Unmanned aerial vehicle fire source point positioning path planning method based on machine learning

Abstract

The invention relates to the technical field of fire source positioning path planning, in particular to an unmanned aerial vehicle fire source point positioning path planning method based on machine learning. Comprising the following steps: firstly, acquiring a suspected fire source estimated position of a mountain forest, and establishing a fire source attraction vector field according to position information and environment information of a suspected fire source area; secondly, generating an accurate initial fire source search path according to influence information in the fire source vector field; in the searching process, a fire source point recognition model is adopted, the model carries out learning from multi-dimensional features according to image information and numerical information, the accuracy of fire source recognition is improved, and the position of a fire source can be adjusted according to a recognition result; and finally, establishing an obstacle repulsion field by using obstacle information of the mountain forest, and combining the obstacle repulsion field with the fire source attraction vector field to form a total influence vector field optimization search path. According to the method, the accuracy of unmanned aerial vehicle fire source point positioning path planning is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fire source positioning path planning, and in particular to a method for unmanned aerial vehicle fire source positioning path planning based on machine learning. Background Art

[0002] Fire is a destructive event caused by combustion, usually involving abnormal phenomena such as flames, smoke and high temperature; fire often occurs in various scenes, such as residential areas, industrial areas, grasslands and forests. Among these scenes, forest fires occur most frequently, especially in summer and autumn every year, when vegetation in forests is more likely to catch fire.

[0003] Once a forest fire occurs, it will cause very serious impacts. For example, the fire destroys the forest vegetation, causing many animals and plants to lose their habitats. The burning vegetation releases carbon dioxide, carbon monoxide, methane and other harmful gases, which have a negative impact on air quality and global climate; fires may damage or destroy infrastructure (such as roads, power facilities, etc.) and private property, increasing reconstruction and repair costs. Therefore, when a fire occurs, it is necessary to promptly discover the fire source and prevent and control it to prevent the fire from further expanding.

[0004] When a forest fire occurs, the dense vegetation in the forest is easy to ignite, which greatly facilitates the spread of the fire. Traditional fire detection methods result in a long reaction time, which leads to a delay in fire fighting. With the rapid development of drone technology and machine learning technology, the use of drones for forest inspections has been widely used. In this process, machine learning models can process and analyze these large amounts of data to automatically identify fire sources, abnormal vegetation, and potential danger areas. In particular, in terms of fire source location and path planning in the forest, the combination of drone technology and machine learning technology can significantly improve the efficiency and accuracy of forest fire source location and path planning. This combination not only improves the speed and quality of data collection, but also enhances the accuracy of fire source identification and the intelligence of path planning, ultimately achieving more efficient and safe forest inspections and fire emergency response.

[0005] Many excellent methods have been proposed in the existing literature to use drones to locate and plan paths for forest fire sources, and have been put into practical use. However, in real scenarios, some existing technologies for fire source location planning lack consideration of obstacles in the forest, which results in inaccurate fire source location planning during the inspection process, resulting in the inability to prevent and control forest fires in a timely manner.

[0006] To this end, the present invention proposes a method for UAV fire source location path planning based on machine learning. Summary of the invention

[0007] The present invention proposes a method for positioning a fire source point of an unmanned aerial vehicle based on machine learning for positioning a fire source point when a forest fire occurs; the present invention mainly proceeds from the following aspects; firstly, the present invention establishes a spatial coordinate set according to the spatial position information of the suspected fire source point when a forest fire occurs; then, the fire source attraction vector field is established according to the spatial distribution of the suspected fire source point and the forest environment characteristics; secondly, the present invention initializes the fire source search path of the unmanned aerial vehicle according to the influence intensity of the suspected fire source point; in the process of searching according to the fire source search path, the present invention proposes a fire source point recognition model for identifying the suspected fire source point; the model identifies the suspected fire source point by The image data and numerical data of the fire source point are analyzed, and a weighted operation is performed by extracting the features of the two different types of data; the weighted features are weighted calculated to improve the accuracy of fire source identification; in addition, the present invention calculates the fire probability of the area that is not a fire source point identified by the fire source point identification model, collects vegetation, environmental factor data and geographic information data in combination with historical fire data, and outputs the results through a logistic regression model after characterization; finally, in order to improve the accuracy of path planning, the present invention regards obstacles in the mountains and forests as repulsive fields, so that the drone can reduce the impact of obstacles during flight and ensure the effectiveness of route planning.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] Step 1: Obtain the estimated location of the suspected fire source in the forest and construct an estimated spatial coordinate set;

[0010] Step 2: Establish a fire source attraction vector field; wherein the function of the fire source attraction vector field is:

[0011]

[0012] in, represents the function of the fire source attraction vector field; P represents the position vector of the UAV; FP j represents the estimated position vector of the suspected fire source point j; g j Indicates the intensity coefficient of the suspected fire source point j; B j represents the influence factor of suspected fire source j; θ j represents the standard deviation of the influence range of the suspected fire source j; q represents the power of the distance; represents a non-zero constant; represents the unit vector calculation function; α() represents the adjustment factor of the mountain forest environment characteristics;

[0013] Step 3: Initialize the fire source search path of the UAV according to the fire source attraction vector field; wherein the generation process of the initial fire source search path includes: obtaining the influence information of each suspected fire source point in the fire source attraction vector field; the influence information represents Wherein, l represents the estimated location of the suspected fire source; represents the impact direction; s represents the impact intensity;

[0014] The process of obtaining the impact intensity includes: obtaining terrain features, vegetation features and meteorological features; calculating the spatial distance according to the estimated location of the suspected fire source; the calculation formula of the spatial distance is: Among them, E <i,j> represents the spatial distance between the suspected fire source point i and the suspected fire source point j; (x i ,y i ,z i ) represents the spatial coordinates of the suspected fire source point i; (x j ,y j ,z j ) represents the spatial coordinates of the suspected fire source point j; the intensity influence constant is obtained according to the historical fire occurrence records, and the basic influence intensity is calculated; the calculation formula of the basic influence intensity is: represents the basic impact intensity of the suspected fire source i; γ i represents the influence intensity constant of the suspected fire source point i; the environmental adjustment factor is defined according to the terrain characteristics, the vegetation characteristics and the meteorological characteristics; wherein the calculation formula of the environmental adjustment factor is: F env =F slope *F vegetation *F weather Among them, F env represents the environmental adjustment factor; F slope Represents the slope factor; F vegetation Represents vegetation factor; F weather represents the meteorological factor; the impact intensity is calculated according to the environmental adjustment factor; the calculation formula of the impact intensity is: s i =s i base *F env ; Among them, s i Represents the impact intensity of the suspected fire source i.

[0015] Further, a direction vector is calculated according to the influencing direction, and a normalization operation is performed on the direction vector to obtain a normalized direction vector;

[0016] Further, the search radius is calculated; the search radius is represented by: R=k*s; wherein R represents the search radius value; k represents the proportional constant;

[0017] Furthermore, the search path step size is set to obtain the initial path point;

[0018] Further, the initial path point is generated, and the generation formula is: P n =l+n*D norm *L; where P n represents the initial path point after n steps; n represents the number of steps; D norm represents the normalized direction vector; L represents the search path step length;

[0019] Furthermore, after screening the generated initial path points for abnormal points, the initial fire source search path is generated.

[0020] Step 4: Obtain relevant data of the current position of the drone through the equipment on the drone; wherein the relevant data includes: image data and numerical data; wherein the image data includes: original image, depth image and multispectral image; the numerical data includes: geographic coordinate data, vegetation data, meteorological data and environmental data.

[0021] Step 5: preprocessing the relevant data to obtain standardized data;

[0022] Among them, the preprocessing includes: image data preprocessing and numerical data preprocessing; wherein, the image data preprocessing includes: removing noise in the image data to obtain denoised image data; cropping the denoised image data to a uniform size to obtain standard-scale image data; enhancing the standard-scale image data to obtain enhanced image data; the numerical data preprocessing includes: processing missing values ​​of the numerical data to obtain filled numerical data; normalizing the supplemented numerical data to obtain normalized numerical data; and performing feature selection on the normalized numerical data to obtain standard numerical data.

[0023] Step 6: Input the standardized data into the fire source point identification model to obtain the fire source identification result;

[0024] Wherein, the fire source identification model includes:

[0025] Input layer: including an image input module and a numerical input module, which are used to receive the image data and numerical data in the pre-processed related data respectively;

[0026] The feature extraction layer includes an image feature extraction module and a numerical feature extraction module, which are respectively used to extract image features and numerical features from the input data of the input layer; wherein the image features include: color features, shape features, texture features, temperature features, smoke features and dynamic features; the numerical features include: temperature data, humidity data and gas concentration data.

[0027] The fire feature weighting layer includes an image feature weighting module and a numerical feature weighting module, which are respectively used to weight the image features and the numerical features that are conducive to fire identification; wherein the weighting formula of the fire feature weighting layer is:

[0028]

[0029] Wherein, WIF represents the weighted feature of the image feature weighting module; i Represents the weight of the i-th image feature; If i represents the i-th image feature; n represents the total number of image features; WVF represents the weighted feature of the numerical feature weighting module; ν j Vf represents the weight of the j-th numerical feature; j represents the jth numerical feature; m represents the total number of numerical features;

[0030] The fire identification layer is used to identify fires based on image weighted features and numerical weighted features;

[0031] An output layer, used for outputting the fire source identification result;

[0032] Furthermore, in the step five, when the fire source identification result is no, the probability of fire occurrence at the current location is obtained; wherein, the specific process of obtaining the fire occurrence probability includes: obtaining the relevant data about vegetation data, environmental factor data and geographic information data; combining the historical fire data of the current location to form a potential fire data set; applying feature engineering to the data in the potential fire data set to obtain a potential fire feature set; and inputting the potential fire feature set into a logistic regression model to obtain the probability of fire occurrence.

[0033] Step 7: adjusting the estimated position of the suspected fire source point according to the fire source identification result, and re-obtaining the impact direction and impact intensity;

[0034] Step 8: Optimizing the initial fire source search path to obtain an optimized fire source search path; the specific process includes:

[0035] Collecting map data of the mountain and forest;

[0036] Further, obstacles in the initial fire source search path are marked according to the obstacle information recorded on the map data, including: the location, type and size of the obstacle;

[0037] Furthermore, an obstacle repulsion field is established according to the annotation content of the obstacle; wherein the function of the obstacle repulsion field is: Among them, P irepresents the position vector of obstacle i; k i A is the constant related to obstacle type i; i represents the size factor of obstacle i; ||PP i || represents the Euclidean distance between the drone position and obstacle i; λ represents the power; ε represents a non-zero constant; σ i represents the standard deviation of the size of obstacle i; exp() represents the Gaussian function;

[0038] Further, the obstacle repulsion field is combined with the fire source attraction vector field to obtain a total influence vector field;

[0039] Further, the total influence vector field is used to optimize the initial fire source search path;

[0040] Further, visually displaying the relevant area in the map data according to the initial fire source search path;

[0041] Further, visually displaying the relevant area in the map data according to the initial fire source search path;

[0042] Further, the relevant area is divided into a plurality of uniform grids using a space grid method;

[0043] Further, the total influence vector field gradient in each of the grids is calculated; wherein the calculation formula of the total influence vector field gradient is:

[0044]

[0045] in, The gradient vector representing the total influence vector field of each network;

[0046] Further, the influence field intensity and direction of the suspected fire source point are evaluated according to the total influence vector field gradient of each of the grids;

[0047] Further, the fire source search path is adjusted using a gradient descent algorithm according to the evaluation result to obtain the optimized fire source search path;

[0048] Step 9: Repeat step 4 to step 8, and stop when the suspected fire source is identified as the real fire source;

[0049] Step 10: Obtain the optimal fire source search path and the identification of each suspected fire source point.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention proposes an initial path planning method for positioning a fire source point by an unmanned aerial vehicle. The method establishes a fire source attraction vector field according to the estimated position of the suspected fire source point for planning the initial path; in the generation process, the planned path points are obtained according to the spatial position information, influence direction and influence intensity of the suspected fire source point of the vector field; wherein, in the process of obtaining the influence intensity, the terrain characteristics, vegetation characteristics and meteorology of the suspected fire source point are comprehensively considered, and a comprehensive fire source search path planning scheme can be provided, so that forest fires can be timely controlled.

[0052] 2. The present invention proposes a path optimization method; this method generates a corresponding repulsive field through known obstacle information; the repulsive field is generated by considering the type, size and position of the obstacle, and the repulsive field is combined with the gravitational field to generate the final influence field; this method can effectively avoid obstacles and obtain the best path through iterative optimization, thereby improving the safety and efficiency of the path and enabling forest fires to be controlled in a timely manner.

[0053] 3. The present invention proposes a fire source point recognition model for fire source point recognition of UAVs during cruising; the model includes: an input layer, a feature extraction layer, a fire feature weighting layer, a fire recognition layer and an output layer; wherein the fire feature weighting layer weights the learned image features and numerical features respectively, and the main purpose of the weighting is to enable the model to better identify the fire source features, so as to improve the accuracy of identifying the real fire source point, provide a reliable judgment basis for forest fires, and facilitate the prevention and control of forest fires. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A flow chart of a method for locating a fire source point of an unmanned aerial vehicle based on machine learning provided in an embodiment of the present invention;

[0055] Figure 2 A structural diagram of a fire source identification model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The frequent occurrence of forest fires has brought serious hazards in many aspects, which not only affect the environment, but also have a profound impact on humans. Forests are densely covered with vegetation and rich in flammable materials. Once a fire occurs, the flames can quickly spread from one area to another. Wind speed and terrain will also aggravate the spread of the fire and expand the scope of the fire. If timely firefighting is not carried out, huge losses will be caused. Therefore, in order to avoid the occurrence of forest fires, relevant staff use drone technology to locate the fire source. Due to the maneuverability, autonomy and safety of drones, they have played a great role in forest fire prevention. At present, many documents have proposed methods for locating and planning paths for forest fire sources using drones, but there are some defects therein; therefore, the present invention proposes a method for locating and planning paths for drone fire sources based on machine learning to solve the problems existing in the prior art; the present invention will be introduced in detail from two embodiments below.

[0058] Embodiment 1:

[0059] In the embodiment of the present application, satellite monitoring is used to find that there is a suspected fire source in the forest A; in order to further check the exact location of the fire source, the method proposed in the present invention is used to locate the fire source and plan the path; the specific implementation steps are as follows: Figure 1 As shown, it includes: S10. Obtain the position vector of the suspected fire source point and construct an estimated spatial coordinate set; S20. Establish the fire source influence vector field; S30. Generate the initial fire source search path; S40. Obtain the relevant data of the current suspected fire source point; S50. Preprocess the relevant data; S60. Identify the fire source; S70. Update the fire source point location information; S80. Optimize the fire source search path; S90. Obtain the fire source point location information and the optimal search path. The following will be based on Figure 1 Explain the contents in

[0060] As shown in Table 1, the estimated position vector of the suspected fire source point in the mountain forest A is obtained by the satellite monitoring system, and an estimated spatial coordinate set is constructed, corresponding to step S10;

[0061] Table 1. Estimated location information of suspected fire sources

[0062]

[0063] According to the content in Table 1, this set represents:

[0064]

[0065] Further, a fire source attraction vector field is constructed, corresponding to step S20; wherein the process of establishing the fire source attraction vector field includes: obtaining the mountain forest environment characteristics of the location of the UAV, and digitizing the obtained mountain forest environment characteristics to obtain the adjustment factor of the mountain forest environment characteristics; according to the location of the UAV and the estimated location of the suspected fire source point, a construction function of the fire source attraction vector field is obtained, and the function is expressed as:

[0066]

[0067] Wherein, P represents the position vector of the UAV; FP j represents the estimated position vector of the suspected fire source point j; g j Indicates the intensity coefficient of the suspected fire source point j; B j represents the influence factor of suspected fire source j; θ i represents the standard deviation of the influence range of the suspected fire source j; q represents the power of the distance; represents a non-zero constant; represents the unit vector calculation function; α() represents the adjustment factor of the mountain forest environment characteristics;

[0068] Furthermore, an initial fire source search path is generated according to the fire source attraction vector field established above, corresponding to step S30; wherein the generation process of the initial fire source search path includes:

[0069] Obtain the influence information of each suspected fire source point in the fire source attraction vector field; the influence information represents Wherein, l represents the estimated location of the suspected fire source; represents the impact direction; s represents the impact intensity;

[0070] The impact intensity is obtained based on the terrain characteristics, vegetation characteristics and meteorological characteristics of the suspected fire source points combined with the spatial distance between the fire source points; the spatial distance is calculated using the spatial Euclidean distance method, and the formula is: Among them, E i,j> represents the spatial distance between the suspected fire source point i and the suspected fire source point j; (x i ,y i ,z i ) represents the spatial coordinates of the suspected fire source point i; (x j ,y j ,z j ) represents the spatial coordinates of the suspected fire source point j; the intensity influence constant is obtained according to the historical fire occurrence records; for example, the spatial distance between the suspected fire source point 1 and the suspected fire source point 2 is:

[0071] The intensity influence constant is obtained according to the historical fire occurrence records, and the basic influence intensity is calculated; the calculation formula of the basic influence intensity is: in, represents the basic impact intensity of the suspected fire source i; γ i represents the influence intensity constant of suspected fire source i;

[0072] The initial search path for the UAV is generated based on the influence strength of the suspected fire source, which is obtained based on the terrain characteristics, vegetation characteristics, and meteorological characteristics of the suspected fire source combined with the spatial distance. This can significantly improve the accuracy and efficiency of the UAV path planning. This method ensures the comprehensiveness and scientificity of path planning.

[0073] The environmental adjustment factor is defined according to the terrain characteristics, vegetation characteristics and meteorological characteristics at the suspected fire source; wherein the calculation formula of the environmental adjustment factor is: F env =F slope *F vegetation *F weather Among them, F env represents the environmental adjustment factor; F slope Represents the slope factor; F vegetation Represents vegetation factor; F weather represents meteorological factors;

[0074] The calculation formula for the slope factor is: F slope =1+k slope *S; where S represents the terrain slope; k slope Indicates the slope adjustment factor;

[0075] The calculation formula for vegetation factor is: F vegetation =k veg *D*(1-H); where k veg represents vegetation adjustment coefficient; D represents vegetation density; H represents vegetation humidity;

[0076] The calculation formula for the meteorological factor is: F weather =k wea *W*T; where k wea represents the adjustment coefficient; W represents the wind speed; T represents the temperature;

[0077] The impact intensity is calculated according to the environmental adjustment factor; the calculation formula of the impact intensity is: s i =s i base *F env ; Among them, s i Represents the impact intensity of the suspected fire source i.

[0078] Furthermore, the influencing direction calculates a direction vector, and performs a normalization operation on the direction vector to obtain a normalized direction vector;

[0079] Further, the search radius is calculated; the search radius is expressed as: R = k*s; wherein R represents the search radius value; k represents a proportional constant; in the embodiment of the present application, the proportional constant is set according to the coverage range of the suspected fire source point, and the larger the coverage range, the larger the proportional constant is set;

[0080] Further, the search path step length is set to obtain the initial path point; the search path step length is set according to the flight distance per unit time of the UAV at a constant speed; wherein the initial path point is a point with the suspected fire source as the center and a circular area formed according to the search radius, with the search path step length as the interval;

[0081] Furthermore, after filtering the generated initial path points for abnormal points, the initial fire source search path is generated. Among them, the abnormal points are mainly removed from points with calculation errors and overlaps;

[0082] In this application example, an initial fire source search path is planned for the positioning path of the UAV to the fire source point; the generation process of the path is planned using the influence information of each suspected fire source point in the fire source attraction vector field, and the search radius and search step are set to obtain the initial path point; the initial fire source search path is finally obtained by screening the initial path points; this method can effectively improve the efficiency and accuracy of the UAV positioning of the fire source point. It has significant advantages in optimizing the search path, improving the search efficiency, reducing time consumption, improving the autonomy of the UAV, enhancing the positioning accuracy and adapting to different environments.

[0083] Furthermore, the drone sets a random starting point according to the initial fire source search path, that is, randomly sets any suspected fire source point as the starting point; the camera installed on the drone uses a sensor to obtain relevant data of the current suspected fire source point, corresponding to Figure 1 Step S40; wherein the relevant data includes: image data and numerical data; the image data includes an original image referring to a scene image of the suspected fire source area, a depth image referring to depth information in the scene of the suspected fire source area, and a multispectral image referring to thermal reactions of different materials in the suspected fire source area; the numerical data includes: geographic coordinate data, vegetation data, meteorological data, and environmental data.

[0084] In the embodiment of the present application, the fire source point is identified by acquiring image data and numerical data; through multi-dimensional information fusion; wherein, the image data provides visual features (such as the color and shape of the flame), and the numerical data provides environmental information (such as temperature and humidity). Combining these two types of data can more comprehensively identify the fire source point and improve the accuracy of identification.

[0085] Further, the relevant data is preprocessed, corresponding to step S50; specifically including: the preprocessing is divided into: image data preprocessing and numerical data preprocessing; wherein, the image data preprocessing includes: removing noise in the image data to obtain denoised image data; performing uniform size cropping on the denoised image data to obtain standard scale image data; enhancing the standard scale image data to obtain enhanced image data; the numerical data preprocessing includes: processing missing values ​​of the numerical data to obtain filled numerical data; normalizing the supplemented numerical data to obtain normalized numerical data; performing feature selection on the normalized numerical data to obtain standard numerical data.

[0086] In the embodiments of the present application, the acquired image data and numerical data are preprocessed respectively, which is beneficial for preprocessing to improve data quality, reduce computational complexity, improve recognition accuracy, enhance robustness, and improve data fusion effect, thereby making fire source identification more accurate, efficient and reliable.

[0087] Further, the fire source is identified at the current suspected fire source location of the drone, corresponding to the above step S60; in the embodiment of the present application, a fire source point identification model is used to identify the fire source point, see Figure 2 ; The specific identification process includes:

[0088] The preprocessed image data are respectively input into the image input module and the numerical input module of the input layer;

[0089] Furthermore, the received data of different types are input into the image feature extraction module and the numerical feature extraction module of the feature extraction layer; the image features include: color features, shape features, texture features, temperature features, smoke features and dynamic features; the numerical features include: temperature data, humidity data and gas concentration data; the gas concentration master data mainly includes the following aspects: carbon dioxide, carbon monoxide, oxygen and sulfur dioxide, etc.;

[0090] In the process of identifying the fire source, the color features, shape features, texture features, temperature features, smoke features and dynamic features in the image data are extracted respectively; the temperature data, humidity data and gas concentration data in the numerical data are extracted respectively; comprehensive decision support is provided for multiple features, helping to more accurately assess the danger of the fire source and the treatment measures.

[0091] The image feature extraction module uses the ResNet-101 network model for feature extraction; the network model uses a 512-layer network structure, including: convolution layer, batch normalization layer, activation layer (such as ReLU) and residual block. The numerical feature module uses the principal component analysis method to obtain the corresponding features;

[0092] Furthermore, the extracted features are input into a fire feature weighting layer for weighting; the fire feature weighting layer includes image feature weighting and numerical feature weighting; as shown in the following expression formula:

[0093]

[0094] Wherein, WIF represents the weighted feature of the image feature weighting module; i Represents the weight of the i-th image feature; If i represents the i-th image feature; n represents the total number of image features; WVF represents the weighted feature of the numerical feature weighting module; ν j Vf represents the weight of the j-th numerical feature; j represents the jth numerical feature; m represents the total number of numerical features;

[0095] Furthermore, the image weighted features and the numerical weighted features are input into the fire identification layer; wherein, in the fire identification layer, the two different types of features are unified into the same dimension for splicing; then, the probability of the fire source point is obtained by the fully connected layer according to the feature weights, which is a binary classification problem; namely, whether it is a fire source point;

[0096] Furthermore, the recognition probability of the fire recognition layer is output by the output layer to determine whether the current suspected fire source point is a real fire source point.

[0097] In the embodiment of the present application, a fire source point recognition model is used to identify suspected fire source points; the model includes: an input layer, a feature extraction layer, a fire feature weighting layer, a fire recognition layer and an output layer; wherein the fire feature weighting layer weights the learned image features and numerical features respectively, and the main purpose of the weighting is to enable the model to better identify the fire source features, so as to improve the accuracy of identifying the real fire source point.

[0098] Furthermore, when the identification result is a real fire source point, the estimated position is adjusted according to the identification situation, the position of the fire source point is accurately determined, and the impact direction and impact intensity of the fire source point are recalculated, corresponding to the above step S70;

[0099] Furthermore, when the identification result is a non-fire source point, the probability of fire occurrence at the current location is calculated, and the initial fire source search path is optimized, corresponding to the above step S80;

[0100] Among them, the process of obtaining the probability of fire occurrence includes: obtaining the relevant data about vegetation data, environmental factor data and geographic information data; combining the historical fire data of the current location to form a potential fire data set; applying feature engineering to the data in the potential fire data set to obtain a potential fire feature set; and inputting the potential fire feature set into a logistic regression model to obtain the probability of fire occurrence.

[0101] In the embodiment of the present application, the probability of fire occurrence is also calculated for fire source points whose identification results are false; a logistic regression model is used to make predictions by combining the currently identified vegetation data, environmental factor data, and geographic information data with historical fire data; this comprehensive analysis helps to more comprehensively assess the probability of fire occurrence, thereby more accurately identifying high-risk areas; and by incorporating the identification results of false fire source points into the calculation of fire occurrence probability, false alarms and false warnings can be effectively reduced.

[0102] Furthermore, the initial fire source search path is optimized, and the specific process includes:

[0103] Collect map data of the forest; mark obstacles in the initial fire source search path according to the obstacle information recorded on the map data, including: the location, type and size of the obstacle; establish an obstacle repulsion field according to the marked content of the obstacle; wherein the function of the obstacle repulsion field is: Among them, P i represents the position vector of obstacle i; k i A is the constant related to obstacle type i; i represents the size factor of obstacle i; ||PP i || represents the Euclidean distance between the drone position and obstacle i; λ represents the power; ε represents a non-zero constant; σ i represents the standard deviation of the size of obstacle i; exp() represents the Gaussian function;

[0104] Combining the obstacle repulsion field with the fire source attraction vector field to obtain a total influence vector field;

[0105] In the embodiment of the present application, the generation of the initial fire source search path is optimized; the optimization process establishes an obstacle repulsion field according to the obstacle information marked in the map data of the mountain forest A; the obstacle repulsion field is combined with the fire source attraction vector field to generate a total influence vector field; the obstacle repulsion field can accurately reflect the distribution of obstacles on the map, avoiding the situation of crossing obstacles in path planning. Combined with the fire source attraction vector field, it can ensure that the path planning not only avoids obstacles, but also heads towards the fire source target, thereby improving the accuracy and effectiveness of the path.

[0106] Further, the total influence vector field is used to optimize the initial fire source search path to obtain the optimized fire source search path; the optimization process of the initial fire source search path using the total influence vector field includes: visually displaying the relevant area in the map data according to the initial fire source search path; dividing the relevant area into a plurality of uniform grids using a spatial grid method; calculating the total influence vector field gradient in each of the grids; wherein the calculation formula of the total influence vector field gradient is:

[0107]

[0108] in, The gradient vector representing the total influence vector field of each network;

[0109] Further, the influence field strength and direction of the suspected fire source point are evaluated according to the total influence vector field gradient of each grid; and the fire source search path is adjusted using a gradient descent algorithm according to the evaluation result;

[0110] In the embodiment of the present application, in the optimization of the initial fire source search path according to the total influence vector field, the corresponding map data is gridded according to the spatial grid method, and the influence vector field gradient in the grid is calculated, through the influence field strength and direction; according to the evaluation results, the fire source search path is adjusted using the gradient descent algorithm to obtain the optimal search path; by combining the spatial grid method with the gradient descent algorithm, accurate, dynamic and effective path planning can be achieved in the fire source search path optimization, thereby improving the overall effect and efficiency of the fire emergency response.

[0111] Furthermore, when the real fire source is identified, the search path optimization of the drone ends, and the optimal fire source search path and the identification status of each suspected fire source are obtained, corresponding to step S90.

[0112] In the embodiment of the present application, the fire source point positioning and path planning of the mountain forest A are realized by the present invention; mainly including the following points: first, the estimated position of the suspected fire source point of the mountain forest A is obtained; secondly, a fire source attraction vector field is established according to the suspected fire source point; wherein the fire source attraction vector field is obtained according to the position of the UAV, the position information of the suspected fire source point and the environmental information of the suspected fire source point; and an initial fire source search path is generated according to the influence intensity of the fire source attraction vector field; the influence intensity is obtained according to the terrain characteristics, vegetation characteristics and meteorological characteristics at the suspected fire source point combined with the spatial distance; this can significantly improve the accuracy and efficiency of the UAV path planning; A fire source point recognition model is used in the UAV search process, and the fire source point is identified by collecting image information and numerical information of the suspected fire source area; the model improves the accuracy of fire source point identification by weighting different types of data features; finally, in order to optimize the search path of the UAV, the present invention proposes a fire source search path optimization method by establishing an obstacle repulsion field according to the obstacle annotation information of the map data of the mountain forest A, and combining the obstacle repulsion field and the fire source attraction vector field to produce a total influence vector field, thereby realizing effective planning of the search path; the above methods can be combined to achieve accurate positioning of forest fire sources and path planning.

[0113] Embodiment 2:

[0114] In the above-mentioned embodiment 1, the scheme of the present invention is used to verify the effect; in order to further illustrate the effect of the present invention, in the embodiment of the present application, the fire source location path planning is performed on the mountain B, and the specific process is basically the same as that of the embodiment 1, including:

[0115] Step 1: Obtain the estimated location of the suspected fire source in Forest B and construct an estimated spatial coordinate set;

[0116] Step 2: Establish the fire source attraction vector field;

[0117] Step 3: Generate the initial fire source search path of the UAV according to the fire source attraction vector field; wherein the generation process of the initial fire source search path includes: obtaining the influence information of each suspected fire source point in the fire source attraction vector field; the influence information represents Wherein, l represents the estimated location of the suspected fire source; Represents the influencing direction; s represents the influencing intensity; calculates the direction vector according to the influencing direction, and normalizes the direction vector to obtain a normalized direction vector; calculates the search radius; the search radius is represented by: R=k*s; wherein R represents the search radius value; k represents the proportional constant; sets the search path step length to obtain the initial path point; after screening the generated initial path points for abnormal points, generates the initial fire source search path.

[0118] Step 4: Obtain relevant data of the current position of the drone through the device on the drone;

[0119] Step 5: preprocessing the relevant data to obtain standardized data;

[0120] Step 6: Input the standardized data into the fire source point identification model to obtain the fire source identification result;

[0121] Step 7: adjusting the estimated position of the suspected fire source point according to the fire source identification result, and re-obtaining the impact direction and impact intensity;

[0122] Step 8: Optimize the initial fire source search path to obtain an optimized fire source search path; the specific optimization process includes:

[0123] Collect map data of the forest; mark obstacles in the initial fire source search path according to the obstacle information recorded on the map data, including: the location, type and size of the obstacle; establish an obstacle repulsion field according to the marked content of the obstacle; wherein the function of the obstacle repulsion field is: Among them, P i represents the position vector of obstacle i; k i A is the constant related to obstacle type i; i represents the size factor of obstacle i; ||PP i || represents the Euclidean distance between the drone position and obstacle i; λ represents the power; ε represents a non-zero constant; σ i represents the standard deviation of the size of obstacle i; exp() represents the Gaussian function;

[0124] Combining the obstacle repulsion field with the fire source attraction vector field to obtain a total influence vector field;

[0125] The initial fire source search path is optimized using the total influence vector field, and the relevant area in the map data is visualized according to the initial fire source search path; the relevant area is divided into a plurality of uniform grids using a spatial grid method; the total influence vector field gradient in each grid is calculated; wherein the calculation formula of the total influence vector field gradient is:

[0126]

[0127] in, The gradient vector representing the total influence vector field within each network;

[0128] According to the total influence vector field gradient of each grid, the influence field strength and direction of the suspected fire source point are evaluated; according to the evaluation result, the fire source search path is adjusted by using the gradient descent algorithm to obtain the optimized fire source search path;

[0129] Step 9: Repeat step 4 to step 8, and stop when the suspected fire source is identified as the real fire source;

[0130] Step 10: Obtain the optimal fire source search path for Forest B and the identification of each suspected fire source point.

[0131] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for locating and planning a fire source point of an unmanned aerial vehicle based on machine learning, characterized in that: The following steps are involved: Step 1: Obtain the estimated location of the suspected fire source in the forest and construct an estimated spatial coordinate set; Step 2: Establish the fire source attraction vector field, whose function is expressed as: in, represents the function of the fire source attraction vector field; P represents the position vector of the UAV; FP j represents the estimated position vector of the suspected fire source point j; g j Indicates the intensity coefficient of the suspected fire source point j; B j represents the influence factor of suspected fire source j; θ j represents the standard deviation of the influence range of the suspected fire source j; q represents the power of the distance; represents a non-zero constant; represents the unit vector calculation function; α() represents the adjustment factor of the mountain forest environment characteristics; Step 3: generating an initial fire source search path of the UAV according to the fire source attraction vector field; Step 4: obtaining relevant data of the current position of the drone through the device on the drone; the relevant data includes: image data and numerical data; Step 5: preprocessing the relevant data to obtain standardized data; Step 6: Input the standardized data into the fire source point identification model to obtain the fire source identification result; Step 7: adjusting the estimated position of the suspected fire source point according to the fire source identification result, and re-obtaining the impact direction and impact intensity; Step 8: Optimizing the initial fire source search path to obtain an optimized fire source search path; Step nine: repeating the steps four to eight, and stopping when the suspected fire source is identified as the real fire source; Step 10: Obtain the optimal fire source search path and the identification of each suspected fire source point.

2. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1 is characterized in that: The generation process of the initial fire source search path includes: obtaining the influence information of each suspected fire source point in the fire source attraction vector field; the influence information represents Wherein, l represents the estimated location of the suspected fire source; Represents the influencing direction; s represents the influencing intensity; calculates the direction vector according to the influencing direction, and normalizes the direction vector to obtain a normalized direction vector; calculates the search radius; the search radius is represented by: R=k*s; wherein R represents the search radius value; k represents the proportional constant; sets the search path step length to obtain the initial path point; after screening the generated initial path points for abnormal points, generates the initial fire source search path.

3. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1 is characterized in that: The process of obtaining the impact intensity includes: obtaining terrain features, vegetation features and meteorological features; calculating the spatial distance according to the estimated position of the suspected fire source point; the calculation formula of the spatial distance is: Among them, E i,j> represents the spatial distance between the suspected fire source point i and the suspected fire source point j; (x i ,y i ,z i ) represents the spatial coordinates of the suspected fire source point i; (x j ,y j ,z j ) represents the spatial coordinates of the suspected fire source point j; the intensity influence constant is obtained according to the historical fire occurrence records, and the basic influence intensity is calculated; the calculation formula of the basic influence intensity is: in, represents the basic impact intensity of the suspected fire source i; γ i represents the influence intensity constant of suspected fire source i; The environmental adjustment factor is defined according to the acquired terrain features, the vegetation features and the meteorological features; wherein the calculation formula of the environmental adjustment factor is: F env =F slope *F vegetation *F weather Among them, F env represents the environmental adjustment factor; F slope Represents the slope factor; F vegetation Represents vegetation factor; F weather represents the meteorological factor; the impact intensity is calculated according to the environmental adjustment factor; the calculation formula of the impact intensity is: Among them, s i Represents the impact intensity of the suspected fire source i.

4. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1 is characterized in that: The image data includes: original image, depth image and multi-spectral image; the numerical data includes: geographic coordinate data, vegetation data, meteorological data and environmental data.

5. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1 is characterized in that: The preprocessing includes: image data preprocessing and numerical data preprocessing; wherein, the image data preprocessing includes: removing noise in the image data to obtain denoised image data; performing uniform size cropping on the denoised image data to obtain standard scale image data; performing enhancement on the standard scale image data to obtain enhanced image data; the numerical data preprocessing includes: processing missing values ​​of the numerical data to obtain filled numerical data; normalizing the supplemented numerical data to obtain normalized numerical data; performing feature selection on the normalized numerical data to obtain standard numerical data.

6. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1 is characterized in that: The fire source point recognition model includes: an input layer: including an image input module and a numerical input module, respectively used to receive the image data and numerical data in the pre-processed related data; a feature extraction layer includes an image feature extraction module and a numerical feature extraction module, respectively used to extract image features and numerical features from the input data of the input layer; a fire feature weighting layer includes an image feature weighting module and a numerical feature weighting module, respectively used to weight the features of the image features and the numerical features that are conducive to fire recognition; wherein the weighting formula of the fire feature weighting layer is: Wherein, WIF represents the weighted feature of the image feature weighting module; i Represents the weight of the i-th image feature; If i represents the i-th image feature; n represents the total number of image features; WVF represents the weighted feature of the numerical feature weighting module; ν j Vf represents the weight of the j-th numerical feature; j represents the jth numerical feature; m represents the total number of numerical features; The fire identification layer is used to identify fires based on image weighted features and numerical weighted features; the output layer is used to output the fire source identification results.

7. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 6 is characterized in that: The image features include: color features, shape features, texture features, temperature features, smoke features and dynamic features; the numerical features include: temperature data, humidity data and gas concentration data.

8. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1, characterized in that: In the step five, when the fire source identification result is no, the probability of fire occurrence at the current location is obtained; wherein the specific process of obtaining the fire occurrence probability includes: obtaining the relevant data about vegetation data, environmental factor data and geographic information data; combining the historical fire data of the current location to form a potential fire data set; applying feature engineering to the data in the potential fire data set to obtain a potential fire feature set; and inputting the potential fire feature set into a logistic regression model to obtain the probability of fire occurrence.

9. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 1, characterized in that: The process of step eight includes: collecting map data of the forest; marking obstacles in the initial fire source search path according to the obstacle information recorded on the map data, including: the location, type and size of the obstacle; establishing an obstacle repulsion field according to the marked content of the obstacle; wherein the function of the obstacle repulsion field is: Among them, P i represents the position vector of obstacle i; k i A is the relevant constant representing the type of obstacle i; i represents the size factor of obstacle i; ||PP i || represents the Euclidean distance between the drone position and obstacle i; λ represents the power; ε represents a non-zero constant; σ i represents the standard deviation related to the size of obstacle i; exp() represents a Gaussian function; the obstacle repulsion field is combined with the fire source attraction vector field to obtain a total influence vector field; the initial fire source search path is optimized using the total influence vector field to obtain the optimized fire source search path.

10. The method for locating a fire source point of an unmanned aerial vehicle based on machine learning according to claim 9, characterized in that: The optimization process of the initial fire source search path using the total influence vector field includes: visually displaying the relevant area in the map data according to the initial fire source search path; dividing the relevant area into a plurality of uniform grids using a spatial grid method; calculating the total influence vector field gradient of each grid; wherein the calculation formula of the total influence vector field gradient is: in, The gradient vector representing the total influence vector field within each network; The influence field strength and direction of the suspected fire source point are evaluated according to the total influence vector field gradient of each grid; and the fire source search path is adjusted using the gradient descent algorithm according to the evaluation result.

Images