An intelligent fire protection system with automatic fire monitoring function

Abstract

The present invention relates to the field of smart fire fighting technology, and discloses a smart fire fighting system with an automatic fire monitoring function, which divides a target area into a plurality of sub-areas, and collects remote sensing images for each sub-area at fixed time intervals; maps each sub-area to a topological node, and constructs a topological edge between the topological nodes corresponding to adjacent sub-areas to obtain a topological structure; analyzes the remote sensing image through a preset image analysis model to obtain a fire score for each sub-area, and maps the fire score to a characteristic value of the corresponding topological node; obtains the characteristic values ​​of the topological nodes of the time series based on the remote sensing image, and combines the environmental information of the target area to obtain an updated model of the topological structure, obtains the sub-area corresponding to the rescue task, and obtains a rescue plan for the rescue task through the updated model of the topological structure.

Description

Technical Field

[0001] The present invention relates to the field of intelligent fire fighting technology, and more specifically, to an intelligent fire fighting system with automatic fire condition monitoring function. Background Art

[0002] When performing rescue missions in fires, it is very important to monitor the fire conditions. For each area in the fire, the change of the fire conditions over time is not only affected by the burning conditions of the area itself, but also often by the correlation of adjacent areas. However, existing rescue missions often only focus on the fire conditions of the areas on the rescue path, but ignore the correlation of adjacent areas on the rescue path, resulting in safety issues in the execution of rescue missions. Summary of the invention

[0003] The present invention provides an intelligent fire fighting system with automatic fire condition monitoring function, which solves the technical problems raised in the background technology.

[0004] The present invention provides an intelligent fire protection system with automatic fire monitoring function, comprising:

[0005] The data acquisition module is used to divide the target area into several sub-areas and collect remote sensing images for each sub-area at fixed time intervals;

[0006] A topology building module is used to map each sub-region into a topological node and build topological edges between topological nodes corresponding to adjacent sub-regions to obtain a topological structure;

[0007] An image analysis module is used to analyze the remote sensing image through a preset image analysis model to obtain a fire score for each sub-area and map the fire score to a characteristic value of a corresponding topological node;

[0008] The topology update module is used to obtain the characteristic values ​​of the topological nodes of the time series based on the remote sensing images, and combine the environmental information of the target area to obtain an updated model of the topological structure;

[0009] The rescue planning module is used to obtain the sub-area corresponding to the rescue mission and obtain the rescue plan of the rescue mission through the update model of the topological structure.

[0010] Furthermore, the target area is divided into several sub-areas, including:

[0011] Obtain the characteristic image of the target area in the historical period, calculate the normalized vegetation index for the characteristic image, and obtain the vegetation matrix;

[0012] Setting vegetation threshold , , replace all elements in the vegetation matrix that are less than the vegetation threshold with 0, and replace all elements in the vegetation matrix that are greater than the vegetation threshold with 1, and obtain a replacement matrix;

[0013] Perform eight-neighborhood detection on element 1 in the replacement matrix to obtain several clusters in the replacement matrix; divide the area corresponding to each cluster in the corresponding target area into a sub-area.

[0014] Further, the preset image analysis models include:

[0015] Obtain several fire pictures and manually annotate the fire pictures to obtain the corresponding fire scores;

[0016] Initialize and generate several occlusion clouds, and randomly add several occlusion clouds to each fire picture to obtain a corresponding set of training samples; the blur rate of the occlusion cloud is , , the size of the blocking cloud is , , Indicates the number of pixels of open fire in the corresponding fire image;

[0017] The fire score corresponding to each group of training samples is used as the sample label, and the image analysis model is trained by combining the training samples. The loss function of the hidden layer of the image analysis model is as follows:

[0018] ;

[0019] in, represents the loss function of the hidden layer of the image analysis model, and represents the first loss weight and the second loss weight, represents the number of samples in any set of training samples, express The index of express The index of Indicates the corresponding group The predicted fire score of samples, Indicates the corresponding group The predicted fire score of samples, represents the fire score of the corresponding group, Indicates the corresponding group The blur ratio of the occluding cloud in samples, Indicates the corresponding group The blur ratio of the occluding cloud in samples, , Represents the natural base.

[0020] Furthermore, the hidden layers of the image analysis model include:

[0021] The image analysis model includes several hidden layers, each hidden layer includes a convolution unit and a transformation unit, the output end of the convolution unit is connected to the input end of the transformation unit, and the output end of the transformation unit is connected to the input end of the convolution unit of the next hidden layer;

[0022] ;

[0023] ;

[0024] ;

[0025] in, Indicates The convolution graph of the Lth convolution unit input to the Lth convolution unit is The first Line The transformed value of the output of the column, Represents the Lth convolution unit to the The first feature map of the input of the transformation unit The first Line The convolution value of the output of the column, It represents the L-th convolution unit on the The first feature map of the input of the transformation unit The first Line The convolution value of the output of the column, represents the fuzzy adjustment weight, represents the first enhancement weight, Represents the remote sensing image Line The fuzziness of the column, Indicates hidden layers, Indicates hidden layers, express The index of Represents the length of the convolution kernel of the convolution unit, express The index of Represents the width of the convolution kernel of the convolution unit, express The index of Indicates The convolutional unit The convolution weights of the channels, Indicates The convolutional unit The convolution bias of the channels, Indicates The transformation unit is The convolutional graph input by the convolutional unit is The first Line The transformed value of the output of the column, represents the second enhancement weight, Represents the number of pixel columns of the remote sensing image, Indicates the number of pixel rows of the remote sensing image.

[0026] Furthermore, the topological structure also includes topological edge weights:

[0027] Constructing the topological edge weight for the uth topological node and the u+1th topological node in the topological structure, including: obtaining the border length of the uth sub-region and the u+1th sub-region , and the environmental information of the target area, which includes the wind angle between the u-th sub-area and the u+1-th sub-area and wind speed , and the normalized vegetation index of the u-th sub-region and the u+1-th sub-region in the vegetation matrix; wherein the u-th sub-region and the u+1-th sub-region are adjacent sub-regions;

[0028] The topological edge weight of the topological edge between the u-th topological node and the u+1-th topological node The calculation formula is as follows:

[0029] ;

[0030] in, represents the normalized vegetation index of the u-th sub-region, represents the normalized vegetation index of the u+1th sub-region, represents the total area of ​​the u-th sub-region and the u+1-th sub-region, represents the first linear weight, represents the first index weight, represents the second linear weight, Represents the second index weight.

[0031] Furthermore, an updated model of the topological structure is obtained, including:

[0032] In a preset time period, the characteristic value of each topological node at each moment in the preset time period is obtained, and an update model of the topological structure is constructed based on the characteristic value. The formula is as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] in, Represents a topological node At the moment The characteristic value of represents the first updated weight, represents the second updated weight, represents the third updated weight, Represents a topological node At the moment The characteristic value of represents the eigenvalue of topological node u+1 at time t, represents the set of adjacent topological nodes of topological node u, represents the neighborhood interaction factor, represents the environmental driving factor, represents the first learning rate, represents the second learning rate, represents the third learning rate, express Activation function;

[0037] Based on the characteristic value of each topological node at each moment in the preset time period, the custom parameters in the updated model of the topological structure are fitted by the gradient descent algorithm to obtain the specific values ​​of the custom parameters; wherein the custom parameters include , , , , and .

[0038] Furthermore, a rescue plan for the rescue mission is obtained, including:

[0039] Set a rescue time period, obtain the characteristic value of each topological node in the topological structure at the initial moment of the rescue time period, and iterate the update model of the topological structure to obtain the characteristic value of each topological node at each moment in the rescue time period;

[0040] Obtain the sub-region corresponding to the rescue task, and generate G rescue plans based on the constraint conditions. Each rescue plan includes entering the target area from any edge sub-region of the target area, and leaving the target area from any edge sub-region of the target area after reaching the sub-region corresponding to the rescue task;

[0041] The constraints include: the unit time moving distance of the rescue plan is ; The total distance of the rescue plan is less than the preset distance threshold; The total time of the rescue plan is less than the preset time threshold; At any moment in the rescue time period, the characteristic value of the topological node corresponding to the sub-area where the rescue plan is located is less than the preset characteristic value;

[0042] Evaluate G rescue plans and select the rescue plan with the highest evaluation value. The evaluation formula is as follows:

[0043] ;

[0044] in, represents the number of moments in the total time of the rescue plan, express The index of represents the total time of the rescue plan, represents the total distance of the rescue solution, Indicates The eigenvalue of the topological node of the rth iteration corresponding to the sub-region at the moment, , and Represent the first rescue weight, the second rescue weight and the third rescue weight respectively, , .

[0045] The beneficial effects of the present invention are: by constructing a dynamic topological structure and a time series feature prediction model, accurate monitoring and prediction of fire spread is achieved, edge weights are dynamically adjusted in combination with environmental information, and strong robustness and adaptability are achieved in complex fire scenarios. The fire dynamics can be evaluated in real time, effectively improving the safety and efficiency of fire monitoring and rescue. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a module diagram of an intelligent fire protection system with automatic fire monitoring function of the present invention. DETAILED DESCRIPTION

[0047] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0048] like Figure 1 As shown, a smart fire protection system with automatic fire monitoring function includes:

[0049] The data acquisition module is used to divide the target area into several sub-areas and collect remote sensing images for each sub-area at fixed time intervals;

[0050] A topology building module is used to map each sub-region into a topological node and build topological edges between topological nodes corresponding to adjacent sub-regions to obtain a topological structure;

[0051] An image analysis module is used to analyze the remote sensing image through a preset image analysis model to obtain a fire score for each sub-area and map the fire score to a characteristic value of a corresponding topological node;

[0052] The topology update module is used to obtain the characteristic values ​​of the topological nodes of the time series based on the remote sensing images, and combine the environmental information of the target area to obtain an updated model of the topological structure;

[0053] The rescue planning module is used to obtain the sub-area corresponding to the rescue mission and obtain the rescue plan of the rescue mission through the update model of the topological structure.

[0054] In one embodiment of the present invention, the target area is divided into several sub-areas, including:

[0055] Obtain the characteristic image of the target area in the historical period, calculate the normalized vegetation index for the characteristic image, and obtain the vegetation matrix;

[0056] Setting vegetation threshold , , replace all elements in the vegetation matrix that are less than the vegetation threshold with 0, and replace all elements in the vegetation matrix that are greater than the vegetation threshold with 1, and obtain a replacement matrix;

[0057] Perform eight-neighborhood detection on element 1 in the replacement matrix to obtain several clusters in the replacement matrix; divide the area corresponding to each cluster in the corresponding target area into a sub-area.

[0058] Specifically, feature images are extracted from historical remote sensing data of the target area (such as satellite images or drone images), and the feature images are processed to calculate the normalized vegetation index. This index is used to indicate the degree of vegetation coverage and is specifically defined as the ratio of the difference between the near-infrared band reflectivity and the red band reflectivity of the feature image to the sum of the near-infrared band reflectivity and the red band reflectivity. The calculation result of the vegetation index is stored in the form of a matrix. Each element in the matrix represents the degree of vegetation coverage at the corresponding position, and the value range is: -1 to 1. Each element in the vegetation matrix is ​​replaced. If the element value is less than the threshold , replace it with 0, indicating a non-vegetated area; if the element value is greater than or equal to the threshold , replace it with 1 to represent the vegetation area. After the replacement, a binary matrix (replacement matrix) is formed, which can clearly represent the distribution of sparse vegetation areas and dense vegetation areas in the target area. If a matrix element value is 1 and is adjacent to pixels with surrounding element values ​​of 1 (such as 8 neighborhoods), these adjacent pixels are considered to belong to the same connected area. Each connected area is identified as a "cluster", representing a vegetation area. The dense vegetation area is clustered using the connected area detection method to ensure that the divided sub-areas have actual geographical significance.

[0059] In one embodiment of the present invention, the preset image analysis model includes:

[0060] Obtain several fire pictures and manually annotate the fire pictures to obtain the corresponding fire scores;

[0061] Initialize and generate several occlusion clouds, and randomly add several occlusion clouds to each fire picture to obtain a corresponding set of training samples; the blur rate of the occlusion cloud is , , the size of the blocking cloud is , , Indicates the number of pixels of open fire in the corresponding fire image;

[0062] The fire score corresponding to each group of training samples is used as the sample label, and the image analysis model is trained by combining the training samples. The loss function of the hidden layer of the image analysis model is as follows:

[0063] ;

[0064] in, represents the loss function of the hidden layer of the image analysis model, and represents the first loss weight and the second loss weight, represents the number of samples in any set of training samples, express The index of express The index of Indicates the corresponding group The predicted fire score of samples, Indicates the corresponding group The predicted fire score of samples, represents the fire score of the corresponding group, Indicates the corresponding group The blur ratio of the occluding cloud in samples, Indicates the corresponding group The blur ratio of the occluding cloud in samples, , Represents the natural base.

[0065] Specifically, in actual fire scenes, fire burning is often accompanied by a large amount of smoke, which will cause occlusion of the fire area in the remote sensing image, thus affecting the accuracy of image analysis. The concentration and distribution of smoke are dynamically adjusted as the fire changes, so traditional image analysis models often find it difficult to effectively extract fire features under the condition of smoke occlusion. In order to overcome this problem, the occlusion rate, a key parameter, is introduced to simulate the dynamic occlusion effect of smoke on the image in the fire scene, and this effect is incorporated into the training and evaluation process of the model, thereby significantly enhancing the model's adaptability to actual complex scenes. The occlusion rate is used to quantify the degree of occlusion of the fire area by smoke, ranging from 0 to 1. When the occlusion rate is close to 0, it means that the smoke is thin and the image is clear; when the occlusion rate is close to 1, it means that the smoke is thick and the fire area is almost completely occluded. In the process of generating training data, the simulated smoke area is superimposed on the fire image, and the occlusion rate is randomly adjusted to generate smoke effects of different concentrations and distributions. This process is designed to simulate various possible scenarios in a real fire environment so that the training data covers a wider range of occlusion situations. The occlusion rate is used as a key parameter to measure the impact of smoke during the training process, guiding the model's learning process under different occlusion conditions. For example, when the smoke is thin (low occlusion rate), the model needs to pay more attention to the accurate extraction of fire characteristics; when the smoke is thick (high occlusion rate), the model needs to be more robust and able to capture fire characteristics from limited information. The smoke occlusion problem is a common challenge in actual fire monitoring, but traditional methods are often difficult to effectively consider during the training phase. By introducing the occlusion rate, the present invention simulates the actual impact of dynamic smoke in a fire on image occlusion, making the model training closer to the real scene. When the occlusion rate is low, the model focuses on learning the feature extraction of the fire area to ensure the high accuracy of the fire score; when the occlusion rate is high, the model reduces the strict requirements for the accuracy of the fire score and enhances the robustness of the fuzzy scene. By randomly simulating smoke scenes with different occlusion rates, the training data covers a variety of complex environments, so that the model can cope with various smoke occlusion conditions in actual deployment and improve its applicability.

[0066] In one embodiment of the present invention, the hidden layer of the image analysis model includes:

[0067] The image analysis model includes several hidden layers, each hidden layer includes a convolution unit and a transformation unit, the output end of the convolution unit is connected to the input end of the transformation unit, and the output end of the transformation unit is connected to the input end of the convolution unit of the next hidden layer;

[0068] ;

[0069] ;

[0070] ;

[0071] in, Indicates The convolution graph of the Lth convolution unit input to the Lth convolution unit is The first Line The transformed value of the output of the column, Represents the Lth convolution unit to the The first feature map of the input of the transformation unit The first Line The convolution value of the output of the column, It represents the L-th convolution unit on the The first feature map of the input of the transformation unit The first Line The convolution value of the output of the column, represents the fuzzy adjustment weight, represents the first enhancement weight, Represents the remote sensing image Line The fuzziness of the column, Indicates hidden layers, Indicates hidden layers, express The index of Represents the length of the convolution kernel of the convolution unit, express The index of Represents the width of the convolution kernel of the convolution unit, express The index of Indicates The convolutional unit The convolution weights of the channels, Indicates The convolutional unit The convolution bias of the channels, Indicates The transformation unit is The convolutional graph input by the convolutional unit is The first Line The transformed value of the output of the column, represents the second enhancement weight, Represents the number of pixel columns of the remote sensing image, Indicates the number of pixel rows of the remote sensing image.

[0072] Specifically, in the image analysis model, the hidden layer is the core module of the model for extracting features and learning complex patterns. In view of the complexity of fire monitoring in remote sensing images, especially the difficulty of feature extraction caused by occlusion (such as smoke) and environmental conditions (such as vegetation coverage, terrain characteristics), the present invention designs a hidden layer structure including convolution units and transformation units, and introduces a unique regulation mechanism to enhance the adaptability and accuracy of the model. The design of the hidden layer should not only consider the spatial characteristics of the image, but also adapt to the scenes of high noise and occlusion in remote sensing data, while highlighting the characteristics of the significant fire area. This demand prompted the present invention to add fuzzy regulation weights and significance enhancement mechanisms in the hidden layer to ensure that the model can accurately extract key features in different scenarios. The fuzzy regulation weight is used to dynamically adjust the focus of feature extraction in the hidden layer, and adaptively optimize the degree of blur (such as smoke concentration, vegetation occlusion) at different positions in the image. When the degree of blur in a certain area of ​​the image is high, the regulation weight will dynamically reduce the dependence on the blurry area, thereby reducing the interference of noise on feature extraction; when a certain area of ​​the image is clear, the regulation weight will strengthen feature extraction, so that the model can make full use of the information of the clear area. Traditional convolutional networks lack adaptive processing capabilities for fuzzy areas. By adjusting the fuzzy weights, they achieve dynamic response to image quality, achieve high precision and high robustness for fire monitoring tasks, and can be widely used in complex fire monitoring scenarios.

[0073] It should be noted that when the output of the transformation unit of the last hidden layer is expanded into a one-dimensional vector and input into the classifier, the classification space of the classifier represents the fire score corresponding to the remote sensing image.

[0074] In one embodiment of the present invention, the topological structure further includes topological edge weights:

[0075] Constructing the topological edge weight for the uth topological node and the u+1th topological node in the topological structure, including: obtaining the border length of the uth sub-region and the u+1th sub-region , and the environmental information of the target area, which includes the wind angle between the u-th sub-area and the u+1-th sub-area and wind speed , and the normalized vegetation index of the u-th sub-region and the u+1-th sub-region in the vegetation matrix; wherein the u-th sub-region and the u+1-th sub-region are adjacent sub-regions;

[0076] The topological edge weight of the topological edge between the u-th topological node and the u+1-th topological node The calculation formula is as follows:

[0077] ;

[0078] in, represents the normalized vegetation index of the u-th sub-region, represents the normalized vegetation index of the u+1th sub-region, represents the total area of ​​the u-th sub-region and the u+1-th sub-region, represents the first linear weight, represents the first index weight, represents the second linear weight, Represents the second index weight.

[0079] Specifically, in fire monitoring, the target area is divided into multiple sub-areas, each sub-area is mapped as a topological node, and the spread of fire between sub-areas requires quantifying the correlation between adjacent areas and the possibility of fire spread by constructing topological edge weights between topological nodes. Traditional fire propagation models usually only focus on the spatial border length, ignoring the key influence of environmental factors (such as wind direction and wind speed) and vegetation characteristics (such as the Normalized Difference Vegetation Index NDVI) on fire propagation. Therefore, the present invention designs a complex topological edge weight calculation formula by comprehensively considering the border length, environmental factors and vegetation characteristics, which fully reflects the multi-dimensional driving force of fire propagation.

[0080] In one embodiment of the present invention, obtaining an updated model of a topological structure includes:

[0081] In a preset time period, the characteristic value of each topological node at each moment in the preset time period is obtained, and an update model of the topological structure is constructed based on the characteristic value. The formula is as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] in, Represents a topological node At the moment The characteristic value of represents the first updated weight, represents the second updated weight, represents the third updated weight, Represents a topological node At the moment The characteristic value of represents the eigenvalue of topological node u+1 at time t, represents the set of adjacent topological nodes of topological node u, represents the neighborhood interaction factor, represents the environmental driving factor, represents the first learning rate, represents the second learning rate, represents the third learning rate, express Activation function;

[0086] Based on the characteristic value of each topological node at each moment in the preset time period, the custom parameters in the updated model of the topological structure are fitted by the gradient descent algorithm to obtain the specific values ​​of the custom parameters; wherein the custom parameters include , , , , and .

[0087] Specifically, in actual fire monitoring, the fire characteristic values ​​of different sub-areas (corresponding to topological nodes) change dynamically over time. This change is not only affected by the characteristics of the area itself (such as the current fire state and environmental driving force), but also significantly affected by the interaction between adjacent areas. Therefore, a topological structure update model is designed to dynamically describe and predict the time series changes of the characteristic values ​​of topological nodes, thereby providing reliable data support for fire propagation modeling and rescue planning. The topological structure update model dynamically updates the characteristic values ​​of each topological node to predict its state at a future moment. This process comprehensively considers the continuity of the characteristic values ​​of the topological node itself, the interaction between adjacent topological nodes, and the influence of external environmental driving forces on the characteristic values ​​of the topological nodes. A series of custom parameters are involved in the model to dynamically adjust the update process of the characteristic values ​​of the topological nodes. The continuity weight parameter controls the contribution of the characteristic values ​​of the topological nodes themselves to the future state. The interaction weight parameter quantifies the influence of neighborhood interactions on the changes in the characteristic values ​​of topological nodes. The environmental driving force weight parameter controls the direct effect of external environmental factors on the characteristic values ​​of topological nodes. The specific values ​​of the parameters are trained through optimization algorithms such as gradient descent to ensure that the model can achieve the best fit in the actual data to the characteristic values ​​of the topological nodes of the topological structure at each moment in the preset time period.

[0088] In one embodiment of the present invention, obtaining a rescue plan for a rescue mission includes:

[0089] Set a rescue time period, obtain the characteristic value of each topological node in the topological structure at the initial moment of the rescue time period, and iterate the update model of the topological structure to obtain the characteristic value of each topological node at each moment in the rescue time period;

[0090] Obtain the sub-region corresponding to the rescue task, and generate G rescue plans based on the constraint conditions. Each rescue plan includes entering the target area from any edge sub-region of the target area, and leaving the target area from any edge sub-region of the target area after reaching the sub-region corresponding to the rescue task;

[0091] The constraints include: the unit time moving distance of the rescue plan is ; The total distance of the rescue plan is less than the preset distance threshold; The total time of the rescue plan is less than the preset time threshold; At any moment in the rescue time period, the characteristic value of the topological node corresponding to the sub-area where the rescue plan is located is less than the preset characteristic value;

[0092] Evaluate G rescue plans and select the rescue plan with the highest evaluation value. The evaluation formula is as follows:

[0093] ;

[0094] in, represents the number of moments in the total time of the rescue plan, express The index of represents the total time of the rescue plan, represents the total distance of the rescue solution, Indicates The eigenvalue of the topological node of the rth iteration corresponding to the sub-region at the moment, , and Represent the first rescue weight, the second rescue weight and the third rescue weight respectively, , .

[0095] Specifically, the proposed rescue mission planning scheme takes the dynamic update model of topological nodes as the core, and generates the fire evolution characteristics of each topological node during the rescue time period by predicting the characteristic values ​​of the topological nodes in the target area in time series. On this basis, a variety of rescue paths that meet the constraints are generated based on the initial conditions. The constraints include the single moving distance and total time limit, and the fire characteristic value cannot exceed the safety threshold. Each scheme is comprehensively evaluated by the evaluation formula. The evaluation formula takes the accumulation of the fire characteristic values ​​of the sub-region (indicating the total amount of fire risk), the total time of the rescue path, and the total distance as key parameters. The priority is determined by weighted summation. The setting of weights ensures that high fire risk areas are avoided first and the rescue time and distance are shortened as much as possible, so as to select the optimal scheme. This scheme combines timeliness, security and flexibility, and is particularly suitable for intelligent rescue mission planning in dynamic fire scenarios.

[0096] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. An intelligent fire protection system with automatic fire monitoring function, characterized in that: include: The data acquisition module is used to divide the target area into several sub-areas and collect remote sensing images for each sub-area at fixed time intervals; A topology building module is used to map each sub-region into a topological node and build topological edges between topological nodes corresponding to adjacent sub-regions to obtain a topological structure; The image analysis module is used to analyze the remote sensing image through a preset image analysis model to obtain the fire score of each sub-area and map the fire score to the characteristic value of the corresponding topological node; wherein the preset image analysis model includes: Obtain several fire pictures and manually annotate the fire pictures to obtain the corresponding fire scores; Initialize and generate several occlusion clouds, and randomly add several occlusion clouds to each fire picture to obtain a corresponding set of training samples; the blur rate of the occlusion cloud is , , the size of the blocking cloud is , , Indicates the number of pixels of open fire in the corresponding fire image; The fire score corresponding to each group of training samples is used as the sample label, and the image analysis model is trained by combining the training samples. The loss function of the hidden layer of the image analysis model is as follows: ； in, represents the loss function of the hidden layer of the image analysis model, and represents the first loss weight and the second loss weight, represents the number of samples in any set of training samples, express The index of express The index of Indicates the first The predicted fire score of samples, Indicates the first The predicted fire score of samples, represents the fire score of the corresponding group, Indicates the first The blur ratio of the occluding cloud in samples, Indicates the first The blur ratio of the occluding cloud in samples, , Represents the natural base; The topology update module is used to obtain the characteristic values ​​of the topological nodes of the time series based on the remote sensing images, and combine the environmental information of the target area to obtain an updated model of the topological structure; The rescue planning module is used to obtain the sub-area corresponding to the rescue mission and obtain the rescue plan of the rescue mission through the update model of the topological structure.

2. According to claim 1, a smart fire protection system with automatic fire monitoring function is characterized in that: Divide the target area into several sub-areas, including: Obtain the characteristic image of the target area in the historical period, calculate the normalized vegetation index for the characteristic image, and obtain the vegetation matrix; Setting vegetation threshold , , replace all elements in the vegetation matrix that are less than the vegetation threshold with 0, and replace all elements in the vegetation matrix that are greater than the vegetation threshold with 1, and obtain a replacement matrix; Perform eight-neighborhood detection on element 1 in the replacement matrix to obtain several clusters in the replacement matrix; divide the area corresponding to each cluster in the corresponding target area into a sub-area.

3. According to claim 1, the intelligent fire protection system with automatic fire monitoring function is characterized in that: The hidden layers of the image analysis model include: The image analysis model includes several hidden layers, each hidden layer includes a convolution unit and a transformation unit, the output end of the convolution unit is connected to the input end of the transformation unit, and the output end of the transformation unit is connected to the input end of the convolution unit of the next hidden layer; ； ； ； in, Indicates The convolution graph of the Lth convolution unit input to the Lth convolution unit is The first Line The transformed value of the output of the column, Represents the Lth convolution unit to the The first feature map of the input of the transformation unit The first Line The convolution value of the output of the column, It represents the L-th convolution unit on the The first feature map of the input of the transformation unit The first Line The convolution value of the output of the column, represents the fuzzy adjustment weight, represents the first enhancement weight, Represents the remote sensing image Line The fuzziness of the column, Indicates hidden layers, Indicates hidden layers, express The index of Represents the length of the convolution kernel of the convolution unit, express The index of Represents the width of the convolution kernel of the convolution unit, express The index of Indicates The convolutional unit The convolution weights of the channels, Indicates The convolutional unit The convolution bias of the channels, Indicates The transformation unit is The convolutional graph input by the convolutional unit is The first Line The transformed value of the output of the column, represents the second enhancement weight, Represents the number of pixel columns of the remote sensing image, Indicates the number of pixel rows of the remote sensing image.

4. According to claim 2, the intelligent fire protection system with automatic fire monitoring function is characterized in that: The topology structure also includes topological edge weights: Constructing the topological edge weight for the uth topological node and the u+1th topological node in the topological structure, including: obtaining the border length of the uth sub-region and the u+1th sub-region , and the environmental information of the target area, which includes the wind angle between the u-th sub-area and the u+1-th sub-area and wind speed , and the normalized vegetation index of the u-th sub-region and the u+1-th sub-region in the vegetation matrix; wherein the u-th sub-region and the u+1-th sub-region are adjacent sub-regions; The topological edge weight of the topological edge between the u-th topological node and the u+1-th topological node The calculation formula is as follows: ； in, represents the normalized vegetation index of the u-th sub-region, represents the normalized vegetation index of the u+1th sub-region, represents the total area of ​​the u-th sub-region and the u+1-th sub-region, represents the first linear weight, represents the first index weight, represents the second linear weight, Represents the second index weight.

5. According to claim 4, the intelligent fire protection system with automatic fire monitoring function is characterized in that: Get an updated model of the topology, including: In a preset time period, the characteristic value of each topological node at each moment in the preset time period is obtained, and an update model of the topological structure is constructed based on the characteristic value. The formula is as follows: ； ； ； in, Represents a topological node At the moment The characteristic value of represents the first updated weight, represents the second updated weight, represents the third updated weight, Represents a topological node At the moment The characteristic value of represents the eigenvalue of topological node u+1 at time t, represents the set of adjacent topological nodes of topological node u, represents the neighborhood interaction factor, represents the environmental driving factor, represents the first learning rate, represents the second learning rate, represents the third learning rate, express Activation function; Based on the characteristic value of each topological node at each moment in the preset time period, the custom parameters in the updated model of the topological structure are fitted by the gradient descent algorithm to obtain the specific values ​​of the custom parameters; wherein the custom parameters include , , , , and .

6. The intelligent fire protection system with automatic fire monitoring function according to claim 5 is characterized in that: Get rescue solutions for rescue missions, including: Set a rescue time period, obtain the characteristic value of each topological node in the topological structure at the initial moment of the rescue time period, and iterate the update model of the topological structure to obtain the characteristic value of each topological node at each moment in the rescue time period; Obtain the sub-region corresponding to the rescue task, and generate G rescue plans based on the constraint conditions. Each rescue plan includes entering the target area from any edge sub-region of the target area, and leaving the target area from any edge sub-region of the target area after reaching the sub-region corresponding to the rescue task; The constraints include: the unit time moving distance of the rescue plan is ; The total distance of the rescue plan is less than the preset distance threshold; The total time of the rescue plan is less than the preset time threshold; At any moment in the rescue time period, the characteristic value of the topological node corresponding to the sub-area where the rescue plan is located is less than the preset characteristic value; Evaluate G rescue plans and select the rescue plan with the highest evaluation value. The evaluation formula is as follows: ； in, represents the number of moments in the total time of the rescue plan, express The index of represents the total time of the rescue plan, represents the total distance of the rescue solution, Indicates The eigenvalue of the topological node of the rth iteration corresponding to the sub-region at the moment, , and Represent the first rescue weight, the second rescue weight and the third rescue weight respectively, , .

Images