A fire emergency command system based on multi-terminal cooperation

Abstract

The application provides a fire emergency command system based on multi-terminal cooperation, comprising a fire scene global data sensing and cloud access module, a cross-terminal cloud edge cooperation and synchronization engine module, a fire emergency big data intelligent decision service module, a multi-terminal permission dynamic adaptation and secure transmission module, a fire task closed-loop cooperation execution module and an emergency data cloud precipitation and iterative optimization module. The core purpose of the application is to solve the technical bottlenecks of data dispersion, insufficient cooperation, inaccurate decision and no iteration capability of the traditional fire emergency command system. By constructing an intelligent closed-loop architecture of multi-terminal cooperation, combining innovative algorithms and secure transmission technology, precise fire situation prediction, optimal allocation of rescue resources, efficient multi-terminal cooperation execution and continuous system iterative optimization are realized, ultimately improving the efficiency, accuracy and safety of fire emergency command, and providing reliable technical support for efficient disposal of complex fire scenes.

Description

Technical Field

[0001] This invention relates to the fields of emergency management, data communication and intelligent decision-making, and in particular to a fire emergency command system based on multi-terminal collaboration. Background Technology

[0002] In the field of fire emergency command, traditional emergency systems have long faced multiple technical bottlenecks, making it difficult to meet the needs of efficient handling of complex fire scenarios.

[0003] In existing technologies, data collection is mostly limited to a single terminal or a local area, lacking full-area coverage. Furthermore, data storage across different terminals (such as command center screens and mobile devices on-site) is scattered, lacking a unified real-time synchronization mechanism, leading to delayed information sharing. Commanders cannot grasp the complete fire situation, resource status, and on-site execution status in real time. Regarding multi-terminal collaboration, traditional systems lack adaptability to command, execution, and equipment terminals, exhibiting poor operational consistency between terminals and significant delays in command issuance and on-site feedback, easily resulting in a disconnect between decision-making and execution. Decision-making relies on commanders' experience and lacks intelligent analysis support based on big data. Fire situation prediction relies solely on simple sensor data inferences, and rescue resource allocation and route planning do not fully consider multi-objective collaborative optimization, resulting in limited decision accuracy and difficulty in balancing rescue efficiency, resource consumption, and personnel safety. In addition, existing systems are mostly statically designed, lacking data accumulation and iterative optimization mechanisms, unable to continuously improve system performance through historical data. Even after long-term use, technical bottlenecks remain, making it difficult to adapt to complex fire scenarios with different building structures and meteorological conditions. Summary of the Invention

[0004] This invention provides a multi-terminal collaborative fire emergency command system. Its core objective is to address the technical bottlenecks of traditional fire emergency command systems, including data fragmentation, insufficient collaboration, inaccurate decision-making, and lack of iterative capabilities. By constructing a multi-terminal collaborative intelligent closed-loop architecture and integrating innovative algorithms and secure transmission technologies, it achieves accurate fire situation prediction, optimal allocation of rescue resources, efficient multi-terminal collaborative execution, and continuous system iteration and optimization. Ultimately, this improves the efficiency, accuracy, and safety of fire emergency command, providing reliable technical support for the efficient handling of complex fire scenarios. To achieve the above objectives, this invention adopts the following technical solution:

[0005] A fire emergency command system based on multi-terminal collaboration includes:

[0006] The fire scene full-domain data perception and cloud access module is used to collect fire emergency related full-domain data and perform preprocessing operations such as noise reduction, format standardization and outlier removal to obtain preprocessed data;

[0007] The cross-terminal cloud-edge collaborative synchronization engine module is used to receive the preprocessed data and synchronize it to all terminals and subsequent modules through a hybrid transmission channel to obtain synchronized data;

[0008] The fire emergency big data intelligent decision service module is used to receive the synchronous data, and to perform analysis and calculation using a graph convolutional long short-term memory network with spatiotemporal attention mechanism enhancement and an improved multi-objective whale optimization algorithm to generate a fire emergency command and decision scheme.

[0009] The multi-terminal permission dynamic adaptation and secure transmission module is used to receive the fire emergency command and decision-making scheme, perform permission filtering and encryption on it, obtain targeted data and transmit it to each terminal and subsequent modules.

[0010] The fire-fighting task closed-loop collaborative execution module is used to receive the directional data, issue tasks based on the data, and collect execution data and feedback data in real time during the task execution process;

[0011] The emergency data cloud accumulation and iterative optimization module is used to receive the execution data and feedback data, and to analyze and optimize them using an adaptive weighted multi-granularity feature fusion algorithm and a dual-feedback reinforcement learning algorithm to obtain optimized algorithm parameters. The optimized algorithm parameters are then fed back to the fire emergency big data intelligent decision service module.

[0012] In this specification, the working process of the graph convolutional long short-term memory network enhanced by the spatiotemporal attention mechanism is as follows: First, spatial features are extracted from the received synchronous data. Key regional features are strengthened and irrelevant regional interference is weakened through spatial attention weights. Then, spatial correlations between regions are mined based on the spatial adjacency matrix. Subsequently, the temporal features of the data are extracted and the situational changes at key time nodes are focused. Finally, temporal dependencies are mined through the long short-term memory network to output fire situation prediction data.

[0013] In this specification, the improved multi-objective whale optimization algorithm works as follows: taking fire situation prediction data output by a graph convolutional long short-term memory network with synchronized data and spatiotemporal attention mechanism as input, a diverse initial population is first generated through chaotic mapping, then three types of objective functions are constructed: rescue response time, rescue resource consumption, and rescue risk. The solution vector is optimized through iterative strategies of surrounding the prey and bubble net attacks, while the adaptive weights of each objective function are dynamically adjusted. Finally, the Pareto optimal solution is selected through non-dominated sorting, and a fire emergency command decision scheme is generated in combination with the actual needs of fire command.

[0014] In this specification, the spatiotemporal attention-enhanced graph convolutional long short-term memory network and the improved multi-objective whale optimization algorithm form a bidirectional interactive and collaborative relationship: the fire situation prediction data output by the spatiotemporal attention-enhanced graph convolutional long short-term memory network is used to correct the route distance parameter of rescue response time and the fire risk probability parameter of rescue risk in the improved multi-objective whale optimization algorithm; after the improved multi-objective whale optimization algorithm is executed, the actual fire situation data collected by the fire-fighting task closed-loop collaborative execution module is fed back to the spatiotemporal attention-enhanced graph convolutional long short-term memory network to verify the prediction accuracy, and the network is triggered to retrain when the prediction error exceeds a set threshold.

[0015] In this specification, the adaptive weighted multi-granularity feature fusion algorithm works as follows: hierarchical feature extraction is performed on the received execution data and feedback data. First, fine-grained features reflecting micro-changes in the data are extracted through one-dimensional convolution. Then, medium-grained features reflecting local patterns are extracted through one-dimensional pooling. Next, coarse-grained features reflecting global trends are extracted through principal component analysis. Subsequently, the signal-to-noise ratio of various granular features is calculated, adaptive weights are assigned according to the signal-to-noise ratio, and finally, the fused feature matrix is ​​obtained by weighted summation.

[0016] In this specification, the working process of the dual-feedback reinforcement learning algorithm is as follows: using the fusion feature matrix output by the adaptive weighted multi-granularity feature fusion algorithm as the state input, a total reward function containing decision feedback reward and execution feedback reward is constructed. The decision feedback reward is obtained by comparing the algorithm's predicted decision result with the expert's optimal decision result. The execution feedback reward is calculated by combining the actual rescue time with the planned rescue time and the actual rescue risk probability. Based on the total reward function, the algorithm parameter adjustment amount is output, and the weight matrix of the spatiotemporal attention mechanism-enhanced graph convolutional long short-term memory network and the adaptive weights of the improved multi-objective whale optimization algorithm are iteratively updated.

[0017] In this specification, the adaptive weighted multi-granularity feature fusion algorithm and the dual-feedback reinforcement learning algorithm form a bidirectional interactive and collaborative relationship: the fusion feature matrix output by the adaptive weighted multi-granularity feature fusion algorithm provides accurate state input for the dual-feedback reinforcement learning algorithm, ensuring the accuracy of the parameter optimization direction; the total reward function value output by the dual-feedback reinforcement learning algorithm is used to correct the adaptive weight calculation of the adaptive weighted multi-granularity feature fusion algorithm. When the total reward function value is lower than a set threshold, the adaptive weighted multi-granularity feature fusion algorithm is triggered to re-extract features and optimize weights.

[0018] In this specification, the training process of the spatiotemporal attention mechanism-enhanced graph convolutional long short-term memory network is as follows: historical fire situation data is collected and divided into training, validation, and test sets according to proportions; the weight matrix and bias terms of the network are randomly initialized; the learning rate and maximum number of iterations are set; the training set data is input into the network for forward propagation to calculate the predicted value; the error between the predicted value and the true value is calculated using the mean squared error loss function; the parameters are updated by backpropagation using the Adam optimization algorithm; the model accuracy is evaluated using the validation set after a set number of iterations; when the validation set loss does not decrease for multiple consecutive iterations, an early stopping mechanism is triggered, and the optimal model parameters are saved.

[0019] In this specification, the objective function optimization process of the improved multi-objective whale optimization algorithm is as follows: the three objective functions of rescue response time, rescue resource consumption, and rescue risk are normalized to eliminate the difference in dimensions. The initial weights are set according to the priority of fire emergency command. During the iteration process, the weights are dynamically adjusted according to the ratio of the current value to the maximum value of each objective function to ensure the coordinated optimization of the three types of objectives and avoid the overall scheme being insufficient due to the optimization of a single objective.

[0020] In this specification, the emergency data cloud accumulation and iterative optimization module is also used to store the entire process data of the system, including global data, preprocessed data, synchronized data, fire emergency command and decision-making schemes, execution data, and feedback data. The adaptive weighted multi-granularity feature fusion algorithm extracts multi-granularity features from the stored historical data, providing rich training data for the dual feedback reinforcement learning algorithm and ensuring the comprehensiveness and effectiveness of algorithm parameter optimization.

[0021] In summary, the present invention has at least the following beneficial effects:

[0022] Significantly improved efficiency of multi-terminal collaboration: Through real-time synchronization of data across the entire domain and closed-loop execution processes, seamless collaboration between the command end, execution end, and equipment end is achieved. There is no delay in the entire process of issuing instructions, providing on-site feedback, and adjusting plans, thus completely solving the problems of information silos and the disconnect between decision-making and execution.

[0023] Significantly improved decision-making accuracy: With the help of two-way interaction of innovative algorithms, fire situation prediction can accurately capture the spatiotemporal evolution pattern, and rescue resource allocation and route planning can achieve multi-objective coordination. The decision-making plan is more in line with the actual needs on site, reducing the risk of blind dispatch.

[0024] Security capabilities have been comprehensively enhanced: Through dynamic permission adaptation, multi-layered encrypted transmission, and operation log traceability, the security and compliance of data transmission and terminal operations are ensured, effectively preventing data leakage, tampering, and unauthorized access, while also ensuring the safety of rescue personnel during operations.

[0025] System adaptability is continuously upgraded: Relying on data cloud accumulation and iterative optimization mechanism, the system can continuously optimize core algorithm parameters through historical data, adapt to different fire types, building structures and meteorological conditions, and its performance steadily improves after long-term use, with a wide range of application scenarios.

[0026] Standardization of emergency response procedures: Construct a closed-loop process of data collection, synchronization, decision-making, execution, and optimization to standardize the operation of all aspects of fire emergency command, reduce reliance on human experience, and improve the standardization and replicability of the response process. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the fire emergency command system based on multi-terminal collaboration involved in this invention.

[0028] Figure 2 This is a flowchart illustrating the fire emergency command system based on multi-terminal collaboration involved in this invention.

[0029] Figure 3 This is a schematic diagram of the algorithm interaction process involved in this invention.

[0030] Figure 4 This is a schematic diagram of the algorithm training and optimization process involved in this invention. Detailed Implementation

[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] refer to Figure 1 and Figure 2 This embodiment provides a fire emergency command system based on multi-terminal collaboration, as detailed below:

[0033] I. System Overview

[0034] In fire emergency command scenarios, traditional systems generally face core pain points such as data dispersion, response delays, insufficient multi-terminal collaboration, and limited decision-making accuracy, which seriously affect fire response efficiency and rescue safety. This system, relying on three core technology foundations—cloud computing, big data services, and internet platforms—constructs an innovative six-module closed-loop architecture: full-domain data perception in fire scenarios, cloud-edge collaborative synchronization, big data intelligent decision-making, dynamic permission-based secure transmission, closed-loop task execution, and cloud-based data accumulation and iteration. This achieves deep collaboration between the fire command center's large screen, commander's PC, rescue personnel's mobile devices, fire rescue equipment terminals, and fire station monitoring terminals.

[0035] The system's core innovation lies in breaking the limitations of the linear process in traditional fire emergency command. Through algorithm combination and a two-way interaction mechanism, it creates a complete closed loop for data collection, synchronization, decision-making, execution, and optimization. All modules are built on a fire emergency internet platform, data storage utilizes a cloud computing distributed cluster, data computation relies on the parallel processing capabilities of big data services, and data transmission is achieved through encrypted internet dedicated lines, ensuring real-time performance, security, and reliability throughout the process. Ultimately, it addresses the core pain points of traditional systems, achieving the goals of accurate fire situation prediction, optimal allocation of rescue resources, efficient multi-terminal collaborative execution, and continuous system iteration and optimization.

[0036] II. System Core Modules and Specific Implementation Details

[0037] (I) Firefighting Scene Full-Domain Data Perception and Cloud Access Module

[0038] 1. Module Function: As the data acquisition entry point of the system, the core function of this module is to comprehensively capture various dynamic and static data in fire emergency scenarios. Through the collaborative mode of edge preprocessing and cloud aggregation, it transforms the scattered raw data into standardized and usable basic data, providing high-quality data support for subsequent cross-terminal synchronization and intelligent decision-making. This is the basic prerequisite for the efficient operation of the entire system.

[0039] 2. Specific Implementation Details: To ensure the comprehensiveness and completeness of data collection, the module covers seven core data sources. Each data type corresponds to key decision-making dimensions for fire emergency response, as shown in the following examples: 1. Fire scene sensor data: temperature, smoke concentration, toxic gas concentration, flame spread rate, and on-site air pressure. This type of data directly reflects the real-time fire situation and is the core basis for fire level determination and spread prediction; 2. Fire rescue equipment status data: real-time location of fire trucks, water tank volume, water pump pressure, ladder extension status, and fire extinguisher inventory. Real-time monitoring of equipment status can avoid resource waste and risks caused by equipment malfunctions during execution; 3. Rescue personnel data: real-time GPS location of rescue personnel, vital signs data (heart rate, blood oxygen saturation, body temperature), and personal protective equipment wearing status. Ensuring the safety of rescue personnel is the primary principle of emergency command. 4. On-site video and image data: Videos from fixed surveillance cameras, drone aerial videos, videos from law enforcement recorders worn by rescue personnel, and on-site images captured by mobile devices. Visualized data can intuitively present the situation at the fire scene and assist commanders in judging the situation; 5. Static data on fire resources: Location of fire stations, distribution of fire hydrants, status of fire lanes, types and quantities of reserve rescue materials. Static data is the basic reference for resource allocation and route planning; 6. Meteorological data: On-site wind speed, wind direction, rainfall, and air humidity. Meteorological conditions directly affect the speed and direction of fire spread and are key factors influencing situation prediction; 7. Basic information data of the jurisdiction: Building structure type, floor height, resident density, and distribution of key protected units. This type of data determines the difficulty of fire handling and rescue priority, and provides a basis for task allocation.

[0040] To address the issues of high noise, inconsistent formats, and low transmission efficiency in raw data, the module adopts a collaborative model of edge node preprocessing + cloud aggregation. Edge nodes are deployed at fire stations or temporary on-site communication vehicles, close to the data acquisition equipment, enabling rapid local processing of raw data. This includes three core operations: first, noise reduction, using a moving average filtering algorithm to remove abnormal fluctuations in sensor signals caused by environmental interference, ensuring data authenticity; second, format standardization, converting heterogeneous data output from different devices (such as binary data from sensors and encoded video data) into JSON format for easier parsing and processing by subsequent modules; and third, outlier removal, filtering data exceeding normal ranges due to sensor malfunctions or equipment abnormalities based on preset reasonable thresholds, preventing erroneous data from influencing decision-making. After preprocessing, the data is uploaded in real-time to a distributed storage cluster on the cloud computing platform via a dedicated internet line. The distributed storage architecture supports parallel writing and reading of massive amounts of data and features data redundancy backup, ensuring zero data loss during transmission and storage, providing data security for the stable operation of subsequent modules. The input to this module is raw data directly acquired by various acquisition devices, without any front-end modules providing data support; the output is preprocessed full-domain fire emergency data stored in the distributed database of the cloud computing platform. This data will serve as the sole data input source for the cross-terminal cloud-edge collaborative synchronization engine module and the fire emergency big data intelligent decision-making service module, directly determining the subsequent synchronization efficiency and decision accuracy.

[0041] (ii) Cross-terminal cloud-edge collaborative synchronization engine module

[0042] 1. Module Function: As the core hub connecting cloud data and edge terminals, the core function of this module is to build a dual-mode synchronous architecture of cloud central control + edge node caching, so as to achieve low latency and high consistency synchronization of pre-processed data among all terminals, break the problem of isolated data and asynchronous updates of each terminal in traditional systems, and ensure that the data obtained by the command end, execution end and monitoring end are completely consistent, laying the foundation for multi-terminal collaborative decision-making and execution.

[0043] 2. Specific Implementation Details: The module supports a comprehensive range of terminal types covering all roles in fire emergency command. Each type of terminal performs specific functions, as shown in the following examples: 1. Fire Command Center Large Screen Terminal: Primarily used for overall situation display, allowing the command team to collectively view information such as fire situation, resource allocation, and task execution progress; 2. Commander PC Terminal: The core operating terminal, used by commanders to formulate decision-making plans, adjust task allocation, and issue instructions; 3. Rescue Personnel Mobile Terminal: On-site execution terminal, used by rescue personnel to receive tasks, provide feedback on execution status, and report on-site conditions; 4. Fire Rescue Equipment Terminal: Equipment control terminal, receiving equipment operation instructions and providing feedback on equipment operating status; 5. Fire Station Monitoring Terminal: Station management terminal, monitoring the status of resources and personnel standby within the station, and receiving material allocation instructions.

[0044] To achieve low-latency, highly reliable multi-device synchronization, the module relies on the fire emergency internet platform and establishes a hybrid transmission channel combining WebSocket long connections and Kafka message queues. This architecture combines the advantages of WebSocket's real-time bidirectional transmission and Kafka's high-concurrency processing, representing an innovative design specifically for fire emergency scenarios. The specific synchronization mechanism is as follows:

[0045] The central control node is deployed on a cloud computing platform, monitoring data updates in the distributed database in real time. When data changes, it categorizes the data according to data type (sensor data, video streams, task commands, status feedback) and distributes it to edge nodes in various regions via a Kafka message queue. Kafka's partitioning mechanism and batch processing capabilities effectively handle data transmission pressure in high-concurrency scenarios, ensuring that data is not lost or duplicated.

[0046] Edge nodes are deployed at various fire stations or temporary on-site communication vehicles, establishing persistent communication links with terminals in the corresponding areas via WebSocket long-lived connections. WebSocket long-lived connections reduce the overhead of establishing and disconnecting connections, enabling real-time data push. At the same time, edge nodes have local caching capabilities. When terminals experience network instability or brief interruptions due to complex on-site environments, edge nodes will temporarily store data distributed from the cloud and resend it in chronological order after the network is restored, ensuring that terminal data is not lost.

[0047] The synchronization triggering mechanism is designed in three categories to comprehensively ensure the consistency of data synchronization: First, data update triggering: when data changes on any terminal or in the cloud, full-terminal synchronization is triggered immediately to ensure that data updates are not delayed; second, timed supplementary synchronization: a full data verification is performed every 100 milliseconds to compare the consistency between the local cached data on each terminal and the cloud data. If there are differences, supplementary synchronization is triggered to avoid missing data synchronization; third, terminal-initiated synchronization request: when a terminal restarts, reconnects to the network, or detects local data anomalies, it actively sends a synchronization request to the edge node to quickly restore data consistency.

[0048] The module takes as input preprocessed data stored in the distributed database of the cloud computing platform by the fire scene full-domain data perception and cloud access module; the output is standardized fire emergency data synchronized to all terminals. After receiving the data, each terminal updates its local cache in real time to ensure that the fire situation, task information and resource status displayed by all terminals at the same time are completely consistent, providing a unified data foundation for subsequent decision-making by commanders and execution of tasks by rescue personnel.

[0049] (III) Fire Emergency Big Data Intelligent Decision Service Module

[0050] 1. Module Function: As the core decision-making hub of the system, this module relies on the big data services of the Internet platform to conduct in-depth analysis of the synchronized full-domain data. Through the bidirectional interaction of two types of innovative algorithms—a graph convolutional long short-term memory network with spatiotemporal attention mechanism enhancement and an improved multi-objective whale optimization algorithm—it achieves accurate prediction of fire situation evolution, optimal allocation of rescue resources, and dynamic planning of rescue routes. It generates complete and executable fire emergency command and decision-making plans, providing core basis for task allocation and directly determining rescue efficiency and safety.

[0051] 2. Specific implementation details:

[0052] (1) Algorithm 1: Graph Convolutional Long Short-Term Memory Network Enhanced with Spatiotemporal Attention Mechanism (STA-GConvLSTM); Traditional fire situation prediction algorithms mostly use a single Long Short-Term Memory Network (LSTM), which can only capture the temporal features of the data and ignore the spatial correlation characteristics in the fire spread process, resulting in limited prediction accuracy, especially with large errors under complex building structures and meteorological conditions. The STA-GConvLSTM algorithm breaks through this limitation and innovatively integrates the spatiotemporal attention mechanism and graph convolutional network to construct a four-layer architecture of spatial attention layer - graph convolutional layer - temporal attention layer - LSTM layer. It can accurately capture the spatial correlation of fire spread (such as the fire transmission relationship in different areas) and effectively mine the temporal evolution pattern (such as the trend of fire change over time).

[0053] The core role of this algorithm in this scheme is to achieve accurate prediction of the fire situation at different times in the future, output key information such as the fire spread range and the spread path of dangerous areas, provide direct data support for the allocation of rescue resources and route planning, and avoid problems such as rescue teams entering dangerous areas or unreasonable resource allocation.

[0054] Model construction process: The overall output formula of the STA-GConvLSTM model is:

[0055] ;

[0056] in, The fire situation data for the future time period T, represented by the model output, includes two core dimensions: the fire spread range and the spread path of the danger zone. To meet the decision-making needs of different rescue stages, the value of T is set to 10 minutes, 30 minutes, and 60 minutes, corresponding to the three stages of emergency response, comprehensive rescue, and long-term prevention and control, respectively. Let be the spatiotemporal feature matrix input at time t, with dimension . N is the number of samples, C is the number of feature channels, covering all features related to fire spread such as temperature, smoke concentration, wind speed and direction, and building structure. H and W are the height and width of the spatial grid, respectively, which divides the fire area into grids with a certain precision to facilitate accurate description of spatial features. Let be a spatial adjacency matrix with dimension . , This represents the number of nodes in the spatial grid. This indicates that grid i and grid j have a spatial relationship (such as adjacent regions). Indicates no correlation; this matrix is ​​used to capture the spatial transmission relationships of fire spread. Here is the spatial attention layer weight matrix, with dimension 1. This is used to adjust the importance of different spatial features; This is the weight matrix for the time attention layer, with dimension 1. , The time step is used to enhance the features of key time nodes. The weight matrix of the graph convolutional layer has dimensions of . , used to extract spatial correlation features; Here is the LSTM layer weight matrix, with dimension 1. It is used to explore patterns in the evolution of time.

[0057] The specific calculation processes of each layer of the model are progressive, forming a complete feature extraction and prediction process:

[0058] The core function of the spatial attention layer is to highlight the spatial features that have a significant impact on fire spread. The spatial attention weights are calculated first. :

[0059] ; The function maps weight values ​​to the interval between 0 and 1, ensuring that the sum of the attention weights for each spatial node is 1, which facilitates the quantification of the importance of different spatial regions. Then, spatial attention weighting is achieved through Hadamard product (element-wise multiplication), resulting in the spatial attention-weighted feature matrix. : This process can effectively enhance the characteristics of key areas (such as the fire core area and flammable areas) and weaken the interference from irrelevant areas.

[0060] Graph convolutional layers are based on spatial adjacency matrices. The feature matrix after spatial attention weighting Perform spatial correlation extraction and output. : The optimized spatial adjacency matrix , It is an identity matrix used to preserve the characteristics of each grid node, avoiding focusing only on neighboring nodes while ignoring its own information; for The degree matrix, This is used to normalize the spatial adjacency matrix and avoid feature bias caused by differences in node degree. The ReLU activation function is given by the formula: By introducing nonlinear features, the model's ability to fit complex spatial relationships is improved.

[0061] The role of the time attention layer is to focus on the time-specific features that play a crucial role in changes in the fire situation and to calculate the time attention weights. : ;

[0062] Subsequently, temporal attention weighting is achieved through Hadamard product to obtain the temporally attention-weighted feature matrix. : ;

[0063] This process can effectively capture key time points in the fire spread process (such as the moment when the fire suddenly changes), and improve the model's accuracy in mining the laws of time evolution.

[0064] The LSTM layer, as the core of the model's temporal prediction, mines the long-term temporal dependencies of the feature matrix through the synergistic effect of the input gate, forget gate, output gate, and cell state. The calculation formulas for each key parameter are as follows:

[0065] ; ;

[0066] ; ;

[0067] ; ;

[0068] in Here are the weight matrices for each gate in the LSTM input layer, all with dimensions [0, 1]. ; Here are the weight matrices for each gate in the LSTM hidden layer, with dimensions 1. ; These are the bias terms for each major category, with all dimensions being... ; This is the hidden layer output at time t-1, used to convey historical time-series features; The candidate cell state at time t is used to store the temporal information of the current time. The hyperbolic tangent activation function is given by the formula: , which is a numerical range used to control the cell state.

[0069] Finally, the output of the STA-GConvLSTM model Output from hidden layer Obtained through output layer mapping: ;in The output layer weight matrix has dimensions of . (2 corresponds to the two output dimensions of spread range and spread path). The output layer bias term has a dimension of This enables the mapping from the feature matrix to fire situation prediction results.

[0070] Model training process: The core objective of model training is to optimize the weight matrices and bias terms using historical data to improve prediction accuracy. The specific steps are as follows:

[0071] 1. Dataset Construction: Collect historical data on fire emergency response over the past 5 years, covering fire situation data under different fire types (such as building fires, forest fires, and chemical fires), different meteorological conditions (such as sunny days, rainy days, and windy days), and different building structures (such as brick-concrete structures, steel structures, and wood structures). Divide the data into training set, validation set, and test set in a ratio of 8:1:1. The training set is used for model parameter learning, the validation set is used to monitor the training effect and prevent overfitting, and the test set is used to finally evaluate the model accuracy.

[0072] 2. Initialize parameters: Randomly initialize Equal weight matrices, initialize all bias terms to a zero vector, and set the learning rate. (Balancing training speed and convergence accuracy), number of iterations (Ensure the model fully learns the data features);

[0073] 3. Forward Propagation: The training data is input into the model in batches, passing sequentially through a spatial attention layer, a graph convolutional layer, a temporal attention layer, and an LSTM layer. The output of each layer is calculated according to the formulas described above, ultimately yielding the predicted fire situation data. ;

[0074] 4. Loss function calculation: The mean squared error loss function is used. Evaluation of predicted values Compared with the true value The error is calculated using the following formula:

[0075] Where M is the number of training samples, this loss function can effectively quantify the deviation between the predicted value and the true value, and provide direction for parameter optimization;

[0076] 5. Backpropagation: The Adam optimization algorithm is used to calculate the gradient of the loss function with respect to each weight matrix and bias term. The parameters are then updated using gradient descent. The update formula is as follows:

[0077] ;in This represents the weights or biases that need to be updated. This represents the current iteration number. For first-order moment estimation, , For loss function pairs gradient, ; For second-order moment estimation, , ; To prevent the minimum value where the denominator is 0, the Adam optimization algorithm can adaptively adjust the learning rate, thereby improving training stability and convergence speed.

[0078] 6. Validation and early stopping: Every 50 iterations, input the validation set data into the model and calculate the validation set loss. If the validation set loss does not decrease for 100 consecutive iterations (error fluctuation is less than 0.0001), stop training to avoid model overfitting and save the current optimal model parameters.

[0079] Model application process: After the model training is completed, it receives the real-time spatiotemporal feature matrix synchronized by the cross-terminal cloud-edge collaborative synchronization engine module. spatial adjacency matrix Following the computational flow of the model construction, the data sequentially passes through a spatial attention layer, a graph convolutional layer, a temporal attention layer, and an LSTM layer, ultimately outputting fire situation prediction data for the next 10, 30, and 60 minutes. These data will serve as direct input parameters for the improved multi-objective whale optimization algorithm, used to refine the objective functions for rescue resource allocation and route planning, providing core support for accurate decision-making.

[0080] (2) Algorithm 2: Improved Multi-Objective Whale Optimization Algorithm (IMOWOA); Traditional rescue resource allocation and route planning algorithms often adopt single-objective optimization (such as optimizing only response time), ignoring key factors such as resource consumption and rescue risk, resulting in insufficient practicality of decision-making schemes. The IMOWOA algorithm breaks through this limitation. Based on the traditional whale optimization algorithm, it introduces adaptive weights and chaotic initialization strategies, and simultaneously optimizes the three core objectives of rescue response time, rescue resource consumption, and rescue risk, to achieve multi-objective collaborative optimization.

[0081] The core function of this algorithm in this scheme is to generate an optimal allocation list of rescue resources and a detailed list of rescue routes based on the fire situation prediction data output by the STA-GConvLSTM algorithm, combined with real-time resource status and geographic information. This ensures that rescue teams can reach the scene and carry out the response with the fastest speed, the lowest consumption, and the least risk, thereby improving rescue efficiency and safety.

[0082] Model construction process: The core of the IMOWOA algorithm is to construct a multi-objective optimization objective function and find the Pareto optimal solution through iterative optimization. The overall optimization objective function is defined as follows:

[0083] ;in The solution vector for rescue resource allocation and route planning has the following dimensions. K represents the number of rescue resource groups. Each solution vector contains three core dimensions: rescue team number, rescue route node sequence, and resource allocation quantity, comprehensively covering key information on resource allocation and route planning. Let be the objective function for rescue response time, representing the total time from the departure of resources to the fire scene, with the objective of minimizing the response time; Let be the objective function for rescue resource consumption, representing the total consumption of human, material, and financial resources during the rescue process, with the goal of minimizing resource consumption; Let be the objective function for rescue risk, representing the probability of risks such as fire and toxic gases faced by rescuers, with the goal of minimizing rescue risk.

[0084] The specific calculation formulas for each objective function fully integrate the output data of the STA-GConvLSTM algorithm and real-time scene information, ensuring that the optimization results closely match the actual situation:

[0085] 1. Objective function for rescue response time: ;in Let be the number of route nodes for the k-th resource group, where each node corresponds to a key location in the route (such as a fire station, a road intersection, or a fire scene). The initial distance of the k-th resource group through the n-th route node is given by the fire situation data output by the STA-GConvLSTM model to address the impact of fire spread on the route. The spread range in the middle is corrected, and the correction formula is as follows: , For correction factor, Let T represent the fire spread area at the nth node in the future time period T. This area is the largest in the region, and adjustments can prevent rescue routes from being blocked by the spread of the fire. The travel speed of the k-th resource passing through the n-th route node is affected by the status of the fire lane (such as whether it is congested or damaged) and the fire situation (such as whether the flames have spread to the route), and is obtained in real time from the cross-terminal cloud-edge collaborative synchronization engine module. The loading time for the kth group of resources is determined based on the resource type (such as fire extinguishing equipment or search and rescue equipment). The waiting time for the resources in group k, such as waiting for other teams to assemble or for traffic control to be lifted; The time weight of the k-th group of resources is determined by the urgency of the fire situation. The more urgent the fire, the greater the weight, ensuring priority scheduling of critical resources.

[0086] 2. Objective function for rescue resource consumption: ;in The number of personnel to be deployed for the kth group of resources is determined based on the scale of the fire and the needs of the response. Cost per unit of time for a single person, including labor costs, wear and tear of protective equipment, etc. The number of equipment (such as fire trucks, ladders, and water pumps) deployed for the k-th group of resources. Cost per unit of time for a single piece of equipment, including fuel consumption, equipment depreciation, etc. The quantity of supplies (such as fire extinguishers, water hoses, and first aid medicines) invested in the k-th resource group. Cost per unit of material consumption; The consumption weight of the k-th resource group is determined by the scarcity of the resource. The more scarce the resource, the higher the weight, thus avoiding excessive consumption.

[0087] 3. Rescue risk objective function: ;in The fire situation data output by the STA-GConvLSTM model represents the fire risk probability of the k-th resource group. The formula for calculating the flame spread rate is as follows: , For risk coefficient, Let be the flame spread speed in the k-th resource operation area. The faster the flame spread speed, the higher the risk probability. The probability of the k-th resource group facing the risk of toxic gas is calculated from on-site sensor data; The probability of the kth group of resources facing the risk of building collapse is determined by combining the building structure type and the duration of the fire. The risk weight of the k-th group of resources is determined by the degree of danger of the work area. The higher the degree of danger, the greater the weight, so as to ensure the safety of rescue personnel is given priority.

[0088] The IMOWOA algorithm achieves multi-objective optimization by iteratively updating the solution vector. The core iterative process includes three stages: prey encirclement, bubble net attack, and adaptive weight adjustment.

[0089] 1. Encirclement Phase: After algorithm initialization, the search range is narrowed down using the following formula to approximate the optimal solution: ; ;in This represents the current iteration number; Let t be the optimal solution vector at time t, which is the optimal resource allocation and route planning scheme found so far. , Decrease linearly from 2 to 0. A random number between 0 and 1. The range of values ​​for determines the search step size. As the number of iterations increases, the step size gradually decreases, achieving precise search. , A random number between 0 and 1. Random weights are used to enhance the algorithm's global search capability.

[0090] 2. Bubble Web Attack Phase (Shrinking Encirclement): The solution vector is further optimized by simulating the bubble web attack behavior of whales using a logarithmic spiral function. ; ; ;in The maximum number of iterations is set to 200; This is the logarithmic spiral shape constant, with a value of 1, which controls the shape of the spiral. The value is a random number between -1 and 1, which ensures that the algorithm performs a local search near the optimal solution, thus improving the optimization accuracy.

[0091] 3. Adaptive Weight Adjustment: To achieve synergistic optimization of the three objectives, adaptive weights are introduced. (i=1,2,3 correspond to three objectives), dynamically adjust the importance of each objective, using the following formula:

[0092] ;in The initial weights are set to [0.4, 0.3, 0.3] based on experience in fire emergency command (response time has the highest weight, followed by risk and consumption). The maximum value of the i-th objective function is obtained through historical data statistics. This formula can dynamically adjust the weights according to the objective function value of the current iteration. When an objective function value is close to the maximum value, the corresponding weight will increase to ensure that the objective is fully optimized and achieve multi-objective collaboration.

[0093] 4. Chaotic Initialization: To improve the diversity of the initial population and avoid the algorithm getting trapped in local optima, the initial population is generated using a Logistic chaotic mapping, with the following formula: ;in Here are the chaos coefficients. As chaotic variables, the initial population generated through chaotic mapping has a more uniform distribution, can cover a wider search space, and improves the algorithm's global optimization capability.

[0094] Model training process: The core objective of model training is to find the Pareto optimal solution for multi-objective collaboration through iterative optimization. The specific steps are as follows:

[0095] 1. Population Initialization: An initial solution vector population is generated using a chaotic initialization strategy. The population size is... To ensure sufficient diversity in the population; the solution vector dimension is determined based on the number of rescue resource groups K, which is dynamically adjusted by the fire scale and the amount of available resources;

[0096] 2. Objective Function Normalization: Since the three objective functions have different dimensions (time, cost, probability), direct optimization will lead to biased results. Therefore, the values ​​of the three objective functions are normalized to the interval between 0 and 1. The normalization formula is as follows: ,in Let be the minimum value of the i-th objective function. To achieve the maximum value, normalization is used to eliminate dimensional differences and ensure that the three objectives are equally important;

[0097] 3. Iterative optimization: Set the maximum number of iterations. The solution vector and adaptive weights are updated in each iteration according to the formulas for surrounding the prey, bubble net attack, and adaptive weight adjustment, gradually approaching the optimal solution.

[0098] 4. Pareto Optimal Solution Selection: The non-dominated sorting method is used to select Pareto optimal solutions. The core of non-dominated sorting is to stratify the solution vectors according to their dominance relationship. The dominance relationship is defined as follows: if the three objective function values ​​of solution vector X1 are all no greater than those of solution vector X2, and at least one objective function value is less than that of X2, then X1 dominates X2. By sorting, solution vectors that are not dominated by other solution vectors are selected to form the Pareto optimal solution set.

[0099] 5. Determining the Optimal Solution: Combining the experience and preferences of fire commanders, the final solution vector for rescue resource allocation and route planning is selected from the Pareto optimal solution set. The empirical preference mainly considers factors such as the urgency of the fire, the scarcity of resources, and the risk tolerance of rescue efforts to ensure that the optimal solution is practically feasible.

[0100] Model application process: When applying the model, first receive the fire situation prediction data output by the STA-GConvLSTM model. This includes information such as the fire spread range and flame spread speed at different future time periods; it also receives real-time information such as fire resource location data, fire lane status data, and traffic data synchronized by the cross-terminal cloud-edge collaborative synchronization engine module. This data is then input into the IMOWOA algorithm, which sequentially completes population initialization, objective function normalization, iterative optimization, and Pareto optimal solution selection, ultimately obtaining the optimal allocation list of rescue resources and detailed rescue routes.

[0101] This algorithm interacts bidirectionally with the STA-GConvLSTM algorithm: the output data of STA-GConvLSTM directly corrects the objective function parameters of IMOWA (such as route distance correction in rescue response time and fire risk probability calculation in rescue risk), ensuring that the optimization results of IMOWA closely reflect changes in the fire situation; simultaneously, the optimization results of IMOWA are fed back to the verification stage of STA-GConvLSTM, specifically judging the prediction accuracy through error calculation. The interaction formula is as follows:

[0102] ;in For prediction error, Fire situation data predicted by STA-GConvLSTM. This refers to the actual fire situation data obtained from on-site feedback by the rescue team after the IMOWOA algorithm is executed. For error weights; when When the error exceeds the threshold, it indicates that the prediction error of STA-GConvLSTM exceeds the threshold, triggering the model to be retrained, thereby improving the subsequent prediction accuracy and achieving synergistic optimization of the two types of algorithms.

[0103] (3) Decision-making scheme generation; Big data services perform real-time computation of two algorithms, STA-GConvLSTM and IMOWOA, and integrate the output results of the two algorithms to generate a complete decision-making scheme containing six core contents. All contents are necessary information for fire emergency command. Examples are as follows: 1. Core fire information: precise coordinates (latitude and longitude) of the fire location, fire type (building fire / forest fire / chemical fire, etc.), current fire level (level 1 / 2 / 3 / 4), and expected spread range (in the next 10 / 30 / 60 minutes); 2. Rescue resource allocation list: number and person in charge of each rescue team, number of the equipped fire truck, type and quantity of rescue materials carried, and departure fire station; 3. Division of rescue tasks: fire-fighting team tasks (fire-fighting area, equipment used, 4. Firefighting sequence; 5. Search and rescue team tasks (search and rescue area, search and rescue priority, rescue method); 6. Evacuation team tasks (evacuation area, evacuation route, evacuation responsible person); 7. Logistics support team tasks (material supply, equipment maintenance, medical support); 8. Detailed rescue route: specific driving route of each fire truck (passing nodes, turning instructions), estimated arrival time, and backup route (to prevent the main route from being blocked); 9. Safety warning information: range of dangerous areas (accurate to the grid), residential areas that need to be evacuated, and warning of toxic gas diffusion (diffusion direction, diffusion speed, protection suggestions); 10. Emergency linkage instructions: specific cooperation content of medical emergency departments (emergency point setting, ambulance dispatch), traffic control departments (controlled road sections, control time), and power departments (power outage areas, power outage time).

[0104] The module takes into account standardized fire emergency data synchronized by the cross-terminal cloud-edge collaborative synchronization engine module, as well as external related data (such as real-time traffic data from the transportation department and emergency medical resource data from the medical department) connected through the fire emergency internet platform. The output is a complete fire emergency command and decision-making plan, which is directly transmitted to the fire task closed-loop collaborative execution module and simultaneously synchronized to all terminals through the cross-terminal cloud-edge collaborative synchronization engine module for commanders to view and confirm, ensuring the consistency and executability of the decision-making plan.

[0105] (iv) Multi-terminal dynamic permission adaptation and secure transmission module

[0106] 1. Module Functions: The core function of this module is to realize dynamic allocation of permissions for multiple terminals and ensure data transmission security. On the one hand, through fine-grained role permission management, it ensures that terminals with different roles can only obtain and operate data within the corresponding permission scope, avoiding data leakage and misoperation. On the other hand, through multiple encryption and protection technologies, it ensures the security and integrity of data during transmission, providing a safe and reliable transmission environment for multi-terminal collaboration.

[0107] 2. Specific Implementation Details: To meet the operational needs of different roles in fire emergency command while ensuring data security, the module defines six core roles. The permissions of each role strictly match its responsibilities, as shown in the examples below: 1. Fire Chief Commander: Possesses the highest permissions, able to view all terminal data (including fire situation, resource status, task execution progress, personnel vital signs, etc.), modify decision-making plans, issue global instructions (such as adjusting rescue strategies, dispatching cross-regional resources), and approve permission changes (such as temporarily elevating the permissions of on-site commanders); 2. On-site Commander: Permissions cover the jurisdiction area, able to view all data within the jurisdiction, adjust local task allocation (such as adjusting the task assignments of rescue teams within the area), and provide feedback on the on-site situation (such as reporting sudden changes in the fire); 3. Rescue Team Leader: Permissions focus on their own team, able to view their own team's... 4. Rescue Team Members: Permissions are limited to individuals. They can view personal task data (such as the rescue area they are responsible for and the task requirements), report personal status (such as vital signs and task completion status), and report on-site conditions (such as the discovery of trapped personnel and changes in fire intensity). 5. Logistics Support Personnel: Permissions are limited to logistics resources. They can view rescue material inventory data, receive material allocation instructions, and report material supply status (such as material delivery status and inventory replenishment progress). 6. System Administrators: Permissions are limited to system maintenance. They are only responsible for system maintenance (such as server repair and software updates) and permission configuration (such as adding roles and adjusting permission scope). They cannot view any emergency command data to ensure data isolation.

[0108] The permission adaptation process is designed as a full-process control system encompassing identity authentication, permission allocation, and real-time verification: When a terminal logs in, it must undergo dual authentication using facial recognition and a username / password. Facial recognition is used to verify the authenticity of the operator's identity, while the username / password is used for system identity verification. This dual authentication effectively prevents unauthorized logins. After successful authentication, the cloud computing platform automatically assigns corresponding permissions to the terminal based on a pre-configured role and permission library and synchronizes the permission list to the terminal's local machine and edge nodes. Every data request from the terminal (such as viewing data, issuing instructions, or providing feedback) must undergo permission verification. Both edge nodes and the cloud central control node perform dual verification to ensure permission matching. Only operations that pass verification can be executed; otherwise, the request is rejected and logged, ensuring comprehensive permission control.

[0109] Data transmission security employs a triple protection strategy to comprehensively safeguard data security: First, encrypted transmission, using SSL / TLS 1.3 encryption protocol throughout the entire process, and AES-256 encryption algorithm to encrypt data before transmission. The encryption key is distributed through an asymmetric encryption algorithm to ensure that the key is not leaked. Second, access protection, with firewalls and intrusion detection systems deployed on the cloud computing platform and edge nodes. The firewall blocks unauthorized IP access, and the intrusion detection system monitors data transmission behavior in real time, immediately blocking and alerting if abnormal access is detected (such as frequent requests or data theft). Third, integrity verification, using the SHA-256 hash verification mechanism. The sending end calculates the hash value of the data and transmits it along with the data. The receiving end recalculates the hash value after receiving the data. If the two hash values ​​match, it means that the data has not been tampered with; if they do not match, the data is rejected and a retransmission is requested to ensure data integrity.

[0110] The module's inputs include data synchronized by the cross-terminal cloud-edge collaborative synchronization engine module, decision-making schemes generated by the fire emergency big data intelligent decision-making service module, and operation requests and authentication information from each terminal. The outputs are directed data after permission filtering (each terminal only receives data within its own permission scope), encrypted transmission data, and permission verification results (pass / reject). Directed data is transmitted to the fire task closed-loop collaborative execution module and each terminal. Encrypted transmission ensures security throughout the data flow process, and permission verification results are used to control whether terminal operations are executed.

[0111] (v) Closed-loop collaborative execution module for fire protection tasks

[0112] 1. Module Function: As the core of the system's execution, the core function of this module is to build a closed-loop collaborative process of decision-making, distribution, execution, feedback, and adjustment. It distributes the decision-making plans generated by the fire emergency big data intelligent decision-making service module to the corresponding terminals according to roles, collects task execution data and feedback information in real time, and triggers dynamic adjustments to the decision-making plans to ensure efficient multi-terminal collaborative execution. This solves the problems of disconnect between task distribution and execution and untimely adjustments in traditional systems.

[0113] 2. Specific Implementation Details: The task distribution mechanism is based on the dynamic adaptation of multi-terminal permissions and the permission filtering results of the secure transmission module. The task dispatch center of the cloud computing platform distributes tasks according to roles, ensuring that each terminal only receives task information related to its own responsibilities: 1. Fire Command Center Large Screen Terminal: Receives the complete decision-making plan and displays the task allocation status (such as task division of each team, resource allocation list) and the overall execution progress (such as the completion percentage of each task and the location of rescue teams) in real time for the command team to monitor in real time; 2. Commander's PC Terminal: Receives the complete decision-making plan and supports manual fine-tuning of local tasks (such as adjusting the task area of ​​rescue teams and supplementary material allocation instructions). The fine-tuned task allocation... The case needs to be verified by permissions and then reissued to the relevant terminals; 3. Mobile terminal of the rescue team leader: Receives the team's task list (such as the rescue area and task objectives) and rescue route (such as driving route and estimated arrival time) to facilitate the overall management of the team's execution; 4. Mobile terminal of the rescue team members: Receives specific individual tasks, such as going to the 5th floor of Building 3 to search for trapped people, carrying demolition tools and first aid medicines. The task information is clear and specific to ensure that the rescue team members execute it accurately; 5. Fire rescue equipment terminal: Receives equipment operation instructions, such as turning on the water pump of the fire truck, adjusting the pressure to 0.8MPa, and extending the ladder to a height of 10 meters. The instructions are standardized to facilitate the precise execution of the equipment.

[0114] The task execution and feedback process is designed as a real-time data collection, encrypted transmission, and full-terminal synchronization workflow: During task execution, each terminal collects execution data and feedback information in real time through various methods, including three categories: automatic terminal collection (e.g., equipment terminals automatically collect water pump pressure, rescue personnel's mobile terminals automatically collect vital signs), manual input by rescue personnel (e.g., rescue team members manually report that the search and rescue of the 5th floor of Building 3 has been completed and no trapped personnel have been found), and automatic recording and transmission from video equipment (e.g., law enforcement recorders record and upload on-site video in real time). The feedback data comprehensively covers the execution status, as shown in the following examples: 1. Rescue team execution progress: percentage of completed tasks, remaining tasks, and estimated completion time; 2. 3. On-site emergencies: Newly discovered fire hazards (such as fires in adjacent buildings), unexpected events (such as partial building collapses, injuries to rescue personnel); 4. Individual task completion status: The execution status of the individual's assigned tasks (not started / in progress / completed), and the quality of completion (such as whether the search and rescue was thorough, and whether the fire extinguishing was effective); 5. Individual vital signs data: Heart rate, blood oxygen saturation, and body temperature of rescue personnel, used to monitor personnel safety; 6. Real-time equipment status: Water volume in fire truck tanks, water pump pressure, ladder extension and retraction status, fire extinguisher inventory, etc., used to determine whether the equipment can continue to work; 7. On-site video transmission: Real-time display of the fire scene situation and the rescue execution process to assist commanders in assessing the situation.

[0115] Feedback data is transmitted back to the cloud computing platform through the encrypted transmission channel of the multi-terminal permission dynamic adaptation and secure transmission module. After receiving the data, the cloud central control node immediately synchronizes it to all authorized terminals through the cross-terminal cloud-edge collaborative synchronization engine module, ensuring that the commander can grasp the execution status in real time and that the rescue personnel can understand the overall progress.

[0116] The dynamic task adjustment mechanism ensures that decision-making plans can quickly adapt to changes on-site: The cloud computing platform receives feedback data in real time and inputs it along with the original decision data into the fire emergency big data intelligent decision-making service module. The module recalculates two types of algorithms, STA-GConvLSTM and IMOWOA, to generate dynamically adjusted task plans. For example, if feedback data shows that the fire spread speed in Building 3 exceeds the prediction and the existing fire-fighting resources are insufficient, the algorithm will recalculate the fire situation, adjust the resource allocation plan, and dispatch rescue teams from nearby fire stations to provide support. After the adjusted plan is filtered by the multi-terminal permission dynamic adaptation and secure transmission module, it is sent to the corresponding terminal to ensure that the adjustment instructions are executed quickly.

[0117] The module takes as input the fire emergency command decision-making plan after permission filtering, task execution data and feedback data from each terminal; and outputs task execution status data, feedback data, and dynamically adjusted task plans. The task execution status data and feedback data are transmitted to the cross-terminal cloud-edge collaborative synchronization engine module (for full-terminal synchronization) and the emergency data cloud accumulation and iterative optimization module (for system iteration). The dynamically adjusted task plans are transmitted to the relevant terminals to ensure that task execution matches the on-site situation in real time.

[0118] (vi) Emergency Data Cloud Accumulation and Iterative Optimization Module

[0119] 1. Module Function: As the core of the system's iteration, the core function of this module is to rely on the fire emergency internet platform and big data services to accumulate the system's entire process data in the cloud. Through the bidirectional interaction of two types of innovative algorithms—adaptive weighted multi-granularity feature fusion algorithm and dual feedback reinforcement learning algorithm—it can mine the value of data, optimize the core algorithm parameters of the fire emergency big data intelligent decision service module, and achieve continuous iterative upgrades of system performance, making the system more and more accurate with the number of uses.

[0120] 2. Specific implementation details:

[0121] (1) Algorithm 1: Adaptive Weighted Multi-Granularity Feature Fusion Algorithm (AWMFF); Traditional data feature extraction algorithms often use single-granularity feature extraction (such as extracting only fine-grained detail features or coarse-grained trend features), resulting in incomplete feature information and affecting subsequent optimization effects. The AWMFF algorithm breaks through this limitation and innovatively integrates three types of features: fine-grained, medium-grained, and coarse-grained. Through adaptive weight allocation, it highlights the role of important features and provides accurate and comprehensive feature support for dual-feedback reinforcement learning algorithms.

[0122] The core function of this algorithm in this scheme is to extract comprehensive and effective features from massive historical data, avoid information omissions caused by single-granularity features, provide high-quality feature input for parameter optimization of the core decision-making algorithm, ensure accurate optimization direction, and improve the system iteration effect.

[0123] Model building process: The core of the AWMFF algorithm is to assign adaptive weights to features of different granularities, thereby achieving effective fusion of multi-granularity features. The overall feature fusion formula is as follows:

[0124] ;in The fused feature matrix has a dimension of . D is the feature dimension, which covers all features related to decision optimization, such as fire situation, resource allocation, task execution, and rescue effect. For particle size type, g=1 represents fine particle size, g=2 represents medium particle size, and g=3 represents coarse particle size; For the adaptive weights of the g-th class granularity, satisfying The weight values ​​are dynamically adjusted based on the importance of the features; Let g be the feature matrix of the g-th granularity, with dimension g. Each type of granularity feature corresponds to information at different levels.

[0125] The extraction process for features at each granularity specifically mines information at different levels, progressively forming a complete feature system:

[0126] 1. Fine-grained characteristics Focusing on the raw, detailed features of the data, which reflect micro-level changes such as real-time fluctuations in fire temperature, every movement of rescue personnel, and subtle changes in equipment status, the extraction formula is as follows: ,in It is a one-dimensional convolution function used to extract local detail features. The kernel size is [size]. Step size, This is the original historical data matrix, containing the original data for the entire system process;

[0127] 2. Medium-grained characteristics Focusing on the local statistical characteristics of the data, statistical summarization is performed based on fine-grained features to reflect local patterns in the data, such as the average fire temperature within a 10-minute period, the time a rescue team spends in a certain area, and the average operating pressure of certain equipment. The extraction formula is as follows: ,in This is a one-dimensional pooling function used for downsampling and statistical analysis of fine-grained features. The pooling kernel size, Step size;

[0128] 3. Coarse-grained characteristics Focusing on the global trend characteristics of the data, a global analysis is conducted based on medium-granularity features to reflect the overall changing patterns of the data. Examples include the changing trends of the fire situation, the overall consumption of rescue resources, and the changing trends of rescue efficiency throughout the rescue process. The extraction formula is as follows: ,in Principal component analysis (PCA) functions are used for dimensionality reduction and extraction of global principal features. Using the principal component dimension, we ensure that core trends are preserved while reducing computational complexity.

[0129] Adaptive weights The calculation is based on the signal-to-noise ratio (SNR) of each granularity feature. The higher the SNR, the more effective information and the less noise the feature has, and the greater its weight. The calculation formula is as follows:

[0130] ;in The signal-to-noise ratio of the g-th granularity feature is calculated using the following formula: , for The mean, for The standard deviation of the signal-to-noise ratio can effectively quantify the quality of features at each granularity, ensuring reasonable weight allocation.

[0131] Model training process: The core objective of model training is to optimize adaptive weights and improve the quality of fused features. The specific steps are as follows:

[0132] 1. Dataset Construction: Collect historical execution data (such as task completion time, rescue effect, and resource consumption) from the fire-fighting task closed-loop collaborative execution module and historical decision data (such as initial decision plan and dynamic adjustment plan) from the fire emergency big data intelligent decision-making service module. Divide the data into training set and test set in a 9:1 ratio. The training set is used for weight optimization, and the test set is used to verify the fusion effect.

[0133] 2. Feature Extraction: Fine-grained, medium-grained, and coarse-grained feature extraction is performed sequentially on the training set data to obtain feature matrices of three granularities. ;

[0134] 3. Weight Optimization: Aiming at the classification accuracy of the fused features, the adaptive weights are optimized using gradient descent. Classification accuracy reflects the predictive ability of the fused features on rescue effectiveness (e.g., success rate of rescue, resource consumption control rate). The optimization formula is as follows: ,in To improve classification accuracy, The learning rate is used to iteratively adjust the weights and improve the classification accuracy of the fused features.

[0135] 4. Validation: Input the test set data into the model, extract the three types of granular features and fuse them according to the optimized weights, calculate the classification accuracy of the fused features. If the accuracy reaches 90% or more, it means that the fusion effect meets the requirements; otherwise, continue to adjust the weights until the target is met.

[0136] Model Application Process: During model application, the system receives task execution data (such as task completion progress, on-site emergencies, and equipment status) from the fire-fighting task closed-loop collaborative execution module and decision data (such as initial decision plans and dynamic adjustment plans) from the fire emergency big data intelligent decision-making service module. This data is then integrated into the original historical data matrix. Subsequently, fine-grained features were extracted sequentially. Medium-grained characteristics coarse-grained characteristics Calculate the signal-to-noise ratio and adaptive weights for various features. Finally, the fused feature matrix is ​​obtained according to the feature fusion formula. This matrix will serve as the state input for the dual-feedback reinforcement learning algorithm, providing precise feature support for the parameter optimization of the core decision-making algorithm.

[0137] (2) Algorithm 2: Dual Feedback Reinforcement Learning Algorithm (DFRL); Traditional reinforcement learning algorithms often use single feedback signals (such as feedback based solely on decision results), resulting in a single optimization direction and insufficient precision in parameter adjustment. The DFRL algorithm breaks through this limitation by innovatively integrating decision feedback and execution feedback signals to construct a more comprehensive reward function and achieve dynamic iterative optimization of the core decision algorithm parameters.

[0138] The core function of this algorithm in this solution is to dynamically optimize the STA-GConvLSTM and IMOWOA algorithm parameters of the fire emergency big data intelligent decision service module based on the fusion features extracted by the AWMFF algorithm. Through the synergistic effect of the two types of feedback signals, it ensures that the parameter adjustment fits the actual application scenario and continuously improves the accuracy and practicality of the decision algorithm.

[0139] Model building process: The core of the DFRL algorithm is to construct a reward function that includes decision feedback and execution feedback. Through interactive iteration of reinforcement learning, the parameters of the target algorithm are optimized. The overall optimization formula is:

[0140] ;in The algorithm parameters to be optimized cover all weight matrices of STA-GConvLSTM. (etc.) and the adaptive weights of IMOWOA (etc.) ); The learning rate is set to 0.002 to balance optimization speed and parameter stability. The reward function is determined by the decision feedback. and performance feedback rewards The components are used to evaluate the effectiveness of parameter adjustments; The state at time t is the fused feature matrix output by the AWMFF algorithm. This indicates that it comprehensively reflects the characteristic information of historical data; The action at time t is the algorithm parameter adjustment amount, used to correct the target algorithm parameters; The state at time t+1 is the algorithm output after parameter adjustment (such as adjusted fire situation prediction data and resource allocation plan), which is used to evaluate the effectiveness of the action.

[0141] The reward function is the core of the DFRL algorithm. By fusing two types of feedback signals, it comprehensively evaluates the effect of parameter adjustment. The specific calculation formula is as follows:

[0142] ;in For decision feedback weights, The implementation feedback weight is higher because the implementation effect directly reflects the actual value of the decision-making solution; The decision feedback reward is used to evaluate the theoretical rationality of the decision-making scheme, and the calculation formula is as follows: , The decision results predicted by the algorithm after parameter adjustment (such as fire situation prediction data and resource allocation plan). The optimal decision outcome as evaluated by experts is assessed by comparing the theoretical decision outcome with the expert's optimal outcome to evaluate the rationality of the decision. To implement feedback rewards, used to evaluate the actual effectiveness of the decision-making plan, the calculation formula is as follows: , For actual rescue time, This represents the planned rescue time output by the algorithm after parameter adjustments. A smaller ratio indicates higher rescue efficiency. The reward is based on the actual probability of rescue risks (such as the probability of rescuers being injured or equipment being damaged), and it can simultaneously take into account both rescue efficiency and safety.

[0143] The state transition formula of the DFRL algorithm describes the change of state after parameter adjustment. The formula is as follows:

[0144] ;in The core algorithm for the fire emergency big data intelligent decision-making service module is either STA-GConvLSTM or IMOWOA; this formula represents the state at time t. Next, perform parameter adjustment actions. Substituting the adjusted parameters into the core algorithm, the state at time t+1 is obtained. This refers to the new decision result, which is used for subsequent reward calculations and parameter optimization.

[0145] Model training process: The core objective of model training is to find the optimal core algorithm parameters through interactive iteration. The specific steps are as follows:

[0146] 1. Agent Initialization: Initialize the state space, action space, and reward function parameters of the DFRL agent. The dimension of the state space is consistent with the dimension of the fusion feature matrix of the AWMFF algorithm to ensure that feature information can be fully received. The action space is the range of values ​​for the algorithm parameter adjustment. A reasonable range is set according to the parameter type (e.g., the adjustment range of the weight matrix is ​​[-0.01, 0.01]) to avoid the algorithm being unstable due to excessive parameter adjustment.

[0147] 2. Interactive Iteration: The fused feature matrix output by the AWMFF algorithm is... As the initial state The agent, based on the current state and exploration strategy ( - Greedy strategy Output Action (Parameter adjustment amount); Substitute the adjusted parameters into the STA-GConvLSTM or IMOWOA algorithm to obtain the new decision result, i.e., the new state. ; Calculate the decision feedback reward based on the new status and historical execution data. and performance feedback rewards And thus receive the total reward ;

[0148] 3. Parameter Update: Calculate the reward function according to the optimization formula for the core algorithm parameters. Gradient, update parameters The number of iterations was set to 500 to ensure that the parameters were fully optimized;

[0149] 4. Convergence verification: When the reward function value fluctuates by less than 0.01 for 50 consecutive iterations, it indicates that the parameters have stabilized and the optimization effect has reached its optimum. Training is then stopped and the current optimal parameters are saved.

[0150] Model application process: When applying the model, the first step is to receive the fused feature matrix output by the AWMFF algorithm. (Right now The agent adjusts the output parameters based on this state. Subsequently, the weight matrix of STA-GConvLSTM and the adaptive weight of IMOWOA are updated according to the optimization formula to complete the parameter optimization of the core decision algorithm. The optimized parameters will be directly applied to the real-time decision-making process of the fire emergency big data intelligent decision service module to improve the accuracy of fire situation prediction and the optimization effect of resource allocation.

[0151] The algorithm achieves bidirectional interaction with the AWMFF algorithm: the reward function calculation of the DFRL algorithm depends on the fusion feature matrix extracted by the AWMFF algorithm. The accuracy of feature fusion directly affects the accuracy of state assessment; simultaneously, the reward value of the DFRL algorithm is fed back to the AWMFF algorithm to correct its adaptive weight calculation, and the interaction formula is as follows: ,in R is the correction coefficient and the reward function value of DFRL. When the reward value is high, it indicates that the quality of the current fused features is good, and the corresponding weight will be appropriately increased to further strengthen the role of effective features. If the feature fusion error (indirectly reflected by the reward value of DFRL) exceeds the threshold (reward value is less than 0.5), the AWMFF model will be retrained to improve the feature fusion accuracy and achieve synergistic optimization of the two types of algorithms.

[0152] (3) Application of data cloud accumulation and optimization results: Data cloud accumulation is the foundation of system iteration. Relying on the distributed storage service of the cloud computing platform, a dedicated cloud database for fire emergency data is established to store various types of data throughout the system process, ensuring that the data is traceable and analyzable. Examples of data types for accumulation are as follows: 1. Raw sensing data: Raw data directly acquired by all acquisition devices (such as raw sensor signals and raw video streams); 2. Pre-processed data: Standardized data processed by the fire scene full-domain data sensing and cloud access module; 3. Synchronization log data: Data synchronization time, synchronization status (success / failure), and synchronization data volume of each terminal; 4. Decision scheme data: Initial decision schemes generated by the fire emergency big data intelligent decision service module and all dynamically adjusted decision schemes; 5. Task execution data: Task issuance time, execution start time, execution completion time, and task completion quality assessment (such as rescue success rate and resource consumption control rate); 6. Feedback data: Feedback information reported by all terminals and records of emergencies; 7. Permission operation log: Login records, data access records, and operation records of each role (such as issuing instructions and adjusting tasks); 8. Secure transmission log: Data transmission time, encryption verification results, and access interception records.

[0153] All data is stored categorized by timestamp and event number, with a retention period of at least 5 years to meet the needs of fire emergency event review and historical data analysis; it also supports multi-dimensional retrieval (such as retrieval by event type, time, and terminal type) to facilitate quick querying and analysis of specific data.

[0154] The optimization results are directly applied to the core decision-making module. The optimized parameters output by the DFRL algorithm (the weight matrix of STA-GConvLSTM and the adaptive weights of IMOWA) are updated in real time to the fire emergency big data intelligent decision-making service module, replacing the original parameters for subsequent fire situation prediction and resource allocation optimization. The feature fusion scheme of the AWMFF algorithm is synchronized to the preprocessing stage of the fire scene full-domain data perception and cloud access module to optimize the data preprocessing strategy and improve the feature extraction effect of the original data. At the same time, the module generates a fire emergency command review report, which includes a review of the entire event handling process, evaluation of the effectiveness of the decision-making scheme, details of algorithm parameter optimization, and suggestions for subsequent improvement. This report is pushed to the PC and large screen terminals of the fire command center to provide a reference for commanders to summarize experience and improve command capabilities. The inputs to this module are the raw and preprocessed data from the fire scene full-domain data perception and cloud access module, the synchronization log data from the cross-terminal cloud-edge collaborative synchronization engine module, the decision scheme data from the fire emergency big data intelligent decision service module, the task execution data and feedback data from the fire task closed-loop collaborative execution module, and the permission operation logs and secure transmission logs from the multi-terminal permission dynamic adaptation and secure transmission module. The outputs are the optimized decision model parameters, the collaborative process optimization scheme, the permission configuration optimization scheme, and the fire emergency command review report, which are transmitted to the corresponding modules to achieve system iteration and upgrade.

[0155] refer to Figure 3 The algorithm's bidirectional interactive process is as follows: synchronous data input → spatiotemporal attention-enhanced graph convolutional long short-term memory network (situation prediction) → improved multi-objective whale optimization algorithm (decision generation) → execution feedback data → adaptive weighted multi-granularity feature fusion algorithm (feature extraction) → dual feedback reinforcement learning algorithm (parameter optimization) → reverse update of the preceding core algorithm.

[0156] refer to Figure 4The algorithm training and optimization process is as follows: historical data preparation → algorithm initialization (weights / parameters) → training set input → forward propagation operation → loss / reward function calculation → backpropagation parameter update → validation set accuracy evaluation → early stopping mechanism triggering → optimal parameter saving → real-time scenario application. This process is a general training and optimization process for spatiotemporal attention-enhanced graph convolutional long short-term memory networks, improved multi-objective whale optimization algorithms, adaptive weighted multi-granularity feature fusion algorithms, and dual-feedback reinforcement learning algorithms. The core training logic revolves around a closed loop of data-driven, iterative optimization, and validation implementation. It is an abstract and unified representation of the core training and optimization logic of the four algorithms, covering the entire process from data preparation to initialization, training, evaluation, optimization, and implementation. It can be fully adapted by only making targeted implementations in local steps (such as loss / reward function type and propagation operation details) according to the characteristics of each algorithm.

[0157] In some embodiments, a graph convolutional long short-term memory network with spatiotemporal attention mechanism enhancement interacts bidirectionally with an adaptive weighted multi-granularity feature fusion algorithm. The fire situation prediction data output by STA-GConvLSTM and the actual fire situation data collected by the fire-fighting task closed-loop collaborative execution module are input into AWMFF as a supplementary source of fine-grained features, enriching the dimensions of the multi-granularity features. The fused feature matrix extracted by AWMFF is fed back to the validation stage of STA-GConvLSTM to optimize the network's feature fitting ability for complex scenes. The interaction correlation formula is: In the formula The fire situation prediction features output by STA-GConvLSTM The adaptive weights for the predicted features are calculated in conjunction with the signal-to-noise ratio of other granular features. The interaction ends when the signal-to-noise ratio of the fused feature matrix output by AWMFF is ≥25dB ​​and the prediction error of STA-GConvLSTM based on the fused features is ≤8%. If both conditions are met, the current round of interaction stops.

[0158] In some embodiments, a graph convolutional long short-term memory network enhanced with spatiotemporal attention mechanism interacts bidirectionally with a dual-feedback reinforcement learning algorithm. DFRL updates the optimized weight matrix and bias terms to STA-GConvLSTM in real time, adjusting the network's spatiotemporal feature extraction accuracy; the prediction error of STA-GConvLSTM serves as a supplement to the DFRL decision feedback reward, correcting the evaluation dimension of the reward function. The interaction correlation formula is:

[0159] In the formula These are the predicted values ​​from STA-GConvLSTM. This is the actual value. The weights represent the prediction error. The interaction ends when the prediction error fluctuation of STA-GConvLSTM is ≤2% over 10 consecutive iterations, and the total reward function value of DFRL is ≥0.85. If both conditions are met continuously, the interaction terminates.

[0160] In some embodiments, the improved multi-objective whale optimization algorithm and the adaptive weighted multi-granularity feature fusion algorithm form a bidirectional interaction. The fused feature matrix output by AWMFF is used to screen the initial population of IMOWA, removing individuals with features that differ too much from the historical best solution, thereby improving the quality of the population; the three objective function values ​​of IMOWA are fed back to AWMFF to correct the adaptive weight allocation logic of the multi-granularity features. The interaction correlation formula is:

[0161] In the formula The three types of objective function values ​​for IMOWOA, The adjusted weights corresponding to the objective function (summing up to 0.1) are used. The interaction ends when the initial population quality rate of IMOWA is ≥90% and the fusion feature classification accuracy of AWMFF is ≥92%.

[0162] In some embodiments, the improved multi-objective whale optimization algorithm and the dual-feedback reinforcement learning algorithm form a bidirectional interaction. DFRL outputs an adaptive weight adjustment of IMOWA based on the total reward function, dynamically optimizing the priority of the three types of objectives; the actual execution effect (rescue efficiency, risk control) corresponding to the Pareto optimal solution of IMOWA serves as the core input for DFRL's execution feedback reward, improving the reward evaluation system. The interaction correlation formula is:

[0163] In the formula This is the weighting adjustment factor. Let be the gradient of the DFRL reward function. The interaction ends when: in 20 consecutive iterations, no new non-dominated solutions are added to the Pareto optimal solution of IMOWA, and the total reward function value of DFRL fluctuates by ≤3%.

Claims

1. A fire emergency command system based on multi-terminal collaboration, characterized in that, include: The fire scene full-domain data perception and cloud access module is used to collect fire emergency related full-domain data and perform preprocessing operations such as noise reduction, format standardization and outlier removal to obtain preprocessed data; The cross-terminal cloud-edge collaborative synchronization engine module is used to receive the preprocessed data and synchronize it to all terminals and subsequent modules through a hybrid transmission channel to obtain synchronized data; The fire emergency big data intelligent decision service module is used to receive the synchronous data, and to perform analysis and calculation using a graph convolutional long short-term memory network with spatiotemporal attention mechanism enhancement and an improved multi-objective whale optimization algorithm to generate a fire emergency command and decision scheme. The multi-terminal permission dynamic adaptation and secure transmission module is used to receive the fire emergency command and decision-making scheme, perform permission filtering and encryption on it, obtain targeted data and transmit it to each terminal and subsequent modules. The fire-fighting task closed-loop collaborative execution module is used to receive the directional data, issue tasks based on the data, and collect execution data and feedback data in real time during the task execution process; The emergency data cloud accumulation and iterative optimization module is used to receive the execution data and feedback data, and to analyze and optimize them using an adaptive weighted multi-granularity feature fusion algorithm and a dual-feedback reinforcement learning algorithm to obtain optimized algorithm parameters. The optimized algorithm parameters are then fed back to the fire emergency big data intelligent decision service module.

2. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The working process of the graph convolutional long short-term memory network enhanced by the spatiotemporal attention mechanism is as follows: First, spatial features are extracted from the received synchronous data. Key regional features are strengthened and irrelevant regional interference is weakened through spatial attention weights. Then, spatial correlations between regions are mined based on the spatial adjacency matrix. Subsequently, the temporal features of the data are extracted and the situational changes at key time nodes are focused on. Finally, temporal dependencies are mined through the long short-term memory network to output fire situation prediction data.

3. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The improved multi-objective whale optimization algorithm works as follows: taking fire situation prediction data output by a graph convolutional long short-term memory network with synchronized data and spatiotemporal attention mechanism as input, it first generates a diverse initial population through chaotic mapping, then constructs three types of objective functions: rescue response time, rescue resource consumption, and rescue risk, optimizes the solution vector through iterative strategies such as prey encirclement and bubble net attack, and dynamically adjusts the adaptive weights of each objective function, finally selects the Pareto optimal solution through non-dominated sorting method, and generates a fire emergency command decision scheme in combination with the actual needs of fire command.

4. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The spatiotemporal attention-enhanced graph convolutional long short-term memory network and the improved multi-objective whale optimization algorithm form a bidirectional interactive and collaborative relationship: the fire situation prediction data output by the spatiotemporal attention-enhanced graph convolutional long short-term memory network is used to correct the route distance parameter of rescue response time and the fire risk probability parameter of rescue risk in the improved multi-objective whale optimization algorithm; after the improved multi-objective whale optimization algorithm is executed, the actual fire situation data collected by the fire-fighting task closed-loop collaborative execution module is fed back to the spatiotemporal attention-enhanced graph convolutional long short-term memory network to verify the prediction accuracy. When the prediction error exceeds the set threshold, the network is triggered to retrain.

5. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The adaptive weighted multi-granularity feature fusion algorithm works as follows: hierarchical feature extraction is performed on the received execution data and feedback data. First, fine-grained features reflecting micro-changes in the data are extracted through one-dimensional convolution. Then, medium-grained features reflecting local patterns are extracted through one-dimensional pooling. Next, coarse-grained features reflecting global trends are extracted through principal component analysis. Then, the signal-to-noise ratio of various granular features is calculated. Adaptive weights are assigned according to the signal-to-noise ratio. Finally, the fused feature matrix is ​​obtained by weighted summation.

6. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The working process of the dual-feedback reinforcement learning algorithm is as follows: taking the fusion feature matrix output by the adaptive weighted multi-granularity feature fusion algorithm as the state input, a total reward function containing decision feedback reward and execution feedback reward is constructed. The decision feedback reward is obtained by comparing the algorithm's predicted decision result with the expert's optimal decision result. The execution feedback reward is calculated by combining the actual rescue time with the planned rescue time and the actual rescue risk probability. Based on the total reward function, the algorithm parameter adjustment amount is output, and the weight matrix of the spatiotemporal attention mechanism-enhanced graph convolutional long short-term memory network and the adaptive weights of the improved multi-objective whale optimization algorithm are iteratively updated.

7. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The adaptive weighted multi-granularity feature fusion algorithm and the dual-feedback reinforcement learning algorithm form a bidirectional interactive and collaborative relationship: the fusion feature matrix output by the adaptive weighted multi-granularity feature fusion algorithm provides accurate state input for the dual-feedback reinforcement learning algorithm, ensuring the accuracy of the parameter optimization direction; the total reward function value output by the dual-feedback reinforcement learning algorithm is used to correct the adaptive weight calculation of the adaptive weighted multi-granularity feature fusion algorithm. When the total reward function value is lower than a set threshold, the adaptive weighted multi-granularity feature fusion algorithm is triggered to re-extract features and optimize weights.

8. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The training process of the graph convolutional long short-term memory network enhanced by the spatiotemporal attention mechanism is as follows: historical fire situation data is collected and divided into training, validation, and test sets according to proportions. The weight matrix and bias terms of the network are randomly initialized, and the learning rate and maximum number of iterations are set. The training set data is input into the network for forward propagation to calculate the predicted value. The error between the predicted value and the true value is calculated using the mean squared error loss function. The parameters are updated by backpropagation using the Adam optimization algorithm. The model accuracy is evaluated through the validation set after a set number of iterations. When the validation set loss does not decrease for multiple consecutive iterations, an early stopping mechanism is triggered to save the optimal model parameters.

9. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The objective function optimization process of the improved multi-objective whale optimization algorithm is as follows: the three objective functions of rescue response time, rescue resource consumption, and rescue risk are normalized to eliminate the difference in dimensions. The initial weights are set according to the priority of fire emergency command. During the iteration process, the weights are dynamically adjusted according to the ratio of the current value to the maximum value of each objective function to ensure the coordinated optimization of the three types of objectives and avoid the overall scheme being insufficient due to the optimization of a single objective.

10. The fire emergency command system based on multi-terminal collaboration according to claim 1, characterized in that, The emergency data cloud accumulation and iterative optimization module is also used to store the entire process data of the system, including global data, preprocessed data, synchronized data, fire emergency command and decision-making schemes, execution data, and feedback data. The adaptive weighted multi-granularity feature fusion algorithm extracts multi-granularity features from the stored historical data, providing rich training data for the dual feedback reinforcement learning algorithm and ensuring the comprehensiveness and effectiveness of algorithm parameter optimization.

Images