Intelligent early warning and alarming system for sudden water regimen disasters based on multi-source perception

Abstract

The invention relates to the technical field of water conservancy monitoring and disaster early warning, in particular to an intelligent early warning and alarming system for sudden water regimen disasters based on multi-source perception. The system comprises a multi-source data acquisition module, a sudden abnormity preliminary screening module, a space-time coupling verification module, a dynamic effectiveness evaluation module and an intelligent early warning execution module. The method comprises the following steps of: acquiring multi-modal heterogeneous sensing data by a system, and identifying data burst jump characteristics; the core of the method is that a rainfall correlation degree and a field video stream are combined to calculate a multi-source signal space-time coupling confidence coefficient, and a physical logic and visual perception dual verification system is established to determine signal validity; when the verification result indicates a real disaster, the system triggers graded alarm according to the evolution trend; according to the invention, the conversion from single threshold alarm to multi-dimensional joint discrimination is realized, the environmental noise interference can be effectively eliminated, and the early warning reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of water conservancy monitoring and disaster early warning technology, specifically to an intelligent early warning and alarm system for sudden water disasters based on multi-source sensing. Background Technology

[0002] With the deepening of water conservancy informatization construction, the demand for water situation monitoring in complex field environments is increasing, and the real-time acquisition and analysis of multi-source heterogeneous data has become a key link in ensuring the safety of water conservancy projects. Currently, existing emergency flood warning methods mainly rely on numerical threshold determination by a single sensor or regular manual inspections. However, in unattended field scenarios, single monitoring devices are highly susceptible to interference from non-hazardous factors such as electron drift, biological disturbance, and obstruction by foreign objects. These environmental noises often cause sudden jumps in sensor data, generating environmental interference signals. In addition, the lack of logical cross-validation of multi-dimensional data makes it difficult for traditional systems to accurately distinguish between real disaster characteristics and random interference signals in complex physical environments, resulting in a high false alarm rate and a lack of adaptive response mechanisms, which seriously affects the reliability of early warnings.

[0003] Therefore, how to effectively eliminate noise interference in complex environments and achieve accurate identification and intelligent early warning of real sudden water disasters has become an urgent problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent early warning and alarm system for sudden flood disasters based on multi-source sensing. Specifically, the technical solution of this invention includes: The multi-source data acquisition module is configured to acquire multimodal heterogeneous sensor data within the monitoring area. The multimodal heterogeneous sensor data includes: water level time series data, rainfall time series data, cross-sectional flow velocity data, and on-site monitoring video stream. The sudden anomaly screening module is configured to identify sudden jump features where the rate of change of data amplitude exceeds a preset jump threshold based on the water level time series data or the cross-sectional flow velocity data, and to mark data segments with the sudden jump features as candidate anomaly signals. The spatiotemporal coupling verification module is configured to respond to the candidate anomaly signal by calling rainfall time-series data associated with the candidate anomaly signal in the time and spatial dimensions, and calculating the spatiotemporal coupling confidence of multi-source signals that characterizes the degree of correlation between rainfall driving and water level change. The dynamic validity assessment module is configured to call the on-site monitoring video stream associated with the candidate abnormal signal, and based on the spatiotemporal coupling confidence of the multi-source signal and the analysis results of the on-site monitoring video stream, the analysis results include identifying whether there are non-aquatic foreign objects obstructing or biological disturbance features at the monitoring point, determining whether the candidate abnormal signal belongs to real disaster features or environmental noise interference, and generating validity verification results. The intelligent early warning execution module is configured to trigger a graded alarm command based on the evolution trend of the sudden change characteristics when the validity verification result indicates real disaster characteristics.

[0005] Preferably, the spatiotemporal coupling verification module performs the following operations to construct the spatiotemporal coupling confidence of multi-source signals: Determine the occurrence time of the candidate abnormal signal and the corresponding monitoring point; Retrieve rainfall time-series data within a preset range upstream of the monitoring point; Based on a preset nonlinear hysteresis effect model, the causal logical relationship between the rainfall time series data and the candidate anomaly signal is analyzed to generate a hydraulic logic matching score. The hydraulic logic matching score is mapped to the base score of the spatiotemporal coupling confidence of the multi-source signal.

[0006] Preferably, the spatiotemporal coupling verification module is further used for: Collect soil saturation parameters and topographic features around the monitoring points; Based on the soil saturation parameter and the topographic feature parameter, the lag time window in the nonlinear hysteresis effect model is dynamically adjusted. Within the lag time window, it is verified whether the cumulative amount of the rainfall time series data meets the minimum physical conditions for triggering the candidate anomaly signal, so as to correct the hydraulic logic matching score.

[0007] Preferably, the dynamic validity assessment module performs the following operations to generate validity verification results: The preset pseudo-anomaly feature recognition algorithm model is invoked to perform image semantic analysis on the on-site monitoring video stream; The system identifies whether the monitoring points are obstructed by non-aquatic objects or exhibit biological disturbance characteristics, and generates an environmental noise probability value that quantifies the degree of environmental interference. Based on a preset weighted fusion strategy, a comprehensive credibility index is calculated by combining the spatiotemporal coupling confidence of the multi-source signals with the probability value of the environmental noise. If the comprehensive credibility index is greater than the preset alarm trigger threshold, a validity verification result indicating the characteristics of the real disaster is generated. If the overall credibility index is less than or equal to the alarm trigger threshold, then a verification result indicating the effectiveness of environmental noise interference is generated and the alarm is suppressed.

[0008] Preferably, the system further includes a dynamic threshold control module, configured as follows: Monitor the duration of rainfall and the trend of upstream water flow in the multimodal heterogeneous sensor data; When it is confirmed that there is continuous rainfall or a continuous rise in upstream water level, calculate the dynamic threshold adaptive convergence time. Within the adaptive convergence time of the dynamic threshold, the alarm trigger threshold is reduced according to a preset attenuation strategy to improve the system's sensitivity to subsequent sudden floods.

[0009] Preferably, when constructing the spatiotemporal coupling confidence score for multi-source signals, the spatiotemporal coupling verification module further performs the following: Detect whether there are missing data in some dimensions of the multimodal heterogeneous sensing data; If data is missing, interpolation is performed using data from similar sensors at nearby monitoring points and historical data from the same period to complete the data. Cross-validation is performed on the completed data, and a confidence penalty factor is introduced according to the proportion of missing data to reduce the weight of the spatiotemporal coupling confidence of the multi-source signal.

[0010] Preferably, the intelligent early warning execution module performs the following operations to trigger a tiered alarm command: Calculate the rate of change and acceleration of the actual disaster characteristics per unit time; The rate of change and acceleration are input into a preset disaster risk assessment mapping table to determine the current water situation's danger level; Based on the stated hazard level, a corresponding alarm response strategy is matched, which includes adjusting the data acquisition frequency, sending alarm information through multiple channels, and activating linkage control equipment.

[0011] Preferably, the system further includes an adaptive feedback optimization module, used for: Record the comparison information between the validity verification results and the subsequent manual review results; When the comparison information shows a false alarm or a missed alarm, the corresponding multimodal heterogeneous sensing data sample is extracted; The pseudo-anomaly feature recognition model is incrementally trained using the multimodal heterogeneous sensing data samples to update the parameters of the pseudo-anomaly feature recognition model.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This system breaks through the limitations of traditional single-sensor threshold alarms. By integrating hardware devices such as radar water level gauges, rain gauges, and Doppler current meters, and combining them with on-site monitoring video streams, a dual verification system of physical logic and visual perception is established. On the one hand, the spatiotemporal coupling confidence is calculated using the physical causal relationship between rainfall and water level; on the other hand, image semantic analysis is used to identify non-water body obstructions or biological disturbances. This mechanism effectively eliminates non-hazardous interference signals caused by sensor electronic drift, spider web obstruction, or floating objects on the water surface, solving the problem that existing technologies cannot distinguish between real disaster characteristics and environmental noise interference, and significantly improving the reliability of the early warning system. 2. This system does not rely solely on a fixed runoff time window. Instead, it dynamically adjusts the lag time window and runoff compensation coefficient in the nonlinear hysteresis effect model by collecting soil saturation parameters and topographic features around the monitoring points. This design can accurately distinguish between different scenarios, such as slow runoff caused by strong soil water absorption in the dry season and rapid runoff caused by soil saturation in the rainy season. It also automatically corrects the analysis logic for steep or gentle slopes. By introducing boundary constraints and dynamic correction mechanisms, it effectively solves the problem that a single fixed model cannot adapt to seasonal changes and complex geomorphic environments, ensuring the robustness and computational accuracy of the algorithm under different hydrological conditions. 3. This system is designed with a dynamic threshold control module, which can monitor the duration of rainfall and the trend of upstream water flow in real time. When continuous rainfall or a continuous rise in water level is confirmed, the system will simulate the sensitization response of a biological nervous system and automatically reduce the alarm trigger threshold according to a preset attenuation strategy. This mechanism allows the system to gradually increase its sensitivity as disaster risks accumulate, effectively preventing slow response to subsequent minor but critical sudden floods due to overly rigid threshold settings, and significantly improving the system's emergency response capability during long-duration rainfall. 4. Through the adaptive feedback optimization module, this system can automatically record the comparison information between the validity verification results and the manual review results. Once a false alarm or missed alarm is detected, the system will automatically extract the corresponding multimodal sensor data as a hard sample and incrementally train the pseudo-anomaly feature recognition model. This self-iterative capability without forgetting the original knowledge enables the system to continuously adapt to the unique environmental characteristics of specific monitoring points, solving the problem that traditional early warning system algorithms cannot be updated after deployment and are difficult to cope with environmental evolution, thus realizing the continuous evolution of algorithm performance. Attached Figure Description

[0013] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0015] Example 1: Please see Figure 1 A smart early warning and alarm system for sudden flood disasters based on multi-source sensing includes: The multi-source data acquisition module is configured to acquire multimodal heterogeneous sensor data within the monitoring area. The multimodal heterogeneous sensor data includes: water level time series data, rainfall time series data, cross-sectional flow velocity data, and on-site monitoring video stream. The sudden anomaly screening module is configured to identify sudden jump characteristics where the rate of change of data amplitude exceeds a preset jump threshold based on water level time series data or cross-sectional flow velocity data, and mark data segments with sudden jump characteristics as candidate anomaly signals. The spatiotemporal coupling verification module is configured to respond to candidate anomaly signals by calling rainfall time-series data associated with the candidate anomaly signals in the time and spatial dimensions, and calculating the spatiotemporal coupling confidence of multi-source signals that characterizes the degree of correlation between rainfall-driven and water level changes. The dynamic validity assessment module is configured to call the on-site monitoring video stream associated with the candidate abnormal signal, and based on the spatiotemporal coupling confidence of the multi-source signal and the analysis results of the on-site monitoring video stream, the analysis results include identifying whether there are non-water body foreign objects blocking or biological disturbance characteristics at the monitoring point, determining whether the candidate abnormal signal belongs to real disaster characteristics or environmental noise interference, and generating validity verification results. The intelligent early warning execution module is configured to trigger a graded alarm command based on the evolution trend of sudden change characteristics when the validity verification result indicates the characteristics of a real disaster.

[0016] Based on this, the multi-source data acquisition module integrates hardware devices such as radar water level gauges, rain gauges, and Doppler current meters to acquire sets of sensing information with different physical dimensions and data structures; among them, water level time series data... Rainfall time series data derived from real-time sampling of radar level gauges or submersible level gauges Intensity information derived from rain gauges or weather radar, and cross-sectional flow velocity data. The water flow velocity is derived from a Doppler current meter, while the on-site monitoring video stream... It consists of real-time image frames captured by visible light or infrared cameras; The initial screening module for sudden anomalies executes rapid capture logic and calculates water level time-series data. or cross-sectional flow velocity data The first derivative, i.e., the rate of change; in response to the rate of change exceeding a preset jump threshold. In this embodiment, to ensure the ability to capture early-stage sudden floods and filter out regular fluctuations, and to maintain consistency with subsequent risk grading logic, the threshold is specifically set as follows: If the system is operating in minute-level sampling mode, this threshold will be automatically converted to an equivalent value. This embodiment preferably uses this threshold setting to prevent the initial screening threshold from being too high, leading to... The medium-risk level data in the interval was incorrectly filtered out. To ensure the closed-loop logic of the entire process, the system identified sudden jump characteristics and marked such data segments that showed numerical abnormalities but had not yet been verified by multi-source logic as candidate abnormal signals. The spatiotemporal coupling verification module, as the core logic verification unit, responds to candidate anomaly signals by calling the associated rainfall time-series data in the time dimension (historical backtracking) and the spatial dimension (upstream catchment area). To calculate the spatiotemporal coupling confidence of multi-source signals, this index quantifies whether the currently monitored water level or flow velocity anomalies have a physical driving source; furthermore, the dynamic effectiveness evaluation module introduces visual perception as a second layer of verification, by analyzing the on-site monitoring video stream. The system determines whether non-aquatic interference exists and assesses the authenticity of the signal by combining spatiotemporal coupling confidence and video analysis results; the intelligent early warning execution module triggers an alarm only when the validity verification results indicate real disaster characteristics, based on the evolution trend. This embodiment breaks through the limitations of traditional single-threshold alarms by constructing a multimodal logic gating mechanism. In complex field environments, it effectively eliminates non-hazardous interference signals caused by sensor drift, biological disturbance, or foreign object obstruction by utilizing the physical causal relationship between rainfall and water level and the dual verification of video semantic analysis. This achieves accurate capture of real water disasters and significantly improves the reliability of the system's early warning in unattended scenarios.

[0017] Example 2: The spatiotemporal coupling verification module performs the following operations to construct the spatiotemporal coupling confidence of multi-source signals: Determine the occurrence time of candidate abnormal signals and their corresponding monitoring points; Retrieve rainfall time-series data within a preset range upstream of the monitoring point; Based on a pre-defined nonlinear hysteresis effect model, the causal logical relationship between rainfall time series data and candidate anomaly signals is analyzed to generate a hydraulic logic matching score. The hydraulic logic matching score is mapped to the basic score of the spatiotemporal coupling confidence of multi-source signals.

[0018] Specifically, because rainfall convergence at the river cross-section involves surface infiltration and confluence processes, the downstream water level response is delayed in time and decays non-linearly in amplitude. The system determines the timing of candidate anomaly signals. and corresponding monitoring points ; The system retrieves monitoring points. Rainfall time series data within the upstream preset range, i.e., the catchment area. The system is based on nonlinear hysteresis effect analysis logic and calculates hydraulic logic matching scores. To quantify the causal relationship between rainfall and water level changes, considering that the numerator is the amount of rainfall. The denominator is the water level. To avoid dimensional confusion in numerical calculations, a unit alignment factor is introduced to address the dimensional differences. The calculation formula is as follows:

[0019] For the discretized implementation of computer programs, the above formula is transformed into a cumulative summation logic:

[0020] in, The sampling time step for rainfall time series data needs to be converted to hours (h). For example, for 10-minute data... This converts rainfall intensity (mm / h) into physical rainfall (mm), solving the problem of dimensional uniformity in discrete calculations; among which, Sourced from rain sensors The upstream rainfall intensity at any given time, physically represented as the input driving force, is measured in mm / h. : Derived from the preset effective confluence time window, its physical meaning is the maximum retrospective duration of rainfall affecting water levels, in hours; The weighting function is derived from a preset time decay function, specifically constructed using an exponential decay model, and its expression is as follows: In the formula Let be a natural constant; where, This is the preset bus attenuation coefficient, typically ranging from 0.1 to 0.8. In this embodiment, 0.3 is preferred. This represents the time difference between the moment rainfall occurred and the current moment; the physical meaning of this function is that the closer the rainfall occurs... The greater the contribution of rainfall at any given moment to the current water level, the greater the contribution becomes, and the contribution decreases exponentially over time; it is dimensionless. : Calculated values ​​derived from water level data, with the physical meaning of The water level change amplitude of the candidate anomaly signal at any given time, in meters; its specific calculation logic is as follows: system backtracking Preset sampling interval before time For example, read the water level at a time of 10 minutes. ,calculate To prevent calculation overflow caused by the denominator approaching zero due to small changes in water level, the system sets a minimum water level change threshold. Take during calculation As the actual denominator value; if the system is configured in continuous monitoring mode, then To trigger the derivative threshold The increase in water level within that time window, thus ensuring the denominator It can quantify the instantaneous change magnitude at the current moment, avoiding numerical mismatch caused by the integration process not being completed; It is worth noting that, in order to prevent the impact of water level fluctuations during extreme flooding scenarios, The maximum value leads to a high hydraulic logic matching score. The value was incorrectly calculated as a minimum, resulting in a missed report. This module has a preset extreme water condition bypass mechanism: before performing the division operation, the system determines... Is it greater than the preset extreme flood threshold? ,For example ;like Then, the hydraulic logic matching score will be forcibly determined. This means directly assigning an extremely high confidence level to the physical correlation, skipping subsequent formula calculations, to ensure absolute capture capability for extreme disasters; if Then continue with the following formula for calculation: The unit alignment coefficient is derived from a preset value and is used to clarify the dimensional conversion logic in the formula. In this embodiment, in order to generate a hydraulic logic matching score in mm / m to directly match the subsequent threshold, Set to 1, dimensionless; to generate standard dimensionless physical ratios, you can set... mm / m is used to offset the millimeter unit of the molecule, but in this mode, the threshold in the subsequent judgment logic... It must be divided by 1000 simultaneously, i.e. adjusted to 0.0015, to prevent logic failure due to the reduction in the magnitude of the value; The system will calculate The values ​​are mapped to the spatiotemporal coupling confidence of multi-source signals through a normalization function. In order to solve The range of 0 to positive infinity prevents it from being directly used as a probability. This embodiment uses the following modified Sigmoid function for mapping:

[0021] in, The preset steepness coefficient has the following dimensions: To adapt The unit, for example, take , This serves as the physical matching baseline threshold; it is worth noting that, It is not a universal constant, but a key physical quantity that needs to be calibrated based on the specific geographical parameters of the monitoring area; its calibration formula is:

[0022] Among them, coefficient (unit: Used to convert dimensionless ratios to The magnitude is used to match the hydraulic logic matching score. Dimensions; This represents the average width of the river channel; As the effective confluence reference length of the upstream main channel used for dimension normalization, in this embodiment, this value is fixed at [value to be filled in] to match millimeter-level rainfall and meter-level water level changes. This value is primarily based on the empirical statistical average of typical hydrological response distances in small and medium-sized watersheds, used for standardization of dimensions. In practical applications, if the monitored watershed scale is large, users are allowed to adjust the value based on the actual physical length of the controlled river section. Make adjustments proportionally; For the catchment area, it is particularly important to emphasize that when substituting into the formula for calculation, it must be uniformly converted to square meters (m²). (Unit, i.e., if the original data is) It needs to be converted to Substitute them in to ensure they match the meter-level units of the molecule and the coefficients in the formula; The average runoff coefficient; This is a dimensionless correction factor for channel roughness and friction loss, with a value range of [value range missing]. In this embodiment, The specific numerical value of this coefficient is obtained by consulting the Manning roughness coefficient table based on the riverbed sediment type, or by inversion calibration based on measured data of historical flood wave propagation time for that river section; for example, in a certain catchment area River width The monitoring points were set with an average runoff coefficient. It is 0.27, after calibration. The value is 1.5 mm / m, or 0.0015 in dimensionless mode; the physical meaning of this function is: when the hydraulic logic matching score is... Significantly greater than the baseline threshold This indicates that the driving force of rainfall is sufficient. It quickly approaches 1; conversely, it responds to The rainfall was large, but the cumulative rainfall was small, resulting in much smaller hour, Approaching 0; This embodiment introduces a hysteresis effect analysis model that includes time integrals, weighting functions, and unit alignment coefficients, and clarifies the conversion logic from physical quantity ratios to confidence probabilities. This enables the identification of real flooding processes that conform to hydraulic laws from the perspective of physical mechanisms. The solution effectively solves the problem of abnormal noise signals caused by rainless self-rising due to equipment electronic faults, ensuring the rigor of the early warning logic at the hydrophysical level.

[0023] Example 3: The spatiotemporal coupling verification module is also used for: Collect soil saturation parameters and topographic features around the monitoring points; Based on soil saturation parameters and topographic features, the lag time window in the nonlinear hysteresis effect model is dynamically adjusted. Within the lag time window, verify whether the cumulative amount of rainfall time series data meets the minimum physical conditions for triggering candidate anomaly signals, in order to correct the hydraulic logic matching score.

[0024] The system collects soil saturation parameters around the monitoring points using soil moisture sensors. The topographic and geomorphological feature parameters are obtained from the geographic information database. In this embodiment, the average slope of the monitoring area is mainly selected. As a key parameter; it is hereby clarified that the lag time window mentioned in this embodiment and the effective confluence time window defined in Embodiment 2 refer to the same technical feature, and both correspond to parameters in the physical model. Considering the drastically different response rates of dry and saturated soils to rainfall, the system implements a dynamic adjustment strategy, adjusting the lag time window in the nonlinear hysteresis effect model. Corrections are made; to prevent the time window calculation from becoming negative due to excessively large parameters under highly sensitive geological conditions, thus causing the integral logic to collapse, this embodiment introduces a boundary constraint function. Corrected lag time window The calculation is as follows:

[0025] in, The baseline runoff time is determined by the watershed area and its physical meaning is the runoff duration under standard conditions, measured in hours (h). The saturation influence coefficient, derived from historical data fitting, physically represents the weight of soil moisture content on runoff velocity and is dimensionless. Specifically, this coefficient is obtained through linear regression analysis of the soil saturation x - rainfall runoff delay time reduction rate y dataset for the same historical period at this monitoring point, with the fitting model being: Solve using the least squares method The dataset was constructed as follows: events with complete rainfall-water level response processes were selected from the hydrological historical records of the monitoring station over the past 3 years, and soil moisture and runoff lag time data pairs at the time of each event were extracted; and only when the determination coefficient of the regression analysis was... Only then was it determined that If the value is valid, otherwise the default empirical value will be used. Among them, the dependent variable The specific calculation formula is defined as follows: In the formula This refers to the time difference between the peak rainfall time and the peak water level response time of a single rainfall event, extracted from historical rainfall and water level records. This monitoring point was located in an area with historically very low soil saturation, i.e. The average confluence lag time under the given conditions, i.e., the baseline value; this formula clarifies the reduction rate. It represents the proportion of time reduction caused by the increased flow velocity at the current humidity relative to dry soil conditions; : This is derived from soil saturation data collected by sensors. Its physical meaning is the degree of water filling in soil pores, ranging from [0,1]. The terrain correction factor is derived from terrain analysis, and its specific value is based on the average slope of the monitored area. The calculation is as follows:

[0026] This is explicitly introduced The coefficient converts angles to radians, where, The average slope of the monitoring area after boundary constraint processing is calculated using the following logic: , here This is the actual measured slope value; this constraint logic is used to prevent issues arising from the tangent function in scenarios involving cliffs or artificial vertical slopes. The correction factor is caused by the tendency to infinity. Overflow or zeroing, here Let be a natural constant; where, This is a preset landform attenuation constant, which is a dimensionless empirical parameter and typically ranges from [value range missing]. This value is positively correlated with the vegetation cover of the watershed surface, for example, taking 0.5; the physical meaning of this formula is slope. The larger the value, the faster the water flow converges; the corresponding correction factor... The smaller the value, the shorter the lag time window; In the adjusted window Within this timeframe, the system re-verifies the accumulated rainfall; if the accumulated rainfall meets the minimum physical conditions, the system invokes the calculation logic of Example 2, using the modified lag time window. Replace the original parameters As the lower limit of integration, that is, the integration interval is changed to In this step, in order to correct the decrease in cumulative rainfall value caused by the shortening of the integration interval, thus affecting the matching score... To address the fundamental deviation from the expected reduction, the system introduces a runoff compensation coefficient. This coefficient is inversely proportional to the time window reduction ratio, and the calculation formula is as follows: It should be noted that, in order to prevent in In the case of extremely low soil saturation leading to extremely rapid runoff, the calculation results may be distorted due to an excessively large compensation coefficient. Therefore, this module sets an upper limit constraint on the compensation. The specific execution logic is as follows: ,in Usually set to This indicates that the compensation mechanism is mainly applicable to Common physical scenarios; the system will be based on Multiply the calculated integral result by To update the hydraulic logic matching score This compensation mechanism ensures that, in scenarios with high soil saturation, although the physical runoff time is short, the system can correctly identify the high driving efficiency of rainfall and prevent the high-risk rapid flooding process from being misjudged as environmental noise due to the truncation of the mathematical integration interval. Conversely, in response to the accumulated amount falling below the minimum physical threshold required to trigger a candidate anomaly signal, i.e., the minimum physical condition in the embodiment, the system forces the hydraulic logic to match the score. Corrected to 0; here, minimum physical threshold The setting logic is as follows: The system sets the basic interception threshold based on the soil permeability characteristics of the monitoring area. For clay areas, it is set to 3mm-5mm, and for sandy areas, it is set to 8mm-12mm. In this embodiment, 5mm is preferred; or a dynamic calculation method can be used, i.e. ,in, To minimize the production flow driving coefficient, this embodiment takes... , This represents the amplitude of water level changes; this logic ensures that subsequent confidence calculations are only performed if the accumulated rainfall is physically sufficient to generate runoff. This embodiment solves the problem that a single fixed model cannot adapt to different hydrogeological conditions, such as strong water absorption in the dry season and rapid runoff in the rainy season, by introducing a dynamic adjustment analysis window based on soil saturation and topographic parameters, and by correcting the integral results with a runoff compensation coefficient. This dynamic correction mechanism and boundary constraints significantly improve the system's early warning accuracy and algorithm robustness in seasonal transitions and complex geomorphological environments.

[0027] Example 4: The dynamic validity assessment module performs the following operations to generate validity verification results: The preset pseudo-anomaly feature recognition algorithm model is invoked to perform image semantic analysis on the on-site monitoring video stream; Identify whether monitoring points have non-aquatic foreign objects obstructing or biological disturbance characteristics, and generate environmental noise probability values ​​that quantify the degree of environmental interference. Based on a preset weighted fusion strategy, the comprehensive credibility index is calculated by combining the spatiotemporal coupling confidence of multi-source signals and the probability value of environmental noise. If the overall credibility index is greater than the preset alarm trigger threshold, a validity verification result indicating the characteristics of the real disaster will be generated. If the overall credibility index is less than or equal to the alarm trigger threshold, a verification result indicating the effectiveness of environmental noise interference is generated and the alarm is suppressed.

[0028] The system invokes a pre-defined pseudo-anomaly feature recognition algorithm model, such as an image classifier based on a convolutional neural network, to analyze the on-site monitoring video stream. Perform image semantic analysis; the model identifies whether there are non-aquatic foreign objects obstructing the image or biological disturbance features; since the original output of the neural network model is usually a multi-dimensional probability distribution vector. Since direct operations with scalars are not possible, the system executes aggregation mapping logic to generate scalar environmental noise probability values ​​that quantify the degree of environmental interference. The specific mapping logic uniformly adopts the complementary probability algorithm: This calculation method can cover all potential foreign object interferences that are not in water bodies but were not explicitly trained on the model, avoiding the risk of missed detection caused by simply accumulating known interference categories; the higher the value, the greater the possibility of non-water body interference. The system detects the signal quality of the on-site monitoring video stream; if the video stream is lost or the image quality score is lower than a preset threshold, a weighting coefficient is forcibly set. That is, it relies entirely on the physical model; if the video stream is normal, then based on the preset weighted fusion strategy, the system combines the spatiotemporal coupling confidence of multi-source signals. With environmental noise probability value Calculate the comprehensive credibility index The formula is as follows:

[0029] in, The weighting coefficient, derived from system configuration, physically represents the trust weight of sensor data relative to visual data and is dimensionless. Considering that physical sensors are generally more stable than visual algorithms in adverse weather conditions, this embodiment suggests that the value range of this coefficient be [range missing]. The preferred value is That is, it focuses on the judgment results of the physical model; : Derived from the confidence level of the spatiotemporal coupling verification module, its physical meaning is the degree of credibility of the physical logic, and it is dimensionless; : The environmental noise probability derived from the video analysis model is a scalar calculated through the above vector aggregation logic. Its physical meaning is the probability of the existence of visual interference, and it is dimensionless. The system will With preset alarm trigger threshold A comparison is performed; in this embodiment, Set as This means that an alarm will only be triggered if the overall credibility exceeds 75%; in response to If the disaster is determined to be real, a validity verification result is generated; in response to The alarm was determined to be caused by environmental noise interference and was suppressed. This embodiment effectively solves the problems of perception blind spots and data type mismatch by fusing heterogeneous data of logic and vision and clarifying the mapping relationship from multi-class probability vectors to single noise indicators. In particular, it uses computer vision technology to identify physical occlusion and biological disturbance, which greatly reduces the false alarm rate caused by complex on-site environment and realizes intelligent alarm identification and rejection.

[0030] Example 5: The system also includes a dynamic threshold control module, configured as follows: Monitoring the duration of rainfall and upstream water inflow trends in multimodal heterogeneous sensor data; When it is confirmed that there is continuous rainfall or a continuous rise in upstream water level, calculate the dynamic threshold adaptive convergence time. Within the adaptive convergence time of the dynamic threshold, the alarm trigger threshold is reduced according to the preset attenuation strategy to improve the system's sensitivity to subsequent sudden floods.

[0031] The system monitors the duration of rainfall in real time from multimodal heterogeneous sensor data. In response to confirmation of continuous rainfall or a sustained rise in upstream water levels, the system sets a dynamic threshold for adaptive convergence time. ;exist During the specified time period, the system executes a sensitization response strategy, lowering the alarm trigger threshold according to a preset attenuation strategy. Generate the dynamic alarm threshold that is in effect at the current moment. The calculation formula is as follows:

[0032] in, : is a natural constant, with a value of approximately 2.71828; The threshold value is derived from the preset configuration. Its physical meaning is the alarm threshold under sunny or normal conditions. It is dimensionless. The sensitivity decay coefficient, derived from empirical settings, physically represents the rate at which the threshold decreases over time, and is measured in units of... Its value range is usually 100%. In this embodiment, the preferred embodiment is... ; The duration of the current rainfall event, derived from a timer, is in units of... Furthermore, even after the rainfall stops, as long as the upstream water level continues to rise, the timer continues to accumulate until it resets when the water level begins to recede. With this strategy, as rainfall continues, the system automatically lowers the alarm threshold, increasing its sensitivity to subsequent minor and sudden flooding events. This embodiment simulates the sensitization response of a biological nervous system, automatically adjusting detection standards during periods of continuous rainfall with high disaster risk. Through a threshold control strategy that decays exponentially over time, it effectively avoids missed detections or slow responses caused by threshold rigidity, significantly improving the system's emergency response capability under persistent severe weather conditions.

[0033] Example 6: When constructing the spatiotemporal coupling confidence score for multi-source signals, the spatiotemporal coupling verification module also performs the following: To detect whether there are missing data in some dimensions of multimodal heterogeneous sensing data; If data is missing, interpolation is performed using data from similar sensors at nearby monitoring points and historical data from the same period to complete the data. Cross-validation is performed on the completed data, and a confidence penalty factor is introduced according to the proportion of missing data to reduce the weight of the spatiotemporal coupling confidence of multi-source signals.

[0034] The system detects whether there are missing dimensions in the multimodal heterogeneous sensor data. In response to missing data, such as lost rainfall data, the system retrieves data from nearby monitoring points (i.e., data from similar sensors within a 5km radius) and historical data from the same period at those points. It then uses Kriging interpolation or inverse distance weighting to complete the data. Based on the completed data, the system performs cross-validation and introduces a confidence penalty factor. The confidence level of spatiotemporal coupling of the original multi-source signals. The confidence level is then reduced by weighting. The calculation is as follows:

[0035] in, The penalty factor, derived from a preset constant, physically represents the weight of the negative impact of missing data on credibility. It is dimensionless, and its value range is typically [value missing]. In this embodiment, the preferred embodiment is... ; The percentage of missing data derived from statistical analysis, physically representing the proportion of missing data out of the total required data, is within a certain range. The specific calculation method adopts the time-dimensional statistical method, and the formula is as follows: ,in, To calculate the number of missing sampling points within the current sliding window, This represents the total number of sampling points that the window should theoretically contain. This step ensures that the confidence level of the final output can objectively reflect the integrity and quality of the data source; This embodiment ensures the robustness of the system when some sensors fail due to extreme weather conditions; by introducing a confidence penalty factor, it maintains basic operational capabilities while preventing misleading alarms caused by over-reliance on interpolated data, demonstrating the system's robustness under data quality fluctuations.

[0036] Example 7: The intelligent early warning execution module performs the following operations to trigger tiered alarm commands: Calculate the rate of change and acceleration of real disaster characteristics per unit time; Input the rate of change and acceleration into a preset disaster risk assessment mapping table to determine the current water situation's hazard level; Based on the level of danger, a corresponding alarm response strategy is matched. The alarm response strategy includes adjusting the data acquisition frequency, sending alarm information through multiple channels, and activating linkage control equipment.

[0037] Based on confirmed disaster characteristic data, the system calculates the rate of water level change. That is, the first derivative and acceleration. That is, the second derivative; where, Indicates the rate of rise in water level. The system will characterize the degree of intensification of the rising water trend; and Input a preset disaster risk assessment mapping table to determine the current water situation's hazard level; the mapping table specifically defines the following numerical limits: setting a rate warning threshold. Acceleration warning threshold It is worth noting that the above thresholds and It is not an absolutely fixed universal standard, but rather a configurable parameter supported by the system; in actual deployment, it allows for regional calibration based on historical disaster data or expert experience in the monitoring area to adapt to rivers with different hydrological characteristics. The specific risk level classification logic is as follows: Extremely high risk level: when and At that time, it was determined that the flood was rising rapidly and the acceleration exceeded the warning threshold; High risk level: when and At that time, it is judged that the water level rises rapidly but the acceleration is low or shows a decelerating trend; medium risk level: when and At that time, it was determined to be a normal rise in water level; if Within this range but If it is, it will be directly upgraded to a high-risk level; low-risk level: when At that time, it was determined to be a minor fluctuation; Based on the determined hazard level, the system matches the corresponding alarm response strategy: for low-risk levels, only the data acquisition frequency is adjusted, such as encrypting it to 1 minute / time; for high-risk levels, multi-channel alarm information is sent and linkage control equipment is activated, such as automatically closing gates or starting drainage pumping stations. This embodiment not only focuses on the triggering of alarms, but also distinguishes the dynamic differences between steady rises and rapid surges by introducing an acceleration index; this refined hierarchical response mechanism enables on-demand allocation of emergency resources and effectively optimizes the efficiency and targeting of disaster response.

[0038] Example 8: The system also includes an adaptive feedback optimization module, used for: Record the comparison information between the validity verification results and the subsequent manual review results; When the comparison information shows false alarms or omissions, the corresponding multimodal heterogeneous sensor data samples are extracted. The pseudo-anomaly feature recognition model is incrementally trained using multimodal heterogeneous sensor data samples to update the model's parameters.

[0039] The system automatically records the validity verification results of each generation and receives the results of subsequent manual review, generating comparison information. In response to the comparison information showing false alarms or omissions, the system automatically extracts the multimodal heterogeneous sensor data corresponding to that time period, especially the video frames and sensor waveforms that caused the misjudgment, and marks them as difficult samples. The system uses these difficult samples to incrementally train the pseudo-anomaly feature recognition model, that is, without forgetting the original knowledge, it uses new data to fine-tune the model parameter weights to update the model's recognition ability. This embodiment constructs a closed-loop mechanism of monitoring, alarm, feedback, and updating, solving the problem of traditional early warning systems being deployed once and fixed for life. By continuously adapting to the unique environmental characteristics of specific monitoring points, such as plant interference in specific seasons, the system achieves self-iteration and optimization of algorithm performance, significantly extending the system's life cycle and scope of application.

[0040] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-source perception-based intelligent early warning and alarm system for sudden water disaster, characterized in that, The method comprises the following steps: A multi-source data acquisition module is configured to acquire multi-modal heterogeneous sensor data in a monitoring area, including water level time series data, rainfall time series data, cross-section flow rate data, and on-site monitoring video stream; An abrupt anomaly preliminary screening module is configured to identify abrupt jump features with a change rate of data amplitude exceeding a preset jump threshold based on the water level time series data or the cross-section flow rate data, and mark data segments with the abrupt jump features as candidate anomaly signals; A space-time coupling verification module is configured to call rainfall time series data associated with the candidate anomaly signals in time and space dimensions in response to the candidate anomaly signals, and calculate a multi-source signal space-time coupling confidence degree representing the degree of association between rainfall driving and water level change; A dynamic effectiveness evaluation module is configured to call the on-site monitoring video stream associated with the candidate anomaly signal, and determine whether the candidate anomaly signal belongs to a real disaster feature or environmental noise interference based on the multi-source signal space-time coupling confidence degree and analysis results of the on-site monitoring video stream, including identifying whether there is a non-water body foreign matter obstruction or biological disturbance feature at the monitoring point, and generating an effectiveness verification result; An intelligent early warning execution module is configured to trigger a hierarchical alarm instruction according to the evolution trend of the abrupt jump feature when the effectiveness verification result indicates a real disaster feature. 2.The multi-source perception based intelligent early warning and alarm system for sudden water disaster according to claim 1, characterized in that, The space-time coupling verification module performs the following operations to construct the multi-source signal space-time coupling confidence degree: Determine the occurrence time of the candidate anomaly signal and the corresponding monitoring point; Retrieve rainfall time series data within a preset range upstream of the monitoring point; Based on a preset nonlinear hysteresis effect model, analyze the causal logic relationship between the rainfall time series data and the candidate anomaly signal, and generate a hydrology logic matching score; Map the hydrology logic matching score to the basic score of the multi-source signal space-time coupling confidence degree. 3.The multi-source perception based intelligent early warning and alarm system for sudden water disaster according to claim 2, characterized in that, The space-time coupling verification module is also used for: Collecting soil saturation parameters and topographic features around the monitoring point; Based on the soil saturation parameters and the topographic features, dynamically adjusting the hysteresis time window in the nonlinear hysteresis effect model; Within the hysteresis time window, verify whether the cumulative amount of rainfall time series data meets the minimum physical condition for triggering the candidate anomaly signal to correct the hydrology logic matching score. 4.The multi-source perception based intelligent early warning and alarm system for sudden water disaster according to claim 1, characterized in that, The dynamic effectiveness evaluation module performs the following operations to generate the effectiveness verification result: Call a preset pseudo-anomaly feature recognition algorithm model to perform image semantic analysis on the on-site monitoring video stream; Identify whether there is a non-water body foreign matter obstruction or biological disturbance feature at the monitoring point, and generate an environmental noise probability value quantifying the degree of environmental interference; Based on a preset weighted fusion strategy, combine the multi-source signal space-time coupling confidence degree and the environmental noise probability value to calculate a comprehensive confidence index; If the comprehensive confidence index is greater than a preset alarm trigger threshold, generate an effectiveness verification result indicating a real disaster feature; If the comprehensive credibility index is less than or equal to the alarm trigger threshold, an effectiveness verification result indicating environmental noise interference is generated and the alarm is suppressed.

5. The intelligent early warning and alarm system for sudden water disaster based on multi-source perception according to claim 4, characterized in that, The system further comprises a dynamic threshold regulation module configured to: monitor the rainfall duration and upstream water trend in the multi-modal heterogeneous sensor data; when it is confirmed that the rainfall or upstream water level is continuously rising, calculate a dynamic threshold adaptive convergence duration; within the dynamic threshold adaptive convergence duration, reduce the alarm trigger threshold according to a preset attenuation strategy to improve the sensitivity of the system to subsequent sudden water conditions. 6.The multi-source perception based intelligent early warning and alarm system for sudden water disaster according to claim 1, characterized in that, The spatio-temporal coupling verification module further performs the following operations when constructing the multi-source signal spatio-temporal coupling confidence: detect whether there is data missing in part of the dimensions of the multi-modal heterogeneous sensor data; if there is data missing, call the same type of sensor data of the adjacent monitoring point and the historical synchronous data for interpolation completion; based on the completed data, cross-check and introduce a confidence penalty factor according to the data missing ratio to reduce the weight of the multi-source signal spatio-temporal coupling confidence. 7.The multi-source perception based intelligent early warning and alarm system for sudden water disaster according to claim 1, wherein, The intelligent early warning execution module performs the following operations to trigger a hierarchical alarm instruction: calculate the change rate and acceleration of the real disaster feature in unit time; input the change rate and acceleration into a preset disaster risk assessment mapping table to determine the danger level of the current water condition; according to the danger level, match the corresponding alarm response strategy, which includes adjusting the data acquisition frequency, sending multi-channel alarm information and starting the linkage control equipment. 8.The multi-source perception based intelligent early warning and alarm system for sudden water disaster according to claim 4, characterized in that, The system further comprises an adaptive feedback optimization module for: record the comparison information of the effectiveness verification result and the subsequent manual review result; when the comparison information shows false positives or false negatives, extract the corresponding multi-modal heterogeneous sensor data sample; use the multi-modal heterogeneous sensor data sample to incrementally train the pseudo abnormal feature recognition model to update the parameters of the pseudo abnormal feature recognition model.

Images