Fire-fighting turbofan cannon jet posture automatic planning method based on multi-source sensing data

Abstract

This invention relates to the field of fire-fighting turbofan monitor control technology, specifically a method for automatically planning the spray attitude of fire-fighting turbofan monitors based on multi-source sensor data. The method includes: acquiring real-time monitoring data of the fire scene, such as 3D laser point clouds, infrared thermal imaging temperature fields, visible light video streams, and atmospheric wind speed and direction, through multiple sensors. After preprocessing, the 3D spatial boundary of the flame and the distribution of high-temperature hot spots are extracted. Spatial fusion is then used to generate a 3D flame thermodynamic composite model containing temperature and geometric attributes. This model, along with other monitoring data, is input into an improved depth deterministic strategy gradient algorithm optimized by fire jet dynamics and operational safety constraints to calculate the optimal spray attitude, including parameters such as pitch angle and azimuth angle. This method can accurately characterize the fire state, improve the accuracy of the spray attitude and fire extinguishing efficiency, and ensure operational safety.

Description

Technical Field

[0001] This invention relates to the field of fire-fighting turbofan monitor control technology, and in particular to an automatic planning method for the spray attitude of fire-fighting turbofan monitors based on multi-source sensor data. Background Technology

[0002] Automatic planning of the spray attitude for fire-fighting turbofan monitors is primarily used to collect fire scene monitoring data, analyze the fire state, and calculate the optimal spray attitude to achieve efficient fire suppression. Existing technologies mostly collect single-type fire monitoring data, analyze only single-dimensional fire information, and do not integrate multi-source sensor data to build a comprehensive fire model; the spray attitude calculation uses conventional algorithms, without considering fire jet dynamics and operational safety constraints, and only outputs basic attitude parameters based on simple fire location information.

[0003] Conventional technical solutions have shortcomings. Single-type monitoring data cannot fully reflect the combustion range, temperature distribution, and on-site environmental conditions of a fire, which can easily lead to biased fire status judgment. Without constructing a fire model that combines temperature attributes and geometric structure, it is difficult to accurately grasp the spatial distribution characteristics of the fire. Conventional algorithms lack targeted constraint optimization, and the calculated spray posture is prone to deviating from the actual fire extinguishing needs, making it unsuitable for complex fire-fighting scenarios, affecting fire extinguishing efficiency, and even posing operational safety hazards.

[0004] The methods for acquiring and processing sensor data need to be optimized, and multi-source monitoring data should be integrated to construct a comprehensive fire model that accurately represents the fire status and spatial distribution. The spray attitude calculation algorithm needs to be improved and optimized in combination with relevant constraints to generate optimal spray attitude parameters that are suitable for actual fire-fighting scenarios, thereby improving fire-fighting efficiency and operational safety. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an automatic planning method for the spray attitude of fire-fighting turbofan cannons based on multi-source sensor data.

[0006] To achieve the above objectives, the present invention employs the following technical solution: an automatic planning method for the spray attitude of a fire-fighting turbofan monitor based on multi-source sensor data, comprising:

[0007] Real-time monitoring data of the target area is obtained by various sensors deployed in the fire scene. The real-time monitoring data includes three-dimensional laser point cloud data, infrared thermal imaging temperature field data, visible light image video stream, and atmospheric wind speed and direction information.

[0008] The three-dimensional laser point cloud data is preprocessed to extract the three-dimensional spatial boundary of the core area of ​​the flame combustion, and the infrared thermal imaging temperature field data is threshold segmented to identify the two-dimensional hot spot distribution in the high-temperature danger area.

[0009] The three-dimensional spatial boundary of the core area of ​​the flame combustion is spatially fused and mapped with the two-dimensional hot spot distribution of the high-temperature hazard area to generate a three-dimensional flame thermal composite model, which includes temperature attributes and geometric structure.

[0010] The three-dimensional flame thermal composite model, the visible light image video stream, and the atmospheric wind speed and direction information are jointly input into the improved depth deterministic strategy gradient algorithm, which is optimized based on fire jet dynamics constraints and operational safety constraints.

[0011] Using the improved depth deterministic strategy gradient algorithm, the optimal spray attitude of the fire-fighting turbofan cannon is calculated. The optimal spray attitude includes the spray pitch angle, spray azimuth angle, turbofan speed, and jet medium mixing ratio.

[0012] As a further aspect of the present invention, the three-dimensional laser point cloud data is preprocessed to extract the three-dimensional spatial boundary of the core region of the flame combustion, including:

[0013] The acquired raw 3D laser point cloud data is subjected to denoising filtering to remove discrete noise points caused by smoke and dust, forming a denoised point cloud.

[0014] The denoised point cloud is spatially voxelized, dividing the continuous space into a regular three-dimensional voxel grid, and the point cloud density within each voxel grid is calculated.

[0015] Voxel grids with point cloud density below a preset threshold are identified as open areas, while voxel grids with point cloud density above a preset threshold are identified as potential combustion material accumulation areas.

[0016] Connectivity analysis is performed on the potential combustion material accumulation region to extract the set of voxels that are spatially connected and have a volume greater than the minimum fireball volume, and each set of voxels is marked as an independent combustion object.

[0017] Based on the spatial point cloud distribution of the independent burning object, its convex hull or minimum circumscribed cube is calculated, and the outer surface of the convex hull or minimum circumscribed cube is defined as the three-dimensional spatial boundary of the core region of the flame combustion.

[0018] As a further aspect of the present invention, threshold segmentation is performed on the infrared thermal imaging temperature field data to identify the two-dimensional hot spot distribution in high-temperature hazardous areas, including:

[0019] Gaussian smoothing filtering is applied to the infrared thermal imaging temperature field data to suppress sensor thermal noise and obtain a smoothed temperature field image.

[0020] The smoothed temperature field image is processed using an adaptive threshold segmentation algorithm, which automatically calculates the segmentation threshold based on the local temperature gradient to extract high-temperature regions where the temperature is higher than the background.

[0021] Morphological opening operations are performed on the extracted high-temperature regions to eliminate small noise points and fill the voids inside the regions, forming a complete high-temperature region connected domain.

[0022] Calculate the temperature statistical characteristics of the connected domains in each high-temperature region, including the average temperature, the maximum temperature, and the temperature distribution variance.

[0023] The connected regions of high-temperature areas where the average temperature exceeds the first dangerous temperature threshold, the highest temperature exceeds the second dangerous temperature threshold, or the temperature distribution variance exceeds the preset variance threshold are marked as two-dimensional hot spots of the high-temperature dangerous area, and their geometric contours and temperature characteristics are recorded.

[0024] As a further aspect of the present invention, the improved depth deterministic strategy gradient algorithm is optimized based on fire jet dynamics constraints and operational safety constraints, and its working principle includes:

[0025] A state space is constructed, the inputs of which are the discrete voxelized representation of the three-dimensional flame thermodynamic composite model, the smoke concentration feature map extracted in real time from the visible light image video stream, and the atmospheric wind speed and direction information;

[0026] A motion space is constructed, and the output of the motion space is a continuous jet attitude control quantity, which includes the jet pitch angle, jet azimuth angle, turbofan speed and jet medium mixing ratio.

[0027] Between the policy network and value network trained by the agent, a fire jet dynamics simulation model is introduced as an environment model. The fire jet dynamics simulation model is used to predict the jet trajectory, impact range and cooling and suppression effect on the flame based on the current state and actions.

[0028] Define a reward function, which is composed of a weighted sum of multiple reward items. The reward items include the proximity reward between the jet landing point and the flame core, the reward for the reduction of the flame model volume per unit time, the safety reward for the jet avoiding dangerous building components, and the penalty for the total amount of medium consumed by the jet.

[0029] After each action is output by the policy network, the action is input into the fire jet dynamics simulation model to obtain the predicted next state and immediate reward. The policy network and value network are then trained offline or online using the data until the network converges. The policy network is the optimized attitude planner.

[0030] As a further aspect of the present invention, the definition of the reward function includes:

[0031] Calculate the Euclidean distance between the jet impact point and the flame core, and use the negative exponential function value of the Euclidean distance as the proximity bonus between the jet impact point and the flame core.

[0032] Compare the changes in the number of voxels marked as flames in the three-dimensional flame thermodynamic composite model before and after the execution action, and multiply the reduced number of voxels by a positive coefficient as the reward for the reduction in the volume of the flame model per unit time.

[0033] The three-dimensional spatial coordinates of dangerous building components that are prohibited from being hit by jets in the fire scene are pre-loaded. The minimum distance between the predicted jet trajectory and the three-dimensional spatial coordinates of all dangerous building components is calculated. When the minimum distance is greater than the safe distance, a positive safety reward is given for the jet to avoid dangerous building components; otherwise, a negative reward is given.

[0034] Accumulate the media consumption caused by all actions within a decision cycle, and multiply the negative value of the media consumption by a coefficient as a penalty for the total amount of media consumed by the jet.

[0035] All reward and penalty items are linearly weighted and summed according to preset weights to obtain the output value of the reward function at each decision step.

[0036] As a further aspect of the present invention, the optimal spray attitude of the fire-fighting turbofan cannon is calculated using the improved depth deterministic strategy gradient algorithm, including:

[0037] The current three-dimensional flame thermal composite model, the current frame smoke features extracted from the visible light image video stream, and the current atmospheric wind speed and direction information are jointly encoded into the input vector required by the state space of the improved depth deterministic strategy gradient algorithm.

[0038] The input vector is input into the policy network of the improved deep deterministic policy gradient algorithm that has been trained.

[0039] The strategy network outputs an action vector, which corresponds to a set of specific values ​​for the jet pitch angle, jet azimuth angle, turbofan speed, and jet medium mixing ratio in the action space.

[0040] The values ​​of each control variable are decoded from the action vector output by the policy network to form the initial jet attitude command at the current moment.

[0041] As a further aspect of the present invention, the method further includes:

[0042] The optimal spraying attitude is dynamically compensated and corrected based on the atmospheric wind speed and direction information to generate a final execution command, driving the fire-fighting turbofan monitor actuator to move, including:

[0043] Obtain the atmospheric wind speed and direction information at the current moment, and analyze the wind speed magnitude, wind direction angle, and short-term fluctuation variance of wind speed from it;

[0044] Based on the fluid dynamics model, the expected deviations in the horizontal and vertical directions of the jet as it travels from the outlet of the fire-fighting turbofan cannon to the predetermined impact point are calculated under the conditions of the wind speed and wind direction angle.

[0045] Based on the short-term fluctuation variance of the wind speed, the dispersion range of the jet during flight due to wind speed fluctuations is estimated, and the dispersion range is used as a safety margin.

[0046] The calculated offset is superimposed in the opposite direction onto the jet azimuth and jet pitch angles in the initial jet attitude command to obtain the corrected attitude angle after wind speed deviation compensation.

[0047] Based on the corrected attitude angle and combined with the safety margin, the turbofan speed is finely adjusted to ensure that the jet can still cover the core area of ​​the target within the predetermined dispersion range, and finally the final execution command containing the compensated parameters is output.

[0048] As a further aspect of the present invention, the method also includes online adjustment of the jet effect evaluation and attitude planning strategy based on sensor data feedback:

[0049] After executing the final execution command to drive the fire-fighting turbofan cannon to spray, new infrared thermal imaging temperature field data and three-dimensional laser point cloud data are continuously acquired.

[0050] Based on the new infrared thermal imaging temperature field data, the area change rate and average temperature change rate of the two-dimensional hot spot in the high-temperature danger zone are calculated.

[0051] Based on the new three-dimensional laser point cloud data, the volume change rate and centroid displacement of the three-dimensional spatial boundary of the flame combustion core region are calculated.

[0052] The area change rate, average temperature change rate, volume change rate, and centroid displacement are used as evaluation indicators for the spraying effect.

[0053] The jetting effect evaluation index is fed back to the improved depth deterministic policy gradient algorithm as a component of the state input for the next round, or as the basis for calculating the immediate reward in the reward function, thereby realizing the online adaptive adjustment of the attitude planning strategy.

[0054] As a further aspect of the present invention, the planning process using the improved deep deterministic policy gradient algorithm also includes a real-time safety monitoring and action intervention mechanism:

[0055] After the policy network outputs the action vector, the jet attitude control quantity decoded from the action vector is input into an independent fast collision detection module;

[0056] The rapid collision detection module loads a three-dimensional model of the fire-fighting turbofan cannon body, a three-dimensional model of the surrounding fixed obstacles, and the three-dimensional flame thermal composite model to simulate the motion of the fire-fighting turbofan cannon body and the jet trajectory under the jet attitude control.

[0057] During the simulation, the test detects whether the fire-fighting turbofan cannon body or jet trajectory geometrically interferes with the three-dimensional model of the surrounding fixed obstacles or the preset prohibited crossing area in the three-dimensional flame thermal composite model.

[0058] If geometric interference is detected, the action vector is immediately intercepted, and a preset safety avoidance strategy is triggered to generate an interference-free alternative action.

[0059] The alternative action or the original action that passes the safety test is taken as the optimal injection posture that can be executed at the end.

[0060] As a further aspect of the present invention, the calculation of the Euclidean distance between the jet impact point and the flame core, and the use of the negative exponential function value of the Euclidean distance as a proximity bonus between the jet impact point and the flame core, includes:

[0061] The spatial coordinates of the predicted jet impact point corresponding to the current action are obtained by simulating the dynamics of the fire jet.

[0062] The three-dimensional flame thermodynamic composite model is queried to obtain the geometric center coordinates of all voxels inside the three-dimensional spatial boundary of the flame combustion core region, and the arithmetic mean of the geometric center coordinates is calculated. The arithmetic mean is then used as the spatial coordinates of the flame core.

[0063] Based on the spatial coordinates of the predicted jet impact point and the spatial coordinates of the flame core, the Euclidean distance in the three-dimensional Cartesian coordinate system is calculated.

[0064] The calculated Euclidean distance is used as an input variable and substituted into a preset exponential decay function. The output value of the preset exponential decay function is the negative exponential function value of the Euclidean distance.

[0065] The negative exponential function value of the Euclidean distance is directly assigned as the specific value of the proximity reward between the jet landing point and the flame core in the reward function.

[0066] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0067] The three-dimensional laser point cloud data acquired by various sensors deployed in the fire scene are preprocessed to extract the three-dimensional spatial boundary of the core flame combustion area. The infrared thermal imaging temperature field data is thresholded to identify the two-dimensional hot spot distribution in the high-temperature hazard area. The three-dimensional spatial boundary of the core flame combustion area and the two-dimensional hot spot distribution in the high-temperature hazard area are spatially fused and mapped to generate a three-dimensional flame thermal composite model containing temperature attributes and geometric structure. This can avoid the limitations of conventional single data acquisition and non-fusion modeling, comprehensively capture the spatial geometric features and temperature distribution patterns of the fire, make the fire state representation more accurate, reduce the spray attitude deviation caused by incomplete fire information, and meet the actual needs of complex fire scenarios.

[0068] An improved deep deterministic strategy gradient algorithm is developed by inputting a three-dimensional flame thermal composite model, visible light image video stream, and atmospheric wind speed and direction information. This improved algorithm is optimized based on fire jet dynamics constraints and operational safety constraints. The optimized algorithm is used to calculate the optimal spray attitude of the fire-fighting turbofan cannon, which includes the spray pitch angle, spray azimuth angle, turbofan speed, and jet medium mixing ratio. This avoids the drawbacks of conventional algorithms, such as lack of constraints and single attitude parameters, making the spray attitude calculation more consistent with the laws of fire jets and operational safety requirements. It accurately matches the spatial distribution of the fire and the on-site environmental conditions, improves the accuracy of jet coverage, enhances the fire extinguishing effect, avoids operational safety hazards, and adapts to fire-fighting scenarios of varying complexity. Attached Figure Description

[0069] Figure 1 This is a flowchart of the automatic planning method for the spray attitude of a fire-fighting turbofan monitor based on multi-source sensor data, as described in this invention.

[0070] Figure 2 A flowchart for extracting the three-dimensional spatial boundary of the core region of flame combustion;

[0071] Figure 3 A flowchart illustrating the working principle of the improved deep deterministic policy gradient algorithm. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0073] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0074] See Figure 1 This invention provides an automatic planning method for the spray attitude of a fire-fighting turbofan monitor based on multi-source sensor data. The specific method includes:

[0075] Real-time monitoring data of the target area is acquired through multiple sensors deployed in the fire scene. This data includes 3D laser point cloud data, infrared thermal imaging temperature field data, visible light image video streams, and atmospheric wind speed and direction information. The 3D laser point cloud data is preprocessed to extract the 3D spatial boundary of the core flame combustion area, and the infrared thermal imaging temperature field data is thresholded to identify the 2D hotspot distribution in the high-temperature hazard area. The 3D spatial boundary of the core flame combustion area and the 2D hotspot distribution in the high-temperature hazard area are spatially fused and mapped to generate a 3D flame thermal composite model, which includes temperature attributes and geometric structure. The 3D flame thermal composite model, visible light image video streams, and atmospheric wind speed and direction information are input into an improved depth-deterministic strategy gradient algorithm, which is optimized based on fire jet dynamics constraints and operational safety constraints. Using the improved depth-deterministic strategy gradient algorithm, the optimal spray attitude of the fire-fighting turbofan monitor is calculated. The optimal spray attitude includes the spray elevation angle, spray azimuth angle, turbofan speed, and jet medium mixing ratio.

[0076] In one embodiment of the present invention, the three-dimensional laser point cloud data is preprocessed to extract the three-dimensional spatial boundary of the core region of the flame combustion, see reference. Figure 2The process includes denoising and filtering the acquired raw 3D laser point cloud data to remove discrete noise points caused by smoke and dust, resulting in a denoised point cloud. The denoised point cloud is then spatially voxelized, dividing the continuous space into a regular 3D voxel grid, and the point cloud density within each voxel grid is calculated. Voxel grids with point cloud densities below a preset threshold are identified as open areas, while those with density above the preset threshold are identified as potential combustion material accumulation areas. Connectivity analysis is performed on potential combustion material accumulation areas to extract spatially connected voxel sets with volumes larger than the minimum fireball volume, and each voxel set is marked as an independent combustion object. Based on the spatial point cloud distribution of these independent combustion objects, their convex hulls or minimum bounding cubes are calculated, and the outer surface of the convex hull or minimum bounding cube is defined as the 3D spatial boundary of the flame combustion core region. Threshold segmentation is performed on the infrared thermal imaging temperature field data to identify the 2D hotspot distribution in high-temperature hazard areas, including applying Gaussian smoothing filtering to the infrared thermal imaging temperature field data to suppress sensor thermal noise and obtain a smoothed temperature field image. An adaptive threshold segmentation algorithm is used to process the smoothed temperature field image. This algorithm automatically calculates the segmentation threshold based on the local temperature gradient, extracting high-temperature regions where the temperature is higher than the background. Morphological opening operations are performed on the extracted high-temperature regions to eliminate small noise points and fill internal voids, forming complete connected domains of the high-temperature regions. The temperature statistical features of each high-temperature region's connected domain are calculated, including average temperature, maximum temperature, and temperature distribution variance. High-temperature regions whose average temperature exceeds the first danger temperature threshold, or whose maximum temperature exceeds the second danger temperature threshold, or whose temperature distribution variance exceeds a preset variance threshold, are marked as two-dimensional hotspots of high-temperature danger zones, and their geometric contours and temperature characteristics are recorded.

[0077] In the implementation, an initial fire occurred in an indoor storage area, where stacked shelves and some open flames were present. A 3D LiDAR scanner positioned above the scene acquired raw 3D LiDAR point cloud data. This raw data contained a large number of discrete noise points caused by smoke diffuse reflection. In the implementation, the acquired raw 3D LiDAR point cloud data underwent denoising filtering. A statistical outlier removal algorithm was used to identify and remove discrete noise points whose distance distribution within the neighborhood of each point exceeded a standard range, forming a denoised point cloud. The denoised point cloud was then spatially voxelized, dividing the entire storage space into a 3D voxel grid with a side length of 0.1 meters. All voxel grids were traversed, and the number of points falling into each voxel grid was counted as the point cloud density. In the implementation, voxel grids with a point cloud density below a threshold of 5 were classified as open areas, while voxel grids with a point cloud density above a threshold of 20 were classified as potential combustion material accumulation areas. A connected component analysis based on three-dimensional adjacency relationships is performed on potential combustion material accumulation areas to extract spatially connected voxel sets occupying a volume exceeding 0.5 cubic meters. Each such voxel set is labeled as an independent combustion object. It can be understood that, based on the spatial point cloud distribution of these independent combustion objects, their spatial convex hulls are calculated, and the outer surface of the triangular mesh of this convex hull is defined as the three-dimensional spatial boundary of the flame combustion core region.

[0078] In some embodiments, an infrared thermal imager working in conjunction with a 3D LiDAR synchronously acquires infrared thermal imaging temperature field data. In a specific implementation, Gaussian smoothing filtering is applied to the infrared thermal imaging temperature field data, and a 5x5 convolution kernel is used to perform convolution operations on the original temperature field image to suppress inherent thermal noise from the sensor, resulting in a smoothed temperature field image. An Otsu adaptive thresholding segmentation algorithm based on local region mean and standard deviation is used to process the smoothed temperature field image. This algorithm automatically calculates the segmentation threshold based on the temperature gradient of each pixel's neighborhood in the image, thereby extracting high-temperature regions with temperatures significantly higher than the background temperature. In some embodiments, morphological opening operations are performed on the extracted high-temperature regions. First, a 3x3 rectangular structuring element is used for erosion to eliminate small noise points, and then the same structuring element is used for dilation to fill the voids formed by noise or temperature measurement errors within the region, ultimately forming a connected high-temperature region with complete boundaries and internal connectivity. Optionally, temperature statistical features are calculated for each high-temperature region connected domain. These features include the average temperature of all pixels within the connected domain, the highest temperature among all pixels, and the temperature distribution variance of all pixel temperature values. It is understandable that a connected region of a high-temperature area whose average temperature exceeds the first dangerous temperature threshold of 300 degrees Celsius, or whose maximum temperature exceeds the second dangerous temperature threshold of 500 degrees Celsius, or whose temperature distribution variance exceeds the preset variance threshold of 1000, is marked as a two-dimensional hot spot of a high-temperature dangerous area, and its geometric contour is recorded as a polygonal contour. At the same time, its average temperature, maximum temperature and temperature distribution variance features are stored in association.

[0079] In practical implementation, the processing of the same indoor warehouse fire scenario demonstrated a parallel processing flow for two types of data. The processing of 3D laser point cloud data focused on identifying high-density clusters of burning objects from their spatial geometry and defining their 3D contours, while the processing of infrared thermal imaging temperature field data emphasized segmenting abnormally high-temperature regions exceeding safety thresholds from the temperature distribution and quantifying their thermal characteristics. At the data comparison level, the 3D spatial boundary provided by the 3D laser point cloud data defined the three-dimensional spatial occupancy and geometric dimensions of the core flame combustion region, while the 2D hotspot distribution provided by the infrared thermal imaging temperature field data provided the temperature intensity and gradient information of the plane in that region. The preprocessing results of these two types of data will be used for spatial fusion in subsequent steps. The 3D spatial boundary provides a framework for 3D registration, while the 2D hotspot distribution needs to be mapped to assign temperature attributes to the corresponding regions within this framework, thus jointly forming the basis of a 3D flame thermal composite model that includes temperature attributes and geometric structure.

[0080] In one embodiment of the present invention, the improved depth deterministic strategy gradient algorithm is optimized based on fire jet dynamics constraints and operational safety constraints, see reference. Figure 3Its working principle includes constructing a state space, the input of which is a discrete voxelized representation of a three-dimensional flame thermodynamic composite model, smoke concentration feature maps extracted in real time from visible light image video streams, and atmospheric wind speed and direction information. An action space is constructed, the output of which is a continuous jet attitude control quantity, including jet pitch angle, jet azimuth angle, turbofan speed, and jet medium mixing ratio. A fire jet dynamics simulation model is introduced as an environment model between the policy network and value network trained on the agent. This model is used to predict the jet trajectory, impact range, and cooling and suppression effect on the flame based on the current state and actions. A reward function is defined, consisting of a weighted sum of multiple reward terms, including the proximity reward between the jet impact point and the flame core, the reward for the reduction in flame model volume per unit time, the safety reward for the jet avoiding dangerous building components, and the penalty for the total amount of medium consumed by the jet. After each action is output by the policy network, the action is input into the fire jet dynamics simulation model to obtain the predicted next state and immediate reward. The policy network and value network are then trained offline or online using the data until the network converges. The policy network is the optimized attitude planner.

[0081] In practical implementation, the improved deep deterministic policy gradient algorithm was trained and deployed in a simulated chemical plant tank fire scenario. This scenario involved multiple closely spaced burning tanks with complex flame morphology and thermal distribution. A state space was constructed, with inputs including a discrete voxel representation of the 3D flame-thermal composite model, smoke concentration feature maps extracted in real-time from visible light video streams, and atmospheric wind speed and direction information. Specifically, the discrete voxel representation of the 3D flame-thermal composite model was achieved by dividing the entire monitoring space into a 0.2-meter-side cubic grid. Each voxel contained a binary "on fire" flag and a normalized temperature value. The smoke concentration feature maps extracted in real-time from the visible light video streams were obtained by processing video frames using a pre-trained convolutional neural network. The network output a single-channel grayscale image with the same resolution as the input image, where pixel values ​​represented the estimated smoke concentration for the corresponding image region. Atmospheric wind speed and direction information was provided by a digital weather station and encoded as a 3D vector containing wind speed magnitude, horizontal wind direction angle, and vertical airflow velocity.

[0082] In some embodiments, an action space is constructed, and the output of the action space is a continuous jet attitude control quantity, including jet pitch angle, jet azimuth angle, turbofan speed, and jet medium mixing ratio. Specifically, the output range of the jet pitch angle is -15 degrees to +60 degrees, the output range of the jet azimuth angle is 0 degrees to 350 degrees, the output range of the turbofan speed is 1000 rpm to 5000 rpm, and the jet medium mixing ratio refers to the volume ratio of foam concentrate to water, with an output range of 1% to 10%. These continuous values ​​together constitute a four-dimensional action vector. A fire jet dynamics simulation model is introduced as an environment model between the policy network and the value network trained on the agent. This fire jet dynamics simulation model is a numerical simulator based on simplified principles of computational fluid dynamics. The fire jet dynamics simulation model simulates the jetting process of fire-fighting media (foam or water) under specific jet pitch angle, jet azimuth angle, turbofan speed and jet medium mixing ratio based on the current state input and the action vector output by the strategy network. It also predicts the jet's trajectory under the action of the wind field, the jet's impact dispersion range on the target plane, and the jet's estimated cooling and suppression effect on the voxel temperature and combustion state in the three-dimensional flame thermodynamic composite model.

[0083] Optionally, a reward function is defined, which consists of a weighted sum of multiple reward items. These items include the reward for the proximity of the jet's impact point to the flame core, the reward for the reduction in the flame model's volume per unit time, the safety reward for the jet avoiding dangerous building components, and the penalty for the total amount of medium consumed by the jet. The specific mathematical expression of the reward function is as follows:

[0084]

[0085] in: This represents the total reward value for each decision step. The bonus represents the proximity of the jet's landing point to the flame core. The reward represents the reduction in the size of the flame model per unit of time. This represents a safety bonus for the jet stream avoiding dangerous building components. This represents the penalty for the total amount of medium consumed by the jet. , , , These are preset weight coefficients, corresponding to the four reward items. After each action output by the policy network, the four-dimensional action vector output by the policy network is input into the fire jet dynamics simulation model to obtain the predicted next state and the immediate reward calculated by the above reward function. It can be understood that during the training phase, the policy network and value network are trained offline or online using data composed of the current state, action, reward, and next state. The network parameters are iteratively updated using a deep reinforcement learning algorithm until the output action of the policy network tends to stabilize and the evaluation of the value network converges. At this point, the trained policy network is the optimized attitude planner.

[0086] In some embodiments, data comparison can clearly demonstrate the differences between the improved deep deterministic policy gradient algorithm and traditional planning methods. Traditional methods may rely on predetermined rules or simple feedback control, while the improved deep deterministic policy gradient algorithm, by constructing a state space and continuous action space containing composite sensing information, can handle high-dimensional, nonlinear fire-fighting scenarios. In specific implementations, traditional methods struggle to quantify the trade-off between the "safety reward of the jet avoiding dangerous building components" and the "reward of the flame model volume reduction per unit time," while the improved deep deterministic policy gradient algorithm uses preset weight coefficients in the reward function. and Explicit encoding is performed, and the policy is automatically optimized through interactive learning with the fire jet dynamics simulation model to maximize long-term rewards. Introducing the fire jet dynamics simulation model as an environmental model is crucial. It allows the agent to learn through extensive trial and error in a safe digital simulation environment, without performing potentially dangerous actions in actual fire scenarios. Simultaneously, the predicted states and rewards provided by the simulation model provide a data foundation for training the policy network and value network. It can be understood that the eventually converged policy network can directly output continuous spray attitude control quantities that satisfy both fire jet dynamics constraints and operational safety constraints, based on real-time perceived 3D flame thermodynamic composite model, smoke concentration feature map, and wind speed and direction information.

[0087] In one embodiment of the invention, a reward function is defined, including calculating the Euclidean distance between the jet impact point and the flame core, and using the negative exponential function value of the Euclidean distance as a reward for the proximity between the jet impact point and the flame core. The number of voxels marked as flames in the three-dimensional flame thermodynamic composite model is compared before and after the action is executed, and the reduced number of voxels is multiplied by a positive coefficient as a reward for the reduction in flame model volume per unit time. The three-dimensional spatial coordinates of dangerous building components in the fire scenario that are prohibited from being struck by the jet are pre-loaded, and the minimum distance between the predicted jet trajectory and the three-dimensional spatial coordinates of all dangerous building components is calculated. When the minimum distance is greater than the safe distance, a positive safety reward is given for the jet avoiding the dangerous building components; otherwise, a negative reward is given. The media consumption caused by all actions within a decision cycle is accumulated, and the negative value of the media consumption is multiplied by a coefficient as a penalty for the total amount of media consumed by the jet. All reward items and penalty items are linearly weighted and summed according to preset weights to obtain the output value of the reward function at each decision step. The Euclidean distance between the jet impact point and the flame core is calculated. The negative exponential function value of the Euclidean distance is used as the proximity reward between the jet impact point and the flame core. This includes simulating the predicted spatial coordinates of the jet impact point corresponding to the current action using a fire jet dynamics simulation model. A 3D flame thermodynamic composite model is queried to obtain the coordinates of the geometric center points of all voxels within the 3D spatial boundary of the flame combustion core region. The arithmetic mean of these geometric center point coordinates is calculated and used as the spatial coordinates of the flame core. Based on the predicted spatial coordinates of the jet impact point and the spatial coordinates of the flame core, the Euclidean distance in a 3D Cartesian coordinate system is calculated. The calculated Euclidean distance is used as an input variable and substituted into a preset exponential decay function. The output value of the preset exponential decay function is the negative exponential function value of the Euclidean distance. This negative exponential function value of the Euclidean distance is directly assigned as the specific numerical value of the proximity reward between the jet impact point and the flame core in the reward function.

[0088] In practical implementation, the process of defining the reward function is concretized in a simulated high-rise building exterior wall fire scenario. In this scenario, flames spread upwards along the building facade, surrounded by dangerous building components such as glass curtain walls and ventilation ducts that require protection. The Euclidean distance between the jet impact point and the flame core is calculated, and the negative exponential function value of the Euclidean distance is used as the proximity reward between the jet impact point and the flame core. This includes obtaining the predicted spatial coordinates of the jet impact point corresponding to the current action through a fire jet dynamics simulation model. A three-dimensional flame thermodynamic composite model is queried to obtain the coordinates of the geometric center points of all voxels marked as combustion states within the three-dimensional spatial boundary of the flame combustion core region. The arithmetic mean of these geometric center point coordinates is calculated and used as the spatial coordinates of the flame core. Based on the predicted spatial coordinates of the jet impact point and the spatial coordinates of the flame core, the Euclidean distance in a three-dimensional Cartesian coordinate system is calculated. The calculated Euclidean distance is used as an input variable and substituted into a preset exponential decay function. The output value of the preset exponential decay function is the negative exponential function value of the Euclidean distance. The proximity reward between the jet impact point and the flame core is then calculated. The calculation formula is: ,in It is the scaling factor for proximity rewards. It is the distance attenuation coefficient. This represents the calculated Euclidean distance between the predicted jet impact point and the flame core. It can be understood that... The value is directly assigned as the specific numerical value of the proximity reward between the corresponding jet landing point and the flame core in the reward function.

[0089] In some embodiments, the change in the number of voxels labeled as flames in the three-dimensional flame thermodynamic composite model before and after the action is performed is compared, and the reduced number of voxels is multiplied by a positive coefficient as the reward for the reduction in flame model volume per unit time. The calculation method is as follows:

[0090]

[0091] in: It is the fire extinguishing reward coefficient. It is the total number of voxels marked as flames in the three-dimensional flame thermodynamic composite model before the action is performed. This refers to the total number of voxels marked as flames in the three-dimensional flame thermodynamic composite model, updated by the fire jet dynamics simulation model after the action is executed. In practice, the three-dimensional spatial coordinates of hazardous building components in the fire scenario that are prohibited from being struck by the jet are pre-loaded. The minimum distance between the predicted jet trajectory and the three-dimensional spatial coordinates of all hazardous building components is calculated. When the minimum distance is greater than the safe distance, a positive safety bonus is given for the jet avoiding the hazardous building component; otherwise, a negative bonus is given. (Safety bonus for jet avoiding hazardous building components) Calculated according to the following rule: if the predicted jet trajectory is the closest distance to the surface of all hazardous building components. Greater than the preset safety distance ,but (C is a positive constant); if ,but ,in This is a penalty coefficient. It accumulates the media consumption caused by all actions within a decision cycle, multiplies the negative value of the media consumption by a coefficient, and serves as the penalty for the total media consumption of the jet. (Jet media consumption total penalty) The calculation method is as follows ,in It is a cost penalty coefficient. It is the total volume of water and foam concentrate consumed within a decision-making cycle.

[0092] Optionally, all reward and penalty items are linearly weighted and summed according to preset weights to obtain the output value of the reward function at each decision step. For the preset reward function weight coefficient configuration in a high-rise building fire simulation training cycle, please refer to Table 1:

[0093] Table 1: Preset Table of Weight Coefficients for Reward Function

[0094]

[0095] It is understandable that the total reward at each decision step... The calculation formula is In practice, data comparison can clarify the role of different reward items. For example, in a decision-making process, if the jet accurately hits the core of the flame but is too close to a dangerous ventilation duct, the calculated reward... The value may be high, but The value may be negative, resulting in a decrease in the total reward. The jet is suppressed. Conversely, if the jet deviates significantly from the flame target to completely avoid hazardous components, then... High value, but The value will be very low, at the same time The value may be zero or negative, and the total reward is... It won't be high either. In some embodiments, the total amount of jet-consumed medium is penalized. Weighting coefficients The relatively low setting indicates that fire extinguishing effectiveness and operational safety are prioritized over minimizing media consumption during training. However, the presence of this penalty term still guides the policy network to avoid meaningless, continuous high-volume spraying. Through this multi-objective weighted summation reward function design, the improved deep deterministic policy gradient algorithm is guided during training to find a spraying attitude policy that balances strike accuracy, fire extinguishing efficiency, obstacle avoidance, and resource consumption.

[0096] In one embodiment of the present invention, an improved deep deterministic policy gradient algorithm is used to calculate the optimal spray attitude of the fire-fighting turbofan monitor. This includes encoding the current three-dimensional flame thermodynamic composite model, the smoke features of the current frame extracted from the visible light image video stream, and the current atmospheric wind speed and direction information into an input vector required by the state space of the improved deep deterministic policy gradient algorithm. The input vector is then input into the trained policy network of the improved deep deterministic policy gradient algorithm. The policy network outputs an action vector, which corresponds to a set of specific values ​​for the spray pitch angle, spray azimuth angle, turbofan speed, and jet medium mixing ratio within the action space. The values ​​of each control variable are decoded from the action vector output by the policy network to form the preliminary spray attitude command for the current moment. The optimal spray attitude is dynamically compensated and corrected based on the atmospheric wind speed and direction information to generate the final execution command, driving the fire-fighting turbofan monitor actuator. This includes acquiring the current atmospheric wind speed and direction information and resolving the wind speed magnitude, wind direction angle, and short-term fluctuation variance of the wind speed. Based on a fluid dynamics model, the expected horizontal and vertical deviations of the jet as it travels from the outlet of the fire-fighting turbofan cannon to the predetermined impact point are calculated under specific wind speed and direction conditions. The dispersion range of the jet due to wind speed fluctuations during flight is estimated based on the short-term variance of wind speed, and this dispersion range is used as a safety margin. The calculated deviations are then superimposed in the opposite direction onto the azimuth and pitch angles of the initial jet attitude command to obtain the corrected attitude angles after wind speed deviation compensation. Based on the corrected attitude angles and the safety margin, the turbofan speed is fine-tuned to ensure that the jet still covers the core area of ​​the target within the predetermined dispersion range. Finally, the final execution command containing the compensated parameters is output.

[0097] In practical implementation, the process of calculating the optimal spray attitude of the fire-fighting turbofan cannon and performing dynamic compensation using an improved deep deterministic policy gradient algorithm is illustrated using a forest fire in a hilly area as an example, where unstable lateral winds exist. The current 3D flame-thermal composite model, the current frame smoke features extracted from the visible light image video stream, and the current atmospheric wind speed and direction information are jointly encoded into the input vector required by the state space of the improved deep deterministic policy gradient algorithm. In the specific implementation, the 3D flame-thermal composite model is discretized into a 20x20x15 voxel grid, with each voxel containing temperature and combustion state information, which is flattened into a 6000-dimensional vector. The current frame smoke features extracted from the visible light image video stream are processed into a 256-dimensional feature vector through an encoder network. The current atmospheric wind speed and direction information includes three scalar values: wind speed magnitude, horizontal wind angle, and vertical wind speed. After normalization, these data are concatenated into an input vector with a total dimension of 6259. The input vector is fed into a pre-trained, improved deep deterministic policy gradient algorithm policy network. This policy network is a multilayer perceptron with 6259 input layer nodes and 4 output layer nodes. The policy network outputs a four-dimensional action vector, corresponding to a specific set of values ​​in the action space for jet pitch angle, jet azimuth angle, turbofan speed, and jet medium mixing ratio. The values ​​of each control variable are decoded from the action vector output by the policy network. For example, the decoded values ​​are a jet pitch angle of 25 degrees, a jet azimuth angle of 120 degrees, a turbofan speed of 3200 rpm, and a jet medium mixing ratio of 3%. These values ​​constitute the initial jet attitude command for the current moment.

[0098] In some embodiments, the optimal spray attitude is dynamically compensated and corrected based on atmospheric wind speed and direction information to generate the final execution command, driving the fire-fighting turbofan monitor actuator. This includes acquiring the current atmospheric wind speed and direction information and analyzing the wind speed magnitude, wind direction angle, and short-term fluctuation variance of the wind speed. In a specific implementation, an ultrasonic anemometer deployed on the fire-fighting turbofan monitor turret provides a set of readings per second, for example, the current wind speed is 5.2 m / s, the wind direction angle is 60 degrees (0 degrees to the east, increasing clockwise), and the variance of the wind speed magnitude over the past 10 seconds is 0.8. Based on the fluid dynamics model, the expected horizontal and vertical offset of the jet as it travels from the fire-fighting turbofan monitor outlet to the predetermined impact point is calculated under the given wind speed and wind direction angle conditions. The formula for calculating the jet offset is... ,in These represent the expected offsets of the jet in the east-west, north-south, and vertical directions, respectively. It refers to the wind speed. It is the wind direction. It is the initial velocity of the jet (determined by the turbine fan speed). It is the density of the jet medium (related to the mixing ratio). This is the estimated time for the jet to travel from the exit point to the target point. It can be understood that, based on the short-term variance of wind speed fluctuations, the dispersion range of the jet due to wind speed fluctuations during flight is estimated, and this dispersion range is used as a safety margin. The calculated offset is then superimposed in the opposite direction onto the jet azimuth and pitch angles in the initial jet attitude command to obtain the corrected attitude angle after wind speed deviation compensation. For example, if the calculated expected offset of the jet on the horizontal plane is 1.5 meters east and 2.0 meters north, then the jet azimuth angle of 120 degrees in the initial jet attitude command is compensated and corrected in the opposite direction, adjusted to... The degree is adjusted, and a similar correction is made to the jet pitch angle.

[0099] Optionally, based on the corrected attitude angle and considering a safety margin, the turbofan rotation speed is fine-tuned to ensure that the jet can still cover the target core area within the predetermined dispersion range. The final execution command, including the compensated parameters, is then output. In the example of a forest fire in a hilly area, wind speed fluctuations cause dispersion of the jet's impact point. Assume the calculated jet dispersion range on the target plane is an ellipse with a major axis of 2 meters and a minor axis of 1 meter. To ensure the jet covers the flame core area, the turbofan rotation speed needs to be fine-tuned to change the initial kinetic energy and dispersion characteristics of the jet. The adjustment logic is: if the spatial scale of the flame core area is larger than the jet dispersion range, the turbofan rotation speed is maintained or slightly reduced to conserve the medium; if the scale of the flame core area is close to or smaller than the jet dispersion range, the turbofan rotation speed is increased to increase the jet velocity and reduce dispersion. The final execution command includes the azimuth and pitch compensated angle values, the fine-tuned turbofan rotation speed value, and the original jet medium mixing ratio. Data comparison clearly demonstrates the necessity of dynamic compensation. Refer to Table 2, which shows the values ​​of the initial attitude command, calculated wind-induced drift, compensation correction, and final execution command under crosswind conditions within a single decision cycle.

[0100] Table 2: Calculation Table for Dynamic Wind Field Compensation Correction

[0101]

[0102] In practice, this complete process, from state encoding and policy network inference to physical compensation, is executed periodically. After each final execution command, the system acquires new sensor data, updates the three-dimensional flame-thermal composite model, smoke characteristics, and wind speed and direction information, forms a new input vector, and initiates the next round of attitude calculation and compensation correction cycle, thereby achieving adaptive and precise strikes against dynamically changing fire and wind fields.

[0103] In one embodiment of the present invention, the online adjustment of the spray effect evaluation and attitude planning strategy based on sensor data feedback includes continuously acquiring new infrared thermal imaging temperature field data and three-dimensional laser point cloud data after executing the final execution command to drive the fire-fighting turbofan cannon to spray. Based on the new infrared thermal imaging temperature field data, the area change rate and average temperature change rate of the two-dimensional hot spot in the high-temperature hazard area are calculated. Based on the new three-dimensional laser point cloud data, the volume change rate and centroid displacement of the three-dimensional spatial boundary of the flame combustion core area are calculated. The area change rate, average temperature change rate, volume change rate, and centroid displacement are used as spray effect evaluation indicators. The spray effect evaluation indicators are fed back to the improved depth deterministic strategy gradient algorithm as a component of the next round of state input, or as the basis for calculating the immediate reward in the reward function, to achieve online adaptive adjustment of the attitude planning strategy. The real-time safety monitoring and action intervention mechanism includes inputting the spray attitude control quantity decoded from the action vector into an independent fast collision detection module after the action vector is output by the strategy network. The rapid collision detection module loads a 3D model of the fire-fighting turbofan monitor, 3D models of surrounding fixed obstacles, and a 3D composite model of flame and thermodynamics to simulate the motion of the monitor and the jet trajectory under spray attitude control. During the simulation, it detects whether the monitor or jet trajectory geometrically interferes with preset prohibited crossing zones in the 3D models of surrounding fixed obstacles or the 3D composite model of flame and thermodynamics. If geometric interference is detected, the action vector is immediately intercepted, and a preset safety avoidance strategy is triggered. This strategy generates an interference-free alternative action. The alternative action or the original action that passes the safety detection is taken as the final executable optimal spray attitude.

[0104] In practical implementation, the system collaboratively operates within a chemical plant fire scenario involving multiple large storage tanks and complex pipe corridors. This is achieved through online adjustment of the spray effect evaluation and attitude planning strategy based on sensor data feedback, along with real-time safety monitoring and action intervention mechanisms. After executing the final command to drive the fire-fighting turbofan cannon, the system continuously acquires new infrared thermal imaging temperature field data and 3D laser point cloud data. Based on the new infrared thermal imaging temperature field data, the system calculates the area change rate and average temperature change rate of the two-dimensional hotspots in the high-temperature hazard area. The area change rate is calculated by comparing the total pixel area of ​​all two-dimensional hotspots marked as high-temperature hazard areas in the current and previous infrared images, while the average temperature change rate is calculated by comparing the average temperature values ​​of these areas in the current and previous frames. Based on the new 3D laser point cloud data, the system calculates the volume change rate and centroid displacement of the 3D spatial boundary of the flame combustion core area. The volume change rate is calculated by comparing the total volume of the voxels marked as flames in the 3D flame thermodynamic composite model at the current and previous moments, while the centroid displacement is obtained by calculating the coordinate change of the 3D geometric center of the flame voxel set in 3D space.

[0105] In some embodiments, the area change rate, average temperature change rate, volume change rate, and centroid displacement are used as evaluation indicators for the spraying effect. These evaluation indicators are quantified into specific numerical values. Comprehensive evaluation indicators for spraying effect. This can be expressed as:

[0106]

[0107] in: The rate of change of area of ​​a two-dimensional hot spot representing a high-temperature hazard zone ( This represents the change in area. (for the initial area) Represents its average temperature change rate ( The change in temperature (Initial average temperature) The rate of volume change representing the core region of flame combustion ( It is the change in volume. (for the initial volume) The modulus representing the centroidal displacement. These are the weighting coefficients assigned to different indicators. It can be understood that the jet effect evaluation indicators are fed back to the improved deep deterministic policy gradient algorithm as part of the next round's state input, or as the basis for calculating the immediate reward in the reward function, thus achieving online adaptive adjustment of the attitude planning strategy. For example, when used as state input, the values ​​of these four evaluation indicators are normalized and concatenated into the state vector; when used as the basis for reward calculation, the volume change rate... Negative values ​​can be directly used as part of the fire extinguishing effect bonus, while centroid displacement Excessive fire could lead to the fire spreading, resulting in a negative reward.

[0108] In practice, real-time safety monitoring and action intervention mechanisms operate in parallel with the aforementioned online adjustment process. After the policy network outputs the action vector, the jet attitude control quantity decoded from the action vector is input into an independent rapid collision detection module. The rapid collision detection module loads the 3D model of the fire-fighting turbofan monitor, the 3D model of the surrounding fixed obstacles, and the 3D flame thermodynamic composite model to simulate the motion of the fire-fighting turbofan monitor and the jet trajectory under the jet attitude control quantity. In the chemical plant scenario, the 3D model of the surrounding fixed obstacles includes precise 3D geometric models of storage tanks, reaction towers, and pipe rack supports. The preset prohibited crossing areas in the 3D flame thermodynamic composite model may include storage tank areas containing sensitive chemicals that are not on fire. The system detects whether the fire-fighting turbofan monitor or jet trajectory geometrically interferes with the 3D model of the surrounding fixed obstacles or the preset prohibited crossing areas in the 3D flame thermodynamic composite model during the simulation. Geometric interference detection is achieved by calculating the shortest distance between the moving envelope and the triangular facets of the obstacle model; interference is determined to have occurred when the shortest distance is less than a safety threshold.

[0109] In some embodiments, if geometric interference is detected, the action vector is immediately intercepted, and a preset safety avoidance strategy is triggered. This strategy generates an interference-free alternative action. In specific implementations, the preset safety avoidance strategy can be a rule-based adjuster. For example, when a collision between the jet trajectory and a reaction tower model is detected, the strategy maintains the turbofan speed and mixing ratio unchanged, but fine-tunes the jet azimuth angle by a fixed step away from the reaction tower until the rapid collision detection module confirms that no interference will occur at the new azimuth angle. Optionally, the alternative action or the original action that passes the safety detection is taken as the final executable optimal jet attitude. Data comparison illustrates the effectiveness of this mechanism. In one decision, the improved deep deterministic policy gradient algorithm network might output a theoretically most efficient action for fire suppression. However, simulation by the rapid collision detection module reveals that the jet trajectory under this action will pass through a high-temperature pipe gallery (marked as a prohibited crossing area), posing a risk of secondary disasters. In this case, the action vector is intercepted, and the safety avoidance strategy generates an alternative action that adjusts the azimuth angle 5 degrees to the left. Although the expected fire extinguishing efficiency of this alternative action is slightly lower than the original action, it completely avoids interference with the utility tunnel, ensuring operational safety. Only actions that have undergone safety arbitration are issued for execution. Simultaneously, the state, alternative actions, and results generated by this decision are recorded for online learning in subsequent improvements to the deep deterministic policy gradient algorithm. This allows the policy network to proactively avoid such unsafe actions in similar future scenarios.

[0110] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for automatic planning of fire-fan-cannon jetting posture of multi-source sensing data, characterized in that, The method includes: Real-time monitoring data of the target area is obtained by various sensors deployed in the fire scene. The real-time monitoring data includes three-dimensional laser point cloud data, infrared thermal imaging temperature field data, visible light image video stream, and atmospheric wind speed and direction information. The three-dimensional laser point cloud data is preprocessed to extract the three-dimensional spatial boundary of the core area of ​​the flame combustion, and the infrared thermal imaging temperature field data is threshold segmented to identify the two-dimensional hot spot distribution in the high-temperature danger area. The three-dimensional spatial boundary of the core area of ​​the flame combustion is spatially fused and mapped with the two-dimensional hot spot distribution of the high-temperature hazard area to generate a three-dimensional flame thermal composite model, which includes temperature attributes and geometric structure. The three-dimensional flame thermal composite model, the visible light image video stream, and the atmospheric wind speed and direction information are jointly input into the improved depth deterministic strategy gradient algorithm, which is optimized based on fire jet dynamics constraints and operational safety constraints. Using the improved depth deterministic strategy gradient algorithm, the optimal spray attitude of the fire-fighting turbofan cannon is calculated. The optimal spray attitude includes the spray pitch angle, spray azimuth angle, turbofan speed, and jet medium mixing ratio.

2. The method of claim 1, wherein, The three-dimensional laser point cloud data is preprocessed to extract the three-dimensional spatial boundary of the core flame combustion region, including: The acquired raw 3D laser point cloud data is subjected to denoising filtering to remove discrete noise points caused by smoke and dust, forming a denoised point cloud. The denoised point cloud is spatially voxelized, dividing the continuous space into a regular three-dimensional voxel grid, and the point cloud density within each voxel grid is calculated. Voxel grids with point cloud density below a preset threshold are identified as open areas, while voxel grids with point cloud density above a preset threshold are identified as potential combustion material accumulation areas. Connectivity analysis is performed on the potential combustion material accumulation region to extract the set of voxels that are spatially connected and have a volume greater than the minimum fireball volume, and each set of voxels is marked as an independent combustion object. Based on the spatial point cloud distribution of the independent burning object, its convex hull or minimum circumscribed cube is calculated, and the outer surface of the convex hull or minimum circumscribed cube is defined as the three-dimensional spatial boundary of the core region of the flame combustion.

3. The method of claim 1, wherein, Threshold segmentation is performed on the infrared thermal imaging temperature field data to identify the two-dimensional hot spot distribution in high-temperature hazardous areas, including: Gaussian smoothing filtering is applied to the infrared thermal imaging temperature field data to suppress sensor thermal noise and obtain a smoothed temperature field image. The smoothed temperature field image is processed using an adaptive threshold segmentation algorithm, which automatically calculates the segmentation threshold based on the local temperature gradient to extract high-temperature regions where the temperature is higher than the background. Morphological opening operations are performed on the extracted high-temperature regions to eliminate small noise points and fill the voids inside the regions, forming a complete high-temperature region connected domain. Calculate the temperature statistical characteristics of the connected domains in each high-temperature region, including the average temperature, the maximum temperature, and the temperature distribution variance. The connected regions of high-temperature areas where the average temperature exceeds the first dangerous temperature threshold, the highest temperature exceeds the second dangerous temperature threshold, or the temperature distribution variance exceeds the preset variance threshold are marked as two-dimensional hot spots of the high-temperature dangerous area, and their geometric contours and temperature characteristics are recorded.

4. The method of claim 1, wherein, The improved depth deterministic strategy gradient algorithm is optimized based on fire jet dynamics constraints and operational safety constraints. Its working principle includes: A state space is constructed, the inputs of which are the discrete voxelized representation of the three-dimensional flame thermodynamic composite model, the smoke concentration feature map extracted in real time from the visible light image video stream, and the atmospheric wind speed and direction information; A motion space is constructed, and the output of the motion space is a continuous jet attitude control quantity, which includes the jet pitch angle, jet azimuth angle, turbofan speed and jet medium mixing ratio. Between the policy network and value network trained by the agent, a fire jet dynamics simulation model is introduced as an environment model. The fire jet dynamics simulation model is used to predict the jet trajectory, impact range and cooling and suppression effect on the flame based on the current state and actions. Define a reward function, which is composed of a weighted sum of multiple reward items. The reward items include the proximity reward between the jet landing point and the flame core, the reward for the reduction of the flame model volume per unit time, the safety reward for the jet avoiding dangerous building components, and the penalty for the total amount of medium consumed by the jet. After each action is output by the policy network, the action is input into the fire jet dynamics simulation model to obtain the predicted next state and immediate reward. The policy network and value network are then trained offline or online using the data until the network converges. The policy network is the optimized attitude planner.

5. The method of claim 4, wherein, The defined reward function includes: Calculate the Euclidean distance between the jet impact point and the flame core, and use the negative exponential function value of the Euclidean distance as the proximity bonus between the jet impact point and the flame core. Compare the changes in the number of voxels marked as flames in the three-dimensional flame thermodynamic composite model before and after the execution action, and multiply the reduced number of voxels by a positive coefficient as the reward for the reduction in the volume of the flame model per unit time. The three-dimensional spatial coordinates of dangerous building components that are prohibited from being hit by jets in the fire scene are pre-loaded. The minimum distance between the predicted jet trajectory and the three-dimensional spatial coordinates of all dangerous building components is calculated. When the minimum distance is greater than the safe distance, a positive safety reward is given for the jet to avoid dangerous building components; otherwise, a negative reward is given. Accumulate the media consumption caused by all actions within a decision cycle, and multiply the negative value of the media consumption by a coefficient as a penalty for the total amount of media consumed by the jet. All reward and penalty items are linearly weighted and summed according to preset weights to obtain the output value of the reward function at each decision step.

6. The method of claim 4, wherein, The optimal spray attitude of the fire-fighting turbofan cannon is calculated using the improved depth deterministic strategy gradient algorithm, including: The current three-dimensional flame thermal composite model, the current frame smoke features extracted from the visible light image video stream, and the current atmospheric wind speed and direction information are jointly encoded into the input vector required by the state space of the improved depth deterministic strategy gradient algorithm. The input vector is input into the policy network of the improved deep deterministic policy gradient algorithm that has been trained. The strategy network outputs an action vector, which corresponds to a set of specific values ​​for the jet pitch angle, jet azimuth angle, turbofan speed, and jet medium mixing ratio in the action space. The values ​​of each control variable are decoded from the action vector output by the policy network to form the initial jet attitude command at the current moment.

7. The method of claim 6, wherein, The method further includes: The optimal spraying attitude is dynamically compensated and corrected based on the atmospheric wind speed and direction information to generate a final execution command, driving the fire-fighting turbofan monitor actuator to move, including: Obtain the atmospheric wind speed and direction information at the current moment, and analyze the wind speed magnitude, wind direction angle, and short-term fluctuation variance of wind speed from it; Based on the fluid dynamics model, the expected deviations in the horizontal and vertical directions of the jet as it travels from the outlet of the fire-fighting turbofan cannon to the predetermined impact point are calculated under the conditions of the wind speed and wind direction angle. Based on the short-term fluctuation variance of the wind speed, the dispersion range of the jet during flight due to wind speed fluctuations is estimated, and the dispersion range is used as a safety margin. The calculated offset is superimposed in the opposite direction onto the jet azimuth and jet pitch angles in the initial jet attitude command to obtain the corrected attitude angle after wind speed deviation compensation. Based on the corrected attitude angle and combined with the safety margin, the turbofan speed is finely adjusted to ensure that the jet can still cover the core area of ​​the target within the predetermined dispersion range, and finally the final execution command containing the compensated parameters is output.

8. The method of claim 3, wherein, The method also includes online adjustment of the jet effect evaluation and attitude planning strategy based on sensor data feedback: After executing the final execution command to drive the fire-fighting turbofan cannon to spray, new infrared thermal imaging temperature field data and three-dimensional laser point cloud data are continuously acquired. Based on the new infrared thermal imaging temperature field data, the area change rate and average temperature change rate of the two-dimensional hot spot in the high-temperature danger zone are calculated. Based on the new three-dimensional laser point cloud data, the volume change rate and centroid displacement of the three-dimensional spatial boundary of the flame combustion core region are calculated. The area change rate, average temperature change rate, volume change rate, and centroid displacement are used as evaluation indicators for the spraying effect. The jetting effect evaluation index is fed back to the improved depth deterministic policy gradient algorithm as a component of the state input for the next round, or as the basis for calculating the immediate reward in the reward function, thereby realizing the online adaptive adjustment of the attitude planning strategy.

9. The method of claim 4, wherein, The planning process using the improved deep deterministic policy gradient algorithm also includes a real-time safety monitoring and action intervention mechanism: After the policy network outputs the action vector, the jet attitude control quantity decoded from the action vector is input into an independent fast collision detection module; The rapid collision detection module loads a three-dimensional model of the fire-fighting turbofan cannon body, a three-dimensional model of the surrounding fixed obstacles, and the three-dimensional flame thermal composite model to simulate the motion of the fire-fighting turbofan cannon body and the jet trajectory under the jet attitude control. During the simulation, the test detects whether the fire-fighting turbofan cannon body or jet trajectory geometrically interferes with the three-dimensional model of the surrounding fixed obstacles or the preset prohibited crossing area in the three-dimensional flame thermal composite model. If geometric interference is detected, the action vector is immediately intercepted, and a preset safety avoidance strategy is triggered to generate an interference-free alternative action. The alternative action or the original action that passes the safety test is taken as the optimal injection posture that can be executed at the end.

10. The method for automatic planning of the spray attitude of a fire-fighting turbofan monitor based on multi-source sensor data according to claim 5, characterized in that, The calculation of the Euclidean distance between the jet impact point and the flame core, using the negative exponential function value of the Euclidean distance as a proximity bonus between the jet impact point and the flame core, includes: The spatial coordinates of the predicted jet impact point corresponding to the current action are obtained by simulating the dynamics of the fire jet. The three-dimensional flame thermodynamic composite model is queried to obtain the geometric center coordinates of all voxels inside the three-dimensional spatial boundary of the flame combustion core region, and the arithmetic mean of the geometric center coordinates is calculated. The arithmetic mean is then used as the spatial coordinates of the flame core. Based on the spatial coordinates of the predicted jet impact point and the spatial coordinates of the flame core, the Euclidean distance in the three-dimensional Cartesian coordinate system is calculated. The calculated Euclidean distance is used as an input variable and substituted into a preset exponential decay function. The output value of the preset exponential decay function is the negative exponential function value of the Euclidean distance. The negative exponential function value of the Euclidean distance is directly assigned as the specific value of the proximity reward between the jet landing point and the flame core in the reward function.

Images