Urban high-rise building fire unmanned vehicle three-dimensional modeling and path planning system

Abstract

The present application relates to a kind of urban high-rise building fire unmanned vehicle three-dimensional modeling and path planning system, specifically relates to unmanned vehicle fire fighting field, the scheme is fused sparse sensor data and quantifies uncertainty by belief field construction module, risk field prediction module is forward-lookingly deduced fire by means of physical information neural network, and time-varying risk probability field is generated;Path planning module combines the risk field with power, communication constraint into multidimensional time-varying cost map, and the optimal trajectory of fast marching tree search with time window is adopted;Perception update module identifies high uncertainty area according to prediction variance, and initiatively detects by information gain and power budget decision, and is updated by using bayes and triggers re-planning;The closed loop formed as a whole makes unmanned vehicle real-time perception fire evolution, initiatively reduces prediction blind area, and forward-lookingly avoids risk, improves survivability and task reliability in high-temperature dense smoke environment.

Description

Technical Field

[0001] This invention relates to the field of drone firefighting, and more specifically, to a three-dimensional modeling and path planning system for drones used in urban high-rise building fires. Background Technology

[0002] With the acceleration of urbanization, the number of high-rise buildings has surged. These buildings have complex internal structures and are densely populated. Once a fire breaks out, the flames can spread rapidly along vertical channels such as exterior wall insulation materials, vertical pipe shafts, ventilation ducts, and cable trays, creating a three-dimensional combustion situation. Simultaneously, unforeseen events such as sudden wind changes, deflagration, or secondary ignition further exacerbate the uncertainty of the fire, making the rescue environment constantly changing. In such scenarios, drones, with their advantages of maneuverability and the ability to carry multiple sensors, are widely used for tasks such as fire reconnaissance, material delivery, and locating trapped personnel. However, high-rise building fires are often accompanied by high temperatures, dense smoke, strong heat radiation, and strong updrafts. When drones are performing continuous reconnaissance or material delivery missions, their initially planned flight paths may be surrounded by the spreading flames within minutes, leading to equipment damage or even secondary disasters. Therefore, enabling drones to predict the dynamic evolution of fires and adjust their paths in real time during flight to avoid risk areas that change over time has become a key bottleneck restricting the practical effectiveness of firefighting drones.

[0003] Currently, UAV path planning commonly employs methods such as Algorithm A, Rapid Exploratory Random Tree (RRT), or Model Predictive Control (MPC). These methods typically assume that the position and shape of environmental obstacles remain unchanged within the planning period, relying solely on current sensor data for static or quasi-static obstacle avoidance. A few studies have introduced a time dimension, but these are limited to handling moving obstacles with known motion patterns, such as other UAVs or vehicles. For threats like fires, which spread continuously and change drastically, existing solutions cannot predict the risk situation in the next few minutes. On the other hand, while sophisticated fire propagation models such as FDS or FLUENT can simulate the fire spread process, their computational demands are enormous, making real-time operation on airborne embedded platforms difficult. Furthermore, the infrared or visible light sensors on UAVs have limited detection ranges and cannot detect changes in distant fire conditions in advance. Therefore, there is an urgent need to develop a lightweight fire propagation prediction method that can rapidly generate a risk probability field for the next 5 to 10 minutes using limited real-time measurement data (such as temperature, wind speed, and combustible material type), and integrate this time-varying risk field as a dynamic constraint into a rolling time-domain path planning framework, thereby enabling UAVs to possess forward-looking autonomous risk avoidance capabilities in fire propagation environments. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a 3D modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs). The system utilizes a belief field construction module, a risk field prediction module, a path planning module, and a perception update module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including:

[0006] The belief field construction module receives multi-source sparse observation data, including temperature distribution data from an infrared thermal imager and wind field data from an anemometer, collected by the UAV's onboard sensors. It performs spatiotemporal alignment and uncertainty measurement to generate the belief field at the current moment. The belief field includes the current risk field on the grid, wind field distribution data, and boundary constraints of the observed point set.

[0007] Risk field prediction module: Using the belief field as the initial and boundary conditions, the physical information neural network is used to approximate the partial differential equation of fire propagation and output the time-varying risk probability field within the future preset time window. The time-varying risk probability field includes the combustion probability of each location at each time and its prediction variance.

[0008] Path planning module: The time-varying risk probability field is merged with the drone's current remaining battery power and current communication link quality into a multi-dimensional time-varying cost map. A fast traversing tree with time windows is used to perform a forward-looking path search on the multi-dimensional time-varying cost map to generate an executable trajectory in the current rolling time domain.

[0009] Perception Update Module: Based on the prediction variance, it identifies hotspot areas with high uncertainty. It decides whether to insert a waypoint for the probe based on the information gain value of the hotspot area and the power budget required to perform the probe. After the probe is performed, it uses Bayesian update to fuse the new temperature observation data obtained by the probe into the belief field, triggering the replanning of the risk field prediction module and the path planning module.

[0010] In a preferred embodiment, the belief field construction module performs the following specific operations when performing spatiotemporal alignment and uncertainty measurement:

[0011] Temperature distribution data and wind field data from multi-source sparse observation data are mapped onto regular grid nodes on the building's exterior surface, respectively.

[0012] For each grid node, a spatial radius and time window are defined with it as the center, and temperature observation index set and wind field observation index set falling within the spatiotemporal neighborhood are selected;

[0013] A weighted kernel regression method was used, with spatial distance, time difference, and the reciprocal of sensor noise parameters as weights, to estimate the temperature value and wind speed vector of each grid node. Based on this, the information density of each grid node was calculated. The information density was obtained by logarithmically weighting the spatial weight, time weight, and sensor noise parameter contribution of each observation in the neighborhood.

[0014] The inverse of the information density is then defined as the uncertainty measure of the grid node. The smaller the uncertainty measure, the more reliable the estimate. Finally, the temperature estimate, wind field estimate, uncertainty measure, and boundary constraints of the observed point set for each grid node are packaged together to generate the belief field.

[0015] In a preferred embodiment, when generating the belief field, an initial risk value mapping is also performed:

[0016] The temperature estimate of each grid node is converted into an initial risk value through a differentiable risk indicator function, which is set to a piecewise linear form: when the temperature estimate is less than or equal to the safe temperature threshold, the initial risk value is zero.

[0017] When the temperature estimate is between the safe temperature threshold and the critical damage temperature threshold, the initial risk value is equal to the difference between the temperature estimate and the safe temperature threshold divided by the difference between the critical damage temperature threshold and the safe temperature threshold.

[0018] When the temperature estimate is greater than or equal to the critical damage temperature threshold, the initial risk value is one; the obtained initial risk value and the wind field estimate together constitute the current risk field in the belief field.

[0019] In a preferred embodiment, the risk field prediction module performs the following specific operations when using a physical information neural network to approximate the partial differential equation of fire propagation:

[0020] Construct a composite loss function, which is obtained by weighted summation of data fitting term, physical residual term, boundary condition term and initial condition term;

[0021] The data fitting term uses the temperature estimate of each grid node in the belief field as a supervision signal and the reciprocal of the uncertainty measure of each grid node as a weight, so that high-confidence observations have a greater influence.

[0022] The physical residual term is randomly sampled from a set of spatial-temporal collocation points within the domain, and the sum of squared residuals of the physical information neural network output with respect to the partial differential equation of fire propagation is calculated.

[0023] The boundary condition terms constrain the network output to be consistent with the preset boundary values ​​on the original geometric boundaries of the building facade;

[0024] The initial condition term forces the network's output at the current moment to be consistent with the temperature estimate in the belief field, and also uses the reciprocal of the uncertainty metric as the weight.

[0025] In a preferred embodiment, after training the physical information neural network is completed, uncertainty quantification and time-varying risk probability field generation are also performed.

[0026] Uncertainty quantification employs the Monte Carlo dropout method, which randomly discards some neurons in the hidden layer of the network during inference, performs multiple forward propagations for each spatiotemporal point, obtains multiple temperature prediction values, and then calculates the fluctuation amplitude of these prediction values. The fluctuation amplitude is the prediction variance in the time-varying risk probability field.

[0027] The time-varying risk probability field generation converts the average temperature prediction from multiple forward propagations into the combustion probability through a differentiable risk indicator function. The risk indicator function is the same as the piecewise linear function used in the belief field construction module, and the fluctuation amplitude is normalized and used as the prediction dispersion. The prediction dispersion is the normalized prediction variance.

[0028] Finally, each spatiotemporal point outputs a tuple containing the combustion probability and the predicted dispersion, covering all moments within the preset future time window.

[0029] In a preferred embodiment, when the path planning module merges the time-varying risk probability field with the UAV's current remaining battery power and current communication link quality into a multi-dimensional time-varying cost map, the specific operation is as follows:

[0030] For each spatiotemporal point, the instantaneous cost is defined as the weighted sum of three terms: the combustion probability, the predicted dispersion, and the communication quality penalty function.

[0031] Combustion probability comes from time-varying risk probability field, prediction dispersion is normalized prediction variance, communication quality penalty function is exponential function with natural constant e as base, the preset communication quality sensitivity coefficient with negative exponential independent variable is multiplied by the current communication link quality, the lower the communication quality, the greater the penalty.

[0032] The weight coefficients of the three items are fixed. By default, the weight of combustion probability is the largest, followed by the weight of predicted dispersion, and the weight of communication quality penalty is the smallest. The weight coefficients can be dynamically adjusted according to task priority.

[0033] In a preferred embodiment, when using a fast progress tree with a time window for prospective path search, the specific operation is as follows:

[0034] The search space is expanded into a state-time space, where each node consists of a spatial location and an arrival time. The cost of an edge is defined as the weighted sum of the integral of the instantaneous cost and the integral of the energy consumption during the movement from the starting node to the target node. The instantaneous cost comes from a multidimensional time-varying cost map, and the energy consumption integral is the integral of the energy consumption rate per unit time along time. The energy consumption rate per unit time is the cube of the velocity vector multiplied by a preset first aerodynamic parameter plus a preset second aerodynamic parameter.

[0035] The feasibility of an edge must meet three conditions simultaneously: the probability of combustion at any time on the path does not exceed the preset hard danger threshold, the cumulative energy consumption does not exceed the available power after deducting the preset safety reserve power from the drone's current remaining power, and the quality of the communication link throughout the path is not lower than the preset minimum communication threshold.

[0036] In a preferred embodiment, when the perception update module determines hotspot areas with high uncertainty based on the prediction variance, the specific operation is as follows:

[0037] Extract the prediction variance of each spatial location from the time-varying risk probability field for all moments within a future preset time window, and calculate the average prediction variance of each spatial location in the time dimension.

[0038] Spatial locations with average prediction variance exceeding a preset hotspot threshold are marked as candidate hotspots;

[0039] Connectivity clustering is performed on candidate hotspots to remove isolated points, resulting in several consecutive hotspot regions, each containing a set of adjacent grid nodes.

[0040] In a preferred embodiment, when deciding whether to insert a probe yaw waypoint based on the information gain value of the hotspot area and the power budget required to perform the probe, the specific operation is as follows:

[0041] For each hotspot region, calculate the uncertainty of the current region, which is equal to the sum of the average prediction variances of all grid nodes in that region;

[0042] Assuming that the prediction variance of all grid nodes in the region drops to a preset baseline level after detection, the expected residual uncertainty is calculated. The expected residual uncertainty is equal to the number of grid nodes in the region multiplied by the preset baseline level.

[0043] Information gain is defined as the current uncertainty minus the expected residual uncertainty.

[0044] Estimate the total energy consumption required to fly from the current location to the center of the hotspot area and back to the original route. The total energy consumption is equal to the round-trip distance multiplied by the energy consumption rate per unit distance. The energy consumption rate per unit distance is obtained by dividing the energy consumption rate per unit time by the preset cruise speed.

[0045] The detection utility function is defined as the information gain value divided by the total energy consumption plus a preset minimum positive value;

[0046] When the detection utility function is greater than the preset utility threshold and the total energy consumption does not exceed the available power after deducting the preset safety reserve power from the current remaining power, a decision is made to insert a detection waypoint.

[0047] In a preferred embodiment, when the newly obtained temperature observation data is fused into the belief field using Bayesian update after the detection is performed, and the replanning of the risk field prediction module and the path planning module is triggered, the specific operation is as follows:

[0048] After the drone flies to the center of the hotspot area, it uses an airborne infrared thermal imager to acquire new temperature observation data of the grid nodes in the area. Each new temperature observation data is accompanied by sensor noise variance, which is the square of the sensor noise parameter.

[0049] After returning to the original route, for each observed grid node, the original uncertainty measure and temperature estimate are fused with the newly observed sensor noise variance and temperature value using the Bayesian update formula to obtain the fused uncertainty measure and temperature estimate.

[0050] The fused uncertainty measure is equal to the reciprocal of the sum of the original uncertainty measure and the reciprocal of the sensor noise variance; the fused temperature estimate is equal to the fused uncertainty measure multiplied by the original temperature estimate divided by the original uncertainty measure plus the weighted sum of the new temperature value divided by the sensor noise variance.

[0051] The updated temperature estimate and uncertainty measure are written back to the corresponding grid node in the belief field, and the new observation point is added to the set of observed points. Using the updated belief field as the new input, the risk field prediction module is immediately triggered to re-perform the physical information neural network deduction, and the path planning module is triggered to re-plan the trajectory.

[0052] The beneficial effects of this invention are as follows: the belief field construction module integrates sparse sensor data and quantifies uncertainty; the risk field prediction module uses a physical information neural network to prospectively extrapolate the fire situation and generate a time-varying risk probability field; the path planning module merges this risk field with power and communication constraints into a multi-dimensional time-varying cost map and uses a fast walking tree with time windows to search for the optimal trajectory; the perception update module identifies high uncertainty areas based on prediction variance, actively detects through information gain and power budget decisions, and uses Bayesian updates to incorporate new observations and trigger replanning; the overall closed loop enables the UAV to perceive the fire evolution in real time, actively reduce prediction blind spots, and proactively avoid risks, thereby improving its survivability and mission reliability in high-temperature and dense smoke environments. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0057] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0058] Example 1

[0059] This embodiment provides, for example Figure 1-2 The system shown is a 3D modeling and path planning system for UAVs used in urban high-rise building fires, specifically including:

[0060] The belief field construction module receives multi-source sparse observation data, including temperature distribution data from an infrared thermal imager and wind field data from an anemometer, collected by the UAV's onboard sensors. It performs spatiotemporal alignment and uncertainty measurement to generate the belief field at the current moment. The belief field includes the current risk field on the grid, wind field distribution data, and boundary constraints of the observed point set.

[0061] Risk field prediction module: Using the belief field as the initial and boundary conditions, the physical information neural network is used to approximate the partial differential equation of fire propagation and output the time-varying risk probability field within the future preset time window. The time-varying risk probability field includes the combustion probability of each location at each time and its prediction variance.

[0062] Path planning module: The time-varying risk probability field is merged with the drone's current remaining battery power and current communication link quality into a multi-dimensional time-varying cost map. A fast traversing tree with time windows is used to perform a forward-looking path search on the multi-dimensional time-varying cost map to generate an executable trajectory in the current rolling time domain.

[0063] Perception Update Module: Based on the prediction variance, it identifies hotspot areas with high uncertainty. It decides whether to insert a waypoint for the probe based on the information gain value of the hotspot area and the power budget required to perform the probe. After the probe is performed, it uses Bayesian update to fuse the new temperature observation data obtained by the probe into the belief field, triggering the replanning of the risk field prediction module and the path planning module.

[0064] In this embodiment, it is specifically necessary to explain that the belief field construction module performs spatiotemporal alignment and uncertainty measurement as follows:

[0065] Temperature distribution data and wind field data from multi-source sparse observation data are mapped onto regular grid nodes on the building's exterior surface. The regular grid nodes on the building's exterior surface are generated as follows: First, a 3D point cloud model or building information model of the building's exterior is obtained. Then, grid nodes are evenly distributed on the exterior surface at a spacing of 1 meter by 1 meter. For corners or uneven areas, the spacing is increased to 0.5 meters by 0.5 meters to ensure that each grid node corresponds to a unique spatial coordinate.

[0066] For each grid node, a spatial radius and time window are defined centered on it, and temperature observation index sets and wind field observation index sets that fall within the spatiotemporal neighborhood are selected. The spatial radius is set to five meters, and the time window is set to thirty seconds. For each grid node, all observation data are traversed, and only observation points with a spatial distance of less than five meters and a time difference of less than thirty seconds are retained to form the temperature observation index set and wind field observation index set for that node. If the number of observation points in the neighborhood of a node is less than three, the spatial radius is expanded to ten meters or the time window is expanded to sixty seconds until the minimum number requirement is met.

[0067] A weighted kernel regression method is used, employing spatial distance, time difference, and the reciprocal of sensor noise parameters as weights to estimate the temperature value and wind speed vector for each grid node. The specific calculation process of weighted kernel regression is as follows: For each grid node, firstly, the spatial weight of each observation point in its neighborhood is calculated. The spatial weight is determined by the Euclidean distance between the observation point and the grid node using a Gaussian kernel function; the closer the distance, the greater the weight. Secondly, the time weight is calculated, determined by the exponential decay function of the time difference between the observation time and the current time; the smaller the time difference, the greater the weight. Then, the sensor accuracy weight is calculated, taken as the reciprocal of the sensor noise parameter corresponding to the observation point; the smaller the noise parameter, the higher the sensor accuracy, and the greater the weight. The above three weights are multiplied to obtain the comprehensive weight for each observation point. Then, a weighted average is applied to the temperature value or wind speed vector components of all observation points to obtain the temperature estimate and wind speed vector estimate for that grid node. If there are only temperature observations but no wind field observations in the neighborhood of a node, the wind field estimate is temporarily missing, awaiting subsequent interpolation to fill the gap.

[0068] Based on this, the information density is calculated for each grid node. The information density is obtained by weighted summation of the logarithmic contributions of the spatial weights, temporal weights, and sensor noise parameters of each observation in the neighborhood. The calculation process of the information density is as follows: For each grid node, all temperature observation points in its neighborhood are traversed, and the contribution value of each observation point to the information density is calculated. The contribution value is equal to the spatial weight of the observation point multiplied by the temporal weight, and then multiplied by the logarithm of the square of the sensor noise parameter with the natural constant e as the base. The contribution values ​​of all observation points are summed to obtain the information density of the grid node. This information density comprehensively reflects the number of observations in the neighborhood, the uniformity of spatial distribution, and the accuracy level of the sensor. The higher the information density, the more reliable the estimation of the node.

[0069] The reciprocal of the information density is defined as the uncertainty measure of the grid node. The smaller the uncertainty measure, the more reliable the estimation. To prevent division by zero due to zero information density, a very small positive value is added to the information density before taking the reciprocal. This positive value is 10 to the power of -8. The uncertainty measure is dimensionless. The smaller its value, the more abundant and high-quality observations are around the grid node, and the more reliable the estimation result is. Conversely, if the uncertainty measure is large, it indicates that there are few observations or high sensor noise near the node, and the estimation result should be used with caution.

[0070] Finally, the temperature estimate, wind field estimate, uncertainty measure, and boundary constraints of the observed point set for each grid node are packaged together to generate a belief field. The boundary constraints of the observed point set refer to the outer contour of the connected region formed by the grid nodes covered by at least one observation, as well as the original geometric boundaries of the building facade (such as windows, balconies, roof edges, etc.). These boundary constraints come from the offline loaded building information model or pre-scanned point cloud and are used to limit the solution domain of the subsequent physical information neural network to avoid invalid inferences in non-flying areas.

[0071] When generating the belief field, an initial risk value mapping is also performed:

[0072] The temperature estimate of each grid node is converted into an initial risk value through a differentiable risk indicator function. The risk indicator function is set to a piecewise linear form: when the temperature estimate is less than or equal to the safe temperature threshold, the initial risk value is zero. The safe temperature threshold is set to 60 degrees Celsius. This value comes from the highest ambient temperature that the UAV's fuselage materials and electronic components can withstand. Below this temperature, it is considered that there is no risk of thermal damage, so the initial risk value is set to zero.

[0073] When the temperature estimate is between the safe temperature threshold and the critical damage temperature threshold, the initial risk value is equal to the difference between the temperature estimate and the safe temperature threshold divided by the difference between the critical damage temperature threshold and the safe temperature threshold. The critical damage temperature threshold is set to 300 degrees Celsius, which is taken from the temperature at which the shell material of a typical fire-fighting drone begins to undergo irreversible deformation or combustion. Within the range of 60 degrees Celsius to 300 degrees Celsius, the initial risk value increases linearly from zero to one, indicating that the degree of risk increases uniformly with increasing temperature. To make the piecewise linear function differentiable at the inflection points, cubic spline interpolation is used in the neighborhood of the two inflection points (60 degrees Celsius and 300 degrees Celsius) for smoothing in actual calculations. The width of the smoothing interval is set to five degrees Celsius to ensure that the entire function is differentiable everywhere in the domain, so as to meet the requirements of the smoothness of the loss function during the backpropagation of the physical information neural network.

[0074] When the estimated temperature is greater than or equal to the critical damage temperature threshold, the initial risk value is one; when the temperature reaches or exceeds 300 degrees Celsius, the location is considered to be in direct contact with the flame or in a high-intensity heat radiation zone, and the drone cannot pass safely, so the initial risk value saturates to one; the obtained initial risk value and the wind field estimate together constitute the current risk field in the belief field; the current risk field is a scalar field with the same grid resolution as the belief field, and each grid node stores an initial risk value and a wind speed vector; this risk field will serve as one of the initial conditions for the subsequent risk field prediction module, used to drive the physical information neural network to predict the future spread trend of the fire.

[0075] In this embodiment, it is specifically necessary to explain that when the risk field prediction module uses a physical information neural network to approximate the solution of the partial differential equation for fire propagation, the specific operation is as follows:

[0076] A composite loss function is constructed, which is obtained by weighted summation of the data fitting term, physical residual term, boundary condition term, and initial condition term. The weighting coefficients of the composite loss function are as follows: the weight coefficient of the data fitting term is 1, the weight coefficient of the physical residual term is 0.1, the weight coefficient of the boundary condition term is 0.01, and the weight coefficient of the initial condition term is 1. The selection principle for these coefficients is to prioritize the accuracy of data fitting and initial conditions, while appropriately constraining physical laws and boundary conditions. If the loss value of any term is too large or too small during training, the weight coefficients can be adaptively adjusted to bring all losses to the same order of magnitude. Specifically, the mean of all losses is calculated every 500 training steps, and then the weights are redistributed proportionally to make the weighted contributions of each loss approximately equal.

[0077] The data fitting term uses the temperature estimate of each grid node in the belief field as a supervision signal and the reciprocal of the uncertainty measure of each grid node as a weight, so that observations with high confidence have a greater influence. The specific calculation process of the data fitting term is as follows: First, all grid nodes in the belief field are traversed. For each grid node, its temperature estimate and uncertainty measure are extracted. Then, the square of the difference between the temperature prediction value and the temperature estimate output by the physical information neural network for the grid node at the current time is calculated. This squared difference is multiplied by the reciprocal of the uncertainty measure of the node, where a very small positive value has been added beforehand to prevent division by zero. This very small positive value is 10 to the power of negative 8. The weighted squared differences of all grid nodes are summed and then divided by the total number of grid nodes to obtain the final value of the data fitting term. This processing makes nodes with smaller uncertainty measures (i.e., higher observation confidence) contribute a larger error signal, guiding the network to prioritize fitting reliable regions.

[0078] The physical residual term is calculated by randomly sampling a set of spatial-temporal collocation points within the domain and calculating the sum of squared residuals of the physical information neural network output with respect to the partial differential equation for fire propagation. The specific calculation process for the physical residual term is as follows: First, ten thousand spatial-temporal collocation points are randomly generated within the solution domain. The spatial coordinates are uniformly distributed within the area covered by the grid nodes on the building's outer surface, and the temporal coordinates are uniformly distributed between the current moment and the end of a preset future time window. For each collocation point, automatic differentiation is used to calculate the partial derivatives of the physical information neural network output with respect to the spatial and temporal coordinates, including the first-order partial derivative of temperature with respect to time, the first-order partial derivative of temperature with respect to spatial coordinates (gradient), and the second-order partial derivative of temperature with respect to spatial coordinates (divergence). These partial derivatives are then substituted into the left-hand side of the partial differential equation for fire propagation to calculate the residual value. The sum of the squared residuals of all collocation points is then divided by the total number of collocation points to obtain the value of the physical residual term. Automatic differentiation is implemented through a backpropagation algorithm, eliminating the need for manual derivation of derivative expressions and ensuring high computational accuracy.

[0079] The boundary condition term ensures that the network output on the original geometric boundary of the building facade is consistent with the preset boundary value. The specific calculation process of the boundary condition term is as follows: First, all boundary mesh nodes are extracted from the original geometric boundary of the building facade. These nodes include window edges, balcony edges, roof edges, and the junction of the building bottom and the ground. For each boundary node, a preset boundary value is set according to the nature of its location: For open boundaries in direct contact with the outside air (such as window openings), a zero gradient condition is used, which requires that the normal derivative of the temperature at this boundary be zero; for solid building walls, an adiabatic condition is used, which requires that the normal derivative of the temperature at this boundary be zero; for areas with a known fixed temperature (such as walls in direct contact with flames), a Dirichlet condition is used to fix the temperature to the temperature estimate of this node in the belief field; then, the sum of squared errors between the output of the physical information neural network on these boundary nodes and the preset boundary value is calculated, and then divided by the total number of boundary nodes to obtain the value of the boundary condition term.

[0080] The initial condition term forces the network's output at the current moment to match the temperature estimate in the belief field, and also uses the reciprocal of the uncertainty metric as the weight. The specific calculation process of the initial condition term is exactly the same as that of the data fitting term, the difference being the scope: the initial condition term only applies to the current moment, while the data fitting term can include observations from historical moments (if any). In practical implementation, since the current belief field temperature estimate is the only supervised data source, the initial condition term and the data fitting term can be merged into a unified loss term; they are described separately here for theoretical clarity. The weight coefficient of the initial condition term is the same as that of the data fitting term, taking a value of one.

[0081] After training the physical information neural network, uncertainty quantification and time-varying risk probability field generation are also performed.

[0082] Uncertainty quantification employs the Monte Carlo dropout method, which randomly discards some neurons in the hidden layer of the network during inference. Multiple forward propagations are performed for each spatiotemporal point to obtain multiple temperature predictions. The fluctuation amplitude of these predictions is then calculated, and this fluctuation amplitude represents the prediction variance in the time-varying risk probability field. The specific implementation of the Monte Carlo dropout method is as follows: After the physical information neural network is trained, it is switched to inference mode. In inference mode, the activation state of the dropout layer is maintained, meaning that hidden layer neurons are randomly discarded with a certain probability. The dropout probability is set to 0.1, meaning that each hidden layer neuron has a 10% probability of being set to zero in each forward propagation. For each spatiotemporal point requiring prediction, fifty forward propagations are performed independently, with different neurons randomly discarded during each forward propagation, resulting in fifty temperature predictions. The average of these fifty predictions is then calculated, followed by the square of the difference between each prediction and the average. These squared differences are summed and divided by forty-nine to obtain the sample variance. The square root of this variance is then taken to obtain the fluctuation amplitude. This fluctuation amplitude reflects the model's uncertainty regarding the prediction results: a larger fluctuation amplitude indicates greater prediction differences under different discarding patterns, suggesting insufficient confidence in the model's prediction for that point; a smaller fluctuation amplitude indicates stable prediction results and sufficient model confidence. The fluctuation amplitude is the prediction variance in the time-varying risk probability field.

[0083] The time-varying risk probability field generation converts the average temperature prediction from multiple forward propagations into a combustion probability using a differentiable risk indicator function. This risk indicator function is the same piecewise linear function used in the belief field construction module. The fluctuation amplitude is normalized and used as the prediction dispersion, which is the normalized prediction variance. The specific process for fluctuation amplitude normalization is as follows: First, find the maximum and minimum fluctuation amplitude values ​​at all spatiotemporal points within the entire future preset time window. Then, for each spatiotemporal point, subtract the minimum value from the fluctuation amplitude, and divide by the difference between the maximum and minimum values ​​to obtain a normalized value between zero and one, which is the prediction dispersion. If the maximum value equals the minimum value, the prediction dispersion for all points is set to zero. The smaller the prediction dispersion, the more reliable the prediction; the larger the prediction dispersion, the less reliable the prediction. This normalization operation allows uncertainties in different scenarios to be compared on a uniform scale, facilitating a reasonable balance between risk and uncertainty in the path planning module.

[0084] Ultimately, each spatiotemporal point outputs a tuple containing the combustion probability and predicted scatter, covering all moments within a future preset time window. The length of the future preset time window is set to 300 seconds, or five minutes. The time step is set to 10 seconds, meaning a risk probability field snapshot is output every 10 seconds. The spatial resolution is consistent with the belief field grid, i.e., one meter by one meter, with 0.5 meters by 0.5 meters in the corner region. Therefore, the entire time-varying risk probability field contains thirty time slices, each containing thousands to tens of thousands of grid nodes, with each node storing a combustion probability value and a predicted scatter value. This amount of data is acceptable on an airborne embedded platform because the physical information neural network only needs to perform a one-time forward propagation to generate predictions for all spatiotemporal points, without the need for point-by-point calculations.

[0085] In this embodiment, it is specifically necessary to explain that when the path planning module merges the time-varying risk probability field with the drone's current remaining battery power and current communication link quality into a multi-dimensional time-varying cost map, the specific operation is as follows:

[0086] For each spatiotemporal point, the instantaneous cost is defined as the weighted sum of three terms: combustion probability, predicted dispersion, and communication quality penalty function. The default values ​​for the three weight coefficients are as follows: the weight of combustion probability is 0.6, the weight of predicted dispersion is 0.2, and the weight of communication quality penalty is 0.2, with the sum always equal to 1. The dynamic adjustment method for the weight coefficients is as follows: when the task priority is set to "high survivability" mode, the weight of combustion probability is increased to 0.8, the weight of predicted dispersion is decreased to 0.1, and the weight of communication quality penalty is decreased to 0.1; when the task priority is set to "high communication" mode, the weight of communication quality penalty is increased to 0.5, the weight of combustion probability is decreased to 0.4, and the weight of predicted dispersion is decreased to 0.1. The task priority can be set remotely by the ground station or automatically switched by the UAV based on the current remaining power: when the remaining power is less than 20% of the total capacity, it automatically switches to "high survivability" mode to reduce energy consumption risk.

[0087] Combustion probability is derived from a time-varying risk probability field, predicted dispersion is the normalized prediction variance, and communication quality penalty function is an exponential function with the natural constant e as the base. The preset communication quality sensitivity coefficient with a negative exponential variable is multiplied by the current communication link quality, and the lower the communication quality, the greater the penalty. The preset value of the communication quality sensitivity coefficient is 0.1, and the unit is the reciprocal of decibel-milliwatt. The communication link quality is represented by the received signal strength indication value, and the unit is decibel-milliwatt, with the value range typically between -120 decibel-milliwatt and -30 decibel-milliwatt.

[0088] The specific calculation process of the communication quality penalty function is as follows: First, obtain the received signal strength indication value at the current location of the UAV. Then, calculate the negative communication quality sensitivity coefficient multiplied by the received signal strength indication value. Finally, use this result as the exponent to calculate the exponential function value with the natural constant e as the base. For example, when the received signal strength indication value is -90 decibels / milliwatts, the exponential variable is -0.1 multiplied by -90 equals 9, and the exponential function value is e to the power of 9, approximately 8,103, at which point the penalty is extremely large. When the received signal strength indication value is -30 decibels / milliwatts, the exponential variable is -0.1 multiplied by -30 equals 3, and the exponential function value is e to the power of 3, approximately 20.09, at which point the penalty is moderate. This exponential function ensures that the penalty increases sharply when the communication quality is extremely poor, forcing the path planning module to prioritize areas with good communication.

[0089] The weight coefficients of the three items are fixed, with the default values ​​showing the highest weight for combustion probability, followed by predicted dispersion, and the lowest weight for communication quality penalty. These weight coefficients can be dynamically adjusted based on task priority. The specific implementation of this dynamic adjustment is as follows: a task priority enumeration variable is maintained in the path planning module, with values ​​including "balanced mode," "high survival mode," and "high communication mode." At the start of each planning phase, the corresponding weight coefficient table is read based on the current task priority. If the task priority changes during flight, the new weight coefficients are automatically adopted in the next rolling time-domain planning cycle without interrupting the current trajectory. This adjustment mechanism allows the system to flexibly adapt to the needs of different rescue phases: high communication mode can be used in the initial reconnaissance phase to ensure data transmission, while automatically switching to high survival mode when approaching the fire source to mitigate risks.

[0090] When using a fast progress tree with a time window for prospective path search, the specific operation is as follows:

[0091] The search space is expanded into a state-time space, where each node consists of a spatial location and an arrival time. The state-time space is discretized as follows: the spatial location is taken from the belief field grid nodes or generated through random sampling, with the number of samples set to 500, and the sampling points are evenly distributed on the outer surface of the building; the time dimension starts from the current moment and extends to the end of a future preset time window at 10-second intervals, for a total of 30 time layers; each node is uniquely identified by a spatial location index and a time layer index, for example, the node number is "grid node number_time layer number"; this discretization method ensures that the search space size is controllable, and the time resolution is consistent with the output step size of the risk field prediction module, which facilitates direct querying of the time-varying risk probability field;

[0092] The cost of an edge is defined as the weighted sum of the integral of the instantaneous cost and the integral of the energy consumption during the movement from the starting node to the target node. The instantaneous cost comes from a multi-dimensional time-varying cost map, and the energy consumption integral is the integral of the energy consumption rate per unit time over time. The energy consumption rate per unit time is the cube of the velocity vector multiplied by a preset first aerodynamic parameter plus a preset second aerodynamic parameter. The specific numerical calculation method for the edge cost is as follows: Assume that the UAV moves at a constant speed along a straight line from the starting node to the target node, and the movement time is the time difference between the two nodes. First, calculate the velocity magnitude, which is equal to the spatial distance divided by the time difference. Then, calculate the energy consumption rate per unit time, which is equal to the cube of the velocity magnitude multiplied by the first aerodynamic parameter plus the second aerodynamic parameter. The preset value of the first aerodynamic parameter is 0.001, with units of joules per second squared and cubic meters per meter. The preset value of the second aerodynamic parameter is 0.5, with units of joules per second. The energy consumption integral is equal to the energy consumption rate per unit time multiplied by the time difference.

[0093] The instantaneous cost integral uses the trapezoidal numerical integration method: the motion time is divided into ten equal sub-intervals, and the instantaneous time value in the multidimensional time-varying cost map is queried at the endpoint of each sub-interval. Then, the integral approximate value is obtained by summing the values ​​according to the trapezoidal area formula. Finally, the energy consumption integral is multiplied by a preset power cost weight (the default value is 0.1) and added to the instantaneous cost integral to obtain the total cost of the edge. The power cost weight can be adjusted according to task requirements. When the remaining power is low, the weight can be increased to prioritize energy saving.

[0094] The feasibility determination of an edge requires the simultaneous fulfillment of three conditions: the probability of combustion at any time on the path does not exceed the preset hard danger threshold; the cumulative energy consumption does not exceed the available power after deducting the preset safety reserve power from the drone's current remaining power; and the communication link quality throughout the path is not lower than the preset minimum communication threshold. The values ​​of each preset parameter are set as follows: the hard danger threshold is 0.95, meaning that when the probability of combustion reaches 95%, it is considered absolutely impassable; the safety reserve power is 20% of the total capacity of the drone's battery. For example, if the total battery capacity is 5,000 mAh, then the safety reserve power is 1,000 mAh, and the available power is 4,000 mAh; the minimum communication threshold is -100 dBmW, meaning that when the received signal strength indicator value is lower than -100 dBmW, communication is considered interrupted, and the path is impassable.

[0095] Feasibility determination is performed simultaneously during edge cost calculation: For the first condition, when calculating the instantaneous cost integral, if the combustion probability at any sub-interval endpoint exceeds 0.95, the edge is directly marked as infeasible and skipped; for the second condition, after calculating the energy consumption integral, if the cumulative energy consumption plus the consumed energy consumption exceeds the available power, it is infeasible; for the third condition, when sampling five points evenly along the entire path, if the communication link quality at any sampling point is lower than -100 dB / mW, it is infeasible; only when all three conditions are met simultaneously is the edge included in the candidate edge set of the fast traversal tree.

[0096] In this embodiment, it should be specifically noted that when the perception update module determines hotspot areas with high uncertainty based on the prediction variance, the specific operation is as follows:

[0097] The prediction variance of each spatial location within a future preset time window is extracted from the time-varying risk probability field. The average prediction variance of each spatial location in the time dimension is calculated. The specific calculation process of the average prediction variance is as follows: First, all grid nodes in the time-varying risk probability field are traversed. For each grid node, the prediction variance value of all time layers (a total of thirty time layers, each layer is ten seconds apart) within the future preset time window is extracted. The sum of these thirty prediction variance values ​​is then divided by thirty to obtain the average prediction variance of the grid node. This average prediction variance reflects the average level of prediction uncertainty of the location within the entire future time window. The larger the value, the less reliable the fire prediction at that location is.

[0098] Spatial locations with average prediction variance exceeding a preset hotspot threshold are marked as candidate hotspots. The preset hotspot threshold is set to 0.6. This threshold is calibrated experimentally: in typical fire simulations, when the average prediction variance exceeds 0.6, the deviation between the prediction results of the physical information neural network and the actual fire spread exceeds 15%. Therefore, it is believed that the area needs to be actively detected to reduce uncertainty.

[0099] Candidate hotspots are clustered using connected components to remove isolated points, resulting in several consecutive hotspot regions. Each hotspot region contains a set of adjacent grid nodes. The specific algorithm for connected component clustering is as follows: a two-pass connected component labeling algorithm based on eight neighborhoods is used. First, all candidate hotspot grid nodes are labeled as foreground pixels, and the remaining nodes are labeled as background. In the first scan, all grid nodes are traversed from left to right and from top to bottom. For each foreground node, its four neighboring nodes (top left, top right, top right, and left) are checked. If there are already labeled foreground nodes, the current node is labeled with the smallest label among them, and the equivalence relation is recorded. If there are no neighboring foreground nodes, a new label is assigned. In the second scan, labels are merged according to the equivalence relation, and all nodes belonging to the same connected component are assigned the same final label. Finally, the number of grid nodes contained in each connected component is counted. If the number of nodes is less than three, it is considered an isolated point and removed. The remaining connected components are the hotspot regions. Each hotspot region is defined by the set of spatial coordinates of all grid nodes it contains, and its geometric center is calculated as the target point for subsequent detection.

[0100] When deciding whether to insert a probe waypoint based on the information gain value of the hotspot area and the power budget required for the probe, the specific operation is as follows:

[0101] For each hotspot region, the uncertainty of the current region is calculated. The uncertainty of the current region is equal to the sum of the average prediction variances of all grid nodes in the region. The calculation process of the uncertainty of the current region is as follows: First, obtain the average prediction variance values ​​of all grid nodes in the hotspot region, and add these values ​​one by one to get the sum. This sum is the uncertainty of the current region. The unit is the same as the prediction variance and it is dimensionless. The larger this value is, the higher the overall prediction uncertainty of the region is, and the more worthy it is to be detected first.

[0102] Assuming that the prediction variance of all grid nodes in the area drops to a preset baseline level after detection, the expected residual uncertainty is calculated. The expected residual uncertainty is equal to the number of grid nodes in the area multiplied by the preset baseline level. The preset baseline level is set to 0.1. This value represents the lower limit of the residual uncertainty caused by sensor noise after one near-field infrared detection. The expected residual uncertainty is equal to the number of grid nodes in the hotspot area multiplied by 0.1, indicating that even after detection, there will still be some uncertainty, but it is much lower than the level before detection.

[0103] Information gain is defined as the current uncertainty minus the expected residual uncertainty; the information gain is the total amount of uncertainty that can be eliminated by probing the region; the larger the value, the higher the information value of probing the region.

[0104] Estimate the total energy consumption required to fly from the current location to the center of the hotspot area and return via the original route. Total energy consumption equals the round-trip distance multiplied by the energy consumption rate per unit distance. The energy consumption rate per unit distance is obtained by dividing the energy consumption rate per unit time by the preset cruising speed. The preset cruising speed is 5 meters per second, or 18 kilometers per hour, which is the typical cruising speed for a fire-fighting drone. The energy consumption rate per unit time is calculated using the formula in the path planning module: the energy consumption rate per unit time equals the cube of the speed multiplied by the first aerodynamic parameter plus the second aerodynamic parameter. When the speed is 5 meters per second, the cube of the speed is 125, multiplied by the first aerodynamic parameter 0.001, resulting in 0.125, and added to the second aerodynamic parameter 0.5, resulting in 0.625 joules per second. The energy consumption rate per unit distance equals the energy consumption rate per unit time divided by the cruising speed, i.e., 0.625 divided by 5, equals 0.125 joules per meter. Total energy consumption equals the round-trip distance multiplied by 0.125 joules per meter. The round-trip distance is twice the Euclidean distance between the center of the hotspot area and the current location.

[0105] The detection utility function is defined as the information gain value divided by the total energy consumption plus a preset minimum positive value. The preset minimum positive value is 10 to the power of negative 8, which is consistent with the value in the belief field construction module above, and is used to prevent division by zero when the total energy consumption is zero. The detection utility function measures the information gain value obtained for each joule of electrical energy consumed, and the unit is unit uncertainty per joule.

[0106] When the detection utility function is greater than the preset utility threshold and the total energy consumption does not exceed the available power after deducting the preset safety reserve power from the current remaining power, a decision is made to insert a detection yaw waypoint. The preset utility threshold is 0.5, meaning that at least 0.5 units of uncertainty can be eliminated per joule before detection is worthwhile. The preset safety reserve power is 20% of the total battery capacity, consistent with the value in the path planning module. The available power is equal to the current remaining power minus the safety reserve power. If both of the above conditions are met, the center of the hotspot area is inserted as a detection yaw waypoint into the current trajectory sequence, and the insertion position is after the waypoint in the current trajectory that is closest to the center and the time closest.

[0107] After the probe is performed, the new temperature observation data obtained from the probe is fused into the belief field using Bayesian updates, and the replanning of the risk field prediction module and the path planning module is triggered. The specific operation is as follows:

[0108] After the drone reaches the center of the hotspot area, it uses an onboard infrared thermal imager to acquire new temperature observation data for the grid nodes in that area. Each new temperature observation data includes a sensor noise variance, which is the square of the sensor noise parameter. The sensor noise parameter is factory-calibrated to two degrees Celsius; therefore, the sensor noise variance is the square of two degrees Celsius, or four square degrees Celsius. After the drone reaches the center of the hotspot area, it hovers and adjusts the gimbal angle so that the infrared thermal imager's field of view covers the entire hotspot area. For each grid node in the area, the infrared thermal imager reads its surface temperature value and records the spatial coordinates of the node. Since the resolution of the infrared thermal imager is higher than the grid spacing, each grid node corresponds to multiple pixels. The average value is taken as the temperature observation value of that node, and the variance of the pixel values ​​is calculated as an estimate of the observation noise. If a node is obstructed or exceeds the field of view, that node is skipped.

[0109] After returning to the original route, for each observed grid node, the Bayesian update formula is used to fuse the original uncertainty measure and temperature estimate with the newly observed sensor noise variance and temperature value, resulting in a fused uncertainty measure and temperature estimate. The specific calculation process of the Bayesian update is as follows: assuming that the original uncertainty measure follows a Gaussian distribution and the new observation also follows a Gaussian distribution, the posterior distribution after fusion is still a Gaussian distribution; the fused uncertainty measure is equal to the reciprocal of the sum of the original uncertainty measure and the reciprocal of the sensor noise variance; the fused temperature estimate is equal to the fused uncertainty measure multiplied by the original temperature estimate divided by the original uncertainty measure plus the weighted sum of the new temperature value divided by the sensor noise variance; this formula is essentially the normalized mean and variance of the product of two Gaussian distributions.

[0110] For example, if the original uncertainty is 0.01, the original temperature estimate is 80 degrees Celsius, the newly observed temperature is 85 degrees Celsius, and the sensor noise variance is 4 squared degrees Celsius, then the fused uncertainty is equal to 1 divided by 0.01 plus 1 divided by 4, which equals 1 divided by 100 plus 0.25, which equals 1 divided by 100.25, approximately equal to 0.00998, which is almost unchanged; the fused temperature estimate is equal to 0.00998 multiplied by 80 divided by 0.01 plus 85 divided by 4, which equals 0.00998 multiplied by 8000 plus 21.25, which equals 0.00998 multiplied by 80021.25, approximately equal to 80.02 degrees Celsius; it can be seen that the new observation has little impact on the temperature estimate because the original observation is already very accurate; if the original uncertainty is large, the weight of the new observation will be significantly increased.

[0111] The fused uncertainty measure is equal to the reciprocal of the sum of the original uncertainty measure and the reciprocal of the sensor noise variance; mathematically, this means that the posterior precision equals the prior precision plus the observation precision; the prior precision is the reciprocal of the original uncertainty measure, and the observation precision is the reciprocal of the sensor noise variance; the higher the posterior precision, the smaller the posterior uncertainty measure.

[0112] The fused temperature estimate is equal to the fused uncertainty multiplied by the original temperature estimate divided by the original uncertainty plus the new temperature value divided by the weighted sum of the sensor noise variance; mathematically, the posterior mean is equal to the weighted average of the prior mean and the observed mean, with the weights being the prior accuracy and the observed accuracy, respectively; the fused uncertainty serves as a normalization factor to ensure that the sum of the weights is one.

[0113] The updated temperature estimate and uncertainty measure are written back to the corresponding grid nodes in the belief field, and the new observation point is added to the set of observed points. The specific process of the write-back operation is as follows: traverse all observed grid nodes, replace the temperature estimate field of the corresponding node in the belief field with the fused temperature estimate, and replace the uncertainty measure field with the fused uncertainty measure; at the same time, add the spatial coordinates of these grid nodes to the set of observed points, and update the boundary constraints of the set of observed points: if the new observation point is located outside the original boundary, then extend the boundary to include the new observation point.

[0114] Using the updated belief field as the new input, the risk field prediction module is immediately triggered to re-perform the physical information neural network deduction, and the path planning module is triggered to re-plan the trajectory. The specific process of triggering replanning is as follows: the perception update module sends a replanning request signal to the risk field prediction module, carrying the updated belief field data pointer; after receiving the request, the risk field prediction module immediately interrupts the current deduction task (if any), and uses the new belief field as the initial and boundary conditions to retrain or fine-tune the physical information neural network to generate a new time-varying risk probability field; at the same time, after receiving the replanning request, the path planning module uses the new time-varying risk probability field, the current remaining power, and the current communication link quality as input to re-execute the multi-dimensional time-varying cost map construction and fast progression tree search to generate a new executable trajectory; the entire replanning process should be completed within five seconds to ensure that the UAV can respond to changes in the fire situation in a timely manner.

[0115] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0116] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0117] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0121] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A 3D modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs), characterized in that, Specifically, it includes: The belief field construction module receives multi-source sparse observation data, including temperature distribution data from an infrared thermal imager and wind field data from an anemometer, collected by the UAV's onboard sensors. It performs spatiotemporal alignment and uncertainty measurement to generate the belief field at the current moment. The belief field includes the current risk field on the grid, wind field distribution data, and boundary constraints of the observed point set. Risk field prediction module: Using the belief field as the initial and boundary conditions, the physical information neural network is used to approximate the partial differential equation of fire propagation and output the time-varying risk probability field within the future preset time window. The time-varying risk probability field includes the combustion probability of each location at each time and its prediction variance. Path planning module: The time-varying risk probability field is merged with the drone's current remaining battery power and current communication link quality into a multi-dimensional time-varying cost map. A fast traversing tree with time windows is used to perform a forward-looking path search on the multi-dimensional time-varying cost map to generate an executable trajectory in the current rolling time domain. Perception Update Module: Based on the prediction variance, it identifies hotspot areas with high uncertainty. It decides whether to insert a waypoint for the probe based on the information gain value of the hotspot area and the power budget required to perform the probe. After the probe is performed, it uses Bayesian update to fuse the new temperature observation data obtained by the probe into the belief field, triggering the replanning of the risk field prediction module and the path planning module.

2. The urban high-rise building fire drone 3D modeling and path planning system according to claim 1, characterized in that: The belief field construction module performs the following operations when performing spatiotemporal alignment and uncertainty measurement: Temperature distribution data and wind field data from multi-source sparse observation data are mapped onto regular grid nodes on the building's exterior surface, respectively. For each grid node, a spatial radius and time window are defined with it as the center, and temperature observation index set and wind field observation index set falling within the spatiotemporal neighborhood are selected; A weighted kernel regression method was used, with spatial distance, time difference, and the reciprocal of sensor noise parameters as weights, to estimate the temperature value and wind speed vector of each grid node. Based on this, the information density of each grid node was calculated. The information density was obtained by logarithmically weighting the spatial weight, time weight, and sensor noise parameter contribution of each observation in the neighborhood. The inverse of the information density is then defined as the uncertainty measure of the grid node. The smaller the uncertainty measure, the more reliable the estimate. Finally, the temperature estimate, wind field estimate, uncertainty measure, and boundary constraints of the observed point set for each grid node are packaged together to generate the belief field.

3. The urban high-rise building fire drone 3D modeling and path planning system according to claim 2, characterized in that: When generating the belief field, an initial risk value mapping is also performed: The temperature estimate of each grid node is converted into an initial risk value through a differentiable risk indicator function, which is set to a piecewise linear form: when the temperature estimate is less than or equal to the safe temperature threshold, the initial risk value is zero. When the temperature estimate is between the safe temperature threshold and the critical damage temperature threshold, the initial risk value is equal to the difference between the temperature estimate and the safe temperature threshold divided by the difference between the critical damage temperature threshold and the safe temperature threshold. When the temperature estimate is greater than or equal to the critical damage temperature threshold, the initial risk value is one; the obtained initial risk value and the wind field estimate together constitute the current risk field in the belief field.

4. The urban high-rise building fire drone 3D modeling and path planning system according to claim 3, characterized in that: When the risk field prediction module uses a physical information neural network to approximate the solution of the partial differential equation for fire propagation, the specific operation is as follows: Construct a composite loss function, which is obtained by weighted summation of data fitting term, physical residual term, boundary condition term and initial condition term; The data fitting term uses the temperature estimate of each grid node in the belief field as a supervision signal and the reciprocal of the uncertainty measure of each grid node as a weight, so that high-confidence observations have a greater influence. The physical residual term is randomly sampled from a set of spatial-temporal collocation points within the domain, and the sum of squared residuals of the physical information neural network output with respect to the partial differential equation of fire propagation is calculated. The boundary condition terms constrain the network output to be consistent with the preset boundary values ​​on the original geometric boundaries of the building facade; The initial condition term forces the network's output at the current moment to be consistent with the temperature estimate in the belief field, and also uses the reciprocal of the uncertainty metric as the weight.

5. A three-dimensional modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs) according to claim 4, characterized in that: After training the physical information neural network, uncertainty quantification and time-varying risk probability field generation are also performed. Uncertainty quantification employs the Monte Carlo dropout method, which randomly discards some neurons in the hidden layer of the network during inference, performs multiple forward propagations for each spatiotemporal point, obtains multiple temperature prediction values, and then calculates the fluctuation amplitude of these prediction values. The fluctuation amplitude is the prediction variance in the time-varying risk probability field. The time-varying risk probability field generation converts the average temperature prediction from multiple forward propagations into the combustion probability through a differentiable risk indicator function. The risk indicator function is the same as the piecewise linear function used in the belief field construction module, and the fluctuation amplitude is normalized and used as the prediction dispersion. The prediction dispersion is the normalized prediction variance. Finally, each spatiotemporal point outputs a tuple containing the combustion probability and the predicted dispersion, covering all moments within the preset future time window.

6. A three-dimensional modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs) according to claim 5, characterized in that: When the path planning module merges the time-varying risk probability field with the drone's current remaining battery power and current communication link quality into a multi-dimensional time-varying cost map, the specific operation is as follows: For each spatiotemporal point, the instantaneous cost is defined as the weighted sum of three terms: the combustion probability, the predicted dispersion, and the communication quality penalty function. Combustion probability comes from time-varying risk probability field, prediction dispersion is normalized prediction variance, communication quality penalty function is exponential function with natural constant e as base, the preset communication quality sensitivity coefficient with negative exponential independent variable is multiplied by the current communication link quality, the lower the communication quality, the greater the penalty. The weight coefficients of the three items are fixed. By default, the weight of combustion probability is the largest, followed by the weight of predicted dispersion, and the weight of communication quality penalty is the smallest. The weight coefficients can be dynamically adjusted according to task priority.

7. A three-dimensional modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs) according to claim 6, characterized in that: When using a fast progress tree with a time window for prospective path search, the specific operation is as follows: The search space is expanded into a state-time space, where each node consists of a spatial location and an arrival time. The cost of an edge is defined as the weighted sum of the integral of the instantaneous cost and the integral of the energy consumption during the movement from the starting node to the target node. The instantaneous cost comes from a multidimensional time-varying cost map, and the energy consumption integral is the integral of the energy consumption rate per unit time along time. The energy consumption rate per unit time is the cube of the velocity vector multiplied by a preset first aerodynamic parameter plus a preset second aerodynamic parameter. The feasibility of an edge must meet three conditions simultaneously: the probability of combustion at any time on the path does not exceed the preset hard danger threshold, the cumulative energy consumption does not exceed the available power after deducting the preset safety reserve power from the drone's current remaining power, and the quality of the communication link throughout the path is not lower than the preset minimum communication threshold.

8. A three-dimensional modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs) according to claim 7, characterized in that: When the perception update module identifies hotspot areas with high uncertainty based on prediction variance, the specific operation is as follows: Extract the prediction variance of each spatial location at all times within a future preset time window from the time-varying risk probability field, and calculate the average prediction variance of each spatial location in the time dimension; Spatial locations with average prediction variance exceeding a preset hotspot threshold are marked as candidate hotspots; Connectivity clustering is performed on candidate hotspots to remove isolated points, resulting in several consecutive hotspot regions, each containing a set of adjacent grid nodes.

9. A three-dimensional modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs) according to claim 8, characterized in that: When deciding whether to insert a probe waypoint based on the information gain value of the hotspot area and the power budget required for the probe, the specific operation is as follows: For each hotspot region, calculate the uncertainty of the current region, which is equal to the sum of the average prediction variances of all grid nodes in that region; Assuming that the prediction variance of all grid nodes in the region drops to a preset baseline level after detection, the expected residual uncertainty is calculated. The expected residual uncertainty is equal to the number of grid nodes in the region multiplied by the preset baseline level. Information gain is defined as the current uncertainty minus the expected residual uncertainty. Estimate the total energy consumption required to fly from the current location to the center of the hotspot area and back to the original route. The total energy consumption is equal to the round-trip distance multiplied by the energy consumption rate per unit distance. The energy consumption rate per unit distance is obtained by dividing the energy consumption rate per unit time by the preset cruise speed. The detection utility function is defined as the information gain value divided by the total energy consumption plus a preset minimum positive value; When the detection utility function is greater than the preset utility threshold and the total energy consumption does not exceed the available power after deducting the preset safety reserve power from the current remaining power, a decision is made to insert a detection waypoint.

10. A three-dimensional modeling and path planning system for urban high-rise building fires using unmanned aerial vehicles (UAVs) according to claim 9, characterized in that: After the probe is performed, the new temperature observation data obtained from the probe is fused into the belief field using Bayesian updates, and the replanning of the risk field prediction module and the path planning module is triggered. The specific operation is as follows: After the drone flies to the center of the hotspot area, it uses an airborne infrared thermal imager to acquire new temperature observation data of the grid nodes in the area. Each new temperature observation data is accompanied by sensor noise variance, which is the square of the sensor noise parameter. After returning to the original route, for each observed grid node, the original uncertainty measure and temperature estimate are fused with the newly observed sensor noise variance and temperature value using the Bayesian update formula to obtain the fused uncertainty measure and temperature estimate. The fused uncertainty measure is equal to the reciprocal of the sum of the original uncertainty measure and the reciprocal of the sensor noise variance; The fused temperature estimate equals the fused uncertainty multiplied by the original temperature estimate divided by the original uncertainty, plus the weighted sum of the new temperature value divided by the sensor noise variance; Write the updated temperature estimate and uncertainty measure back to the corresponding grid node in the belief field, and add the new observation point to the set of observed points; Using the updated belief field as the new input, the risk field prediction module is immediately triggered to re-perform the physical information neural network deduction, and the path planning module is triggered to re-plan the trajectory.

Images