Fire extinguishing and rescue dispatching method based on agent cooperation

Abstract

The application discloses a fire extinguishing and rescue scheduling method based on agent cooperation, relates to the technical field of multi-agent system and distributed artificial intelligence, and comprises the following steps: S1, outputting a fire field environment space-time state tensor; S2, introducing a differentiable aggregation phase change mechanism by improving a PredRNN model, and performing space-time evolution modulation and differentiable aggregation phase change mapping under the constraint of a fire spread driving force; S3, calculating a fire prediction arrival time; S4, performing target distribution and route calculation on a UAV agent; S5, generating a candidate truncated attack task set; S6, generating a fire extinguishing and rescue scheduling strategy; S7, outputting a time sequence linkage execution instruction set; and S8, performing iterative updating of the fire extinguishing and rescue scheduling strategy. The application overcomes the limitations of traditional methods, such as fire prediction distortion, weak space-time cooperation correlation and neglecting dynamic conflict constraints, and provides an efficient solution for agent cooperative fire extinguishing and rescue scheduling.

Description

Technical Field

[0001] This invention relates to the fields of multi-agent systems and distributed artificial intelligence technology, and in particular to a fire fighting and rescue dispatching method based on agent collaboration. Background Technology

[0002] With the explosive growth of multi-source fire scene observation data and the increasing complexity of multi-agent collaborative rescue needs, classical scheduling systems face severe computational challenges in the spatiotemporal evolution extrapolation and multi-agent collaborative control of dynamic fire scene environments. Existing fire spread prediction and scheduling algorithms, such as standard recurrent neural networks or heuristic task allocation models, while improving the efficiency of fire estimation and flight path calculation by utilizing historical observation data and basic kinematic constraints, mainly rely on linear extrapolation of single environmental parameters and local path planning of independent nodes for spatiotemporal state updates. Methods based solely on linear extrapolation and local path planning ignore the complex spatiotemporal evolution modulation mechanisms, differentiable aggregation phase transition physical characteristics, and deep spatiotemporal collaborative relationships between UAVs and fire trucks implicit in multi-source fire scene data. This leads to high-dimensional state space feature distortion when extrapolating the fire line expansion boundary in future periods and generating candidate cutoff strike task sets, thus limiting the accuracy of fire spread pre-positioning selection and multi-agent spatiotemporal collaborative matching. Furthermore, classical scheduling methods often struggle to fully utilize the kinematic constraint distribution of multi-agent systems and the spatiotemporal grid occupancy map to constrain the trajectory reconstruction space. This results in a large number of spatially overlapping conflicts when dealing with multi-agent conflict detection and collision avoidance reconstruction, increasing computational overhead and reducing the iterative update efficiency of fire-fighting and rescue scheduling strategies.

[0003] Therefore, how to provide a fire fighting and rescue dispatch method based on intelligent agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a fire-fighting and rescue scheduling method based on agent collaboration. Through a closed-loop iterative update process based on deviation evaluation results, it performs dynamic evaluation of spatiotemporal deviations and multi-agent conflict detection on the fire line expansion boundary output by the improved PredRNN model and the actual operating states of multiple agents. It extracts the spatiotemporal deviations including the actual and predicted operating states, and calculates an adaptive deviation tolerance threshold based on statistical distribution characteristics to output the deviation evaluation results. The deviation evaluation results are embedded as feedback compensation constraints in the strategy feedback compensation calculation. By calculating the update amount of the scheduling strategy parameters with respect to the spatiotemporal deviation gradient, the fire-fighting and rescue scheduling strategy is corrected using an iterative mechanism until the adaptive deviation tolerance threshold convergence condition is met. This process, by establishing a closed-loop feedback path from "deviation feature measurement" to "scheduling strategy parameters," effectively eliminates spatiotemporal trajectory overlap conflicts during multi-agent execution, ensuring that the generated fire-fighting and rescue scheduling strategy can dynamically maintain spatiotemporal collaborative correlation under the evolution of the fire environment's spatiotemporal state. This achieves the technical effect of improving the accuracy of multi-agent collaborative interception while maintaining efficient collision avoidance and reconstruction performance. The improved PredRNN model introduces a differentiable aggregation phase transition mechanism, decouples and performs differential calculations on the evolutionary motion feature tensor to generate local net response features and maps them to a state probability distribution matrix. Based on the fire spread driving force, it performs anisotropic direction-weighted diffusion evolution, which overcomes the limitations of traditional methods such as fire prediction distortion, weak spatiotemporal coordination and neglect of dynamic conflict constraints, and provides an efficient solution for intelligent agent collaborative fire fighting and rescue scheduling.

[0005] A firefighting and rescue dispatching method based on intelligent agent collaboration according to an embodiment of the present invention specifically includes: S1. Acquire multi-source observation data of the fire scene and perform spatiotemporal benchmark alignment and gridded reconstruction to output the spatiotemporal state tensor of the fire scene environment. S2. Input the spatiotemporal state tensor of the fire environment into the improved PredRNN model, introduce a differentiable aggregation phase transition mechanism, perform spatiotemporal evolution modulation and differentiable aggregation phase transition mapping under the constraint of fire spread driving force, and deduce the fire line expansion boundary for future periods. S3. Select and generate a set of fire spread precursor points from the fire line expansion boundary, and calculate the predicted arrival time of the fire corresponding to the set of fire spread precursor points. S4. Based on the set of fire spread precursor points, target allocation and flight path calculation are performed on the UAV agent, and the predicted guidance spatiotemporal trajectory set of the UAV and the corresponding UAV arrival time are output. S5. Obtain fire truck accessibility constraints, combine fire prediction arrival time with UAV arrival time to calculate cutoff strike points associated with the fire spread precursor point set, deduce fire truck arrival time and generate candidate cutoff strike task set. S6. Perform spatiotemporal collaborative matching and conflict resolution on the candidate interception strike mission set and the UAV predicted guidance spatiotemporal trajectory set to generate fire fighting and rescue dispatch strategy. S7. Perform spatiotemporal grid mapping and multi-agent conflict detection on the fire fighting and rescue dispatch strategy, and perform trajectory collision avoidance reconstruction based on the detection results, and output a set of time-series linkage execution instructions; S8. Execute the time-series linkage execution instruction set, obtain the actual operating status of the multi-agent, calculate the spatiotemporal deviation between the actual operating status of the multi-agent and the predicted operating status corresponding to the time-series linkage execution instruction set, and perform iterative updates of the fire-fighting and rescue dispatch strategy based on the deviation evaluation results of the spatiotemporal deviation.

[0006] Optionally, S1 specifically includes: S11. Obtain multi-source raw observation data of the fire scene, extract environmental parameters and observation spatiotemporal labels from the multi-source raw observation data of the fire scene; based on the spatial distribution characteristics and time span statistical distribution of the observation spatiotemporal labels, calculate the adaptive spatiotemporal alignment benchmark, project the environmental parameters onto the adaptive spatiotemporal alignment benchmark, and output the benchmark aligned multi-source parameter sequence. S12. Based on the spatial distribution density statistical characteristics of the benchmark-aligned multi-source parameter sequence, calculate the adaptive spatial grid resolution; construct a spatial grid index according to the adaptive spatial grid resolution, and use the spatial grid index to perform spatial interpolation calculation on the benchmark-aligned multi-source parameter sequence to output the gridded environmental parameter matrix. S13. Extract the timestamp sequence of the gridded environmental parameter matrix. Based on the timestamp sequence, splice the gridded environmental parameter matrices of different environmental types along the feature channel dimension and stack them along the time dimension to output the fire environment spatiotemporal state tensor.

[0007] Optionally, the improved PredRNN model includes a spatiotemporal feature encoding layer, a spatiotemporal memory layer, a spatial feature interaction modulation layer, and a differentiable aggregated phase transition layer: The spatiotemporal feature encoding layer is used to perform local convolution calculation on the fire environment spatiotemporal state tensor to extract spatial distribution features, and extract environmental parameters from the fire environment spatiotemporal state tensor to calculate the fire spread driving force; the spatial distribution features and the fire spread driving force are channel aligned and concatenated, and then dimensionality reduction calculation is performed through linear matrix mapping to generate an initial spatiotemporal hidden state tensor sequence. The spatiotemporal memory layer is used to perform temporal state evolution calculations in the time dimension on the initial spatiotemporal hidden state tensor sequence to generate global spatiotemporal semantic features; at the same time, it performs bidirectional state interaction calculations between layers, passing the global spatiotemporal semantic features downward to constrain the extraction of local spatiotemporal detail features, and feeding back the obtained local spatiotemporal detail features upward; by fusing the global spatiotemporal semantic features and the local spatiotemporal detail features, a memory tensor sequence that fuses global and local spatiotemporal dependencies is output. The spatial feature interaction modulation layer is used to map the memory tensor sequence that fuses global and local spatiotemporal dependencies into a multi-channel spatiotemporal state tensor; an external driving constraint feature vector is generated based on the fire spread driving force; a multi-mode spatial neighborhood feature aggregation operation is performed on the multi-channel spatiotemporal state tensor to obtain spatial interaction aggregation features; the spatial interaction aggregation features and the external driving constraint feature vector are fused and calculated to generate an evolutionary motion feature tensor; The differentiable aggregation phase transition layer is used to introduce a differentiable aggregation phase transition mechanism, decouple and perform differential calculation on the evolutionary motion feature tensor, generate local net response features and map them into a state probability distribution matrix; perform anisotropic direction-weighted diffusion evolution on the state probability distribution matrix based on the fire spread driving force; couple the evolved state probability distribution matrix with the local net response features, determine the dynamic phase transition critical value based on the statistical distribution of the coupled features and perform differentiable state mapping to generate target boundary prediction results, which are output as the fire line expansion boundary.

[0008] Optionally, the differentiable aggregation phase transition mechanism specifically includes: Perform channel-dimensional feature decoupling mapping on the evolutionary motion feature tensor to obtain positive response features and negative inhibition features; Spatial projection and pixel-by-pixel difference calculations are performed on the positive response features and the negative suppression features to generate local net response features; Differentiable normalization calculation is performed on the local net response features to smoothly map the continuous values ​​of the local net response features to the probability interval, thereby generating a state probability distribution matrix. The direction decoupling calculation is performed on the fire spread driving force to extract the driving direction vector and driving intensity scalar; a spatial neighborhood direction mask is constructed based on the driving direction vector, and an asymmetric weight allocation calculation is performed on the spatial neighborhood direction mask in combination with the driving intensity scalar to generate an anisotropic spatial transfer weight matrix. The spatial neighborhood convolution calculation of the state probability distribution matrix is ​​performed on the anisotropic spatial transition weight matrix to obtain the diffusion update probability matrix; The diffusion update probability matrix and the local net response features are fused to obtain the comprehensive response state features; The dynamic state critical value is calculated based on the statistical distribution of the comprehensive response state characteristics; a differentiable threshold activation calculation is performed on the comprehensive response state characteristics; when the comprehensive response state characteristics are greater than the dynamic state critical value, a nonlinear state mapping is performed to generate the target boundary prediction result, which is output as the fire line extension boundary.

[0009] Optionally, S3 specifically includes: S31. Extract the discrete point set and local curvature features of the fire line expansion boundary; calculate the adaptive curvature filtering threshold based on the statistical distribution of the local curvature features; perform a filtering operation on the discrete point set of the boundary using the adaptive curvature filtering threshold, and output the set of fire spread precursor points; S32. Extract the temporal resolution of the fire line expansion boundary and extract the spatial displacement vector of the fire spread precursor point set between adjacent time frames; based on the spatial displacement vector and the temporal resolution of the fire line expansion boundary, calculate the local spread rate of the fire spread precursor point set. S33. Obtain the fire ignition point and calculate the spatial distance from the current fire ignition point to each preceding point in the set of fire spread precursor points; based on the quotient of spatial distance and local spread rate, calculate and output the predicted arrival time of the fire.

[0010] Optionally, S4 specifically includes: S41. Obtain the initial state distribution of the UAV agent and the spatial topological features of the fire spread front point set; calculate the adaptive target allocation weight matrix based on the statistical distribution of the spatial topological features; use the adaptive target allocation weight matrix to perform global optimal matching between the UAV agent and the fire spread front point set, and output the front point-UAV mapping pair. S42. Based on the forward point-UAV mapping pair, extract the spatial connectivity graph from the UAV agent to the corresponding set of forward points of fire spread; perform shortest path search and trajectory smoothing calculation in the spatial connectivity graph, and output the set of predicted spatiotemporal trajectories for UAV guidance. S43. Extract the end node timestamp of the UAV prediction guidance spatiotemporal trajectory set; obtain the kinematic dynamic constraints of the UAV agent; perform time compensation calculation based on the end node timestamp and the kinematic dynamic constraints of the UAV agent; and output the arrival time of the UAV.

[0011] Optionally, S5 specifically includes: S51. Obtain basic performance parameters of fire trucks, fire scene terrain access data, and initial position distribution of fire trucks; S52. Calculate the adaptive traffic speed field based on the statistical distribution characteristics of fire scene terrain access data; combine the basic performance parameters of the fire truck, and use the adaptive traffic speed field and the initial position distribution of the fire truck to calculate the accessibility constraints of the fire truck. S53. Extract the time difference between the predicted arrival time of the fire and the arrival time of the drone; calculate the time truncation safety margin based on the time difference and the accessibility constraints of the fire truck; use the time truncation safety margin to perform spatial reverse offset calculation on the set of fire spread precursor points and output the truncation strike point. S54. Based on the constraints of the cutoff point and the accessibility of the fire truck, perform path time estimation and output the arrival time of the fire truck; perform spatiotemporal correlation coding on the cutoff point, the arrival time of the fire truck, the predicted arrival time of the fire, and the arrival time of the drone, and output a candidate cutoff strike task set.

[0012] Optionally, S6 specifically includes: S61. Extract the spatiotemporal overlap features between the UAV predicted guidance spatiotemporal trajectory set and the candidate truncation strike mission set; calculate the adaptive collaborative matching weights based on the statistical distribution of the spatiotemporal overlap features; use the adaptive collaborative matching weights to perform a global association mapping between the UAV predicted guidance spatiotemporal trajectory set and the candidate truncation strike mission set, and output the collaborative matching relationship set; S62. Extract spatiotemporal conflict features based on the collaborative matching relationship set; calculate the adaptive conflict resolution penalty factor based on the statistical distribution of the spatiotemporal conflict features; use the adaptive conflict resolution penalty factor to calculate the trajectory offset and task rearrangement of the collaborative matching relationship set, and output a conflict-free collaborative task sequence. S63. Based on the time-series cascading arrangement and execution resource binding of conflict-free collaborative task sequences, output fire-fighting and rescue scheduling strategies.

[0013] Optionally, S7 specifically includes: S71. Obtain the fire fighting and rescue dispatch strategy and the kinematic constraint distribution of the multi-agent system; calculate the adaptive spatiotemporal grid resolution based on the statistical characteristics of the kinematic constraint distribution of the multi-agent system; perform spatiotemporal grid mapping on the fire fighting and rescue dispatch strategy using the adaptive spatiotemporal grid resolution, and output the spatiotemporal grid occupancy map. S72. Extract multi-agent spatial overlap features based on spatiotemporal grid occupancy map; calculate adaptive conflict detection threshold based on statistical distribution of multi-agent spatial overlap features; perform multi-agent conflict detection on spatiotemporal grid occupancy map using adaptive conflict detection threshold, and output conflict detection results; S73. Extract the trajectory segment to be reconstructed based on the collision detection results; calculate the adaptive trajectory offset step size based on the statistical distribution characteristics of the trajectory segment to be reconstructed; perform trajectory anti-collision reconstruction on the trajectory segment to be reconstructed using the adaptive trajectory offset step size, and output the timing linkage execution instruction set.

[0014] Optionally, S8 specifically includes: S81. Obtain the actual operating state of the multi-agent system and the time-series linkage execution instruction set; extract the predicted operating state corresponding to the time-series linkage execution instruction set; perform time alignment and spatial registration calculation between the actual operating state and the predicted operating state of the multi-agent system, and output the spatiotemporal deviation. S82. Extract the statistical distribution characteristics of spatiotemporal deviation; calculate the adaptive deviation tolerance threshold based on the statistical distribution characteristics of spatiotemporal deviation; perform dynamic evaluation calculation on spatiotemporal deviation using the adaptive deviation tolerance threshold, and output the deviation evaluation results; S83. Obtain the fire fighting and rescue dispatch strategy; based on the deviation assessment results and the fire fighting and rescue dispatch strategy execution strategy feedback compensation calculation, output the iteratively updated fire fighting and rescue dispatch strategy.

[0015] The beneficial effects of this invention are: (1) By constructing an improved PredRNN model and a differentiable aggregated phase transition mapping process, a deep fusion of the spatiotemporal evolution characteristics of the fire environment and the physical mechanism of phase transition was achieved. The spatial feature interaction modulation layer maps the memory tensor sequence that integrates global and local spatiotemporal dependencies into a multi-channel spatiotemporal state tensor. Based on the fire spread driving force, it generates external driving constraint feature vectors and performs multi-mode spatial neighborhood feature aggregation to generate an evolutionary motion feature tensor containing external constraints. The differentiable aggregated phase transition layer performs decoupling and differential calculation on the evolutionary motion feature tensor, generates local net response features and maps them into a state probability distribution matrix. It combines the driving direction vector and the driving intensity scalar to construct an anisotropic spatial transfer weight matrix and performs direction-weighted diffusion evolution. The above process smoothly maps the continuous spatiotemporal state field to the probability interval and applies anisotropic physical constraints, accurately capturing the critical phase transition characteristics in the fire spread process, and outputting the fire line expansion boundary with clear physical boundaries and dynamic evolution trends.

[0016] (2) By constructing a spatiotemporal collaborative matching and spatiotemporal grid anti-collision reconstruction system, a two-level safety barrier of multi-agent task-level association and underlying trajectory execution was established. The spatiotemporal overlap features between the UAV prediction guidance spatiotemporal trajectory set and the candidate truncation strike task set were extracted, adaptive collaborative matching weights and adaptive conflict resolution penalty factors were calculated, and trajectory offset and task rearrangement were performed to output a conflict-free collaborative task sequence. Based on the kinematic constraint distribution of multi-agents, adaptive spatiotemporal grid resolution was calculated to perform spatiotemporal grid mapping on the fire fighting and rescue scheduling strategy. The trajectory segments to be reconstructed were extracted using the adaptive conflict detection threshold, and trajectory anti-collision reconstruction was performed by combining the adaptive trajectory offset step size calculated based on the statistical distribution characteristics of the trajectory segments. The above process resolves task-level temporal conflicts through adaptive weight allocation, and eliminates spatial overlap by combining dynamic grid mapping and step size compensation, outputting a temporal linkage execution instruction set that takes into account both spatiotemporal collaborative efficiency and physical anti-collision safety. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is an overall flowchart of a fire fighting and rescue dispatching method based on intelligent agent collaboration proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the improved PredRNN model for a fire fighting and rescue dispatching method based on agent collaboration proposed in this invention. Detailed Implementation

[0018] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 A fire fighting and rescue dispatching method based on intelligent agent collaboration, specifically including: S1. Acquire multi-source observation data of the fire scene and perform spatiotemporal benchmark alignment and gridded reconstruction, and output the spatiotemporal state tensor of the fire scene environment. S2. Input the spatiotemporal state tensor of the fire environment into the improved PredRNN model, introduce a differentiable aggregation phase transition mechanism, perform spatiotemporal evolution modulation and differentiable aggregation phase transition mapping under the constraint of fire spread driving force, and deduce the fire line expansion boundary for future periods. S3. Select and generate a set of fire spread precursor points from the fire line expansion boundary, and calculate the predicted arrival time of the fire corresponding to the set of fire spread precursor points. S4. Based on the set of fire spread precursor points, target allocation and flight path calculation are performed on the UAV agent, and the predicted guidance spatiotemporal trajectory set of the UAV and the corresponding UAV arrival time are output. S5. Obtain fire truck accessibility constraints, combine fire prediction arrival time with UAV arrival time to calculate cutoff strike points associated with the fire spread precursor point set, deduce fire truck arrival time and generate candidate cutoff strike task set; S6. Perform spatiotemporal collaborative matching and conflict resolution on the candidate interception strike mission set and the UAV predicted guidance spatiotemporal trajectory set to generate fire fighting and rescue dispatch strategy. S7. Perform spatiotemporal grid mapping and multi-agent conflict detection on the fire fighting and rescue dispatch strategy, and perform trajectory collision avoidance reconstruction based on the detection results, and output a set of time-series linkage execution instructions; S8. Execute the time-series linkage execution instruction set, obtain the actual operating status of the multi-agent, calculate the spatiotemporal deviation between the actual operating status of the multi-agent and the predicted operating status corresponding to the time-series linkage execution instruction set, and perform iterative updates of the fire-fighting and rescue dispatch strategy based on the deviation evaluation results of the spatiotemporal deviation.

[0020] In this embodiment, S1 specifically includes: S11. Read the original multi-source observation data of the fire site and extract the environmental parameters from the original multi-source observation data. The environmental parameters include temperature values ​​in degrees Celsius, wind speed values ​​in meters per second, and relative humidity values ​​in percentage. At the same time, extract the observation spatiotemporal labels corresponding to the environmental parameters. Calculate the sum of the maximum longitude span and the maximum latitude span of the observation spatiotemporal labels as the spatial distribution feature, and calculate the difference between the maximum and minimum observation times as the time span statistical distribution. Set the benchmark time interval to 5 minutes and the benchmark spatial step size to 100 meters. Divide the spatial distribution feature by the benchmark spatial step size and round up, and divide the time span statistical distribution by the benchmark time interval and round up. Combine these to construct an adaptive spatiotemporal alignment benchmark. Use a bilinear interpolation function to project the environmental parameters onto the adaptive spatiotemporal alignment benchmark, calculate the weighted average value of the values ​​of adjacent grid points, and output the benchmark aligned multi-source parameter sequence.

[0021] S12. The number of data points per unit area in the statistical benchmark-aligned multi-source parameter sequence is used as the spatial distribution density statistical feature. The density threshold is set to 10 per square kilometer. The spatial distribution density statistical feature is divided by the density threshold and rounded up to obtain the adaptive spatial grid resolution. The adaptive spatial grid resolution is set to 50 meters. A spatial grid index is constructed according to the adaptive spatial grid resolution. The inverse distance weighted interpolation function is used to perform spatial interpolation calculation on the benchmark-aligned multi-source parameter sequence. The square of the inverse distance between the target grid point and the surrounding known data points is calculated as the weight. The values ​​of each known data point are multiplied by the corresponding weight, summed, and divided by the total weight to output the gridded environmental parameter matrix.

[0022] S13. Extract the timestamp sequence of the gridded environmental parameter matrix. Based on the chronological order of the timestamp sequence, concatenate the gridded environmental parameter matrices of three different environmental types (temperature, wind speed, and relative humidity) along the feature channel dimension to generate a channel concatenation matrix. Stack the channel concatenation matrix along the time dimension according to the chronological order of the timestamp sequence to generate an initial spatiotemporal tensor. Perform nonlinear mapping calculation on the initial spatiotemporal tensor using the ReLU activation function. Determine the value of each element in the initial spatiotemporal tensor. If the value of an element is less than 0, set the value of that element to 0. If the value of an element is greater than or equal to 0, keep the value of that element unchanged. Output the fire environment spatiotemporal state tensor.

[0023] In this embodiment, the improved PredRNN model includes a spatiotemporal feature encoding layer, a spatiotemporal memory layer, a spatial feature interaction modulation layer, and a differentiable aggregated phase transition layer: The spatiotemporal feature encoding layer is used to read the spatiotemporal state tensor of the fire environment. A 3x3 convolution kernel with a stride of 1 grid is set to perform local convolution calculation on the fire environment spatiotemporal state tensor. After processing with the ReLU activation function, if the calculation result is less than 0, it is set to 0; if it is greater than or equal to 0, it remains unchanged to extract spatial distribution features. Environmental parameters are extracted from the fire environment spatiotemporal state tensor by multiplying the wind speed value by a weight coefficient of 0.5, adding the temperature value by a weight coefficient of 0.3, and adding the relative humidity value by a weight coefficient of -0.2 to calculate the driving force of fire spread. The spatial distribution features and the driving force of fire spread are concatenated by channel alignment to obtain a concatenated feature tensor. A dimension reduction weight matrix is ​​set, and the concatenated feature tensor is multiplied by the dimension reduction weight matrix to perform linear mapping dimension reduction calculation to generate the initial spatiotemporal hidden state tensor sequence.

[0024] The spatiotemporal memory layer is used to read the initial spatiotemporal hidden state tensor sequence. Long short-term memory units are constructed to perform temporal state evolution calculations on the initial spatiotemporal hidden state tensor sequence. The hidden state of the previous time step is multiplied by the forget gate weight and the current input feature is multiplied by the input gate weight to generate global spatiotemporal semantic features. A bidirectional network with two hidden layers is constructed to perform bidirectional state interaction calculations between layers. The global spatiotemporal semantic features are passed down, and the dot product attention score between the global spatiotemporal semantic features and the local grid features is calculated as a constraint coefficient. This constraint coefficient is multiplied by the local grid features to constrain the extraction of local spatiotemporal detail features, and the obtained local spatiotemporal detail features are fed back up. The global spatiotemporal semantic features and the local spatiotemporal detail features are added by channel, and the calculation result is mapped to between -1 and 1 through the Tanh activation function, outputting a memory tensor sequence that fuses global and local spatiotemporal dependencies.

[0025] The spatial feature interaction modulation layer is used to read the memory tensor sequence that fuses global and local spatiotemporal dependencies. It maps the sequence into a multi-channel spatiotemporal state tensor using a 1x1 convolutional kernel, with 64 output channels. The layer reads the fire spread driving force, unfolds it into a one-dimensional vector along the spatial dimension, and multiplies it by a driving weight matrix of dimension 64 to generate an external driving constraint feature vector. A 3x3 neighborhood window is set to perform multi-mode spatial neighborhood feature aggregation on the multi-channel spatiotemporal state tensor. The average value of the center point and its eight neighboring points within the window is calculated as the spatial interaction aggregation feature. The spatial interaction aggregation feature and the external driving constraint feature vector are fused by element-wise multiplication to generate the evolutionary motion feature tensor.

[0026] Differentiable aggregation phase transition layer is used to introduce a differentiable aggregation phase transition mechanism, decouple and differentially calculate the evolutionary motion feature tensor, generate local net response features and map them to a state probability distribution matrix; perform anisotropic direction-weighted diffusion evolution on the state probability distribution matrix based on the fire spread driving force; couple the evolved state probability distribution matrix with the local net response features, determine the dynamic phase transition critical value based on the statistical distribution of the coupled features and perform differentiable state mapping to generate target boundary prediction results, which are output as the fire line expansion boundary.

[0027] In its implementation, this invention inherits the basic framework of the traditional PredRNN network, which performs local convolution on the spatiotemporal state tensor to extract spatial distribution features and performs temporal state evolution along the time dimension. However, it has undergone a deep transformation based on this framework, which deeply integrates the physical mechanism.

[0028] The core of the transformation lies in the introduction of a differentiable aggregated phase transition layer and a spatial feature interaction modulation layer. This abandons the limitation of traditional models that rely solely on the linear propagation of hidden states. Instead, it performs bidirectional state interaction between layers in the spatiotemporal memory layer to integrate global semantics and local detailed features. In the spatial feature interaction modulation layer, external driving constraint feature vectors are generated based on the fire spread driving force to perform multi-mode spatial neighborhood feature aggregation. More importantly, the evolutionary motion feature tensor is decoupled and differentially calculated through the differentiable aggregated phase transition mechanism to generate local net response features and map them into a state probability distribution matrix. Then, anisotropic direction-weighted diffusion evolution is performed based on the driving force, and the phase transition critical value is dynamically determined based on the statistical distribution of coupling features to perform differentiable state mapping.

[0029] This transformation completely solves the problems of traditional fire simulation methods, such as lack of physical phase transition constraints, difficulty in capturing sudden fire line boundaries, and anisotropic diffusion distortion caused by complex wind fields and terrain. It realizes the leap from pure data-driven to physical-guided differentiable computation, significantly improving the physical rationality and spatiotemporal accuracy of fire line expansion boundary prediction during the dynamic evolution of the fire scene, and providing a highly reliable forward situational awareness foundation for subsequent multi-agent collaborative scheduling.

[0030] In this embodiment, the microaggregable phase transition mechanism specifically includes: Read the evolutionary motion feature tensor, construct a convolutional layer with 2 output channels, set the kernel size to 3 by 3 grids, and set the stride to 1 grid. Perform channel-dimensional feature decoupling mapping on the evolutionary motion feature tensor. Define the first output channel feature as the positive response feature and the second output channel feature as the negative suppression feature to obtain the positive response feature and negative suppression feature.

[0031] Read the positive response features and negative suppression features, perform spatial projection on the positive response features and negative suppression features to align them one by one on the spatial grid coordinates; calculate the pixel value of each pixel of the positive response feature after spatial projection minus the corresponding pixel value of the negative suppression feature, perform pixel-by-pixel difference calculation, and use the difference result as the local net response feature.

[0032] Read the local net response features, calculate the difference between each pixel value in the local net response features and the global mean, divide it by the global standard deviation, and perform differentiable normalization calculation; input the normalized values ​​into the Sigmoid activation function to smoothly map the continuous values ​​of the local net response features to the probability interval between 0 and 1, and generate the state probability distribution matrix.

[0033] The driving force of fire spread is read, the wind direction angle is used as the driving direction vector, and the wind speed in meters per second is used as the driving intensity scalar. The direction decoupling calculation of the driving force of fire spread is performed. Based on the driving direction vector, a 3x3 spatial neighborhood direction mask is constructed, and the mesh mask value for the downwind direction is set to 0.8, the mesh mask value for the upwind direction is 0.1, and the mesh mask value for the crosswind direction is 0.4. The mesh mask value of each direction is multiplied by the driving intensity scalar, and the spatial neighborhood direction mask is subjected to asymmetric weight allocation calculation in combination with the driving intensity scalar to generate an anisotropic spatial transfer weight matrix.

[0034] Read the anisotropic spatial transition weight matrix and the state probability distribution matrix. Use the anisotropic spatial transition weight matrix as the convolution kernel. Perform spatial neighborhood convolution calculation on the state probability distribution matrix using the anisotropic spatial transition weight matrix. Multiply the weight values ​​of the region covered by the convolution kernel with the corresponding values ​​of the state probability distribution matrix and sum them to obtain the diffusion update probability matrix.

[0035] Read the diffusion update probability matrix and the local net response features, concatenate the diffusion update probability matrix and the local net response features along the feature channel dimension, perform feature fusion calculation, and obtain the comprehensive response state features.

[0036] Read the comprehensive response state features, calculate the global mean of all elements within the comprehensive response state features plus twice the global standard deviation as the dynamic state critical value; perform differentiable threshold activation calculation on the comprehensive response state features. When the value of the comprehensive response state features is greater than the dynamic state critical value, input the value into the ReLU activation function to perform nonlinear state mapping and retain the original value. When the value of the comprehensive response state features is less than or equal to the dynamic state critical value, set the value to 0, generate the target boundary prediction result, and output it as the fire line extension boundary.

[0037] In this embodiment, S3 specifically includes: S31. Read the fire line expansion boundary, extract a discrete point every 10 meters along the boundary line, and summarize the coordinates of all discrete points into a boundary discrete point set; calculate the reciprocal of the radius of the circumcircle formed by each point in the boundary discrete point set and its two adjacent points, and use the obtained value as the local curvature feature; calculate the mean of all local curvature features, set it to 0.05 per meter, and the standard deviation to 0.02 per meter, and calculate the sum of the mean and twice the standard deviation to obtain the adaptive curvature filtering threshold of 0.09 per meter; traverse the boundary discrete point set, and determine whether the local curvature feature value corresponding to each point is greater than the adaptive curvature filtering threshold of 0.09 per meter. If it is greater, keep the point; if it is less than or equal to, remove the point. Use this filtering operation to output the set of fire spread precursor points.

[0038] S32. Read the fire line expansion boundary and extract the time difference between the prediction results of two adjacent frames as 10 minutes, which is taken as the time resolution of the fire line expansion boundary; read the fire spread precursor point set, calculate the square difference between the horizontal coordinate of the precursor point in the current frame and the horizontal coordinate of the previous frame, plus the square difference between the vertical coordinate, and then take the square root of the result to obtain the spatial displacement vector of the fire spread precursor point set between adjacent time frames; calculate the quotient of the spatial displacement vector divided by the time resolution of the fire line expansion boundary of 10 minutes, and take this quotient as the local spread rate of the fire spread precursor point set, in meters per minute.

[0039] S33. Obtain the coordinate data of the fire ignition point, read the set of fire spread precursor points, calculate the square of the difference between the horizontal coordinate of the fire ignition point and the horizontal coordinate of the precursor points, add the square of the difference between the vertical coordinate of the fire ignition point and the vertical coordinate of the precursor points, take the square root of the result, calculate the spatial distance from the current fire ignition point to each precursor point in the set of fire spread precursor points, in meters; read the local spread rate, calculate the quotient of the spatial distance divided by the local spread rate, use the quotient as the predicted arrival time of the fire, in minutes, and output the predicted arrival time of the fire.

[0040] In this embodiment, S4 specifically includes: S41. Obtain the initial state distribution of the UAV agent, read the set of fire spread precursor points, calculate the square of the difference between the current position of the UAV and each precursor point, plus the square of the difference between the x-coordinates, and take the square root of the result to obtain the spatial distance as the spatial topological feature; calculate the global mean of all spatial topological features as 500 meters and the global standard deviation as 100 meters, calculate the quotient of 1 plus the spatial topological feature divided by the global standard deviation, input the quotient into the Sigmoid activation function to map the value to between 0 and 1, and use it as the element value of the adaptive target allocation weight matrix; use the Hungarian algorithm, with the optimization objective of minimizing the sum of the element values ​​of the adaptive target allocation weight matrix, perform global optimal matching on the UAV agent and the set of fire spread precursor points, and output the precursor point-UAV mapping pair.

[0041] S42. Read the preceding point-UAV mapping pair. Starting from the current position of the UAV and ending at the mapped preceding point position, construct a node network based on the distribution of environmental obstacles. Extract the spatial connectivity graph from the UAV agent to the corresponding set of preceding points for fire spread. Within the spatial connectivity graph, set the UAV's cruising speed to 15 meters per second. Calculate the flight time between adjacent nodes using the Dijkstra algorithm. Perform shortest path search to obtain the initial discrete path point sequence. Use the cubic Bézier curve function to perform interpolation fitting calculations on adjacent nodes in the initial discrete path point sequence to complete the trajectory smoothing calculation. Assign a corresponding timestamp to each smoothed path point and output the UAV predicted guidance spatiotemporal trajectory set.

[0042] S43. Read the UAV predicted guidance spatiotemporal trajectory set, extract the time value corresponding to the last path point in the sequence as the end node timestamp; obtain the kinematic dynamic constraints of the UAV agent, set the maximum horizontal deceleration of the UAV to 2 meters per second squared, the hovering attitude adjustment time to 5 seconds, calculate the time taken for the UAV to decelerate from a cruising speed of 15 meters per second to 0, add the hovering attitude adjustment time of 5 seconds to this time to obtain the dynamic compensation time; add the time value of the end node timestamp to the dynamic compensation time, perform time compensation calculation, and output the arrival time of the UAV.

[0043] In this embodiment, S5 specifically includes: S51. Obtain the basic performance parameters of the fire truck, the terrain access data of the fire scene, and the initial position distribution of the fire truck; set the maximum driving speed of the fire truck in the basic performance parameters to 60 kilometers per hour and the maximum climbing angle to 30 degrees; obtain the terrain access data of the fire scene, and extract the terrain slope value in degrees and the surface roughness value of each grid; obtain the initial position distribution of the fire truck, and set its longitude coordinate to 116.4 degrees and its latitude coordinate to 39.9 degrees.

[0044] S52. The global mean of the terrain slope values ​​of all grids in the statistical fireground terrain access data is 10 degrees, and the global mean of the surface roughness values ​​is 0.2. Based on this statistical distribution characteristic, an adaptive access speed field is calculated. The speed benchmark value is set to 60 kilometers per hour. The speed benchmark value is multiplied by 1 (in parentheses), minus the terrain slope divided by the maximum climbing angle of 30 degrees multiplied by 0.5, and then minus the surface roughness multiplied by 0.5. The result is used as the access speed value of each grid. Combining the basic performance parameters of the fire truck, using the adaptive access speed field and the initial position distribution of the fire truck, and setting a time limit of 15 minutes, the Dijkstra algorithm is used to calculate the set of all grid points that can be reached within 15 minutes from the initial position distribution point of the fire truck. The spatial area formed by this set is used as the accessibility constraint of the fire truck.

[0045] S53. Extract the predicted arrival time of the fire and set it to 30 minutes and the arrival time of the drone to 10 minutes. Calculate the difference between the predicted arrival time of the fire (30 minutes) and the arrival time of the drone (10 minutes) to obtain a time difference of 20 minutes. Based on the time difference of 20 minutes and the accessibility constraint of the fire truck, set the average time taken for the fire truck to reach the edge of the accessibility constraint to 5 minutes. Calculate the time difference of 20 minutes and the average time taken to 5 minutes to obtain a time truncation safety margin of 15 minutes. Read the set of fire spread precursor points, set the local spread rate to 5 meters per minute, calculate the local spread rate of 5 meters per minute multiplied by the time truncation safety margin of 15 minutes to obtain an offset distance of 75 meters. Move the coordinates of each precursor point in the set of fire spread precursor points 75 meters in the opposite direction of the fire spread direction. Perform spatial reverse offset calculation and output the cutoff strike point.

[0046] S54. Based on the constraints of the cutoff strike point and the fire truck's accessibility, the Dijkstra algorithm is used to perform path time estimation in an adaptive traffic speed field. The calculated path time is set to 12 minutes, and the fire truck arrival time is output as 12 minutes. Spatiotemporal correlation encoding is performed on the cutoff strike point, the fire truck arrival time (12 minutes), the predicted fire arrival time (30 minutes), and the drone arrival time (10 minutes). The latitude and longitude coordinates of the cutoff strike point, the fire truck arrival time, the predicted fire arrival time, and the drone arrival time are concatenated into a feature vector. The feature vector is then input into the ReLU activation function. Elements greater than 0 are retained, and elements less than 0 are set to 0. The candidate cutoff strike task set is output.

[0047] In the specific implementation process, this invention inherits the basic environmental and kinematic constraints of fire truck basic performance parameters, fire site terrain access data and initial position distribution in the traditional dispatching method, but has made in-depth modifications to the processing mechanism. The core is to deeply couple dynamic fire prediction with the time-series collaboration of air and ground multi-agents.

[0048] Specifically, the traditional static target point direct allocation mode was abandoned, and instead an adaptive traffic speed field was calculated based on the statistical distribution characteristics of fire terrain access data. This led to the derivation of fire truck accessibility constraints that better fit the actual fire terrain. Furthermore, the temporal difference between the predicted fire arrival time and the arrival time of the UAV was extracted, and the temporal truncation safety margin was calculated in conjunction with the accessibility constraints. This margin was then creatively used to perform spatial reverse offset calculations on the set of fire spread precursor points to dynamically solve the truncation strike points.

[0049] This transformation completely solves the problem of "arrival is already outdated" spatiotemporal disconnect caused by the speed difference between drones and fire trucks and the dynamic spread of fire in traditional methods. It avoids fire trucks missing their targets or being delayed, and ensures that fire trucks can accurately intercept the fire line before the fire arrives. Finally, by performing spatiotemporal correlation coding on the interception point and the arrival time of multiple parties to generate a candidate interception and strike task set, it realizes efficient linkage of air and ground cross-platform resources and precise dynamic interception and strike, which significantly improves the scientific nature and actual success rate of fire fighting and rescue dispatch strategies.

[0050] In this embodiment, S6 specifically includes: S61. Read the UAV predicted guidance spatiotemporal trajectory set and the candidate interception strike mission set. Calculate the sum of the squares of the spatial distance (in meters) and time difference (in seconds) of each mission point in the UAV predicted guidance spatiotemporal trajectory set and the candidate interception strike mission set. Take the square root of the result and use the result as the spatiotemporal overlap feature. Calculate the global mean of all spatiotemporal overlap features as 100 meters and the global standard deviation as 50 meters. Calculate the quotient of 1 divided by the 1 in parentheses plus the spatiotemporal overlap feature divided by the global standard deviation. Input the calculation result into the Sigmoid activation function to map it between 0 and 1, and output the adaptive collaborative matching weight. Using the adaptive collaborative matching weight, set the matching weight threshold to 0.5. When the adaptive collaborative matching weight is greater than 0.5, the matching is considered successful. Perform a global association mapping on the UAV predicted guidance spatiotemporal trajectory set and the candidate interception strike mission set, and output the collaborative matching relationship set.

[0051] S62. Read the cooperative matching relationship set, calculate the three-dimensional spatial distance between each UAV predicted guidance spatiotemporal trajectory set and the candidate interception strike mission set at the same timestamp. If the spatial distance is less than the set safe distance threshold of 20 meters, then the spatial distance value is taken as the spatiotemporal conflict feature. The global mean of the spatiotemporal conflict feature is 10 meters. Calculate 1 plus the quotient of the spatiotemporal conflict feature divided by the global mean, and use the calculation result as the adaptive conflict resolution penalty factor. Using the adaptive conflict resolution penalty factor, set the offset step size to 5 meters, and offset the coordinate points of the UAV predicted guidance spatiotemporal trajectory set with conflict in the cooperative matching relationship set along the right direction perpendicular to the original heading by the offset step size multiplied by the size of the adaptive conflict resolution penalty factor. At the same time, delay the execution time of the candidate interception strike mission set by 2 seconds, perform trajectory offset and mission rearrangement calculation, and output the conflict-free cooperative mission sequence.

[0052] S63. Read the conflict-free collaborative task sequence, extract the timestamp and spatial coordinates of each task in the conflict-free collaborative task sequence, sort the tasks in order of timestamp from earliest to latest, calculate the difference between the timestamps of adjacent tasks as the execution interval time, connect adjacent tasks with an interval time greater than 0, and perform time-series cascading arrangement; bind each task in the conflict-free collaborative task sequence with the corresponding drone or fire truck identifier, set the drone resource capacity to 1 task and the fire truck resource capacity to 1 task, determine whether the task allocation quantity is less than or equal to the resource capacity, if the condition is met, the binding is successful, and output the fire fighting and rescue dispatch strategy.

[0053] In its implementation, this invention inherits the basic process of multi-agent task allocation and trajectory conflict detection in traditional scheduling methods, but it has been deeply modified on this basis. The core technical point is to place the UAV prediction guidance spatiotemporal trajectory set and the candidate interception strike task set in a unified spatiotemporal dimension for dynamic solution. It abandons the traditional static allocation mode based on a single distance or fixed priority, and instead extracts the spatiotemporal overlap features between the two and calculates the adaptive collaborative matching weight based on its statistical distribution to perform global association mapping. At the same time, it dynamically calculates the adaptive conflict resolution penalty factor based on the extracted spatiotemporal conflict features, and uses this factor to adaptively perform trajectory offset and task rearrangement calculation on the collaborative matching relationship set.

[0054] This transformation completely solves the problems of resource competition and spatiotemporal overlap conflicts among air and ground multi-agents in complex fire environments in traditional methods, avoids physical collisions and task blockages during execution, and finally outputs conflict-free collaborative task sequences and fire fighting and rescue scheduling strategies through temporal cascading orchestration and execution resource binding, significantly improving the safety of multi-agent collaborative execution and the temporal continuity of the overall rescue mission.

[0055] In this embodiment, S7 specifically includes: S71. Obtain the fire-fighting and rescue dispatch strategy and the kinematic constraint distribution of multiple agents. Set the maximum speed of the multiple agents to 10 meters per second and the minimum turning radius to 5 meters. Calculate the average maximum speed of the kinematic constraint distribution of the multiple agents to 8 meters per second and the average minimum turning radius to 4 meters. Based on these statistical characteristics, calculate the adaptive spatiotemporal grid resolution. Set the spatial reference step size to 2 meters and the time reference step size to 1 second. Calculate the spatial resolution value of 0.25 meters by multiplying the spatial reference step size by the reciprocal of the average maximum speed. Calculate the time resolution value of 0.25 seconds by multiplying the time reference step size by the reciprocal of the average minimum turning radius. Combine the two as the adaptive spatiotemporal grid resolution. Use the adaptive spatiotemporal grid resolution to perform spatiotemporal grid mapping on the trajectory points and execution times in the fire-fighting and rescue dispatch strategy. Divide each trajectory point into grids according to a spatial resolution of 0.25 meters and divide the execution time into time slices according to a time resolution of 0.25 seconds. Statistically analyze the agent occupancy of each grid in each time slice and output the spatiotemporal grid occupancy map.

[0056] S72. Read the spatiotemporal grid occupancy map and calculate the three-dimensional spatial distance between the center points of different agent grids within the same time slice. If this distance is less than the set overlap distance threshold of 1 meter, then this distance value is used as the multi-agent spatial overlap feature. The global mean of the multi-agent spatial overlap feature is 0.5 meters and the global standard deviation is 0.1 meters. The sum of the global mean and twice the global standard deviation is calculated to get 0.7 meters. 0.7 meters is input into the Sigmoid activation function to map it to between 0 and 1 and multiplied by the baseline safety threshold of 5 meters to calculate the adaptive conflict detection threshold of 3.5 meters. Multi-agent conflict detection is performed on the spatiotemporal grid occupancy map using the adaptive conflict detection threshold of 3.5 meters. It is determined whether the distance between the center points of different agents within the same time slice in the spatiotemporal grid occupancy map is less than the adaptive conflict detection threshold of 3.5 meters. If it is less, it is marked as a conflict state. All spatiotemporal grid sequences marked as conflict states are summarized and the conflict detection results are output.

[0057] S73. Read the collision detection results and extract the trajectory segments marked as collision states from the collision detection results as the trajectory segments to be reconstructed. Calculate the duration of each collision point in the trajectory segment to be reconstructed. The mean of all durations is 2 seconds and the standard deviation is 0.5 seconds. Calculate the mean plus twice the standard deviation to get 3 seconds. Input 3 seconds into the ReLU activation function to retain the original value and multiply it by the unit time offset reference of 1 meter per second to calculate the adaptive trajectory offset step size of 3 meters. Use the adaptive trajectory offset step size of 3 meters to perform trajectory anti-collision reconstruction on the trajectory segment to be reconstructed. Offset the spatial coordinates of the collision points in the trajectory segment to be reconstructed by 3 meters perpendicular to the original motion direction to generate a new collision-free trajectory segment to replace the original trajectory segment to be reconstructed. Combine all the reconstructed trajectory segments in time sequence and output the time-series linkage execution instruction set.

[0058] In this embodiment, S8 specifically includes: S81. Obtain the actual operating state of the multi-agent system and the time-series linkage execution instruction set. Set the actual operating state of the multi-agent system to include a timestamp of 10 seconds and spatial coordinates of 5 meters x and 5 meters y. Extract the predicted operating state corresponding to the time-series linkage execution instruction set. Set the predicted operating state to include a timestamp of 10 seconds and spatial coordinates of 6 meters x and 5 meters y. Align the execution time of the actual operating state and the predicted operating state of the multi-agent system. Calculate the difference between the actual timestamp and the predicted timestamp to be 0 seconds. Perform spatial registration calculation. Calculate the difference between the actual spatial coordinate x of 5 meters and the predicted spatial coordinate x of 6 meters to obtain the x-coordinate deviation of -1 meter. Calculate the difference between the actual spatial coordinate y of 5 meters and the predicted spatial coordinate y of 5 meters to obtain the y-coordinate deviation of 0 meters. Calculate the square root of the square of the x-coordinate deviation plus the square of the y-coordinate deviation to obtain the deviation distance of 1 meter. Combine the time difference of 0 seconds and the deviation distance of 1 meter to output the spatiotemporal deviation.

[0059] S82. Read the spatiotemporal deviation and calculate the global mean of all deviation distances in the spatiotemporal deviation as 2 meters and the global standard deviation as 0.5 meters, which are used as the statistical distribution characteristics of the spatiotemporal deviation. Calculate the adaptive deviation tolerance threshold based on the statistical distribution characteristics of the spatiotemporal deviation. Calculate the sum of the global mean of 2 meters and twice the global standard deviation of 1 meter to get 3 meters. Input 3 meters into the Sigmoid activation function, map it to the range of 0 to 1, and multiply it by the baseline tolerance coefficient of 10 meters to calculate the adaptive deviation tolerance threshold of 9.5 meters. Use the adaptive deviation tolerance threshold of 9.5 meters to perform dynamic evaluation calculation on the spatiotemporal deviation, and determine whether the deviation distance of 1 meter in the spatiotemporal deviation is greater than the adaptive deviation tolerance threshold of 9.5 meters. If the result is 1 meter less than or equal to 9.5 meters, mark it as a normal state and output the deviation evaluation result as a normal state.

[0060] S83. Obtain the fire-fighting and rescue dispatch strategy and read the deviation assessment results; based on the deviation assessment results and the fire-fighting and rescue dispatch strategy execution strategy feedback compensation calculation, judge the deviation assessment results. When the deviation assessment results are in an abnormal state, set the compensation correction coefficient to 1.2, and multiply the execution time in the fire-fighting and rescue dispatch strategy by the compensation correction coefficient 1.2 for delay adjustment; when the deviation assessment results are in a normal state, set the compensation correction coefficient to 1.0, multiply the execution time in the fire-fighting and rescue dispatch strategy by the compensation correction coefficient 1.0 to remain unchanged, recombine the time parameters after compensation correction with the spatial path parameters in the original fire-fighting and rescue dispatch strategy, and output the iteratively updated fire-fighting and rescue dispatch strategy.

[0061] Example 1: To verify the feasibility of an agent-based collaborative firefighting and rescue dispatching method in forest fire suppression, the proposed method was applied to the intelligent firefighting and rescue command system of a provincial forest fire brigade (hereinafter referred to as "Brigade F"). In traditional forest fire suppression dispatching systems, manual command based on static fire spread models or heuristic task allocation algorithms are typically used. These methods not only struggle to accurately predict future fire line expansion boundaries in complex wind fields and terrain environments, but also fail to accurately handle the spatiotemporal coordination conflicts between drones and fire trucks, easily leading to errors in the calculation of cutoff points or agent collisions. To address these issues, Brigade F decided to adopt an agent-based collaborative firefighting and rescue dispatching method.

[0062] During implementation, Brigade F first utilized multi-source sensors deployed on reconnaissance drones and high-altitude satellites to acquire raw, multi-source observation data of the fire scene. Environmental parameters and observation spatiotemporal labels were extracted. Based on the spatial distribution characteristics and temporal statistical distribution of the observation spatiotemporal labels, an adaptive spatiotemporal alignment benchmark was calculated. The environmental parameters were projected and stacked along the time dimension to output the fire scene environmental spatiotemporal state tensor. Simultaneously, Brigade F's command experts precisely calibrated the collected terrain accessibility data and basic performance parameters of the fire trucks, serving as the benchmark for solving the fire truck accessibility constraints.

[0063] Team F improved the PredRNN model by using a spatiotemporal feature encoding layer to extract spatial distribution features and fire spread driving forces. A spatiotemporal memory layer performs bidirectional state interaction calculations between layers, outputting a memory tensor sequence that integrates global and local spatiotemporal dependencies. A spatial feature interaction modulation layer generates external driving constraint feature vectors based on fire spread driving forces, performs multi-mode spatial neighborhood feature aggregation operations, and generates an evolutionary motion feature tensor. A differentiable aggregated phase transition layer decouples and performs differential calculations on the evolutionary motion feature tensor, generating local net response features and mapping them to a state probability distribution matrix. Combining the driving direction vector and driving intensity scalar, it performs anisotropic direction-weighted diffusion evolution, accurately outputting the fire line expansion boundary with dynamic phase transition characteristics.

[0064] In the core scheduling generation and execution phase, the system extracts a set of fire spread precursor points from the fire line expansion boundary, calculates cutoff strike points by combining the predicted fire arrival time and the arrival time of UAVs, and generates a candidate cutoff strike task set. By extracting spatiotemporal overlap and spatiotemporal conflict features, it calculates an adaptive conflict resolution penalty factor to execute trajectory offset and task rearrangement, outputting a conflict-free collaborative task sequence and fire-fighting and rescue scheduling strategy. Subsequently, based on the kinematic constraint distribution of multi-agent systems, it calculates an adaptive spatiotemporal grid resolution, extracts the trajectory segments to be reconstructed using an adaptive conflict detection threshold, and performs trajectory collision avoidance reconstruction by combining an adaptive trajectory offset step size, outputting a time-series coordinated execution instruction set.

[0065] In the closed-loop iteration and compensation phase, the system acquires the actual operating states of multiple agents and the predicted operating states corresponding to the time-series coordinated execution instruction set, and calculates the spatiotemporal deviation based on execution time alignment and spatial registration. An adaptive deviation tolerance threshold is calculated based on the statistical distribution characteristics of the spatiotemporal deviation, and the deviation evaluation result is output. Finally, the deviation evaluation result is fed back to the fire-fighting and rescue dispatch strategy execution strategy for compensation calculation, outputting an iteratively updated fire-fighting and rescue dispatch strategy, thus achieving a closed-loop transition from fire prediction to agent collaborative execution.

[0066] During implementation, the technical team of Brigade F discovered that, compared with traditional static fire simulation and conventional dispatching methods, the agent-based collaborative firefighting and rescue dispatching method significantly improved the accuracy of fire situational awareness and the safety of multi-agent dispatching. Conventional methods cannot quantify spatiotemporal conflicts among multiple agents and have poor performance in identifying fire line boundaries in complex fire environments. However, the above process, through a differentiable aggregation phase transition mechanism, spatiotemporal grid anti-collision reconstruction, and deviation feedback compensation, effectively achieves accurate simulation of the dynamic evolution of the fire line and zero-conflict collaborative dispatching of multiple agents.

[0067] To further verify the actual performance, the brigade conducted a detailed comparative test between the agent-based firefighting and rescue dispatch method and the conventional method. Specific performance data is shown in Table 1. Table 1 Performance Comparison of Firefighting and Rescue Dispatch Methods of the General Team F

[0068] As shown in Table 1, the system performance was comprehensively improved after applying the agent-based collaborative firefighting and rescue dispatching method. The accuracy of fire line expansion boundary prediction increased from 78.5% with conventional methods to 95.2%, and the calculation error of interception points decreased from 25.6 meters to 3.1 meters, significantly improving the accuracy of fire situation awareness and providing a reliable basis for subsequent multi-agent dispatching. The multi-agent trajectory conflict rate decreased from 15.2% to 0.5%, effectively avoiding agent collision accidents. The time taken to generate dispatching strategies was significantly reduced from 120 seconds to 35 seconds, significantly enhancing the system's timeliness. In addition, the firefighting task completion rate increased from 72.0% to 94.5%, and the average arrival delay of fire trucks decreased from 15.4 minutes to 2.2 minutes, significantly reducing rescue response time. The frequency of manual intervention in command and dispatch also decreased significantly, from 12 times / hour to 1 time / hour.

[0069] By adopting an agent-based collaborative firefighting and rescue dispatching method, the General Team F successfully achieved accurate simulation of dynamic fire lines and zero-conflict collaborative dispatching of multiple agents. This effectively reduced the risk of fire spreading out of control, ensured the safety of rescue personnel and equipment, significantly improved the intelligence and digitalization level of forest firefighting and rescue, significantly reduced the workload of command personnel, enhanced the stability and robustness of the dispatching system, and provided strong technical support for the construction of smart fire protection.

[0070] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fire fighting and rescue dispatching method based on intelligent agent collaboration, characterized in that, Includes the following steps: S1. Acquire multi-source observation data of the fire scene and perform spatiotemporal benchmark alignment and gridded reconstruction, and output the spatiotemporal state tensor of the fire scene environment. S2. Input the spatiotemporal state tensor of the fire environment into the improved PredRNN model, introduce a differentiable aggregation phase transition mechanism, perform spatiotemporal evolution modulation and differentiable aggregation phase transition mapping under the constraint of fire spread driving force, and deduce the fire line expansion boundary for future periods. S3. Select and generate a set of fire spread precursor points from the fire line expansion boundary, and calculate the predicted arrival time of the fire corresponding to the set of fire spread precursor points. S4. Based on the set of fire spread precursor points, target allocation and flight path calculation are performed on the UAV agent, and the predicted guidance spatiotemporal trajectory set of the UAV and the corresponding UAV arrival time are output. S5. Obtain fire truck accessibility constraints, combine fire prediction arrival time with UAV arrival time to calculate cutoff strike points associated with the fire spread precursor point set, deduce fire truck arrival time and generate candidate cutoff strike task set; S6. Perform spatiotemporal collaborative matching and conflict resolution on the candidate interception strike mission set and the UAV predicted guidance spatiotemporal trajectory set to generate fire fighting and rescue dispatch strategy. S7. Perform spatiotemporal grid mapping and multi-agent conflict detection on the fire fighting and rescue dispatch strategy, and perform trajectory collision avoidance reconstruction based on the detection results, and output a set of time-series linkage execution instructions; S8. Execute the time-series linkage execution instruction set, obtain the actual operating status of the multi-agent, calculate the spatiotemporal deviation between the actual operating status of the multi-agent and the predicted operating status corresponding to the time-series linkage execution instruction set, and perform iterative updates of the fire-fighting and rescue dispatch strategy based on the deviation evaluation results of the spatiotemporal deviation.

2. The fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, S1 specifically includes: S11. Obtain multi-source raw observation data of the fire scene, extract environmental parameters and observation spatiotemporal labels from the multi-source raw observation data of the fire scene; based on the spatial distribution characteristics and time span statistical distribution of the observation spatiotemporal labels, calculate the adaptive spatiotemporal alignment benchmark, project the environmental parameters onto the adaptive spatiotemporal alignment benchmark, and output the benchmark aligned multi-source parameter sequence. S12. Based on the spatial distribution density statistical characteristics of the benchmark-aligned multi-source parameter sequence, calculate the adaptive spatial grid resolution; construct a spatial grid index according to the adaptive spatial grid resolution, and use the spatial grid index to perform spatial interpolation calculation on the benchmark-aligned multi-source parameter sequence to output the gridded environmental parameter matrix. S13. Extract the timestamp sequence of the gridded environmental parameter matrix. Based on the timestamp sequence, splice the gridded environmental parameter matrices of different environmental types along the feature channel dimension and stack them along the time dimension to output the fire environment spatiotemporal state tensor.

3. The fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, The improved PredRNN model includes a spatiotemporal feature encoding layer, a spatiotemporal memory layer, a spatial feature interaction modulation layer, and a differentiable aggregated phase transition layer: The spatiotemporal feature encoding layer is used to perform local convolution calculation on the fire environment spatiotemporal state tensor to extract spatial distribution features, and extract environmental parameters from the fire environment spatiotemporal state tensor to calculate the fire spread driving force; the spatial distribution features and the fire spread driving force are channel aligned and concatenated, and then dimensionality reduction calculation is performed through linear matrix mapping to generate an initial spatiotemporal hidden state tensor sequence. The spatiotemporal memory layer is used to perform temporal state evolution calculations in the time dimension on the initial spatiotemporal hidden state tensor sequence to generate global spatiotemporal semantic features; at the same time, it performs bidirectional state interaction calculations between layers, passing the global spatiotemporal semantic features downward to constrain the extraction of local spatiotemporal detail features, and feeding back the obtained local spatiotemporal detail features upward; by fusing the global spatiotemporal semantic features and the local spatiotemporal detail features, a memory tensor sequence that fuses global and local spatiotemporal dependencies is output. The spatial feature interaction modulation layer is used to map the memory tensor sequence that fuses global and local spatiotemporal dependencies into a multi-channel spatiotemporal state tensor; an external driving constraint feature vector is generated based on the fire spread driving force; a multi-mode spatial neighborhood feature aggregation operation is performed on the multi-channel spatiotemporal state tensor to obtain spatial interaction aggregation features; the spatial interaction aggregation features and the external driving constraint feature vector are fused and calculated to generate an evolutionary motion feature tensor; The differentiable aggregation phase transition layer is used to introduce a differentiable aggregation phase transition mechanism, decouple and perform differential calculation on the evolutionary motion feature tensor, generate local net response features and map them into a state probability distribution matrix; perform anisotropic direction-weighted diffusion evolution on the state probability distribution matrix based on the fire spread driving force; couple the evolved state probability distribution matrix with the local net response features, determine the dynamic phase transition critical value based on the statistical distribution of the coupled features and perform differentiable state mapping to generate target boundary prediction results, which are output as the fire line expansion boundary.

4. The fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 3, characterized in that, The specific mechanisms for differentiable aggregation phase transitions include: Perform channel-dimensional feature decoupling mapping on the evolutionary motion feature tensor to obtain positive response features and negative inhibition features; Spatial projection and pixel-by-pixel difference calculations are performed on the positive response features and the negative suppression features to generate local net response features; Differentiable normalization calculation is performed on the local net response features to smoothly map the continuous values ​​of the local net response features to the probability interval, thereby generating a state probability distribution matrix. The direction decoupling calculation is performed on the fire spread driving force to extract the driving direction vector and driving intensity scalar; a spatial neighborhood direction mask is constructed based on the driving direction vector, and an asymmetric weight allocation calculation is performed on the spatial neighborhood direction mask in combination with the driving intensity scalar to generate an anisotropic spatial transfer weight matrix. The spatial neighborhood convolution calculation of the state probability distribution matrix is ​​performed on the anisotropic spatial transition weight matrix to obtain the diffusion update probability matrix; The diffusion update probability matrix and the local net response features are fused to obtain the comprehensive response state features; The dynamic state critical value is calculated based on the statistical distribution of the comprehensive response state characteristics; a differentiable threshold activation calculation is performed on the comprehensive response state characteristics; when the comprehensive response state characteristics are greater than the dynamic state critical value, a nonlinear state mapping is performed to generate the target boundary prediction result, which is output as the fire line extension boundary.

5. The fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, S3 specifically includes: S31. Extract the discrete point set and local curvature features of the fire line expansion boundary; calculate the adaptive curvature filtering threshold based on the statistical distribution of the local curvature features; perform a filtering operation on the discrete point set of the boundary using the adaptive curvature filtering threshold, and output the set of fire spread precursor points; S32. Extract the temporal resolution of the fire line expansion boundary and extract the spatial displacement vector of the fire spread precursor point set between adjacent time frames; based on the spatial displacement vector and the temporal resolution of the fire line expansion boundary, calculate the local spread rate of the fire spread precursor point set. S33. Obtain the fire ignition point and calculate the spatial distance from the current fire ignition point to each preceding point in the set of fire spread precursor points; based on the quotient of spatial distance and local spread rate, calculate and output the predicted arrival time of the fire.

6. The fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, S4 specifically includes: S41. Obtain the initial state distribution of the UAV agent and the spatial topological features of the fire spread front point set; calculate the adaptive target allocation weight matrix based on the statistical distribution of the spatial topological features; use the adaptive target allocation weight matrix to perform global optimal matching between the UAV agent and the fire spread front point set, and output the front point-UAV mapping pair. S42. Based on the forward point-UAV mapping pair, extract the spatial connectivity graph from the UAV agent to the corresponding set of forward points of fire spread; perform shortest path search and trajectory smoothing calculation in the spatial connectivity graph, and output the set of predicted spatiotemporal trajectories for UAV guidance. S43. Extract the end node timestamp of the UAV prediction guidance spatiotemporal trajectory set; obtain the kinematic dynamic constraints of the UAV agent; perform time compensation calculation based on the end node timestamp and the kinematic dynamic constraints of the UAV agent; and output the arrival time of the UAV.

7. A fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, S5 specifically includes: S51. Obtain basic performance parameters of fire trucks, fire scene terrain access data, and initial position distribution of fire trucks; S52. Calculate the adaptive traffic speed field based on the statistical distribution characteristics of fire scene terrain access data; combine the basic performance parameters of the fire truck, and use the adaptive traffic speed field and the initial position distribution of the fire truck to calculate the accessibility constraints of the fire truck. S53. Extract the time difference between the predicted arrival time of the fire and the arrival time of the drone; calculate the time truncation safety margin based on the time difference and the accessibility constraints of the fire truck; use the time truncation safety margin to perform spatial reverse offset calculation on the set of fire spread precursor points and output the truncation strike point. S54. Based on the constraints of the cutoff point and the accessibility of the fire truck, perform path time estimation and output the arrival time of the fire truck; perform spatiotemporal correlation coding on the cutoff point, the arrival time of the fire truck, the predicted arrival time of the fire, and the arrival time of the drone, and output a candidate cutoff strike task set.

8. A fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, S6 specifically includes: S61. Extract the spatiotemporal overlap features between the UAV predicted guidance spatiotemporal trajectory set and the candidate truncation strike mission set; calculate the adaptive collaborative matching weights based on the statistical distribution of the spatiotemporal overlap features; use the adaptive collaborative matching weights to perform a global association mapping between the UAV predicted guidance spatiotemporal trajectory set and the candidate truncation strike mission set, and output the collaborative matching relationship set; S62. Extract spatiotemporal conflict features based on the collaborative matching relationship set; calculate the adaptive conflict resolution penalty factor based on the statistical distribution of the spatiotemporal conflict features; use the adaptive conflict resolution penalty factor to calculate the trajectory offset and task rearrangement of the collaborative matching relationship set, and output a conflict-free collaborative task sequence. S63. Based on the time-series cascading arrangement and execution resource binding of conflict-free collaborative task sequences, output fire-fighting and rescue scheduling strategies.

9. A fire fighting and rescue dispatching method based on intelligent agent collaboration according to claim 1, characterized in that, Specifically, S7 includes: S71. Obtain the fire fighting and rescue dispatch strategy and the kinematic constraint distribution of the multi-agent system; calculate the adaptive spatiotemporal grid resolution based on the statistical characteristics of the kinematic constraint distribution of the multi-agent system; perform spatiotemporal grid mapping on the fire fighting and rescue dispatch strategy using the adaptive spatiotemporal grid resolution, and output the spatiotemporal grid occupancy map. S72. Extract multi-agent spatial overlap features based on spatiotemporal grid occupancy map; calculate adaptive conflict detection threshold based on statistical distribution of multi-agent spatial overlap features; perform multi-agent conflict detection on spatiotemporal grid occupancy map using adaptive conflict detection threshold, and output conflict detection results; S73. Extract the trajectory segment to be reconstructed based on the collision detection results; calculate the adaptive trajectory offset step size based on the statistical distribution characteristics of the trajectory segment to be reconstructed; perform trajectory anti-collision reconstruction on the trajectory segment to be reconstructed using the adaptive trajectory offset step size, and output the timing linkage execution instruction set.

10. A fire fighting and rescue dispatching method based on agent collaboration according to claim 1, characterized in that, S8 specifically includes: S81. Obtain the actual operating state of the multi-agent system and the time-series linkage execution instruction set; extract the predicted operating state corresponding to the time-series linkage execution instruction set; perform time alignment and spatial registration calculation between the actual operating state and the predicted operating state of the multi-agent system, and output the spatiotemporal deviation. S82. Extract the statistical distribution characteristics of spatiotemporal deviation; calculate the adaptive deviation tolerance threshold based on the statistical distribution characteristics of spatiotemporal deviation; perform dynamic evaluation calculation on spatiotemporal deviation using the adaptive deviation tolerance threshold, and output the deviation evaluation results; S83. Obtain the fire fighting and rescue dispatch strategy; based on the deviation assessment results and the fire fighting and rescue dispatch strategy execution strategy feedback compensation calculation, output the iteratively updated fire fighting and rescue dispatch strategy.

Images