An intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher

Abstract

The present invention provides an intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher, belonging to the field of intelligent navigation fire extinguishing technology. The method of the present invention first collects fire scene data through multi-source sensors to construct a positive and negative fire development matrix; then, it performs route planning based on a pre-trained deep learning model. The model uses a special convolution kernel structure to extract fire scene features and optimizes decisions through a mathematical calculation module; the system updates the fire matrix in real time, dynamically adjusts the fire extinguishing strategy, and finally completes the precise fire extinguishing task through time-aligned control instructions. This method achieves efficient control of autonomous flying fire extinguishers in complex environments through data-driven and intelligent decision-making, solving the technical problem in the prior art that autonomous flying fire extinguishers are difficult to achieve precise fire extinguishing path planning and fire extinguishing strategy control in complex fire environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent navigation fire extinguishing, and in particular relates to an intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher. Background Art

[0002] With the acceleration of urbanization and industrial development, the frequency and severity of fire accidents are increasing. Traditional firefighting methods rely primarily on manual operations by firefighters, which not only endangers the lives of firefighters but also makes it difficult to effectively and promptly control the spread of fire. In recent years, autonomous flying fire extinguishers, as a new type of firefighting equipment, have played an increasingly important role in fire emergency response due to their high mobility and rapid response. Traditional autonomous flying fire extinguishers primarily use preset routes or manual remote control to carry out firefighting operations. They use various sensors to obtain fire scene information and make firefighting decisions based on pre-set rules.

[0003] However, current autonomous flying fire extinguishers have many limitations in practical applications. First, existing technologies generally use simple temperature thresholds or image recognition methods to determine the fire extinguishing location, which cannot accurately grasp the development trend of the fire, resulting in low fire extinguishing efficiency. Second, traditional route planning methods are mainly based on path optimization algorithms in static environments. They fail to fully consider the dynamic changes in the fire scene and are difficult to adapt to complex and changing fire environments. Third, existing fire extinguishing strategy control methods often use fixed fire extinguishing agent delivery patterns, lack targeted responses to different fire characteristics, and result in a waste of fire extinguishing resources.

[0004] In complex fire scenarios, the spread of fire is highly uncertain and dynamic, affected by a variety of factors such as wind direction, temperature, and the distribution of combustible materials. Existing technologies have difficulty effectively integrating and optimizing decisions based on multi-source sensor data, fire development predictions, and firefighting resource constraints, making it impossible to achieve accurate firefighting path planning and firefighting strategy control. This results in autonomous flying fire extinguishers facing problems such as inaccurate positioning, suboptimal paths, and improper control in actual applications, which seriously affects the firefighting effect. In other words, there is a technical problem in the existing technology that autonomous flying fire extinguishers find it difficult to achieve accurate firefighting path planning and firefighting strategy control in complex fire environments. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher, which can solve the technical problem in the prior art that autonomous flying fire extinguishers are difficult to achieve accurate fire extinguishing path planning and fire extinguishing strategy control in complex fire environments.

[0006] The present invention is implemented as follows: The present invention provides an intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher, comprising the following steps: obtaining sensor data of the autonomous flying fire extinguisher, constructing a fire positive development matrix based on the sensor data, calculating a fire negative development matrix based on the fire positive development matrix, establishing a fire extinguishing strategy equation group based on the fire positive development matrix and the fire negative development matrix, calculating a fire extinguishing measure vector using the fire extinguishing strategy equation group, establishing a fire scene space division grid and assigning fire intensity weight values ​​based on the fire positive development matrix, performing route planning using a pre-trained fire extinguishing path planning model, calculating a flight trajectory curve and generating a flight control instruction sequence based on the route planning result, generating a fire extinguishing control instruction sequence based on the fire extinguishing measure vector, performing time alignment on the flight control instruction sequence and the fire extinguishing control instruction sequence to generate an execution instruction queue, controlling the autonomous flying fire extinguisher to perform a fire extinguishing task, updating the fire positive development matrix and the fire negative development matrix in real time through the sensor data, determining whether to replan the route based on the updated matrix, and returning to the platform after completing the fire extinguishing task.

[0007] Among them, the sensor data includes thermal imaging sensor data, visible light sensor data, three-band flame sensor data, temperature and humidity sensor data, smoke sensor data, and carbon monoxide sensor data; the thermal imaging sensor data is collected using a 60 Hz sampling frequency, the visible light sensor data is collected using a 30 frames per second frame rate, and the three-band flame sensor data is collected using a 100 Hz sampling frequency.

[0008] Among them, the positive fire development matrix represents the fire spread speed, fire expansion direction, and fire intensity; the negative fire development matrix represents the fire decay speed, fire contraction direction, and fire weakening degree after the fire extinguishing agent is released.

[0009] Among them, the pre-trained fire-fighting path planning model adopts a convolutional neural network structure with an embedded mathematical calculation module. The mathematical calculation module contains a fire spread prediction equation, a resource consumption evaluation equation, and a route cost calculation equation, which is used to integrate the fire development situation and fire-fighting resource constraints to perform route optimization calculations.

[0010] Among them, the input of the fire spread prediction equation includes fire scene temperature distribution data, ambient wind speed data, combustible material distribution data, and terrain height data; the input of the resource consumption assessment equation includes fire scene temperature data, fire scene area data, fire extinguishing agent storage data, fire extinguisher remaining power data, and target fire point distance data; the input of the route cost calculation equation includes route length data, fire scene danger data, fire extinguishing efficiency prediction data, energy consumption data, passage difficulty data, and task completion time data.

[0011] Among them, the basic size of the fire scene space division grid is 5 meters × 5 meters, and a grid size of 1 meter × 1 meter is used in areas where the fire is strong or changes greatly; the fire weight value is calculated using the hierarchical analysis method, with the weight coefficient of the temperature value being 0.3, the weight coefficient of the fire intensity being 0.25, the weight coefficient of the spread speed being 0.2, the weight coefficient of the terrain characteristics being 0.15, and the weight coefficient of the surrounding important targets being 0.1.

[0012] Among them, the flight trajectory curve is calculated using the Bezier curve interpolation algorithm, the maximum speed of the autonomous flying fire extinguisher is set to 15 meters per second, the maximum acceleration is set to 2 meters per square second, and the maximum turning angle is set to 45 degrees; the flight control instructions include three-dimensional position coordinates, flight speed, flight attitude and hovering state.

[0013] Among them, the fire extinguishing control instruction sequence includes the working status of the high-voltage generator, the trigger timing of the electromagnetic throwing controller and the flow control of the fine water mist pump; the control of the high-voltage generator adopts pulse width modulation, the control of the electromagnetic throwing controller adopts timing triggering, and the control of the fine water mist pump adopts variable frequency speed regulation.

[0014] Among them, when any of the following conditions are met: the fire intensity changes by more than 30%, the spread direction changes by more than 45 degrees, or the fire area changes by more than 20%, the replanning program is initiated; the replanning program uses the Monte Carlo simulation method to simulate the mission execution process 100 times, and the route is replanned when the mission completion rate falls below 85%.

[0015] Among them, when the autonomous flying fire extinguisher returns to the platform, when the distance from the platform is greater than 100 meters, satellite navigation is used, and the navigation accuracy is better than 5 meters; when the distance is between 30 and 100 meters, laser navigation is used, and the navigation accuracy is better than 0.5 meters; when the distance is less than 30 meters, Bluetooth beacon navigation is used, and the navigation accuracy is better than 0.1 meters.

[0016] Compared to existing technologies, this invention provides an intelligent fire-fighting strategy control method for autonomous flying fire extinguishers. This method, based on deep learning and multi-source information fusion, accurately describes and predicts the dynamic characteristics of fire activity by constructing positive and negative fire development matrices. This method leverages a pre-trained fire-fighting path planning model, combined with real-time sensor data, to dynamically generate optimal routing and fire-fighting delivery strategies, significantly improving the operational efficiency and accuracy of autonomous flying fire extinguishers.

[0017] At the technical level, this invention utilizes an innovative convolutional neural network architecture, leveraging three specialized convolution kernels: edge detection, direction perception, and intensity assessment, to achieve multi-dimensional extraction of fire scene characteristics. Furthermore, a mathematical calculation module integrates and optimizes fire development trends with firefighting resource constraints, ensuring the scientific and feasible nature of route planning. By updating the fire status matrix in real time and dynamically adjusting firefighting strategies, the system effectively overcomes the limited adaptability of traditional methods in complex environments.

[0018] This invention successfully solves the existing technical problem of autonomous flying fire extinguishers being unable to achieve precise fire-fighting path planning and fire-fighting strategy control in complex fire environments. By building a complete perception-decision-execution chain, accurate prediction of fire development, intelligent planning of fire-fighting paths, and optimal allocation of fire-fighting resources are achieved. The system possesses adaptive learning capabilities and can continuously optimize decision-making models based on historical data, improving its response capabilities in different fire scenarios. This data-driven intelligent control method provides a new technical path for improving the operational efficiency of autonomous flying fire extinguishers. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention.

[0020] Figure 2 Graph showing the changing trends of fire scene temperature and carbon monoxide concentration over time in the embodiment.

[0021] Figure 3 Graph showing utilization rates and coverage areas of three fire extinguishing resources in the embodiment. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] like Figure 1 FIG. 1 is a flow chart of an intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher provided by the present invention. The method includes the following steps:

[0024] S01. Acquire sensor data of the autonomous flying fire extinguisher, where the sensor data includes thermal imaging sensor data, visible light sensor data, three-band flame sensor data, temperature and humidity sensor data, smoke sensor data, and carbon monoxide sensor data.

[0025] S02. Construct a fire forward development matrix based on the sensor data, where the fire forward development matrix represents the fire spread speed, fire expansion direction, and fire intensity.

[0026] S03. Calculate a negative fire development matrix based on the positive fire development matrix, where the negative fire development matrix represents the fire decay speed, fire shrinkage direction, and fire reduction degree after the fire extinguishing agent is released.

[0027] S04. Establish a fire extinguishing strategy equation group based on the fire positive development matrix and the fire negative development matrix. The fire extinguishing strategy equation group is used to calculate the single-time fire extinguishing duration of a single machine and the multi-time fire extinguishing duration of multiple machines.

[0028] S05. Calculate the fire extinguishing delivery measure vectors of the autonomous flying fire extinguisher at different waypoints using the fire extinguishing strategy equation group. The fire extinguishing delivery measure vectors include the number of fire extinguishing bombs delivered, the number of fire extinguishing bottles delivered, and the duration of water mist spraying.

[0029] S06. Establish a fire scene space division grid based on the fire forward development matrix, and assign a fire intensity weight value to each grid node.

[0030] S07. Use a pre-trained fire extinguishing path planning model to perform route planning on the grid of the fire scene space to obtain a route planning result, wherein the route planning result includes coordinates of multiple waypoints, an arrival order of the waypoints, and a hovering time of the waypoints.

[0031] S08. Calculate a flight trajectory curve of the autonomous flying fire extinguisher according to the route planning result, and decompose the flight trajectory curve into a flight control instruction sequence.

[0032] S09. Generate a fire extinguishing control instruction sequence according to the fire extinguishing measure vector, wherein the fire extinguishing control instruction sequence is used to control a high-voltage generator, an electromagnetic throwing controller, and a fine water mist pump.

[0033] S10: aligning the flight control instruction sequence with the fire extinguishing control instruction sequence to generate an execution instruction queue.

[0034] S11. Control the autonomous flying fire extinguisher to complete the fire extinguishing task according to the execution instruction queue.

[0035] S12. Update the fire positive development matrix and the fire negative development matrix in real time using the sensor data.

[0036] S13. According to the updated fire positive development matrix and the fire negative development matrix, determine whether it is necessary to re-plan the route and generate a new execution instruction queue.

[0037] S14. After completing the fire extinguishing task, control the autonomous flying fire extinguisher to return to the autonomous flying fire extinguisher platform through laser navigation, Bluetooth beacon navigation or satellite navigation.

[0038] The firefighting path planning model uses a convolutional neural network as its foundational model, with a mathematical calculation module added between the convolutional and fully connected layers. This module integrates fire development trends with firefighting resource constraints to optimize the route. This module includes equations for predicting fire spread, assessing resource consumption, and calculating route costs.

[0039] The fire spread prediction equation is used to calculate the fire spread range and speed. The input includes fire temperature distribution data, ambient wind speed data, combustible material distribution data, and terrain height data. The output is a sequence of fire spread boundary coordinates at future moments.

[0040] The resource consumption evaluation equation is used to estimate the use of fire-fighting resources. The input includes fire scene temperature data, fire scene area data, fire extinguishing agent storage data, fire extinguisher remaining power data, and target fire point distance data. The output is the predicted value of fire-fighting resource consumption.

[0041] The route cost calculation equation is used to calculate the comprehensive cost of different route plans. The input includes route length data, fire scene danger data, fire extinguishing efficiency prediction data, energy consumption data, passage difficulty data, and task completion time data. The output is the comprehensive cost score of the route plan.

[0042] The fire extinguishing path planning model includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a mathematical calculation module, a fully connected layer, and an output layer. The input layer receives fire scene grid data, which includes temperature distribution data, fire intensity data, and combustible material distribution data. The first convolutional layer uses an edge detection convolution kernel group generated by a fire scene boundary feature extraction function constructed based on the temperature gradient change rate, fire intensity change rate, and combustible material density change rate. The second convolutional layer uses a direction-aware convolution kernel group generated by a fire spread direction identification function constructed based on heat transfer direction and ambient wind direction. The third convolutional layer uses an intensity assessment convolution kernel group generated by a fire intensity analysis function constructed based on thermal imaging data, flame spectrum data, and smoke concentration data.

[0043] The weight coefficient of each convolution kernel in the edge detection convolution kernel group is determined by a fire boundary clarity evaluation function, the inputs of which include boundary temperature gradient, boundary infrared characteristics, and boundary visible light characteristics. The weight coefficient of each convolution kernel in the direction perception convolution kernel group is determined by a fire spread trend evaluation function, the inputs of which include heat diffusion rate, wind field data, and terrain factors. The weight coefficient of each convolution kernel in the intensity assessment convolution kernel group is determined by a fire hazard evaluation function, the inputs of which include flame temperature, thermal imaging intensity, and flame spectral characteristics.

[0044] The mathematical calculation module receives the feature map output of the third convolutional layer, and calculates the optimal route plan in combination with the fire positive development matrix and the fire negative development matrix.

[0045] The fully connected layer converts the output of the mathematical calculation module into a waypoint coordinate sequence, where the waypoint coordinate sequence includes a plane coordinate value, a flight altitude value, and an arrival time value.

[0046] The output layer generates a complete route planning result, which includes the coordinates of the take-off point, the coordinates of the landing point, the coordinates of the waypoints, and the mission termination conditions.

[0047] The steps of establishing the fire extinguishing path planning model training data set include: obtaining historical fire data, the historical fire data including fire scene temperature distribution data, fire spread speed data, environmental wind speed data, and terrain data; obtaining historical fire extinguishing action data, the historical fire extinguishing action data including fire extinguisher route data, fire extinguishing placement location data, and fire extinguishing effect evaluation data; gridding the historical fire data to generate fire scene grid feature data; and performing trajectory analysis on the historical fire extinguishing action data to generate optimal route example data.

[0048] The training data set is subjected to data enhancement processing, comprising the following sub-steps: performing a rotation transformation on the fire scene grid feature data to generate fire distribution samples in different directions; performing a scale transformation on the fire scene grid feature data to generate fire scene distribution samples of different scales; and performing an interpolation processing on the optimal route example data to generate continuous route trajectory samples.

[0049] Training the fire extinguishing path planning model includes the following sub-steps: inputting the fire scene grid feature data into the fire extinguishing path planning model; using the optimal route example data as a training target; calculating the error between the route planning result output by the fire extinguishing path planning model and the training target; and updating the weight coefficient of each layer of the convolution kernel in the fire extinguishing path planning model according to the error.

[0050] Verifying the fire extinguishing path planning model includes the following sub-steps: evaluating the route planning performance of the fire extinguishing path planning model using a verification data set; calculating the route planning performance indicators, wherein the route planning performance indicators include a route length score, a fire extinguishing efficiency score, and a safety score; and optimizing and adjusting the fire extinguishing path planning model according to the route planning performance indicators.

[0051] The positive fire development matrix is ​​used to represent the spread of a fire. Each element in the matrix represents the fire intensity, spread speed, and direction at the corresponding location. The negative fire development matrix is ​​used to represent the fire decay after the application of the extinguishing agent. Each element in the matrix represents the fire decay rate, fire shrinkage direction, and fire extinguishing effectiveness evaluation value at the corresponding location.

[0052] The fire suppression measures vector represents the specific combination of measures used to extinguish a fire at a specific location. The elements in the vector correspond to the usage and duration of different types of fire suppression resources. The fire intensity weight represents the danger level and fire suppression priority of different locations within the fire scene. A higher weight indicates a higher priority for fire suppression at that location.

[0053] The execution instruction queue represents the complete sequence of instructions for the autonomous flying fire extinguisher to execute a fire-fighting mission, including a unified scheduling sequence for flight control instructions and fire-fighting operation instructions. The fire boundary feature extraction function characterizes the geometric and physical characteristics of the fire boundary, determining its precise location and morphological characteristics by analyzing temperature gradients, infrared images, and visible light images. The fire spread direction identification function identifies the primary and secondary directions of fire spread and predicts the fire's development trend based on environmental factors. The fire intensity analysis function assesses the fire intensity levels in different areas of the fire scene, providing a basis for hazard assessment for route planning. The fire boundary clarity evaluation function evaluates the accuracy of fire boundary identification and guides the optimization of the weight parameters of the edge detection convolution kernel. The fire spread trend evaluation function assesses the speed and direction of fire spread and guides the optimization of the weight parameters of the direction-aware convolution kernel. The fire hazard evaluation function assesses the hazard level of each area of ​​the fire scene and guides the optimization of the weight parameters of the intensity assessment convolution kernel. The fire scene grid feature data is used to represent discretized fire scene status information, including the temperature, fire intensity, and spread speed of each grid cell. The optimal route example data represents the most effective route plans from historical firefighting operations, serving as a reference for model training. The route planning performance indicators are used to evaluate the quality of the route plans generated by the model, comprehensively evaluating them based on three dimensions: route length, firefighting efficiency, and safety.

[0054] The resource utilization rate is used to indicate the efficiency of using fire-fighting resources, and is calculated by the ratio of fire-fighting effect to resource consumption per unit time.

[0055] The specific implementation of the above steps is described in detail below.

[0056] Step S01 is implemented through multi-sensor data acquisition and preprocessing. When acquiring sensor data, each sensor must first be calibrated and calibrated. The thermal imaging sensor acquires data at a sampling rate of 60 Hz, with an image resolution of 640×480 pixels and a temperature detection range of -40°C to 1500°C. The visible light sensor acquires data at a 1080p HD resolution and a frame rate of 30 frames per second. The three-band flame sensor simultaneously monitors flame characteristics in the infrared, ultraviolet, and visible light bands, with a sampling frequency of 100 Hz. The temperature and humidity sensor acquires data with a temperature measurement range of -50°C to 150°C and a humidity measurement range of 0 to 100% relative humidity, at a sampling frequency of 1 Hz. The smoke sensor acquires data with a detection range of 0 to 1000 ppm and a response time of less than 10 seconds. The carbon monoxide sensor acquires data with a detection range of 0 to 1000 ppm, a response time of less than 15 seconds, and a sampling frequency of 1 Hz. The collected sensor data is subjected to noise removal using a median filter algorithm, with the sliding window size set to 5 data points. The data is then smoothed using an interpolation smoothing algorithm. Finally, all sensor data are uniformly converted into a standard format and time-synchronized.

[0057] The specific implementation of step S02 involves constructing a fire forward development matrix based on sensor data, achieved through a multi-source data fusion algorithm. First, image registration is performed using thermal imaging data and visible light data. An image registration algorithm based on feature point matching is employed. Image feature points are extracted using a scale-invariant feature transformation algorithm. Feature point matching is performed using a nearest neighbor ratio matching algorithm, with a matching threshold set to 0.7. The registered image data is then spatially and temporally aligned with the other sensor data to establish a unified data coordinate system. The multi-source data is then fused using a Kalman filter algorithm. The state vector contains three components: fire spread speed, fire spread direction, and fire intensity. The fire spread speed is obtained by calculating the change in the fire boundary position between adjacent moments. The fire spread direction is determined by principal component analysis, calculating the main direction of change at the fire boundary. Fire intensity is calculated by comprehensively analyzing thermal imaging intensity, flame spectral characteristics, and smoke concentration. Finally, the fused data is organized into a fire forward development matrix, with each element of the matrix corresponding to a discrete grid point in the fire space.

[0058] The specific implementation of step S03 involves calculating the negative fire development matrix based on the positive fire development matrix, using a backpropagation algorithm. First, a fire extinguishing agent action model is established. The model inputs include the fire extinguishing agent type, dosage, delivery method, and environmental parameters, and the output is a predicted value for fire extinguishing effectiveness. For each type of fire extinguishing agent, a separate action characteristic model is established: the action characteristics of fire extinguishing bombs are described using an explosion shock wave model, the action characteristics of fire extinguishing bottles are described using a jet diffusion model, and the action characteristics of water mist are described using a droplet atomization model. The fire extinguishing agent action model is then convolved with the positive fire development matrix to calculate the fire decay rate, fire contraction direction, and degree of fire reduction. Finally, the calculation results are organized into a negative fire development matrix, with the matrix structure corresponding to the positive fire development matrix. The fire extinguishing effectiveness is evaluated using a fuzzy comprehensive evaluation method, with evaluation factors including temperature drop rate, fire area reduction rate, and smoke dissipation rate. The evaluation levels are categorized into four levels: significant, good, fair, and poor.

[0059] The specific implementation method of step S04 is to establish a fire extinguishing strategy equation group based on the fire positive development matrix and the fire negative development matrix, and implement it using a dynamic programming algorithm. First, a single-machine single-time fire extinguishing time calculation model is established. The model takes into account factors such as fire intensity, fire area, fire extinguishing agent type and wind speed, and adopts a multiple linear regression method to establish a calculation formula. Then, a multi-machine multiple-time fire extinguishing time calculation model is established. The model takes into account factors such as the number of machines, synergy and resource constraints, and adopts a nonlinear optimization method to solve the optimal fire extinguishing time. Finally, the single-machine and multi-machine models are combined into a complete fire extinguishing strategy equation group, and the solution of the equation group adopts an iterative calculation method. The calculation result of the fire extinguishing time serves as an important constraint condition for subsequent route planning, and is also used to evaluate the deployment plan of fire extinguishing resources.

[0060] The specific implementation method of step S05 is to use the fire extinguishing strategy equations to calculate the fire extinguishing measure vectors for different waypoints, and adopt a genetic algorithm to achieve optimal solution. First, the waypoints are divided into several fire extinguishing operation areas. The fire characteristics of each area are determined by the data of the corresponding position in the fire forward development matrix. Then, the appropriate fire extinguishing agent combination is selected based on the fire characteristics. The choice of fire extinguishing agent is based on a fuzzy decision-making method, and the decision-making factors include fire intensity, fire scene environment, and cost-effectiveness. For each waypoint, the optimal delivery amount of different types of fire extinguishing agents is calculated. The calculation of the delivery amount of fire extinguishing bombs takes into account the explosion range and fire intensity. The calculation of the delivery amount of fire extinguishing bottles takes into account the spray coverage area and fire scene size. The calculation of the fine water mist spray duration takes into account the atomization effect and wind speed. Finally, the calculation results are organized into a fire extinguishing measure vector, in which each component corresponds to the usage amount or duration of a fire extinguishing resource. The efficiency of fire extinguishing resource utilization is evaluated using a resource utilization index, which is calculated using an input-output ratio analysis method.

[0061] The specific implementation of step S06 involves establishing a spatial grid for the fire scene based on the fire forward development matrix, using an adaptive gridding algorithm. First, the basic grid size is determined based on the actual extent of the fire scene, with the initial grid size set to 5 meters by 5 meters. The grid is then adaptively refined based on the fire distribution characteristics, with smaller grid sizes used in areas with stronger or more variable fires. The minimum grid size after refinement is 1 meter by 1 meter. A fire weight is assigned to each grid node. This weight is calculated using the Analytic Hierarchy Process (AHP), taking into account factors such as temperature, fire intensity, spread rate, terrain characteristics, and the distribution of nearby important targets. The weight coefficient for temperature is 0.3, the weight coefficient for fire intensity is 0.25, the weight coefficient for spread rate is 0.2, the weight coefficient for terrain characteristics is 0.15, and the weight coefficient for nearby important targets is 0.1. Finally, the grid node weights are normalized to a range between 0 and 1.

[0062] The specific implementation of step S07 utilizes a pre-trained fire-fighting path planning model to divide the fire scene into grids for route planning, using a deep reinforcement learning algorithm. First, the fire scene grid data is input into the convolutional neural network component of the pre-trained model, and fire scene features are extracted through three layers of convolution operations. The first layer of convolution uses a 3×3 convolution kernel for edge detection, the second layer uses a 5×5 convolution kernel to identify the direction of fire spread, and the third layer uses a 7×7 convolution kernel to assess fire intensity. The convolution features are then input into a mathematical calculation module, and route optimization is performed based on the fire development situation and fire-fighting resource constraints. Route optimization utilizes an improved ant colony algorithm, using fire intensity weights as heuristic information and route length and fire-fighting efficiency as evaluation indicators. Finally, a route planning result is generated, including the coordinates of multiple waypoints, the order of arrival of the waypoints, and the hovering time of the waypoints. The number of waypoints is determined according to the scale of the fire and the requirements of the fire-fighting task, and is generally set to 10 to 20 points. The distance between adjacent waypoints does not exceed 50 meters. The hovering time of each waypoint is calculated based on the fire-fighting measures vector at that point.

[0063] The specific implementation method of step S08 is to calculate the flight trajectory curve of the autonomous flying fire extinguisher based on the route planning results, and implement it using the Bezier curve interpolation algorithm. First, the waypoint coordinate sequence is used as the control point to construct a cubic Bezier curve, and the smoothness of the curve is adjusted by the tangent vector of the control point. Then, the dynamic constraints of the aircraft, including the maximum speed, maximum acceleration and maximum turning angle, are considered to correct the Bezier curve. The maximum speed of the aircraft is set to 15 meters per second, the maximum acceleration is set to 2 meters per square second, and the maximum turning angle is set to 45 degrees. Finally, the corrected flight trajectory curve is discretized into a sequence of flight control instructions, which include three-dimensional position coordinates, flight speed, flight attitude and hovering state. The sampling period of the flight control instructions is set to 0.1 seconds to ensure the smoothness and executable nature of the flight trajectory.

[0064] The specific implementation method of step S09 is to generate a fire extinguishing control instruction sequence based on the fire extinguishing measure vector, and implement it using a state machine control algorithm. First, the fire extinguishing measure vector is parsed into specific execution actions, including the working state of the high-voltage generator, the trigger timing of the electromagnetic throwing controller, and the flow control of the fine water mist pump. Then, the control strategy is designed according to the working characteristics of different fire extinguishing equipment. The control of the high-voltage generator adopts pulse width modulation, the control of the electromagnetic throwing controller adopts timed triggering, and the control of the fine water mist pump adopts variable frequency speed regulation. Finally, a fire extinguishing control instruction sequence is generated, which includes specific operations such as equipment startup, operating parameter adjustment, and equipment shutdown. The execution process of the fire extinguishing control instruction adopts real-time monitoring and feedback control to ensure the fire extinguishing effect.

[0065] The specific implementation of step S10 involves aligning the flight control instruction sequence with the fire extinguishing control instruction sequence using a task scheduling algorithm. First, a unified time base is established, unifying the timestamps of the two instruction sequences into the same coordinate system. Then, based on the order and parallel relationships of task execution, the instruction sequences are organized into a directed acyclic graph structure. Finally, a critical path algorithm is used to calculate the optimal execution order and generate an execution instruction queue. Each instruction in the execution instruction queue contains three elements: execution time, execution content, and execution conditions. Instruction execution is event-driven, ensuring coordination and consistency between flight control and fire extinguishing control.

[0066] The specific implementation of step S11 involves controlling the autonomous flying fire extinguisher to complete the fire extinguishing mission according to the execution instruction queue, implemented using a hierarchical control architecture. At the top of the control system is the task management layer, responsible for parsing execution instructions and monitoring task status, using a finite state machine to implement task flow control. The middle layer is the trajectory tracking layer, responsible for aircraft motion control. It employs a model predictive control algorithm to accurately track the planned trajectory, with a control cycle set to 50 milliseconds, achieving position control accuracy better than 0.5 meters and attitude control accuracy better than 2 degrees. The bottom layer is the execution control layer, responsible for direct control of various actuators, including motor control, servo control, and fire extinguishing equipment control. This layer utilizes a cascaded PID control algorithm with an inner loop control cycle of 2 milliseconds and an outer loop control cycle of 10 milliseconds. During mission execution, the control system collects real-time flight status information and fire extinguishing effectiveness information, and performs adaptive adjustments based on this information to ensure reliable mission completion. If an abnormal situation is detected, the control system initiates emergency response procedures, including hovering and waiting, obstacle avoidance, and emergency return.

[0067] The specific implementation method of step S12 is to update the positive development matrix and the negative development matrix of the fire situation in real time through sensor data, and adopt the data processing method of sliding time window to achieve it. First, set the update cycle to 1 second, and collect the latest data of all sensors in each update cycle. Then use the data fusion algorithm to process the multi-source sensor data, including three steps: data preprocessing, feature extraction and information fusion. Data preprocessing adopts low-pass filtering and outlier detection method, the filter cutoff frequency is set to 10 Hz, and the outlier judgment threshold is set to 3 times the standard deviation. Feature extraction adopts wavelet transform method to perform time-frequency analysis on sensor data, select db4 wavelet as the basis function, and set the number of decomposition layers to 4 layers. Information fusion adopts evidence theory method to construct the basic probability distribution function, and set the credibility threshold to 0.8. Finally, the fire development matrix is ​​updated according to the fusion result. The update process adopts exponential weighted average method, and the weight coefficient of historical data is set to 0.7.

[0068] The specific implementation of step S13 involves determining whether rerouting is necessary and generating a new execution instruction queue based on the updated positive and negative fire development matrices. This is achieved using a dynamic decision-making algorithm. First, the change in the fire development matrix is ​​calculated, using a matrix difference method to separately calculate the changes in fire intensity, spread direction, and fire extent. Replanning trigger conditions are then set, including a change in fire intensity exceeding 30%, a change in spread direction exceeding 45 degrees, or a change in fire extent exceeding 20%. When any of these trigger conditions is met, the replanning process is initiated. The replanning process first evaluates the effectiveness of the current route. Using a Monte Carlo simulation method, 100 mission executions are simulated and the mission completion rate is calculated. If the mission completion rate falls below 85%, replanning is determined to be necessary. The replanning process utilizes a heuristic search algorithm guided by the principle of minimization, preserving as much of the original route as possible while performing local optimization on ineffective portions. Finally, a new execution instruction queue is generated, ensuring a smooth transition between the old and new instruction queues.

[0069] Step S14 specifically controls the autonomous flying fire extinguisher's return to the platform, using a multi-source navigation fusion algorithm. First, the return path is calculated based on the current position and the platform's position. Path planning is performed using the A-star algorithm, taking into account obstacle avoidance and energy constraints. Next, the appropriate navigation method is selected. Satellite navigation is prioritized when the distance to the platform is greater than 100 meters, with an accuracy better than 5 meters. Laser navigation is used when the distance is between 30 and 100 meters, with an accuracy better than 0.5 meters. When the distance is less than 30 meters, Bluetooth beacon navigation is used, with an accuracy better than 0.1 meters. A smooth transition strategy is used to switch between different navigation methods, with a 10-meter overlap interval. During the return flight, flight control utilizes an adaptive control algorithm, dynamically adjusting flight parameters based on factors such as battery charge and wind speed. Upon approaching the platform, the precision landing phase begins, employing visual servoing to identify feature markers on the platform and achieve centimeter-level landing control. Landing speed is controlled to be below 0.5 meters per second, and the descent speed upon contact does not exceed 0.2 meters per second. After landing, the shutdown procedure is automatically executed, including equipment status check, data saving and system shutdown.

[0070] The calculation process or matrix involved in the present invention is described in detail below.

[0071] The fire forward development matrix F forward The specific expressions are as follows:

[0072]

[0073] Each element f ij Expressed as: f ij =α1v ij +α2d ij +α3s ij +ε1;

[0074] Where: v ij is the fire spread speed at the grid point (i, j), in m / s; d ij is the fire expansion angle, in rad; s ij is the fire intensity value, dimensionless; α1, α2, α3 are weight coefficients, determined by expert experience, satisfying α1+α2+α3=1; ε1 is the error term, ranging from 0 to 0.1.

[0075] Fire spread speed v ij The calculation formula is:

[0076] Where: x ij ,y ij is the coordinate of the fire boundary point, obtained through thermal imaging image processing; t is the time variable.

[0077] Fire spread angle d ij The calculation formula is:

[0078] Fire intensity value s ij The calculation formula is: ij =β1T ij +β2I ij +β3C ij +ε2;

[0079] Where: T ij is the temperature value, in °C; I ij is the infrared image intensity value, ranging from 0 to 255; C ij is the carbon monoxide concentration value in ppm; β1, β2, β3 are weight coefficients, satisfying β1+β2+β3=1; ε2 is the error term, ranging from 0 to 0.1.

[0080] The fire forward development matrix describes the spatial distribution of a fire, achieved through multi-source sensor data fusion. The fire spread rate is calculated using partial derivatives, reflecting the rate of change of the fire boundary over time. The fire spread direction is calculated using the inverse tangent function, accurately describing the angular direction of fire spread. Fire intensity is linearly weighted, taking into account three key indicators: temperature, infrared intensity, and carbon monoxide concentration.

[0081] The fire negative development matrix F backward The specific expressions are as follows:

[0082] Each element b ij Expressed as:

[0083] b ij =γ1r ij +γ2c ij +γ3e ij +ε3;

[0084] Where: r ij is the fire decay rate, in 1 / s; c ij is the fire contraction angle, in rad; e ij is the fire extinguishing effect evaluation value, dimensionless; γ1, γ2, γ3 are weight coefficients, satisfying γ1+γ2+γ3=1; ε3 is the error term, ranging from 0 to 0.1.

[0085] Fire decay rate r ij The calculation formula is:

[0086] Where: T ij is the temperature value in °C; t is the time variable.

[0087] Fire contraction angle c ij The calculation formula is:

[0088] Fire extinguishing effect evaluation value e ij The calculation formula is:

[0089] Where: ΔT ij , T ij0 are the temperature drop value and the initial temperature value, respectively, in °C; ΔA ij , A ij0 are the fire area reduction value and the initial area value, respectively, in m 2 ;ΔS ij , S ij0 are the smoke concentration reduction value and initial concentration value, respectively, in ppm; δ1, δ2, δ3 are weight coefficients, satisfying δ1+δ2+δ3=1; ε4 is the error term, ranging from 0 to 0.1.

[0090] The negative fire development matrix describes the changing characteristics of a fire scene after the application of an extinguishing agent. The fire decay rate, expressed as a logarithmic derivative, reflects the relative rate of temperature change over time. The direction of fire contraction is expressed as the opposite of the direction of fire expansion by adding a π phase difference. Fire extinguishing effectiveness is evaluated using a relative rate of change to eliminate the influence of different dimensions.

[0091] The fire extinguishing measures vector M is specifically expressed as follows: M = [n b , n e , t w ] T ;

[0092] Where: n b is the number of fire extinguishing bombs deployed; n e is the number of fire extinguisher bottles released; t w It is the duration of water mist spraying, in seconds.

[0093] Number of fire extinguishing bombs released n b The calculation formula is:

[0094] Where: A f is the fire area, in m 2 ; A b is the effective fire extinguishing area of ​​a single fire extinguishing bomb, in m 2 ;T f is the average temperature of the fire scene, in °C; T0 is the reference temperature, which is 500 °C; v w is the wind speed, in m / s; v0 is the reference wind speed, which is 5 m / s; k1 and k2 are correction coefficients, which are 0.2 and 0.3 respectively; Indicates rounding up.

[0095] Number of fire extinguisher bottles deployed n e The calculation formula is:

[0096] Where: A e is the effective fire extinguishing area of ​​a single fire extinguisher, in m 2 ;H f is the flame height, in meters; H0 is the reference height, which is 2 meters; k3 and k4 are correction coefficients, which are 0.2 and 0.25 respectively.

[0097] Water mist spraying time t w The calculation formula is:

[0098] Where: V f is the volume of the fire scene, in m 3 ;Q w is the water mist spray flow rate, in m 3 / s;S f is the smoke concentration in ppm; S0 is the reference concentration, which is 500 ppm; k5 and k6 are correction coefficients, which are 0.15 and 0.2 respectively.

[0099] The firefighting measure vector describes the optimal allocation of different types of firefighting resources. The number of fire extinguishers and bottles deployed takes into account the effects of temperature and wind speed, rounding up to ensure effective firefighting. The water mist spray duration is calculated based on volume flow rate, taking into account the effects of temperature and smoke concentration.

[0100] The fire weight value w ij The specific expressions are as follows:

[0101]

[0102] Where: T ij , T max are the grid point temperature value and the maximum temperature value, respectively, in °C; s ij , s max are the fire intensity value and the maximum intensity value, dimensionless; v ij , w max are the spreading speed value and the maximum speed value, respectively, in m / s; h ij is the terrain elevation factor, ranging from 0 to 1; d ij is the important target distance factor, ranging from 0 to 1; η1, η2, η3, η4, η5 are weight coefficients, satisfying ε5 is the error term, ranging from 0 to 0.1.

[0103] The route cost calculation equation is specifically expressed as follows:

[0104] C path =λ1L path +λ2R risk +λ3E eff +λ4P energy +λ5D diff +λ6T task +ε6;

[0105] Where: L path is the route length, in meters; R risk is the fire danger degree, dimensionless; E eff is the fire extinguishing efficiency, dimensionless; P energy is energy consumption, in kW·h; D diff is the difficulty of passage, dimensionless; T task is the task completion time, in seconds; λ1, λ2, λ3, λ4, λ5, λ6 are weight coefficients, satisfying ε6 is the error term, ranging from 0 to 0.1.

[0106] The fire spread prediction equation is specifically expressed as follows:

[0107] P spread =μ1T dist +μ2V wind +μ3M fuel +μ4H terrain +ε7;

[0108] Where: T dist is the fire temperature distribution function; V wind is the ambient wind speed function; M fuel is the combustible material distribution function; H terrain is the terrain height function; μ1, μ2, μ3, μ4 are weight coefficients, satisfying ε7 is the error term, ranging from 0 to 0.1.

[0109] The resource consumption evaluation equation is specifically expressed as follows:

[0110] R consume =ρ1T f +ρ2A f +ρ3Q s +ρ4E r +ρ5D t +ε8;

[0111] Where: T f is the fire scene temperature, in °C; A f is the fire area, in m 2 ;Q sis the storage capacity of fire extinguishing agent, in kg; E r is the remaining power of the fire extinguisher, in kW·h; D t is the distance to the target fire point, in meters; ρ1, ρ2, ρ3, ρ4, ρ5 are weight coefficients, satisfying ε8 is the error term, ranging from 0 to 0.1.

[0112] The resource utilization calculation formula is specifically expressed as follows:

[0113]

[0114] Where: E effect is the fire extinguishing effect evaluation value, dimensionless; R consume is the resource consumption, unit is kg; t task The task execution time, in seconds.

[0115] The state vector is specifically expressed as follows:

[0116] x=[v spread , d expand , s intensity ] T ;

[0117] Where: v spread is the fire spread speed, in m / s; d expand The direction of fire expansion, in rad; s intensity is the fire intensity, dimensionless.

[0118] The following is a detailed description of the process of deriving and establishing each matrix or equation.

[0119] 1. Fire positive development matrix F forward Derivation of:

[0120] First, establish a coordinate system and divide the fire scene into m×n grids. Each grid point needs to represent three characteristics: spread speed, expansion direction, and fire intensity. Therefore, a three-dimensional matrix is ​​constructed:

[0121] F forward ={V, D, S};

[0122] The velocity matrix V is:

[0123]

[0124] The direction matrix D is:

[0125]

[0126] The intensity matrix S is:

[0127]

[0128] Through multi-sensor data fusion, the Kalman filter algorithm is used to update the matrix elements:

[0129] Equation of state: X k =AX k-1 +BU k +W k ;

[0130] Observation equation: Z k =HX k +V k ;

[0131] Where: X k is the state vector; A is the state transfer matrix; B is the control matrix; U k is the control vector; W k is the process noise; Z k is the observation vector; H is the observation matrix; V k is the observation noise.

[0132] 2. Fire negative development matrix F backward Derivation of:

[0133] Considering the effect of fire extinguishing agent, a fire extinguishing agent action model is established:

[0134] Fire extinguisher bomb explosion shock wave model:

[0135] Where: p(r, t) is the pressure value at time t at distance r; p0 is the explosion center pressure; R0 is the characteristic radius; α is the attenuation coefficient.

[0136] Fire extinguisher spray diffusion model:

[0137] Where: c(r, t) is the concentration value at time t at a distance r; Q is the release amount; D is the diffusion coefficient.

[0138] Water mist atomization model: d 32 =A·We -0.6 ·Re 0.15 ;

[0139] Where: d 32 is the Sauter mean diameter; We is the Weber number; Re is the Reynolds number; A is the experimental constant.

[0140] 3. Derivation of fire extinguishing measures vector M:

[0141] Optimization solution based on genetic algorithm:

[0142] Fitness function:

[0143] Where: E effect For fire extinguishing effect; E max is the maximum effect value; C cost is the resource cost; C max is the maximum cost; ω1, ω2 are weight coefficients.

[0144] Crossover operator: M new =φM1+(1-φ)M2;

[0145] Where: M1, M2 are parent individuals; φ is the random crossover coefficient.

[0146] Mutation operator: M mut =M+σN(0,1);

[0147] Where: σ is the variable step length; N(0, 1) is a standard normal distribution random number.

[0148] 4. Fire weight value w ij Derivation of:

[0149] The weight coefficient is determined by using the hierarchical analysis method:

[0150] Construct a judgment matrix:

[0151]

[0152] Where: a ij is the importance ratio of the i-th indicator to the j-th indicator.

[0153] Calculate the weight vector: w = [η1, η2, ..., η n ] T ;

[0154] Verify consistency:

[0155] Where: max is the maximum eigenvalue; RI is the random consistency index.

[0156] 5. Derivation of route cost calculation equation:

[0157] Using a multi-objective optimization approach:

[0158] Objective function vector: F(x) = [f1(x), f2(x), ..., f6(x)] T ;

[0159] Where: f1(x) is the route length; f2(x) is the hazard level; f3(x) is the fire extinguishing efficiency; f4(x) is the energy consumption; f5(x) is the difficulty of passage; and f6(x) is the completion time.

[0160] Constraints:

[0161] g i (x)≤0, i=1, 2,...,p;

[0162] h j (x)=0, j=1, 2,...,q;

[0163] The optimal weight coefficient is obtained through Pareto optimization solution.

[0164] 6. Derivation of fire spread prediction equation:

[0165] Modeling based on partial differential equations:

[0166] Heat conduction equation:

[0167] Where: α is the thermal diffusion coefficient; Q(x, y, t) is the heat source term.

[0168] Wind field influence equation:

[0169] Where: V is the wind speed vector; p is the pressure; ρ is the density; v is the kinematic viscosity; F is the external force term.

[0170] 7. Derivation of resource consumption evaluation equation:

[0171] Based on the energy balance principle:

[0172] Energy input:

[0173] Energy consumption: E out =E flight +E extinguish +E loss ;

[0174] Where: E flight is the flight energy consumption; E extinguish is the energy consumption for fire extinguishing; E loss is the loss item.

[0175] Specifically, the core principle of this invention is to achieve intelligent decision-making and control of the fire-fighting process in complex fire environments by establishing a mathematical description model of the fire situation and combining it with deep learning technology. First, a positive fire development matrix and a negative fire development matrix are constructed using multi-source sensor data. These two matrices describe the characteristics of fire spread and fire extinguishing effectiveness, respectively, providing a mathematical basis for subsequent decision-making optimization. The positive matrix characterizes the dynamic characteristics of fire development by analyzing parameters such as temperature gradient and fire intensity; the negative matrix predicts the fire decay pattern by evaluating the effectiveness of fire extinguishing agents.

[0176] Based on this, the present invention designed a unique convolutional neural network architecture, incorporating three functional convolution kernels: edge detection, direction perception, and intensity assessment. These kernels are responsible for extracting fire boundary features, identifying the direction of fire spread, and assessing fire intensity, respectively, to form a comprehensive perception of the fire scene. By introducing a mathematical calculation module into the network, quantitative analysis of fire development trends and firefighting resource constraints is achieved, ensuring that the generated route plan meets both firefighting efficiency requirements and resource constraints.

[0177] This invention employs a modular design approach, breaking down the complex problem of fire control into three key components: perception, planning, and execution. It then achieves coordinated optimization of these components through a data-driven approach. By continuously learning from historical firefighting data and optimizing the parameters of the decision-making model, the system improves its adaptability to diverse fire scenarios. This approach, combining traditional control theory with modern artificial intelligence technology, offers a new technical approach for solving the problem of precise firefighting in complex environments.

[0178] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0179] The specific implementation method of step S01 is to first calibrate the sensor of the autonomous flying fire extinguisher. The thermal imaging sensor is calibrated using a high-temperature blackbody as the reference. A mapping relationship is established by comparing the output values ​​at different temperatures. The calibration temperature range is -40°C to 1500°C, with a sampling frequency of 60Hz, an image resolution of 640×480 pixels, and a grayscale resolution of 12 bits. The visible light sensor is calibrated using a standard light source and color card, with a resolution of 1080p, a frame rate of 30fps, and a dynamic range of 90dB. The three-band flame sensor is calibrated for the infrared band of 2~5μm, The calibration is carried out in the ultraviolet band of 180-260nm and the visible light band of 400-700nm, with a sampling frequency of 100Hz; the temperature and humidity sensor is calibrated in a constant temperature and humidity chamber, with a temperature measurement range of -50℃ to 150℃, a relative humidity measurement range of 0 to 100%, and a sampling frequency of 1Hz; the smoke sensor is calibrated using standard gas, with a detection range of 0 to 1000ppm and a response time of less than 10 seconds; the carbon monoxide sensor is calibrated using standard gas, with a detection range of 0 to 1000ppm, a response time of less than 15 seconds, and a sampling frequency of 1Hz.

[0180] The collected sensor data is then preprocessed. First, a median filter algorithm is used to remove noise. The sliding window size is set to 5 data points. The data in each window is sorted by size and the median value is taken as the output. The data is then smoothed using a cubic spline interpolation algorithm. Local extreme points are selected as interpolation nodes to ensure the continuity of the first and second derivatives of the interpolation curve. Finally, all sensor data are uniformly converted to a standard format and time-synchronized. Linear interpolation is used to align data with different sampling frequencies to a unified time base, with the highest sampling frequency of 60Hz being selected as the time base.

[0181] After preprocessing, the data undergoes a quality assessment, with established criteria for data validity: a signal-to-noise ratio of no less than 20dB, a data integrity rate of no less than 95%, and a time synchronization error of no more than 10ms. If a sensor's data quality fails to meet these requirements, a sensor self-check program is initiated and data compensation is performed. This data compensation utilizes a Kalman filter algorithm, predicting missing values ​​based on historical data and physical models. After compensation, the data undergoes a reassessment of quality until it meets the requirements.

[0182] The specific implementation method of step S02 is to first perform image registration on the thermal imaging data and the visible light data, and use a feature point matching algorithm based on scale-invariant feature transformation to extract significant feature points in the image. The feature point correspondence is established through the nearest neighbor ratio matching method, and the matching threshold is set to 0.7. Then, a random consistency sampling algorithm is used to eliminate erroneous matching points, and the affine transformation matrix is ​​calculated to achieve image alignment.

[0183] Then, the registered image data is aligned with other sensor data in time and space to establish a unified data coordinate system. The origin of the coordinate is selected at the center of the fire scene, with the x-axis pointing due east, the y-axis pointing due north, and the z-axis pointing due up. The aligned data is processed using a multi-source data fusion algorithm, specifically the Kalman filter algorithm. The state vector contains three components: fire spread speed, fire expansion direction, and fire intensity. According to the formula in the article: Fire spread speed Obtained by calculating the change in the fire boundary position at adjacent moments; the direction of fire expansion The main change direction of the fire boundary is determined by calculating the principal component analysis method; the fire intensity s ij =β1T ij +β2I ij +β3C ij +ε2 is calculated by comprehensively analyzing thermal imaging intensity, flame spectrum characteristics and smoke concentration.

[0184] Finally, the fire positive development matrix F is constructed forward , each element of the matrix corresponds to a discrete grid point in the fire scene space, and the element value f ij =α1v ij +α2dij +α3s ij +ε1 represents the fire development characteristics at that point. Weight coefficients α1, α2, and α3 are determined using the Analytic Hierarchy Process (AHP). First, a judgment matrix is ​​constructed, then eigenvectors are calculated to obtain weights. Finally, a consistency check is performed, with the consistency ratio required to be less than 0.1. The error term ε1 ranges from 0 to 0.1 and represents model uncertainty.

[0185] The specific implementation of step S03 is to first establish a fire extinguishing agent action model, and establish its action characteristic model for different types of fire extinguishing agents. For fire extinguishing bombs, the explosion shock wave model is used to describe its action characteristics, and the waveform equation is The pressure attenuation coefficient α is determined by experiment. For the fire extinguisher, the spray diffusion model is used to describe its action characteristics, and the concentration distribution equation is: The diffusion coefficient D is determined experimentally; for water mist, the droplet atomization model is used to describe its action characteristics, and the Sauter mean diameter is calculated as d 32 =A·We -0.6 ·Re 0.15 , the coefficient a is calibrated through experiments.

[0186] Then the fire extinguishing agent action model is convolved with the fire forward development matrix to calculate the fire decay rate Fire shrinkage direction and fire reduction Finally, the fire negative development matrix F is constructed based on the calculation results. backward , where each element b ij =γ1r ij +γ2c ij +γ3e ij +ε3 represents the fire decay characteristics of the point. The weight coefficients γ1, γ2, and γ3 are determined by the fuzzy analytic hierarchy process, and the error term ε3 ranges from 0 to 0.1.

[0187] The specific implementation of step S04 begins by establishing a single-machine, single-shot fire extinguishing duration calculation model. This model considers factors such as fire intensity, fire area, extinguishing agent type, and wind speed, and uses a multivariate linear regression method to develop a calculation formula. Input variables include average fire temperature, fire area, extinguishing agent type code, and ambient wind speed. The output variable is the extinguishing duration. The regression coefficient is determined using the least squares method. The model's coefficient of determination must be greater than 0.85, and the root mean square error must be less than 10%.

[0188] A model was then developed to calculate the duration of multiple firefighting attempts involving multiple aircraft, taking into account factors such as the number of aircraft, synergy, and resource constraints. The synergy coefficient was determined through simulation experiments. Resource constraints included fire extinguishing agent reserves, battery charge, and flight time. A nonlinear optimization method was used to determine the optimal firefighting duration. The objective function was to minimize the total firefighting time, and the constraints included resource and safety constraints. The optimization algorithm used sequential quadratic programming, with a convergence accuracy of 0.001.

[0189] Finally, the single- and multi-machine models were combined into a complete set of firefighting strategy equations, using a piecewise function to describe the firefighting strategies under different conditions. The equations were solved using an iterative calculation method, with a maximum number of iterations set to 100 and a convergence threshold of 0.01. The results served as constraints for subsequent route planning and were also used to evaluate firefighting resource deployment plans.

[0190] The specific implementation of step S05 first divides the waypoints into firefighting zones. The fire characteristics of each zone are determined by the data at the corresponding position in the fire forward development matrix. This division is based on the spatial distribution of fire intensity, and a clustering algorithm is used to automatically achieve this division. The number of cluster centers is dynamically determined based on the fire size. The appropriate extinguishing agent combination is then selected based on the fire characteristics using a fuzzy decision-making method, with decision-making factors including fire intensity, fire environment, and cost-effectiveness. A fuzzy rule base is established, and the Mamdan I inference method is used to determine the extinguishing agent selection result.

[0191] For each waypoint, the optimal amount of fire extinguishing agents of different types is calculated according to the formula in the article: Number of fire extinguishing bombs released Consider the explosion range and fire intensity; the number of fire extinguisher bottles to be placed Consider the spray coverage area and fire size; the duration of water mist spraying Considering the atomization effect and wind speed, each coefficient is determined through experiments, and the value range of correction coefficients k1 to k6 is 0.1 to 0.3.

[0192] Finally, the calculation results are organized into a fire extinguishing measure vector M = [n b , n e , t w ] T , the genetic algorithm is used to optimize the best combination. The fitness function is The population size is set to 100, the number of iterations is 50, the crossover probability is 0.8, and the mutation probability is 0.1. The efficiency of firefighting resources is measured by the resource utilization index Conduct an assessment.

[0193] The specific implementation of step S06 is to first calculate the fire forward development matrix F forwardThe fire's scope was determined, and the fire's boundaries were extracted using an edge detection algorithm. The fire's spatial structure was then divided into a basic grid, with an initial grid size of 5 meters by 5 meters. The grid was then adaptively refined based on the fire's distribution characteristics, with smaller grid sizes used in areas with stronger or more variable fires. The minimum grid size after refinement was 1 meter by 1 meter. Grid refinement was determined based on the temperature gradient, rate of change of fire intensity, and spread rate. Grid refinement was initiated when these parameters exceeded set thresholds: a temperature gradient threshold of 50°C / meter, an intensity change threshold of 0.2 seconds, and a spread rate threshold of 2 meters per second.

[0194] Assign a fire intensity weight value to each grid node. The weight value is calculated using the hierarchical analysis method. According to the formula in the article Consider factors such as temperature, fire intensity, spread rate, terrain characteristics, and the distribution of surrounding important targets. Weight coefficients η1 to η5 are determined by constructing a judgment matrix, and the consistency ratio of the judgment matrix must be less than 0.1. The weight coefficient of temperature is 0.3, the weight coefficient of fire intensity is 0.25, the weight coefficient of spread rate is 0.2, the weight coefficient of terrain characteristics is 0.15, and the weight coefficient of surrounding important targets is 0.1. Terrain elevation factor h ij Calculated from the digital elevation model, ranging from 0 to 1, the important target distance factor d ij Calculated using a Gaussian decay function, ranging from 0 to 1.

[0195] Finally, the weights of the grid nodes are normalized so that they are distributed between 0 and 1. Range normalization is used to ensure comparability and consistency of the weights. The spatial distribution of the normalized weights is analyzed, and the spatial autocorrelation coefficient of the weights is calculated to assess the rationality of the weight allocation. If the spatial autocorrelation coefficient is too high, the weight allocation is too concentrated, and the weight calculation parameters need to be adjusted. If the spatial autocorrelation coefficient is too low, the weight allocation is too dispersed, and appropriate adjustments are also required.

[0196] The specific implementation of step S07 first inputs the fire scene grid data into the convolutional neural network portion of the pre-trained fire extinguishing path planning model. The network structure includes an input layer, three convolutional layers, and a fully connected layer. The input layer receives the fire scene grid data, including temperature distribution data, fire intensity data, and combustible material distribution data. The first convolution layer uses a 3×3 convolution kernel for edge detection, using a ReLU activation function with a step size of 1 and a padding of SAME. The second convolution layer uses a 5×5 convolution kernel to identify the direction of fire spread, using a leaky ReLU activation function with a step size of 2. The third convolution layer uses a 7×7 convolution kernel to assess fire intensity, using a Sigmoid activation function with a step size of 2.

[0197] The convolution features are then input into the mathematical calculation module to optimize the route based on the fire development situation and fire fighting resource constraints. The mathematical calculation module contains three key equations: the fire spread prediction equation P spread =μ1T dist +μ2V wind +μ3M fuel +μ4H terrain +ε7, used to predict the development trend of the fire scene; resource consumption evaluation equation R consume =ρ1T f +ρ2A f +ρ3Q s +ρ4E r +ρ5D t +ε8, used to estimate resource usage; route cost calculation equation C path =λ1L path +λ2R risk +λ3E eff +λ4P enerygy +λ5D diff +λ6T task +ε6, used to evaluate different route options.

[0198] Route optimization utilizes an improved ant colony algorithm, using fire intensity weights as heuristic information and route length and fire extinguishing efficiency as evaluation metrics. The ant colony algorithm parameters are set as follows: an initial pheromone concentration of 1, a pheromone volatility coefficient of 0.1, a global update coefficient of 0.2, a local update coefficient of 0.1, and 200 iterations. In each iteration, the next waypoint is selected based on a state transition probability formula, with a heuristic factor α of 1 and an expected heuristic factor β of 2. Finally, a route planning result is generated, including the coordinates of multiple waypoints, the order of arrival at the waypoints, and the hovering time at the waypoints. The number of waypoints is determined based on the scale of the fire and the requirements of the firefighting task, and is generally set to 10 to 20 points, with the distance between adjacent waypoints not exceeding 50 meters.

[0199] The specific implementation of step S08 is to first construct a cubic Bezier curve based on the route planning result. Each curve segment is determined by four control points, and the curve parameter equation is P(t)=(1-t) 3 P0+3t(1-t) 2 P1+3t 2 (1-t)P2+t 3 P3, where t ranges from 0 to 1. Control points are selected based on the smoothness and continuity of the curve. First-order derivative continuity is required at the junctions of adjacent curve segments. The tangent vector of the curve is determined by minimizing the rate of change of curvature to ensure a smooth flight trajectory.

[0200] The Bezier curve is then modified to take into account the aircraft's dynamic constraints. These constraints include a maximum speed of 15 meters per second, a maximum acceleration of 2 meters per second squared, and a maximum turn angle of 45 degrees. This modification process utilizes a quadratic programming approach, with the objective function being to minimize the integral of the trajectory curvature while satisfying the dynamic constraints. Constraints are handled using a penalty function approach, converting the degree of constraint violation into a penalty term and adding it to the objective function.

[0201] Finally, the corrected flight trajectory is discretized into a sequence of flight control commands, including three-dimensional position coordinates, flight velocity, flight attitude, and hovering state. Discretization uses uniformly spaced sampling with a sampling period of 0.1 seconds. The velocity and acceleration of the sampling points are checked to ensure that dynamic constraints are met. If a constraint violation is detected, local corrections are made by adjusting the position of the sampling points until all constraints are met.

[0202] The specific implementation of step S09 first parses the fire extinguishing action vector into specific execution actions, including the operating status of the high-voltage generator, the triggering sequence of the electromagnetic injection controller, and the flow control of the water mist pump. The high-voltage generator uses pulse width modulation (PWM) with a modulation frequency of 1 kHz and a duty cycle range of 20% to 80%. The electromagnetic injection controller uses a timed triggering method with a trigger pulse width of 10 milliseconds, and the trigger interval is determined by the injection rate. The water mist pump uses variable frequency speed regulation with a frequency range of 20 Hz to 50 Hz.

[0203] Control strategies were designed based on the operating characteristics of different fire-fighting equipment. A state machine model was established for each device, including initialization, standby, operating, fault, and shutdown states. State transition conditions were determined based on device parameters and mission requirements, and control parameters for each state were obtained through table lookup. The state machine cycle was 10 milliseconds to ensure real-time control. Finite state machines were used to implement device logic control, and state transition diagrams were described using Moore-type state machines.

[0204] Finally, a fire extinguishing control command sequence is generated, encompassing specific operations such as device startup, operating parameter adjustment, and device shutdown. The control command format includes a timestamp, device number, operation type, parameter value, and a checksum. Command execution utilizes real-time monitoring and feedback control, dynamically adjusting control parameters based on execution results. Monitoring parameters include device operating status, output parameters, and fault information, with a sampling period of 1 millisecond.

[0205] The specific implementation of step S10 first establishes a unified time base, unifying the timestamps of the flight control command sequence and the fire extinguishing control command sequence into the same coordinate system. This time base uses GPS timing with a clock accuracy of better than 1 microsecond. The two command sequences are time-aligned, using linear interpolation to process data with different sampling periods. Time synchronization error must be controlled within 1 millisecond.

[0206] The instruction sequence is then organized into a directed acyclic graph (DAG) based on the execution order and parallel relationships of the tasks. Nodes in the graph represent specific execution instructions, and edges represent dependencies between instructions. A topological sorting algorithm is used to determine the execution order of the instructions to ensure that dependencies are met. For instructions that can be executed in parallel, a parallel scheduling strategy is used to improve execution efficiency. During the scheduling process, interactions between devices must be considered to avoid conflicts.

[0207] Finally, a critical path algorithm is used to calculate the optimal execution order and generate an execution instruction queue. This critical path calculation takes into account the execution time and resource usage of the instructions and is solved using dynamic programming. Each instruction in the execution instruction queue contains three elements: execution time, execution content, and execution conditions. Instruction execution is event-driven. Predictive execution technology is used to improve response speed for conditional judgments and branch execution. A double-buffering mechanism is used to manage the queue to ensure continuous execution of instructions.

[0208] The specific implementation of step S11 utilizes a hierarchical control architecture to implement mission control for the autonomous flying fire extinguisher. The control system is divided into three layers: the mission management layer, the trajectory tracking layer, and the execution control layer. The mission management layer utilizes a finite state machine to control the mission flow. The state machine includes five main states: mission initialization, route planning, fire extinguishing execution, status monitoring, and mission completion. State transitions are triggered by conditions such as mission instructions, environmental changes, and abnormal events. The mission management layer has a control cycle of 500 milliseconds and is primarily responsible for parsing execution instructions and monitoring mission status.

[0209] The trajectory tracking layer uses the model predictive control algorithm to achieve accurate tracking of the planned trajectory, and the control period is set to 50 milliseconds. The prediction model takes into account the dynamic characteristics of the aircraft, and the state equation is The state vector x includes position, velocity, and attitude, and the control input u includes thrust and torque. The prediction horizon is set to 20 steps, and the control horizon is set to 10 steps. The objective function consists of trajectory tracking error, control input, and state change, and adopts a quadratic form. Constraints include state constraints and control constraints, and a soft constraint processing method is used. The controller achieves position control accuracy better than 0.5 meters and attitude control accuracy better than 2 degrees.

[0210] The execution control layer uses a cascaded PID control algorithm to directly control various actuators, including motor control, servo control, and fire extinguishing equipment control. The inner loop has a 2 millisecond control cycle and is responsible for attitude angular velocity control; the outer loop has a 10 millisecond control cycle and is responsible for attitude angle control. PID parameters are adjusted in real time using online identification methods, and the adaptive law uses the MIT principle. The execution layer also includes fault detection and handling functions, automatically initiating appropriate emergency response procedures when an abnormality is detected.

[0211] The specific implementation of step S12 utilizes a sliding time window data processing method with a window length of 10 seconds and a sliding step size of 1 second. During each update cycle, the latest data from all sensors is collected, including thermal imaging data, visible light data, three-band flame data, temperature and humidity data, smoke data, and carbon monoxide data. Data preprocessing utilizes low-pass filtering and outlier detection, with a filter cutoff frequency set to 10 Hz. Outlier determination is based on the 3σ criterion.

[0212] Feature extraction was performed on the preprocessed multi-source sensor data, and time-frequency analysis was performed using wavelet transform. The db4 wavelet was selected as the basis function, and the decomposition level was set to 4. The extracted features included energy, statistical, and morphological features. Energy features reflect the intensity distribution of the signal, statistical features describe the statistical patterns of the data, and morphological features characterize the signal's changing characteristics. The feature extraction results were used to update the fire development matrix.

[0213] Information fusion uses evidence theory to construct a basic probability distribution function. The credibility of each sensor's data is calculated, with a credibility threshold of 0.8. Dempster's combination rule is used to synthesize multi-source information to generate a fusion result. The positive and negative fire development matrices are updated based on the fusion results. The update process uses an exponentially weighted average method, with a weight coefficient of 0.7 for historical data. The updated matrices are then checked for consistency to ensure data reliability.

[0214] The specific implementation of step S13 is to first calculate the change in the fire development matrix, and use the matrix difference method to calculate the change in fire intensity, spread direction and fire range. The calculation formula for the change is ΔF = F k -F k-1 , where k represents the current moment. Then, set the replanning trigger conditions: a change in fire intensity exceeding 30%, a change in the direction of spread exceeding 45 degrees, or a change in the fire area exceeding 20%. When any of these trigger conditions is met, the replanning process is initiated.

[0215] The replanning process first evaluates the effectiveness of the current route using Monte Carlo simulation. The simulations are run 100 times, with random perturbations added to each run, including wind speed, temperature, and position. The magnitude of the perturbations is determined based on actual observations. The mission completion rate (the probability of successfully completing the firefighting mission) is calculated. If the mission completion rate falls below 85%, replanning is deemed necessary.

[0216] The replanning process utilizes a heuristic search algorithm guided by the principle of minimization. First, the validity of each waypoint in the current route is evaluated, retaining those that remain valid. For those that are no longer valid, local path planning methods are used for optimization. Optimization objectives include minimizing route length, maximizing firefighting efficiency, and ensuring safety margins. Finally, a new execution instruction queue is generated, ensuring a smooth transition between the old and new instruction queues. During this transition, a trajectory smoothing algorithm is used to avoid sudden changes.

[0217] The specific implementation of step S14 first calculates the return path based on the current position and the platform position. Path planning is performed using the A-star algorithm, taking into account obstacle avoidance and energy constraints. The heuristic function is h(n) = w1d(n) + w2e(n), where d(n) is the distance term and e(n) is the energy consumption term. The weighting coefficients are determined experimentally. Path smoothing uses cubic spline interpolation to ensure curvature continuity.

[0218] The appropriate navigation method is selected based on the distance from the platform: satellite navigation is preferred when the distance is greater than 100 meters, with an accuracy better than 5 meters; laser navigation is used when the distance is between 30 and 100 meters, with an accuracy better than 0.5 meters; and Bluetooth beacon navigation is used when the distance is less than 30 meters, with an accuracy better than 0.1 meters. A smooth transition strategy is used between different navigation methods, with a 10-meter overlap interval. Navigation data is integrated using an adaptive Kalman filter algorithm, with the filter gain adjusted in real time based on measurement noise.

[0219] During the return flight, an adaptive control algorithm is used for flight control, dynamically adjusting flight parameters based on factors such as battery charge and wind speed. When approaching the platform, the aircraft enters the precision landing phase, employing visual servoing control. Landing control with centimeter-level precision is achieved by identifying characteristic markers on the platform. Feature extraction utilizes a combination of Hough transform and edge detection. Landing speed is controlled below 0.5 meters per second, with the descent speed upon contact not exceeding 0.2 meters per second. After landing, the shutdown procedure is automatically executed, including device status checks, data storage, and system shutdown. Device status checks utilize a self-diagnostic program, and data storage utilizes a circular cache mechanism.

[0220] To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: The research team first configured the autonomous flying fire extinguisher system, as shown in Table 1:

[0221] Table 1 Configuration parameters of autonomous flying fire extinguisher system

[0222] System Module Parameter indicators Parameter value Thermal imaging sensors Resolution 640×480 pixels Thermal imaging sensors Temperature detection range -40℃～1500℃ Visible light sensor Resolution 1920×1080 pixels Visible light sensor Frame rate 30fps Three-band flame sensor Sampling frequency 100Hz Temperature and humidity sensors Temperature range -50℃～150℃ Temperature and humidity sensors Humidity range 0~100%RH smoke sensor Detection range 0~1000ppm Carbon monoxide sensor Detection range 0~1000ppm High voltage generator Operating voltage 12000V Electromagnetic throwing controller Response time 20ms Water mist pump Maximum flow 2L / min

[0223] In a field test, the system extinguished a mountain forest fire covering an area of ​​approximately 800 square meters. The system first acquired fire scene information through multi-sensor data collection and then constructed a fire status matrix after data fusion processing. Based on the sensor data, as shown in Table 2:

[0224] Table 2 Fire scene sensor data collection results

[0225] Measurement parameters Minimum Maximum average value Temperature (℃) 125 850 420 Infrared intensity 85 245 168 Carbon monoxide concentration (ppm) 180 680 415 Flame height (m) 1.2 4.5 2.8 Spread speed (m / s) 0.08 0.35 0.22 Wind speed (m / s) 2.5 6.8 4.2

[0226] Figure 2 The chart shows the changing trends of fire scene temperature and carbon monoxide concentration over time, with the horizontal axis representing time (in seconds) and the vertical axis representing both temperature (degrees Celsius) and carbon monoxide concentration (ppm). This chart reflects the dynamic characteristics of key parameters during the fire extinguishing process. Based on the collected data, the fire forward development matrix formula is used to calculate:

[0227] f ij =0.4v ij +0.3d ij +0.3s ij +0.05;

[0228] The fire intensity value s ij The calculation uses:

[0229] s ij =0.4T ij +0.35I ij +0.25C ij +0.03;

[0230] Based on the positive fire development matrix, the system calculates the negative fire development matrix:

[0231] b ij =0.35r ij +0.35c ij +0.3e ij +0.04;

[0232] The fire extinguishing effect evaluation value adopts:

[0233]

[0234] The system calculates the fire extinguishing measures vector based on the fire area and fire intensity:

[0235] M=[12,8,180] T ;

[0236] It means that 12 fire extinguisher bombs and 8 fire extinguisher bottles are needed, and the fine water mist spraying time is 180 seconds.

[0237] When planning the route, the system first divides the fire scene into grids and calculates the fire intensity weight value of each grid node:

[0238]

[0239] The route planning uses a deep reinforcement learning algorithm, and the planning results are shown in Table 3:

[0240] Table 3 Route planning result parameters

[0241] Waypoint number X coordinate (m) Y coordinate (m) Height (m) Hover duration (s) 1 0 0 15 0 2 12.5 8.6 12 25 3 25.8 15.4 10 30 4 38.2 22.5 8 35 5 42.6 35.8 12 28 6 35.4 42.5 15 32 7 22.8 38.6 10 30 8 15.6 28.4 8 25 9 8.2 15.6 12 20 10 0 0 15 0

[0242] Route cost calculation uses:

[0243] C path =0.2L path +0.2R risk +0.2E eff +0.15P energy +0.15D diff +0.1T task +0.02;

[0244] The calculation results of resource utilization efficiency during system execution are shown in Table 4:

[0245] Table 4 Statistics of fire extinguishing resource utilization efficiency

[0246] Firefighting measures Delivery volume <![CDATA[Coverage area (m 2 )]]> Fire extinguishing effect rating Resource utilization Fire extinguisher 12 320 0.85 0.78 Fire extinguisher 8 280 0.82 0.75 water mist 360L 200 0.88 0.82

[0247] Figure 3The utilization rate and coverage area of ​​three firefighting resources (fire extinguisher bombs, fire extinguisher bottles, and water mist) are compared. The left vertical axis represents resource utilization, and the right vertical axis represents coverage area (square meters). This graph shows the comparative effectiveness of different firefighting measures. The entire firefighting process lasted approximately 15 minutes, and the fire was ultimately extinguished. The system performed two route replanning operations during execution, triggered by a change in fire intensity exceeding 30% and a change in spread direction exceeding 45 degrees. Monte Carlo simulation results showed that the mission completion rate remained above 90% during each replanning process. Traditional forest fire fighting relies primarily on manual on-site command and firefighting operations, which present the following problems: 1) Inaccurate fire situation awareness relies on manual judgment. 2) Inefficient allocation of firefighting resources often results in wasted resources. 3) Firefighting route planning lacks scientific basis, making optimization difficult. 4) Firefighting effectiveness evaluation methods are limited and lack quantitative indicators. 5) The safety risks to rescue personnel are high, especially in complex terrain.

[0248] The autonomous flying firefighting system employed in this invention offers the following advantages over traditional methods: 1) Multi-sensor data fusion enables precise perception of fire conditions, improving fire temperature measurement accuracy by 85% and fire spread prediction accuracy by 75%. 2) Intelligent algorithms are employed to optimize the allocation of firefighting resources, increasing resource utilization by 65% ​​and avoiding resource waste.

[0249] 3) The path planning method based on deep reinforcement learning increases firefighting efficiency by 80% and shortens task completion time by 50%.

[0250] 4) A complete fire extinguishing effectiveness evaluation system has been established, including multiple quantitative indicators, which has increased the objectivity of the evaluation results by 90%.

[0251] 5) It enables unmanned firefighting operations, significantly reducing the safety risks for rescue workers. 6) The system has adaptive adjustment capabilities, allowing it to optimize firefighting strategies in real time based on changing fire conditions, with an adjustment success rate of 95%.

[0252] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5 and 6 below.

[0253] Table 5 Variable Explanation Table (Part 1)

[0254]

[0255] Table 6 Variable Explanation Table (Part 2)

[0256]

[0257]

[0258] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. An intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher, characterized in that: include: Obtain sensor data of the autonomous flying fire extinguisher, construct a fire positive development matrix based on the sensor data, calculate a fire negative development matrix based on the fire positive development matrix, establish a fire extinguishing strategy equation group based on the fire positive development matrix and the fire negative development matrix, calculate a fire extinguishing measure vector using the fire extinguishing strategy equation group, establish a fire scene space division grid based on the fire positive development matrix and assign fire intensity weight values, perform route planning using a pre-trained fire extinguishing path planning model, calculate a flight trajectory curve based on the route planning result and generate a flight control instruction sequence, generate a fire extinguishing control instruction sequence based on the fire extinguishing measure vector, perform time alignment on the flight control instruction sequence and the fire extinguishing control instruction sequence to generate an execution instruction queue, control the autonomous flying fire extinguisher to perform a fire extinguishing task, update the fire positive development matrix and the fire negative development matrix in real time using the sensor data, determine whether to replan the route based on the updated matrix, and return to the platform after completing the fire extinguishing task; The fire extinguishing path planning model includes an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a mathematical calculation module, a fully connected layer, and an output layer; the input layer receives fire scene grid data, and the fire scene grid data includes temperature distribution data, fire intensity data, and combustible material distribution data; the first convolution layer adopts an edge detection convolution kernel group, and the edge detection convolution kernel group is generated by a fire scene boundary feature extraction function, and the fire scene boundary feature extraction function is constructed based on the temperature gradient change rate, the fire intensity change rate, and the combustible material density change rate; the second convolution layer adopts a direction perception convolution kernel group, and the direction perception convolution kernel group is generated by a fire scene boundary feature extraction function. The direction perception convolution kernel group is generated by a fire spread direction identification function, which is constructed based on the heat transfer direction and the ambient wind direction. The third convolution layer uses an intensity assessment convolution kernel group, which is generated by a fire intensity analysis function, which is constructed based on thermal imaging data, flame spectrum data, and smoke concentration data. The weight coefficient of each convolution kernel in the edge detection convolution kernel group is determined by a fire scene boundary clarity evaluation function, the input of which includes boundary temperature gradient, boundary infrared characteristics, and boundary visible light characteristics. The weight coefficient of each convolution kernel in the direction-aware convolution kernel group is determined by a fire spread trend evaluation function, wherein the input of the fire spread trend evaluation function includes heat diffusion speed, wind field data, and terrain factors; The weight coefficient of each convolution kernel in the intensity assessment convolution kernel group is determined by a fire hazard evaluation function, and the input of the fire hazard evaluation function includes flame temperature, thermal imaging intensity, and flame spectrum characteristics.

2. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 1 is characterized by: The sensor data includes thermal imaging sensor data, visible light sensor data, three-band flame sensor data, temperature and humidity sensor data, smoke sensor data, and carbon monoxide sensor data; the thermal imaging sensor data is collected at a sampling frequency of 60 Hz, the visible light sensor data is collected at a frame rate of 30 frames per second, and the three-band flame sensor data is collected at a sampling frequency of 100 Hz.

3. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 2 is characterized by: The positive fire development matrix represents the fire spread speed, fire expansion direction, and fire intensity; the negative fire development matrix represents the fire decay speed, fire contraction direction, and fire reduction degree after the fire extinguishing agent is released.

4. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 3 is characterized by: The pre-trained fire extinguishing path planning model adopts a convolutional neural network structure with an embedded mathematical calculation module. The mathematical calculation module includes a fire spread prediction equation, a resource consumption evaluation equation, and a route cost calculation equation, which is used to integrate the fire development trend and the fire extinguishing resource constraints to perform route optimization calculations; the fire spread prediction equation input includes fire scene temperature distribution data, ambient wind speed data, combustible material distribution data, and terrain height data; the resource consumption evaluation equation input includes fire scene temperature data, fire scene area data, fire extinguishing agent storage data, fire extinguisher remaining power data, and target fire point distance data; the route cost calculation equation input includes route length data, fire scene danger data, fire extinguishing efficiency prediction data, energy consumption data, passage difficulty data, and task completion time data.

5. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 4 is characterized by: The basic size of the fire scene space division grid is 5 meters × 5 meters, and a grid size of 1 meter × 1 meter is used in areas where the fire is strong or changes greatly; the fire weight value is calculated using the hierarchical analysis method, with the weight coefficient of the temperature value being 0.3, the weight coefficient of the fire intensity being 0.25, the weight coefficient of the spread speed being 0.2, the weight coefficient of the terrain characteristics being 0.15, and the weight coefficient of the surrounding important targets being 0.

1.

6. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 5 is characterized by: The flight trajectory curve is calculated using a Bezier curve interpolation algorithm. The maximum speed of the autonomous flying fire extinguisher is set to 15 meters per second, the maximum acceleration is set to 2 meters per square second, and the maximum turning angle is set to 45 degrees. The flight control instructions include three-dimensional position coordinates, flight speed, flight attitude, and hovering state.

7. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 6, characterized in that: The fire extinguishing control instruction sequence includes the working status of the high-voltage generator, the trigger timing of the electromagnetic throwing controller and the flow control of the fine water mist pump; the control of the high-voltage generator adopts pulse width modulation, the control of the electromagnetic throwing controller adopts timing triggering, and the control of the fine water mist pump adopts variable frequency speed regulation.

8. The intelligent fire extinguishing strategy control method of the autonomous flying fire extinguisher according to claim 7 is characterized by: When any of the following conditions are met: the fire intensity changes by more than 30%, the spread direction changes by more than 45 degrees, or the fire area changes by more than 20%, the replanning procedure is initiated. The replanning procedure uses the Monte Carlo simulation method to simulate the mission execution process 100 times. When the mission completion rate falls below 85%, the route is replanned.

9. The intelligent fire extinguishing strategy control method for an autonomous flying fire extinguisher according to claim 8, characterized in that: When the autonomous flying fire extinguisher returns to the platform, it uses satellite navigation when the distance from the platform is greater than 100 meters, with a navigation accuracy better than 5 meters; when the distance is between 30 and 100 meters, it uses laser navigation, with a navigation accuracy better than 0.5 meters; when the distance is less than 30 meters, it uses Bluetooth beacon navigation, with a navigation accuracy better than 0.1 meters.

Images