Geological disaster real-time monitoring system based on Internet of Things

Abstract

The invention discloses a geological disaster real-time monitoring system based on the Internet of Things, and relates to the technical field of geological disaster intelligent early warning. Original data such as earth surface displacement, rainfall and earth sound signals are acquired through a multi-source sensing acquisition module; the edge intelligent processing module extracts disturbance characteristics and calculates disturbance entropy and an energy index; constructing a regional topological graph by using a graph neural network, and predicting a propagation path and a risk probability of an induction factor; fusing meteorological prediction and topographic factors to construct a multi-dimensional mapping model, and identifying a disaster development trend; judging whether a critical state is entered or not based on a causal reasoning map and triggering early warning; the sampling frequency and the communication priority of the sensor are dynamically adjusted in combination with the historical anomaly similarity, and resource optimization configuration is achieved; the system has the advantages of high recognition precision, high response speed, low energy consumption and high deployment adaptability, and is suitable for intelligent monitoring and early warning of geological disasters such as landslide and debris flow.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster intelligent early warning, in particular to a geological disaster real-time monitoring system based on Internet of Things. BACKGROUND

[0002] At present, geological disasters (such as landslides, collapses, debris flows, etc.) frequently occur in complex terrain and inconvenient transportation areas such as high mountains and valleys, mine area slopes, and earthquake belts. These areas often have characteristics such as few people, weak infrastructure, and long response time. Traditional geological disaster monitoring methods rely on a limited number of manual patrols, simple sensor deployment (such as inclinometers, rain gauges, crack meters, etc.), and regular data collection, lacking real-time sensing capabilities and spatial dynamic correlation modeling of disaster triggering processes, making it difficult to timely detect disaster precursors and predict disaster development paths, resulting in response lag, high false alarm rate, and extensive coverage blind spots.

[0003] With the development of Internet of Things technology, Low Power Wide Area Network (LPWAN), edge computing, and sensor ad hoc network technology have been gradually introduced into geological monitoring. Although the frequency and regional coverage of data collection have been improved, there are still serious technical bottlenecks in multi-dimensional heterogeneous data fusion, weak disaster signal extraction, and causal modeling between disaster trigger factors. For example, in heavy rainfall-induced landslides, the non-linear correlation between rainfall intensity, soil moisture content, and surface displacement is complex, and existing models cannot identify potential dangerous areas in the early stages of disturbance. At the same time, edge nodes are limited by computing power and energy consumption, and cannot perform complex feature extraction and behavior prediction on real-time data, causing micro-change signals to be lost in background noise, severely restricting the accuracy and initiative of the geological disaster early warning system. SUMMARY

[0004] The purpose of the present application is to provide a geological disaster real-time monitoring system based on Internet of Things to solve the problems in the background art.

[0005] To achieve the above purpose, the present application provides the following technical solution: a geological disaster real-time monitoring system based on Internet of Things, comprising: A multi-source sensing acquisition module: deploying Internet of Things sensor network nodes in the target area to collect raw data including rainfall, ground displacement, soil moisture content, ground temperature, and ground sound signals; An edge intelligent processing module: sending the raw data to the edge node, which performs feature extraction based on local low-power algorithms to obtain a micro-variation anomaly signal vector in the time-frequency domain, and calculates signal disturbance entropy and instantaneous energy index; A regional risk modeling module: constructing a spatial topology graph between nodes, combining historical geological structure models and real-time sensing data, predicting the propagation path and influence probability of key trigger factors in the region through a graph neural network, and generating a disaster potential risk graph. Multi-source fusion analysis module: based on the risk map, the key node features of high-risk areas are extracted, further fusion of meteorological prediction data and DEM elevation data is carried out, and a multi-dimensional mapping relationship between disaster trigger factors and terrain response factors is established; Early warning module: a disaster evolution chain model is constructed by using the causal reasoning graph, whether the potential disaster enters the critical conversion stage is judged, and if it is judged as a critical state, an early warning information is sent; Resource optimization module: according to the similarity of the historical abnormal sequence of the monitoring point and the current disturbance mode, the sampling frequency and communication priority of the sensor are adaptively adjusted.

[0006] Preferably, the edge intelligent processing module comprises: The obtained original data packet is synchronized and calibrated according to the time stamp and space node identifier, and is divided into fixed-length analysis segments by using a sliding window segmentation method; The seismic signal and displacement signal in each analysis segment are subjected to fast Fourier transform and wavelet packet decomposition, and the energy distribution spectrum of the signal in the time-frequency domain is extracted; The disturbance entropy and instantaneous energy index of the micro-variation signal are calculated based on the energy distribution spectrum, and the potential abnormal disturbance event is identified.

[0007] Preferably, the disturbance entropy and instantaneous energy index of the micro-variation signal based on the energy distribution spectrum comprise: The energy distribution spectrum obtained by spectrum analysis is normalized to construct the energy probability density function in the corresponding time period; The Shannon disturbance entropy is calculated based on the probability density function, which is used to measure the energy uncertainty and disturbance intensity change of the signal in the frequency domain; The instantaneous power of the continuous segment signal is extracted in the time domain by using a sliding time window, and the instantaneous energy index is calculated according to the weighted average method.

[0008] Preferably, the regional risk modeling module comprises: Based on the spatial position information and monitoring factor type of each sensing node, a sensor spatial topology graph in the region is constructed, in which the node represents the sensor unit and the edge represents the geological correlation path; Fusion of historical geological structure model and current real-time sensing data, each node is assigned a multi-dimensional feature vector; The topology graph is trained by using a graph neural network to learn the propagation relationship of the geological induction factors between nodes, and the disaster influence probability of different nodes is predicted; The prediction result is mapped to a geographic information system to generate a regional disaster potential risk map.

[0009] Preferably, the topological graph is trained using a graph neural network to learn the propagation relationship of geological triggering factors between nodes, and to predict the disaster impact probability of different nodes, including: The constructed spatial topological graph is represented as a graph structure data G=(V,E,X), where V is a node set, E is an edge set, and X is a multi-dimensional geological feature matrix of the nodes; A graph convolutional neural network is selected as the core model, the adjacency matrix of X is normalized, and the node features are embedded and aggregated using inter-layer propagation rules; A cross-entropy loss function is constructed using the historical abnormal frequency of the nodes as the supervision label, and the graph neural network is supervised and trained to optimize the parameters iteratively; After training, the influence probability value of the geological triggering factor of each node is output, and the nodes are spatially sorted according to the probability.

[0010] Preferably, the multi-source fusion analysis module comprises: Based on the node set in the disaster potential risk map whose predicted probability is higher than the preset threshold, the corresponding key spatial position and its geological feature information are extracted; Receive rainfall intensity, wind speed and temperature change data in the future time period, and map them to the area where the high-risk nodes are located; Combine the digital elevation model to obtain the slope, slope direction and drainage flow direction topographic response factors of the high-risk area; Construct a multi-dimensional mapping model based on decision tree to learn the synergistic relationship between disaster triggering factors and topographic response factors.

[0011] Preferably, the warning module comprises: Based on the characteristics of high-risk nodes, a causal reasoning graph is constructed between disaster triggering factors and response factors, and the edge weight in the graph represents the conditional probability relationship between variables; According to the Bayesian network structure learning algorithm, the graph is parameter trained to obtain the probability path of the chain propagation of disaster triggering factors to the response state; Set the threshold condition of the critical transition stage of the disaster, when the joint probability value of a certain state node exceeds the preset threshold, it is determined to enter the critical state.

[0012] Preferably, the resource optimization module comprises: similarity comparison of the current disturbance feature vector sequence with the sample sequences in the historical abnormal sequence library, to obtain a normalized similarity score S; hierarchical judgment according to the similarity S and the preset threshold S_high, S_mid: when S >= S_high, it is judged as a high similarity state, the sampling frequency is raised from the reference frequency F_base to the high priority frequency F_high, and the node is assigned to the communication priority 1; when S_mid <= S < S_high, the sampling frequency is raised to the medium frequency F_mid, and the communication priority 2 is assigned; otherwise, F_base and priority 3 are maintained.

[0013] In the above technical solution, the technical effects and advantages provided by the present application are as follows: 1、The present application realizes the full-link intelligent closed loop from data acquisition, edge analysis, risk modeling to early warning response by constructing a geological disaster monitoring system that integrates multi-source perception, graph neural network analysis, causal reasoning and resource optimization. By deploying low-power feature extraction algorithms on edge nodes, the system can identify early disturbance micro-signal and respond quickly, significantly improving the accuracy of disaster precursor identification and the real-time of response, effectively avoiding the problems of traditional systems such as heavy dependence on central computing and response lag.

[0014] 2、The present application further introduces graph neural network and Bayesian causal reasoning model to dynamically capture the propagation path of disaster inducing factors in the region and its probability deduction process, realizing the visual prediction of disaster evolution chain. At the same time, the resource optimization module dynamically adjusts the sensor sampling and communication strategy based on the historical abnormal similarity, greatly reducing the system power consumption and communication burden while ensuring the monitoring reliability, making the present application have stronger stability, self-adaptability and long-term operation ability in the field, suitable for risk monitoring and early warning needs in various complex geological environments. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0016] Figure 1 The system module flowchart of the present application. DETAILED DESCRIPTION

[0017] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0018] Embodiments, please refer to Figure 1 The geological disaster real-time monitoring system based on the Internet of Things comprises: The multi-source sensing acquisition module: the Internet of Things sensing network nodes are arranged in the target area to collect original data including rainfall, ground displacement, soil moisture content, ground temperature and ground sound signals; The edge intelligent processing module: the original data are sent to the edge node, the edge node performs feature extraction based on a local low-power algorithm, obtains a time-frequency domain micro-variation abnormal signal vector, and calculates signal disturbance entropy and instantaneous energy index; The regional risk modeling module: a spatial topological graph between nodes is constructed, historical geological structure models and real-time sensing data are combined, a graph neural network is used to predict the propagation path and influence probability of key induced factors in the region, and a disaster potential risk graph is generated; The multi-source fusion analysis module: key node features of a high-risk area are extracted based on the risk graph, meteorological prediction data and DEM elevation data are further fused, and a multi-dimensional mapping relationship between disaster triggering factors and terrain response factors is established; The early warning module: a disaster evolution chain model is constructed by using a causal reasoning graph, whether a potential disaster enters a critical conversion stage is judged, and if the critical state is judged, early warning information is sent; The resource optimization module: according to the similarity of the historical abnormal sequence of the monitoring point and the current disturbance mode, the sampling frequency and the communication priority of the sensor are adaptively adjusted.

[0019] The multi-source sensing acquisition module provided in the embodiments of the present application is applied to a geological disaster real-time monitoring method based on the Internet of Things, and is used for realizing high-frequency and real-time collection of multi-dimensional geological parameters of a target area. First, the Internet of Things sensing nodes are arranged in a geological disaster risk area to be monitored. The nodes are spatially optimized according to the topographic features, historical disaster distribution and potential disaster evolution path. The nodes can form a sensing network supporting self-organization, self-healing and remote management through a low-power wide-area network (such as LoRa, NB-IoT).

[0020] The various sensor nodes are divided into different types according to functions, including meteorological sensor nodes, geological structure sensor nodes and energy disturbance sensor nodes. According to the types of geological disasters (such as landslides, mudslides, collapses and the like), automatic selection and fixed-point deployment are performed.

[0021] Rainfall is an important external driving factor for inducing geological disasters such as landslides and mudslides, and by deploying a rain gauge, data such as rainfall intensity, cumulative rainfall and short-time storm index within a unit time in a target region are collected.

[0022] In the historical landslide area or the fissure development area, a surface displacement sensor such as a GNSS displacement meter, a laser range finder or an optical fiber inclinometer is arranged to capture the surface deformation and the slope movement trend.

[0023] In order to monitor the influence of underground water on the stability of the slope, a soil moisture content sensor (such as a TDR sensor) is arranged along the depth direction of the slope profile to obtain the moisture content change of the soil layer at different depths in real time, reflect the rainfall infiltration, water storage and seepage conditions, and is an important criterion for the critical evolution stage of the landslide.

[0024] The soil temperature sensor is used to monitor the soil temperature of the shallow and deep layers, and analyze the abnormal fluctuations in the coupling process of heat conduction and seepage. The mutation or periodic disturbance of the soil temperature signal is often related to the enhanced underground water activity and structure stress release, and is an important supplementary index for early weak signal extraction.

[0025] In the key area, a ground sound monitoring unit (such as a ground sound sensor or an acoustic emission array) is arranged to capture the high-frequency ground sound signals generated by the micro-fissure expansion and stress accumulation in the rock-soil body. By monitoring the changes of the ground sound energy and frequency distribution in real time, the “no appearance” hidden disaster process is assisted to be identified.

[0026] The multi-source sensor nodes send the collected data to the edge computing nodes in a timely manner through a wireless communication protocol.

[0027] The edge intelligent processing module provided by the embodiment of the application realizes in-situ feature extraction and micro-variation anomaly identification of the original data by deploying a lightweight computing model and a low-power algorithm at the edge node of the monitoring area.

[0028] The edge intelligent processing module includes original data calibration and slicing processing, time-frequency domain feature extraction, disturbance entropy and instantaneous energy index calculation, anomaly identification and feature vector output and the like. The specific implementation steps are as follows: The original data packet uploaded by the multi-source sensor acquisition module includes data streams from rain, surface displacement, soil moisture content, ground sound signals and soil temperature sensors. Due to the differences in data collection frequency and inconsistent time stamps of different sensor nodes, the edge node needs to perform unified calibration processing on the original data.

[0029] Specifically, first, all data streams are time-aligned according to timestamps (accurate to milliseconds), missing sampling points are filled using linear interpolation, and data are mapped to node coordinates in the network topology according to spatial node numbers.

[0030] After calibration, fixed-length sliding time windows are used to process continuous data. The window length can be dynamically set according to the specific type of geological disaster, and the sliding step is generally twice the original sampling period. For example, if the geophone sampling frequency is 100 Hz, the window length is set to 10 seconds, corresponding to 1000 sampling points, and the sliding step is set to 2 seconds, realizing real-time feature extraction at the segment level.

[0031] In each sliding window segment, frequency domain feature extraction is performed on the geophone signal and the ground displacement signal. These two types of signals have obvious disturbance response in disaster precursors, so they are used as main variables for subsequent energy spectrum construction. The specific processing method is as follows: First, perform fast Fourier transform (FFT) on the signal to obtain the amplitude distribution in the frequency domain; then, use wavelet packet decomposition (WPD) to further decompose the signal into multi-scale multi-band sub-signals to improve frequency resolution and adaptability to non-stationary signals; finally, calculate the energy value of each frequency band sub-signal and aggregate to construct the time-frequency energy distribution spectrum of the segment.

[0032] Let W_ij be the jth frequency band sub-signal obtained by wavelet decomposition in the ith sliding window, and define its energy E_ij as the sum of squares of all sampling points (i.e., the sum of squares of all sampling points). The energy distribution spectrum of the entire signal can be represented as a matrix E = [E_11, E_12,..., E_mn], where m is the number of time segments and n is the number of frequency bands.

[0033] To identify the weak disturbance features before the occurrence of geological disasters, the disturbance entropy and instantaneous energy index are introduced as micro-variation anomaly criteria.

[0034] The calculation method of disturbance entropy is as follows: first, normalize the energy values in the energy distribution spectrum, convert each frequency band energy value E_ij to energy probability P_ij, and define it as: (sum of all frequency bands in the current window); then, calculate the disturbance entropy H_t of the normalized spectrum using the Shannon entropy definition, which is specifically expressed as: The larger the value, the more dispersed the signal energy distribution and the more chaotic the spectrum, indicating the presence of disturbance. Under normal circumstances, signal energy is concentrated in one or two main frequency bands, and the disturbance entropy is low; when a landslide or geophone precursor appears, the spectrum spreads, causing the entropy value to rise sharply.

[0035] The transient energy index calculation method is as follows: in the time domain, the original signal in each sliding window is calculated to obtain a transient power sequence, that is, the square value of the amplitude of each sampling point, then a local weighted average is performed in a sliding sub-window (such as 1 second) to calculate a local energy mean value. The transient energy index I_e is defined as: ; wherein A_k is the amplitude of the kth sampling point, and W_k is the weight thereof (which is set according to the distance from the center time, for example, a Gaussian weight or a trigonometric function weight).

[0036] The disturbance entropy H_t calculated in each sliding window and the transient energy index I_e form a two-dimensional feature vector , which is compared with a historical background model to determine whether it is a potential abnormal disturbance.

[0037] A dynamic double-threshold identification algorithm is used, and an entropy threshold H_thresh and an energy threshold I_thresh are set, and the abnormal segment is identified through the following judgment logic: if and , it is determined as a high-risk disturbance; if any value exceeds the threshold, it is determined as a moderate disturbance and enters an observation state; if none of them exceeds, it is considered normal. The threshold can be set through historical sample data, and adaptive threshold adjustment for different regions and geological types is supported.

[0038] In the embodiment of the present application, the regional risk modeling module constructs a spatial topology atlas of sensor nodes, combines historical geological structure models and real-time sensing data, uses a graph neural network to model and predict the potential propagation path and influence probability of a geological inducing factor, and further generates a regional-level disaster potential risk atlas.

[0039] First, according to the spatial positions (such as longitude and latitude, elevation, geological profile position, etc.) of the sensor nodes deployed in the target monitoring region and the monitoring factor types (such as rainfall, displacement, ground sound, water content, etc.) thereof, a sensor spatial topology atlas for describing the propagation relationship of geological inducing factors is constructed. The topology graph is defined as a graph structure data G=(V,E,X) in the form of an undirected graph, wherein: V represents a node set, corresponding to various types of sensor units; E represents an edge set, used to represent possible geological physical coupling paths (such as the same landslide body, the same rock layer, the crack network, the fault plane extension direction, etc.) between nodes; X represents a node feature matrix, each row of which corresponds to a multi-dimensional feature vector of a node, which will be defined in detail later.

[0040] The edge establishment rule is based on the following strategy: if the Euclidean distance between two nodes is less than a preset geological propagation radius R (for example, 50 meters); or both are in the same geological unit (such as a slip zone, a fault zone, or a drainage network); or have a significant historical response coupling relationship (based on mutual information MI value greater than a threshold of 0.6); then the two nodes are considered to have a potential geological induced factor propagation path, and an edge is established in the graph structure to connect them.

[0041] To quantify the perception and response capabilities of the model for each node, each node V_i is assigned a d-dimensional feature vector X_i representing its multi-dimensional attributes in the current state. The construction of X_i combines historical geological models and current real-time perception data, specifically including the following components: terrain factors: elevation value, slope angle, slope direction; geological factors: soil type, weathering degree, groundwater depth; dynamic factors: average rainfall in the past 7 days, average ground vibration disturbance entropy, displacement rate, water content rate; historical anomaly factors: abnormal event frequency and duration in the past 30 days; environmental factors: ground vegetation coverage, ground temperature gradient.

[0042] These features are normalized to the [0, 1] interval to ensure comparability of different dimensional features in neural network training. X ultimately forms an Nxd feature matrix, where N is the number of nodes in the graph and d is the number of feature dimensions for each node.

[0043] The present application selects a graph convolutional neural network (GCN) as the core prediction model, which is trained on the constructed graph structure data G to learn the spatio-temporal pattern of geological induced factor propagation between nodes.

[0044] The graph convolutional neural network operates on the graph structure based on the adjacency matrix, and each layer in the model performs the following propagation formula: ; Where: A is the node adjacency matrix, representing the node connection relationship; D is the degree matrix, which normalizes A; represents the node feature input matrix of the lth layer, is the training weight parameter of the lth layer; is an activation function, and ReLU function is selected.

[0045] Through multi-layer graph convolution propagation, the output feature vector of each node integrates its own attributes and the information of adjacent nodes, realizing the "diffusion" propagation and aggregation of information in the spatial graph.

[0046] To train the graph neural network, a supervised label needs to be introduced. The present application takes the frequency or risk level of geological anomalies occurring in each node within a historical time window as a supervised signal to construct a node-level label vector , where: , represents the node There are high-risk behaviors such as landslides, displacement mutations, etc. in historical data; , represents that the node is not an abnormal point. The cross-entropy loss function is used as the objective function for training, which is: ; Where, is the node The prediction probability value output by the graph neural network represents the probability of inducing disasters in its current state. Through the back propagation and gradient descent method (such as the Adam optimizer), the model parameters are continuously iterated until the loss function converges.

[0047] After training, the model outputs the corresponding geological disaster impact probability for each node feature input. The higher the probability value, the more likely the node will become a disaster-inducing core or response area.

[0048] Map the prediction probability of all nodes to the actual geographic space, and combine the terrain data (such as DEM model) and geological fault map to realize two-dimensional or three-dimensional visualization, and build a regional-level "disaster potential risk map".

[0049] The embodiment of the present application provides a multi-source fusion analysis module, which is used for considering the geological structure, future meteorological change trend and topographic features of the high-risk area on the basis of the existing disaster potential risk map, constructing a multi-dimensional coupling mapping model between inducing factors and response factors, and realizing dynamic prediction of the development trend and spatial influence range of geological disasters.

[0050] First, based on the disaster potential risk map output by the regional risk modeling module, identify nodes with a prediction probability higher than a set threshold P_thresh to form a high-risk node set S_high. Wherein, P_thresh is a disaster risk judgment threshold preset by the system, which can be set according to historical data experience or ROC curve analysis method, and generally takes a value range of 0.6 to 0.8.

[0051] For each high-risk node, the following key geological feature information is extracted from its original feature vector: stratigraphic structure (such as clay layer, clastic rock, weathering zone); displacement change rate and direction; ground vibration disturbance intensity index; soil moisture content fluctuation range; cumulative number of historical abnormal events.

[0052] Obtain weather prediction data of the monitoring area within T hours (T is generally 6 to 48 hours) in the future, including: rainfall intensity (unit: mm / hour); average wind speed and direction (unit: m / s); temperature change curve (unit: Celsius); The weather prediction data is mapped to the area where the high-risk node is located in a grid form using a spatial interpolation algorithm such as inverse distance weighting (IDW) or Kriging interpolation, realizing the accurate projection of weather parameters from the global scale to the local node. Each node finally obtains a set of future weather change sequences, which are used to simulate the continuous change process of potential induced stress.

[0053] Combined with the digital elevation model (DEM), the terrain response characteristics of the location of each high-risk node are obtained, including but not limited to: slope (expressed in degrees or percentages): indicating the degree of inclination of the ground surface, which is a key factor affecting the probability of landslide and debris flow occurrence; aspect (expressed in 0 to 360 degrees): used to evaluate the influence direction of rainfall, wind, and sunlight on a specific slope surface; drainage direction and convergence area: calculated by D8 algorithm or D∞ algorithm, used to judge the water flow aggregation trend after rainfall, identify possible sliding path or valley water accumulation area.

[0054] All terrain features are extracted in GIS raster format and then converted to point data associated with node coordinates to realize the binding of high-precision terrain information and risk nodes.

[0055] After obtaining geological, meteorological, and terrain information, the system constructs a unified input feature vector F_i for each high-risk node, which has the following three types of variables: disaster trigger factors: rainfall intensity change, wind speed extreme value, temperature mutation, water content daily variation amplitude; terrain response factors: slope, aspect, convergence index, slope foot length; historical dynamic characteristics: disturbance entropy mean, displacement fluctuation frequency, historical disaster label.

[0056] The node feature set is input into the mapping model based on the decision tree (Decision Tree) algorithm, and the model training target is to judge the possibility of the current node entering the critical state of disaster under the superposition of meteorological disturbance.

[0057] The decision tree model training steps are as follows: F_i is used as input, the prediction target is set to the node entering the "critical state" (label 1) or "non-critical state" (label 0); information gain or Gini coefficient is used as the division criterion to recursively construct tree structure nodes; supervised training is performed through historical monitoring samples, and cross-validation method is used to prevent overfitting; the output is the disaster trigger possibility score corresponding to each node, with a range of 0 to 1.

[0058] The early warning module provided by the embodiment of the application is used for dynamically judging the development state of potential geological disasters in a monitoring area, and automatically triggering an early warning release mechanism when the disaster evolution reaches a critical state, so that accurate and efficient active disaster response is realized. The module is based on the causal inference theory, and constructs a logical dependency relationship between disaster inducing factors and response factors. By means of a Bayesian network, conditional probability inference is realized, and a complete disaster evolution chain identification mechanism is formed.

[0059] Firstly, the module extracts a set of disaster inducing factors and response factors based on the high-risk node features identified by the previous module. The inducing factors include but are not limited to: sustained rainfall intensity, soil moisture content rate, sudden increase in ground disturbance entropy, displacement rate change, etc. The response factors include surface crack enlargement, structural stress anomaly, historical landslide precursor, etc.

[0060] According to domain knowledge and historical data statistics, a causal inference graph (Causal Inference Graph) of the disaster evolution process is constructed. The graph is represented by a directed acyclic graph, in which: nodes represent various types of inducing and response variables; directed edges represent the causal influence paths between variables; each edge is assigned a conditional probability value, representing the influence degree of the predecessor variable on the state of the successor variable. For example, the edge from “continuous 3-hour rainfall exceeding 20 mm” to “rapid increase in soil moisture content” may have a high edge weight value of 0.85, indicating a strong dependency relationship.

[0061] In order to obtain the real probability estimate of the edge weight in the causal graph, the application introduces a Bayesian network structure learning algorithm to train the causal graph. The Bayesian network takes the joint probability distribution of nodes as the core, and can realize the inference and probability update of the latent state. The training process includes the following steps: A large number of historical monitoring samples are used as a training set to construct a variable state data set D, which is in the form of a multi-dimensional feature vector, where each dimension represents the state value of a factor at a certain time; an optimal graph structure is generated by using a structure learning algorithm (such as K2 algorithm, greedy search or heuristic scoring), that is, it is determined whether there is a causal connection between variables; a parameter learning algorithm (such as maximum likelihood estimation or Bayesian estimation) is used to calculate the conditional probability distribution between each pair of variables to form a complete Bayesian network parameter table. The finally obtained model is in the form of: ; wherein, represents the variable The conditional probability of a variable given the states of all its parent nodes represents the joint distribution of all variables in the Bayesian network.

[0062] In the real-time monitoring stage, the system inputs the observation data of the current high-risk node into the Bayesian network, and calculates the occurrence probability of the response factor state according to the joint probability inference mechanism of the network. For example, given the current values of rainfall, water content and disturbance entropy, the joint probability value P (landslide precursor = 1 | current input) of “landslide precursor has appeared” is calculated.

[0063] A disaster critical state threshold T critical is set, representing the minimum probability threshold for triggering a certain response state. If the calculation result meets , it is determined that the current disaster has entered the critical conversion stage. The T critical can be set according to different disaster types, for example, the recommended threshold for landslide disaster is 0.7, the threshold for collapse is 0.65, and the threshold for debris flow is 0.75, and can be calibrated by expert experience or historical false alarm rate.

[0064] In addition, multiple critical thresholds can be set to divide the disaster warning levels, such as: warning level I (mild): P ∈ [0.5, 0.65); warning level II (moderate): P ∈ [0.65, 0.8); warning level III (severe): P ≥ 0.8.

[0065] Once it is judged according to the Bayesian inference result that the disaster evolution has reached or exceeded the critical state, the warning release mechanism is triggered. The warning information includes the following contents: disaster type (such as landslide, debris flow); high-risk area positioning information (latitude and longitude, village name, monitoring point number); current warning level; recommended prevention measures (such as risk avoidance, evacuation, traffic control, etc.); data support basis (related index change trend, probability score).

[0066] The resource optimization module provided in the embodiment of the application dynamically adjusts the sampling frequency and communication priority of the sensor based on the similarity of the current disturbance characteristics of the monitoring point and the historical abnormal behavior, so as to prolong the system running time and resource use efficiency without sacrificing the precursor identification accuracy.

[0067] The current disturbance feature vector sequence of the monitoring point is extracted in each sliding time window, and the feature vector includes key parameters such as geophone disturbance entropy, displacement rate and water content change rate, and is denoted as a current sequence vector set .

[0068] The matching historical abnormal sequence library Each historical sample sequence is a multi-dimensional time series of a geological anomaly. To measure the similarity between the current disturbance and the historical anomaly, the dynamic time warping (DTW) algorithm or the normalized cosine similarity is used for comparison: DTW can handle the problem of inconsistent sequence time alignment, and the minimum path distance between the two sequences is calculated as the basis for similarity; the cosine similarity calculates the cosine value of the vector angle, and the result ranges from 0 to 1, and the larger the value, the more similar it is.

[0069] After comparing all historical samples, the highest similarity score S_max is taken as the normalized similarity score S of the current disturbance, which is defined in the range of 0 to 1, and subsequent classification judgment and control strategy execution are performed according to S.

[0070] According to the comparison result of the similarity score S and the preset threshold S_high and S_mid, the system divides the monitoring node into three risk levels, corresponding to different resource scheduling strategies: When S≥S_high (such as S_high=0.8), it is determined that the current disturbance is highly similar to the historical abnormal state, and a high priority strategy is enabled. The node sampling frequency is increased from the default reference frequency F_base to the high frequency mode F_high, F_high is 5 to 50 times the value of F_base, and the specific value is configured by the disaster type. At the same time, the node communication priority is set to level 1, ensuring that the node reporting data has the highest scheduling priority in the transmission link.

[0071] When S_mid≤S<S_high (such as S_mid=0.5), it is determined to be a medium similarity state, the sampling frequency is increased to the intermediate frequency F_mid (2 to 5 times F_base), and the communication priority is set to level 2; when S<S_mid, the current sampling frequency F_base and communication priority 3 are maintained, and resources are saved. The sampling frequency and priority adjustment strategy table is as follows:

[0072] To avoid the node running out of energy due to frequent high-frequency sampling, before executing the sampling frequency promotion strategy, it is necessary to judge whether the remaining power E_res of the monitoring node meets the minimum sampling guarantee threshold E_min. E_min is the minimum power lower limit configured, which is generally determined according to the sensor model, sampling power consumption and required running time, for example, 2.5 volts or 20% capacity. The specific strategy is as follows: if E_res≥E_min, the sampling frequency is allowed to be promoted according to the strategy; if E_res <e_min，将自动启用降采样策略或事件触发式上报机制：即仅在扰动特征发生快速变化、超过局部扰动门限δt（如扰动熵变幅超过0.2）时进行短时间高频采样，其余时段进入低功耗监听模式；The above description is only the specific implementation of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.< / e_min，将自动启用降采样策略或事件触发式上报机制：即仅在扰动特征发生快速变化、超过局部扰动门限δt（如扰动熵变幅超过0.2）时进行短时间高频采样，其余时段进入低功耗监听模式；

Claims

1. A geological disaster real-time monitoring system based on Internet of Things, characterized in that: The method comprises the following steps: A multi-source sensing acquisition module: deploying Internet of Things sensor network nodes in the target area to collect raw data including rainfall, ground displacement, soil moisture content, ground temperature, and ground sound signals; An edge intelligent processing module: sending the raw data to the edge node, which extracts features based on local low-power algorithms to obtain time-frequency domain micro-variation anomaly signal vectors, and calculates signal disturbance entropy and instantaneous energy index; A regional risk modeling module: constructing a spatial topology graph between nodes, combining historical geological structure models and real-time sensing data, predicting the propagation path and influence probability of key triggering factors in the region through a graph neural network, and generating a disaster potential risk graph; A multi-source fusion analysis module: extracting key node features of high-risk areas based on the risk graph, further integrating meteorological prediction data and DEM elevation data, and establishing a multi-dimensional mapping relationship between disaster triggering factors and terrain response factors; An early warning module: using a causal reasoning graph to construct a disaster evolution chain model to determine whether a potential disaster has entered a critical conversion stage, and sending a warning message if it is determined to be in a critical state; A resource optimization module: adaptively adjusting the sampling frequency and communication priority of the sensor based on the similarity of the historical anomaly sequence of the monitoring point and the current disturbance pattern. 2.The real-time monitoring system for geological disasters based on Internet of Things according to claim 1, characterized in that: The edge intelligent processing module comprises: Synchronize and calibrate the obtained raw data packets according to the time stamp and spatial node identifier, and divide them into fixed-length analysis segments using a sliding window segmentation method; Perform fast Fourier transform and wavelet packet decomposition on the ground sound signals and displacement signals in each analysis segment to extract the energy distribution spectrum of the signals in the time-frequency domain; Calculate the disturbance entropy and instantaneous energy index of the micro-variation signal based on the energy distribution spectrum, and identify potential abnormal disturbance events. 3.The real-time monitoring system for geological disasters based on Internet of Things according to claim 2, characterized in that: The calculation of the disturbance entropy and instantaneous energy index of the micro-variation signal based on the energy distribution spectrum comprises: Normalize the energy distribution spectrum obtained through spectral analysis to construct an energy probability density function for the corresponding time period; Calculate the Shannon disturbance entropy based on the probability density function to measure the energy uncertainty and disturbance intensity variation of the signal in the frequency domain; Extract the instantaneous power of the continuous segment signal using a sliding time window in the time domain, and calculate the instantaneous energy index according to the weighted average method. 4.The real-time monitoring system for geological disasters based on Internet of Things according to claim 1, characterized in that: The regional risk modeling module comprises: Based on the spatial location information and monitoring factor type of each sensor node, construct a sensor spatial topology graph in the region, where the nodes represent sensor units and the edges represent geological correlation paths; Integrate historical geological structure models and current real-time sensing data to assign a multi-dimensional feature vector to each node; Train the topology graph using a graph neural network to learn the propagation relationship of geological triggering factors between nodes and predict the disaster impact probability of different nodes; Map the prediction results to a geographic information system to generate a regional disaster potential risk graph. 5.The real-time monitoring system for geological disasters based on Internet of Things according to claim 4, characterized in that: Training the topology graph using a graph neural network to learn the propagation relationship of geological triggering factors between nodes and predicting the disaster impact probability of different nodes comprises: representing the constructed spatial topology graph as graph structure data wherein V is a node set, E is an edge set, and X is a multi-dimensional geological feature matrix of the nodes The graph convolutional neural network is selected as a core model, the adjacency matrix of X is normalized, and the interlayer propagation rule is used to embed and aggregate the node features; The node historical abnormal frequency is defined as a supervision label, and a cross-entropy loss function is constructed to supervise the training of the graph neural network and iteratively optimize the parameters; After the training is completed, the influence probability value of the geological inducing factor of each node is output, and the nodes are spatially sorted according to the probability. 6.The real-time monitoring system for geological disasters based on Internet of Things according to claim 1, characterized in that: The multi-source fusion analysis module comprises: Based on the node set with a prediction probability higher than a preset threshold in the disaster potential risk map, the corresponding key spatial position and its geological feature information are extracted; Receive rainfall intensity, wind speed and temperature change data in the future time period, and map them to the area where the high-risk nodes are located; Combined with the digital elevation model, the slope, slope direction and drainage direction topographic response factors of the high-risk area are obtained; A multi-dimensional mapping model based on decision tree is constructed to learn the synergistic relationship between disaster triggering factors and topographic response factors. 7.The real-time monitoring system for geological disasters based on Internet of Things according to claim 1, characterized in that: The early warning module comprises: Based on the characteristics of high-risk nodes, a causal reasoning graph is constructed between disaster inducing factors and response factors, and the edge weight of the graph represents the conditional probability relationship between variables; According to the Bayesian network structure learning algorithm, the graph is trained to obtain the probability path of the chain propagation of disaster inducing factors to the response state; Set the threshold condition of the critical conversion stage of the disaster, when the joint probability value of a certain state node exceeds the preset threshold, it is determined to enter the critical state. 8.The real-time monitoring system for geological disasters based on Internet of Things according to claim 1, characterized in that: The resource optimization module comprises: The current disturbance feature vector sequence is compared with the sample sequence in the historical abnormal sequence library to obtain a normalized similarity score S; according to the similarity S and the preset threshold S_high, S_mid, the similarity is classified and determined: when S≥S_high, it is determined to be highly similar, the sampling frequency is increased from the base frequency F_base to the high priority frequency F_high, and the node is assigned to the communication priority 1; when S_mid≤S<S_high, the sampling frequency is increased to the medium frequency F_mid, and the communication priority 2 is assigned; otherwise, F_base and priority 3 are maintained.

Images