A visual hierarchical early warning method and system

Abstract

The application discloses a visual hierarchical early warning method and system, wherein the method comprises the following steps: firstly, collecting a multi-source perception original data set, transmitting the original data set to an edge side for preprocessing, extracting an evidence feature vector of a suspected risk point, splicing a light-weight original fusion feature vector, and uploading the light-weight original fusion feature vector to a cloud; secondly, adopting a multi-competitive hypothesis intervention to reconstruct a cause-effect reasoning to complete risk early warning grade division; thirdly, constructing a complete early warning evidence chain, calculating a hash digest, and storing the hash digest in a consortium blockchain; fourthly, executing differential point rendering according to a point position and a risk grade, and generating a regional risk heat map by adopting a kernel density estimation algorithm; fifthly, implementing hierarchical differential linkage treatment according to the early warning grade; sixthly, taking a mean square error loss function as a target to perform grid search optimization, updating parameters, and issuing the parameters to an edge gateway; and seventhly, outputting standardized fire-fighting risk statistical data set as compliance digital assets after the evidence chain is desensitized. The application can reduce the false alarm rate in a complex scene, the research and judgment process is traceable, and cloud-edge collaborative closed-loop early warning is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire risk monitoring and early warning technology, and in particular to a visualized, graded early warning method and system. Background Technology

[0002] With the continuous implementation of smart fire protection construction in China and the large-scale deployment of 5G IoT sensing terminals, massive amounts of data from smoke detectors, water pressure sensors, electrical monitoring, and AI video recognition provide a data foundation for early warning of fire risks. Existing fire warning technologies generally rely on single sensors and fixed thresholds to trigger alarms, only mining statistical correlations between data. They cannot distinguish between three different types of triggers: real fire hazards, environmental interference, and equipment operational disturbances. Under conditions such as dust in mountainous areas and complex electromagnetic interference in old residential areas, they are prone to generating a large number of false alarms.

[0003] Traditional early warning systems are "black box outputs," only showing the risk outcome without preserving a complete chain of evidence for analysis. After an alarm occurs, it's impossible to trace the root cause of false alarms, making it difficult to guide the iteration of early warning rules. Furthermore, most solutions rely on uploading all raw sensor data to the cloud for centralized processing, which can lead to uplink network bandwidth congestion and excessive cloud computing load when dealing with a large number of terminals. Fixed early warning grading rules reuse the same set of thresholds across different scenarios, resulting in a one-size-fits-all approach to alarm pushes and device linkage actions, leading to delayed emergency responses for low-risk scenarios and lagging instructions for high-risk scenarios.

[0004] Existing models such as XGBoost and Bayesian network risk assessment are statistical learning models that train on the relationships between samples but lack causal reasoning capabilities. They also exhibit poor generalization ability in complex scenarios with insufficient training sample coverage, such as mountainous tourism areas and old residential communities. Furthermore, they lack a complete process encompassing cloud-edge collaborative lightweight feature extraction, intervention reconstruction with multiple hypothesis reasoning, GIS evidence pop-up visualization, blockchain evidence storage, and business-driven model closed-loop iteration. This makes them unable to meet the high-concurrency access requirements of tens of thousands of fire protection terminals across the province. Therefore, there is an urgent need to invent a visualized, tiered early warning method and system to address these technical shortcomings. Summary of the Invention

[0005] The purpose of this invention is to provide a visualized hierarchical early warning method and system that can overcome the shortcomings of existing fire early warning systems, such as high false alarm rates, black box judgments without traceability, high cloud bandwidth pressure, homogenized handling strategies, and poor generalization in complex scenarios. It adopts multiple competing hypotheses to intervene and reconstruct causal reasoning, and realizes lightweight feature extraction, risk scoring and hierarchical classification, blockchain hash digest evidence storage, GIS visualization, hierarchical linkage handling and parameter closed-loop iteration, which is suitable for various complex fire monitoring scenarios.

[0006] To achieve the above objectives, the present invention provides a visualized hierarchical early warning method comprising the following steps: S1. Collect the raw dataset of multi-source sensing and transmit it to the edge measurement; S2. After preprocessing the raw data transmitted to the edge, multi-source evidence feature vectors are extracted from suspected risk points, spliced ​​and fused to obtain lightweight raw fused feature vectors, and then uploaded to the cloud. S3. Reconstruct causal reasoning by intervening with multiple competing hypotheses and classify risk warning levels; S4. Construct a complete chain of evidence for a single early warning, calculate its SHA-256 hash digest, and store it on the consortium blockchain. S5. Perform differentiated point marking rendering based on point location and risk warning level, and generate regional risk heat map using kernel density estimation algorithm; S6. Implement differentiated and coordinated responses based on risk warning levels, including differentiated message push, work order issuance, and on-site hardware coordination. S7. Read the manually reviewed and labeled sample set, perform grid search optimization with the mean square error loss function as the optimization objective, update the parameters and then send them to each edge gateway for synchronization. S8. De-identify the evidence chain data, filter sensitive fields, and output a standardized fire risk statistics dataset to provide compliant data services to external parties as a digital asset for fire analysis and assessment.

[0007] Preferably, in S1, the on-site physical sensing information of the IoT fire protection terminal is collected, and the multi-source sensing raw dataset is output. And transmit the original multi-source sensing dataset to the edge: ; ; In the formula, For the original dataset of multi-source perception, Indicates the first Terminal sampling records; This indicates the total number of terminal sampling records. Sampling time; The location's geographical coordinates; For sensor measurement array, This is a fault code from the device's self-test. To identify confidence levels.

[0008] Preferably, in S1, the IoT fire protection terminal includes a smart smoke detector, a combustible gas detector, a fire water pressure monitoring terminal, an electrical fire monitoring terminal, and an AI video flame recognition camera.

[0009] Preferably, S2 is as follows: S21, Firstly adopt The criterion for identifying and removing accidental noise sampling points is expressed as follows: ; In the formula, The mean value is obtained by sampling through a sliding window. The standard deviation of the sliding window sampling; S22. Obtain the standardized dataset using max-min normalization. The normalization formula is as follows: ; In the formula, These are the normalized feature components; These are the original measured values ​​of the characteristic components to be normalized. This represents the maximum value within the statistical window for this feature component. This is the minimum value within the statistical window of this feature component; S23, Read the scene adaptation threshold library Perform local primary screening, threshold library The threshold values ​​for each indicator are divided into four scenarios: high-rise buildings, old residential areas, industrial parks, and mountainous cultural tourism scenic areas. The data is then compared point-by-point with the standardized dataset. If all features fall within the corresponding scene threshold range, the dataset is deemed risk-free, and the dataset is standardized. Stored locally at the edge, without being uploaded to the cloud; If at least one feature exceeds the scene threshold, it is marked as a suspected risk point, triggering lightweight feature extraction. The extracted features include physical evidence feature vectors. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations ; S24. Transform the physical evidence feature vector Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations After concatenation and fusion, a lightweight original fused feature vector is obtained: ; Lightweight original fused feature vector Upload to the cloud to achieve bandwidth offloading.

[0010] Preferably, S3 includes the following steps: S31. For the original fused feature vector Decoupling yields physical evidence feature vectors. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations ; S32. Employing multiple competing hypotheses for intervention and reconstruction, simulating feature performance under different hypotheses: ; In the formula, This indicates the characteristics of fire hazard hypothesis intervention and reconstruction; This indicates the characteristics of environmental disturbance hypothesis intervention reconstruction; This indicates that the characteristics of the intervention and reconstruction are based on the assumption of operating condition fluctuations; The number of dimensions of physical evidence features; The number of dimensions for evidence of environmental interference; The number of dimensions for evidence of operating condition fluctuations; It is a zero-dimensional vector of physical features; It is a zero vector of environmental feature dimensions; The zero vector represents the dimension of the working condition features; S33. Use cosine similarity to measure the degree of matching between the original features and the reconstructed features, and calculate the explanatory power of each hypothesis: ; in, Explanation degree of fire hazard assumptions; Explanation degree for the environmental disturbance hypothesis; Explanatory power for the assumption of operating condition fluctuations; Represents the L2 norm; S34. Determine the authenticity of potential hazards and set a threshold for hypothetical determination. ,when and If the alarm is triggered, it is considered a genuine fire hazard; otherwise, it is considered a false alarm. Calculate the overall risk score : ; in, For sensor confidence weights, ; This is a scene correction factor. ; ; Risk warning levels will be output according to the following rules. : .

[0011] Preferably, in S4, a complete chain of evidence for a single early warning is constructed, recording all key information from this assessment to form the original complete chain of evidence. : ; in, System timestamp; The coordinates of the location are latitude and longitude. To unify the timestamps after alignment; Mark as a false alarm; The hash digest is calculated using the following formula: ; Store the hash digest on the consortium blockchain, and store the original complete chain of evidence. Stored in the local evidence chain storage module database.

[0012] Preferably, in S5, based on the latitude and longitude of the location Risk warning level Perform differentiated point rendering: Marked in blue; Marked in yellow; Marked in orange; High-risk layers are highlighted with a red flashing indicator and placed at the top. A kernel density estimation algorithm is used to generate a regional risk heatmap, which visually displays the spatial distribution of risks. ; In the formula, Indicates position Risk density estimate; This indicates the total number of risk points involved in the heatmap calculation; Indicates bandwidth; Represents the kernel function; Indicates the first Geographical location of each risk point.

[0013] Preferably, in S6, the hierarchical differentiated linkage handling rule is as follows: The platform only sends risk warnings to local grid workers, without hardware integration, and incorporates them into daily inspection tasks. We pay close attention to risks and use in-station messages and SMS to push information to grid members and enterprise safety administrators to automatically generate hazard verification work orders, requiring on-site verification within a specified time limit; In case of risk, the system will send out warnings through multiple channels, including in-station messages, SMS, and APP, and coordinate with on-site sound and light warning devices to issue emergency response work orders and go to the site for verification as soon as possible. In the event of an emergency, the system will notify the fire and rescue brigade and local duty stations, coordinate emergency broadcasts and audible and visual alarms, and provide rescue route planning and suggestions for dispatching surrounding rescue resources.

[0014] Preferably, in S7, the set of manually reviewed and annotated samples is read. ,in, For the first Manually labeled training samples; The mean squared error loss function is used to measure the deviation between the model's output risk score and the human-generated true label. The expression for the loss function is as follows: ; in, This represents the total number of training samples; For the first Each manual sample outputs a comprehensive risk score; The training sample index numbers are manually labeled, and their values ​​are set to... ; by The optimization objective is to perform a grid search to find the optimal sensor confidence weights. Scene correction coefficient Scene adaptation threshold library .

[0015] This invention also provides a visualized hierarchical early warning system, including a perception and acquisition module, an edge-side preprocessing module, a cloud-based causal analysis and hierarchical module, an evidence chain storage module, a blockchain evidence storage subunit, a GIS visualization and rendering module, a hierarchical disposal linkage module, an iterative optimization module, and a digital asset interface unit. The sensing and acquisition module is used to collect raw datasets from multiple sensing sources. The output is transmitted to the edge-side preprocessing module; The edge-side preprocessing module is deployed on the edge gateway and includes a data cleaning unit, a time-aligned preprocessing unit, a normalization unit, a local initial screening unit, and a lightweight feature extraction unit. The cloud-based causal analysis and grading module includes a physical causal guidance unit, a multi-competitive hypothesis reasoning unit, a risk scoring unit, and a grading unit. The evidence chain storage module stores a complete single early warning evidence chain. The blockchain evidence storage subunit calculates the hash digest of the evidence chain. The hash digest is uploaded to the consortium blockchain; The GIS visualization rendering module includes a point rendering unit, a multi-layer linkage unit, an evidence pop-up unit, and a risk heat map generation sub-unit. The tiered response and coordination module has a built-in tiered response strategy library, which is based on the risk warning level. Perform differentiated message push, work order issuance, and hardware linkage to match business processing actions; The iterative optimization module reads the manually reviewed sample set and uses the mean squared error loss function as the optimization objective to iteratively update the sensor confidence weights through grid search. Scene correction coefficient Scene adaptation threshold library Update the parameters and send them to the edge-side preprocessing module for synchronization; The digital asset interface unit anonymizes evidence chain data, outputs standardized fire risk statistics sets, and provides compliant digital asset data services to external parties.

[0016] Therefore, the present invention employs the above-mentioned visualized hierarchical early warning method and system, and the beneficial technical effects are as follows: (1) This invention eliminates the timing misalignment caused by inconsistent sampling periods of different terminals by adding multi-source sensing time alignment preprocessing at the edge end; completes anomaly filtering and local coarse screening, the scenario adaptation threshold library refers to the national standard and distinguishes the initial parameters of multiple scenarios, and the threshold supports automatic iterative updates of business samples; only lightweight features are uploaded, which significantly reduces the uplink network bandwidth pressure; the cloud adopts multiple competing assumptions to reconstruct causal reasoning, distinguishes real fire hazards, environmental interference, and working condition disturbances from the causal dimension, gets rid of the constraints of simple statistical correlation, effectively reduces the false alarm rate of early warning under complex working conditions in mountainous areas and old residential areas, and improves the reliability of early warning results.

[0017] (2) This invention retains the complete evidence chain of the entire early warning process. The evidence chain records aligned timestamps and uses hash digests for on-chain storage, taking into account both storage costs and anti-tampering audit traceability capabilities. GIS realizes risk classification rendering and visualization of judgment evidence pop-up windows, solving the problem of traditional early warning black boxes. The four-level risk level matches differentiated handling strategies, realizing alarm push, hardware linkage and accurate matching of risk levels, shortening the time for emergency judgment and handling.

[0018] (3) This invention utilizes on-site manual verification samples and uses loss function optimization to complete the closed-loop iteration of weights and scenario thresholds. The system can continuously adapt to local scenarios according to actual business. It sets up a digital asset desensitization output process, analyzes and transforms data into compliant fire protection digital assets for external services, and adapts to the entire business model of provincial smart fire protection digital asset operation.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of an embodiment of a visualized hierarchical early warning method of the present invention; Figure 2 This is a cloud-based multi-competitive hypothesis intervention reasoning flowchart of an embodiment of a visualized hierarchical early warning method of the present invention; Figure 3 This is an overall structural block diagram of an embodiment of a visual hierarchical early warning system according to the present invention; Figure 4 This is a comparison chart of test indicators for different algorithm models of a visualized hierarchical early warning method and system embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figure 1 As shown, a visualized hierarchical early warning method includes the following steps: S1. Collects on-site physical perception information from IoT fire protection terminals such as smart smoke detectors, combustible gas detectors, fire water pressure monitoring terminals, electrical fire monitoring terminals, and AI video flame recognition cameras. The sampling periods for different terminals vary: electrical monitoring terminals have shorter sampling periods, while smoke sensors and AI video recognition outputs have longer sampling periods. Outputs a multi-source perception raw dataset. And transmit the original multi-source sensing dataset to the edge: ; ; In the formula, For the original dataset of multi-source perception, Indicates the first Terminal sampling records; This indicates the total number of terminal sampling records. Sampling time; The location's geographical coordinates; For sensor measurement array, This is a fault code from the device's self-test. To identify confidence levels, This is the index number of the original terminal sampling record.

[0024] In this step, the multi-source sensing raw dataset contains all basic information such as time, location, measurement values, device status, and AI recognition results, providing a complete data source for subsequent end-to-end analysis. Furthermore, the sensing acquisition module will... The data is transmitted to the edge-side preprocessing module, where no complex calculations are performed at the perception layer, thus preserving the original observation information.

[0025] S2. After preprocessing the raw data transmitted to the edge, multi-source evidence feature vectors are extracted from suspected risk points, and the resulting lightweight raw fused feature vectors are obtained by splicing and fusing them and then uploaded to the cloud.

[0026] This invention deploys an edge gateway locally at the monitoring point to perform preprocessing. The aim is to filter out a large number of normal, risk-free samples, resolving timing misalignment issues caused by inconsistent sampling periods from multiple sensors. Only lightweight features of suspected risks are uploaded to the cloud, achieving computational and bandwidth offloading and avoiding network congestion caused by uploading massive amounts of raw data to the cloud. It is important to note that the local initial screening at the edge only serves as a coarse screening to mark suspected risks; it does not output the final alarm result. All true / false alarm determinations are handled by a cloud-based causal reasoning model, specifically: S21. Because sensors can experience transient noise fluctuations, this invention employs... The criteria identify and remove randomly changing noise sampling points to prevent noisy data from entering subsequent analysis processes and causing misjudgments. The criterion expression is: ; In the formula, The mean value is obtained by sampling through a sliding window. The standard deviation of the sliding window sampling is given.

[0027] Set the duration of the time-aligned sliding window. The optimal value for the project is 10 seconds; based on duration... The continuous timeline is divided into several consecutive time slices, each of which is assigned a representative moment, and then an aligned unified timestamp is generated. , You can select the midpoint or the start time of the window for the corresponding time slice.

[0028] Different fire protection IoT terminals have different sampling periods, and the output times of electrical equipment, smoke detectors, and AI video recognition are not completely synchronized. Forcibly stitching together observations collected at different times into the same set of feature vectors would result in temporal misalignment, directly undermining the validity of subsequent causal inference inputs. This step performs temporal alignment at the edge, mapping the observations of each sensor to the current time slice and aligning them to the representative time. When the sensor observes data within the current segment window, the distance within this window is selected. The most recent valid sampling point; if the current slice window has no valid observations from the sensor, then backtrack and use the most recent valid sampling point saved from the previous adjacent time slice to fill the gap; ensuring the same original fused feature vector. All feature components belong to the same time slice. After time alignment, an intermediate dataset with time dimension alignment is obtained.

[0029] S22. Employ max-min normalization to eliminate dimensional differences between different sensors and map each dimension of the features to... Dimensionless intervals yield standardized datasets. The normalization formula is as follows: ; In the formula, These are the normalized feature components; These are the original measured values ​​of the characteristic components to be normalized. This represents the maximum value within the statistical window for this feature component. This is the minimum value within the statistical window for this feature component.

[0030] S23. First, read the scene adaptation threshold library. Perform local primary screening, threshold library The internal database distinguishes four scenarios: high-rise buildings, old residential areas, industrial parks, and mountain cultural tourism scenic spots. The threshold values ​​for each indicator are shown in Table 1, which represents the normalized initial coarse screening thresholds for the four scenarios stored in the database. Table 1 Scene Adaptation Threshold Library Initial preset parameter table

[0031] As shown in Table 1, high-rise buildings experience less indoor environmental interference, so the threshold settings are more sensitive; older residential areas have aging wiring and stronger electromagnetic interference, so the thresholds are appropriately relaxed; industrial parks have many flammable and explosive hazards, so the threshold settings are relatively strict; mountainous cultural and tourism scenic areas have prominent outdoor dust and light disturbances, so the thresholds related to smoke and dust are significantly raised. During system operation, the iterative optimization module will automatically perform grid search updates on the entire threshold library based on samples manually reviewed by business personnel to adapt to the actual on-site environment. This threshold is only used for local coarse screening at the edge, only marking suspected risk points, and does not directly output alarms. The final determination of the authenticity of the hidden danger is completed by the cloud-based multi-competitive hypothesis intervention and reconstruction of the causal reasoning model.

[0032] Then, the standardized dataset is compared point by point: If all features fall within the corresponding scene threshold range, the dataset is deemed risk-free, and the dataset is standardized. It is stored locally at the edge and not uploaded to the cloud.

[0033] If at least one feature exceeds the scene threshold, it is marked as a suspected risk point, triggering lightweight feature extraction. The extracted features include physical evidence feature vectors. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations .

[0034] Since the monitoring points are in a normal state most of the time, this invention directly filters normal samples through local initial screening, and only suspected abnormalities are entered into subsequent cloud-based in-depth analysis, which greatly reduces the amount of data uploaded.

[0035] S24. Transform the physical evidence feature vector (Smoke, combustible gas, electrical overload, AI flame confidence, etc.) Environmental interference evidence feature vector (Temperature, humidity, dust, light, etc.) and feature vectors of evidence for fluctuations in operating conditions After concatenating and fusing (equipment fault codes, power supply offset, and communication packet loss rate), a lightweight original fused feature vector is obtained: .

[0036] This step does not transfer the large original video file, but only the lightweight original fused feature vector that has already been decoupled. Transmitting data to the cloud reduces bandwidth consumption while fully preserving the three dimensions of evidence: physical, environmental, and operational conditions, ensuring complete input for causal reasoning in the cloud.

[0037] S3. The cloud receives lightweight features of suspected risks uploaded from the edge. Multiple competing hypotheses are used to reconstruct causal reasoning. Through intervention, feature components of a certain type of factor are stripped away. The explanatory power of different hypotheses for observed features is compared. After distinguishing between hidden dangers, environmental interference, and equipment failures, a risk warning level is determined. Figure 2 As shown, it includes the following steps: S31. For the original fused feature vector Decoupling by dimension yields the physical evidence feature vector. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations This is to prepare for intervention and restructuring.

[0038] S32. Employing multiple competing hypotheses for intervention and reconstruction, simulating feature performance under different hypotheses: ; In the formula, This indicates the characteristics of fire hazard hypothesis intervention and reconstruction, which retain only physical evidence of fire hazards and simulates only fire hazards without environmental or operational disturbances. This indicates that the environmental disturbance hypothesis intervenes in the reconstruction of features, retaining only the environmental disturbance features; This indicates that the operating condition fluctuation assumption intervenes in the reconstruction characteristics, retaining only the operating condition fluctuation characteristics. The number of dimensions of physical evidence features; The number of dimensions for evidence of environmental interference; The number of dimensions for evidence of operating condition fluctuations; It is a zero-dimensional vector of physical features; It is a zero vector of environmental feature dimensions; It is the zero vector of the working condition feature dimension.

[0039] S33. Cosine similarity is used to measure the degree of matching between the original features and the reconstructed features. The higher the similarity, the better the hypothesis can explain the current observation phenomenon. This value corresponds to the explanatory power of the hypothesis. The calculation formula is as follows: ; in, Explanation degree of fire hazard assumptions; Explanation degree for the environmental disturbance hypothesis; Explanatory power for the assumption of operating condition fluctuations; This represents the L2 norm.

[0040] S34. Determine the authenticity of potential hazards and set a threshold for hypothetical determination. This represents the minimum explanatory power required for the fire hazard hypothesis compared to the other two types of hypotheses. and If the alarm is triggered, it is considered a genuine fire hazard; otherwise, it is considered a false alarm.

[0041] Taking into account the interpretability of the assumptions, the reliability of the sensors, and the differences in the scene environment, a comprehensive risk score is calculated. : ; in, For sensor confidence weights, The more historical faults a device has, the lower its reliability and the smaller its weight. The scene correction coefficient is used to compensate for the inherent interference levels in different scenes. ; The higher the value, the higher the risk.

[0042] Risk warning levels will be output according to the following rules. The continuous scores are mapped to four levels of business feasibility, directly guiding subsequent processing procedures. .

[0043] S4. To address the black-box problem in early warning, all intermediate calculation results of this early warning are completely preserved to form an evidence chain. Considering the large amount of data in the original evidence chain, the complete large-volume data is not uploaded to the chain; only the hash digest is uploaded to the chain for evidence storage, while the original evidence chain is stored in the database. The hash digest has irreversible and tamper-proof characteristics. During later audits, the evidence chain hash is recalculated and compared with the on-chain digest to verify whether the evidence chain has been tampered with, balancing storage costs and tamper-proof audit capabilities.

[0044] Specifically: First, construct a complete chain of evidence for a single early warning, recording all key information from this assessment to form the original complete chain of evidence. : ; in, System timestamp; The coordinates of the location are latitude and longitude. To unify the timestamps after alignment; Mark as a false alarm; right The core metadata fields are used to calculate their hash digests, using the following formula: ; Store the hash digest on the consortium blockchain, and store the original complete chain of evidence. Stored in the local evidence chain storage module database.

[0045] S5. Based on the location of the monitoring points and the risk warning level, perform differentiated point marking and rendering, and generate a regional risk heat map using a kernel density estimation algorithm. Specifically: Based on the latitude and longitude of the point Risk warning level Perform differentiated point rendering: Marked in blue.

[0046] It is marked in yellow.

[0047] Marked in orange.

[0048] High-risk layers are marked with a red flashing indicator and placed at the top.

[0049] Additionally, it allows overlaying fire protection facility layers, AI video surveillance layers, and rescue resource layers, supporting one-click retrieval of on-site videos and information on nearby mini fire stations. Evidence pop-up unit: Clicking the risk marker on the map loads the evidence chain data in a pop-up window, displaying physical, environmental, and operational characteristics line by line, along with three sets of hypothesis explanations, comprehensive risk scores, risk levels, and aligned timestamps. It supports retrieving on-chain hash digests to verify the integrity of evidence.

[0050] A kernel density estimation algorithm is used to generate a regional risk heatmap, which visually displays the spatial distribution of risks. ; In the formula, Indicates position Risk density estimate; This indicates the total number of risk points involved in the heatmap calculation; Indicates bandwidth; Represents the kernel function; Indicates the first Geographical location of each risk point.

[0051] S6. Implement differentiated and tiered response measures based on risk warning levels, including differentiated message push, work order issuance, and on-site hardware linkage, to avoid excessive alarms that disturb users in low-risk situations, while ensuring that high-risk alarms are notified in a timely manner, and to connect algorithm output with actual business actions.

[0052] The tiered and differentiated joint response rules are as follows: Risk warnings are issued in real time, but only the platform sends messages to local grid workers; there is no hardware linkage, and the information is incorporated into daily inspection tasks.

[0053] We pay close attention to risks and use in-station messages and SMS to push information to grid workers and enterprise safety administrators to automatically generate hazard verification work orders, requiring on-site verification within a specified time.

[0054] In case of risk, the system will send out warnings through multiple channels, including in-station messages, SMS, and APP, and coordinate with on-site sound and light warning devices to issue emergency response work orders and promptly go to the site for verification.

[0055] In the event of an emergency, the system will notify the fire and rescue brigade and local duty stations, coordinate emergency broadcasts and audible and visual alarms, and provide rescue route planning and suggestions for dispatching surrounding rescue resources.

[0056] S7. Real-world business operations generate a large number of manually reviewed samples (business personnel verify on-site whether alarms are real hidden dangers or false alarms). This invention uses real business samples to reverse optimize edge and cloud parameters, allowing the system to continuously adapt to local scenarios as business operations progress.

[0057] First, read the manually reviewed and annotated sample set. ,in, For the first Each training sample is manually labeled; a real potential hazard is assigned a value of 1, and a false alarm is assigned a value of 0.

[0058] The mean squared error loss function is used to measure the deviation between the model's output risk score and the human-generated true label. The expression for the loss function is as follows: ; in, This represents the total number of training samples; For the first Each manual sample outputs a comprehensive risk score; The training sample index numbers are manually labeled, and their values ​​are set to... .

[0059] This invention is based on The optimization objective is to perform a grid search to find the optimal sensor confidence weights. Scene correction coefficient Scene adaptation threshold library After the update, the entire set of parameters will be distributed and synchronized to the local screening units of each edge gateway to complete the iteration.

[0060] S8. De-identify the evidence chain data, filter sensitive fields, and output a standardized fire risk statistics dataset to provide compliant data services to external parties as a digital asset for fire analysis and assessment.

[0061] like Figure 3 As shown, the present invention also provides a visualized hierarchical early warning system for implementing the above method, including a perception and acquisition module, an edge-side preprocessing module, a cloud-based causal judgment and hierarchical module, an evidence chain storage module, a blockchain evidence storage subunit, a GIS visualization and rendering module, a hierarchical disposal linkage module, an iterative optimization module, and a digital asset interface unit.

[0062] The system connects to 5G IoT fire protection terminals, which include intelligent smoke detectors, combustible gas detectors, fire water pressure monitoring terminals, electrical fire monitoring terminals, and AI video flame recognition cameras. The sensing and acquisition module is used to collect multi-source raw data sets from the 5G IoT fire protection terminals. The output is transmitted to the edge-side preprocessing module.

[0063] The edge-side preprocessing module is deployed on the edge gateway and includes a data cleaning unit, a time-aligned preprocessing unit, a normalization unit, a local initial screening unit, and a lightweight feature extraction unit. The data cleaning unit performs 3σ anomaly removal. The time-aligned preprocessing unit configures the duration of the time-aligned sliding window. Complete multi-source sensor timing alignment; the normalization unit performs max-min normalization to output a standardized dataset. The local initial screening unit has a built-in scene adaptation threshold library. Local primary screening is completed to filter out risk-free samples; the lightweight feature extraction unit extracts feature vectors of physical evidence. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations splicing Only lightweight features are uploaded afterward to achieve network bandwidth offloading.

[0064] The cloud-based causal analysis and grading module includes a physical causal guidance unit, a multi-competitive hypothesis reasoning unit, a risk scoring unit, and a grading unit. The multi-competitive hypothesis reasoning unit performs intervention and reconstruction to generate reconstructed features, and uses cosine similarity calculation to obtain the explanatory power of the fire hazard hypothesis. Explanation of environmental disturbance hypothesis Explanatory power of operating condition fluctuation assumptions .

[0065] The risk scoring unit calculates the overall risk score. The risk warning level is output by the risk classification unit. .

[0066] The evidence chain storage module stores a complete single early warning evidence chain. The blockchain evidence storage subunit calculates the hash digest of the evidence chain. The hash digest is uploaded to the consortium blockchain.

[0067] The GIS visualization rendering module includes a point rendering unit, a multi-layer linkage unit, an evidence pop-up unit, and a risk heatmap generation sub-unit. The risk heatmap generation sub-unit uses a kernel density estimation algorithm to generate regional risk heatmaps. The evidence pop-up unit can display aligned and unified timestamps. It takes into account both point-to-point analysis and macro-regional risk assessment.

[0068] The GIS visualization rendering module's multi-layer linkage unit can overlay fire protection facility layers, AI video surveillance layers, and rescue resource layers, and supports one-click retrieval of location monitoring videos to view the distribution information of surrounding mini fire stations and fire water sources.

[0069] The evidence pop-up unit reads data from the database. The pop-up page fully displays the characteristics of physical evidence, environmental interference evidence, operating condition fluctuation evidence, the explanatory power of the three sets of hypotheses, the comprehensive risk score, the risk level, and the aligned unified timestamp. It supports retrieving the hash digest of blockchain evidence storage to complete the verification of evidence consistency and tamper-proof.

[0070] The tiered response and coordination module has a built-in tiered response strategy library, which is based on risk level. It executes differentiated message push, work order issuance, and hardware linkage to match business processing actions.

[0071] The iterative optimization module reads the set of manually reviewed samples. , For the first The training samples are manually labeled, with labels categorized into real hazard labels and false alarm labels. The mean squared error loss function is used as the optimization objective, and the sensor confidence weights are iteratively updated using a grid search. Scene correction coefficient Scene adaptation threshold library The updated parameters are sent and synchronized to the edge-side preprocessing module to achieve closed-loop iteration of business processes.

[0072] The digital asset interface unit desensitizes evidence chain data, filters out sensitive fields such as precise latitude and longitude coordinates of locations and unique hardware identifiers of equipment; and outputs only a standardized fire risk statistics set that retains statistical dimensions, providing compliant digital asset data services to external parties.

[0073] Example 1: This embodiment was deployed in a mountainous cultural tourism scenic area. The outdoor environment in the mountains is characterized by dust, high temperatures, and varying light levels. The different sampling periods of various sensors make this suitable for verifying the time-aligned preprocessing and environmental disturbance suppression capabilities of this invention. The deployed equipment includes an outdoor 5G smart smoke detector, an infrared forest fire monitoring terminal, and temperature, humidity, and dust sensors. As shown in Table 2, the test dataset contains 1260 samples, including real fire hazards, false alarms due to environmental interference, and false alarms due to equipment operating condition fluctuations, simulating various alarm sources in real-world business scenarios.

[0074] Table 2. Sample Distribution of Test Dataset

[0075] System parameter configuration for this embodiment: Duration of time-aligned sliding window Assuming a threshold for judgment Mountainous cultural tourism scene correction coefficient Sensor confidence weights Based on the equipment's historical fault records Dynamic assignment; risk score ranges are strictly implemented according to the grade division units.

[0076] The edge-side preprocessing module receives the raw dataset output by the terminal. ,go through Outlier filtering; multi-source sensor time alignment preprocessing is performed, with 10-second windows used for partitioning, and missing observations within the window are filled in as close as possible to eliminate sampling time misalignment between different devices; then max-min normalization is performed to obtain a standardized dataset. The local initial judgment unit reads the scene adaptation threshold library specific to mountainous cultural tourism scenic spots. .

[0077] Scenario 1: Dust disturbance, dust index temporarily exceeds limit.

[0078] After normalization, feature vectors are extracted: physical evidence characteristics The smoke content was 0.30, which did not exceed the smoke threshold of 0.42 for mountain cultural tourism scenic areas.

[0079] Environmental disturbance characteristics Among them, the dust component was 0.57, which is slightly higher than the dust threshold of 0.55 for mountain cultural tourism scenic areas, triggering local preliminary screening at the edge and marking it as a suspected risk point.

[0080] Operating condition fluctuation characteristics The equipment is operating normally without any obvious abnormalities.

[0081] Concatenation yields a lightweight vector Uploaded to the cloud, the cloud-based multi-competitive hypothesis reasoning unit performs intervention reconstruction, generating three sets of reconstructed features, calculating cosine similarity, and obtaining exemplary explanatory power: .

[0082] Calculate the difference The difference is less than the assumed judgment threshold. If the conditions for determining a genuine fire hazard are not met, it is determined to be a false alarm and the alarm is set. .

[0083] The confidence weight of the sensor at this location Mountain scene correction factor ; Result in a comprehensive risk score: .

[0084] Output Level 1 risk warning.

[0085] Construct a complete chain of evidence The chain of evidence records aligned timestamps. Calculate hash digest The hash digest is uploaded to the blockchain, and the original evidence chain is stored in the database. A GIS map renders blue markers; a pop-up window shows a high dust index of 0.57 in the environmental disturbance features, along with the aligned timestamp. This clearly indicates that the alarm originated from environmental disturbances; the tiered response and linkage module only pushes platform messages to grid members and will not trigger SMS or audible / visual alarms, thus avoiding false alarms and wasting emergency resources.

[0086] Scenario 2: The potential danger of smoke rising from a real mountain forest.

[0087] Sensors collect real smoke signals, and after edge time alignment and normalization, suspected risks are marked. Uploaded to the cloud; cloud computing yields an exemplary interpretability: . , This meets the criteria for determining genuine hidden dangers.

[0088] The confidence weight of the sensor at this location Scene correction coefficient 0; The overall risk score is: ; Output Level IV emergency.

[0089] The GIS map displays points marked in red flashing, and a pop-up window fully showcases the physical evidence characteristics, the explanatory power of three sets of hypotheses, and the aligned timestamps. The tiered response and linkage module pushes emergency alarms to the fire and rescue brigade through multiple channels, links with the scenic area's emergency broadcast system, and outputs rescue route planning and suggestions for dispatching surrounding rescue resources.

[0090] After the business review was completed, all 1260 samples with manual review labels were sent to the iterative optimization module; using the loss function Minimize the target by performing a grid search optimization and updating the threshold library for mountain tourism scenarios. Sensor confidence weights The updated parameters are then sent to each edge gateway, completing a closed-loop iteration. The system's ability to recognize similar mountainous scenes is further optimized in the future.

[0091] Comparative experiment: The causal inference model reconstructed by the multi-competitive hypothesis intervention of this invention (with added time-aligned preprocessing) was compared with the traditional XGBoost fire risk assessment model and the Bayesian network early warning model on the same test dataset of 1260 mountainous cultural tourism scenic spots. The comparison results are shown in Table 3 and... Figure 4 As shown.

[0092] The formulas for calculating the evaluation indicators are as follows: ; ; ; ; ; in, For accuracy; This is a true example, indicating that the sample itself does indeed have a fire hazard, and the model determines that there is a hazard (early warning). A true negative example means that the sample itself has no fire hazard, and the model determines that there is no hazard (no alarm is triggered). A false positive (false alarm) indicates that the sample itself has no fire hazard, but the model determines that there is a hazard and issues an alarm. A false negative (missed alarm) indicates that a fire hazard actually exists in the sample, but the model determines that it is safe and no alarm is triggered, thus the danger is missed. for Fraction; For accuracy; This refers to the recall rate.

[0093] Table 3 Comparison of Test Metrics for Different Algorithm Models

[0094] As shown in Table 3, traditional Bayesian networks and XGBoost are statistical learning models that lack multi-source sensing time alignment, making them susceptible to temporal misalignment interference. Furthermore, they cannot distinguish disturbances caused by environmental and operational conditions, resulting in a high false alarm rate in complex mountainous environments. This invention adds time alignment preprocessing at the edge to eliminate temporal misalignment, and then reconstructs multi-hypothesis reasoning through causal intervention to identify alarm triggers from a causal perspective, thus improving accuracy. The score has improved significantly, the false alarm rate has dropped dramatically, and it is well-suited for complex fire monitoring scenarios in mountainous cultural and tourism scenic areas.

[0095] After the test is completed, the system anonymizes all risk statistics data and outputs them to the fire protection digital asset system through the digital asset interface unit, outputting the anonymized statistical assets to support compliance data services in the insurance and urban planning industries.

[0096] Therefore, this invention employs the aforementioned visualized, tiered early warning method and system. By intervening with multiple competing hypotheses to reconstruct causal reasoning, it distinguishes between fire hazards, environmental disturbances, and operational fluctuations, reducing false alarm rates in complex scenarios. Blockchain-based evidence storage ensures the auditable and traceable assessment process, while GIS visualization and tiered collaborative response establish a seamless emergency response chain. On-site verification samples enable closed-loop model iteration, outputting compliant fire safety digital assets, resulting in high early warning reliability and strong generalization capabilities.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A visual, hierarchical early warning method, characterized in that, Includes the following steps: S1. Collect the raw dataset of multi-source sensing and transmit it to the edge measurement; S2. After preprocessing the raw data transmitted to the edge, multi-source evidence feature vectors are extracted from suspected risk points, spliced ​​and fused to obtain lightweight raw fused feature vectors, and then uploaded to the cloud. S3. Reconstruct causal reasoning by intervening with multiple competing hypotheses and classify risk warning levels; S4. Construct a complete chain of evidence for a single early warning, calculate its SHA-256 hash digest, and store it on the consortium blockchain. S5. Perform differentiated point marking rendering based on point location and risk warning level, and generate regional risk heat map using kernel density estimation algorithm; S6. Implement differentiated and coordinated responses based on risk warning levels, including differentiated message push, work order issuance, and on-site hardware coordination. S7. Read the manually reviewed and labeled sample set, perform grid search optimization with the mean square error loss function as the optimization objective, update the parameters and then send them to each edge gateway for synchronization. S8. De-identify the evidence chain data, filter sensitive fields, and output a standardized fire risk statistics dataset to provide compliant data services to external parties as a digital asset for fire analysis and assessment.

2. The visualized hierarchical early warning method according to claim 1, characterized in that, In S1, on-site physical sensing information from IoT fire protection terminals is collected, and multi-source sensing raw datasets are output. And transmit the original multi-source sensing dataset to the edge: ； ； In the formula, For the original dataset of multi-source perception, Indicates the first Terminal sampling records; This indicates the total number of terminal sampling records. Sampling time; The location's geographical coordinates; For sensor measurement array, This is a fault code from the device's self-test. To identify confidence levels.

3. The visualized hierarchical early warning method according to claim 2, characterized in that, In S1, the IoT fire protection terminal includes a smart smoke detector, a combustible gas detector, a fire water pressure monitoring terminal, an electrical fire monitoring terminal, and an AI video flame recognition camera.

4. The visualized hierarchical early warning method according to claim 3, characterized in that, S2 specifically refers to: S21, Firstly adopt The criterion for identifying and removing accidental noise sampling points is expressed as follows: ； In the formula, The mean value is obtained by sampling through a sliding window. The standard deviation of the sliding window sampling; S22. Obtain the standardized dataset using max-min normalization. The normalization formula is as follows: ； In the formula, These are the normalized feature components; These are the original measured values ​​of the characteristic components to be normalized. This represents the maximum value within the statistical window for this feature component. This is the minimum value within the statistical window of this feature component; S23, Read the scene adaptation threshold library Perform local primary screening, threshold library The threshold values ​​for each indicator are divided into four scenarios: high-rise buildings, old residential areas, industrial parks, and mountainous cultural tourism scenic areas. The data is then compared point-by-point with the standardized dataset. If all features fall within the corresponding scene threshold range, the dataset is deemed risk-free, and the dataset is standardized. Stored locally at the edge, without being uploaded to the cloud; If at least one feature exceeds the scene threshold, it is marked as a suspected risk point, triggering lightweight feature extraction. The extracted features include physical evidence feature vectors. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations ; S24. Transform the physical evidence feature vector Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations After concatenation and fusion, a lightweight original fused feature vector is obtained: ； Lightweight original fused feature vector Upload to the cloud to achieve bandwidth offloading.

5. The visualized hierarchical early warning method according to claim 4, characterized in that, S3 includes the following steps: S31. For the original fused feature vector Decoupling yields physical evidence feature vectors. Feature vector of environmental interference evidence Evidence vector of operating condition fluctuations ; S32. Employing multiple competing hypotheses for intervention and reconstruction, simulating feature performance under different hypotheses: ； In the formula, This indicates the characteristics of fire hazard hypothesis intervention and reconstruction; This indicates the characteristics of environmental disturbance hypothesis intervention reconstruction; This indicates that the characteristics of the intervention and reconstruction are based on the assumption of operating condition fluctuations; The number of dimensions of physical evidence features; The number of dimensions for evidence of environmental interference; The number of dimensions for evidence of operating condition fluctuations; It is a zero-dimensional vector of physical features; It is a zero vector of environmental feature dimensions; The zero vector represents the dimension of the working condition features; S33. Use cosine similarity to measure the degree of matching between the original features and the reconstructed features, and calculate the explanatory power of each hypothesis: ； in, Explanation degree of fire hazard assumptions; Explanation degree for the environmental disturbance hypothesis; Explanatory power for the assumption of operating condition fluctuations; Represents the L2 norm; S34. Determine the authenticity of potential hazards and set a threshold for hypothetical determination. ,when and If the alarm is triggered, it is considered a genuine fire hazard; otherwise, it is considered a false alarm. Calculate the overall risk score : ； in, For sensor confidence weights, ; This is a scene correction factor. ; ; Risk warning levels will be output according to the following rules. : 。 6. The visualized hierarchical early warning method according to claim 5, characterized in that, In S4, a complete chain of evidence for a single early warning is constructed, recording all key information from this assessment to form the original complete chain of evidence. : ； in, System timestamp; The coordinates of the location are latitude and longitude. To unify the timestamps after alignment; Mark as a false alarm; The hash digest is calculated using the following formula: ； Store the hash digest on the consortium blockchain, and store the original complete chain of evidence. Stored in the local evidence chain storage module database.

7. The visualized hierarchical early warning method according to claim 6, characterized in that, In S5, based on the latitude and longitude of the location Risk warning level Perform differentiated point rendering: Marked in blue; Marked in yellow; Marked in orange; High-risk layers are highlighted with a red flashing indicator and placed at the top. A kernel density estimation algorithm is used to generate a regional risk heatmap, which visually displays the spatial distribution of risks. ； In the formula, Indicates position Risk density estimate; This indicates the total number of risk points involved in the heatmap calculation; Indicates bandwidth; Represents the kernel function; Indicates the first Geographical location of each risk point.

8. The visualized hierarchical early warning method according to claim 7, characterized in that, In S6, the tiered and differentiated linkage handling rules are as follows: The platform only sends risk warnings to local grid workers, without hardware integration, and incorporates them into daily inspection tasks. We pay close attention to risks and use in-station messages and SMS to push information to grid members and enterprise safety administrators to automatically generate hazard verification work orders, requiring on-site verification within a specified time limit; In case of risk, the system will send out warnings through multiple channels, including in-station messages, SMS, and APP, and coordinate with on-site sound and light warning devices to issue emergency response work orders and go to the site for verification as soon as possible. In the event of an emergency, the system will notify the fire and rescue brigade and local duty stations, coordinate emergency broadcasts and audible and visual alarms, and provide rescue route planning and suggestions for dispatching surrounding rescue resources.

9. A visualized hierarchical early warning method according to claim 8, characterized in that, In S7, read the manually reviewed and labeled sample set. ,in, For the first Manually labeled training samples; The mean squared error loss function is used to measure the deviation between the model's output risk score and the human-generated true label. The expression for the loss function is as follows: ； in, This represents the total number of training samples; For the first Each manual sample outputs a comprehensive risk score; The training sample index numbers are manually labeled, and their values ​​are set to... ; by The optimization objective is to perform a grid search to find the optimal sensor confidence weights. Scene correction coefficient Scene adaptation threshold library .

10. A visualized hierarchical early warning system for implementing the method according to any one of claims 1-9, characterized in that, It includes a perception and acquisition module, an edge-side preprocessing module, a cloud-based causal analysis and grading module, an evidence chain storage module, a blockchain evidence storage sub-unit, a GIS visualization and rendering module, a tiered disposal linkage module, an iterative optimization module, and a digital asset interface unit; The sensing and acquisition module is used to collect raw datasets from multiple sensing sources. The output is transmitted to the edge-side preprocessing module; The edge-side preprocessing module is deployed on the edge gateway and includes a data cleaning unit, a time-aligned preprocessing unit, a normalization unit, a local initial screening unit, and a lightweight feature extraction unit. The cloud-based causal analysis and grading module includes a physical causal guidance unit, a multi-competitive hypothesis reasoning unit, a risk scoring unit, and a grading unit. The evidence chain storage module stores a complete single early warning evidence chain. The blockchain evidence storage subunit calculates the hash digest of the evidence chain. The hash digest is uploaded to the consortium blockchain; The GIS visualization rendering module includes a point rendering unit, a multi-layer linkage unit, an evidence pop-up unit, and a risk heat map generation sub-unit. The tiered response and coordination module has a built-in tiered response strategy library, which is based on the risk warning level. Perform differentiated message push, work order issuance, and hardware linkage to match business processing actions; The iterative optimization module reads the manually reviewed sample set and uses the mean squared error loss function as the optimization objective to iteratively update the sensor confidence weights through grid search. Scene correction coefficient Scene adaptation threshold library Update the parameters and send them to the edge-side preprocessing module for synchronization; The digital asset interface unit anonymizes evidence chain data, outputs standardized fire risk statistics sets, and provides compliant digital asset data services to external parties.

Images