Method and system for controlling perfluorohexanone fire extinguishing robot based on risk perception

Abstract

The invention relates to the technical field of fire extinguishing robot control, in particular to a perfluorohexanone fire extinguishing robot control method and system based on risk perception. The method comprises the following steps: firstly, carrying out cross-view feature interaction processing on a visible light image, an infrared image and three-dimensional point cloud data at a fire position, constructing a local environment model, and combining with a building global structure model to obtain a fire environment three-dimensional model; then, when robot path planning is carried out based on the fire environment three-dimensional model, the heat radiation risk, the structure stability risk and the traffic difficulty risk are considered, and an initial traffic path is obtained; and finally, the robot arrives at the fire extinguishing position according to the initial passing path, in the fire extinguishing process, a target decision based on reinforcement learning is carried out, and according to a decision result, the robot executes fire fighting or self-protection continuously, so that the problem of fire fighting and self-protection conflict when the robot faces a fire disaster is solved, and the safety of the fire extinguishing robot is guaranteed while the fire extinguishing quality is improved.

Description

Technical Field

[0001] This invention relates to the field of fire extinguishing robot control technology, specifically to a perfluorohexanone fire extinguishing robot control method and system based on risk perception. Background Technology

[0002] In fire rescue, perfluorohexanone has great potential for application in the field of fire-fighting robots due to its environmentally friendly, highly efficient, and non-damaging properties. However, existing fire-fighting robot control methods cannot fully utilize its performance, which restricts the quality of rescue and the safety of the robots.

[0003] Existing firefighting robot control methods suffer from three major problems: First, the accuracy of fire scene perception is insufficient, relying mostly on single visible light or infrared image data, resulting in large deviations in fire point location and ignoring the overall structural relationships of the building, thus failing to provide comprehensive data support for decision-making. Second, path planning often only considers distance factors, easily leading to the planning of high-risk paths, and robots are prone to getting stuck in dangerous areas during execution. Third, the intelligence of firefighting decision-making is low, using fixed rules to control the "firefighting-movement" action. When faced with sudden situations such as fire spread or overheating of the robot body, the robot cannot balance the conflict between "firefighting" and "self-preservation," easily leading to incomplete firefighting or robot damage. Summary of the Invention

[0004] The purpose of this invention is to provide a control method and system for a perfluorohexanone fire extinguishing robot based on risk perception.

[0005] The technical solution of this invention is as follows: A risk-aware-based control method for a perfluorohexanone fire extinguishing robot includes the following operations: S1. Acquire visible light images, infrared images, and 3D point cloud data of the fire location; after separating the illumination and reflection components of the visible light image at the fire location, perform channel attention processing to obtain a visible light enhancement map; after correcting the infrared image, fuse it with the visible light enhancement map to obtain a visual infrared fusion map; the visual infrared fusion map and the 3D point cloud data undergo cross-view feature interaction processing to obtain a dual-view fusion feature map; based on the dual-view fusion feature map, obtain the fire location identification result, fuse it with the 3D point cloud data, and construct a local environment model; fuse the local environment model with the global building structure model to obtain a 3D fire environment model; S2. Based on the risks of thermal radiation, structural stability, and accessibility, mark the robot-accessible areas in the 3D model of the fire environment; perform path planning on the robot-accessible areas to obtain the initial access path; S3. The robot travels to the preset fire extinguishing location according to the initial travel path, aims at the fire point to extinguish the fire, and makes a target decision based on reinforcement learning in real time based on the environmental state, robot state and task progress, and obtains the decision result. The robot executes the action according to the decision result.

[0006] In S2, the robot's passable area is formed by planar passable grids in the three-dimensional model of the fire environment. The passable cost of the passable grid is less than the passable cost threshold. The passable cost is obtained based on the grid's thermal radiation risk, structural stability risk, and passability risk. The thermal radiation risk is obtained based on the radiation flux density. The structural stability risk is obtained based on the structural bearing capacity coefficient. The passability risk is obtained based on the ratio of the channel width to the robot width.

[0007] Path planning in S2 is based on the A* algorithm, and the evaluation function formula is as follows: , , , , , For nodes n The cost of assessment From the starting point to the node n The actual cost, For nodes n Risk and cost For nodes n The risks and costs of reaching the target point For risk calibration coefficient, Starting point s To the node n The path, For the first in the path p The passage cost of each grid cell. For nodes n The standardized value of thermal radiation risk, or the standardized value of structural stability risk, or the standardized value of accessibility risk of the grid cell. For the first k Weighting coefficients for risk classes The average basic travel cost, For nodes n To the target point o The number of grid cells, For target point o The risks and costs.

[0008] In the reinforcement learning-based target decision-making of S3, the reward function is constructed based on the fire extinguishing benefit, the risk cost of the robot body, and the resource saving incentive; the fire extinguishing benefit is based on the percentage of completed sub-tasks and the area of ​​the fire source; the risk cost of the robot body is based on the risk cost of the robot's location and the temperature of the robot body; and the resource saving incentive is based on the robot's remaining power and the robot's remaining spray volume.

[0009] As the robot travels along the initial path, it compares the detected point cloud data with the corresponding positions in the 3D model of the fire environment in real time; if it encounters a new obstacle, it plans a path based on the remaining passable area.

[0010] The cross-view feature interaction processing operation in S1 is as follows: project the 3D point cloud data onto the bird's-eye view to generate a spatial feature map, fuse it with the visual infrared map, and then perform cross-attention processing after convolution processing to obtain a dual-view fused feature map.

[0011] In S1, the operation to obtain the fire point identification result is as follows: preset three types of target query vectors: flame, obstacle, and escape route. The dual-view fused feature map is matched with the target query through the Transformer decoder, and the three-dimensional coordinates, category and confidence of the target are output to extract the fire point identification result.

[0012] A risk-aware perfluorohexanone (PFH) fire extinguishing robot control system, used to implement the aforementioned risk-aware perfluorohexanone (PFH) fire extinguishing robot control method, includes: The fire environment 3D model generation module is used to acquire visible light images, infrared images, and 3D point cloud data of the fire location; after separating the illumination and reflection components of the visible light image at the fire location, channel attention processing is performed to obtain a visible light enhancement map; the infrared image is corrected and then fused with the visible light enhancement map to obtain a visual infrared fusion map; the visual infrared fusion map and the 3D point cloud data are processed through cross-view feature interaction to obtain a dual-view fusion feature map; based on the dual-view fusion feature map, the fire location identification result is obtained, which is fused with the 3D point cloud data to construct a local environment model; the local environment model is then fused with the global building structure model to obtain a 3D fire environment model; The initial travel path generation module is used to mark the robot-accessible area in the 3D model of the fire environment based on thermal radiation risk, structural stability risk, and access difficulty risk; and to perform path planning on the robot-accessible area to obtain the initial travel path. The decision result generation and execution module is used for the robot to travel to the preset fire extinguishing location according to the initial travel path, aim at the fire point to extinguish the fire, and make target decisions based on reinforcement learning in real time based on the environmental state, robot state and task progress during the fire extinguishing process, and obtain decision results. The robot then executes actions according to the decision results.

[0013] A risk-aware perfluorohexanone (PFH) fire extinguishing robot control device includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the aforementioned risk-aware perfluorohexanone (PFH) fire extinguishing robot control method.

[0014] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described risk-aware perfluorohexanone fire extinguishing robot control method.

[0015] The beneficial effects of this invention are as follows: This invention provides a risk-aware-based control method for a perfluorohexanone (PFH) fire extinguishing robot. First, visible light images, infrared images, and 3D point cloud data of the fire location are acquired. The visible light images, after illumination reflection separation and channel attention enhancement, provide clearer details. The infrared images, after correction, provide more accurate temperature information. The fusion of these two images combines visual details and thermal radiation characteristics to obtain a visual-infrared fusion image. This image is then processed with the 3D point cloud data through cross-view feature interaction, improving the accuracy of fire point identification while constructing a local environment model. Combined with a global building structure model, the resulting 3D fire environment model accurately recreates key information such as the fire location and temperature, while also considering the overall building structure, providing comprehensive and reliable 3D data support for fire extinguishing decisions and path planning. Then, the robot's path is planned based on the 3D fire environment model. When planning, considering risks such as thermal radiation, structural stability, and difficulty of passage, the robot can accurately eliminate dangerous areas with high radiation, easy collapse, and high resistance, directly avoiding ineffective exploration. The resulting initial passage path not only matches the robot's passage capability but also avoids core risk points in advance, significantly reducing planning iteration costs and robot execution risks. Finally, the robot travels to the preset fire extinguishing location according to the initial passage path, extinguishes the fire, and makes target decisions based on reinforcement learning in real time based on environmental conditions, robot status, and task progress. Based on the decision results, the robot can choose to continue firefighting or leave the fire site for self-preservation, resolving the "firefighting-self-preservation" conflict problem when the robot faces a fire. This improves the quality of firefighting while ensuring the robot's safety and enhancing the intelligence of the firefighting robot. Attached Figure Description

[0016] The solutions and advantages of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] In the attached diagram: Figure 1 This is a schematic diagram of the logic of the method in this embodiment. Detailed Implementation

[0018] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0019] Example 1 This embodiment provides a risk-aware-based control method for perfluorohexanone (PFH) fire extinguishing robots. See [link to relevant documentation]. Figure 1 This includes the following operations: S1. Acquire visible light images, infrared images, and 3D point cloud data of the fire location; after separating the illumination and reflection components of the visible light image at the fire location, perform channel attention processing to obtain a visible light enhancement map; after correcting the infrared image, fuse it with the visible light enhancement map to obtain a visual infrared fusion map; the visual infrared fusion map and the 3D point cloud data undergo cross-view feature interaction processing to obtain a dual-view fusion feature map; based on the dual-view fusion feature map, obtain the fire location identification result, fuse it with the 3D point cloud data, and construct a local environment model; fuse the local environment model with the global building structure model to obtain a 3D fire environment model; S2. Based on the risks of thermal radiation, structural stability, and accessibility, mark the robot-accessible areas in the 3D model of the fire environment; perform path planning on the robot-accessible areas to obtain the initial access path; S3. The robot travels to the preset fire extinguishing location according to the initial travel path, aims at the fire point to extinguish the fire, and makes a target decision based on reinforcement learning in real time based on the environmental state, robot state and task progress, and obtains the decision result. The robot executes the action according to the decision result.

[0020] The specific steps are detailed below.

[0021] S1. Acquire visible light images, infrared images, and 3D point cloud data of the fire location; after separating the illumination and reflection components of the visible light image at the fire location, perform channel attention processing to obtain a visible light enhancement map; after correcting the infrared image, fuse it with the visible light enhancement map to obtain a visual infrared fusion map; the visual infrared fusion map and the 3D point cloud data undergo cross-view feature interaction processing to obtain a dual-view fusion feature map; based on the dual-view fusion feature map, obtain the fire location identification result, fuse it with the 3D point cloud data, and construct a local environment model; fuse the local environment model with the global building structure model to obtain a 3D fire environment model.

[0022] Visible light images, after being separated by illumination reflection and enhanced by channel attention, provide clearer details. Infrared images, after correction, provide more accurate temperature information. The fusion of these two images combines visual details with thermal radiation characteristics to obtain a visual-infrared fusion image. This image is then processed with 3D point cloud data through cross-view feature interaction to improve the accuracy of fire point identification while constructing a local environment model. Combined with the global building structure model, the final 3D fire environment model accurately restores key information such as the location and temperature of the fire point, while also taking into account the overall structural relationship of the building. This provides comprehensive and reliable 3D data support for fire fighting decisions and path planning.

[0023] First, we acquire visible light images of the fire location that capture the outline of flames and the texture of obstacles, infrared images that can penetrate dense smoke to identify high-temperature points, and three-dimensional point cloud data that reflects the characteristics of three-dimensional spatial structure.

[0024] Then, after separating the illumination and reflection components of the visible light image at the fire point, channel attention processing is performed to enhance the red feature channel of the flame and suppress the grayscale noise of the smoke, resulting in a visible light enhancement image.

[0025] Next, the infrared image is corrected and then fused with the visible light enhancement image to enhance the flame texture details at the fire point, resulting in a visual infrared fusion image.

[0026] The details of the correction process are as follows.

[0027] Step 1: Apply Gaussian filtering to the infrared image to smooth random noise while preserving flame edge details, resulting in a filtered infrared image. Set a grayscale threshold and perform grayscale binarization on the filtered infrared image. Areas with grayscale values ​​greater than the grayscale threshold are marked as suspected fire areas, and areas with grayscale values ​​less than the grayscale threshold are marked as background areas, resulting in a region location map.

[0028] Step 2: Based on the infrared detection range parameters, pixels in the infrared image whose grayscale values ​​reach the detector's saturation grayscale value at suspected fire areas are marked as saturated distortion pixels. At the same time, in the infrared image at suspected fire areas, areas with grayscale standard deviation less than the first grayscale standard deviation threshold and grayscale difference from the flame core area greater than the second grayscale standard deviation threshold (the first grayscale standard deviation threshold is less than the second grayscale standard deviation threshold) are selected and marked as smoke interference areas.

[0029] Step 3: For each saturated distortion pixel, construct a neighborhood window centered on the saturated distortion pixel. When the window boundary exceeds the image range, use mirror filling. Remove interference pixels (pixels in the smoke interference area and pixels with sudden gray value changes (pixels with local variance > local variance threshold) within the neighborhood window to eliminate instantaneous interference from flame flickering). The remaining pixels form the reference pixel set corresponding to each saturated distortion pixel.

[0030] Step 4: Convert the reference pixel set corresponding to each saturated distorted pixel into temperature values, obtain the temperature change pattern of the reference area, remove pixels whose absolute temperature gradient (including horizontal and vertical gradients) is greater than the corresponding absolute temperature gradient threshold (the product of twice the standard deviation and the mean of the corresponding reference pixel set), and select the pixel closest to the saturated distorted pixel and with the highest temperature as the reference point for the corresponding saturated distorted pixel; construct the corresponding linear gradient extrapolation model based on the relative position of the reference point and the corresponding saturated distorted pixel, and use the corresponding linear gradient extrapolation model to perform temperature recovery on the saturated distorted pixel to obtain the temperature recovery value of each saturated distorted pixel.

[0031] The calculation formula for the linear gradient extrapolation model is as follows: , For saturated distortion pixels P Temperature recovery value, As a reference point P 0 Temperature value, , These represent the average horizontal temperature gradient and average vertical temperature gradient of the neighborhood of the saturated distortion pixel, respectively. , Each is a reference point P 0 With corresponding saturation distortion pixels P The difference in lateral position and the difference in longitudinal position, x , y ), ( x 0 , y 0 ) are saturation distortion pixels P and benchmark P 0 The coordinates are given, and sign() is the sign function to ensure that the gradient direction is consistent with the position, such as saturation distortion pixels. P At the benchmark point P 0 When on the right side, =1, the value rules for sign() can be set and adjusted according to actual needs.

[0032] Step 5: Using the neighborhood temperature interpolation method, with the reference pixel set of the smoke interference area as the benchmark, calculate the temperature value of each pixel in the smoke interference area through bilinear interpolation, and use it as the pixel correction temperature of the smoke interference area; convert the temperature recovery value of the saturated distortion pixel area, the correction temperature of the smoke interference area and the temperature of the background area back to grayscale values, and integrate them into a complete corrected infrared image, which is used to perform the operation of fusion with the visible light enhancement image.

[0033] Furthermore, the visual infrared fusion map and the 3D point cloud data undergo cross-view feature interaction processing to strengthen the correlation between spatial location and semantic information, resulting in a dual-view fusion feature map. Specifically, the 3D point cloud data is projected onto the bird's-eye view to generate a spatial feature map. This spatial feature map, along with the visual infrared fusion map (which serves as the perspective feature map), is then processed by convolution to extract features. Finally, cross-attention processing is performed to achieve cross-view feature interaction, resulting in a dual-view fusion feature map.

[0034] Next, based on the dual-view fusion feature map, the fire point identification result is obtained and fused with the 3D point cloud data (marking the fire point identification result on the 3D point cloud data) to construct a local environment model. Specifically, the operation of obtaining the fire point identification result based on the dual-view fusion feature map involves designing three types of target query vectors: flames, obstacles, and escape routes. A Transformer decoder is used to match the dual-view fusion feature map with the target queries, outputting the target's 3D coordinates, category, and confidence score, thus extracting the fire point identification result.

[0035] Finally, the local environment model is updated in real time and then merged with the global building structure model built based on the BIM model to obtain a three-dimensional fire environment model.

[0036] S2. Based on the risks of thermal radiation, structural stability, and accessibility, mark the robot-accessible areas in the 3D model of the fire environment; perform path planning on the robot-accessible areas to obtain the initial access path.

[0037] Based on the passable areas in the 3D model of the fire environment marked by three risk categories—thermal radiation, structural stability, and accessibility—the system can accurately eliminate dangerous areas with high radiation, easy collapse, and high resistance, ensuring the safety and feasibility of the area. On this basis, path planning can directly avoid ineffective exploration. The initial passage path obtained not only matches the robot's passage capability but also avoids core risk points in advance, providing a safe and efficient basic framework for subsequent path optimization and significantly reducing the planning iteration cost and robot execution risk.

[0038] First, the three-dimensional model of the fire environment is rasterized into a plane. Based on the risks of thermal radiation, structural stability, and accessibility, the access cost of each grid is calculated. Grids with access costs less than the access cost threshold are designated as accessible grids, thereby further marking the robot-accessible areas in the three-dimensional model of the fire environment.

[0039] The formula for calculating the passage cost of a grid is as follows: , For the first i The passage cost of each grid cell. For the first iThe base passage cost per grid cell is determined by the scene type; the larger and more expansive the area, the lower the base passage cost. For the first i The risk cost of each grid For the first i The standardized values ​​for thermal radiation risk, structural stability risk, or accessibility risk of each grid cell are obtained by standardizing the thermal radiation risk, structural stability risk, or accessibility risk respectively. For the first k Weighting coefficients for risk categories (thermal radiation risk, structural stability risk, or difficulty of passage risk).

[0040] Among them, thermal radiation risk is obtained based on radiation flux density, structural stability risk is obtained based on structural bearing capacity coefficient, which can be calculated based on the material high temperature resistance curve of the global structural model of the building, combined with real-time temperature calculation of remaining bearing capacity; passage difficulty risk is obtained based on the ratio of passage width to robot width.

[0041] Then, in the area where the robot can travel, path planning is performed, and the path with the lowest risk assessment value is selected as the initial travel path.

[0042] Path planning is based on the A* algorithm. In this embodiment, to avoid planning the "shortest but most dangerous" path, a risk cost is introduced on the basis of the traditional A* algorithm, and the evaluation function is reconstructed as follows: , For nodes n The cost of assessment From the starting point to the node n The actual cost, For nodes n Risk and cost For nodes n The risks and costs of reaching the target point (the pre-set fire extinguishing location). The risk calibration coefficient is a weighting adjustment term that balances risk and distance.

[0043] From start point to node n The actual cost The calculation formula is as follows: , Starting point s To the node n The path, For the first in the path p The passage cost per grid.

[0044] noden The formula for calculating risk cost is as follows: , Larger values ​​indicate a larger node n The higher the risk of the grid, the more... Multiplication can realize the actual cost premium of high-risk paths, guiding the algorithm to prioritize avoiding high-risk areas.

[0045] node n To the target point o Risk cost The calculation formula is as follows: , The average basic travel cost, For nodes n To the target point o The design ensures the number of grid cells. Do not overestimate the minimum actual cost in a risk-free scenario.

[0046] Risk calibration coefficient It is a node n、 The risk cost of the target point is obtained, and the calculation formula is as follows: , For target point o The risks and costs.

[0047] S3. The robot travels to the preset fire extinguishing location according to the initial travel path, aims at the fire point to extinguish the fire, and makes a target decision based on reinforcement learning in real time based on the environmental state, robot state and task progress, and obtains the decision result. The robot executes the action according to the decision result.

[0048] The robot travels along an initial path to the preset fire extinguishing location, targets the fire, and extinguishes it. During the fire extinguishing process, it makes target decisions based on reinforcement learning in real time, taking into account the environmental conditions, robot status, and task progress. Based on the decision results, the robot executes actions to either continue firefighting or leave the fire site to protect itself. This resolves the "firefighting-self-preservation" conflict problem when the robot faces a fire, ensuring the quality of fire extinguishing while improving the robot's self-preservation ability and enhancing its intelligence.

[0049] In reinforcement learning-based target decision-making, the state space includes environmental state (fire source size, risk value distribution), robot state (battery power, remaining spray volume, body temperature), and task progress (percentage of completed sub-tasks). The action space defines five basic actions: forward, backward, turning, spraying, and hazard avoidance. Each action includes specific parameters (such as spray flow rate and turning angle). Compared to traditional reinforcement learning-based target decision-making methods, this embodiment considers fire extinguishing benefits, body risk costs, and resource conservation incentives in the reward function to ensure that the generated decision results (actions) in target decision-making conform to the robot's fire extinguishing efficiency, ensuring the robot's survivability in extreme situations and reducing losses.

[0050] In goal decision-making based on reinforcement learning Q The formula for calculating the value is as follows: , , for t Moment Q value, , They are respectively t The state and actions (decision results) at any given moment. As a discount factor, for t+ 1 moment Q value, , They are respectively t+ The state and actions (decision results) at time 1. For the reward function, , , These are respectively the benefits of firefighting, the risks associated with the aircraft, and the incentives for resource conservation. , , These are weighted by firefighting benefits, weighted by airframe risk costs, and weighted by resource conservation incentives.

[0051] Firefighting benefits It is based on the percentage of completed sub-tasks and the area of ​​the fire source, and the calculation formula is as follows: , , They are respectively t , t+ The percentage of sub-tasks completed at time 1. , t, respectively t+ The equivalent area of ​​the fire source at moment 1 is calculated using logarithmic terms in the formula, which ensures diminishing marginal returns to fire source shrinkage and avoids excessive spraying that wastes resources. , These are the task progress weight and the fire source control weight, respectively.

[0052] Airframe risk and cost It is based on the risk cost of the robot's location (grid) and the robot's body temperature, and the calculation formula is as follows: , for t Risk cost of the grid cell where the robot is located at time +1 for t +1 moment body temperature , These are the safe temperature threshold and the shutdown temperature threshold, respectively. As a risk penalty weight, This represents the environmental risk coefficient.

[0053] Resource conservation incentives It is based on the robot's remaining battery power and remaining spray volume, and the calculation formula is as follows: , , They are respectively t The robot's remaining battery power at time +1 and its total battery power at the initial time. , They are respectively t The remaining spray volume of the robot at time +1 and the total amount at the initial time are used in the formula. The exponential term in the formula ensures that the more resources remaining, the stronger the incentive, which conforms to the principle of "low consumption first". For dynamic adjustment items, In the early stages of the mission, the weighting of firefighting benefits is increased; in the later stages, the weighting of resource conservation is increased, achieving phase adaptation. This is the resource incentive weight.

[0054] In addition, during the process of the robot traveling according to the initial passage path, the detection point cloud data is compared and matched with the corresponding positions in the three-dimensional model of the fire environment in real time; if new obstacles are encountered (new debris caused by the fire that does not exist in the three-dimensional model of the fire environment when the initial passage path was planned), the path is planned based on the remaining passable area, which enhances the real-time performance, flexibility and accuracy of path planning.

[0055] This embodiment provides a risk-aware perfluorohexanone fire extinguishing robot control system, used to implement the aforementioned risk-aware perfluorohexanone fire extinguishing robot control method, including: The fire environment 3D model generation module is used to acquire visible light images, infrared images, and 3D point cloud data of the fire location; after separating the illumination and reflection components of the visible light image at the fire location, channel attention processing is performed to obtain a visible light enhancement map; the infrared image is corrected and then fused with the visible light enhancement map to obtain a visual infrared fusion map; the visual infrared fusion map and the 3D point cloud data are processed through cross-view feature interaction to obtain a dual-view fusion feature map; based on the dual-view fusion feature map, the fire location identification result is obtained, which is fused with the 3D point cloud data to construct a local environment model; the local environment model is then fused with the global building structure model to obtain a 3D fire environment model; The initial travel path generation module is used to mark the robot-accessible area in the 3D model of the fire environment based on thermal radiation risk, structural stability risk, and access difficulty risk; and to perform path planning on the robot-accessible area to obtain the initial travel path. The decision result generation and execution module is used for the robot to travel to the preset fire extinguishing location according to the initial travel path, aim at the fire point to extinguish the fire, and make target decisions based on reinforcement learning in real time based on the environmental state, robot state and task progress during the fire extinguishing process, and obtain decision results. The robot then executes actions according to the decision results.

[0056] This embodiment provides a risk-aware perfluorohexanone fire extinguishing robot control device, including a processor and a memory, wherein the processor executes the computer program stored in the memory to implement the above-mentioned risk-aware perfluorohexanone fire extinguishing robot control method.

[0057] This embodiment provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described risk-aware perfluorohexanone fire extinguishing robot control method.

[0058] This embodiment provides a risk-aware-based control method for a perfluorohexanone (PFH) fire extinguishing robot. First, visible light images, infrared images, and 3D point cloud data of the fire location are acquired. The visible light images, after illumination reflection separation and channel attention enhancement, provide clearer details. The infrared images, after correction, provide more accurate temperature information. The fusion of these two images combines visual details and thermal radiation characteristics to obtain a visual-infrared fusion image. This image is then processed with the 3D point cloud data through cross-view feature interaction, improving the accuracy of fire point identification while constructing a local environment model. Combined with a global building structure model, the resulting 3D fire environment model accurately recreates key information such as the fire point's location and temperature, while also considering the overall building structure, providing comprehensive and reliable 3D data support for fire extinguishing decisions and path planning. Then, the robot's path is planned based on the 3D fire environment model. When planning, considering risks such as thermal radiation, structural stability, and difficulty of passage, the robot can accurately eliminate dangerous areas with high radiation, easy collapse, and high resistance, directly avoiding ineffective exploration. The resulting initial passage path not only matches the robot's passage capability but also avoids core risk points in advance, significantly reducing planning iteration costs and robot execution risks. Finally, the robot travels to the preset fire extinguishing location according to the initial passage path, extinguishes the fire, and makes target decisions based on reinforcement learning in real time based on environmental conditions, robot status, and task progress. Based on the decision results, the robot can choose to continue firefighting or leave the fire site for self-preservation, resolving the "firefighting-self-preservation" conflict problem when the robot faces a fire. This improves the quality of firefighting while ensuring the robot's safety and enhancing the intelligence of the firefighting robot.

Claims

1. A risk-aware based perfluorocyclohexanone fire extinguishing robot control method, characterized in that, The method comprises the following steps: S1, acquiring visible light images, infrared images and three-dimensional point cloud data at the fire site; After separating the illumination and reflection components of the visible light image at the fire point, channel attention processing is performed to obtain a visible light enhanced image; after correcting the infrared image, the infrared image is fused with the visible light enhanced image to obtain a visual infrared fusion image; the visual infrared fusion image and the three-dimensional point cloud data are processed through cross-view feature interaction to obtain a dual-view fusion feature map; Based on the dual-view fusion feature map, a fire point recognition result is obtained, which is fused with the three-dimensional point cloud data to construct a local environment model; the local environment model is fused with a global structure model of the building to obtain a fire environment three-dimensional model; S2, based on thermal radiation risk, structural stability risk and passage difficulty risk, marking the robot passable area in the fire environment three-dimensional model; On the robot passable area, path planning is performed to obtain an initial passage path; S3, the robot travels to a preset fire extinguishing position according to the initial passage path, aims at the fire point for fire extinguishing, and in the process of fire extinguishing, real-time target decision based on reinforcement learning is performed based on the environment state, the robot state and the task progress to obtain a decision result, and the robot executes actions according to the decision result.

2. The risk perception based perfluorocyclohexanone fire extinguishing robot control method according to claim 1, characterized by, The robot passable area in S2 is formed by the planar passable grids in the fire environment three-dimensional model, and the passable cost of the passable grid is less than a passable cost threshold, and the passable cost is obtained based on the thermal radiation risk, the structural stability risk and the passage difficulty risk of the grid; The thermal radiation risk is obtained based on the radiation flux density; The structural stability risk is obtained based on the structural bearing capacity coefficient; The passage difficulty risk is obtained based on the ratio of the passage width to the robot width.

3. The risk perception based perfluorocyclohexanone fire extinguishing robot control method according to claim 1, wherein, In S2, the path planning is realized based on the A* algorithm, and the evaluation function formula is as follows: ， ， ， ， ， For nodes n The cost of assessment From the starting point to the node n The actual cost, For nodes n Risk and cost For nodes n The risks and costs of reaching the target point For risk calibration coefficient, Starting point s To the node n The path, For the first in the path p The passage cost of each grid cell. For nodes n The standardized value of thermal radiation risk, or the standardized value of structural stability risk, or the standardized value of accessibility risk of the grid cell. For the first k Weighting coefficients for risk classes The average basic travel cost, For nodes n To the target point o The number of grid cells, For target point o The risks and costs.

4. The risk perception based perfluorohexone fire extinguishing robot control method according to claim 1, wherein, In the target decision based on reinforcement learning in S3, the reward function is constructed based on the fire extinguishing benefit, the robot body risk cost and the resource saving incentive; The fire extinguishing benefit is obtained based on the completed subtask proportion and the fire source area; The robot body risk cost is obtained based on the robot position risk cost and the robot body temperature; The resource saving incentive is obtained based on the remaining electric quantity of the robot and the remaining spraying quantity of the robot.

5. The risk perception based perfluorohexone fire extinguishing robot control method according to claim 1, wherein, In the process of traveling of the robot according to the initial passage path, the detected point cloud data is compared with the corresponding position of the fire environment three-dimensional model in real time; if a new obstacle is encountered, path planning is performed based on the remaining passable area.

6. The risk perception based perfluorohexone fire extinguishing robot control method according to claim 1, wherein, In S1, the operation of cross-view feature interaction processing is as follows: The three-dimensional point cloud data is projected onto the bird's eye view to generate a spatial feature map, which is respectively processed through convolution and then cross attention processing to obtain a dual-view fusion feature map.

7. The risk perception based perfluorocyclohexanone fire extinguishing robot control method according to claim 1, wherein, In S1, the operation of obtaining the fire point recognition result is as follows: Three types of target query vectors of fire, obstacle and escape passage are preset, the dual-view fusion feature map is matched with the target query through a Transformer decoder to output the three-dimensional coordinates, the category and the confidence of the target, and the fire point recognition result is obtained by extraction.

8. A risk-aware perfluorocyclohexanone extinguishing robot control system for implementing the risk-aware perfluorocyclohexanone extinguishing robot control method of claim 1, characterized by, It comprises: a fire environment three-dimensional model generation module for acquiring visible light images, infrared images and three-dimensional point cloud data at the fire site; After separating the illumination and reflection components of the visible light image at the fire point, channel attention processing is performed to obtain a visible light enhanced image; after correcting the infrared image, the infrared image is fused with the visible light enhanced image to obtain a visual infrared fusion image; the visual infrared fusion image and the three-dimensional point cloud data are processed through cross-view angle feature interaction to obtain a dual-view fusion feature map; based on the dual-view fusion feature map, a fire point recognition result is obtained, which is fused with the three-dimensional point cloud data to construct a local environment model; the local environment model is fused with a building global structure model to obtain a fire environment three-dimensional model; An initial passage path generation module is configured to mark a robot passable area in the fire environment three-dimensional model based on thermal radiation risk, structural stability risk, and passage difficulty risk; Path planning is performed on the robot passable area to obtain an initial passage path; A decision result generation and execution module is configured to cause the robot to travel to a preset fire extinguishing position according to the initial passage path, align the fire point, and perform fire extinguishing, and in the fire extinguishing process, perform target decision based on reinforcement learning in real time based on an environment state, a robot state, and a task progress to obtain a decision result, and the robot performs an action according to the decision result.

9. A risk-aware perfluorohexone fire extinguishing robot control device, characterized by, A processor and a memory are included, wherein the processor implements the risk perception based perfluorocyclohexanone fire extinguishing robot control method according to any one of claims 1-7 when executing a computer program stored in the memory.

10. A computer-readable storage medium, characterized in that, A computer program is stored, wherein the computer program is executed by a processor to implement the risk perception based perfluorocyclohexanone fire extinguishing robot control method according to any one of claims 1-7.

Images