topology network structure of the monitoring points output in step 1, the edge connections of the graph are determined. Edge connections are established between monitoring points with spatial proximity, hydrological connectivity, or geological similarity. The edge weights are calculated comprehensively based on factors such as spatial distance, elevation difference, and geological similarity. For spatial distance weight, the geographical distance between two monitoring points is calculated; the closer the distance, the greater the weight. For elevation difference weight, the elevation difference between two monitoring points is calculated; the smaller the elevation difference, the better the terrain continuity, and the greater the weight. For geological similarity weight, the similarity of geological parameters such as soil type, rock structure, and slope conditions between two monitoring points is compared; the higher the similarity, the greater the weight. The three types of weights are normalized and then weighted and summed to obtain the final edge weight values. Through the above processing, a complete spatiotemporal graph structure containing node features and edge weights is obtained. This graph structure comprehensively represents the evolutionary characteristics of the disaster gestation process in three dimensions: short-term mutation, medium-term trend, and long-term accumulation, as well as the spatial correlation between monitoring points, providing comprehensive feature input and graph structure support for subsequent disaster gestation status assessment.

[0084] Based on the spatiotemporal graph structure of monitoring points, the node features and edge connections of the graph are directly obtained. Node features are deep temporal feature vectors, and edge weights are calculated comprehensively based on factors such as spatial distance, elevation difference, and geological similarity. A spatiotemporal graph neural network architecture is constructed, comprising three core modules: a graph convolutional layer, a temporal attention layer, and a risk assessment layer. The graph convolutional layer aggregates neighbor node information through a message passing mechanism to learn the spatial association patterns between monitoring points. The graph convolution operation is implemented through the following steps: calculating the degree matrix of the graph, where the diagonal elements of the degree matrix represent the number of edges connected to each node; performing a power-1 (-2) operation on the degree matrix to obtain a normalized matrix; then performing matrix multiplication operations sequentially with the adjacency matrix and the current layer node feature matrix; then performing matrix multiplication operations with the learnable weight matrix; and finally, processing through an activation function to obtain the node feature representation of the next layer. The temporal attention layer employs a multi-head attention mechanism to learn the importance weights of features at different time steps, capturing the temporal evolution patterns of the disaster gestation process. The risk assessment layer maps the output features of graph convolution and temporal attention to disaster risk probability values, and uses the sigmoid activation function to output risk assessment results between 0 and 1.

[0085] Step 4.2: Perform graph neural network training and intelligent disaster risk perception;

[0086] All historical disaster cases were extracted from the historical database to construct a training dataset for a graph neural network. For each historical disaster case, graph structure data of monitoring points at different times before the disaster was extracted, including the temporal feature vectors of each node and the topological connectivity of the graph. At the time of the disaster, the corresponding graph structure was marked as high-risk, with a label value of 1. For different time intervals before the disaster, risk label values ​​were calculated using an exponential decay function based on time distance; the closer to the disaster, the closer the risk label value was to 1, and the farther away, the closer the risk label value was to 0. An end-to-end graph neural network was used for training, employing a binary cross-entropy loss combined with a graph regularization term to ensure the network could accurately predict risks while maintaining the smoothness of the graph structure. During training, the graph convolutional layer learned the spatial dependencies between monitoring points, the temporal attention layer learned the importance weights of the time dimension, and the risk assessment layer learned the nonlinear mapping from graph features to risk probabilities. The network parameters were optimized using a backpropagation algorithm, enabling the network to intelligently perceive the disaster development trend based on the current graph structure of the monitoring points. After training, the graph structure constructed from real-time monitoring data is input into the network, and the network outputs the real-time risk probability value of each monitoring point. The risk probability value reflects the current disaster hazard level of the monitoring point.

[0087] Step 4.3, Dynamic determination of adaptive threshold based on graph attention mechanism;

[0088] By leveraging attention weights in graph neural networks, adaptive adjustment of warning thresholds is achieved. The graph attention mechanism automatically calculates the attention weights of each monitoring point to its neighboring nodes during the learning process; the weights reflect the strength and influence of the association between nodes. For a specific monitoring point, its attention weight distribution in the graph network is extracted, reflecting its importance and sensitivity within the spatial network. Combining the historical risk probability distribution characteristics of the monitoring point, its baseline risk probability and standard deviation under normal conditions are calculated. Based on the attention weights and historical risk distribution, a Bayesian method is used to dynamically adjust the warning threshold. For critical monitoring points with high attention weights, indicating their important position in the disaster propagation network, their warning thresholds are appropriately lowered to improve warning sensitivity. For peripheral monitoring points with low attention weights, their warning thresholds are appropriately raised to reduce false alarm rates. The threshold adjustment process is performed as follows: First, the baseline warning threshold for the monitoring point is obtained as the starting point for adjustment. Then, the normalized attention weight value for that monitoring point is extracted. Next, the attention weight adjustment factor is calculated by subtracting the product of the first adjustment coefficient and the attention weight from 1. Then, the noise suppression adjustment factor is calculated by multiplying the second adjustment coefficient and the noise suppression factor. Finally, the baseline threshold is multiplied by the sum of the attention weight adjustment factor and the noise suppression adjustment factor to obtain the adaptive warning threshold for that monitoring point. Through this graph attention-guided adaptive threshold mechanism, personalized and dynamic adjustment of the warning threshold for each monitoring point is achieved.

[0089] Step 4.4: Real-time assessment of the current disaster risk status level at each monitoring point;

[0090] For any monitoring point, obtain the real-time risk probability value calculated in step 4.2, and retrieve the adaptive early warning threshold determined for that monitoring point in step 4.3. Compare the current risk probability value with the threshold boundary of that monitoring point. When the risk probability value is lower than the lower limit of the attention state threshold, the current situation is determined to be normal. When the risk probability value is within the attention state threshold range, the current situation is determined to be attention state. When the risk probability value is within the early warning state threshold range, the current situation is determined to be early warning state. When the risk probability value exceeds the lower limit of the danger state threshold, the current situation is determined to be danger state. Record the determination result along with the current time, monitoring point identifier, and risk probability value to generate the real-time disaster risk situation assessment result for that monitoring point. Repeat the above process for all monitoring points within the monitoring area to obtain the set of real-time risk situation level determination results for each monitoring point.

[0091] The propagation analysis module analyzes the spatial coordination relationship and disaster propagation patterns among monitoring points based on the risk assessment results of each monitoring point, and obtains the regional propagation assessment results;

[0092] Step 5.1: Extract the spatial topology network structure of the monitoring points;

[0093] The spatial topology network of monitoring points, based on the output of step 1, is invoked. This network uses monitoring points as nodes and pairs of monitoring points with potential disaster propagation paths as directed edges. Spatial location information of each node in the network is extracted, including latitude and longitude coordinates and elevation data. Attribute information of each directed edge in the network is extracted, including the identifiers of the starting and ending nodes connected by the edge. The starting node is a monitoring point with higher elevation upstream, and the ending node is a monitoring point with lower elevation downstream. The weight attribute of the edge is extracted; the weight value is calculated from the spatial distance and elevation difference between the two nodes. Further, combined with watershed river system distribution data, the surface runoff paths corresponding to each directed edge are identified to determine whether there are water flow convergence channels and potential debris flow transport paths between two monitoring points. For monitoring point pairs with clear runoff path connections, the corresponding directed edges are retained in the topology network; for monitoring point pairs without runoff connectivity, the corresponding edges are removed from the topology network. After screening and optimization, a spatial topology network structure of monitoring points reflecting the true hydrological connectivity and disaster propagation paths is obtained.

[0094] Step 5.2: Perform spatiotemporal collaborative anomaly identification of upstream and downstream monitoring points;

[0095] Traverse all monitoring nodes in the topology network and extract the current risk status level of each node based on the output of step 4. Identify monitoring points with a current risk status level of warning or danger and mark these monitoring points as anomaly source nodes. For each anomaly source node, find the set of downstream nodes directly connected to this node through directed edges in the topology network. A downstream node is defined as the terminal node pointed to by a directed edge starting from the anomaly source node. Obtain the real-time monitoring data and time series data of risk probability values ​​for each downstream node. Extract the moment when the risk probability value of the anomaly source node reaches the warning threshold and record it as the anomaly trigger moment. Track the trajectory of the risk probability value change of each downstream node in the time period after the anomaly trigger moment, which is set as a 6-hour window after the anomaly trigger moment. Analyze the changing trend of the risk probability value sequence of downstream nodes and calculate the time derivative of the risk probability value to represent the rate of increase. When the risk probability value of a downstream node shows an upward trend after the anomaly trigger and the rate of increase exceeds the normal fluctuation rate threshold of the node, it is determined that the downstream node is affected by the upstream anomaly. The time difference between the moment when the risk probability value of the upstream anomaly source node rises and the moment when the risk probability value of the downstream node begins to rise is calculated. This time difference characterizes the time lag of disaster propagation along the path. Using time series correlation analysis, the lagged cross-correlation coefficients of the risk probability value sequences of the upstream and downstream nodes are calculated. The time lag corresponding to the peak of the correlation coefficient reflects the disaster propagation delay between the two nodes. When the lagged cross-correlation coefficient exceeds a set correlation threshold, and there is a reasonable temporal causal relationship between the rise in the risk probability value of the downstream node and the anomaly of the upstream node, a disaster propagation symptom is identified, confirming a disaster synergistic correlation between the upstream and downstream nodes. This yields the disaster propagation symptom identification results and the synergistic correlation between the upstream and downstream nodes.

[0096] Step 5.3: Predict the disaster's propagation path, impact range, and arrival time;

[0097] Based on the disaster propagation signs and collaborative relationships identified in step 5.2, a disaster propagation path prediction model is constructed. For the identified anomalous source node, all downstream nodes reachable from this node through the topological network are extracted. Reachability is defined as the set of nodes reachable from the anomalous source node along a directed edge path. Based on the connectivity of the topological network, propagation paths from the anomalous source node to each downstream node are identified. Each path consists of a series of continuous directed edges. For each propagation path, the spatial distance and elevation difference information of each directed edge on the path are extracted. The total length of the propagation path is obtained by summing the distances of all edges on the path, and the total elevation drop of the path is obtained by summing the elevation differences of all edges on the path. The propagation velocity records of historical debris flow disasters in the basin are queried from the historical disaster database. The historical propagation velocity data records the propagation time and average velocity of debris flows from the occurrence location to observation points at different distances. Current rainfall intensity monitoring data is extracted, and the similarity between the current rainfall conditions and the rainfall conditions during historical disasters is analyzed. Based on the statistical characteristics of propagation velocity under similar historical rainfall conditions, combined with the slope conditions of the current path, a propagation velocity estimation model is used to calculate the expected propagation velocity under the current conditions. The propagation velocity estimation model considers the influence of factors such as path slope, rainfall intensity, and source conditions on propagation velocity, and calculates it using empirical formulas or statistical regression models. Based on the expected propagation velocity and the total length of the propagation path, the estimated propagation time from the anomalous source node to each downstream node is calculated. The propagation time is obtained by calculating the ratio of path length to propagation velocity. Adding the current time of the anomalous source node to the propagation time yields the estimated time window for the disaster to reach each downstream node. Combining watershed topographic boundary and catchment area data, the spatial range of potentially affected areas along the disaster propagation path is identified. Using buffer analysis, an impact buffer of a certain width is generated with the propagation path as the center line. The width of the buffer is determined based on the lateral diffusion range of historical disasters. The impact buffer is overlaid and analyzed with data on roads, buildings, and population distribution in the geographic information system to identify potentially affected important targets and densely populated areas. Integrating information such as the propagation path, downstream affected nodes, estimated arrival time, impact range boundaries, and potentially affected targets, a regional propagation assessment result containing complete spatiotemporal information is generated.

[0098] In some embodiments, since areas with complex topography and geological conditions may have multiple potential disaster propagation paths, single-path prediction may miss secondary propagation directions. Therefore, a multi-path probability assessment method can be adopted to comprehensively identify possible disaster propagation directions and assess the probability of occurrence of each path. Specifically, for anomaly source nodes, all possible downstream paths originating from that node are identified in the topology network, including main runoff paths and secondary branch paths. For each path, the disaster propagation probability is calculated based on factors such as the path's slope, width, source distribution, and the frequency of historical disasters along that path. The path propagation probability reflects the likelihood of a disaster occurring along that path; paths with high probabilities are considered primary threat directions, and paths with low probabilities are considered secondary threat directions. All paths are sorted from high to low propagation probabilities, and paths with probabilities exceeding a set threshold are selected as key concern paths. Propagation time and impact range predictions are performed for each key concern path, generating a multi-path propagation situation assessment result. When issuing early warning information, prediction information for both primary and secondary threat paths is provided simultaneously, offering a comprehensive situational reference for emergency decision-making.

[0099] The early warning processing module generates early warning decisions and processes them in real time based on the risk assessment results of each monitoring point and the regional propagation assessment results, thus obtaining intelligent early warning processing results.

[0100] Step 6.1: Establish multi-level early warning decision rules and generate early warning information;

[0101] A rule base for early warning decisions is constructed, comprising local early warning rules based on single-point risk status and regional early warning rules based on multi-point collaborative status. The local early warning rule's judgment logic is to trigger a local early warning decision for a single monitoring point when its risk status level reaches the early warning state. The generated early warning information includes a general early warning level, the current time as the trigger time, the early warning monitoring point's identifier and location information, the current risk probability value, and a description of the risk status level. The regional early warning rule's judgment logic is to trigger a regional joint early warning decision when multiple spatially adjacent monitoring points simultaneously reach the early warning state, and step 5 identifies a disaster collaborative relationship or propagation path between these monitoring points. The criteria for judging spatial adjacency are that the distance between monitoring points is less than a set spatial proximity threshold or that there is a topological network connection between the monitoring points. The generated regional early warning information includes an upgrade of the early warning level to a severe early warning level, the identifier of the early warning source monitoring point set, a description of the identified disaster propagation path, the spatial boundary coordinates of the potential impact range, and the expected time window for the disaster to reach downstream monitoring points. When the risk status of an upstream monitoring point reaches a dangerous state, and the risk probability value of a downstream monitoring point shows a rapid upward trend, exceeding the emergency threshold, the highest level of emergency early warning decision is triggered. Emergency warning information includes an emergency warning level identifier, hazard source monitoring point information, a description of the impending or ongoing disaster situation, the affected area requiring immediate evacuation, and recommended emergency response measures. Various early warning decision rules are applied to the real-time monitoring situation assessment results. Based on the current risk status level and spatial coordination relationships of each monitoring point, corresponding early warning rules are matched to generate early warning decisions and information of the appropriate level, resulting in a tiered and categorized set of early warning decision results.

[0102] Step 6.2, Real-time data processing and rapid early warning response of edge computing nodes;

[0103] Edge computing nodes are deployed at each monitoring point. The edge node hardware configuration includes a data acquisition interface module, a local storage module, a computing processing module, and a wireless communication module. The data acquisition interface module of the edge node connects to various locally deployed sensors, receiving raw monitoring data in real time. The edge node performs data quality assessment, intelligent repair, and spatiotemporal benchmark unification processing based on the output of step 1. It performs quality checks on the received raw data, identifies and repairs abnormal data, converts data timestamps to a unified time benchmark, and completes data preprocessing. The local storage module of the edge node stores the historical monitoring data of the monitoring point for the most recent two hours, with sufficient storage capacity to support short-term time window analysis. The edge node calls the locally stored weight parameter configuration file for the monitoring point, which is distributed from the cloud and contains the monitoring indicator weight coefficients based on the output of step 2. The edge node performs short-term time window time series feature extraction based on the output of step 3. It calculates short-term basic statistical features and specific features based on the monitoring data of the most recent two hours, generating a short-term feature vector.

[0104] The edge node invokes a simplified risk assessment model stored locally. This model is a lightweight version of the complete cloud model, adapted to the computing resource limitations of edge devices. Short-term feature vectors are input into the simplified model to calculate the short-term risk probability value at the current moment. The edge node locally stores the fast response threshold parameter for this monitoring point. When the calculated short-term risk probability value exceeds the fast response threshold, an emergency anomaly is determined. The edge node immediately activates its local audible and visual warning device, issuing a field warning signal. The audible and visual warning device includes a high-decibel siren and a high-brightness warning light, alerting on-site personnel and passing vehicles. Simultaneously, the edge node packages and uploads the warning information and processed monitoring data to the cloud server via its wireless communication module. Data upload uses an encrypted transmission protocol to ensure data security. The entire processing flow from anomaly detection to triggering a local warning is completed locally on the edge node, with processing latency controlled within seconds, achieving a rapid warning response from milliseconds to seconds, and obtaining the local rapid warning result from the edge node.

[0105] Step 6.3, Global In-Depth Analysis and Collaborative Early Warning Decision-Making on Cloud Servers;

[0106] The cloud server receives monitoring data, local early warning information, and node status information uploaded by all edge nodes. The cloud server's storage system stores complete historical monitoring data, historical disaster case data, and early warning event records for all monitoring points from system deployment to the present. The cloud computing resources are configured with a high-performance server cluster, possessing powerful data processing and model calculation capabilities. The cloud executes a complete data processing and analysis workflow, performing a differentiated weighting system construction and dynamic adjustment based on the output of step 2 on the received monitoring data, calculating accurate dynamic weight coefficients based on complete historical data and real-time status. It then performs multi-timescale deep temporal feature extraction based on the output of step 3, simultaneously analyzing monitoring data across short-term, medium-term, and long-term time windows, extracting comprehensive temporal features and performing multi-scale fusion to generate a complete deep temporal feature vector. Finally, it performs a disaster risk situation quantitative assessment based on the output of step 4, using a complete graph neural network model to calculate accurate risk probability values ​​and applying an adaptive threshold system to determine the risk situation level of each monitoring point. Finally, it performs spatial correlation analysis and disaster propagation path prediction based on the output of step 5, analyzing spatial collaborative relationships based on the global monitoring point network, identifying disaster propagation signs, and predicting propagation paths and impact ranges. Based on the global analysis results, the cloud executes multi-level early warning decision generation based on the output of step 6.1. According to the single-point situation and the multi-point collaborative situation, it generates local early warning, regional early warning, or emergency early warning decisions. The cloud distributes the early warning decision information to the edge nodes of all relevant monitoring points via the communication network. After receiving the early warning decision, the edge nodes update their local early warning status and activate the corresponding level of early warning response. Simultaneously, the cloud pushes the early warning information to the early warning information release platform. The early warning platform releases early warning notifications to the monitoring area management department, emergency response department, and the public through various channels such as SMS, web pages, and mobile applications. Based on the latest monitoring data and early warning effect feedback, the cloud periodically optimizes model parameters, packages the optimized weight parameters, model parameters, and threshold parameters into a configuration update file, and distributes it to each edge node via the communication network for parameter synchronization. After receiving the configuration update file, the edge nodes replace the old parameter configuration stored locally, realizing remote updating and version synchronization of model parameters. By combining rapid local response at edge nodes with in-depth global analysis in the cloud, the system ensures both the real-time nature of early warnings to meet the needs of rapid response in emergency situations and the accuracy of early warnings through in-depth analysis using the powerful computing capabilities of the cloud, resulting in intelligent early warning processing results through edge-cloud collaboration.

[0107] The feedback optimization module verifies and provides feedback on the intelligent early warning processing results, identifies problem monitoring points and optimizes them to obtain an optimized configuration scheme;

[0108] Step 7.1: Establish a database for recording and verifying early warning events;

[0109] Establish an early warning event database, which records detailed information about every early warning event triggered during system operation. For each triggered early warning event, record the timestamp of the early warning trigger time, the early warning level (local, regional, or emergency), the early warning monitoring point identifier and location information, the risk probability value of the monitoring point and related monitoring points at the time of the early warning trigger, a snapshot of the monitoring data at the time of the early warning trigger including the values ​​of each monitoring indicator, and the specific content of the early warning decision including the early warning scope and recommended measures.

[0110] Establish an early warning verification and tracking mechanism to continuously track the actual situation in the monitored area after each early warning event is issued. Set a 24-hour verification time window for each warning, during which continuous monitoring is conducted to determine if an actual disaster event occurs. Record whether a debris flow disaster occurs within the warning area through various means, including on-site manual inspections, video surveillance, and sensor anomaly alarms. If a disaster occurs, record the actual time of occurrence, location, scale, impact range, and damage. Verify and label the early warning event based on the comparison between the warning results and the actual situation. If the system issues an early warning and a disaster does occur within the verification time window, with the disaster location matching the warning area and the time falling within a reasonable response window after the warning, the early warning event is labeled as a correct warning, indicating its accuracy and effectiveness. If an actual disaster occurs within the verification time window but the system did not issue an early warning in advance, or if the warning was issued too late to provide sufficient lead time, it is labeled as a missed event, indicating the system failed to identify disaster signs in time. If the system issues an early warning but no actual disaster occurs within the verification time window, it is labeled as a false alarm, indicating the system generated a false warning. By continuously recording and verifying early warning events, a historical early warning event database containing early warning decision information and verification result annotations is established.

[0111] Step 7.2: Perform early warning performance assessment and problem monitoring point identification;

[0112] Based on the early warning event database, the system's early warning performance is quantitatively evaluated according to a set evaluation period, which can be monthly or quarterly. Within an evaluation period, the total number of early warnings triggered by the system is counted, along with the number of events marked as correct and false alarms. Simultaneously, the total number of actual disasters occurring in the monitored area within the evaluation period is counted, along with the number of events marked as missed. An early warning accuracy rate is calculated, which is the ratio of correct early warnings to the total number of early warnings. This rate reflects the proportion of correct early warnings issued by the system; a higher accuracy rate indicates higher reliability. A false alarm rate is calculated, which is the ratio of missed alarms to the total number of actual disasters. This rate reflects the proportion of actual disasters that the system failed to warn of; a lower false alarm rate indicates a stronger ability to detect disasters. A false alarm rate is calculated, which is the ratio of false alarms to the total number of early warnings. This rate reflects the proportion of false alarms; a lower false alarm rate indicates higher accuracy. These performance indicators are calculated for the entire system to obtain the overall early warning performance evaluation result.

[0113] Further performance evaluations were conducted for each monitoring point, and the early warning events and verification results triggered by each monitoring point were statistically analyzed. The early warning accuracy, false negative rate, and false positive rate for each monitoring point were calculated. Performance standards were set: when the early warning accuracy of a monitoring point was lower than the set lower threshold, or the false positive rate was higher than the set upper threshold, or the false negative rate was higher than the set upper threshold, the monitoring point was marked as a problem monitoring point, indicating that its early warning performance was substandard and required optimization. After completing the performance evaluation for all monitoring points, the set of monitoring points with performance problems was identified, resulting in the overall system and individual monitoring point early warning performance evaluation results, as well as a list of problem monitoring points.

[0114] Step 7.3: Perform reverse optimization of model parameters for the problem monitoring points;

[0115] For each monitoring point in the problem monitoring point set output in step 7.2, analyze the reasons for its early warning performance problems and perform targeted parameter optimization. For monitoring points with high false alarm rates, extract all false alarm event records for that monitoring point and analyze the monitoring data characteristics and risk probability value characteristics when the false alarm events occur. Through comparative analysis, it was found that although the risk probability value of the false alarm events exceeded the early warning threshold, no disaster actually occurred, indicating that the early warning threshold setting for that monitoring point was too low, causing normal fluctuations to be misjudged as an early warning state.

[0116] The adaptive early warning threshold for this monitoring point, based on the output of step 4.3, is adjusted upwards. The threshold boundary values ​​for early warning and danger states are adjusted upwards by a certain proportion, determined by the degree of false alarm rate exceeding the limit; the higher the false alarm rate, the greater the threshold increase. For monitoring points with high missed alarm rates, all missed event records for that monitoring point are extracted, and the monitoring data and feature extraction results before the disaster occur in the missed events are analyzed. The analysis reveals that abnormal changes in certain key monitoring indicators in missed events were not fully perceived by the model, indicating that the weight coefficients of these indicators are insufficient. Monitoring indicators showing significant abnormal changes in missed events are identified, and the intelligent indicator weight set for each monitoring point, based on the output of step 2.3, is adjusted and optimized to increase the weight coefficients of these key disaster-causing factor indicators. By increasing the weight of key indicators, the model's sensitivity to disaster precursor signals is enhanced. For monitoring points with generally low early warning accuracy, both false alarms and missed alarms exist, indicating that the risk assessment model itself has biases. Historical disaster case data and historical monitoring data for this monitoring point were re-extracted. The spatiotemporal graph neural network risk assessment model based on the outputs of steps 4.1 and 4.2 was retrained using the latest accumulated data samples. The model's mapping function parameters were optimized to improve the model's fitting accuracy for the disaster characteristics of this monitoring point. For each problematic monitoring point, corresponding optimization measures such as threshold adjustment, weight optimization, or model retraining were implemented according to its specific performance problem type to obtain an optimized configuration scheme, including updated intelligent indicator weight coefficients, adaptive early warning thresholds, and graph neural network model parameters.

[0117] Step 7.4: Perform model parameter update deployment and continuous iterative optimization;

[0118] The optimized parameter configuration scheme based on the output of step 7.3 is deployed and updated to the system runtime environment. The cloud server writes the updated parameters for each problem monitoring point into the system parameter database, replacing the original old parameter configuration. A parameter update configuration file is generated in the cloud, containing the monitoring point identifier, the updated parameter type, and the updated parameter value. The configuration file is distributed to the edge computing nodes of the corresponding monitoring points via the communication network. After receiving the configuration file, the edge nodes perform local parameter updates, replacing the locally stored intelligent indicator weight coefficients, adaptive early warning threshold parameters, and graph neural network model parameters with the updated versions, completing the edge-end parameter synchronization. After the parameter update is completed, the system continues to run with the new parameter configuration, performing analysis and early warning based on the updated intelligent indicator weight set, adaptive early warning threshold, and graph neural network model on subsequent monitoring data.

[0119] The system continuously executes the iterative process from steps 7.1 to 7.4, accumulating new early warning events and verification feedback data during operation. It periodically performs performance evaluations and parameter optimizations. Through a feedback-driven parameter adjustment mechanism, system performance continuously improves with runtime and data accumulation. Through the aforementioned early warning effect evaluation feedback and model adaptive optimization mechanism, the system possesses continuous learning and self-evolution capabilities, achieving continuous improvement in early warning accuracy and reliability. This results in an adaptive optimization parameter set, which includes optimized intelligent indicator weight coefficients, adaptive early warning thresholds, and graph neural network model parameters, enabling continuous performance improvement of the early warning system.

[0120] In one embodiment of the present invention, a specific example is provided:

[0121] Taking the K235 to K258 section of a mountainous expressway in a certain province as an example area, there are 8 key monitoring sections along this section, numbered MP01 to MP08.

[0122] Application scenario data example:

[0123] The initial data collected by the system at monitoring points during the flood season on July 15, 2025 are shown in Table 1:

[0124] Table 1: Example of initial data collection at monitoring points;

[0125]

[0126] Examples of processed feature data are shown in Table 2:

[0127] Table 2: Examples of processed feature data;

[0128]

[0129] Based on the complete processing flow of steps 1 to 7, the system performs quality assessment, spatiotemporal alignment, feature extraction, situation assessment, and spatial correlation analysis on the aforementioned monitoring data. The system identifies monitoring point MP04 as being in a dangerous state, and monitoring point MP02 as being in a warning state, with an upstream-downstream spatial correlation between MP04 and MP02. The system generates a regional joint early warning decision, with warning information including the disaster source location as monitoring point MP04, the predicted propagation path along the basin downstream to MP02 and further downstream areas, the estimated time window for arrival at MP02 being 30 to 60 minutes after the warning is issued, and a recommendation to implement traffic control on the K238 to K242 section. Forty-five minutes after the warning is issued, a debris flow disaster occurs in the area upstream of monitoring point MP04, moving downstream along the predicted path. Due to the timely warning and the early implementation of road control measures, casualties and vehicle damage were successfully avoided, verifying the accuracy and effectiveness of the system's early warning.

[0130] The debris flow geological disaster early warning and monitoring system provided by this invention realizes real-time data acquisition, intelligent analysis, and collaborative early warning for multiple monitoring sections along mountainous highways. Through multi-source data quality assurance and intelligent repair mechanisms, the system effectively solves the problems of missing and abnormal monitoring data, ensuring the reliability of analytical input. The differentiated weighting system construction method based on transfer learning enables the system to adapt to monitoring points with scarce data, improving the model's generalization ability and applicability. Multi-timescale deep feature extraction and adaptive threshold mechanisms enhance the system's comprehensive perception of the disaster development process and its personalized early warning capabilities. Spatial correlation analysis and propagation path prediction functions realize the leap from single-point early warning to regional collaborative early warning, providing precise spatiotemporal early warning information for emergency decision-making. The edge cloud collaborative architecture improves the depth of analysis while ensuring the real-time nature of early warnings, and the early warning effect evaluation feedback mechanism enables the system to have continuous learning and self-optimization capabilities. This system provides important technical support and decision-making basis for the effective prevention and emergency response to debris flow geological disasters.

[0131] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A debris flow geological disaster early warning and monitoring system, characterized in that, Includes the following steps: The data preprocessing module collects raw monitoring data, performs preprocessing, and obtains a high-quality multi-source monitoring dataset; it extracts spatial information of monitoring points, calculates the spatial distance between monitoring points, and obtains the spatial topology network structure of monitoring points. The weight allocation module, based on a high-quality multi-source monitoring dataset, performs feature learning and multi-modal feature weight allocation to obtain an intelligent indicator weight set. The feature fusion module extracts statistical features and time-series features based on an intelligent indicator weight set and a high-quality multi-source monitoring dataset, and then performs feature fusion to obtain a deep time-series feature vector. The risk assessment module, based on deep temporal feature vectors and the spatial topology network structure of monitoring points, performs intelligent perception of disaster situation and quantitative risk assessment, and obtains the risk assessment results of each monitoring point; The propagation analysis module analyzes the spatial coordination relationship and disaster propagation patterns among monitoring points based on the risk assessment results of each monitoring point, and obtains the regional propagation assessment results; The early warning processing module generates early warning decisions and processes them in real time based on the risk assessment results of each monitoring point and the regional propagation assessment results, thus obtaining intelligent early warning processing results. The feedback optimization module verifies and provides feedback on the intelligent early warning processing results, identifies problem monitoring points and optimizes them to obtain an optimized configuration scheme.

2. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps for collecting and preprocessing raw monitoring data include: performing integrity and rationality quality checks on the monitoring data; extracting the timestamp field, data value field, sensor identifier field, and data type field from the received monitoring data records; and performing a data integrity verification process. For the data value field, theoretical upper and lower limits are obtained based on the physical measurement principle and measurement range constraints of the sensor. An outlier detection method based on statistical distribution is used to establish a normal value distribution range. The current data value is compared with the theoretical boundary and the statistical distribution range. When the value exceeds the theoretical boundary or deviates from the statistical distribution range by more than a preset deviation threshold, the data is marked as an outlier.

3. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps of extracting spatial information of monitoring points and calculating spatial distances between monitoring points include: performing a unified transformation of the spatiotemporal reference, extracting spatial coordinate records and coordinate system type information of each monitoring point, and using a coordinate system transformation algorithm to transform coordinates under different coordinate systems to a unified spatial reference coordinate system; calculating the spatial distance between any two monitoring points based on the location information of the monitoring points under the unified coordinate system, extracting digital elevation model data of the monitoring area to obtain the elevation information of each monitoring point, and calculating the relative elevation difference between monitoring points; combining the watershed distribution data and surface runoff path data of the monitoring area to analyze the hydrological connectivity between monitoring points, and constructing a spatial topology network of monitoring points based on spatial distance, elevation difference, and hydrological connectivity information.

4. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps of feature learning and multimodal feature weight allocation include: constructing a multimodal deep learning network architecture that integrates the physical evolution mechanism of debris flow, adopting a multi-branch structure, with each branch corresponding to the physical evolution process of a class of monitoring indicators, and incorporating the constraints of the corresponding physical process into the design of the activation function and loss function of the branch network; performing deep feature learning and adaptive weight optimization of multimodal monitoring data, learning the intrinsic correlation between different monitoring indicators through a self-attention mechanism, and linearly combining the feature representations of each monitoring indicator through learnable weight coefficients.

5. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps for extracting statistical and temporal features include: constructing multi-scale time windows and extracting basic statistical features; setting time window scales of short-term, medium-term, and long-term; calculating basic statistical feature parameters for the temporal data of each monitoring indicator within each time window; extracting specific temporal features for various monitoring indicators; extracting instantaneous rainfall intensity peak features and rainfall duration features for rainfall monitoring indicators; extracting instantaneous displacement velocity features and displacement increment sequence features for surface displacement monitoring indicators; and extracting moisture growth rate features and saturation change features for soil moisture monitoring indicators.

6. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps for intelligent disaster situation perception include: constructing a spatiotemporal graph neural network architecture and a risk assessment model, constructing each monitoring point in the monitoring area as a graph structure, with the monitoring points as nodes in the graph, and the node features being deep temporal feature vectors; determining the edge connection relationships of the graph based on the spatial topology network structure of the monitoring points; constructing a spatiotemporal graph neural network architecture, including a graph convolutional layer, a temporal attention layer, and a risk assessment layer; the graph convolutional layer aggregates neighbor node information through a message passing mechanism, and the temporal attention layer uses a multi-head attention mechanism to learn the importance weights of features at different time steps.

7. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps of the risk quantification assessment include: dynamically determining the adaptive threshold based on the graph attention mechanism, and adaptively adjusting the warning threshold using the attention weight information in the graph neural network; judging the current disaster risk status level of each monitoring point in real time, comparing the current risk probability value with the threshold boundary of the monitoring point, and judging it as normal state, attention state, warning state and dangerous state respectively.

8. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps for analyzing the spatial coordination relationship and disaster propagation patterns among monitoring points include: performing spatiotemporal coordination anomaly identification of upstream and downstream monitoring points; identifying monitoring points with current risk status at warning or danger level as anomaly source nodes; finding a set of downstream nodes for each anomaly source node; tracking the change trajectory of risk probability values ​​of each downstream node within a time period after the anomaly triggering time; predicting the disaster propagation path, impact range, and arrival time; and calculating the estimated propagation time from the anomaly source node to each downstream node based on the expected propagation speed and total propagation path length.

9. The debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps for generating and processing early warning decisions in real time include: establishing multi-level early warning decision rules and generating early warning information; constructing an early warning decision rule library, which includes local early warning rules based on single-point risk status and regional early warning rules based on multi-point collaborative status; triggering local early warning decisions when the risk status level of a single monitoring point reaches the early warning state; triggering regional joint early warning decisions when a preset number of spatially adjacent monitoring points simultaneously reach the early warning state and there is a disaster collaborative relationship or propagation path; deploying edge computing node devices at each monitoring point; the edge nodes perform data quality assessment, intelligent repair, and spatiotemporal benchmark unification processing; calling locally stored simplified risk assessment models to calculate short-term risk probability values; and activating local audible and visual early warning devices when the short-term risk probability exceeds the rapid response threshold.

10. A debris flow geological disaster early warning and monitoring system according to claim 1, characterized in that, The steps for verifying and providing feedback on the intelligent early warning processing results, identifying problem monitoring points, and optimizing them include: establishing an early warning event record and verification feedback database to record detailed information of every early warning event triggered during system operation; setting an early warning verification time window; continuously monitoring the occurrence of actual disaster events within the early warning verification time window; verifying and labeling early warning events based on the comparison between early warning results and actual conditions; calculating early warning accuracy, false alarm rate, and false alarm rate indicators; identifying a monitoring point as a problem monitoring point when its early warning accuracy is lower than a preset lower accuracy threshold, its false alarm rate is higher than a preset upper false alarm rate threshold, or its false alarm rate is higher than a preset upper false alarm rate threshold; and performing reverse optimization of model parameters for the problem monitoring points.

Images