A fire prediction, detection, adjustment and optimization method and system

Abstract

The present invention discloses a fire prediction detection adjustment optimization method and system, which relates to the field of adjustment prediction technology, and comprises the following steps: obtaining fire characteristic data of the area to be measured, performing feature extraction and splicing, and constructing a fire characteristic vector; performing initial node division in the area to be measured, dynamically adjusting the node density according to the fire characteristics and geographical location of the nodes, extracting spatiotemporal features, and constructing a multi-scale adjacency matrix; obtaining the angle between the propagation directions and the propagation time difference between the nodes, establishing a mapping model, obtaining the directional sensitivity and correcting the adjacency matrix; inputting the corrected adjacency matrix and the fire characteristic vector into a spatiotemporal graph convolutional network model, and outputting the fire risk level; the present application improves the accuracy and real-time performance of fire risk prediction by dynamically adjusting the number and density of nodes, combining the multi-scale adjacency matrix with the spatiotemporal graph convolutional network model, and solves the problems of inaccurate node division and insufficient spatiotemporal feature processing in fire prediction.

Description

Technical Field

[0001] The present invention relates to the field of regulation prediction technology, and more specifically, to a fire prediction, detection, regulation and optimization method and system. Background Art

[0002] With the acceleration of urbanization and the increasing number of high-rise buildings, the frequency and severity of fire accidents are also increasing. Traditional fire detection and warning systems often rely on data collection from single sensors, such as smoke, temperature, or humidity sensors. Although these systems can provide early warning of fires, they have certain limitations in practical application. The process of fire spread is complex and involves the dynamic changes of multiple factors, and traditional methods are often unable to fully and accurately capture this information. Therefore, the development of a system that can accurately predict and detect fires by fusing multi-source information and using dynamic adjustment and optimization strategies has become an urgent need in the fire safety field.

[0003] The development of a fire involves multiple factors, such as temperature, humidity, smoke, gas concentration, and the distribution of fire sources. Furthermore, the spread of a fire is not affected by a single factor alone, but by the intertwined influences of environmental conditions, building structure, and spatial layout. Existing fire detection systems often rely on fixed thresholds to determine whether a fire has occurred, but this approach often overlooks the dynamic and complex nature of fire occurrence and spread. For example, an abnormally high temperature or increased smoke concentration may be an early sign of a fire, but in some cases, these phenomena may also be due to other reasons, making traditional systems prone to false alarms.

[0004] To overcome the shortcomings of traditional fire detection methods, fire prediction technologies based on multimodal sensor data fusion and intelligent algorithms have become a research hotspot in recent years. Multi-source information intelligent fusion methods combine multiple data sources, such as temperature and humidity sensors, smoke sensors, gas sensors, and video surveillance, to more comprehensively capture various information about fire occurrences. Fusion algorithms integrate this multi-source data into a unified feature vector, providing accurate input for subsequent fire risk prediction. However, optimizing the combination of these sensors and adjusting their operating strategies to maximize fire prediction accuracy remains an urgent challenge.

[0005] For example, the invention patent with announcement number: CN113741258B discloses a rail transit station fire monitoring system based on the Internet of Things and its optimization method, which belongs to the field of rail transit safety technology. This optimization method models a typical rail transit station and uses the software FDS to carry out numerical simulation, analyzes the temporal and spatial variation of smoke distribution under different fire scenarios, and establishes a fire development prediction model that couples multiple factors. In view of the development characteristics of rail transit fires and smoke, the control focus of fire detection characteristic parameters is clarified, and the selection and distribution of detectors are optimized. In addition, the present invention changes the distributed alarm into a centralized monitoring alarm, adds associated detectors in places where false alarms are prone to occur in rail transit stations, establishes an associated algorithm, reduces false alarms caused by dust interference and other human factors, and improves the accuracy of the alarm.

[0006] For example, the utility model patent with announcement number CN205582179U discloses a remote forest fire monitoring system based on satellite remote sensing images, which includes a disaster monitoring module, a disaster prediction module, a central processing unit, and an alarm center. The disaster monitoring module and the disaster prediction module are both informationally connected to the central processing unit, which is in turn informationally connected to the alarm center. The remote forest fire monitoring system based on satellite remote sensing images provided by this utility model can provide targeted forest fire warnings based on real-time forest area monitoring and forest fire warning models, combined with multiple real-time conditions such as thermal images, temperature, and humidity in the forest area. At the same time, when a forest fire occurs, it can also promptly determine the location and extent of the fire, allowing firefighters to prepare for firefighting work in a targeted manner.

[0007] The above disclosed technical solutions have at least the following technical problems:

[0008] The prediction effect of traditional fire prediction methods is not ideal due to inaccurate node division and insufficient processing of spatiotemporal features.

[0009] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0010] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a fire prediction, detection, adjustment and optimization method and system. By dynamically adjusting the number and density of nodes and combining a multi-scale adjacency matrix with a spatiotemporal graph convolutional network model, the accuracy and real-time performance of fire risk prediction are effectively improved, and the problem of unsatisfactory prediction results caused by inaccurate node division and insufficient spatiotemporal feature processing in traditional fire prediction methods is solved.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] A fire prediction, detection, and adjustment optimization method comprises the following steps: obtaining fire characteristic data of a test area, performing feature extraction and splicing, and constructing a fire characteristic vector; performing initial node division on the test area, dynamically adjusting the node density based on the fire characteristics and geographic location of the nodes, extracting spatiotemporal features, and constructing a multi-scale adjacency matrix; obtaining the angle between the propagation directions and the propagation time difference between the nodes, establishing a mapping model, obtaining directional sensitivity, and correcting the adjacency matrix; and inputting the corrected adjacency matrix and fire characteristic vector into a spatiotemporal graph convolutional network model to output a fire risk level.

[0013] In a preferred embodiment, the angle between the propagation directions and the propagation time difference between nodes is obtained, a mapping model is established, and directional sensitivity is obtained, specifically as follows: the propagation inertia direction vector of each node is extracted based on the fire characteristic data, and the angle between the propagation direction and the spatial direction of the adjacent node is obtained, and the directional sensitivity is output; the characteristic change time points of the node pairs in historical fire events are obtained, the propagation time difference between the nodes is obtained, and the relationship data between the propagation time difference and the directional angle is obtained in combination with the spatial direction relationship of the nodes, and a propagation delay-sensitivity mapping model is established; the actual propagation time difference is obtained, and the directional sensitivity is output based on the mapping model.

[0014] In a preferred embodiment, the modified adjacency matrix is ​​specifically as follows: traverse all valid connection edges in the adjacency matrix, and output the directional sensitivity of the valid edges based on the propagation delay-sensitivity mapping model; obtain the weights of the valid edges in the adjacency matrix, and multiply them by the directional sensitivity to obtain the modified edge weights; and obtain the modified adjacency matrix based on the modified edge weights.

[0015] In a preferred embodiment, the fire feature data of the area to be measured is obtained, and feature extraction and splicing are performed to construct a fire feature vector, specifically as follows: historical fire data is obtained and preprocessed; feature extraction is performed on the preprocessed historical fire data; the feature data is normalized, and different types of normalized feature vectors are spliced ​​to form a complete fire feature vector; and feature dimensionality reduction is performed on the fire feature vector to obtain an initial fire feature vector.

[0016] In a preferred embodiment, the method performs initial node division on the area to be measured, dynamically adjusts the node density according to the fire characteristics and geographical location of the nodes, extracts spatiotemporal features, and constructs a multi-scale adjacency matrix, specifically as follows: obtains boundary coordinate information of the area to be measured, performs initial division on the area according to the boundary coordinate information, divides the area into equally spaced rectangular grids, and sets the initial node density; obtains node feature data of the divided nodes, and outputs an initial fire risk value according to the node feature data; dynamically adjusts the number and density of nodes according to the initial fire risk value, and extracts spatiotemporal features between nodes for the adjusted nodes; constructs adjacency matrices at different scales according to the spatiotemporal features, and fuses the adjacency matrices at different scales to form a multi-scale adjacency matrix.

[0017] In a preferred embodiment, the number and density of nodes are dynamically adjusted according to the initial fire risk value, specifically as follows: based on the initial fire risk value, determine whether to split and merge nodes; if the initial fire risk value exceeds a second threshold, perform quadtree splitting on the nodes; if the initial fire risk value is lower than a first threshold and the fire risk values ​​of adjacent nodes are all lower than the first threshold, merge several low-risk nodes into a single node.

[0018] In a preferred embodiment, the corrected adjacency matrix and fire feature vector are input into the spatiotemporal graph convolutional network model to output the fire risk level, specifically as follows: obtaining input data, processing the input data through graph convolution operation, capturing the spatial correlation between nodes based on the multi-scale adjacency matrix, and performing feature aggregation; inputting the feature-aggregated data into the time convolution layer to capture the changing trend of node features in the time dimension; dividing the historical fire data into training set, validation set and test set, and using the time sliding window method to generate time series samples; inputting the training set into the preset fire risk model for training, and during the model training process, weighting samples of different fire risk levels through the weighted cross entropy loss function; outputting the fire risk probability of each node through the trained model to form a fire risk level assessment result; if it is detected that the fire risk probability of the node exceeds the preset threshold, the real-time warning system is immediately triggered.

[0019] A system for a fire prediction, detection, and adjustment optimization method includes a feature extraction module, a node division module, and a fire risk output module, with connections between the modules; the feature extraction module is used to obtain fire feature data of the area to be tested, perform feature extraction and splicing, and construct a fire feature vector; the node division module is used to perform initial node division on the area to be tested, dynamically adjust the node density according to the fire characteristics and geographical location of the nodes, extract spatiotemporal features, and construct a multi-scale adjacency matrix; the correction module is used to obtain the angle and propagation time difference between the propagation directions of the nodes, establish a mapping model, obtain directional sensitivity, and correct the adjacency matrix; the fire risk output module is used to input the corrected adjacency matrix and fire feature vector into a spatiotemporal graph convolutional network model to output the fire risk level.

[0020] The technical effects and advantages of the fire prediction, detection, adjustment and optimization method and system of the present invention are as follows:

[0021] 1. This invention achieves refined prediction of fire risk through the comprehensive application of feature extraction, dynamic node partitioning, and spatiotemporal graph convolutional networks. The feature extraction module efficiently mines key information from historical fire data and constructs comprehensive fire feature vectors, providing a solid foundation for subsequent analysis. The dynamic node partitioning module intelligently adjusts the number and density of partitions based on the fire characteristics and geographic location of the nodes, effectively capturing the spatiotemporal distribution characteristics of fire risk and making the early warning system more sensitive and accurate.

[0022] 2. This invention fully considers the complexity and diversity of fire risk through a multi-scale adjacency matrix construction method. By fusing adjacency matrices at different scales, the system can more comprehensively capture the spatial correlations between nodes, thereby improving the accuracy of fire risk prediction. This approach not only improves the robustness of the system but also enables it to maintain stable performance in the face of complex and changing fire scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The figure is a flow chart of a fire prediction, detection, adjustment and optimization method according to the present invention.

[0024] Figure 2 This is a schematic diagram of the system structure of a fire prediction, detection, adjustment and optimization method of the present invention.

[0025] Figure 3 It is a fire risk level curve. DETAILED DESCRIPTION

[0026] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] Example 1, Figure 1 The present invention provides a fire prediction, detection, adjustment and optimization method, comprising the following steps:

[0028] S1, obtain the fire feature data of the area to be tested, perform feature extraction and splicing, and construct a fire feature vector.

[0029] In this embodiment, fire feature data of the area to be measured is obtained, and feature extraction and splicing are performed to construct a fire feature vector, as follows:

[0030] Obtain and preprocess historical fire data, deleting duplicate data, abnormal data, and fault data. Use median or mean filtering to remove noise. For small amounts of missing data, use linear interpolation or KNN interpolation to fill in missing data. For large amounts of missing data, consider using historical trend data or time series prediction models to complete the missing data. Perform time alignment to ensure that multimodal data has the same timestamp.

[0031] Performing feature extraction on the pre-processed historical fire data, wherein the feature extraction includes smoke and gas concentration features, temperature features, spatial features, image features, and frequency domain features;

[0032] In order to eliminate the dimension difference, all feature data are normalized, and the normalized feature vectors of different types are concatenated to form a complete fire feature vector;

[0033] Principal component analysis or linear discriminant analysis is used to reduce the dimension of the fire feature vector, reduce redundant data, and obtain the initial fire feature vector.

[0034] S2, performs initial node division on the tested area, dynamically adjusts the node density according to the fire characteristics and geographical location of the nodes, extracts spatiotemporal features, and constructs a multi-scale adjacency matrix.

[0035] In this embodiment, the area to be tested is initially divided into nodes, the node density is dynamically adjusted according to the fire characteristics and geographical location of the nodes, and the spatiotemporal features are extracted to construct a multi-scale adjacency matrix, as follows:

[0036] Obtain the boundary coordinate information of the area to be measured, divide the area into equal-spaced rectangular grids based on the boundary coordinate information, and set the initial node density;

[0037] Obtain node characteristic data of the nodes according to the divided nodes, and output an initial fire risk value according to the node characteristic data, wherein the node characteristic data includes smoke concentration, temperature, and harmful gas concentration;

[0038] Dynamically adjust the number and density of nodes according to the initial fire risk value;

[0039] Extracting spatiotemporal features between nodes from the adjusted nodes, wherein the spatiotemporal features include feature similarity, spatial distance, and time series changes between adjacent nodes;

[0040] Constructing adjacency matrices at different scales according to spatiotemporal features, wherein the adjacency matrices at different scales include a spatial adjacency matrix, a feature adjacency matrix, and a temporal adjacency matrix;

[0041] The adjacency matrices at different scales are fused to form a multi-scale adjacency matrix.

[0042] In this embodiment, the number and density of nodes are dynamically adjusted according to the initial fire risk value, as follows:

[0043] Determine whether to split or merge nodes based on the initial fire risk value;

[0044] When the initial fire risk value exceeds the second threshold, the node is quadtree split;

[0045] Under the condition that the initial fire risk value is lower than the first threshold and the fire risk values ​​of adjacent nodes are all lower than the first threshold, several low-risk nodes are merged into a single node, and the average coordinate is used as the new node coordinate.

[0046] The calculation formula of the initial fire risk value is as follows:

[0047]

[0048] Where: is the initial fire risk value of node i, is the smoke concentration, is the temperature, is the concentration of harmful gases, 、 and is the weight parameter.

[0049] The calculation formula of the spatial adjacency matrix is ​​as follows:

[0050]

[0051] Where: is the spatial relationship matrix value between node i and node j, indicating their similarity in space and fire characteristics, is the spatial distance between node i and node j, is the fire feature vector of node i, is the fire feature vector of node j, It is the scale factor of spatial distance, which is used to control the influence of spatial distance on similarity. is the scale factor of fire characteristic difference, which is used to control the influence of node fire characteristic difference on similarity. is the base of the natural logarithm function.

[0052] The calculation formula for quadtree splitting is as follows:

[0053]

[0054] Where: is the geographical coordinate longitude of node i, is the geographic coordinate latitude of node i, is the horizontal division spacing, is the vertical division spacing.

[0055] S3, obtains the angle between the propagation directions and the propagation time difference between nodes, establishes a mapping model, obtains the directional sensitivity and corrects the adjacency matrix.

[0056] In this embodiment, the angle between the propagation directions and the propagation time difference between the nodes are obtained, a mapping model is established, the directional sensitivity is obtained, and the adjacency matrix is ​​corrected, as follows:

[0057] Extract the propagation inertia direction vector of each node based on the fire characteristic data, obtain the angle between the propagation direction and the spatial direction of the adjacent node, and output the directional sensitivity;

[0058] Obtain the characteristic change time points of node pairs in historical fire events, obtain the propagation time difference between nodes, combine the spatial direction relationship of nodes, obtain the relationship data between the propagation time difference and the direction angle, and establish a propagation delay-sensitivity mapping model;

[0059] The actual propagation time difference is obtained and the directional sensitivity is output based on the mapping model.

[0060] The propagation delay-sensitivity mapping model is as follows:

[0061]

[0062] Where: is the directional sensitivity, is the initial sensitivity, the initial value of the directional sensitivity when Δt=0 (i.e. the maximum enhancement when there is no delay in propagation), is the attenuation control coefficient, which is used to control the rate at which the sensitivity decreases as the propagation delay increases. is the node propagation time difference, is the base of natural logarithms.

[0063] In this embodiment, the adjacency matrix is ​​modified as follows:

[0064] Traverse all valid connection edges in the adjacency matrix and output the directional sensitivity of the valid edges based on the propagation delay-sensitivity mapping model;

[0065] Get the weight of the effective edge in the adjacency matrix and multiply it by the direction sensitivity to get the corrected edge weight;

[0066] According to the corrected edge weights, the corrected adjacency matrix is ​​obtained.

[0067] S4, inputs the corrected adjacency matrix and fire feature vector into the spatiotemporal graph convolutional network model and outputs the fire risk level.

[0068] In this embodiment, the modified adjacency matrix and fire feature vector are input into the spatiotemporal graph convolutional network model to output the fire risk level, as follows:

[0069] Acquiring input data, wherein the input data includes a fire feature vector and a multi-scale adjacency matrix;

[0070] The input data is processed through graph convolution operations, the spatial correlation between nodes is captured based on the multi-scale adjacency matrix, and feature aggregation is performed;

[0071] The feature-aggregated data is input into the temporal convolution layer to capture the changing trend of node features in the time dimension;

[0072] Based on the spatial attention mechanism, the weight of feature aggregation is adaptively adjusted according to the fire characteristics of the nodes and the multi-scale information of the adjacency matrix, highlighting the characteristic information of key areas and abnormal nodes;

[0073] Based on the temporal attention mechanism, dynamic weights are assigned to features at different time steps to further optimize the expressiveness of temporal features.

[0074] The historical fire data is divided into training, validation, and test sets, and a time sliding window method is used to generate time series samples to ensure that the model can learn the spatiotemporal dependencies of fire characteristics.

[0075] During model training, a weighted cross-entropy loss function is used to weight samples of different fire risk levels to improve the model's ability to detect minority fire samples.

[0076] The trained model outputs the fire risk probability of each node to form the fire risk level assessment result, such as Figure 3 As shown:

[0077] Horizontal axis (X-axis): represents the month, from January to December, reflecting the time span of the whole year;

[0078] Vertical axis (Y axis): represents the fire risk level. Although the unit is not specified, it is usually a relative score or index (e.g. 0 to 100 points).

[0079] Broken line: The red broken line marked as "Fire Risk Level" shows the monthly changes in risk level values;

[0080] The risk gradually increases at the beginning of the year (January-February), indicating that there may be increased fire hazards in late winter and early spring (such as frequent heating and electrical appliance use);

[0081] In spring (March-April), the fluctuations dropped slightly, but remained at a high level. From early summer to midsummer (May-August), the prices continued to rise significantly, reaching the highest point of the year in August. This may be related to factors such as hot and dry weather and peak electricity consumption.

[0082] The risk gradually decreases in autumn (September-October). As the climate cools and precipitation increases, the fire risk eases. The fire risk decreases further at the end of the year (November-December), but it is still higher than the level at the beginning of the year, which may be related to the increase in fire and electricity use in winter.

[0083] If the fire risk probability of a node is detected to exceed the preset threshold, or an abnormal characteristic mutation occurs, the real-time early warning system will be triggered immediately, and an alarm message will be sent to relevant personnel to respond to the fire risk in a timely manner.

[0084] Example 2, Figure 2 The present invention provides a system for a fire prediction, detection, adjustment and optimization method, including a feature extraction module, a node division module and a fire risk output module, wherein the modules are connected;

[0085] The feature extraction module is used to obtain the fire feature data of the area to be tested, perform feature extraction and splicing, and construct a fire feature vector;

[0086] The node division module is used to perform initial node division on the test area, dynamically adjust the node density according to the fire characteristics and geographical location of the nodes, extract spatiotemporal features, and construct a multi-scale adjacency matrix;

[0087] The correction module is used to obtain the angle and propagation time difference between nodes, establish a mapping model, obtain directional sensitivity and correct the adjacency matrix;

[0088] The fire risk output module is used to input the corrected adjacency matrix and fire feature vector into the spatiotemporal graph convolutional network model and output the fire risk level.

[0089] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0090] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0091] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0093] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0094] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fire prediction, detection, adjustment and optimization method, characterized in that: The steps include: Obtain fire feature data of the area to be tested, perform feature extraction and splicing, and construct a fire feature vector; Perform initial node division on the test area, dynamically adjust the node density based on the fire characteristics and geographical location of the nodes, extract spatiotemporal features, and construct a multi-scale adjacency matrix; Obtain the angle and propagation time difference between nodes, establish a mapping model, obtain directional sensitivity and correct the adjacency matrix; The modified adjacency matrix and fire feature vector are input into the spatiotemporal graph convolutional network model to output the fire risk level. The angle between the propagation directions and the propagation time difference between the nodes are obtained, a mapping model is established, and the directional sensitivity is obtained as follows: Extract the propagation inertia direction vector of each node based on the fire characteristic data, obtain the angle between the propagation direction and the spatial direction of the adjacent node, and output the directional sensitivity; Obtain the characteristic change time points of node pairs in historical fire events, obtain the propagation time difference between nodes, combine the spatial direction relationship of nodes, obtain the relationship data between the propagation time difference and the direction angle, and establish a propagation delay-sensitivity mapping model; Obtain the actual propagation time difference and output the directional sensitivity based on the mapping model; The modified adjacency matrix is ​​as follows: Traverse all valid connection edges in the adjacency matrix and output the directional sensitivity of the valid edges based on the propagation delay-sensitivity mapping model; Get the weight of the effective edge in the adjacency matrix and multiply it with the directional sensitivity to obtain the corrected edge weight; According to the corrected edge weights, the corrected adjacency matrix is ​​obtained.

2. The fire prediction, detection, adjustment and optimization method according to claim 1, characterized in that: The fire feature data of the area to be tested is obtained, and feature extraction and splicing are performed to construct a fire feature vector, as follows: Acquire historical fire data and pre-process the historical fire data; Perform feature extraction on pre-processed historical fire data; Normalize the feature data and concatenate the normalized feature vectors of different types to form a complete fire feature vector; Perform feature dimensionality reduction on the fire feature vector to obtain the initial fire feature vector.

3. The fire prediction, detection, adjustment and optimization method according to claim 2, characterized in that: The initial node division is performed on the test area, the node density is dynamically adjusted according to the fire characteristics and geographical location of the nodes, and the spatiotemporal characteristics are extracted to construct a multi-scale adjacency matrix, as follows: Obtain the boundary coordinate information of the area to be measured, divide the area into equal-spaced rectangular grids based on the boundary coordinate information, and set the initial node density; According to the divided nodes, node characteristic data of the nodes are obtained, and an initial fire risk value is output according to the node characteristic data; Dynamically adjust the number and density of nodes according to the initial fire risk value, and extract the spatiotemporal characteristics between nodes after adjustment; Adjacency matrices at different scales are constructed according to the spatiotemporal characteristics, and the adjacency matrices at different scales are fused to form a multi-scale adjacency matrix.

4. The fire prediction, detection, adjustment and optimization method according to claim 3, characterized in that: The number and density of nodes are dynamically adjusted according to the initial fire risk value as follows: Determine whether to split or merge nodes based on the initial fire risk value; When the initial fire risk value exceeds the second threshold, the node is quadtree split; Under the condition that the initial fire risk value is lower than the first threshold and the fire risk values ​​of adjacent nodes are all lower than the first threshold, several low-risk nodes are merged into a single node.

5. The fire prediction, detection, adjustment and optimization method according to claim 4, characterized in that: The modified adjacency matrix and fire feature vector are input into the spatiotemporal graph convolutional network model to output the fire risk level, as follows: Obtain input data, process it through graph convolution operations, capture the spatial correlation between nodes based on multi-scale adjacency matrices, and perform feature aggregation; The feature-aggregated data is input into the temporal convolution layer to capture the changing trend of node features in the time dimension; According to the changing trend of node features in the time dimension, the historical fire data is divided into training set, validation set and test set, and the time sliding window method is used to generate time series samples; The training set is input into the preset fire risk model for training. During the model training process, the samples with different fire risk levels are weighted by the weighted cross entropy loss function. The trained model outputs the fire risk probability of each node to form a fire risk level assessment result; If the fire risk probability of a detected node exceeds the preset threshold, the real-time warning system will be triggered immediately.

6. The fire prediction, detection, adjustment and optimization method according to claim 5, characterized in that: The adjacency matrix at different scales includes a spatial adjacency matrix, and the calculation formula of the spatial adjacency matrix is ​​as follows: Where: Aij is the spatial relationship matrix value between node i and node j, indicating their similarity in space and fire characteristics; dij is the spatial distance between node i and node j; Fi is the fire feature vector of node i; Fj is the fire feature vector of node j; σ is the scale factor of spatial distance, used to control the influence of spatial distance on similarity; γ is the scale factor of fire feature difference, used to control the influence of node fire feature difference on similarity; and e is the base of the natural logarithm function.

7. The fire prediction, detection, adjustment and optimization method according to claim 6, characterized in that: The calculation formula for the quadtree split is as follows: Where: x i is the geographical coordinate longitude of node i, y i is the geographic coordinate latitude of node i, Δx is the horizontal division spacing, and Δy is the vertical division spacing.

8. A system using a fire prediction, detection, adjustment and optimization method according to any one of claims 1 to 7, characterized in that: It includes feature extraction module, node division module and fire risk output module, and there are connections between modules; The feature extraction module is used to obtain the fire feature data of the area to be tested, perform feature extraction and splicing, and construct a fire feature vector; The node division module is used to perform initial node division on the test area, dynamically adjust the node density according to the fire characteristics and geographical location of the nodes, extract spatiotemporal features, and construct a multi-scale adjacency matrix; The correction module is used to obtain the angle and propagation time difference between nodes, establish a mapping model, obtain directional sensitivity and correct the adjacency matrix; The fire risk output module is used to input the corrected adjacency matrix and fire feature vector into the spatiotemporal graph convolutional network model and output the fire risk level.

Images