A multi-disaster integrated perception early warning method and system

Abstract

This invention relates to the field of disaster monitoring and early warning technology, specifically to a multi-hazard integrated sensing and early warning method and system. The method includes multi-source data acquisition, preprocessing, fusion, and risk calculation; generating disaster risk values ​​through local inference using lightweight random forest and LSTM models; dynamic threshold adjustment and early warning linkage; initializing early warning thresholds and transmitting early warning commands via adaptive communication scheduling; optimizing response results through dual indicators of response timeliness and emergency action effectiveness, and further optimizing decision parameters using parameter fine-tuning formulas. The system includes a distributed integrated acquisition sound, light, and electricity alarm station, a multi-source data fusion module, an edge computing early warning module, etc., forming a closed loop with in-home sensor alarms. This invention significantly improves the comprehensiveness of disaster monitoring, the timeliness of early warning response, and the reliability of communication transmission, while enhancing the effect of in-home early warning linkage, adapting to various complex application scenarios, and providing efficient and reliable technical support for disaster emergency prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of disaster monitoring and early warning technology, specifically relating to a multi-hazard integrated sensing and early warning method and system. Background Technology

[0002] In existing disaster prevention and control scenarios, the design logic of traditional monitoring equipment is mostly centered around a single disaster. For example, debris flow monitoring equipment is only equipped with geological sensors such as tilt angle and fissure sensors, while flash flood monitoring equipment focuses only on hydrological sensors such as water level and flow velocity sensors. These devices are deployed independently and have limited functions, only capable of collecting basic elements for specific disasters. With the increasing frequency of extreme weather events, scenarios involving multiple disasters simultaneously, such as flash floods and debris flows, and landslides accompanied by collapses, are becoming more common. The singular design of traditional equipment is no longer sufficient to meet the monitoring and early warning needs of complex disasters. The industry urgently needs an integrated solution that can achieve collaborative monitoring and rapid response for multiple disasters. However, traditional disaster monitoring equipment has several key shortcomings that severely restrict the effectiveness of monitoring and early warning:

[0003] (1) The sensor functions are scattered and the data is isolated. The monitoring equipment for different types of disasters collects data separately, which makes it impossible to achieve the intercommunication and correlation analysis of multi-dimensional data such as rainfall, water level, and deformation, forming "data islands" and making it difficult to accurately identify the risk of multi-hazard coupling. (2) Data processing relies on a centralized platform. All collected data must be remotely transmitted to the central platform before it can be analyzed. The transmission delay is high (often reaching minutes or more). Disasters such as flash floods and mudslides are extremely sudden, and the delay problem directly leads to untimely warnings and missing the best evacuation time. (3) There is a gap in the home alarm process. The existing home alarm equipment only has simple sound and light alarm functions. It cannot be linked with the front-end monitoring terminal to receive accurate early warning instructions (such as disaster type, scope of impact, escape route) or provide feedback on the activity status of people in the house, making it difficult to form an early warning closed loop. (4) The equipment has poor environmental adaptability and a single power supply method (mostly relying on mains power or single-path solar energy). It is prone to power outages in the wild when there is no mains power or continuous rain. Furthermore, the interface compatibility is weak, making it difficult to flexibly connect to third-party sensors or emergency equipment, and it is difficult to operate stably in harsh outdoor environments for a long time. (5) Data transmission and security are potentially problematic. The system often uses a single communication method (such as relying solely on 4G), which is prone to interruption in areas with no signal coverage, such as mountainous regions. Furthermore, data transmission is mostly in plaintext and lacks encryption mechanisms, posing a risk of being tampered with or stolen.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems in the existing technology, the present invention provides a multi-hazard integrated sensing and early warning method and system, which solves the problems of single-hazard monitoring, early warning delay, unreliable communication and gap in home early warning in traditional disaster monitoring equipment.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: Firstly, a multi-hazard integrated sensing and early warning method includes: S1. Multi-source data acquisition and preprocessing: Collect core disaster-causing data, environmental auxiliary data, and equipment status data through the multi-hazard monitoring sensor group of the distributed integrated audio-visual alarm station; S2. Multi-source data fusion and risk level calculation: Dynamic weight allocation and weighted fusion are performed on the data streams preprocessed in step S1 to generate a comprehensive disaster risk index; S3. Dynamic Threshold Adjustment and Early Warning Linkage: Based on the regional historical disaster dataset, the baseline thresholds for each level of early warning are initialized using the percentile method; corresponding response actions are executed in conjunction with the early warning level and the status of the communication link; at the same time, early warning instructions are transmitted through dynamic switching between at least two communication links; the distributed integrated acquisition sound and light alarm station is linked with the in-home sensor alarm; the in-home sensor alarm transmits back the status data of the people inside the house; the data transmission process adopts a dual encryption mechanism to ensure security; S4. Response Result Optimization: The effectiveness of the early warning response is evaluated by preset evaluation indicators, and the decision parameters and linkage triggering conditions are iteratively optimized by using parameter fine-tuning formulas based on the evaluation results.

[0007] Furthermore, the core disaster-causing data includes: rainfall data, water level data, soil moisture content data, dip angle data, flow velocity data, and fissure data; The sampling frequency for rainfall and water level data is 1 time / minute, the sampling frequency for soil moisture content and tilt angle data is 1 time / 5 minutes, and the sampling frequency for flow velocity and fissure data is 1 time / 10 minutes. Environmental auxiliary data includes: wind speed data, humidity data, and slope data; The environmental auxiliary data is collected once every 10 minutes. Device status data includes: sensor power data, communication module signal strength data, and CPU / memory usage data; The device status data is collected once every 30 minutes.

[0008] Furthermore, multi-source data acquisition and preprocessing specifically include: Outlier Removal: For abnormal data generated by sensors due to environmental interference or malfunction in disaster monitoring scenarios, outliers are removed using the 3σ principle. Data calibration: To address the impact of environmental factors on disaster-causing data, a linear formula is used to correct for data bias. Time-series data segmentation: In view of the characteristic that disaster risks accumulate over time, time-series disaster-causing data are segmented into time window samples of fixed length to adapt to model inference; Data standardization: To avoid the impact of differences in the magnitude of data from different dimensions on model inference, the Min-Max standardization formula is used to unify the data to a preset numerical range.

[0009] Furthermore, the outlier removal adopts... The principle is to remove outliers, and the specific formula is as follows: For a dataset of a certain indicator Calculate the mean and standard deviation If the data satisfy If the value is 0, it is considered an outlier and removed; among them,

[0010]

[0011] For the i-th collected data, The number of samples in the dataset. The mean of the data. The standard deviation of the data; The linear formula for the data calibration is:

[0012] in, The calibrated soil moisture content (%). The original soil moisture content (%) was measured. Humidity influence coefficient; The relative humidity of the environment (%). The formula for the number of samples in time series data segmentation is: The kth sample is ;in, Let i be the time series data at time i. This represents the total number of time points in the time series data. The length of the time window; The total number of samples; The sample number; The Min-Max standardization formula is:

[0013] in, The data is standardized. This is the original collected data; This is the historical minimum value of this indicator; This is the historical maximum value of this indicator.

[0014] Furthermore, step S2 also includes: Lightweight processing is performed on the pre-trained model to adapt it to the low-power embedded environment of the edge terminal. The lightweight processing includes: selecting an appropriate lightweight algorithm based on the risk factor type in multi-hazard monitoring, removing calculation branches in the algorithm whose contribution is lower than a preset threshold, converting the data precision from high-order floating-point numbers to low-order floating-point numbers or integers, and adapting it to the embedded deployment framework; the lightweight algorithm is used to perform inference calculations on the pre-processed standardized data and output a disaster risk value within a preset score range.

[0015] Furthermore, the lightweight tree-based classification algorithm calculates the risk value using a weighted voting method, with the specific formula as follows:

[0016] in, The risk value output by the random forest. For the number of decision trees, Let be the weight of the i-th decision tree; The risk category score for the i-th decision tree; The lightweight temporal inference algorithm calculates risk values ​​through cell state updates and mapping to fully connected layers: the cell state update process includes parameter calculations for the forget gate, input gate, and output gate to preserve key temporal features; the risk value is mapped to a preset score range through the fully connected layer, with the specific formula as follows:

[0017] in, This is the weight matrix. For bias terms, Hide the current state.

[0018] Furthermore, step S3 specifically includes: The threshold initialization adopts the percentile method, and the baseline threshold for each level of early warning is determined based on the set of historical disaster risk values ​​in the region. Each level of early warning corresponds to a preset percentile parameter. The communication link status is determined by periodically detecting signals to determine whether it is smooth or interrupted; the response actions include mandatory local actions that are not affected by the link status, as well as additional actions such as synchronizing data to the remote platform when the link is smooth, encrypting and storing data locally when the link is interrupted, and retransmitting the data after the link is restored.

[0019] Furthermore, the specific formula for threshold initialization is as follows:

[0020] in, This is the baseline threshold for the k-th level warning. This is a set of historical disaster risk values ​​for the region. The percentiles corresponding to each level of warning are as follows: This is the percentile function.

[0021] Furthermore, the dynamic switching transmission warning command in step S3 is implemented through adaptive communication scheduling. This adaptive communication scheduling includes data priority sorting, link status monitoring, scheduling decision-making, and store-and-forward. Specific steps include: Classify the data to be transmitted according to business importance and assign priority levels, and build a priority queue; Real-time detection of the quality and availability of each communication link, generating link status assessment results; Based on the data priority and link status assessment results, the communication link for data transmission is determined. When all communication links are interrupted, data is cached locally and retransmitted according to priority once the links are restored.

[0022] Furthermore, the data priority sorting specifically includes: The data is categorized into four types based on business type: critical early warning data, important status data, routine monitoring data, and auxiliary and management data, corresponding to four priority levels. A priority queue is constructed using a min-heap data structure. The data packets with the highest priority level are transmitted first. The data packet encapsulation includes data payload, priority identifier, timestamp, retry count and destination address metadata.

[0023] Furthermore, the link status monitoring and scheduling decision specifically includes: The 4G link determines its status by periodically sending TCP / UDP heartbeat packets, monitoring signal strength, network latency, and data packet success rate, and is divided into three levels: smooth, poor, and interrupted. LoRa links determine their status by reading radio frequency parameters and monitoring the success rate of confirmation messages, and are divided into three levels: smooth, poor, and interrupted. The scheduling decision logic is as follows: critical early warning data is transmitted using the 4G link first, and switches to the LoRa link when the 4G link is interrupted; important status data uses 4G when the 4G link is normal, and switches to LoRa when the 4G link is poor or interrupted and the LoRa link is smooth; routine monitoring data and auxiliary management data are transmitted via the selected transmission link or delayed transmission according to the link status and resource usage.

[0024] Furthermore, the store-and-forward mechanism specifically includes: When all communication links are interrupted, data packets are stored in a wait-to-retransmit queue in non-volatile memory. Once any communication link is restored, data in the retransmission queue will be retransmitted first, according to data priority.

[0025] Furthermore, the preset evaluation indicators include: response timeliness rate. Effectiveness of emergency response By fine-tuning the decision parameters, the frequency of audible and visual alarms and the trigger threshold of emergency equipment are adjusted to achieve iterative optimization of decision rules; The response timeliness The specific formula is:

[0026] in, This refers to the number of times a local action is completed within a preset time after an alert is triggered. Total number of warnings; The effectiveness of the emergency response The specific formula is:

[0027] in, This refers to the number of times disaster losses are reduced after coordinated actions. This represents the total number of times the emergency action was triggered. The specific formula for fine-tuning the decision parameters is as follows:

[0028] in, The adjusted decision parameters, The decision parameters before adjustment To adjust the coefficient, Evaluation indicators.

[0029] Secondly, a multi-hazard integrated sensing and early warning system includes: A distributed integrated acquisition sound, light, and electricity alarm station integrates a multi-hazard monitoring sensor group and an edge data preprocessing module; the multi-hazard monitoring sensor group is used to collect core disaster-causing data, environmental auxiliary data, and equipment status data, and the edge data preprocessing module is used to perform preprocessing on the collected data; The multi-source data fusion module is connected to the distributed integrated acquisition sound, light, and electricity alarm station. It is used to perform dynamic weight allocation and weighted fusion on the preprocessed data stream output by the edge data preprocessing module to generate a comprehensive disaster risk index. An edge computing early warning module is connected to the multi-source data fusion module and is used to calculate the disaster risk level based on the comprehensive disaster risk index. The dynamic risk model module, connected to the edge computing early warning module, is used to initialize the baseline thresholds for each level of early warning based on the regional historical disaster dataset using the percentile method. A hybrid communication redundancy module is connected to the distributed integrated acquisition sound, light, and electricity alarm station, and supports at least two communication links. It is used to dynamically switch the transmission channel according to the status of the communication link in order to transmit early warning instructions. An in-home sensor alarm is connected to the hybrid communication redundancy module to receive the warning command and execute the alarm operation, and to transmit the status data of the people inside the house. The secure transmission module, embedded in the hybrid communication redundancy module, employs a dual encryption mechanism to ensure the security of the transmission of early warning commands and monitoring data. The decision optimization module is connected to the edge computing early warning module and the in-home sensor alarm, respectively. It is used to evaluate the early warning response effect through preset evaluation indicators, and to iteratively optimize the decision parameters and linkage triggering conditions based on the evaluation results using parameter fine-tuning formulas.

[0030] Compared with existing technologies, the multi-hazard integrated sensing and early warning method and system provided by this invention includes: multi-source data acquisition, preprocessing, fusion, and risk calculation; generating disaster risk values ​​through local inference using lightweight random forest and LSTM models; dynamic threshold adjustment and early warning linkage; initializing early warning thresholds and transmitting early warning commands through adaptive communication scheduling; optimizing response results through dual indicators of response timeliness and emergency action effectiveness, and further optimizing decision parameters using parameter fine-tuning formulas; the system includes a distributed integrated acquisition sound, light, and electricity alarm station, a multi-source data fusion module, an edge computing early warning module, etc., forming a closed loop with in-home sensor alarms; this invention can significantly improve the comprehensiveness of disaster monitoring, the timeliness of early warning response, and the reliability of communication transmission, while strengthening the effect of in-home early warning linkage, adapting to various complex application scenarios, and providing efficient and reliable technical support for disaster emergency prevention and control. Attached Figure Description

[0031] Figure 1 A schematic diagram of a data transmission link provided in an embodiment of the present invention; Figure 2 A flowchart for continuous link status monitoring provided in an embodiment of the present invention; Figure 3 A flowchart of the data priority processing procedure provided in this embodiment of the invention; Figure 4 A flowchart of scheduling decision-making and execution provided for embodiments of the present invention; Figure 5 A flowchart of an integrated perception and early warning system provided in an embodiment of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0033] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0034] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0035] Example 1 See Figure 5 , Figure 5 This is a flowchart of a multi-hazard integrated sensing and early warning method proposed in this invention. Specific steps may include: S1. Multi-source data acquisition and preprocessing: Collecting core disaster-causing data, environmental auxiliary data, and equipment status data through a distributed integrated acquisition system of multi-hazard monitoring sensors in the acoustic, optical, and electronic alarm station; specifically including: S11. Data Input: Distributed integrated acquisition of sound, light, and electricity alarm stations is used for distributed integrated acquisition. Dedicated sensors are deployed for the core disaster-causing factors of the two types of disasters. The acquisition frequency is dynamically adjusted according to the disaster risk, and the acquisition frequency is increased during periods of heavy rainfall. For specific data types, please refer to Table 1. Table 1

[0036] Referring to Table 1, the core disaster-causing data include: rainfall data, water level data, soil moisture content data, tilt angle data, flow velocity data, and fissure data; the collection frequency of rainfall data and water level data is 1 time / minute, the collection frequency of soil moisture content data and tilt angle data is 1 time / 5 minutes, and the collection frequency of flow velocity data and fissure data is 1 time / 10 minutes. Environmental auxiliary data includes: wind speed data, humidity data, and slope data; the environmental auxiliary data is collected once every 10 minutes. The device status data includes: sensor power data, communication module signal strength data, and CPU / memory usage data; the device status data is collected once every 30 minutes.

[0037] S12. Data Preprocessing: Preprocessing is completed locally on the edge terminal (embedded motherboard of the distributed alarm station), without relying on the network. It focuses on eliminating specific interferences in scenarios such as flash floods / mudslides where sensors are covered by heavy rain or mud. Specifically, it includes: S121. Outlier Removal: For abnormal data generated by sensors due to environmental interference or malfunction in disaster monitoring scenarios, the following measures are taken: The principle is to remove outliers; Specifically, disaster monitoring scenarios include abnormal water levels caused by sensors being destroyed by torrential rain, and abnormal tilt angles caused by sensors being covered by mud and sand in debris flow scenarios. The principle is to eliminate extreme values ​​caused by sensor malfunctions or environmental interference in flash flood / mudslide scenarios. Outlier removal adopts... The principle is to remove outliers, and the specific formula is as follows: For a dataset of a certain indicator Calculate the mean and standard deviation If the data satisfy If the value is 0, it is considered an outlier and removed; among them,

[0038]

[0039] For the i-th collected data, The number of samples in the dataset. The mean of the data. The standard deviation of the data; S122. Data calibration: To address the impact of environmental factors on disaster-causing data, a linear formula is used to correct the deviation of disaster-causing data. Specifically, to address the impact of humidity on soil moisture content in flash flood scenarios and the impact of slope on tilt angle monitoring in debris flow scenarios, a linear calibration formula is used to correct the influence of environmental factors on disaster-causing data. Taking soil moisture content calibration as an example, the specific formula is as follows:

[0040] in, The calibrated soil moisture content (%). The original soil moisture content (%) was measured. Humidity influence coefficient; The relative humidity of the environment (%). S123. Time-series data segmentation: In view of the characteristic that disaster risks accumulate over time, time-series disaster-causing data is segmented into time window samples of fixed length to adapt to model inference. Specifically, considering the characteristic of flash flood / debris flow risk accumulating over time, time-series data (such as rainfall and water level) are divided into fixed-length time window samples for LSTM inference: For time series data The formula for the number of samples in time series data segmentation is: The kth sample is ;in, Let i be the time series data at time i. This represents the total number of time points in the time series data. The length of the time window; The total number of samples; The sample number; S124. Data Standardization: To avoid the impact of differences in the magnitude of data from different dimensions on model inference, the Min-Max standardization formula is used to unify the data to a preset numerical range.

[0041] Specifically, data from different dimensions are standardized to the [0,1] interval to avoid the impact of magnitude differences on model inference, using the Min-Max standardization formula:

[0042] in, The data is standardized. This is the original collected data; This is the historical minimum value of this indicator; This is the historical maximum value of this indicator.

[0043] Model inference is the core of edge computing. For the risk characteristics of flash floods and debris flows, two algorithms, Random Forest (static multi-factor inference) and LSTM (temporal cumulative inference), are adapted respectively. Local real-time computing is achieved through lightweight deployment, and the threshold is dynamically adjusted by combining scenario-based historical data to reduce false alarms and missed alarms.

[0044] S2. Multi-source data fusion and risk level calculation: Dynamic weight allocation and weighted fusion are performed on the preprocessed data stream from step S1 to generate a comprehensive disaster risk index; specifically including: S21. Lightweight Model Deployment: Perform lightweight processing on the pre-trained model to adapt it to the low-power embedded environment of edge terminals. The lightweight processing includes: Based on the risk factor types in multi-hazard monitoring, a suitable lightweight algorithm is selected, calculation branches with contributions below a preset threshold are removed, data precision is converted from high-order floating-point numbers to low-order floating-point numbers or integers, and the algorithm is adapted to an embedded deployment framework. The lightweight algorithm is then used to perform inference calculations on the preprocessed standardized data, outputting a disaster risk value within a preset score range, as detailed in Table 2. Table 2 Lightweighting Treatment

[0045] S22. Scenario-based risk value calculation: Preprocessed standardized data is input into the lightweight model, and a scenario-based risk value (0-10 points, the higher the score, the higher the risk) is output. The reasoning process and formulas of the two algorithms are as follows: S221. Random Forest Inference: Used for static multi-factor risk. Random forests use multiple decision trees to vote on static factors such as tilt angle, soil moisture content, and fracture displacement, outputting debris flow risk values. Inference process and formula: Single decision tree decision-making: Each decision tree divides risk categories based on a certain feature threshold, for example: Decision Tree 1: If "tilt angle > 5° and soil moisture content > 40%" → Risk Category 3 (High Risk); otherwise → Risk Category 1 (Low Risk) Decision Tree 2: If "crack displacement > 10mm and dip angle > 3°" → Risk Category 3 (High Risk); otherwise → Risk Category 2 (Medium Risk) Multi-tree voting aggregation: Calculates the final risk value using a "weighted voting formula". The specific formula is as follows:

[0046] in, The risk value output by the random forest. For the number of decision trees, Let be the weight of the i-th decision tree; The risk category score for the i-th decision tree; S222, LSTM Inference: Used for time-series cumulative risk. LSTM captures the cumulative effect of time-series data such as rainfall and water level through memory units, and outputs flash flood risk values. The inference process and core formula are as follows: LSTM cell state update: Cell state is updated collaboratively through forgetting gate, input gate, and output gate, preserving key temporal information (such as 6 consecutive hours of torrential rain). The forgetting gate formula determines how many historical states are forgotten. The specific formula is as follows:

[0047] The update ratio of new information is controlled by an input gate formula. The specific formula is as follows:

[0048]

[0049] By fusing historical and new information through a cell state formula, the specific formula is as follows:

[0050] The timing characteristics are reflected by the output gate formula, which is as follows:

[0051]

[0052] In the above formula, This is the sigmoid activation function (output 0-1). The hyperbolic tangent function (output) , This is the weight matrix (obtained through pre-training). For bias terms, Hide the state from the previous / current time step. Input features for the current time step. This represents the cell state at the previous / current moment. S223, Risk Value Output: Hiding the LSTM state through a fully connected layer. Mapped to a flash flood risk value of 0-10, the specific formula is as follows:

[0053] in, This is the weight matrix. For bias terms, Hide the current state.

[0054] S3. Dynamic Threshold Adjustment and Early Warning Linkage: Based on the regional historical disaster dataset, the baseline thresholds for each level of early warning are initialized using the percentile method; corresponding response actions are executed based on the early warning level and communication link status; simultaneously, early warning commands are transmitted through dynamic switching between at least two communication links; distributed integrated acquisition of sound, light, and electrical alarm stations is linked with in-home sensor alarms, with the in-home sensor alarms transmitting back the status data of people inside the home; data transmission employs a dual encryption mechanism to ensure security; see reference. Figure 1 As shown, it specifically includes: S31. Threshold Initialization: Threshold initialization adopts the percentile method, determining the baseline threshold for each level of early warning based on the set of historical disaster risk values ​​for the region. Each level of early warning corresponds to a preset percentile parameter; the specific formula is as follows:

[0055] in, This is the baseline threshold for the k-th level warning (k=0~4, corresponding to no warning to emergency warning). This is a set of historical disaster risk values ​​for the region. The percentiles corresponding to each level of warning (Level 0 = 10%, Level 1 = 30%, Level 2 = 50%, Level 3 = 80%, Level 4 = 95%). This is the percentile function.

[0056] S32. Response Action Decision: The communication link status is determined by periodically detecting signals to determine whether it is active or interrupted; response actions include mandatory local actions unaffected by link status, as well as additional actions such as synchronizing data to the remote platform when the link is active, and encrypting and storing data locally and retransmitting it after the link is restored when the link is interrupted. See Table 3 for the complete decision matrix and key references. Table 3 Decision Matrix and Key Explanations

[0057] Adaptive communication scheduling for early warning commands: The transmission of the aforementioned early warning commands and monitoring data is achieved through the adaptive communication scheduling function of the hybrid communication redundancy module. This function is completed collaboratively by three core components: the data queue manager, the link status monitor, and the scheduling decision-maker. The specific process is as follows: S321. Data Priority Sorting: The terminal classifies transmitted data and assigns priority levels according to business logic, constructing a minimum-heap priority queue (with the highest priority data at the top of the heap) to ensure that critical data is transmitted first. (See also...) Figure 3 As shown, data classification and priority are defined as follows: Highest priority (CRITICAL): Critical early warning data, including real-time alarm signals for flash floods / mudslides and alarm data for displacement / rainfall exceeding thresholds, requiring extremely low latency and extremely high reliability; High priority (HIGH): Important status data, including device self-test anomalies, battery level alarms, and sensor fault data, requiring low latency and high reliability; Medium priority (NORMAL): Regular monitoring data, including timed collection of sensor readings such as rainfall, water level, and displacement, with a certain delay allowed; Low priority: Auxiliary and management data, including device logs, configuration query responses, and remote upgrade confirmation packets, can be transmitted when the link is idle.

[0058] The data priority queue is constructed using a min-heap data structure. The core feature of this structure is that the top of the heap always contains the data packet with the lowest priority value (i.e., the highest actual priority level), ensuring that critical data is scheduled first. Each data packet to be sent is encapsulated into a unified structure, which contains: Metadata such as data payload (e.g., early warning commands, sensor monitoring readings), priority identifiers (pre-defined categories corresponding to key early warning data, important status data, etc.), timestamps (data generation time), retry counts (records of the number of retries after transmission failure), and destination addresses (e.g., central platform IP, in-home alarm device identifier) ​​support the orderliness and traceability of data transmission.

[0059] When the application layer generates data that needs to be transmitted, such as sensor readings or alarm trigger signals, the system first matches the corresponding priority label according to the data type, and then inserts the encapsulated structure into the min-heap queue. When the scheduling decision-maker is ready to execute data transmission, it will directly retrieve the data packet with the highest current priority from the head of the queue (i.e., the top of the heap) without traversing the entire queue.

[0060] This minimum-heap-based scheduling mechanism ensures that even if low-priority data is generated and inserted into the queue before high-priority data, subsequently generated high-priority data can be "jumped" to the head of the queue through dynamic adjustment of the heap structure. This effectively avoids transmission delays caused by low-priority data blocking and ensures the timeliness of transmission of critical data (such as real-time alarm signals for flash floods / mudslides).

[0061] S322, Link Status Monitoring: Real-time detection of the quality and availability of 4G and LoRa links, providing a basis for scheduling decisions. S3221, 4G link status detection: Status detection is achieved through a dual mechanism of periodic heartbeat detection and data transmission feedback. S32211. Send small TCP / UDP heartbeat packets to the access server deployed on the cloud platform every 30 seconds, and directly determine the link connectivity by whether the heartbeat packets time out. S32211. After each data transmission via 4G, the link transmission result is updated in real time based on whether a TCP-ACK or application layer confirmation response is received. Specifically, the 4G link quality is comprehensively evaluated through four status indicators, including connectivity reflecting the basic connection status, RSRP and RSRQ (LTE standard indicators) reflecting signal stability, network latency (Ping round-trip time) measuring transmission efficiency, and the success rate of the most recent N data packets reflecting the long-term reliability of the link. S3222, LoRa Link State Probing: Focusing on Wireless Radio Frequency Characteristics and Protocol Acknowledgment Mechanisms for State Probing S32221: Periodically read the core RF parameters of the LoRa module, and intuitively judge the signal strength and anti-interference capability of the wireless link by receiving the signal strength indicator (RSSI) and signal-to-noise ratio (SNR); S32222 utilizes the ConfirmedData message mechanism natively supported by the LoRaWAN protocol. After sending data, it waits for downlink confirmation from the network server. If no confirmation response is received multiple times consecutively, it is determined that the link quality can no longer meet the transmission requirements. Its status assessment mainly relies on two types of core indicators: RSSI and SNR, which directly reflect the quality of the wireless signal, and the confirmation message success rate, which reflects the data transmission confirmation rate, to ensure accurate capture of the real-time availability of the LoRa link.

[0062] S323, State Decision Algorithm: See [link / reference] Figure 2As shown, accurate judgment is achieved by real-time monitoring of multi-dimensional core indicators of the communication link. These indicators include Received Signal Power Ratio (RSRP), Received Signal Strength Indicator (RSS), network latency, and packet loss rate. The algorithm performs comprehensive analysis and weighted evaluation of these indicators, which not only determines whether the link is available, but also further classifies the corresponding quality level. Ultimately, it provides accurate and reliable link status input to the adaptive communication scheduling module, laying the foundation for subsequent transmission link selection.

[0063] In practical scenarios, link status determination needs to be combined with on-site testing and calibration of key thresholds. Taking the signal received power (RSRP) of a 4G communication module as an example, the optimal threshold for signal strength is determined through on-site testing. Generally, an RSRP > -90dBm is considered a good signal, and an RSRP < -110dBm is considered an unusable signal. The specific determination process consists of two steps: First, based on the periodic heartbeat packets from the 4G module, check whether the most recent heartbeat was successful (i.e., whether a response was received within the preset timeout period). If the heartbeat fails, the link status is directly determined to be out of order (DOWN). Second, if the heartbeat is successful, the average RSRP is further calculated and checked. If the average RSRP > -90dBm, the link status is determined to be good (GOOD). If the average RSRP < -110dBm, the link status is still determined to be out of order (DOWN). If the average RSRP is between -110dBm and -90dBm, the link status is determined to be poor (POOR).

[0064] S324, Comprehensive Judgment of the Scheduling Decision Maker: See [link / reference] Figure 4 As shown, the scheduler combines data priority and link status to make the final transmission decision. Its decision-making logic can be described using a state machine or a decision table, see Table 4: Table 4 Sending Decision Logic Table

[0065] S324, Store-and-Forward Mechanism: The core means to ensure no data loss in network outage scenarios. Its storage logic focuses on data packet preservation during a complete link outage. When the scheduling decision-maker determines that "data needs to be sent but all communication links are in an outage (DOWN) state," the data packet will not be directly discarded, but will be automatically transferred to the non-volatile memory of the hybrid communication redundancy module. This memory has a preset "waiting retransmission queue" specifically used to temporarily store data to be transmitted during the network outage. Because the non-volatile memory has the characteristic of not losing data when power is off, even if the edge terminal encounters a temporary power outage, the data in the queue (such as flash flood warning instructions and critical monitoring data) can still be safely retained, fundamentally avoiding the loss of important data due to link interruption.

[0066] Its forwarding logic revolves around the retransmission process after the link is restored, ensuring the orderly transmission of temporarily stored data. When the status of any communication link (4G or LoRa) recovers from down to good or poor, the scheduling decision-maker will prioritize triggering the "waiting retransmission queue" check mechanism, rather than directly transmitting newly generated data. During the retransmission process, the data priority sorting rules are strictly followed, namely, in the order of "critical early warning data > important status data > routine monitoring data > auxiliary and management data", the data in the queue is retransmitted one by one to the target server (such as the central platform, regional emergency command center), which not only ensures the timeliness of high-priority data transmission during the network outage, but also maintains consistency with the priority logic of the overall adaptive communication scheduling.

[0067] Meanwhile, the secure transmission mechanism employs AES encryption and SSL / TLS dual-link encryption (HTTPS+MQTT) to prevent data tampering and theft. This module uses AES-256 to encrypt the data itself, ensuring data confidentiality; then, it uses the TLS / SSL protocol to establish a secure transmission channel, encrypting and authenticating the transmission process (two-way certificate verification). Simultaneously, a hash algorithm (such as SHA-256) is used to generate a data digest, which the receiver verifies to ensure data integrity. This addresses the security vulnerabilities of traditional monitoring equipment's plaintext data transmission, ensuring the security and reliability of monitoring data throughout the entire process from collection to reporting.

[0068] S4. Response Result Optimization: The effectiveness of the early warning response is evaluated using preset assessment indicators. Based on the assessment results, a parameter fine-tuning formula is used to iteratively optimize the decision parameters and linkage triggering conditions. Specifically, this includes: S41. Decision-making effectiveness pre-set evaluation indicators: The evaluation will use two core indicators—response timeliness and emergency response effectiveness—to assess whether the decision has achieved its intended goals: Response Timeliness Effectiveness of emergency response By fine-tuning the decision parameters, the frequency of audible and visual alarms and the trigger threshold of emergency equipment are adjusted to achieve iterative optimization of decision rules; Response timeliness The specific formula is:

[0069] in, This refers to the number of times a local action is completed within a preset time after an alert is triggered. Total number of warnings; Effectiveness of emergency actions The specific formula is:

[0070] in, This refers to the number of times disaster losses are reduced after coordinated actions. This represents the total number of times the emergency action was triggered. The specific formula for fine-tuning the decision parameters is as follows:

[0071] in, The adjusted decision parameters, The decision parameters before adjustment To adjust the coefficient, Evaluation indicators.

[0072] Example 2 This invention proposes a multi-hazard integrated sensing and early warning system, comprising: A distributed integrated acquisition sound, light, and electricity alarm station integrates a multi-hazard monitoring sensor group and an edge data preprocessing module; the multi-hazard monitoring sensor group is used to collect core disaster-causing data, environmental auxiliary data, and equipment status data, and the edge data preprocessing module is used to perform preprocessing on the collected data; The multi-source data fusion module is connected to the distributed integrated acquisition sound, light, and electricity alarm station. It is used to perform dynamic weight allocation and weighted fusion on the preprocessed data stream output by the edge data preprocessing module to generate a comprehensive disaster risk index. An edge computing early warning module is connected to the multi-source data fusion module and is used to calculate the disaster risk level based on the comprehensive disaster risk index. The dynamic risk model module, connected to the edge computing early warning module, is used to initialize the baseline thresholds for each level of early warning based on the regional historical disaster dataset using the percentile method. A hybrid communication redundancy module is connected to the distributed integrated acquisition sound, light, and electricity alarm station, and supports at least two communication links. It is used to dynamically switch the transmission channel according to the status of the communication link in order to transmit early warning instructions. An in-home sensor alarm is connected to the hybrid communication redundancy module to receive the warning command and execute the alarm operation, and to transmit the status data of the people inside the house. The secure transmission module, embedded in the hybrid communication redundancy module, employs a dual encryption mechanism to ensure the security of the transmission of early warning commands and monitoring data. The decision optimization module is connected to the edge computing early warning module and the in-home sensor alarm, respectively. It is used to evaluate the early warning response effect through preset evaluation indicators, and to iteratively optimize the decision parameters and linkage triggering conditions based on the evaluation results using parameter fine-tuning formulas.

[0073] Example 3 This embodiment employs a multi-hazard integrated sensing and early warning method and system proposed in this invention, ensuring that the system is adaptable to the monitoring and early warning needs of multiple hazards such as flash floods, debris flows, and landslides, and can operate stably in complex field environments. Specifically, it includes: A1. Hardware Deployment and Configuration A11. Deployment of Distributed Integrated Acquisition Sound, Light, and Optical Alarm Stations: Distributed integrated acquisition sound, light, and optical alarm stations are installed based on differentiated site selection according to disaster type. The core configuration and deployment logic are as follows: Multi-hazard monitoring sensor system installation: For debris flow monitoring, alarm stations are deployed in high-risk areas such as mountain slopes and valleys, equipped with tilt sensors (to monitor the tilt angle of the mountain in real time), crack sensors (to capture changes in crack width), tipping bucket rain gauges (to collect rainfall in real time), and GNSS positioning modules (to monitor minute surface displacements); For flash flood monitoring, alarm stations are deployed at intervals of 500-1000 meters on both sides of the river, equipped with ultrasonic or pressure-type water level sensors (to monitor changes in river water level), Doppler radar current meters (to measure water flow velocity), and 1080P high-definition video cameras (to monitor water flow patterns and river conditions), and simultaneously equipped with GNSS modules to monitor dynamic changes in river topography.

[0074] Power supply: It adopts a dual-mode intelligent switching power supply of "solar power + mains power", and is equipped with a 100W photovoltaic panel and a 50Ah high-capacity lithium battery. In normal scenarios, solar power is used first, and it automatically switches to mains power in extreme environments such as continuous rain or no sunlight. The device has a built-in power management chip, which supports low-power sleep mode and has a standby power consumption of ≤1W, ensuring long-term stable operation in the wild without mains power.

[0075] Communication module configuration: The communication adopts a redundant design of "LoRa self-organizing network + 4G public network". In mountainous areas without public network signal, a LoRa self-organizing network is built with the 433MHz frequency band. The alarm station acts as a relay node to upload data to the gateway node with 4G signal level by level, with a transmission distance of ≤5km. By default, data is uploaded through the 4G network first. If the 4G signal is interrupted, it will automatically switch to LoRa or ZigBee self-organizing network and start local caching (storage capacity ≥32GB) to support 72 hours of data temporary storage to avoid data loss due to network failure.

[0076] Interface connectivity: Equipped with a rich array of expansion interfaces, external sensors (such as tipping bucket rain gauges) can be connected via an RS485 interface, video signals can be transmitted via an HDMI or Ethernet interface, and alarm devices such as warning lights and horns can be controlled via a switch interface, making it compatible with third-party monitoring equipment and improving system scalability.

[0077] A12. Deployment of In-Home Sensor Alarms: In-home sensor alarms focus on "last-mile" early warning. Installation and configuration are as follows: Indoor installation: The alarm is fixed at the entrance of the resident's house, the human infrared sensor is pointed towards the core activity area of ​​the house, and the personnel counting module (based on binocular camera or infrared beam) covers the main passages such as the entrance door and living room to ensure accurate monitoring of the activity status of people in the house.

[0078] Power supply: Prioritizes connection to 220V AC mains power. It has a built-in 2000mAh backup battery and automatically switches to battery power after AC mains power failure. The battery life is ≥48 hours, avoiding interruption of the warning due to power failure.

[0079] Linkage Configuration: By binding with the nearest distributed integrated acquisition sound, light and electricity alarm station through the 4G / WWAN network, it can receive early warning instructions issued by the alarm station in real time (such as red warning for flash floods and yellow warning for mudslides). At the same time, it can feed back information such as the number of people in the house and their activity status to the upper-level platform, forming a closed loop from early warning to feedback.

[0080] A2. Function Implementation and Parameter Settings A21, Data Acquisition and Edge Computing Data acquisition parameters: The sensor sampling frequency is dynamically adjusted according to the disaster risk. In normal mode, rainfall and water level data are collected once per minute, soil moisture content and tilt angle data are collected once every 5 minutes, and flow velocity and fissure data are collected once every 10 minutes. When a rainstorm (rainfall ≥ 10 mm / h) is detected, it automatically switches to rainstorm mode, and the sampling frequency of rainfall and water level is increased to once per 10 seconds. The tilt angle and acceleration sensors are collected at a high frequency of 10 times per second to accurately capture the precursor signals of geological disasters.

[0081] Data preprocessing: Preprocessing is completed locally on the edge terminal (embedded motherboard) of the distributed alarm station, without relying on an external network: Abnormal data caused by sensor failure or environmental interference is eliminated by the 3σ principle (such as abnormal water levels caused by sensors being destroyed by rainstorms); the influence of environmental factors is corrected by linear formulas (such as calibrating the original collected data based on the influence of humidity on soil moisture content); the time series data such as rainfall and water level are divided into fixed time windows (60 minutes for flash floods and 120 minutes for debris flows) to adapt to subsequent model inference; the multi-dimensional data are unified to the [0,1] interval by Min-Max standardization to eliminate the influence of magnitude differences on the model.

[0082] Edge computing algorithms: The debris flow early warning model achieves early warning by calculating the exponent of the product of deformation rate (tilt angle change Δθ / Δt) and rainfall intensity (mm / h). When the exponent is ≥50 (empirical threshold), an alarm is triggered. The flash flood early warning model calculates the flow rate based on the water level-flow velocity relationship formula Q=A×V (Q is the flow rate, A is the cross-sectional area of ​​the river channel, and V is the flow velocity). It dynamically adjusts the alarm water level threshold by combining historical flood data. When the actual water level reaches the threshold, an early warning is activated.

[0083] A22. Warning Triggering and Response Local alarm station response: When edge computing determines that the disaster risk level is ≥3 (out of 5), the local audible and visual alarm is immediately activated—the siren volume is ≥90dB and the warning light flashes at a frequency of 2Hz; at the same time, the video acquisition module is triggered to record 1080P resolution, ≥5 minutes of on-site footage, which is then stored and uploaded to the cloud platform; in addition, the warning signal is pushed to in-home sensor alarms within 1 kilometer of the surrounding area via LoRa broadcast, realizing "local warning + surrounding linkage".

[0084] In-home alarm linkage: After receiving the early warning command from the alarm station, the in-home sensor alarm activates a multi-modal alarm: the voice broadcast module plays customized early warning content in a loop (such as "A flash flood has occurred in the southeast direction. Please evacuate immediately to the refuge square on the north side"), and the red warning light flashes simultaneously; the personnel counting module counts the number of people in the house in real time. If the number of people in the house is not cleared within 10 minutes, it automatically reports the "suspected personnel stranded" status to the rescue platform, providing a basis for rescue decision-making.

[0085] A23. Security and Disaster Recovery Mechanisms Data encryption: Locally stored data is protected using the AES-256 encryption algorithm, and data transmitted over the public network is encrypted using the HTTPS protocol. The MQTT communication protocol is further enhanced with TLS 1.2 two-way authentication to prevent data from being tampered with or stolen. At the same time, the SHA-256 hash algorithm is used to generate data digests, and the recipient verifies the data integrity by verifying the digests.

[0086] Resume transmission after network outage: When communication links such as 4G and LoRa are interrupted, the data is temporarily stored in local non-volatile memory; after any link is restored, the warning data in the cache is retransmitted in the order of timestamps to ensure that critical warning information is not lost.

[0087] A3. Examples of Multi-Scenario Applications A31. Debris Flow Monitoring and Early Warning Scenarios Deployment: In high-risk debris flow valley areas, deploy 3 distributed integrated acquisition sound, light, and electricity alarm stations at intervals of ≤500 meters to form a triangular monitoring network and achieve full coverage of the area.

[0088] Triggering conditions: A single station monitors a change in tilt angle ≥5° / hour, and rainfall ≥30mm within 1 hour; multi-station data fusion analysis shows a consistent trend in mountain deformation (spatial correlation ≥80%), which is judged as a high risk of debris flow.

[0089] Response process: The alarm station activates the local audible and visual alarm and pushes the warning information to all home alarms within 1 kilometer through the LoRa self-organizing network; the upper-level platform generates a heat map of the disaster impact range based on multi-station data, and the command center issues precise evacuation instructions to the home alarms through the 4G network, and simultaneously dispatches rescue forces to high-risk areas.

[0090] A32. Nighttime monitoring and indoor response scenarios for flash floods Data collection: Distributed alarm stations along the river monitored a 2-meter rise in the river level and a flow velocity of ≥3m / s within 2 hours, reaching the conditions for triggering a flash flood warning.

[0091] Edge computing: Combining historical flood models, it predicts that the flood will spread to downstream residential areas in 2 hours and immediately generates a red alert for flash floods.

[0092] In-home linkage: The in-home alarm automatically increases the alarm volume to ≥100dB at night (ensuring that residents can hear it while sleeping) and repeatedly broadcasts "Flood is approaching, please evacuate to higher ground on the east side of the house"; the personnel counting module continuously monitors the status inside the house, and if no personnel activity is detected for 10 consecutive minutes, it automatically marks "evacuated" to reduce the ineffective investment of rescue resources.

[0093] A33, Monitoring Scenario for Landslide Hazard Points Sensor configuration: Crack sensors with an accuracy of 0.1 mm are installed at the rear edge of the landslide body, and GNSS modules with a positioning accuracy of ±2 mm are deployed at the front edge to monitor the dynamics of the landslide body in real time.

[0094] Warning logic: When the crack sensor detects a crack width expansion of ≥10mm in a single day and the GNSS module captures a horizontal displacement of ≥5mm / day, a level-two yellow warning is triggered.

[0095] Response mechanism: The alarm station activates the timed reporting mode (uploading monitoring data to the platform once every 10 minutes), and the in-home alarm reminds residents by flashing a yellow indicator light and broadcasting a low-frequency voice message ("There is a landslide hazard in the surrounding area. Please pay attention to changes in the ground around your house") to avoid excessive panic.

[0096] A4. Maintenance and Calibration A41. Equipment Self-Test The device automatically starts a self-test program at 3:00 AM every day to check the working status of sensors (such as whether the rain gauge is counting normally and whether the water level sensor is accurate), battery power (to ensure sufficient backup battery power), and storage space (to clean up redundant data and ensure cache capacity). If an abnormality is detected (such as sensor failure or insufficient power), a maintenance work order is sent to the maintenance platform via the 4G network to notify staff to handle it in a timely manner.

[0097] A42, Sensor Calibration Rain gauge calibration: Manual calibration is performed once a month. Through simulated rainfall tests (controlling rainfall intensity and duration), the measurement error of the rain gauge is controlled within ±3% to ensure accurate rainfall data.

[0098] GNSS module calibration: Differential calibration is performed annually with the national GNSS reference station to eliminate accumulated errors from long-term operation and ensure the accuracy of displacement monitoring.

[0099] A43, Model Optimization Every quarter, historical disaster data and monitoring data accumulated on the cloud platform are transmitted back to the edge terminal of the distributed alarm station to update the parameters of the early warning model (such as adjusting the debris flow deformation rate threshold and flash flood alarm water level) so that the model can continuously adapt to the regional disaster characteristics and reduce the false alarm and missed alarm rates.

[0100] A5. Implementation Effectiveness Verification: Through simulation testing and real-world application, the system implementation effectiveness is as follows: Debris flow early warning efficiency: In simulated rainfall experiments, when the mountain deformation rate reaches the early warning threshold, the system triggers a local alarm within 3 seconds, and the response delay of the in-home sensor alarm is ≤1 second, and the early warning timeliness meets the requirements for rapid evacuation.

[0101] Communication reliability: In mountainous areas where 4G signals are blocked, LoRa self-organizing network achieves a data packet transmission success rate of ≥95% within 5 kilometers. Local cached data can be saved for ≥72 hours in the event of network outage, and the communication stability meets the requirements of field applications.

[0102] Indoor response rate: Through multimodal linkage of voice and light alarms, the early warning response rate of residents in the sleep state at night has been increased from 40% of traditional equipment to 85%, effectively solving the pain point of untimely reception of early warning information.

[0103] In summary, the present invention has the following advantages: 1. By integrating multiple sensors and combining them with multi-source data fusion algorithms, simultaneous monitoring and cross-validation of multiple disasters are achieved, breaking the traditional data silo problem and increasing disaster monitoring coverage by 40%. 2. By deploying lightweight random forest and LSTM models at the edge, disaster risk levels are calculated locally in real time, and the early warning response time is controlled within 3 seconds, avoiding the delay problem of centralized processing; 3. By adopting a hybrid LoRa and 4G communication mode combined with an adaptive scheduling mechanism, including data priority sorting and store-and-forward functions, the communication link reliability reaches 99.9%, and data can be cached for at least 72 hours when the network is down, solving the data transmission problem in complex outdoor environments; 4. By dynamically adjusting thresholds using the percentile method and optimizing both response timeliness and emergency response effectiveness, the system adapts to the disaster characteristics of different regions, reduces the false alarm and missed alarm rates of early warnings, and ensures that the effectiveness of emergency response is no less than 99%. 5. Construct a system linking distributed alarm stations and in-home alarm devices, equipped with multi-modal alarms and personnel status feedback functions, achieving an in-home alarm response rate of over 90% and a nighttime response rate of 85%, forming a complete early warning closed loop; 6. It adopts a dual-mode power supply design of solar power and mains power and is equipped with an interface expansion module, which supports the long-term operation of the equipment in the field without mains power, and is suitable for various application scenarios such as mountain villages.

[0104] The above specific 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 examples, 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 scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-hazard integrated sensing and early warning method, characterized in that, include: S1. Multi-source data acquisition and preprocessing: Collect core disaster-causing data, environmental auxiliary data, and equipment status data through the multi-hazard monitoring sensor group of the distributed integrated audio-visual alarm station; S2. Multi-source data fusion and risk level calculation: Dynamic weight allocation and weighted fusion are performed on the data streams preprocessed in step S1 to generate a comprehensive disaster risk index; S3. Dynamic threshold adjustment and early warning linkage: Based on the regional historical disaster dataset, the baseline thresholds for each level of early warning are initialized using the percentile method; The system combines the warning level with the communication link status to execute corresponding response actions. At the same time, it dynamically switches between at least two communication links to transmit warning instructions. The distributed integrated acquisition sound, light and electricity alarm station is linked with the in-home sensor alarm, and the in-home sensor alarm transmits back the status data of the people in the house. Data transmission employs a dual encryption mechanism to ensure security. S4. Response Result Optimization: The effectiveness of the early warning response is evaluated by preset evaluation indicators, and the decision parameters and linkage triggering conditions are iteratively optimized by using parameter fine-tuning formulas based on the evaluation results.

2. The multi-hazard integrated sensing and early warning method according to claim 1, characterized in that, The core disaster-causing data include: rainfall data, water level data, soil moisture content data, tilt angle data, flow velocity data, and fissure data; The sampling frequency for rainfall and water level data is 1 time / minute, the sampling frequency for soil moisture content and tilt angle data is 1 time / 5 minutes, and the sampling frequency for flow velocity and fissure data is 1 time / 10 minutes. Environmental auxiliary data includes: wind speed data, humidity data, and slope data; The environmental auxiliary data is collected once every 10 minutes. Device status data includes: sensor power data, communication module signal strength data, and CPU / memory usage data; The device status data is collected once every 30 minutes.

3. The multi-hazard integrated sensing and early warning method according to claim 1, characterized in that, Multi-source data acquisition and preprocessing specifically include: Outlier Removal: For abnormal data generated by sensors due to environmental interference or malfunction in disaster monitoring scenarios, outliers are removed using the 3σ principle. Data calibration: To address the impact of environmental factors on disaster-causing data, a linear formula is used to correct for data bias. Time-series data segmentation: In view of the characteristic that disaster risks accumulate over time, time-series disaster-causing data are segmented into time window samples of fixed length to adapt to model inference; Data standardization: To avoid the impact of differences in the magnitude of data from different dimensions on model inference, the Min-Max standardization formula is used to unify the data to a preset numerical range.

4. The multi-hazard integrated sensing and early warning method according to claim 3, characterized in that, The outlier removal method is as follows: The principle is to remove outliers, and the specific formula is as follows: For a dataset of a certain indicator Calculate the mean and standard deviation If data satisfy If the value is 0, it is considered an outlier and removed; among them, For the i-th collected data, The number of samples in the dataset. The mean of the data. The standard deviation of the data; The linear formula for the data calibration is: in, The calibrated soil moisture content (%). The original soil moisture content (%) was measured. Humidity influence coefficient; The relative humidity of the environment (%). The formula for the number of samples in time series data segmentation is: The kth sample is ;in, Let i be the time series data at time i. This represents the total number of time points in the time series data. The length of the time window; The total number of samples; The sample number; The Min-Max standardization formula is: in, The data is standardized. This is the original collected data; This is the historical minimum value of this indicator; This is the historical maximum value of this indicator.

5. The multi-hazard integrated sensing and early warning method according to claim 1, characterized in that, Step S2 also includes: Lightweight processing is performed on the pre-trained model to adapt it to the low-power embedded environment of the edge terminal. The lightweight processing includes: selecting an appropriate lightweight algorithm based on the risk factor type in multi-hazard monitoring, removing calculation branches in the algorithm whose contribution is lower than a preset threshold, converting the data precision from high-order floating-point numbers to low-order floating-point numbers or integers, and adapting it to the embedded deployment framework; the lightweight algorithm is used to perform inference calculations on the pre-processed standardized data and output a disaster risk value within a preset score range.

6. The multi-hazard integrated sensing and early warning method according to claim 5, characterized in that, The lightweight tree-based classification algorithm calculates the risk value using a weighted voting method, with the specific formula as follows: in, The risk value output by the random forest. For the number of decision trees, Let be the weight of the i-th decision tree; The risk category score for the i-th decision tree; The lightweight temporal inference algorithm calculates risk values ​​through cell state updates and mapping to fully connected layers: the cell state update process includes parameter calculations for the forget gate, input gate, and output gate to preserve key temporal features; the risk value is mapped to a preset score range through the fully connected layer, with the specific formula as follows: in, This is the weight matrix. For bias terms, Hide the current state.

7. The multi-hazard integrated sensing and early warning method according to claim 1, characterized in that, Step S3 specifically includes: The threshold initialization adopts the percentile method, and the baseline threshold for each level of early warning is determined based on the set of historical disaster risk values ​​in the region. Each level of early warning corresponds to a preset percentile parameter. The communication link status is determined by periodically detecting signals to determine whether it is smooth or interrupted; the response actions include mandatory local actions that are not affected by the link status, as well as additional actions such as synchronizing data to the remote platform when the link is smooth, encrypting and storing data locally when the link is interrupted, and retransmitting the data after the link is restored.

8. The multi-hazard integrated sensing and early warning method according to claim 7, characterized in that, The specific formula for threshold initialization is as follows: in, This is the baseline threshold for the k-th level warning. This is a set of historical disaster risk values ​​for the region. The percentiles corresponding to each level of warning are as follows: This is the percentile function.

9. The multi-hazard integrated sensing and early warning method according to claim 7, characterized in that, In step S3, the dynamic switching transmission warning command is implemented through adaptive communication scheduling. This adaptive communication scheduling includes data priority sorting, link status monitoring, scheduling decision-making, and store-and-forward. Specific steps include: Classify the data to be transmitted according to business importance and assign priority levels, and build a priority queue; Real-time detection of the quality and availability of each communication link, generating link status assessment results; Based on the data priority and link status assessment results, the communication link for data transmission is determined. When all communication links are interrupted, data is cached locally and retransmitted according to priority once the links are restored.

10. The multi-hazard integrated sensing and early warning method according to claim 9, characterized in that, The data priority sorting specifically includes: The data is categorized into four types based on business type: critical early warning data, important status data, routine monitoring data, and auxiliary and management data, corresponding to four priority levels. A priority queue is constructed using a min-heap data structure. The data packets with the highest priority level are transmitted first. The data packet encapsulation includes data payload, priority identifier, timestamp, retry count and destination address metadata.

11. The multi-hazard integrated sensing and early warning method according to claim 9, characterized in that, The link status monitoring and scheduling decision-making specifically includes: The 4G link determines its status by periodically sending TCP / UDP heartbeat packets, monitoring signal strength, network latency, and data packet success rate, and is divided into three levels: smooth, poor, and interrupted. LoRa links determine their status by reading radio frequency parameters and monitoring the success rate of confirmation messages, and are divided into three levels: smooth, poor, and interrupted. The scheduling decision logic is as follows: critical early warning data is transmitted using the 4G link first, and switches to the LoRa link when the 4G link is interrupted; important status data uses 4G when the 4G link is normal, and switches to LoRa when the 4G link is poor or interrupted and the LoRa link is smooth; routine monitoring data and auxiliary management data are transmitted via the selected transmission link or delayed transmission according to the link status and resource usage.

12. The multi-hazard integrated sensing and early warning method according to claim 9, characterized in that, The store-and-forward mechanism specifically includes: When all communication links are interrupted, data packets are stored in a wait-to-retransmit queue in non-volatile memory. Once any communication link is restored, data in the retransmission queue will be retransmitted first, according to data priority.

13. The multi-hazard integrated sensing and early warning method according to claim 1, characterized in that, The preset evaluation indicators include: response timeliness. Effectiveness of emergency response By fine-tuning the decision parameters, the frequency of audible and visual alarms and the trigger threshold of emergency equipment are adjusted to achieve iterative optimization of decision rules; The response timeliness The specific formula is: in, This refers to the number of times a local action is completed within a preset time after an alert is triggered. Total number of warnings; The effectiveness of the emergency response The specific formula is: in, This refers to the number of times disaster losses are reduced after coordinated actions. This represents the total number of times the emergency action was triggered. The specific formula for fine-tuning the decision parameters is as follows: in, The adjusted decision parameters, The decision parameters before adjustment. To adjust the coefficient, Evaluation indicators.

14. A multi-hazard integrated sensing and early warning system, characterized in that, include: A distributed integrated acquisition sound, light, and electricity alarm station, which integrates a multi-hazard monitoring sensor group and an edge data preprocessing module; The multi-hazard monitoring sensor group is used to collect core disaster-causing data, environmental auxiliary data, and equipment status data; the edge data preprocessing module is used to perform preprocessing on the collected data. The multi-source data fusion module is connected to the distributed integrated acquisition sound, light, and electricity alarm station. It is used to perform dynamic weight allocation and weighted fusion on the preprocessed data stream output by the edge data preprocessing module to generate a comprehensive disaster risk index. An edge computing early warning module is connected to the multi-source data fusion module and is used to calculate the disaster risk level based on the comprehensive disaster risk index. The dynamic risk model module, connected to the edge computing early warning module, is used to initialize the baseline thresholds for each level of early warning based on the regional historical disaster dataset using the percentile method. A hybrid communication redundancy module is connected to the distributed integrated acquisition sound, light, and electricity alarm station, and supports at least two communication links. It is used to dynamically switch the transmission channel according to the status of the communication link in order to transmit early warning instructions. An in-home sensor alarm is connected to the hybrid communication redundancy module to receive the warning command and execute the alarm operation, and to transmit the status data of the people inside the house. The secure transmission module, embedded in the hybrid communication redundancy module, employs a dual encryption mechanism to ensure the security of the transmission of early warning commands and monitoring data. The decision optimization module is connected to the edge computing early warning module and the in-home sensor alarm, respectively. It is used to evaluate the early warning response effect through preset evaluation indicators, and to iteratively optimize the decision parameters and linkage triggering conditions based on the evaluation results using parameter fine-tuning formulas.

Images