Geological disaster early warning system and device based on multi-modal data fusion

Abstract

The invention relates to the technical field of alarm devices, and particularly discloses a geological disaster early warning system and device based on multi-modal data fusion, and the system is provided with a time sequence feature extraction module, a time sequence feature prediction module, a trend feature prediction module, a risk trend probability fusion module and an early warning device linkage module. The time sequence feature extraction module gathers multi-source data in real time, completes timestamp time synchronization and STL (Standard Temporary Language) de-seasonal operation at an edge gateway, and generates robust time sequence features; the time sequence feature prediction module analyzes the time sequence features and outputs time sequence window prediction of each node; the trend feature prediction module obtains trend window prediction by combining historical data with the window prediction as a context; the risk trend probability fusion module uses reinforcement learning as evidence fusion to generate comprehensive risk trend probability and probability confidence; the early warning device linkage module ranks according to the comprehensive risk and probability confidence, the intelligent defense terminal differentiates the linkage device and records a receipt, and urgent first and then complete available for the weak network are achieved.

Description

Technical Field

[0001] This invention relates to the field of alarm device technology, specifically to a geological disaster early warning system and device based on multimodal data fusion. Background Technology

[0002] Existing geological disaster early warning systems generally follow a closed loop of monitoring, quality control, feature identification, judgment, grading, linkage, and review: Multi-source monitoring, including rainfall, moisture content, pore pressure, displacement / tilt, microseismic activity, and remote sensing, is deployed at potential hazard points and high-risk zones and integrated into the network, simultaneously performing noise reduction, interpolation, resampling, and clock correction; based on the data, mechanisms and empirical features such as rainfall and seepage, deformation rate / acceleration, and simplified limit equilibrium are constructed, and risk probabilities are output according to single criteria (rain threshold / deformation / mechanism) or statistical / machine learning fusion; the results are mapped to early warning levels (blue / yellow / orange / red) and hysteresis and cancellation conditions are set, while risk heat maps and exposure overlay maps are generated on GIS; information dissemination and on-site handling (barriering, drainage, traffic diversion, pre-evacuation) are automatically triggered according to the grade; afterwards, the hit rate, lead time, and false alarm rate are evaluated through event playback, and thresholds and models are iteratively corrected for continuous optimization.

[0003] For example, Chinese invention patent application CN116246429A discloses a geological disaster early warning system, belonging to the field of geological disaster early warning technology. It includes a remote sensing image acquisition module, a ground elevation data acquisition module, a meteorological data acquisition module, a data processing module, a data analysis module, a monitoring module, and an early warning module. The data processing module fuses multi-temporal high-resolution remote sensing image data and ground elevation information of the target area to obtain a three-dimensional geological model and performs time-series analysis to obtain a temporal change sequence of geological conditions in the target area. The monitoring module divides the target area into key monitoring areas based on the temporal change sequence of geological conditions and conducts dynamic monitoring. The early warning module determines whether to issue a geological disaster early warning based on the dynamic monitoring results and meteorological information.

[0004] For example, Chinese invention patent application CN120318988A discloses a geological disaster early warning method and a geological disaster early warning system, including the following steps: S1: acquiring geological data of the target area; S2: performing quality control and interpolation processing on the geological data acquired in S1 to obtain standardized spatial-temporal continuous data; S3: calculating the stress state of the geological body in the target area based on the standardized spatial-temporal continuous data; S4: calculating the potential instability points in the area and their probability of occurrence; S5: training a disaster evolution prediction model and outputting the probability of disaster occurrence and its regional distribution; S6: generating graded early warning information.

[0005] Based on the above technical solutions, it was found that the actual application scenarios of geological disaster early warning are characterized by multi-source heterogeneity and strong uncertainty, and are often accompanied by extreme weather. In addition, the pre-emptive impact of communication disturbances such as weak networks / disconnections leads to problems such as data delays, missing data, and inconsistent timestamps. Most existing geological disaster early warning technologies rely on threshold triggering of single or a few indicators, and a simple mapping between the number of indicators and the alarm level. This inherently has shortcomings in reliable fusion of multi-source information and insufficient weight calibration. Under the combined effect of the above scenarios and data characteristics, not only are there problems such as disaster classification jitter and distortion, poor adaptability of fixed thresholds to site and season, but also an inability to effectively cope with the data transmission and processing needs under weak network conditions, making it difficult to maintain a continuous and reliable judgment link. Ultimately, this directly leads to a series of negative impacts such as increased false alarm and false negative rates, decreased early warning lead time, misallocation of response resources, and increased risk exposure. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a geological disaster early warning system and device based on multimodal data fusion, which can effectively solve the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a geological disaster early warning system based on multimodal data fusion, comprising: a time-series feature extraction module, used to extract multi-source sensing datasets of the geological disaster monitoring area in real time, transmit them to a local edge gateway for timestamp synchronization and time-series feature extraction, and obtain the time-series features of the monitoring area; a time-series feature prediction module, used to abstract the monitoring area into a graph neural network, limit the influence weights of the graph neural network edges through physical associations, analyze the time-series features in combination with the influence weights, and output the prediction results under the time-series feature window of each node of the graph neural network; and a trend feature prediction module, used for... Historical monitoring data of the monitoring area is extracted, and the prediction results based on time-series features and time-series feature windows are input into the time-series prediction model to output the prediction results of each node under the trend feature window. The risk trend probability fusion module is used to fuse the prediction results of the reinforcement learning model based on the time-series feature window and the prediction results under the trend feature window to obtain the comprehensive risk trend probability and probability confidence of the monitoring data of each node. The early warning device linkage module is used to rank the risk probability based on the comprehensive risk trend probability and probability confidence of the monitoring data of each node. The geological disaster intelligent prevention terminal completes the geological disaster early warning by linking the early warning device with the risk probability ranking.

[0008] A second aspect of this invention provides a geological disaster early warning device based on multimodal data fusion, comprising: a monitoring sensor, a geological disaster intelligent prevention terminal, a signal probe component, a safety component, an alarm component, an interaction component, and a communication component; the monitoring sensor is used to monitor the geological disaster monitoring area and form a multi-source sensing dataset; the geological disaster intelligent prevention terminal is used to generate execution commands based on the probability of comprehensive risk trends and to link the early warning device to issue geological disaster warnings; the signal probe component is used to monitor unknown signals in the geological disaster monitoring area and provide signal distribution data for the geological disaster intelligent prevention terminal; Safety components provide vibration or voice alerts; alarm components issue audible and visual warnings based on risk alert instructions; interaction components issue voice commands to frontline personnel to guide evacuation; and communication components send execution instructions in weak network conditions.

[0009] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) This invention provides a geological disaster early warning system and device based on multimodal data fusion. The time series feature extraction module gathers multi-source data in real time, completes time stamp synchronization and STL deseasoning at the edge gateway, and produces robust time series features. The time series feature prediction module abstracts the region into a GNN, sets edge weights according to physical association (directed waterway, travel time, co-occurrence, and cover similarity), analyzes time series features in combination with weights, and outputs time series window predictions for each node. The trend feature prediction module takes historical data + the above window predictions as context, inputs the time series model (TFT / LSTM), and obtains trend window predictions. The risk trend probability fusion module uses reinforcement learning to perform weight / evidence fusion, and generates comprehensive risk trend probability and probability confidence under physical prior and data quality constraints. The early warning device linkage module sorts according to comprehensive risk and probability confidence, and the intelligent defense terminal differentiates the linkage devices and records the receipts, realizing the priority of emergency and full coverage for weak networks.

[0010] (2) This invention obtains the beneficial effects of physical association assessment parameters of adjacent nodes by using parameters such as directional waterway reachability indication, dominant confluence ratio, co-occurrence rate, and slope-land cover similarity. The edge weight is constrained by structural connectivity (whether there is a road and how fast to travel) and calibrated by historical evidence and geomorphological consistency. Compared with the existing approach that is mostly based on fixed buffer zones or empirical adjacency, this scheme can gate pseudo-correlation and strengthen real transmission, significantly reducing misjudgments of seemingly adjacent but unconnected nodes in rainstorm-type landslide / debris flow chain scenarios.

[0011] (3) This invention uses graph neural network edge weights, context embedding for time series prediction, prior and reinforcement learning model cost functions for fusion layer, and time quality is linked in three places: features, fusion and linkage. This cross-module consistent parameter closed loop avoids the drift of the calibration of each module in the traditional process, reduces the cost of repeated calibration of data and thresholds, improves generalization, and makes abnormal / late data synchronously deweighted throughout the entire link, rather than locally and rigidly removed.

[0012] (4) Existing solutions are generally simplified in terms of grade, empirical in terms of threshold, lack of integration, and lack of strategy for weak networks. This process replaces single-point threshold judgment with STL deseasoning, graph neural network spatial propagation, temporal extrapolation, evidence fusion, and probability confidence link. It can capture precursors earlier and quantify uncertainty, and ensure minimum handling capability in weak networks through minimum available messages and back pressure / breakpoint continuation. At the same time, the physical interpretability of edge weights and full-process metadata (time quality, evidence composition, receipt) make the early warning traceable, auditable, and continuously self-calibrated. Compared with traditional black box or single-source methods, it has substantial improvements in advance, stability, false alarm rate and engineering feasibility. Attached Figure Description

[0013] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the system module connections of the present invention.

[0015] Figure 2 This is a closed-loop diagram for geological disaster early warning.

[0016] Figure 3 This is a closed-loop flowchart for geological disaster early warning.

[0017] Figure 4 The flowchart is adaptively adjusted to execute the instruction closed loop. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] In this embodiment of the invention, the closed-loop process for geological disaster early warning is specifically based on Figure 2 As shown, Figure 2This diagram illustrates a closed-loop system for geological disaster early warning, with a process encompassing multi-source monitoring, edge preprocessing, central intelligent analysis, coordinated dissemination, and terminal feedback. Data from rain gauges, soil / pore pressure sensors, and deformation sensors within the monitoring area first reaches the local edge gateway. After time synchronization, quality labeling, and lightweight aggregation, the gateway uploads the data to the intelligent prevention terminal via the main link (cellular / private network) or a backup link (satellite / BeiDou short message service). At the central side, time-series feature extraction (e.g., STL deseasoning) and graph neural networks combined with time-series prediction yield the risk trends and spread of each node. The system generates graded early warnings and confidence levels at the fusion layer; then it generates execution commands and sends them to on-site early warning devices (audio-visual alarms, intercom / radio, electronic displays, mobile push notifications, etc.). In the event of a weak network, it triggers key actions with the minimum available message first; it simultaneously conducts multi-channel releases and broadcasts to the public and management departments; each device and mobile terminal sends execution status feedback and supplementary on-site information back to the center, and the center adaptively adjusts the distribution strategy based on the feedback (switching links, downgrading commands, resuming interrupted transmissions), forming a traceable and closed-loop geological disaster early warning and emergency response system.

[0021] The specific functional flowchart is as follows: Figure 3 As shown, Figure 3 This is a closed-loop flowchart for geological disaster early warning.

[0022] Reference Figure 1 As shown, the first aspect of this invention provides a geological disaster early warning system based on multimodal data fusion, comprising: a time-series feature extraction module, a time-series feature prediction module, a trend feature prediction module, a risk trend probability fusion module, an early warning device linkage module, and a geological data management library. The geological data management library is used to store preset values ​​for various parameters.

[0023] It should be explained that the preset relationships stored in the aforementioned geological data management database include, but are not limited to, preset, matching, and mapping relationships. Specifically, in this embodiment, the acquisition method of these preset relationships takes the mapping relationship between physical association assessment parameters and edge influence weights as an example. First, relevant physical association assessment parameters (such as water system connectivity, co-occurrence of historical disasters, and slope-land cover similarity) are extracted from historical monitoring data and physical models. These parameters reflect the spatial association and influence degree between adjacent nodes. Next, through physical modeling and expert knowledge, a mapping relationship between the physical association assessment parameters and edge influence weights is constructed, establishing a mapping relationship set. This mapping relationship set can be further optimized using data-driven methods (such as regression analysis and machine learning) to ensure the accuracy and adaptability of the mapping relationship. Finally, these mapping relationships are stored in the geological data management database as the basic data for querying and inference.

[0024] The time series feature extraction module is connected to the time series feature prediction module, the time series feature prediction module is connected to the trend feature prediction module, the trend feature prediction module is connected to the risk trend probability fusion module, the risk trend probability fusion module is connected to the early warning device linkage module, and the time series feature extraction module, the time series feature prediction module, the trend feature prediction module, the risk trend probability fusion module, and the early warning device linkage module are all connected to the geological data management database.

[0025] The temporal feature extraction module is used to extract multi-source sensing datasets of geological disaster monitoring areas in real time, transmit them to the local edge gateway for timestamp synchronization and temporal feature extraction, and obtain the temporal features of the monitoring area.

[0026] The aforementioned multi-source sensing datasets are, specifically, collections of multi-dimensional data related to the same target or scene, gathered from different types of sensors and data sources. These data sources may include environmental sensors (such as meteorological, soil, and photovoltaic data), remote sensing data (such as satellite imagery and radar echo maps), and ground monitoring equipment (such as displacement sensors, seismic monitors, and cameras). By fusing information from different sensing devices and channels, multi-source sensing datasets can provide a more comprehensive and accurate description, helping to conduct more efficient monitoring, prediction, and decision-making in complex environments. Especially in tasks such as disaster early warning and environmental monitoring, they can overcome the limitations of single data sources and improve the reliability and response speed of the system.

[0027] Specifically, the local edge gateway performs timestamp synchronization and time-series feature extraction to obtain the time-series features of the monitored area. The specific analysis process is as follows: The multi-source sensing dataset is collected from various monitoring sensors in the geological disaster monitoring area.

[0028] Based on the data transmission frequency and data packaging settings, each monitoring sensor in the geological disaster monitoring area uploads multi-source sensing data to the local edge gateway.

[0029] The acquisition timestamps of each monitoring sensor in the geological disaster monitoring area and the receiving timestamps of the local edge gateway are obtained simultaneously. The acquisition timestamps are extracted from the execution records of the monitoring sensors, and the receiving timestamps are extracted from the receiving records of the local edge gateway. The acquisition timestamps of the monitoring sensors and the receiving timestamps of the local edge gateway are compared and calculated. Based on the comparison and calculation results, the timestamp synchronization process of the monitoring sensors is configured.

[0030] The specific configuration process is as follows: The upper bound of the reference timestamp of the monitoring sensor is obtained by reverse calculation based on the received timestamp of the local edge gateway. The difference between the acquisition timestamp of the monitoring sensor and the upper bound of the reference timestamp is processed to obtain the timestamp deviation of the monitoring sensor, which is then compared with a predefined timestamp deviation threshold. It needs to be explained that the above reverse calculation specifically involves the gateway marking the received time t1 with the aligned UTC when it receives a data packet, while simultaneously maintaining a one-way link delay estimate lat with the device (half the quantile of the RTT distribution obtained from the heartbeat / handshake, plus the packet / queue processing delay lat', and using the jitter standard deviation to provide tolerance). Based on this, the upper bound of the time of the event is calculated: t = t1 (lat+lat')+ ( ≈ 2σ lat ), where t refers to the actual time of sensor sampling / event occurrence, σ lat This refers to the standard deviation of jitter in link latency.

[0031] That is, the event cannot be later than the reception time minus the reasonable transmission and processing delay; when the event time reported by the device exceeds the upper limit, the unified event time is clamped back to the upper limit, and this deviation is fed back to update the delay and clock deviation estimate.

[0032] If the timestamp deviation of the monitoring sensor is less than or equal to the timestamp deviation threshold, the timestamp synchronization of the monitoring sensor is configured based on the reference timestamp. If the timestamp deviation of the monitoring sensor is greater than the timestamp deviation threshold, the timestamp synchronization of the monitoring sensor is configured based on the upper bound of the reference timestamp.

[0033] When configuring time stamp synchronization for monitoring sensors based on a reference timestamp, a globally unified reference time source (such as UTC or PTP) is first selected. The monitoring sensors are then periodically calibrated using this time source to ensure that the event times recorded by the sensors are consistent with the reference time. For weak network conditions or situations with significant latency, when configuring synchronization based on the upper bound of the reference timestamp, the sensors will limit event times to within this upper bound to avoid time discrepancies caused by latency. Simultaneously, real-time monitoring data is used to correct the sensor's local clock deviation and frequency offset, ensuring that time synchronization accuracy is maintained even in weak network environments.

[0034] After the timestamps are synchronized, the multi-source sensing dataset is decomposed using STL to obtain the temporal features of the monitored area.

[0035] The multi-source sensing dataset (rainfall, soil moisture / pore pressure, deformation, radar echo / satellite grid after node-based statistical analysis) is aligned with a uniform time step and timestamp, imputed for missing data, and denoised. Then, for each monitoring node and each indicator sequence, STL (Seasonal-Trend decomposition using Loess) is applied to decompose it to obtain long-term trend terms, periodic / seasonal terms, and random disturbance / anomaly terms. Based on this, time-series features are derived, such as trend slope / curvature, periodic amplitude and phase, anomaly intensity and quantile residuals, short-window increments and momentum, rolling statistics (mean / variance), extreme values ​​and duration, lag correlation and cross-correlation, etc. Among them, time-series features refer to numerical descriptors formed by deseasoning and robustly refining the original time series, which can be used for modeling and early warning, and are used for subsequent GNN spatial fusion, time-series prediction, and risk classification.

[0036] Furthermore, the data transmission setting frequency and data packaging setting window are analyzed in detail as follows: A heartbeat detection mechanism is configured between the monitoring sensor and the local edge gateway, denoted as the transmitted heartbeat.

[0037] The return duration of the transmission heartbeat is obtained, which can be extracted from the execution record of the local edge gateway. The data transmission status between the monitoring sensor and the local edge gateway is determined based on the return duration of the transmission heartbeat.

[0038] The specific determination is as follows: The return time of the transmission heartbeat is compared with the predefined return time allowable range. If the return time of the transmission heartbeat falls within the return time allowable range, the data transmission status is determined to be normal transmission.

[0039] If the heartbeat transmission time does not fall within the permitted transmission time range: If the return time of the transmission heartbeat is less than the minimum value of the permitted return time interval but greater than or equal to the return time limit, the data transmission status is determined to be weak network transmission. If the return time of the transmission heartbeat is less than the return time limit, the data transmission status is determined to be abnormal transmission.

[0040] It should be explained that the data relationship between the backhaul definition duration and the minimum value of the backhaul duration permitted range is that the backhaul definition duration is less than the minimum value of the backhaul duration permitted range.

[0041] Data transmission status, including normal transmission, weak network transmission, and abnormal transmission.

[0042] If the data transmission status shows normal transmission, then the data transmission reference frequency is recorded as the data transmission set frequency, and the data packaging reference window is recorded as the data packaging set window.

[0043] If the data transmission status shows weak network transmission, the frequency correction factor is multiplied by the data transmission reference frequency to obtain the data transmission correction frequency, which is recorded as the data transmission set frequency. The data transmission correction factor is multiplied by the data packaging reference window to obtain the data packaging correction window, which is recorded as the data packaging set window.

[0044] Specifically, the frequency correction factor is obtained by taking the difference between the return time of the transmission heartbeat and the minimum value of the permitted return time interval. The return time deviation of the transmission heartbeat is obtained by processing the difference between the return time deviation and the frequency correction factor. The frequency correction factor is obtained by matching the correlation mapping relationship set between the return time deviation and the data transmission correction factor.

[0045] If the data transmission status shows abnormal transmission, the data transmission definition frequency is recorded as the data transmission setting frequency, the data packaging definition window is recorded as the data packaging setting window, and the monitoring sensor only uploads the primary key value corresponding to the multi-source sensing data.

[0046] Under both normal and weak network transmission conditions, the local edge gateway implements backpressure control for multi-source data: when the outbound bandwidth is detected to be lower than the threshold or the queue level / round-trip delay / packet loss rate exceeds the limit, the gateway immediately sends instructions to the device to reduce the sending window and the sampling / return frequency, and rearranges the traffic according to the importance of the data, giving priority to the passage of structured summaries and key feature values ​​(such as aggregated statistics, alarm-related triples, timestamps and quality markers), and delaying or pausing the original high-frequency / large-volume data (such as full time series, images / videos, diagnostic logs); after the bandwidth is restored and stabilized for a period of observation, the original window and frequency are gradually restored according to the predetermined gradual relaxation curve, while recording receipts and queue indicators for adaptive calibration of thresholds and policies.

[0047] The temporal feature prediction module is used to abstract the monitoring area into a graph neural network. The influence weights of the graph neural network edges are limited by physical associations. The temporal features are analyzed in combination with the influence weights. The graph neural network outputs the prediction results under the temporal feature window of each node.

[0048] The monitoring area is abstracted as a graph neural network (GNN), where each node represents a monitoring unit (such as a geological monitoring station or grid cell), and the edges between nodes represent the physical relationships between these units (such as water flow, seismic wave propagation, and precipitation impact). The influence weights of the graph neural network edges are defined by combining these physical relationships; that is, the weights of the edges are set according to the geographical, geological, or environmental characteristics of neighboring nodes to characterize the degree of influence between different nodes. For example, the influence of a heavy precipitation area may have a greater weight on nodes along a river system, while its influence is smaller in areas far from the river system. The graph neural network analyzes temporal characteristics through these influence weights of physical relationships, that is, it uses the historical temporal data of neighboring nodes and their weights to predict the future temporal sequence of the target node.

[0049] Specifically, by combining the temporal features analyzed by the influence weights, the graph neural network outputs the prediction results under the temporal feature window of each node. The specific analysis process is as follows: The specific analysis process affecting the weights is as follows: Each monitoring station in the monitoring area is abstracted as a node in a graph neural network, and the connection between adjacent monitoring stations through the road network is abstracted as an edge in a graph neural network.

[0050] Physical association data of adjacent nodes are extracted, including directional waterway reachability indicators of adjacent nodes, dominant confluence ratio of adjacent nodes, co-occurrence rate of adjacent nodes, and slope-land cover similarity of adjacent nodes, to determine the physical association assessment parameters of adjacent nodes.

[0051] Based on the mapping set between physical association evaluation parameters and edge influence weights, the physical association evaluation parameters of adjacent nodes are brought into the mapping set to obtain the edge influence weights.

[0052] By analyzing temporal features through a graph neural network, the spatial relationships between nodes are learned, and the prediction results under the temporal feature window are output by combining the edge influence weights.

[0053] Graph neural networks (GNNs) first take the edge weights of the graph and temporal data as input. Each node's temporal data is fused with the temporal features of its neighboring nodes through the GNN's message passing mechanism, using weighted summation to transmit and update information. The influence weights in this process regulate the intensity of information transmission between neighboring nodes, ensuring that nodes with strong physical connections have a greater predictive impact on the target node. The network continuously aggregates information from its neighborhood and updates node features through multi-layer graph convolution operations, ultimately outputting two key temporal prediction results at each node: the first component of the risk trend probability (representing the probability of a disaster occurring at that node in the near future) and the first component of the risk trend spread probability (representing the probability of a disaster spreading from that node to neighboring nodes in the near future). These two results provide information on the dynamic changes and spatial propagation of risk at each node within the temporal feature window, providing timely early warning and response strategies for the decision-making system.

[0054] The prediction results under the time series feature window of each node include the first component of the risk trend probability of each node and the first component of the risk trend spread probability of each node.

[0055] Furthermore, the physical association parameters of adjacent nodes are determined, and the specific analysis process is as follows: The topographic elevation grid of the monitoring area is obtained. The topographic elevation grid can be downloaded from a publicly available satellite digital elevation model. Based on the topographic elevation grid, the water system of adjacent nodes is extracted. If there is a directed confluence path from node i to adjacent node j, the directed waterway reachability indicator of the adjacent node is recorded as 1, otherwise it is recorded as 0.

[0056] The method involves extracting water systems from adjacent nodes based on a topographic elevation raster. Using a topographic elevation raster (DEM) as input, the method first fills depressions to eliminate digital pits, calculates the flow direction to determine which cell the water flows to for each cell, calculates the cumulative flow to obtain the upstream catchment area for each cell, and binarizes the water system line according to a set threshold (such as upstream area or cumulative number of cells). It can also calculate the river order. Based on the spatial location of adjacent monitoring nodes, the nodes are mapped / attached to the nearest water system line, and the distance / travel time along the waterway and the upstream and downstream reachability are calculated on the water system line. Based on this, a directed water system adjacency relationship is constructed between adjacent nodes (who is downstream of whom, whether they are the same tributary / main stream). It can further divide the watershed into sub-basins and statistically analyze the along-path attributes (slope, order, catchment area, etc.) of each pair of adjacent nodes to form the hydrological priors required for subsequent edge weight calculation.

[0057] A water system refers to a directional drainage network formed on the terrain by surface runoff following the law of water flowing downhill. It includes the main channels and tributaries of rivers / valleys, their connection relationships, and the corresponding sub-basin boundaries. In terms of data structure, it can be regarded as a directed graph (nodes are sinks / intersections and monitoring nodes, and edges are connected segments along the waterway), used to depict the physical relationships such as hydraulic accessibility, transmission intensity, and arrival delay from upstream to downstream.

[0058] Based on the water system extraction of neighboring nodes, the shortest hydraulic path from the current node i to the neighboring node j is extracted. Based on the shortest hydraulic path, the set of current node i pixels that can flow into the shortest hydraulic path and eventually reach the neighboring node j is calculated. Based on the set of current node i pixels, the dominant flow ratio of neighboring nodes is obtained.

[0059] After extracting the river system using the DEM, the shortest hydraulic path P from the current node i to the adjacent node j is first found in the directed river network. ij (Prioritizing minimum travel time, approximating with the largest catchment trunk when parameters are missing); then performing reverse confluence tracing along this path: using P ij Using the upstream river segment as the framework, all upstream pixels that flow into this path and eventually reach j are traced back according to the flow direction grid to obtain the set Ω. ij Then, the water catchment polygon Ω of node i. i Taking the intersection, the set of pixels is Ω. i ∩Ω ij Therefore, the dominant confluence ratio can be calculated: C ij =Area(Ω) i ∩Ω ij ) / Area(Ω i ); where Area(·) refers to the geometric area.

[0060] The data preprocessing includes normalization and de-unitization. The data preprocessing includes the directional waterway reachability indication of adjacent nodes, the dominant confluence ratio of adjacent nodes, the co-occurrence rate of adjacent nodes, and the slope-land cover similarity of adjacent nodes.

[0061] By introducing weight elements and associating and merging them with the data preprocessing results, the physical association evaluation parameters of adjacent nodes are determined.

[0062] The specific analysis process is as follows: In the formula, W ij The physical association evaluation parameter for adjacent nodes is , where i is the current node's number, j is the adjacent node's number, and Cr... ij Cs is a directed waterway reachability indicator for adjacent nodes. ij P represents the dominant flow ratio of adjacent nodes. ij Es represents the co-occurrence rate of adjacent nodes within the same window. ij The slope-land cover similarity between adjacent nodes. For the predefined directed waterway reachability indicators in the geological data management database, the corresponding weight elements are... The weight element corresponding to the predefined dominant confluence ratio in the geological data management database. The weight element corresponding to the predefined co-occurrence rate in the geological data management database. The weight element is a predefined slope-land cover similarity value in the geological data management database.

[0063] The co-occurrence rate within a given time window is expressed as the frequency with which two adjacent nodes both experience a risk event within the same window. A risk event is a definite time / interval that will be included in the co-occurrence statistics or trigger an early warning calculation; it can be either a sensor reading exceeding a threshold or a risk threshold given by the model.

[0064] The slope-land cover similarity between adjacent nodes refers to the degree of similarity between two adjacent nodes in terms of slope morphology and land cover characteristics, reflecting the consistency between their instability mechanism and runoff and sediment production behavior.

[0065] In this embodiment, multivariate analysis specifically considers the correlation between these parameters. If there is a directional waterway reachability indicator (1 / 0), it acts as a gate for the correlation degree. When it is not reachable, the overall correlation degree is forcibly reduced to zero or significantly reduced. Under the premise of reachability, the larger the proportion of dominant confluence, the more upstream runoff can reach the downstream along the main channel, the correlation degree is monotonically enhanced, and it often works together with the travel time weight. The higher the co-occurrence rate in the same window, the more likely the two nodes are to rise and fall together in the same time window. Even after controlling for the combined driving force of seasonality and rainstorms, it still enhances the correlation degree. The higher the slope-land cover similarity, the more consistent the instability mechanism and runoff and sediment production behavior of the two locations are. It can be used as a prior amplification factor to improve the effectiveness of the first two items.

[0066] Specifically, the analysis process for the time series feature window is as follows: The average value of the data transmission set frequency of each monitoring sensor is extracted and processed to obtain the average value of the data transmission set frequency of the monitoring sensor. This average value is then matched with the data sampling step size corresponding to the predefined data transmission set frequency average value interval to determine the data sampling step size of the monitoring sensor.

[0067] Obtain the minimum number of samples for the graph convolutional network. The minimum number of samples is extracted from the initial setting record of the graph convolutional network and multiplied by the step size corresponding to the data sampling of the monitoring sensor. This result is recorded as the lower limit of the temporal feature window.

[0068] The soil moisture at the current time point is compared with the soil moisture at the previous time point. The soil moisture can be extracted from the monitoring records of the monitoring sensors to obtain the soil moisture fluctuation coefficient of the node. The standard deviation of the soil moisture fluctuation coefficient of each node is processed to obtain and record the abnormal fluctuation coefficient of soil moisture in the monitoring area.

[0069] Based on the correlation mapping set between the abnormal fluctuation coefficient and the upper limit of the window, the abnormal fluctuation coefficient of soil moisture in the monitoring area is brought into the correlation mapping set to obtain the upper limit of the window, which is denoted as the upper limit of the time series feature window.

[0070] The trend feature prediction module is used to extract historical monitoring data of the monitoring area, and based on the time series features and the prediction results under the time series feature window, it is input into the time series prediction model to output the prediction results under the trend feature window of each node.

[0071] The time-series prediction model first extracts historical monitoring data of the monitored area and combines it with the prediction results under the time-series features and time-series feature windows as input. This input data includes historical time-series data for each node, observations within the time-series feature window, and the first component of the risk trend probability and the first component of the spread probability output by the graph neural network. This data is processed by a time-series prediction model (such as LSTM). The model captures historical trends and short-term changes through the recursive characteristics of time series and makes predictions for future periods. Through learning over multiple time steps, the model progressively extracts the second component of the risk trend probability (representing the probability of a disaster occurring at the node within a certain future time) and the second component of the risk trend spread probability (representing the probability of a disaster spreading from that node to adjacent nodes within a certain future time) for each node in the future period. Through training and optimization of the time-series prediction model, more accurate disaster occurrence and spread probabilities can be output, providing more precise spatiotemporal dynamic predictions for disaster early warning.

[0072] Furthermore, the trend feature window, specifically the analysis process is as follows: Based on the prediction results under the time series feature window of each node, the first component of the risk trend probability of each node is compared and calculated with the predefined threshold of the first component of the risk trend probability, and the first adjustment factor of the window is obtained by matching, which is denoted as the first adjustment factor of the trend feature window.

[0073] Specifically, the difference between the first component of the risk trend probability of each node and the threshold of the first component of the risk trend probability is processed to obtain the margin of the first component of the risk trend probability of each node. This margin is then matched with the first adjustment factor of the window corresponding to the predefined margin interval of the first component of the risk trend probability of each node to determine the interval to which the margin of the first component of the risk trend probability of each node belongs, and the first adjustment factor of the window corresponding to that interval is obtained.

[0074] The first component of the risk trend spread probability of each node is compared and calculated with the predefined threshold of the first component of the risk trend spread probability, and the second adjustment factor of the window is obtained by matching. This is denoted as the second adjustment factor of the trend feature window.

[0075] Specifically, the difference between the first component of the risk trend spread probability of each node and the threshold of the first component of the risk trend spread probability is processed to obtain the margin of the first component of the risk trend spread probability of each node. This margin is then matched with the predefined threshold range of the first component of the risk trend spread probability to determine the range to which the margin of the first component of the risk trend spread probability of each node belongs, and the second adjustment factor of the window corresponding to that range is obtained.

[0076] The trend feature window is limited by the first adjustment factor and the second adjustment factor of the trend feature window.

[0077] Specifically, the second adjustment factor of the trend feature window is multiplied by the upper limit of the initial trend feature window to obtain and record the upper limit of the trend feature window, and the first adjustment factor of the trend feature window is multiplied by the lower limit of the initial trend feature window to obtain and record the lower limit of the trend feature window.

[0078] The risk trend probability fusion module is used to fuse the prediction results of the reinforcement learning model based on the time-series feature window and the prediction results based on the trend feature window to obtain the comprehensive risk trend probability and probability confidence of the monitoring data of each node.

[0079] The reinforcement learning model fuses the prediction results under the time-series feature window (the first component of risk trend probability and the first component of contagion probability) with the prediction results under the trend feature window (the second component of risk trend probability and the second component of contagion probability). Through methods such as weighted averaging or evidence-based fusion, the two prediction results are combined to output the comprehensive risk trend probability and probability confidence score of each node's monitoring data. During the fusion process, the reinforcement learning model assigns different weights to each prediction result based on the historical prediction accuracy of each node, the credibility of the time-series data, and the physical correlation weights, thereby achieving more accurate risk prediction. Ultimately, the comprehensive risk trend probability and probability confidence score output by the model can provide multi-dimensional information for disaster early warning, not only accurately predicting risks but also indicating the credibility of the prediction results, enabling decision-makers to make more reasonable responses.

[0080] It's important to explain that the necessity of the two predictions lies in the complementarity of time-series and trend features: time-series features reflect the short-term trends of each node, helping to capture sudden abnormal fluctuations, while trend features focus on regular changes over a longer timescale, providing long-term risk assessment. Combining these two types of features can compensate for the limitations of single-feature predictions, enhancing the system's adaptability to complex and dynamic environments. By fusing the two predictions, the reinforcement learning model not only improves prediction accuracy and reduces false positives and false negatives, but also provides risk predictions with probabilistic confidence levels, offering more reliable support for emergency response and decision-making, thereby improving the response efficiency and reliability of the disaster early warning system.

[0081] The early warning device linkage module is used to rank the risk probability based on the comprehensive risk trend probability and probability confidence of the monitoring data of each node. The geological disaster intelligent prevention terminal completes the geological disaster early warning through the risk probability ranking and differentiated linkage early warning device.

[0082] Specifically, the intelligent geological disaster prevention terminal completes geological disaster early warning through a risk probability ranking and differentiated linkage early warning device. The specific analysis process is as follows: Extract the comprehensive risk trend probability from the monitoring data of each node, and arrange the comprehensive risk trend probabilities of the monitoring data of each node in descending order to obtain the initial sequence of comprehensive risk trend probabilities.

[0083] Extract the probability confidence level of the monitoring data of each node. Based on the association mapping set between the probability confidence level and the sequence influence parameter, substitute the probability confidence level of the monitoring data of each node into the association mapping set to obtain the sequence influence parameter, which is denoted as the sequence influence parameter of each node.

[0084] Multiply the sequence influence parameter of each node by the corresponding ordinal number of the initial sequence of the comprehensive risk trend probability to obtain the corrected ordinal number of the initial sequence of the comprehensive risk trend probability. Based on the corrected ordinal number of the initial sequence of the comprehensive risk trend probability, sort them in descending order to form the comprehensive risk trend probability sequence.

[0085] Based on the comprehensive risk trend probability sequence, a predetermined number of nodes corresponding to the comprehensive risk trend probability are extracted and denoted as the early warning node set. The geological disaster intelligent prevention terminal links with the early warning devices belonging to the early warning node set to issue geological disaster early warnings. The early warning device includes a signal probe component, a safety component, an alarm component, an interaction component, and a communication component. The geological disaster intelligent prevention terminal issues execution instructions to each early warning device.

[0086] After all early warning devices execute the corresponding instructions, they send back execution status data to the geological disaster intelligent prevention terminal. The geological disaster intelligent prevention terminal then adjusts the corresponding execution instruction issuance process based on the sent-back execution status data.

[0087] After receiving feedback from various early warning devices, the intelligent geological disaster prevention terminal adaptively adjusts the data transmission process using a three-dimensional closed-loop system encompassing status, data link, and risk. A detailed flowchart is shown below. Figure 4 As shown, Figure 4To implement the closed-loop adaptive adjustment flowchart for command execution: For devices that have successfully executed commands, reduce the frequency of repeated transmissions and enable deduplication / idempotency marking; for unacknowledged / timeout devices, increase priority, switch to backup links, downgrade from rich commands to minimum available messages, and enable fragmentation combined with ACK retransmission and exponential backoff; for execution failures with identifiable causes (low battery, busy, sensor malfunction), change to delayed retry / batch merging or reassign to same-domain backup devices (geographically proximate / functionally redundant), and issue local offline emergency actions; for devices with link jitter, tighten the transmission window and frequency, and resume transmission after bandwidth recovery; for targets that fail multiple times, trigger manual / command confirmation and multi-channel linkage (such as simultaneously driving alarms and intercoms), and dynamically update device health scores and policy whitelists so that subsequent scheduling rearranges queues and throttling thresholds according to health, risk level, and geographical coverage, ensuring timely access to the highest risk points with minimal bandwidth.

[0088] Furthermore, the intelligent geological disaster prevention terminal issues execution commands to each early warning device, including configuration for issuing execution commands under weak network conditions: The intelligent geological disaster prevention terminal issues each execution command according to the command priority type. The process of setting the command priority type is as follows: the command priority type is obtained by mapping the comprehensive risk trend probability of node monitoring data, and it is recorded as the command priority type.

[0089] Instruction priority types include P0 instruction priority, P1 instruction priority, P2 instruction priority, and P3 instruction priority.

[0090] The communication component of the geological disaster intelligent prevention terminal maintains the priority of P0 commands, which are always issued first; P1 commands are compressed into the minimum available message first; and P2 and P3 commands are placed in the backup queue for delayed issuance.

[0091] If the network is completely unavailable during the instruction issuance process of the geological disaster intelligent prevention terminal, the geological disaster intelligent prevention terminal will execute the pre-stored local emergency instructions on the early warning device. After the instructions are executed locally, the execution status will be returned. The returned data will be returned according to the priority of the instruction type. The geological disaster intelligent prevention terminal will confirm the execution result based on the returned execution status.

[0092] The second aspect of the present invention provides a geological disaster early warning device based on multimodal data fusion, comprising: a monitoring sensor, a geological disaster intelligent prevention terminal, a signal probe component, a safety component, an alarm component, an interaction component, and a communication component.

[0093] Monitoring sensors are used to monitor geological disaster monitoring areas and form multi-source sensing datasets.

[0094] The intelligent geological disaster prevention terminal is used to generate and execute instructions based on the probability of comprehensive risk trends, and to link early warning devices to provide geological disaster warnings.

[0095] The terminal aggregates satellite remote sensing, ground sensor, and meteorological data, and performs real-time analysis using a geological disaster AI model based on DeepSeek meteorological coupling. Voice interaction supports spoken queries and command issuance. After the model outputs the comprehensive risk trend probability and probability confidence level, the terminal automatically generates execution commands according to the warning level (blue, yellow, orange, red), triggering and confirming warning devices such as sirens, safety helmet reminders, walkie-talkie commands, and BeiDou / SMS broadcasts. In weak network scenarios, it prioritizes sending the minimum available message to ensure timeliness. Simultaneously, the system overlays basic geographic and population exposure information with the spatial distribution and propagation trend of risks, automatically generating scenario-based emergency recommendations, including tiered prevention and control actions, pre-deployment lists of supplies, key location verification tasks, and visualized evacuation routes, providing them to the command center for one-click issuance and tracking. The overall architecture achieves a closed loop of multi-source perception, intelligent fusion, hierarchical linkage, and explainable backtracking: it can not only carry out continuous monitoring and rolling early warning for potential hazards, and promptly remind departments and the public to take risk avoidance measures, but also quickly form emergency plans and evacuation routes based on AI analysis results in the event of an emergency, assisting decision-makers to complete the response efficiently and controllably.

[0096] The signal probe component is used to passively scan and collect the strength and frequency of unknown wireless signals (such as WIFI / BLE / weak cellular beacons). For example, a WIFI probe outputs rasterized signal distribution data and transmits it back to the geological disaster intelligent prevention terminal. Based on this, the terminal generates a signal heat map and determines the area of ​​suspected human activity, focusing on searching for detectable WIFI signals in dangerous areas, which serves as important evidence that someone is in a dangerous area. The component supports multi-point collaborative measurement and time difference of arrival / received signal strength positioning to improve spatial resolution. In the case of weak network, only the minimum usable digest (timestamp, location, signal strength quantile, probability confidence) is uploaded, and privacy is protected by anonymized / hashable device identifiers, thereby providing real-time and low-cost population distribution perception capabilities for emergency evacuation and rescue priority.

[0097] Safety components used to provide vibration or voice alerts, such as safety helmets.

[0098] Alarm components are used to issue audible and visual warnings based on risk alert instructions. Examples include alarm systems.

[0099] Interactive components are used to issue voice commands to frontline personnel to guide evacuation. Examples include walkie-talkies, which provide real-time two-way voice communication and operate on their own radio frequencies, ensuring stable communication even in areas without network access (such as outdoors, tunnels, and remote areas).

[0100] The communication component is used to send execution commands in weak network conditions. For example, with BeiDou communication, when there is a network but it is unstable, it adopts multi-link adaptive (cellular / LoRa / NB-IoT priority, automatic fallback to BeiDou / satellite short message), enables priority queues and minimum available message (MAM) format, compresses and fragments execution commands, and includes HMAC checksum and ACK retransmission. When WIFI and mobile networks are interrupted, the component switches to an emergency channel (dedicated radio / direct walkie-talkie connection, relay network or satellite short message) to maintain low-bandwidth reliable delivery of personal safety commands, and enables store-and-forward and breakpoint resume recovery locally. End-to-end encryption and timestamp / quality tag (Q_time) are carried in the packet throughout the process to ensure that the commands are traceable and verifiable, and at the same time, the link health status and receipt are fed back to the smart defense terminal to support continuous linkage and controllable degradation of emergency command under extreme conditions.

[0101] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A geological disaster early warning system based on multi-modal data fusion, characterized in that, The method comprises the following steps: a time sequence feature extraction module is used to extract a multi-source perception data set of a geological disaster monitoring area in real time, and the data set is transmitted to a local edge gateway for time stamp synchronization and time sequence feature extraction, so as to obtain time sequence features of the monitoring area; a time sequence feature prediction module is used to abstract the monitoring area into a graph neural network, limit the influence weight of the edge of the graph neural network through physical correlation, analyze the time sequence features in combination with the influence weight, and output the prediction result of each node in the time sequence feature window by the graph neural network; a trend feature prediction module is used to extract historical monitoring data of the monitoring area, input the time sequence features and the prediction result in the time sequence feature window into a time sequence prediction model, and output the prediction result of each node in the trend feature window; a risk trend probability fusion module is used to fuse the prediction result in the time sequence feature window and the prediction result in the trend feature window by a reinforcement learning model, so as to obtain the comprehensive risk trend probability and probability confidence of the monitoring data of each node; an early warning device linkage module is used to sort the risk probability based on the comprehensive risk trend probability and probability confidence of the monitoring data of each node, and a geological disaster intelligent prevention terminal differentiates the linkage of early warning devices through the risk probability sorting difference, so as to complete the geological disaster early warning.

2. The geological disaster early warning system based on multi-modal data fusion according to claim 1, characterized in that: The local edge gateway is used to synchronize the time stamp and extract the time sequence features, and the specific analysis process is as follows: The multi-source perception data set is collected based on each monitoring sensor of the geological disaster monitoring area; Each monitoring sensor of the geological disaster monitoring area transmits the multi-source perception data to the local edge gateway based on the data transmission setting frequency and the data packaging setting window; The collection time stamp of each monitoring sensor of the geological disaster monitoring area and the receiving time stamp of the local edge gateway are obtained, and the collection time stamp of the monitoring sensor and the receiving time stamp of the local edge gateway are compared and calculated, and the time stamp synchronization process of the monitoring sensor is configured based on the comparison result; After the time stamp synchronization, the multi-source perception data set is decomposed by STL to obtain the time sequence features of the monitoring area.

3. The geological disaster early warning system based on multi-modal data fusion according to claim 2, characterized in that: The specific analysis process of the data transmission setting frequency and the data packaging setting window is as follows: A heartbeat probe is configured between the monitoring sensor and the local edge gateway, which is denoted as transmission heartbeat; The return time length of the transmission heartbeat is obtained, and the data transmission state between the monitoring sensor and the local edge gateway is determined based on the return time length of the transmission heartbeat; The data transmission state includes normal transmission, weak network transmission and abnormal transmission; If the data transmission state shows normal transmission, the data transmission reference frequency is denoted as the data transmission setting frequency, and the data packaging reference window is denoted as the data packaging setting window; If the data transmission state shows weak network transmission, the data transmission reference frequency is multiplied by the frequency correction factor to obtain the data transmission correction frequency, which is denoted as the data transmission setting frequency, and the data transmission correction factor is multiplied by the data packaging reference window to obtain the data packaging correction window, which is denoted as the data packaging setting window; If the data transmission state shows abnormal transmission, the data transmission defined frequency is denoted as the data transmission setting frequency, and the data packaging defined window is denoted as the data packaging setting window, and the monitoring sensor only uploads the primary key value corresponding to the multi-source perception data.

4. The geological disaster early warning system based on multi-modal data fusion according to claim 1, characterized in that: The graph neural network combines the influence weights to analyze the temporal features, and outputs the prediction results under the temporal feature window of each node. The specific analysis process is as follows: The specific analysis process for the influencing weights is as follows: Each monitoring station in the monitoring area is abstracted as a node in a graph neural network, and the connection between adjacent monitoring stations through the road network is abstracted as an edge in a graph neural network. Extract physical association data of adjacent nodes, including directional waterway reachability indicators of adjacent nodes, dominant confluence ratio of adjacent nodes, co-occurrence rate of adjacent nodes, and slope-land cover similarity of adjacent nodes, and determine the physical association assessment parameters of adjacent nodes. Based on the mapping set between physical association evaluation parameters and edge influence weights, the physical association evaluation parameters of adjacent nodes are brought into the mapping set to obtain the edge influence weights. The spatial relationships between nodes are learned by analyzing temporal features through graph convolutional networks, and the prediction results under the temporal feature window are output by combining edge influence weights. The prediction results under the time-series feature window of each node include the first component of the risk trend probability of each node and the first component of the risk trend propagation probability of each node.

5. The geological disaster early warning system based on multi-modal data fusion according to claim 4, characterized in that: The specific analysis process for the physical association evaluation parameters used to determine adjacent nodes is as follows: Obtain the topographic elevation grid of the monitoring area, extract the water system of adjacent nodes based on the topographic elevation grid, and if there is a directed confluence path from node i to adjacent node j, record the directed waterway reachability indicator of the adjacent node as 1, otherwise record it as 0. Based on the water system extraction of neighboring nodes, the shortest hydraulic path from the current node i to the neighboring node j is extracted. Based on the shortest hydraulic path, the set of current node i pixels that can flow into the shortest hydraulic path and eventually reach the neighboring node j is calculated. Based on the set of current node i pixels, the dominant flow ratio of neighboring nodes is obtained. The data preprocessing includes normalization and de-unitization. The data preprocessing includes the directional waterway reachability indication of adjacent nodes, the dominant confluence ratio of adjacent nodes, the co-occurrence rate of adjacent nodes, and the slope-land cover similarity of adjacent nodes. By introducing weight elements and associating and merging them with the data preprocessing results, the physical association evaluation parameters of adjacent nodes are determined.

6. The geological disaster early warning system based on multi-modal data fusion according to claim 4, characterized in that: The specific analysis process for the time-series feature window is as follows: The average value of the data transmission set frequency of each monitoring sensor is extracted and processed to obtain the average value of the data transmission set frequency of the monitoring sensor. This average value is then matched with the data sampling step size corresponding to the predefined data transmission set frequency average value interval to determine the data sampling step size of the monitoring sensor. The minimum number of samples for the graph convolutional network is obtained and multiplied by the step size corresponding to the data sampling of the monitoring sensor to obtain and record as the lower limit of the temporal feature window; The soil moisture of the node at the current time point is compared with the soil moisture of the node at the previous time point to obtain the soil moisture fluctuation coefficient of the node. The standard deviation of the soil moisture fluctuation coefficient of each node is processed to obtain and record the abnormal fluctuation coefficient of soil moisture in the monitoring area. Based on the correlation mapping set between the abnormal fluctuation coefficient and the upper limit of the window, the abnormal fluctuation coefficient of soil moisture in the monitoring area is brought into the correlation mapping set to obtain the upper limit of the window, which is denoted as the upper limit of the time series feature window.

7. The geological disaster early warning system based on multi-modal data fusion according to claim 1, characterized in that: The specific analysis process for the trend feature window is as follows: Based on the prediction results under the time series feature window of each node, the first component of the risk trend probability of each node is compared and calculated with the predefined threshold of the first component of the risk trend probability, and the first adjustment factor of the window is obtained by matching, which is denoted as the first adjustment factor of the trend feature window. The first component of the risk trend spread probability of each node is compared and calculated with the predefined threshold of the first component of the risk trend spread probability, and the second adjustment factor of the window is obtained by matching, which is denoted as the second adjustment factor of the trend feature window. The trend feature window is limited by the first adjustment factor and the second adjustment factor of the trend feature window.

8. The geological disaster early warning system based on multi-modal data fusion according to claim 1, characterized in that: The intelligent geological disaster prevention terminal completes geological disaster early warning through a risk probability ranking and differentiated linkage early warning device. The specific analysis process is as follows: Based on the comprehensive risk trend probability sequence, a predetermined number of nodes corresponding to the comprehensive risk trend probability are extracted and denoted as the early warning node set. The geological disaster intelligent prevention terminal links with the early warning devices belonging to the early warning node set to issue geological disaster early warnings. The early warning device includes a signal probe component, a safety component, an alarm component, an interaction component, and a communication component. The geological disaster intelligent prevention terminal issues execution instructions to each early warning device. After all early warning devices execute the corresponding instructions, they send back execution status data to the geological disaster intelligent prevention terminal. The geological disaster intelligent prevention terminal then adjusts the corresponding execution instruction issuance process based on the sent-back execution status data.

9. The geological disaster early warning system based on multi-modal data fusion according to claim 8, characterized in that: The intelligent geological disaster prevention terminal issues execution commands to each early warning device, and also includes configuration for issuing execution commands under weak network conditions. The intelligent geological disaster prevention terminal will issue commands according to the priority type set in the command settings: The process of setting the priority type of the instruction is as follows: the instruction priority type is obtained based on the comprehensive risk trend probability mapping of the node monitoring data, and is denoted as the instruction setting priority type. The instruction priority types include P0 instruction priority, P1 instruction priority, P2 instruction priority, and P3 instruction priority. The geological disaster intelligent prevention terminal access communication component maintains the priority of P0 commands, always sending them first; P1 commands are compressed into the minimum available message first; and P2 and P3 commands are entered into the backup queue for delayed sending. If the network is completely unavailable during the instruction issuance process of the geological disaster intelligent prevention terminal, the geological disaster intelligent prevention terminal will execute the pre-stored local emergency instructions on the early warning device. After the instructions are executed locally, the execution status will be returned. The returned data will be returned according to the priority of the instruction type. The geological disaster intelligent prevention terminal will confirm the execution result based on the returned execution status.

10. A geological disaster early warning device based on multi-modal data fusion, applied to the geological disaster early warning system based on multi-modal data fusion according to any one of claims 1-9, characterized in that: include: Monitoring sensors, intelligent geological disaster prevention terminals, signal probe components, safety components, alarm components, interactive components, and communication components; The monitoring sensors are used to monitor geological disaster monitoring areas and form a multi-source sensing dataset; The intelligent geological disaster prevention terminal is used to generate execution commands based on the probability of comprehensive risk trends and to link with early warning devices to provide geological disaster early warning. The signal probe component is used to monitor unknown signals in the geological disaster monitoring area and provide signal distribution data for the intelligent geological disaster prevention terminal. The safety component is used to provide vibration or voice alerts; The alarm component is used to issue audible and visual warnings based on risk alert instructions; The interactive component is used to issue voice commands to frontline personnel to guide evacuation; The communication component is used to send execution instructions in a weak network environment.

Images