Fire identification method and system based on smoke analysis algorithm

Abstract

The invention discloses a fire identification method and system based on a smoke analysis algorithm, belongs to the technical field of forest fire identification, and is used for solving the problems that the existing forest fire identification algorithm is easily interfered by multiple aspects in a forest complex environment, so that the misjudgment rate is increased, and the forest fire identification efficiency is increased. And a potential fire point and a fire spreading trend cannot be quickly locked when a fire occurs. Comprising the steps of collecting multi-dimensional monitoring data of a monitoring forest region; performing cell division on the monitoring forest region based on the multi-dimensional monitoring data, and generating a dynamic smoke threshold value of each cell; constructing a cell dynamic association library according to the multi-dimensional monitoring data of each cell; performing smoke identification on the multi-dimensional monitoring data, and determining a target cell where a fire occurs based on a dynamic smoke threshold value; performing spreading dynamic simulation on the multi-dimensional monitoring data of the target cell to obtain spreading simulation data; and obtaining an associated cell of the target cell in the cell dynamic association library, determining the associated cell as a potential fire area, and generating a thermodynamic diagram of the potential fire area.

Description

Technical Field

[0001] The present invention relates to the technical field of forest fire recognition, and particularly to a fire recognition method and system based on a smoke analysis algorithm. Background Art

[0002] Forest fires are natural or man-made phenomena occurring in natural ecosystems. The large quantity of combustibles and the changeable environment in forests result in the characteristics of forest fires, such as strong suddenness, great destructiveness, and difficulty in extinguishing. With the development of remote sensing technology and deep learning technology, more and more related technologies have been applied to forest fire prevention and control. However, the current related technologies have certain limitations and the following key technical bottlenecks exist:

[0003] 1. Traditional forest smoke analysis algorithms identify fires by analyzing the scattering characteristics of smoke or absorption peaks at specific wavelengths. However, in the forest environment, the spectral characteristics of fog, morning dew reflection, and fire smoke highly overlap, and it is difficult for traditional algorithms to accurately distinguish them through texture or color. The water film on the leaf surface produces specular reflection under low-angle sunlight irradiation, and its polarization characteristics are similar to the scattered light of smoke particles, which is also easily misjudged as a fire point by the imaging system.

[0004] 2. In the forest environment, certain substances may spontaneously combust and cause fires under specific conditions. In this case, if a fire occurs in the forest, there may be more than one potential ignition point at the same time. Traditional fire recognition algorithms only identify and detect the positions where smoke has already appeared in the forest, and trigger an alarm if a fire is detected. However, they cannot judge whether there are other potential ignition points based on the fire position and environmental information, and send the inference result to the fire department in time. Moreover, the fire spreads very quickly in the forest. If the spread path and speed of the fire cannot be calculated in advance, the fire handling efficiency will be greatly reduced.

[0005] In summary, the current forest fire recognition algorithms are vulnerable to various interferences in the complex forest environment, resulting in an increased misjudgment rate, and unable to quickly lock in potential ignition points and the fire spread trend when a fire occurs, leading to a reduction in fire handling efficiency. Summary of the Invention

[0006] The embodiments of the present invention provide a fire recognition method and system based on a smoke analysis algorithm to solve the following technical problems: The current forest fire recognition algorithms are vulnerable to various interferences in the complex forest environment, resulting in an increased misjudgment rate, and unable to quickly lock in potential ignition points and the fire spread trend when a fire occurs, leading to a reduction in fire handling efficiency.

[0007] The embodiments of the present invention adopt the following technical solutions:

[0008] On the one hand, an embodiment of the present invention provides a fire recognition method based on a smoke analysis algorithm, the method comprising: collecting multi-dimensional monitoring data of a monitored forest area through a multi-spectral polarization imaging system; wherein the multi-dimensional monitoring data at least includes forest three-dimensional point cloud data, environmental parameters, forest polarization images, and forest multi-spectral images;

[0009] Based on the multi-dimensional monitoring data, dividing the monitored forest area into small areas, and generating a dynamic smoke threshold for each small area;

[0010] Construct a dynamic association library for each small area according to the multi-dimensional monitoring data of each small area;

[0011] Construct a two-stream network architecture, perform smoke recognition on the multi-dimensional monitoring data, determine the target small area where a fire occurs based on the dynamic smoke threshold, and simultaneously send an alarm message to a monitoring terminal;

[0012] Construct a multi-physical field coupling model, and perform a spread dynamic simulation on the multi-dimensional monitoring data of the target small area to obtain spread simulation data;

[0013] Obtain the associated small areas of the target small area in the dynamic association library of the small areas, determine the associated small areas as potential fire areas, and generate a heat map of the potential fire areas;

[0014] Send the simulated spread path data and the heat map of the fire area to the monitoring terminal simultaneously.

[0015] In a feasible implementation manner, collecting the multi-dimensional monitoring data of the monitored forest area through a multi-spectral polarization imaging system specifically includes:

[0016] Deploy a lidar and an environmental detector in the monitored forest area to respectively obtain the forest three-dimensional point cloud data and the environmental parameters;

[0017] Collect the forest polarization images through a polarization camera carried by a drone; meanwhile, collect the forest multi-spectral images through a short-wave infrared camera and a long-wave infrared camera carried by the drone;

[0018] The lidar, the environmental detector, the polarization camera, the short-wave infrared camera, and the long-wave infrared camera constitute the multi-spectral polarization imaging system;

[0019] Perform point cloud filtering, polarization image correction, and multi-spectral radiation correction on the collected various data respectively to obtain the multi-dimensional monitoring data.

[0020] In a feasible implementation manner, dividing the monitored forest area into small areas based on the multi-dimensional monitoring data specifically includes:

[0021] Randomly select a number of target points in the monitored forest area, and extract the fire impact factors of each target point from the environmental parameters; wherein, the fire impact factors at least include: terrain data, vegetation data, meteorological data, and lighting data;

[0022] Select K target points from the target points as the initial clustering centers, and perform clustering analysis according to the fire impact factors to obtain K clustering clusters;

[0023] Take the connection line of the outermost target points in each clustering cluster as the cell boundary line to divide the monitored forest area into cells.

[0024] In a feasible implementation manner, generating the dynamic smoke threshold of each cell specifically includes:

[0025] Construct a lightweight LSTM network and train it based on a training set composed of environmental parameters of different forest environments to obtain a dynamic smoke threshold prediction model; wherein, the environmental parameters at least include acquisition time, temperature, humidity, fog concentration, and wind force;

[0026] Input the environmental parameters of each cell in the monitored forest area into the dynamic smoke threshold prediction model to predict the current background smoke baseline of each cell, that is, the dynamic smoke threshold.

[0027] In a feasible implementation manner, constructing a cell dynamic association library according to the multi-dimensional monitoring data of each cell specifically includes:

[0028] Calculate the fire risk index of each cell at different time periods according to the environmental parameters and historical fire data of each cell;

[0029] Calculate the Euclidean distance of the fire risk indexes of each cell in the same time period, determine the associated cells of each cell whose Euclidean distance from each cell is less than a preset threshold, and construct an associated cell table for each cell in each time period to obtain the cell dynamic association library.

[0030] In a feasible implementation manner, calculating the fire risk index of each cell at different time periods according to the environmental parameters and historical fire data of each cell specifically includes:

[0031] Extract the fire impact factors from the environmental parameters collected in each cell at each time period; wherein, the fire impact factors at least include: terrain data, vegetation data, meteorological data, and lighting data;

[0032] Set the fire influence range according to expert experience, and determine the terrain risk value, vegetation flammability risk value, historical fire frequency of each community, as well as the meteorological driving risk value and human activity index of each time period according to the fire influence range where the specific value of the fire influence factor is located.

[0033] Based on the preset importance weights, perform weighted calculations on the terrain risk value, vegetation flammability risk value, historical fire frequency, as well as the meteorological driving risk value and human activity index of each time period to obtain the fire risk index of each community for each time period.

[0034] In a feasible implementation manner, construct a two-stream network architecture to perform smoke recognition on the multi-dimensional monitoring data, and determine the target community where a fire occurs based on the dynamic smoke threshold, specifically including:

[0035] Construct a spatial stream neural network based on the YOLOv8 network and train it through a dataset composed of forest polarization images and forest multi-spectral images.

[0036] Construct a spatio-temporal stream neural network based on a 3D convolutional network and train the spatio-temporal stream neural network after voxelizing the forest three-dimensional point cloud dataset.

[0037] Connect the output ends of the geometric stream neural network and the spatio-temporal stream neural network to a cross-attention module respectively to form the two-stream neural network architecture.

[0038] Input the multi-dimensional monitoring data of each community collected in real time into the two-stream neural network architecture for smoke recognition to obtain the current smoke concentration of each community.

[0039] If the current smoke concentration of the current community is higher than the dynamic smoke threshold of this community for the current time period, then determine the current community as the target community where a fire occurs.

[0040] In a feasible implementation manner, construct a multi-physical field coupling model and perform a spread dynamic simulation on the multi-dimensional monitoring data of the target community to obtain spread simulation data, specifically including:

[0041] Establish a three-dimensional coupling equation of the combustion kinetics equation, fluid mechanics equation, and heat conduction equation to obtain the multi-physical field coupling model.

[0042] Input the multi-dimensional monitoring data of the target community into the multi-physical field coupling model for spread path simulation to obtain the spread simulation path.

[0043] Based on the Rothermel spread model, calculate the spread speed of the target community.

[0044] Combine the spreading simulation path and the spreading speed into the spreading simulation data.

[0045] In a feasible implementation manner, obtain the associated cells of the target cell in the cell dynamic association library, determine the associated cells as potential fire areas, and generate a thermal map of the potential fire areas, specifically including:

[0046] Based on the current time period, search for the associated cells of the target cell in the cell dynamic association library and determine all of them as potential fire areas;

[0047] Obtain the multi-dimensional monitoring data of the potential fire areas, perform multi-spectral analysis on the forest multi-spectral images among them, and generate a spectral thermal map;

[0048] Identify and label the spectral anomaly areas in the spectral thermal map;

[0049] Determine the labeled spectral thermal map as the thermal map of the potential fire areas.

[0050] On the other hand, an embodiment of the present invention also provides a fire recognition system based on a smoke analysis algorithm, and the system includes:

[0051] A multi-dimensional data monitoring module, configured to collect multi-dimensional monitoring data of a monitored forest area through a multi-spectral polarization imaging system; wherein, the multi-dimensional monitoring data at least includes forest three-dimensional point cloud data, environmental parameters, forest polarization images, and forest multi-spectral images;

[0052] A dynamic maintenance module, configured to divide the monitored forest area into cells based on the multi-dimensional monitoring data, and generate a dynamic smoke threshold for each cell; construct a cell dynamic association library according to the multi-dimensional monitoring data of each cell;

[0053] A fire recognition module, configured to construct a two-stream network architecture, perform smoke recognition on the multi-dimensional monitoring data, and determine the target cell where a fire occurs based on the dynamic smoke threshold; construct a multi-physical field coupling model, and perform a spreading dynamic simulation on the multi-dimensional monitoring data of the target cell to obtain spreading simulation data;

[0054] A potential fire area recognition module, configured to obtain the associated cells of the target cell in the cell dynamic association library, determine the associated cells as potential fire areas, and generate a thermal map of the potential fire areas;

[0055] A fire data reporting module, configured to simultaneously send the simulated spreading path data, the thermal map of the fire area, and an alarm message to a monitoring terminal.

[0056] Compared with the prior art, a fire recognition method and system based on a smoke analysis algorithm provided by an embodiment of the present invention have the following beneficial effects:

[0057] First, the present invention integrates forest three-dimensional point cloud data, environmental parameters, polarization images, and multi-spectral images through a multi-spectral polarization imaging system to construct a full-element forest environment perception system. This system breaks through the limitations of traditional single-parameter detection, improves the accuracy of smoke recognition, and reduces the interference misjudgment rate of fog / dew.

[0058] The present invention also proposes a dynamic smoke threshold generation method based on cell division, establishes a detection benchmark adaptable to environmental changes, and significantly improves the detection sensitivity compared with the fixed threshold algorithm. The innovatively constructed dual-stream network architecture increases the extraction dimension of smoke features and shortens the detection response time.

[0059] The present invention shifts fire recognition from passive monitoring to active prediction, establishes a digital simulation system for fire spread, anticipates potential ignition points in advance for early fire prevention and control, forms an all-round monitoring and prevention network integrating space, air, and ground, and significantly improves the digital prevention and control capabilities of large-scale forests. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0061] Figure 1 is a flowchart of a fire recognition method based on a smoke analysis algorithm provided by an embodiment of the present invention;

[0062] Figure 2 is a schematic structural diagram of a fire recognition system based on a smoke analysis algorithm provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0064] An embodiment of the present invention provides a fire recognition method based on a smoke analysis algorithm, as Figure 1 shown, the fire recognition method based on the smoke analysis algorithm specifically includes steps S101 - S106:

[0065] S101. Collect multi-dimensional monitoring data of the monitored forest area through a multi-spectral polarization imaging system. Among them, the multi-dimensional monitoring data at least includes forest three-dimensional point cloud data, environmental parameters, forest polarization images, and forest multi-spectral images.

[0066] Specifically, deploy lidar and environmental detectors in the monitored forest area to obtain forest three-dimensional point cloud data and environmental parameters respectively.

[0067] In addition, carry a polarization camera by a drone to collect forest polarization images; also carry a short-wave infrared camera and a long-wave infrared camera by the drone to collect forest multi-spectral images. The drone conducts regular inspections over the monitored forest area and sends the collected images to the edge computing terminal for subsequent processing at regular intervals.

[0068] The above-mentioned lidar, environmental detector, polarization camera, short-wave infrared camera, and long-wave infrared camera constitute a multi-spectral polarization imaging system.

[0069] Further, after collecting various data, perform point cloud filtering on the collected forest three-dimensional point cloud data, polarization image correction on the forest polarization images, and multi-spectral radiation correction on the forest multi-spectral images to obtain multi-dimensional monitoring data.

[0070] S102. Based on the multi-dimensional monitoring data, divide the monitored forest area into small areas and generate a dynamic smoke threshold for each small area.

[0071] Specifically, randomly select several target points in the monitored forest area and extract the fire impact factors of each target point from the environmental parameters. Among them, the fire impact factors at least include: terrain data, vegetation data, meteorological data, and lighting data.

[0072] In one embodiment, the terrain data includes data such as slope, aspect, and elevation, the vegetation data includes data such as vegetation moisture content, tree species, and canopy density, the meteorological data includes data such as temperature, humidity, wind speed, and wind direction, and the lighting data includes data such as light intensity and ultraviolet intensity.

[0073] Further, select K target points as the initial clustering centers among the target points and perform clustering analysis according to the fire impact factors to obtain K clustering clusters. Use the connection of the outermost target points in each clustering cluster as the small area boundary line to divide the monitored forest area into small areas.

[0074] Further, construct a lightweight LSTM network and train it based on a training set composed of environmental parameters in different forest environments to obtain a dynamic smoke threshold prediction model. Among them, the environmental parameters at least include acquisition time, temperature, humidity, fog concentration, and wind force.

[0075] Then, input the environmental parameters of each sub-region in the monitored forest area into the dynamic smoke threshold prediction model to predict the current background smoke baseline of each sub-region, that is, the dynamic smoke threshold.

[0076] S103. Construct a dynamic association library of sub-regions according to the multi-dimensional monitoring data of each sub-region.

[0077] Specifically, calculate the fire risk index of each sub-region at different time periods according to the environmental parameters and historical fire data of each sub-region.

[0078] Then calculate the Euclidean distance of the fire risk indices of each sub-region in the same time period, determine the associated sub-regions of each sub-region whose Euclidean distance from each sub-region is less than the preset threshold, and construct an associated sub-region table for each sub-region in each time period to obtain the dynamic association library of sub-regions.

[0079] As a feasible implementation method, calculate the fire risk index of each sub-region at different time periods according to the environmental parameters and historical fire data of each sub-region. The specific implementation method is as follows:

[0080] Extract fire impact factors from the environmental parameters collected in each sub-region at each time period; among them, the fire impact factors at least include: terrain data, vegetation data, meteorological data, and illumination data.

[0081] Then set the fire impact interval according to expert experience, that is, the influence value interval of each fire impact factor on the fire. If each fire impact factor falls into a certain interval, then the value of the impact factor is the target value corresponding to the interval, and the target value corresponding to each interval is set according to expert experience. According to the above method, determine the terrain risk value, vegetation flammability risk value, historical fire frequency of each sub-region, as well as the meteorological driving risk value and human activity index of each time period according to the fire impact interval where the specific value of the fire impact factor is located.

[0082] Finally, based on the preset importance weights, perform weighted calculation on the terrain risk value, vegetation flammability risk value, historical fire frequency, as well as the meteorological driving risk value and human activity index of each time period to obtain the fire risk index of each sub-region at each time period.

[0083] In one embodiment, the fire risk index = 0.35 × terrain risk value + 0.25 × vegetation flammability risk value + 0.2 × meteorological driving risk value + 0.15 × human activity index + 0.05 × historical fire frequency.

[0084] S104. Construct a two-stream network architecture, perform smoke recognition on the multi-dimensional monitoring data, determine the target sub-region where a fire occurs based on the dynamic smoke threshold, and send an alarm message to the monitoring terminal at the same time.

[0085] Specifically, a spatial flow neural network is constructed based on the YOLOv8 network and trained with a dataset composed of forest polarization images and forest multispectral images. A spatio-temporal flow neural network is constructed based on a 3D convolutional network and trained after voxelizing the forest three-dimensional point cloud dataset. The output ends of the geometric flow neural network and the spatio-temporal flow neural network are respectively connected to a cross-attention module to form a two-stream neural network architecture.

[0086] Furthermore, the multi-dimensional monitoring data of each community collected in real time is input into the two-stream neural network architecture for smoke recognition to obtain the current smoke concentration of each community. If the current smoke concentration of the current community is higher than the dynamic smoke threshold of the current time period of this community, the current community is determined as the target community where a fire has occurred. And an alarm message is immediately sent to the monitoring terminal.

[0087] S105. Construct a multi-physical-field coupling model and perform a spread dynamic simulation on the multi-dimensional monitoring data of the target community to obtain spread simulation data.

[0088] Specifically, a three-dimensional coupling equation of the combustion kinetics equation, the fluid mechanics equation, and the heat conduction equation is established to obtain the multi-physical-field coupling model.

[0089] Furthermore, the multi-dimensional monitoring data of the target community is input into the multi-physical-field coupling model for a spread path simulation to obtain a spread simulation path. Then, based on the Rothermel spread model, the spread speed of the target community is calculated. Finally, the spread simulation path and the spread speed are combined into spread simulation data.

[0090] S106. Obtain the associated communities of the target community in the community dynamic association library, determine the associated communities as potential fire areas, and generate a heat map of the potential fire areas; send the simulated spread path data and the heat map of the fire area to the monitoring terminal simultaneously.

[0091] Specifically, based on the current time period, search for the associated communities of the target community in the community dynamic association library and determine all of them as potential fire areas. Obtain the multi-dimensional monitoring data of the potential fire areas, and perform multi-spectral analysis on the forest multispectral images among them to generate a spectral heat map.

[0092] Furthermore, identify and label the spectral anomaly areas in the spectral heat map; determine the labeled spectral heat map as the heat map of the potential fire areas.

[0093] Finally, send the simulated spread path data and the heat map of the fire area to the monitoring terminal simultaneously for the reference and decision-making of firefighters.

[0094] In addition, an embodiment of the present invention also provides a fire recognition system based on a smoke analysis algorithm, as Figure 2As shown in the figure, the fire recognition system 200 based on the smoke analysis algorithm specifically includes:

[0095] A multi-dimensional data monitoring module 210, configured to collect multi-dimensional monitoring data of the monitored forest area through a multi-spectral polarization imaging system; wherein, the multi-dimensional monitoring data at least includes forest three-dimensional point cloud data, environmental parameters, forest polarization images, and forest multi-spectral images;

[0096] A dynamic maintenance module 220, configured to divide the monitored forest area into small areas based on the multi-dimensional monitoring data, and generate a dynamic smoke threshold for each small area; construct a dynamic association library for each small area according to the multi-dimensional monitoring data of each small area;

[0097] A fire recognition module 230, configured to construct a two-stream network architecture, perform smoke recognition on the multi-dimensional monitoring data, and determine the target small area where a fire occurs based on the dynamic smoke threshold; construct a multi-physical field coupling model, and perform a spread dynamic simulation on the multi-dimensional monitoring data of the target small area to obtain spread simulation data;

[0098] A potential fire area recognition module 240, configured to obtain the associated small areas of the target small area in the dynamic association library of the small areas, determine the associated small areas as potential fire areas, and generate a thermal map of the potential fire areas;

[0099] A fire data reporting module 250, configured to simultaneously send the simulated spread path data, the thermal map of the fire area, and an alarm message to a monitoring terminal.

[0100] Each embodiment in the present invention is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the embodiments of the device, equipment, and non-volatile computer storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0101] The specific embodiments of the present invention are described above. Additionally, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In certain embodiments, multi-tasking and parallel processing are also possible or may be advantageous.

[0102] The above are only the embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the embodiments of the present invention shall be included within the protection scope of the present invention.

Claims

1. A fire recognition method based on a smoke analysis algorithm, characterized in that, The method includes: Collecting multi-dimensional monitoring data of the monitored forest area through a multi-spectral polarization imaging system; wherein, the multi-dimensional monitoring data at least includes forest three-dimensional point cloud data, environmental parameters, forest polarization images, and forest multi-spectral images; Based on the multi-dimensional monitoring data, dividing the monitored forest area into small areas and generating a dynamic smoke threshold for each small area; Constructing a small area dynamic association library according to the multi-dimensional monitoring data of each small area; Constructing a two-stream network architecture to identify smoke from the multi-dimensional monitoring data, determining the target small area where a fire occurs based on the dynamic smoke threshold, and sending an alarm message to the monitoring terminal at the same time; Constructing a multi-physical field coupling model and performing a spread dynamic simulation on the multi-dimensional monitoring data of the target small area to obtain spread simulation data; Obtaining the associated small areas of the target small area in the small area dynamic association library, determining the associated small areas as potential fire areas, and generating a heat map of the potential fire areas; Sending the simulated spread path data and the heat map of the fire area to the monitoring terminal at the same time.

2. The fire recognition method based on a smoke analysis algorithm according to claim 1, characterized in that, Collecting multi-dimensional monitoring data of the monitored forest area through a multi-spectral polarization imaging system, specifically including: Deploying a lidar and an environmental detector in the monitored forest area to obtain the forest three-dimensional point cloud data and the environmental parameters respectively; Collecting the forest polarization images through a polarization camera carried by a drone; at the same time, collecting the forest multi-spectral images through a short-wave infrared camera and a long-wave infrared camera carried by the drone; The lidar, environmental detector, polarization camera, short-wave infrared camera, and long-wave infrared camera constitute the multi-spectral polarization imaging system; Performing point cloud filtering, polarization image correction, and multi-spectral radiation correction on the collected various data respectively to obtain the multi-dimensional monitoring data.

3. The fire recognition method based on a smoke analysis algorithm according to claim 1, characterized in that, Based on the multi-dimensional monitoring data, dividing the monitored forest area into small areas, specifically including: Randomly selecting several target points in the monitored forest area and extracting the fire influence factors of each target point from the environmental parameters; wherein, the fire influence factors at least include: terrain data, vegetation data, meteorological data, and lighting data; Selecting K target points as the initial clustering centers among the target points and performing clustering analysis according to the fire influence factors to obtain K clustering clusters; Taking the connection of the outermost target points in each clustering cluster as the small area boundary line to divide the monitored forest area into small areas.

4. The fire recognition method based on a smoke analysis algorithm according to claim 1, characterized in that Generating a dynamic smoke threshold for each small area, specifically including: Constructing a lightweight LSTM network and training it based on a training set composed of environmental parameters in different forest environments to obtain a dynamic smoke threshold prediction model; wherein, the environmental parameters at least include collection time, temperature, humidity, fog concentration, and wind force; Inputting the environmental parameters of each small area in the monitored forest area into the dynamic smoke threshold prediction model to predict the current background smoke baseline of each small area, that is, the dynamic smoke threshold.

5. A fire recognition method based on a smoke analysis algorithm according to claim 1, characterized in that, Constructing a small area dynamic association library according to the multi-dimensional monitoring data of each small area, specifically including: Calculating the fire risk index of each small area at different time periods according to the environmental parameters and historical fire data of each small area; Calculate the Euclidean distance of the fire risk index for each community within the same time period, determine the associated communities of each community whose Euclidean distance is less than a preset threshold, construct an associated community table for each community in each time period, and obtain the community dynamic association library.

6. The fire recognition method based on a smoke analysis algorithm according to claim 5, characterized in that, Calculate the fire risk index for each community at different time periods based on the environmental parameters and historical fire data of each community, specifically including: Extract the fire impact factors from the environmental parameters collected in each community at each time period; among them, the fire impact factors at least include: terrain data, vegetation data, meteorological data, and lighting data; Set the fire impact interval according to expert experience, and determine the terrain risk value, vegetation flammability risk value, historical fire frequency, meteorological driving risk value, and human activity index for each community at each time period according to the fire impact interval where the specific value of the fire impact factor is located; Based on the preset importance weights, perform weighted calculations on the terrain risk value, vegetation flammability risk value, historical fire frequency, meteorological driving risk value, and human activity index for each time period to obtain the fire risk index for each community at each time period.

7. A fire recognition method based on a smoke analysis algorithm according to claim 1, characterized in that, Construct a two-stream network architecture to identify smoke from the multi-dimensional monitoring data, and determine the target community where a fire occurs based on the dynamic smoke threshold, specifically including: Construct a spatial stream neural network based on the YOLOv8 network and train it with a dataset composed of forest polarization images and forest multi-spectral images; Construct a spatio-temporal stream neural network based on a 3D convolutional network and train the spatio-temporal stream neural network after voxelizing the forest three-dimensional point cloud dataset; Connect the output ends of the geometric stream neural network and the spatio-temporal stream neural network to a cross-attention module respectively to form the two-stream neural network architecture; Input the multi-dimensional monitoring data of each community collected in real time into the two-stream neural network architecture for smoke identification to obtain the current smoke concentration of each community; If the current smoke concentration of the current community is higher than the dynamic smoke threshold of the current time period of this community, then determine the current community as the target community where a fire occurs.

8. The fire recognition method based on a smoke analysis algorithm according to claim 1, wherein, Construct a multi-physical field coupling model and perform a spread dynamic simulation on the multi-dimensional monitoring data of the target community to obtain spread simulation data, specifically including: Establish a three-dimensional coupling equation of the combustion kinetics equation, fluid mechanics equation, and heat conduction equation to obtain the multi-physical field coupling model; Input the multi-dimensional monitoring data of the target community into the multi-physical field coupling model for spread path simulation to obtain a spread simulation path; Calculate the spread speed of the target community based on the Rothermel spread model; Combine the spread simulation path and the spread speed into the spread simulation data.

9. A fire recognition method based on a smoke analysis algorithm according to claim 1, characterized in that, Obtain the associated communities of the target community in the community dynamic association library, determine the associated communities as potential fire areas, and generate a heat map of the potential fire areas, specifically including: Based on the current time period, search for the associated communities of the target community in the community dynamic association library and determine all of them as potential fire areas; Obtain the multi-dimensional monitoring data of the potential fire area, perform multi-spectral analysis on the forest multi-spectral images therein, and generate a spectral heat map; Identify and label the spectral anomaly areas in the spectral heat map; Determine the labeled spectral heat map as the heat map of the potential fire area.

10. A fire recognition system based on a smoke analysis algorithm, characterized in that, The system includes: A multi-dimensional data monitoring module for collecting multi-dimensional monitoring data of the monitored forest area through a multi-spectral polarization imaging system; wherein, the multi-dimensional monitoring data at least includes forest three-dimensional point cloud data, environmental parameters, forest polarization images, and forest multi-spectral images; A dynamic maintenance module for dividing the monitored forest area into small areas based on the multi-dimensional monitoring data and generating a dynamic smoke threshold for each small area; constructing a small area dynamic association library according to the multi-dimensional monitoring data of each small area; A fire identification module for constructing a two-stream network architecture to identify smoke in the multi-dimensional monitoring data and determining the target small area where a fire occurs based on the dynamic smoke threshold; constructing a multi-physical field coupling model and performing a spread dynamic simulation on the multi-dimensional monitoring data of the target small area to obtain spread simulation data; A potential fire area identification module for obtaining the associated small areas of the target small area in the small area dynamic association library, determining the associated small areas as potential fire areas, and generating a heat map of the potential fire area; A fire data reporting module for simultaneously sending the simulated spread path data, the heat map of the fire area, and the alarm information to the monitoring terminal.

Images