Fire-fighting intelligent early warning system and method based on internet of things

Abstract

The application discloses a fire-fighting intelligent early warning system and method based on Internet of Things, and belongs to the technical field of fire-fighting intelligent early warning. Temperature, smoke concentration and CO concentration data are acquired by an Internet of Things detector, site position, regional function division and article flammability distribution data are synchronously collected, and a composite attribute label containing spatial coordinates, function division and flammability grade is generated after preprocessing. A device operation state space-time feature network is constructed based on device historical data, multi-dimensional time sequence features are extracted, a regional risk transmission coefficient is calculated in combination with the composite attribute label, a differentiated early warning threshold matrix is generated by fusing multiple parameters, a virtual monitoring grid is divided according to the correlation degree of regional function and flammability grade, a suspected risk area is identified, and early warning is triggered after time-space cross verification determines that the suspected risk area is an effective risk. When early warning is triggered, a fire spread is simulated in combination with site topological data, a dynamic risk map is generated, early warning information is pushed, and an emergency plan is generated.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire warning technology, specifically an intelligent fire warning system and method based on the Internet of Things. Background Technology

[0002] In recent years, with the rapid development of IoT technology, fire alarm systems have gradually evolved towards intelligence and networking. Traditional fire monitoring mainly relies on independent devices such as smoke detectors and temperature sensors, triggering alarms through preset fixed thresholds. While these systems can play a certain role in simple scenarios, in complex environments (such as factories, warehouses, and large commercial complexes), due to dynamic environmental changes, differences in equipment operating status, and the influence of building structures, their accuracy and timeliness in providing early warnings often fail to meet actual needs.

[0003] Despite existing technologies that have improved upon this by employing multi-sensor data fusion and machine learning algorithm optimization, several key issues remain: First, there is a lack of coordinated analysis of temporal and spatial characteristics, such as failing to dynamically adjust early warning strategies in conjunction with equipment operating cycles or ignoring the spread trend of local anomalies within a spatial grid. Second, the ability to predict fire propagation paths is insufficient, resulting in the system only being able to passively respond to fires that have already occurred, rather than predicting the direction of risk spread in advance, thus affecting the accuracy and efficiency of emergency measures. These limitations make it difficult for traditional early warning systems to respond promptly and accurately in complex environments, especially in complex locations such as large-scale factories and commercial buildings, where the speed and complexity of fire spread increase the difficulty of early warning and emergency response. Summary of the Invention

[0004] The purpose of this invention is to provide a fire intelligent early warning system and method based on the Internet of Things to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a fire intelligent early warning method based on the Internet of Things, comprising:

[0006] S1: Real-time monitoring data of temperature, smoke concentration and CO concentration are acquired through IoT detectors distributed throughout the site; basic information of the site is collected simultaneously, including site location, functional division of areas and distribution of flammability of items; the collected raw data is preprocessed; and composite attribute tags containing spatial coordinates, functional zoning and flammability level are generated.

[0007] S2: Establish a spatiotemporal feature network for equipment operation status. Input historical operation data of core production equipment, environmental control equipment and fire-fighting equipment in the site into an autoencoder network to extract multidimensional temporal features. Combine this with composite attribute labels to construct a regional risk propagation impact map. Calculate the risk transmission coefficient between different regions through a graph attention mechanism. Integrate multidimensional temporal features, risk transmission coefficients and real-time environmental parameters to generate a differentiated early warning threshold matrix that dynamically changes with time and regional characteristics.

[0008] S3: Based on composite attribute tags, feature values ​​are assigned to regional functional types and flammability levels of items, and correlation indices are calculated to divide virtual monitoring grids. Each grid is configured with an independent monitoring point group, a three-dimensional time-series data array is constructed, and anomaly comprehensive values ​​are calculated to identify and mark suspected risk areas. A differentiated early warning threshold matrix is ​​called for spatiotemporal cross-validation. Only when the time period of grid anomaly occurrence matches the risk period predicted by the spatiotemporal feature network of equipment operation status is it determined as a valid risk and an early warning is triggered.

[0009] S4: Store site topology data. When an effective warning is triggered, combine the current environmental parameters to simulate the fire spread trajectory, generate a two-dimensional dynamic risk map with risk level, propagation path and fire resource location marked, push warning information to relevant terminals and generate emergency response plan.

[0010] Furthermore, S1 includes:

[0011] S1.1: IoT detectors deployed in various areas of the site collect environmental data in real time according to a preset sampling frequency. Temperature detectors, smoke concentration detectors, and CO concentration detectors record corresponding parameter values ​​according to their respective set collection cycles. All detectors transmit the collected data to the data processing center. The received raw monitoring data is preprocessed, and instantaneous pulse noise is removed by filtering algorithms. The values ​​of different types of detectors are normalized to convert the temperature, smoke concentration, and CO concentration to a preset value range.

[0012] S1.2: The data processing center synchronously receives the basic information set I={I1,I2,I3} uploaded by the site management system, where I1 is the site location information, I2 is the regional functional division result, and I3 is the flammability distribution data of items; it performs coordinate mapping transformation on the site location information I1 to generate a location matrix P={P1,P2,…,Pm} in the standard coordinate system, where Pm represents the coordinate vector of the m-th region; it associates the regional functional division result I2 with the location matrix P to construct a functional location mapping table Mf={<P1,F1> ,<P2,F2> ,…,<Pm,Fm>}, where Fm represents the functional type of the m-th region;

[0013] S1.3: According to the preset flammability rating classification rule R={R1,R2,R3}, where R1 is highly flammable, R2 is moderately flammable, and R3 is lowly flammable, the flammability distribution data I3 of the items is classified and processed to generate a flammability rating matrix H={H1,H2,…,Hm}, where Hm represents the flammability rating of the items in the m-th region; the flammability rating matrix H is merged with the functional location mapping table Mf, and the composite attribute label T={ for each region is marked on the electronic map.<P1,F1,H1> ,<P2,F2,H2> ,...,<Pm,Fm,Hm>}; Establish an index structure Idx that links monitoring data with basic information.<D1,T1> ,<D2,T2> ,...,<Dn,Tn>}, where Dn represents the real-time data set of the nth monitoring point, and Tn is the composite attribute label corresponding to the monitoring point; through the spatial neighborhood analysis algorithm, the nearest neighbor regional attribute label is matched for each monitoring point. When the monitoring point is located at the regional boundary, the weighted fusion method is used to calculate its comprehensive attribute value, which is the fusion result of the functional type weight and the flammability level weight.

[0014] Furthermore, S2 includes:

[0015] S2.1: Construct a spatiotemporal feature network Gs=(Vs,Es) for equipment operation status, where the node set Vs={vs1,vs2,...,vsn} represents the equipment status vector at different time periods, and vsn represents the comprehensive operation parameter vector of the core production equipment, environmental control equipment, and fire-fighting equipment in the site during the nth time period; the edge set Es represents the temporal correlation between time periods, and the weight of the edge reflects the correlation strength of the equipment status between adjacent time periods; perform sliding window segmentation on the historical equipment operation data, and extract multidimensional features in each window, including temperature features, fluctuation features, peak features, and temporal features; input the extracted original feature vectors into an autoencoder network for feature dimensionality reduction and reconstruction to generate a low-dimensional spatiotemporal feature vector STV={stv1,stv2,...,stvj}, where stvj represents the core feature vector of the jth time window;

[0016] S2.2: Based on the composite attribute label T, construct the risk propagation impact graph IG=(N,A), where the node set N represents different areas within the site, the edge set A represents the risk transmission path between areas, and the edge weights represent the risk transmission coefficients between areas. The risk transmission coefficients are calculated from the area function type F, the flammability level of the item H, and the spatial distance attenuation factor. Using a graph attention mechanism, calculate the risk impact weight matrix W of each area on the surrounding areas.

[0017] Step S2.3: Design a dynamic threshold generation neural network. Its input layer includes the spatio-temporal feature vector STV, the risk impact weight matrix W, and the environmental parameter vector E(t) at the current time period. The hidden layer captures the temporal dependence relationship through a gated recurrent unit and combines the attention mechanism to perform weighted fusion on the input features. The output layer generates a differential warning threshold matrix Th(t) = {Th1(t), Th2(t),..., Thm(t)}; where Thm(t) represents the warning threshold of the mth region within the t time period, and the calculation formula is: Thm(t) = BaseThm × f(stvj, Wmj, E(t)); where BaseThm is the baseline threshold, Wmj represents the risk impact weight of the mth region on the surrounding regions; f is a non-linear mapping function obtained by training with historical data, and the weight coefficients of spatio-temporal features, risk propagation, and environmental parameters are automatically allocated through the attention mechanism. When a fire actually occurs in a certain region but no warning is triggered for multiple consecutive time periods, the model parameters are adjusted through an inverse reinforcement learning algorithm to increase the risk weight of this region during this time period.

[0018] Further, S3 includes:

[0019] S3.1: Based on the composite attribute label T, calculate the correlation index between the function type and the flammability level in each region: Assign characteristic values to the function type and the flammability level of each region. The characteristic value of the function type is denoted as Fvg, and the characteristic value of the flammability level is denoted as Hvg. Calculate the correlation index Rg through the formula: Rg = |Fvg × Hvg| / (1 + |Fvg - Hvg|); and the value range of Rg is [0, 1]. The larger the value, the stronger the risk correlation between the function type and the flammability level of this region.

[0020] Divide the site into initial uniform units U0 = {u01, u02,..., u0g}, where u0g represents the gth initial uniform unit, and each unit corresponds to a unique spatial range. Calculate the average value R0g of the correlation index of all regions within each initial unit.

[0021] Set the information entropy threshold Et, calculate the information entropy E of the current unit set as E = -Σ(Rg × log2Rg). If E > Et, split the unit with the R0g difference exceeding the threshold ΔR into two sub-units and recalculate the average value of the correlation index of the sub-units; if E < Et, merge the adjacent units with the R0g difference less than ΔR, and repeat the splitting and merging operations until the information entropy E converges to the interval [Et - ΔE, Et + ΔE].

[0022] The final grid set is formed, satisfying the condition that the grid area Sg is inversely proportional to the mean correlation index Rg: Sg = S0 / Rg, where S0 is the baseline area; at the same time, the following constraints are set: for high correlation regions Rg≥θ1, the grid area does not exceed λ1 times the baseline area; for low correlation regions Rg≤θ2, the grid area is not less than λ2 times the baseline area; where θ1 and θ2 are the set high correlation threshold and low correlation threshold, respectively, and θ1>θ2; λ1 and λ2 are the area coefficients of high correlation regions and low correlation regions, respectively, and λ1<1, λ2>1.

[0023] S3.2: Configure an independent data acquisition point group for each grid. The acquisition point group includes three types of monitoring points: temperature, smoke concentration, and CO concentration. The spatial deviation of each monitoring point from its grid does not exceed 1 / 5 of the grid's side length. Construct a three-dimensional time series array Dg(t)={Tg(t),Yg(t),Cg(t)} for the real-time monitoring data of each grid, where Tg(t) is the temperature time series, Yg(t) is the smoke concentration time series, and Cg(t) is the CO concentration time series. Calculate the dynamic baseline values ​​of each parameter. Based on the grid's historical normal data, generate the baseline values ​​Tg0(t), Yg0(t), and Cg0(t) for each moment using a moving average algorithm. The baseline values ​​are dynamically updated according to the ambient temperature and equipment operating status. Extract abnormal feature indicators, including parameter deviation (the relative deviation between the current value and the baseline value), acceleration of change (the rate of change of deviation per unit time), and parameter consistency (the consistency coefficient of the deviation trends of the three types of parameters).

[0024] An anomaly comprehensive value Ag is generated by weighted fusion of anomaly feature indicators. The weight coefficients are dynamically adjusted according to the grid correlation index Rg. The higher Rg is, the greater the weight of the temperature parameter. When Ag exceeds the anomaly threshold corresponding to the grid type for x consecutive sampling periods, the grid is marked as a suspected risk area.

[0025] S3.3 performs spatiotemporal cross-validation:

[0026] Time dimension verification: Extract the anomaly start time t0 of the suspected risk area, combine it with the differentiated early warning threshold matrix Th(t), obtain the threshold Th(t0) of the grid in the time period to which t0 belongs, calculate the ratio of the anomaly comprehensive value Ag to Th(t0), and if the ratio is greater than or equal to the set threshold, the time verification is passed;

[0027] Spatial dimension verification: Analyze the abnormal correlation between the suspected risk area and the three adjacent grids, calculate the spatial correlation index, which is the ratio of the time difference of the occurrence of anomalies in adjacent grids to the spatial distance. When the correlation index of at least two adjacent grids is less than or equal to the preset propagation coefficient, the spatial verification is passed.

[0028] Only when both time and space dimensions are verified simultaneously is it identified as a valid risk area and an early warning is triggered; otherwise, a secondary verification is performed, retrieving the device operation logs and detector status data of the grid to rule out anomalies caused by device malfunctions or environmental interference.

[0029] Furthermore, S4 includes:

[0030] S4.1: Pre-store the site's topological structure data and establish a correlation mapping with the composite attribute label T generated in S1.3 to form a fusion model that includes spatial features and risk attributes; the topological structure data specifically includes:

[0031] Building infrastructure data: wall locations, floor structure, and fire compartment boundaries in the building plan, linked to the coordinate vectors of the location matrix P; Pipeline network data: ventilation duct routing, vent locations, and fan parameters, linked to the area function type F; Fire protection facility data: fire door locations, sprinkler head distribution, and fire extinguisher storage locations, bound to the flammability rating H of the corresponding area; The above data is integrated into a topological relationship set Topo={ using a spatial association algorithm.<P1,F1,H1,Str1,Pipe1,Fac1> ,...,<Pm,Fm,Hm,Strm,Pipem,Facm>}, where Strm, Pipem, and Facm represent the building structure, pipeline network, and fire protection facilities characteristics of the m-th region, respectively;

[0032] S4.2: When S3 triggers an effective warning, the associated data of the risk area in the topology set Topo is called, and the fire spread simulation is started in combination with the current environmental parameters: Taking the grid coordinates where the warning is triggered as the starting point, the composite attribute label Tm and topological features of the area are extracted; the combustion rate coefficient is determined based on Hm, and the heat diffusion rate is calculated in combination with the current environmental parameters; the smoke propagation path is corrected by ventilation parameters, and the propagation blocking effect is judged according to the current status of the fire door; the simulation results generate a spread time series matrix M(t), which records the expected temperature, smoke concentration and risk level of each area at time t. The risk level is determined by fusing the personnel density coefficient of Fm and the combustion intensity coefficient of Hm.

[0033] S4.3: A two-dimensional risk map is generated based on the propagation time series matrix M(t). The map layer is aligned with the virtual monitoring grid divided in S3.1 to achieve spatial correlation between the warning area and the simulation results. The map annotations include: Core risk area: corresponding to the effective risk grid marked in S3, with the current temperature, smoke concentration and ignition time marked; Propagation path zone: according to the propagation direction of each time period in M(t), the fire spread trajectory along the pipes and channels is marked, as well as the corresponding propagation speed; Dynamic risk markers: each grid is marked with the expected time window and risk level of being affected; Fire resource points: the location of the fire-fighting facilities closest to the risk area is extracted from the Topo, and their type and availability status are marked; Finally, the above risk map information is integrated to generate an emergency response plan and push the warning information to relevant terminals.

[0034] A fire early warning system based on the Internet of Things includes a data fusion processing module, a dynamic risk analysis module, a grid monitoring and verification module, and a fire response module.

[0035] The data fusion and processing module acquires real-time monitoring data on temperature, smoke concentration, and CO concentration through IoT detectors distributed throughout the site; it simultaneously collects basic site information, including site location, functional division of areas, and distribution of flammability of items; it preprocesses the collected raw data; and it fuses and generates composite attribute tags that include spatial coordinates, functional zoning, and flammability rating.

[0036] The dynamic risk analysis module establishes a spatiotemporal feature network of equipment operation status. It inputs historical operation data of core production equipment, environmental control equipment, and fire-fighting equipment in the site into an autoencoder network to extract multidimensional temporal features. It then combines composite attribute labels to construct a regional risk propagation impact map and calculates the risk transmission coefficient between different regions through a graph attention mechanism. Finally, it integrates multidimensional temporal features, risk transmission coefficients, and real-time environmental parameters to generate a differentiated early warning threshold matrix that dynamically changes with time and regional characteristics.

[0037] The grid monitoring verification module assigns feature values ​​to regional functional types and flammability levels of items based on composite attribute tags and calculates correlation indices to divide virtual monitoring grids. It configures independent monitoring point groups for each grid, constructs a three-dimensional time-series data array, calculates comprehensive anomaly values, identifies and marks suspected risk areas, and calls a differentiated early warning threshold matrix for spatiotemporal cross-verification. Only when the time period of grid anomaly occurrence matches the risk period predicted by the spatiotemporal feature network of equipment operation status is it determined as a valid risk and an early warning is triggered.

[0038] The fire response module stores site topology data. When an effective warning is triggered, it combines the current environmental parameters to simulate the fire spread trajectory, generate a two-dimensional dynamic risk map that marks the risk level, propagation path and location of fire-fighting resources, pushes the warning information to relevant terminals and generates an emergency response plan.

[0039] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0040] This invention deeply integrates environmental monitoring and site information, binding real-time data such as temperature and smoke with attribute tags such as regional functions and flammability ratings. It combines an autoencoder and graph attention mechanism to extract equipment operating characteristics and calculate regional risk transmission coefficients, generating a dynamically changing warning threshold matrix that transcends the limitations of traditional fixed-threshold warnings, making risk assessment more aligned with real-world scenarios. Simultaneously, it achieves adaptive grid division through a correlation index, and combines a three-dimensional time-series data array with anomaly comprehensive value calculation. Through a spatiotemporal cross-validation mechanism, it effectively improves the accuracy of risk identification and reduces false alarms and missed alarms. Furthermore, this invention achieves a shift from passive response to proactive prediction in the emergency response phase. By linking pre-stored site topology data with real-time warning information and simulating fire spread trajectories using environmental parameters, it generates a dynamic risk map, improving the timeliness and accuracy of fire warnings and significantly enhancing the comprehensive prevention and control capabilities against fire risks. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0042] Figure 1 This is a flowchart of a fire intelligent early warning method based on the Internet of Things. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 The present invention provides a technical solution: a fire intelligent early warning method based on the Internet of Things, comprising:

[0045] S1: Real-time monitoring data of temperature, smoke concentration and CO concentration are acquired through IoT detectors distributed throughout the site; basic information of the site is collected simultaneously, including site location, functional division of areas and distribution of flammability of items; the collected raw data is preprocessed; and composite attribute tags containing spatial coordinates, functional zoning and flammability level are generated.

[0046] S2: Establish a spatiotemporal feature network for equipment operation status. Input historical operation data of core production equipment, environmental control equipment and fire-fighting equipment in the site into an autoencoder network to extract multidimensional temporal features. Combine this with composite attribute labels to construct a regional risk propagation impact map. Calculate the risk transmission coefficient between different regions through a graph attention mechanism. Integrate multidimensional temporal features, risk transmission coefficients and real-time environmental parameters to generate a differentiated early warning threshold matrix that dynamically changes with time and regional characteristics.

[0047] S3: Based on composite attribute tags, feature values ​​are assigned to regional functional types and flammability levels of items, and correlation indices are calculated to divide virtual monitoring grids. Each grid is configured with an independent monitoring point group, a three-dimensional time-series data array is constructed, and anomaly comprehensive values ​​are calculated to identify and mark suspected risk areas. A differentiated early warning threshold matrix is ​​called for spatiotemporal cross-validation. Only when the time period of grid anomaly occurrence matches the risk period predicted by the spatiotemporal feature network of equipment operation status is it determined as a valid risk and an early warning is triggered.

[0048] S4: Store site topology data. When an effective warning is triggered, combine the current environmental parameters to simulate the fire spread trajectory, generate a two-dimensional dynamic risk map with risk level, propagation path and fire resource location marked, push warning information to relevant terminals and generate emergency response plan.

[0049] S1 includes:

[0050] S1.1: IoT detectors deployed in various areas of the site collect environmental data in real time according to a preset sampling frequency. Temperature detectors, smoke concentration detectors, and CO concentration detectors record corresponding parameter values ​​according to their respective set collection cycles. All detectors transmit the collected data to the data processing center. The received raw monitoring data is preprocessed, and instantaneous pulse noise is removed by a filtering algorithm. The values ​​of different types of detectors are normalized, and the temperature, smoke concentration, and CO concentration are uniformly converted to a preset value range of [0,1]. For the temperature parameter, it is linearly normalized based on the site's historical highest temperature Tmax and lowest temperature Tmin using the formula (current temperature - Tmin) / (Tmax - Tmin). The smoke concentration and CO concentration are normalized based on their corresponding safety thresholds Csafe using the formula: current concentration / Csafe. When the concentration exceeds the threshold, it is treated as 1.

[0051] S1.2: The data processing center synchronously receives the basic information set I={I1,I2,I3} uploaded by the site management system, where I1 is the site location information, I2 is the regional functional division result, and I3 is the flammability distribution data of items; it performs coordinate mapping transformation on the site location information I1 to generate a location matrix P={P1,P2,…,Pm} in the standard coordinate system, where Pm represents the coordinate vector of the m-th region; it associates the regional functional division result I2 with the location matrix P to construct a functional location mapping table Mf={<P1,F1> ,<P2,F2> ,…,<Pm,Fm>}, where Fm represents the functional type of the m-th region;

[0052] S1.3: According to the preset flammability rating classification rule R={R1,R2,R3}, where R1 is highly flammable, R2 is moderately flammable, and R3 is lowly flammable, the flammability distribution data I3 of the items is classified and processed to generate a flammability rating matrix H={H1,H2,…,Hm}, where Hm represents the flammability rating of the items in the m-th region; the flammability rating matrix H is merged with the functional location mapping table Mf, and the composite attribute label T={ for each region is marked on the electronic map.<P1,F1,H1> ,<P2,F2,H2> ,...,<Pm,Fm,Hm>}; Establish an index structure Idx that links monitoring data with basic information.<D1,T1> ,<D2,T2> ,...,<Dn,Tn>}, where Dn represents the real-time data set of the nth monitoring point, and Tn is the composite attribute label corresponding to the monitoring point; through the spatial neighborhood analysis algorithm, the nearest neighbor regional attribute label is matched for each monitoring point; when the monitoring point is located at the regional boundary, the weighted fusion method is used to calculate its comprehensive attribute value, which is the fusion result of the functional type weight and the flammability level weight;

[0053] When a monitoring point is located at the boundary of an area, the weight calculation for the weighted fusion is based on the distance from the monitoring point to the center of each adjacent area. The closer the distance, the greater the weight. The weight coefficient is obtained by normalizing the inverse of the distance. For example, if the distance to area A is d1 and the distance to area B is d2, then the weights are d2 / (d1+d2) and d1 / (d1+d2) respectively. The comprehensive attribute value includes the fusion result of functional type weight and flammability level weight. For example, if the monitoring point is adjacent to area A (functional type: warehouse, flammability level: R1) and area B (functional type: office, flammability level: R3), then the functional type in the comprehensive attribute value is fused according to the distance weight, and the flammability level is calculated by weighting the level values ​​of R1 and R3, such as R1=3, R2=2, R3=1.

[0054] Furthermore, S2 includes:

[0055] S2.1: Construct a spatiotemporal feature network Gs=(Vs,Es) for equipment operation status, where the node set Vs={vs1,vs2,...,vsn} represents the equipment status vector at different time periods, and vsn represents the comprehensive operation parameter vector of the core production equipment, environmental control equipment, and fire-fighting equipment in the site during the nth time period; the edge set Es represents the temporal correlation between time periods, and the weight of the edge reflects the correlation strength of the equipment status between adjacent time periods; perform sliding window segmentation on the historical equipment operation data, and extract multidimensional features in each window, including temperature features, fluctuation features, peak features, and temporal features; input the extracted original feature vectors into an autoencoder network for feature dimensionality reduction and reconstruction to generate a low-dimensional spatiotemporal feature vector STV={stv1,stv2,...,stvj}, where stvj represents the core feature vector of the jth time window;

[0056] S2.2: Based on the composite attribute label T, construct the risk propagation impact graph IG=(N,A), where the node set N represents different areas within the site, the edge set A represents the risk transmission path between areas, and the edge weights represent the risk transmission coefficients between areas. The risk transmission coefficients are calculated from the area function type F, the flammability level of the item H, and the spatial distance attenuation factor. Using a graph attention mechanism, calculate the risk impact weight matrix W of each area on the surrounding areas.

[0057] Step S2.3: Design a dynamic threshold generation neural network. Its input layer includes a spatiotemporal feature vector STV, a risk impact weight matrix W, and an environmental parameter vector E(t) for the current time period. The hidden layer captures temporal dependencies through gated recurrent units and combines an attention mechanism to perform weighted fusion of input features. The output layer generates a differentiated early warning threshold matrix Th(t) = {Th1(t), Th2(t), ..., Thm(t)}. Thm(t) represents the early warning threshold of the m-th region within time period t, and its calculation formula is: Thm(t) = BaseThm × f(stvj, Wmj, E(t)). Where BaseThm is the baseline threshold, Wmj represents the risk impact weight of the m-th region on surrounding regions, and f is a nonlinear mapping function trained using historical data. The attention mechanism automatically assigns weight coefficients to spatiotemporal features, risk propagation, and environmental parameters. When a fire actually occurs in a region but no early warning is triggered for several consecutive time periods, the model parameters are adjusted using an inverse reinforcement learning algorithm to increase the risk weight of that region during that time period.

[0058] Furthermore, S3 includes:

[0059] S3.1: Based on the composite attribute label T, calculate the correlation index between the function type and the flammability level in each area: Assign characteristic values to the function type and the flammability level of each area. The characteristic value of the function type is denoted as Fvg, and the characteristic value of the flammability level is denoted as Hvg. Through the formula: Rg = |Fvg × Hvg| / (1 + |Fvg - Hvg|); calculate the correlation index Rg, and the value range of Rg is [0, 1]. The larger the value, the stronger the risk correlation between the function type and the flammability level in this area;

[0060] Divide the site into initial uniform units U0 = {u01, u02,..., u0g}, where u0g represents the g-th initial uniform unit, and each unit corresponds to a unique spatial range. Calculate the average correlation index R0g of all areas within each initial unit;

[0061] Set the information entropy threshold Et, and calculate the information entropy E of the current unit set as E = -Σ(Rg × log2Rg). If E > Et, split the units with R0g difference exceeding the threshold ΔR into two sub-units and recalculate the average correlation index of the sub-units; if E < Et, merge the adjacent units with R0g difference less than ΔR. Repeat the splitting and merging operations until the information entropy E converges to the interval [Et - ΔE, Et + ΔE];

[0062] Finally, form a grid set, where the grid area Sg is inversely proportional to the average correlation index Rg: Sg = S0 / Rg, where S0 is the reference area; at the same time, set the constraint conditions: for the high-correlation areas with Rg ≥ θ1, the grid area does not exceed λ1 times the reference area; for the low-correlation areas with Rg ≤ θ2, the grid area is not less than λ2 times the reference area; where θ1 and θ2 are the set high-correlation threshold and low-correlation threshold respectively, and θ1 > θ2; λ1 and λ2 are the high-correlation area coefficient and low-correlation area coefficient respectively, and λ1 < 1, λ2 > 1.

[0063] S3.2: Configure an independent data acquisition point group for each grid. The acquisition point group includes three types of monitoring points: temperature, smoke concentration, and CO concentration. The spatial deviation of each monitoring point from its grid does not exceed 1 / 5 of the grid's side length. Construct a three-dimensional time series array Dg(t)={Tg(t),Yg(t),Cg(t)} for the real-time monitoring data of each grid, where Tg(t) is the temperature time series, Yg(t) is the smoke concentration time series, and Cg(t) is the CO concentration time series. Calculate the dynamic baseline values ​​of each parameter. Based on the grid's historical normal data, generate the baseline values ​​Tg0(t), Yg0(t), and Cg0(t) for each moment using a moving average algorithm. The baseline values ​​are dynamically updated according to the ambient temperature and equipment operating status. Extract abnormal feature indicators, including parameter deviation (the relative deviation between the current value and the baseline value), acceleration of change (the rate of change of deviation per unit time), and parameter consistency (the consistency coefficient of the deviation trends of the three types of parameters).

[0064] An anomaly comprehensive value Ag is generated by weighted fusion of anomaly feature indicators. The weight coefficients are dynamically adjusted according to the grid correlation index Rg. The higher Rg is, the greater the weight of the temperature parameter. When Ag exceeds the anomaly threshold corresponding to the grid type for x consecutive sampling periods, the grid is marked as a suspected risk area.

[0065] S3.3 performs spatiotemporal cross-validation:

[0066] Time dimension verification: Extract the anomaly start time t0 of the suspected risk area, combine it with the differentiated early warning threshold matrix Th(t), obtain the threshold Th(t0) of the grid in the time period to which t0 belongs, calculate the ratio of the anomaly comprehensive value Ag to Th(t0), and if the ratio is greater than or equal to the set threshold, the time verification is passed;

[0067] Spatial dimension verification: Analyze the abnormal correlation between the suspected risk area and the three adjacent grids, calculate the spatial correlation index, which is the ratio of the time difference of the occurrence of anomalies in adjacent grids to the spatial distance. When the correlation index of at least two adjacent grids is less than or equal to the preset propagation coefficient, the spatial verification is passed.

[0068] Only when both time and space dimensions are verified simultaneously is it identified as a valid risk area and an early warning is triggered; otherwise, a secondary verification is performed, retrieving the device operation logs and detector status data of the grid to rule out anomalies caused by device malfunctions or environmental interference.

[0069] Furthermore, S4 includes:

[0070] S4.1: Pre-store the site's topological structure data and establish a correlation mapping with the composite attribute label T generated in S1.3 to form a fusion model that includes spatial features and risk attributes; the topological structure data specifically includes:

[0071] Building infrastructure data: wall locations, floor structure, and fire compartment boundaries in the building plan, linked to the coordinate vectors of the location matrix P; Pipeline network data: ventilation duct routing, vent locations, and fan parameters, linked to the area function type F; Fire protection facility data: fire door locations, sprinkler head distribution, and fire extinguisher storage locations, bound to the flammability rating H of the corresponding area; The above data is integrated into a topological relationship set Topo={ using a spatial association algorithm.<P1,F1,H1,Str1,Pipe1,Fac1> ,...,<Pm,Fm,Hm,Strm,Pipem,Facm>}, where Strm, Pipem, and Facm represent the building structure, pipeline network, and fire protection facilities characteristics of the m-th region, respectively;

[0072] S4.2: When S3 triggers a valid warning, the associated data of the risk area in the topology set Topo is called, and the fire spread simulation is started in combination with the current environmental parameters: Taking the grid coordinates where the warning is triggered as the starting point, the composite attribute label Tm and topological features of the area are extracted; the combustion rate coefficient is determined based on Hm, and the heat diffusion rate is calculated in combination with the current environmental parameters; the smoke propagation path is corrected by the ventilation parameters of the ventilation duct direction, and the propagation blocking effect is judged according to the current status of the fire door; the simulation results generate a spread time series matrix M(t), which records the expected temperature, smoke concentration and risk level of each area at time t. The risk level is determined by fusing the personnel density coefficient of Fm and the combustion intensity coefficient of Hm.

[0073] S4.3: A two-dimensional risk map is generated based on the propagation time series matrix M(t). The map layer is aligned with the virtual monitoring grid divided in S3.1 to achieve spatial correlation between the warning area and the simulation results. The map annotations include: Core risk area: corresponding to the effective risk grid marked in S3, with the current temperature, smoke concentration and ignition time marked; Propagation path zone: according to the propagation direction of each time period in M(t), the fire spread trajectory along the pipes and channels is marked, as well as the corresponding propagation speed; Dynamic risk markers: each grid is marked with the expected time window and risk level of being affected; Fire resource points: the location of the fire-fighting facilities closest to the risk area is extracted from the Topo, and their type and availability status are marked; Finally, the above risk map information is integrated to generate an emergency response plan and push the warning information to relevant terminals.

[0074] An embodiment of the present invention: Taking a large industrial park as an example, the park includes different functional areas such as production workshops, storage areas, and office areas, and stores items with different flammability levels such as chemical raw materials and finished products; First, temperature, smoke concentration and CO concentration data are collected in real time by IoT detectors distributed in each area, and the park's location coordinates, functional zoning map and flammability list of items are collected simultaneously. After preprocessing, composite attribute tags are generated, such as "Production Workshop A (coordinates X1, Y1) - Functional type: Industrial production - Flammability level: Highly flammable (R1)";

[0075] Based on composite attribute tags, further analysis of equipment operation data is conducted: historical operating parameters of equipment such as workshop reactors, ventilation systems, and fire pumps, such as reactor temperature and fan speed, are input into an autoencoder to extract temporal characteristics such as temperature fluctuations and peak changes of the equipment at different times; combined with the functions and flammability ratings of each area, the risk transmission coefficient between areas is calculated through a graph attention mechanism: the transmission coefficient from the storage area to the production workshop is set to 0.8 due to the close distance and high material correlation, and finally a differentiated early warning threshold matrix is ​​generated. For example, the temperature early warning threshold for the production workshop during the high-load period from 9:00 to 18:00 every day is set to 55℃, while it is adjusted to 45℃ during the low-load period at night. Due to the high flammability rating, the threshold for the storage area is 10℃ lower than that for the office area.

[0076] Based on the functional type and flammability level correlation index, the regions are adaptively segmented. The correlation index between the production workshop (functional characteristic value 0.9) and highly flammable materials (characteristic value 0.8) is 0.78, and it is divided into a dense grid of 10m×10m. The correlation index between the office area (functional characteristic value 0.3) and low-flammable materials (characteristic value 0.2) is 0.06, and it is divided into a sparse grid of 30m×30m. Each grid is configured with an independent monitoring point group, and a three-dimensional time-series data array of temperature, smoke, and CO is constructed in real time. By calculating the deviation from the dynamic baseline, the acceleration of change, and the parameter synergy, an abnormal comprehensive value Ag is generated. When the Ag of a certain production workshop grid exceeds the threshold for three consecutive sampling cycles, it is marked. The area was identified as a suspected risk zone. The anomaly in this grid occurred at 9:30 AM (a period of high equipment load, threshold 55°C), and the Ag-to-threshold ratio reached 1.2 (verified by time). Simultaneously, two adjacent grids showed correlated anomalies within 5 minutes (verified by spatial data). This was deemed a valid risk and an alert was triggered. After the alert was activated, pre-stored park topology data (including wall fire resistance ratings, ventilation duct routing, and sprinkler head distribution) was retrieved. Combined with current humidity and wind speed, the fire spread was simulated: starting from the grid where the fire originated, a combustion rate coefficient of 1.2 was determined based on the high flammability rating (Hm=R1). The smoke diffusion path was corrected using ventilation duct parameters, generating a spread time series matrix, predicting that the fire would spread to adjacent storage areas within 10 minutes.

[0077] The final generated two-dimensional dynamic risk map marks the core risk area: current temperature 62℃, smoke concentration 0.3mg / m³, propagation path along the northwest-facing pipeline: speed 0.5m / min, estimated impact time for each grid: 10 minutes for the storage area, nearest fire-fighting resources: fire extinguishers are available 30m away, sprinkler systems are available 50m away, and an emergency response plan is automatically generated: priority is given to activating the sprinkler system in the fire area and downwind, and an early warning is pushed to the park's fire control room, realizing a fully intelligent response from risk monitoring to emergency handling.

[0078] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A fire intelligent early warning method based on the Internet of Things, characterized in that: The method includes: S1: Real-time monitoring data of temperature, smoke concentration and CO concentration are acquired through IoT detectors distributed throughout the site; basic information of the site is collected simultaneously, including site location, functional division of areas and distribution of flammability of items; the collected raw data is preprocessed; and composite attribute tags containing spatial coordinates, functional zoning and flammability level are generated. S2: Establish a spatiotemporal feature network for equipment operation status. Input historical operation data of core production equipment, environmental control equipment and fire-fighting equipment in the site into an autoencoder network to extract multidimensional temporal features. Combine this with composite attribute labels to construct a regional risk propagation impact map. Calculate the risk transmission coefficient between different regions through a graph attention mechanism. Integrate multidimensional temporal features, risk transmission coefficients and real-time environmental parameters to generate a differentiated early warning threshold matrix that dynamically changes with time and regional characteristics. S3: Based on composite attribute tags, feature values ​​are assigned to regional functional types and flammability levels of items, and correlation indices are calculated to divide virtual monitoring grids. Each grid is configured with an independent monitoring point group, a three-dimensional time-series data array is constructed, and anomaly comprehensive values ​​are calculated to identify and mark suspected risk areas. A differentiated early warning threshold matrix is ​​called for spatiotemporal cross-validation. Only when the time period of grid anomaly occurrence matches the risk period predicted by the spatiotemporal feature network of equipment operation status is it determined as a valid risk and an early warning is triggered. S4: Store site topology data. When an effective warning is triggered, combine the current environmental parameters to simulate the fire spread trajectory, generate a two-dimensional dynamic risk map with risk level, propagation path and fire resource location marked, push warning information to relevant terminals and generate emergency response plan.

2. The fire intelligent early warning method based on the Internet of Things according to claim 1, characterized in that: S1 includes: S1.1: IoT detectors deployed in various areas of the site collect environmental data in real time according to a preset sampling frequency. Temperature detectors, smoke concentration detectors, and CO concentration detectors record corresponding parameter values ​​according to their respective set collection cycles. All detectors transmit the collected data to the data processing center. The received raw monitoring data is preprocessed, and instantaneous pulse noise is removed by filtering algorithms. The values ​​of different types of detectors are normalized to convert the temperature, smoke concentration, and CO concentration to a preset value range. S1.2: The data processing center synchronously receives the basic information set I={I1,I2,I3} uploaded by the site management system, where I1 is the site location information, I2 is the regional functional division result, and I3 is the flammability distribution data of items; it performs coordinate mapping transformation on the site location information I1 to generate a location matrix P={P1,P2,…,Pm} in the standard coordinate system, where Pm represents the coordinate vector of the m-th region; it associates the regional functional division result I2 with the location matrix P to construct a functional location mapping table Mf={<P1,F1> ,<P2,F2> ,…,<Pm,Fm>}, where Fm represents the functional type of the m-th region; S1.3: According to the preset flammability rating classification rule R={R1,R2,R3}, where R1 is highly flammable, R2 is moderately flammable, and R3 is lowly flammable, the flammability distribution data I3 of the items is classified and processed to generate a flammability rating matrix H={H1,H2,…,Hm}, where Hm represents the flammability rating of the items in the m-th region; the flammability rating matrix H is merged with the functional location mapping table Mf, and the composite attribute label T={ for each region is marked on the electronic map.<P1,F1,H1> ,<P2,F2,H2> ,...,<Pm,Fm,Hm>}; Establish an index structure Idx that links monitoring data with basic information.<D1,T1> ,<D2,T2> ,...,<Dn,Tn>}, where Dn represents the real-time data set of the nth monitoring point, and Tn is the composite attribute label corresponding to the monitoring point; through the spatial neighborhood analysis algorithm, the nearest neighbor regional attribute label is matched for each monitoring point. When the monitoring point is located at the regional boundary, the weighted fusion method is used to calculate its comprehensive attribute value, which is the fusion result of the functional type weight and the flammability level weight.

3. The fire intelligent early warning method based on the Internet of Things according to claim 2, characterized in that: S2 includes: S2.1: Construct a spatiotemporal feature network Gs=(Vs,Es) for equipment operation status, where the node set Vs={vs1,vs2,...,vsn} represents the equipment status vector at different time periods, and vsn represents the comprehensive operation parameter vector of the core production equipment, environmental control equipment, and fire-fighting equipment in the site during the nth time period; the edge set Es represents the temporal correlation between time periods, and the weight of the edge reflects the correlation strength of the equipment status between adjacent time periods; perform sliding window segmentation on the historical equipment operation data, and extract multidimensional features in each window, including temperature features, fluctuation features, peak features, and temporal features; input the extracted original feature vectors into an autoencoder network for feature dimensionality reduction and reconstruction to generate a low-dimensional spatiotemporal feature vector STV={stv1,stv2,...,stvj}, where stvj represents the core feature vector of the jth time window; S2.2: Based on the composite attribute label T, construct the risk propagation impact graph IG=(N,A), where the node set N represents different areas within the site, the edge set A represents the risk transmission path between areas, and the edge weights represent the risk transmission coefficients between areas. The risk transmission coefficients are calculated from the area function type F, the flammability level of the item H, and the spatial distance attenuation factor. Using a graph attention mechanism, calculate the risk impact weight matrix W of each area on the surrounding areas. Step S2.3: Design a dynamic threshold generation neural network. Its input layer includes the spatio-temporal feature vector STV, the risk impact weight matrix W, and the environmental parameter vector E(t) of the current period. The hidden layer captures the temporal dependence relationship through a gated recurrent unit and combines the attention mechanism to perform weighted fusion on the input features. The output layer generates a differential warning threshold matrix Th(t) = {Th1(t), Th2(t),..., Thm(t)}; where Thm(t) represents the warning threshold of the m-th region within the t period, and the calculation formula is: Thm(t) = BaseThm × f(stvj, Wmj, E(t)); where BaseThm is the benchmark threshold, Wmj represents the risk impact weight of the m-th region on the surrounding regions; f is a non-linear mapping function obtained by training with historical data, and the weight coefficients of spatio-temporal features, risk propagation, and environmental parameters are automatically allocated through the attention mechanism; when a certain region does not trigger a warning for multiple consecutive periods but a fire actually occurs, the model parameters are adjusted through the inverse reinforcement learning algorithm to increase the risk weight of this region in this period.

4. The fire intelligent early warning method based on the Internet of Things according to claim 3, characterized in that: S3 includes: S3.1: Based on the composite attribute label T, calculate the correlation index of the functional type and the flammability level within each region: Assign feature values to the functional type and flammability level of each region. The functional type feature value is denoted as Fvg, and the flammability level feature value is denoted as Hvg. Through the formula: Rg = |Fvg × Hvg| / (1 + |Fvg - Hvg|); calculate the correlation index Rg, and the value range of Rg is [0, 1]; Divide the site into initial uniform units U0 = {u01, u02,..., u0g}, where u0g represents the g-th initial uniform unit, and each unit corresponds to a unique spatial range. Calculate the mean value R0g of the correlation index of all regions within each initial unit; Set the information entropy threshold Et, calculate the information entropy E of the current unit set E = -Σ(Rg × log2Rg). If E > Et, split the unit with the R0g difference exceeding the threshold ΔR into two sub-units and recalculate the mean value of the correlation index of the sub-units; if E < Et, merge the adjacent units with the R0g difference less than ΔR, and repeat the splitting and merging operations until the information entropy E converges to the interval [Et - ΔE, Et + ΔE]; Finally, form a grid set that satisfies the grid area Sg being inversely proportional to the mean value of the correlation index Rg: Sg = S0 / Rg, where S0 is the benchmark area; at the same time, set the constraint conditions: for the high-correlation region with Rg ≥ θ1, its grid area does not exceed λ1 times the benchmark area; for the low-correlation region with Rg ≤ θ2, the grid area is not less than λ2 times the benchmark area; where θ1 and θ2 are the set high-correlation threshold and low-correlation threshold respectively, and θ1 > θ2; λ1 and λ2 are the high-correlation region area coefficient and low-correlation region area coefficient respectively, and λ1 < 1, λ2 > 1.

5. A fire intelligent early warning method based on the Internet of Things according to claim 4, characterized in that: S3 also includes: S3.2: Configure an independent data acquisition point group for each grid. The acquisition point group includes three types of monitoring points: temperature, smoke concentration, and CO concentration. The spatial deviation between each monitoring point and its grid does not exceed 1 / 5 of the grid's side length. Construct a three-dimensional time series array Dg(t)={Tg(t),Yg(t),Cg(t)} for the real-time monitoring data of each grid, where Tg(t) is the temperature time series, Yg(t) is the smoke concentration time series, and Cg(t) is the CO concentration time series. Calculate the dynamic baseline values ​​of each parameter. Based on the grid's historical normal data, generate the baseline values ​​Tg0(t), Yg0(t), and Cg0(t) for each moment using a moving average algorithm. Extract abnormal feature indicators, including parameter deviation, acceleration of change, and parameter synergy. Generate an abnormal comprehensive value Ag by weighted fusion of abnormal feature indicators. The weight coefficients are dynamically adjusted according to the grid correlation index Rg. The higher the Rg, the greater the weight of the temperature parameter. When Ag exceeds the abnormal threshold corresponding to the grid type for x consecutive sampling periods, mark the grid as a suspected risk area.

6. A fire intelligent early warning method based on the Internet of Things according to claim 5, characterized in that: S3 further includes: S3.3 performs spatiotemporal cross-validation: Time dimension verification: Extract the anomaly start time t0 of the suspected risk area, combine it with the differentiated early warning threshold matrix Th(t), obtain the threshold Th(t0) of the grid in the time period to which t0 belongs, calculate the ratio of the anomaly comprehensive value Ag to Th(t0), and if the ratio is greater than or equal to the set threshold, the time verification is passed; Spatial dimension verification: Analyze the abnormal correlation between the suspected risk area and the three adjacent grids, calculate the spatial correlation index, which is the ratio of the time difference of the occurrence of anomalies in adjacent grids to the spatial distance. When the correlation index of at least two adjacent grids is less than or equal to the preset propagation coefficient, the spatial verification is passed. Only when both time and space dimensions are verified simultaneously is it identified as a valid risk area and an early warning is triggered; otherwise, a secondary verification is performed, retrieving the device operation logs and detector status data of the grid to rule out anomalies caused by device malfunctions or environmental interference.

7. A fire intelligent early warning method based on the Internet of Things according to claim 6, characterized in that: S4 includes: S4.1: Pre-store the site's topological structure data and establish an association mapping with the composite attribute label T. The topological structure data specifically includes: Building infrastructure data: wall locations, floor structure, and fire compartment boundaries in the building plan, linked to the coordinate vectors of the location matrix P; Pipeline network data: ventilation duct routing, vent locations, and fan parameters, linked to the area function type F; Fire protection facility data: fire door locations, sprinkler head distribution, and fire extinguisher storage locations, bound to the flammability rating H of the corresponding area; The above data is integrated into a topological relationship set Topo={ using a spatial association algorithm.<P1,F1,H1,Str1,Pipe1,Fac1> ,...,<Pm,Fm,Hm,Strm,Pipem,Facm>}, where Strm, Pipem, and Facm represent the building structure, pipeline network, and fire protection facilities characteristics of the m-th region, respectively; S4.2: When S3 triggers an effective warning, the associated data of the risk area in the topology set Topo is called, and the fire spread simulation is started in combination with the current environmental parameters: Taking the grid coordinates where the warning is triggered as the starting point, the composite attribute label Tm and topological features of the area are extracted; the combustion rate coefficient is determined based on Hm, and the heat diffusion rate is calculated in combination with the current environmental parameters; the smoke propagation path is corrected by ventilation parameters, and the propagation blocking effect is judged according to the current status of the fire door; the simulation results generate a spread time series matrix M(t), which records the expected temperature, smoke concentration and risk level of each area at time t. The risk level is determined by fusing the personnel density coefficient of Fm and the combustion intensity coefficient of Hm. S4.3: A two-dimensional risk map is generated based on the propagation time series matrix M(t). The map layer is aligned with the virtual monitoring grid divided in S3.1 to achieve spatial correlation between the warning area and the simulation results. The map annotations include: Core risk area: corresponding to the effective risk grid marked in S3, with the current temperature, smoke concentration and ignition time marked; Propagation path zone: according to the propagation direction of each time period in M(t), the fire spread trajectory along the pipes and channels is marked, as well as the corresponding propagation speed; Dynamic risk markers: each grid is marked with the expected time window and risk level of being affected; Fire resource points: the location of the fire-fighting facilities closest to the risk area is extracted from the Topo, and their type and availability status are marked; Finally, the above risk map information is integrated to generate an emergency response plan and push the warning information to relevant terminals.

8. A fire intelligent early warning system based on the Internet of Things, characterized in that: The system includes a data fusion processing module, a dynamic risk analysis module, a grid monitoring and verification module, and a fire response module; The data fusion processing module acquires real-time monitoring data on temperature, smoke concentration, and CO concentration through IoT detectors distributed throughout the site; it simultaneously collects basic site information, including site location, functional division of areas, and distribution of flammability of items; it preprocesses the collected raw data; and it fuses and generates composite attribute tags containing spatial coordinates, functional zones, and flammability levels. The dynamic risk analysis module establishes a spatiotemporal feature network of equipment operation status, inputs historical operation data of core production equipment, environmental control equipment and fire-fighting equipment in the site into an autoencoder network to extract multi-dimensional temporal features, and constructs a regional risk propagation impact map by combining composite attribute labels, and calculates the risk transmission coefficient between different areas through a graph attention mechanism. By integrating multi-dimensional time-series features, risk transmission coefficients, and real-time environmental parameters, a differentiated early warning threshold matrix that dynamically changes with time and regional characteristics is generated. The grid monitoring and verification module assigns feature values ​​and calculates correlation indices based on composite attribute tags to regional functional types and flammability levels of items, thereby dividing virtual monitoring grids. It configures independent monitoring point groups for each grid, constructs a three-dimensional time-series data array, calculates comprehensive anomaly values, identifies and marks suspected risk areas, and calls a differentiated early warning threshold matrix for spatiotemporal cross-verification. Only when the time period of a grid anomaly matches the risk period predicted by the spatiotemporal feature network of equipment operating status is it determined to be a valid risk and an early warning is triggered. The fire response module stores site topology data. When an effective warning is triggered, it simulates the fire spread trajectory by combining the current environmental parameters, generates a two-dimensional dynamic risk map that marks the risk level, propagation path and location of fire-fighting resources, pushes the warning information to relevant terminals and generates an emergency response plan.

Images