Unmanned aerial vehicle cluster cooperative fire extinguishing method and system for building fire

Abstract

The embodiment of the application relates to a method for building fire extinguishing by a drone cluster, comprising: controlling a window breaking drone to break an outer window glass at a target building to form a fire extinguishing passage; obtaining fire environment information after the window breaking and calculating a fire flashover prediction time; comparing the flashover prediction time with a preset fire extinguishing operation preparation time to determine a fire extinguishing operation mode; when the flashover prediction time is greater than a sum of the fire extinguishing operation preparation time and a safety time margin, controlling a dry powder fire extinguishing drone to send dry powder fire extinguishing agent to an indoor fire source area for flame suppression, and controlling a foam fire extinguishing drone to spray foam fire extinguishing agent to the fire source area for sustained cooling after the dry powder fire extinguishing; and when the flashover prediction time is less than or equal to the sum of the fire extinguishing operation preparation time and the safety time margin, controlling the foam fire extinguishing drone to spray foam fire extinguishing agent to the fire source area for fire extinguishing. The application also provides a drone cluster fire extinguishing system.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) swarm technology, and in particular to a method and system for collaborative firefighting using UAV swarms in building fires. Background Technology

[0002] With the rapid pace of modern urbanization, the proliferation of super high-rise buildings has brought severe challenges to fire prevention and control. Due to the extreme vertical height of these buildings and the highly variable wind speeds in the external environment, traditional ground-based firefighting equipment often struggles to effectively cover the fire area using elevated platforms or water pump pressure. In recent years, drone firefighting technology has gained widespread attention as a flexible high-altitude response method. However, existing drone firefighting solutions still face significant bottlenecks when dealing with super high-rise fires.

[0003] On the one hand, most high-rise buildings use enclosed glass curtain wall structures, leaving the fire scene in a semi-enclosed or oxygen-deficient state. When drones perform window-breaking operations to create fire-fighting channels, a large influx of fresh air can instantly induce backfire or even transform into a violent flashover in a short period of time, seriously threatening the structural safety of the building and the conduct of rescue operations. Existing technologies generally lack dynamic perception of the fire's evolution trend after window breaking and cannot adjust combat tactics in real time according to the evolving logic of flashover risk.

[0004] On the other hand, fighting fires in ultra-high-rise buildings faces a contradiction between the payload capacity of drones and their firefighting effectiveness. While dry powder fire extinguishing bombs can suppress flames quickly, they lack the ability to continuously cool the fire scene, resulting in a high rate of reignition. On the other hand, while foam spraying has a good physical cooling effect, its response speed is relatively slow in dealing with the risk of flashover in the early stages of window breaches. In current technical solutions, there is a lack of in-depth temporal coordination and mission planning among various functional drones, making it difficult to form an integrated collaborative combat system of "window breaching - rapid suppression - long-term cooling," resulting in low agent utilization and difficulty in achieving precise handling of complex fire environments. Summary of the Invention

[0005] One objective of this application is to provide a method and system for collaborative firefighting using drone swarms in building fires, at least to address the aforementioned problems.

[0006] To achieve the above objectives, some embodiments of this application provide a method for coordinated firefighting using a swarm of drones in building fires, including: Control a glass-breaking drone at the location of the target building's exterior window to shatter the window glass and create a fire extinguishing path; After breaking a window, fire environmental information is acquired and the fire flashover prediction time is calculated. This process includes: using a methane laser remote sensing detector mounted on a reconnaissance drone to monitor the methane concentration in the smoke escaping from the broken window, and acquiring real-time indoor flame images to extract the heat release rate; using the methane concentration and heat release rate as input sequences, inputting them into a pre-trained long short-term memory network prediction model, and calculating the time from the current moment until the expected flashover occurs, which is then used as the fire flashover prediction time. The firefighting operation mode is determined by comparing the predicted flashover time with the preset firefighting preparation time. When the predicted flashover time is greater than the sum of the fire extinguishing preparation time and the safety margin, the dry powder fire extinguishing drone is controlled to deliver dry powder fire extinguishing agent to the indoor fire source area to suppress the flame, and after the dry powder fire extinguishing, the foam fire extinguishing drone is controlled to spray foam fire extinguishing agent to the fire source area for continuous cooling. When the predicted flashover time is less than or equal to the sum of the fire extinguishing preparation time and the safety margin, control the foam fire extinguishing drone to spray foam extinguishing agent towards the fire source area to extinguish the fire.

[0007] Some embodiments of this application also provide a drone swarm collaborative firefighting system for building fires, including: The intelligent dispatch center is used to make decisions on firefighting operation modes and issue operation instructions to the drone swarm; The reconnaissance drone is used to collect fire environment information and transmit flame image information and smoke concentration information to the intelligent dispatch center. Glass-breaking drones are used to shatter the exterior windows of target buildings under the control of the intelligent dispatch center. Dry powder fire extinguishing drones are used to deliver dry powder fire extinguishing bombs to indoor fire sources in standard operating mode. A foam fire extinguishing drone is used to spray foam fire extinguishing agent onto the fire source area.

[0008] Compared with related technologies, the solution provided in this application systematically solves the risk of flashback or even deflagration induced by the influx of fresh air after the glass curtain wall is broken in a fire in a super high-rise building by constructing an integrated collaborative combat system of "breaking windows - suppression - cooling". By connecting different functions of UAVs in the cluster in a timely manner, the fire extinguishing process is divided into two stages: instantaneous chemical suppression and long-term physical cooling. Under the condition of limited UAV payload, the fire extinguishing efficiency is maximized and the dynamic adaptation to the fire scene environment is achieved. Attached Figure Description

[0009] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0010] Figure 1 This is a flowchart of the drone swarm collaborative firefighting method provided in the embodiments of this disclosure.

[0011] Figure 2 This is another perspective flowchart of the drone swarm collaborative firefighting method provided in the embodiments of this disclosure.

[0012] Figure 3 This is another perspective flowchart of the drone swarm collaborative firefighting method provided in the embodiments of this disclosure.

[0013] Figure 4 This is a schematic diagram of the structure of the drone swarm collaborative firefighting system provided in the embodiments of this disclosure.

[0014] Figure 5 This is a schematic diagram of ground radiative heat flux monitoring data at a fire site provided in an embodiment of this disclosure.

[0015] Figure 6 This is a schematic diagram of the timing operation of a drone swarm collaborative firefighting according to an embodiment of this disclosure. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0018] In this disclosure, the terms "upper," "lower," "inner," "middle," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for better description of the embodiments of this disclosure and their implementations, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to require them to be constructed and operated in a specific orientation. Furthermore, some of the aforementioned terms may be used to indicate other meanings besides orientation or positional relationship; for example, the term "upper" may in some cases indicate a dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in the embodiments of this disclosure according to the specific circumstances.

[0019] Furthermore, the terms "set up," "connect," and "fix" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this disclosure according to the specific circumstances.

[0020] Unless otherwise stated, the term "multiple" means two or more.

[0021] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0022] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present disclosure can be combined with each other.

[0024] Combination Figures 1 to 6 As shown in the embodiments of this disclosure, a method for coordinated firefighting by drone swarms for building fires includes: Control a glass-breaking drone at the location of the target building's exterior window to shatter the window glass and create a fire extinguishing path; After breaking a window, fire environmental information is acquired and the fire flashover prediction time is calculated. This process includes: using a methane laser remote sensing detector mounted on a reconnaissance drone to monitor the methane concentration in the smoke escaping from the broken window, and acquiring real-time indoor flame images to extract the heat release rate; using the methane concentration and heat release rate as input sequences, inputting them into a pre-trained long short-term memory network prediction model, and calculating the time from the current moment until the expected flashover occurs, which is then used as the fire flashover prediction time. The firefighting operation mode is determined by comparing the predicted flashover time with the preset firefighting preparation time. When the predicted flashover time is greater than the sum of the fire extinguishing preparation time and the safety margin, the dry powder fire extinguishing drone is controlled to deliver dry powder fire extinguishing agent to the indoor fire source area to suppress the flame, and after the dry powder fire extinguishing, the foam fire extinguishing drone is controlled to spray foam fire extinguishing agent to the fire source area for continuous cooling. When the predicted flashover time is less than or equal to the sum of the fire extinguishing preparation time and the safety margin, control the foam fire extinguishing drone to spray foam extinguishing agent towards the fire source area to extinguish the fire.

[0025] The drone swarm collaborative firefighting method provided in this disclosure systematically solves the risk of flashback or even deflagration induced by the influx of fresh air after the glass curtain wall is broken in a fire in a super high-rise building by constructing an integrated collaborative combat system of "window breaking - suppression - cooling". By connecting the different functions of drones in the swarm in a timely manner, the firefighting process is divided into two stages: instantaneous chemical suppression and long-term physical cooling. Under the condition of limited drone payload, the firefighting efficiency is maximized and the dynamic adaptation to the fire scene environment is achieved.

[0026] This embodiment uses the flashover prediction time as the core judgment indicator and introduces multi-level logical comparison between fire extinguishing preparation time and safety time margin. This realizes the transformation of fire extinguishing strategy from experience-driven to data-driven, ensuring that in a highly uncertain fire scene environment, the system can match the optimal operation mode in real time according to the remaining safety window, thereby minimizing response time while ensuring operational safety.

[0027] Regarding the synergistic effect of the agents, this embodiment fully leverages the physicochemical advantages of dry powder and foam extinguishing agents. Dry powder extinguishing agents utilize their chemical inhibition capabilities to rapidly reduce the instantaneous radiant heat of the flames in the first stage, gaining a valuable tactical buffer period for subsequent operations. Meanwhile, the subsequent spraying of foam extinguishing agents focuses on long-term physical cooling and coverage of the fire source area, effectively compensating for the weakness of dry powder in preventing reignition. This phased, progressive firefighting tactic not only significantly improves the overall firefighting efficiency of a single flight payload but also ensures complete suppression of the fire, providing a reliable technical means for the rapid response of drones to fires in high-rise buildings.

[0028] In some embodiments, fire environment information includes at least flame image information and fire smoke composition information.

[0029] By integrating chemical composition monitoring and physical radiation analysis, the multidimensionality and real-time nature of situational awareness in ultra-high-rise fire zones have been significantly improved. The application of methane laser remote sensing technology enables reconnaissance drones to non-contactly acquire the concentration of characteristic gases escaping from breached windows without entering the core area of ​​high-temperature airflow, effectively solving the pain point of traditional contact sensors being prone to failure or response lag in extreme high-temperature environments. At the same time, combined with the heat release rate extracted from flame images, the system is provided with key physical quantities that intuitively reflect the intensity of energy evolution of indoor fire sources.

[0030] Building upon this foundation, a pre-trained Long Short-Term Memory (LSTM) predictive model was introduced, enabling a technological leap from static monitoring to dynamic trend extrapolation. Since flashover is a temporal evolution driven by fuel pyrolysis, ventilation conditions, and feedback radiation, the LSTM model can fully uncover the deep correlation between smoke concentration and heat release rate over time, capturing the acceleration characteristics of fire development. This quantitative predictive mechanism provides a scientific timeline for switching subsequent coordinated firefighting modes, ensuring that cluster operation commands are issued within the crucial window before drastic changes occur in the fire situation, greatly enhancing the predictability and safety of firefighting operations.

[0031] Optionally, before acquiring the heat release rate from real-time indoor flame images, an adaptive image data source selection step is also included: Use the camera of a reconnaissance drone to determine the state of the detached glass after the exterior window is broken; If it is determined that the exterior window glass has completely detached, then the camera is used to capture real-time images of the flames inside the room. If it is determined that the exterior window glass has not completely detached, the dispatch system will automatically connect to the target building's internal fire monitoring network to obtain real-time indoor flame images.

[0032] By employing adaptive switching logic for image data sources, this invention effectively addresses the potential physical obstruction issues that may arise during the demolition of exterior windows in high-rise buildings. In real-world scenarios, the effectiveness of glass-breaking drones against curtain wall glass is somewhat uncertain due to the material and remaining frame structure. Relying solely on external viewpoints for image acquisition can easily lead to obstructed visibility. This embodiment addresses this by real-time determination of glass detachment status and automatically accessing the building's internal fire monitoring network when external visibility is limited. This ensures continuous fire scene awareness and fundamentally mitigates the risk of dispatch decision failures caused by interruptions in acquiring critical parameters.

[0033] Meanwhile, because the internal monitoring network has a closer observation perspective to the fire source, it can effectively avoid the interference of intense smoke turbulence at the windows on images taken from the outside, thus providing higher-fidelity visual material for subsequent models. In this way, a deep integration of dynamic monitoring by drone swarms and fixed fire protection facilities in buildings is achieved, enabling the system to maintain strong environmental adaptability and robust situational analysis capabilities even under complex demolition conditions.

[0034] Optionally, before controlling the glass-breaking drone to shatter the window glass at the target building's exterior window location, a non-contact fire source positioning step is also included: Infrared thermal images of the unexposed side of the exterior window glass were collected using a reconnaissance drone. The infrared thermal image is input into an image recognition model pre-trained based on a residual network to extract thermal image features and output the fire source location number, thereby locking the spatial coordinates of the indoor fire source area.

[0035] By employing an infrared thermal imaging identification mechanism based on residual networks, non-contact and precise pre-positioning of indoor fire sources was achieved. In the early stages of fires in high-rise buildings, indoor fire sources are often obscured by high-strength glass curtain walls, making it difficult for traditional visible light detection to penetrate the curtain walls and obtain internal information. This embodiment utilizes infrared thermal imaging technology to capture the temperature field distribution characteristics of the unexposed side of the glass, enabling the sensing of abnormal areas of internal thermal radiation through the curtain wall. This solves the pain point of "not being able to see or accurately locate" before breaking the window, providing a reliable physical basis for subsequent glass-breaking drones to select the optimal operation point.

[0036] Simultaneously, by extracting deep features from thermal images using a pre-trained residual network model, the system can automatically filter out interference from building exterior wall reflections and environmental clutter, transforming the complex continuous temperature distribution into intuitive fire source location numbers and spatial coordinates. This automated identification process significantly reduces the decision-making cycle from fire reconnaissance to demolition, ensuring that window-breaking actions are precisely targeted at the core area of ​​the fire. This pre-positioning step not only improves the hit rate of subsequent fire extinguishing agent delivery but also tactically guarantees the efficient connection between window breaking and suppression, gaining a crucial first-strike opportunity to control the spread of the fire.

[0037] Optionally, the delivery mass of dry powder extinguishing agent and the spray flow rate of foam extinguishing agent to the indoor fire source area are determined by a fire extinguishing performance optimization model.

[0038] By optimizing the fire extinguishing efficiency model, the contradiction between the limited payload and highly complex fire extinguishing requirements of UAVs in ultra-high-rise firefighting missions was resolved. Due to the physical limitations of the UAV platform's takeoff weight and agent carrying capacity, blindly increasing the deployment of a single extinguishing agent often leads to diminishing marginal effectiveness. This embodiment, based on measured data extracted from full-scale fire tests, pre-defines the coupling relationship between fire suppression and cooling factors in the model, enabling the scientific calculation of the optimal ratio of dry powder to foam. This achieves a dynamic balance between rapid local fire suppression and long-term global cooling under limited load conditions.

[0039] In some embodiments, the fire extinguishing performance optimization model is constructed based on full-scale fire extinguishing test data conducted for typical fire scenarios, and is used to determine the optimal ratio of dry powder extinguishing agent to foam extinguishing agent under UAV payload constraints. In other words, the fire extinguishing performance optimization model pre-sets an optimal ratio scheme based on full-scale fire test verification. This optimal ratio scheme uses the UAV platform payload boundary as a constraint, and takes fire suppression factor, cooling factor, and reignition prevention status as comprehensive evaluation indicators to determine an empirical value for the agent ratio that balances rapid local fire suppression and long-term global cooling.

[0040] For example, based on full-scale test data, for a typical 21-square-meter standard protection scenario, the recommended preset mass of dry powder extinguishing agent is 3 kg, while the preset spray flow rate of foam extinguishing agent is 2 L / s. This ratio is not arbitrarily set, but is based on a fire extinguishing efficiency optimization model. While ensuring fire suppression efficiency, it also takes into account the load constraints of the drone platform and avoids diminishing marginal efficiency caused by increasing the mass of dry powder.

[0041] In some embodiments, the delivery time of the dry powder fire extinguishing bomb is determined by the intelligent dispatch center based on the preparation time for the dry powder bomb operation.

[0042] By dynamically locking the delivery time based on the dry powder extinguishing bomb's preparation time through an intelligent dispatch center, a deep coupling between firefighting operations and the fire's evolution window is achieved. In high-rise building fires, every second after a window is broken is crucial in determining whether the fire will abruptly escalate from localized combustion to full-blown flashover. The timing of the dry powder extinguishing agent's intervention directly determines the upper limit of its chemical suppression effectiveness. By incorporating the critical engineering parameter of preparation time into the dispatch logic, the system can preemptively offset the inherent time required for UAV flight preparation, attitude adjustment, and ammunition unlocking, ensuring that the dry powder extinguishing bomb detonates precisely and on time within the golden suppression period after the window is broken. This technically eliminates the risk of fire suppression failure due to delayed commands.

[0043] Meanwhile, this time benchmark, determined by the intelligent dispatch center, lays the logical foundation for the tiered intervention of subsequent foam fire-fighting drones. The precise timing of dry powder delivery allows the system to meticulously plan the establishment sequence of subsequent foam spraying links, using the completion time of this action as the origin, thus forming a compact and non-interfering task chain on the timeline. This task planning method, based on the actual response capabilities of the equipment, not only improves the reliability of fire-fighting agent delivery in complex facade environments but also ensures that the entire drone swarm can efficiently complete the smooth transition from instantaneous fire suppression to long-term cooling within a very short tactical window.

[0044] Optionally, if the predicted combustion time is greater than the sum of the fire extinguishing preparation time and the safety time margin, within the first time window from the completion of the window breaking to the end of the fire extinguishing preparation time, the dry powder fire extinguishing drone is controlled to deliver and detonate the dry powder fire extinguishing agent; within the second time window starting from the fixed delay time after the end of the fire extinguishing preparation time, the foam fire extinguishing drone is controlled to start and spray the foam fire extinguishing agent.

[0045] By precisely segmenting the timing of firefighting operations, deep synergy between dry powder suppression and foam cooling is achieved at the tactical level. Under relatively controllable fire conditions (sufficient time for flashover prediction), the system utilizes the first time window, leveraging the rapid chemical suppression rate of dry powder extinguishing agents to quickly reduce the intensity of fire source radiation and contain flame spread within the golden period after the window is broken. This proactive suppression action effectively offsets the risk of reignition caused by gas replenishment after the window is broken.

[0046] Meanwhile, by setting a fixed delay time and opening a second time window, a seamless connection between foam extinguishing operations and the dry powder action phase was ensured. After the dry powder agent initially controlled the fire, the foam drone intervened, utilizing the excellent physical coverage and continuous cooling properties of the foam extinguishing agent to thoroughly encapsulate and deeply cool the remaining fire core. This tiered operational sequence not only avoided physical interference between different extinguishing agents during the spraying process but also greatly improved the thoroughness of handling high-rise indoor fires through the combination of instantaneous fire control and long-term prevention of reignition, ensuring the continuity and logical rigor of the firefighting operation in terms of time.

[0047] In some embodiments, the first time window is defined as the effective suppression time window of the dry powder bomb determined based on full-scale fire tests, and the window period is set to 80 to 120 seconds; the objective of the dry powder extinguishing agent is to delay the arrival time of flashover through chemical inhibition within the window period, thereby opening up tactical preparation time for the long-term cooling operation of the subsequent foam extinguishing agent.

[0048] For example, in this embodiment, the "first time window" (dry powder bomb delivery window) and the "second time window" (foam injection start-up window) are not fixed values, but are dynamically defined based on the real-time flashover prediction results. The specific logic is as follows: Preparation time for dry powder bomb drone operations (e.g., 90 seconds). Safety time margin (e.g., 10 seconds); In standard mode, there is a fixed delay (e.g., 60 seconds) between the completion of dry powder bomb firing and the start of foam spraying. In emergency mode, the system requires a maximum reaction time (100s) for the foam fire extinguishing drone to begin spraying.

[0049] when In standard operating mode, the first time window is "from the completion of the window breaking (t=0) to t= End. The system immediately issues the dry powder bomb launch command at t=0. The second time window is from t= Start, until t= + End. The system planning bubble at t= + "Start spraying at the moment."

[0050] when ≤ + In emergency operation mode, the first time window is closed (dry powder bombs are not fired), and the second time window is from the completion of the window breaking (t=0) to t= End. The system immediately issues the highest priority command to the foam drone at t=0, requiring it to complete the command at t= "Start spraying before the designated time."

[0051] For example: if the window is broken =120 seconds, since 120 > (90 + 10), enter standard mode. The command is: immediately fire dry powder bomb, and begin spraying foam at 160 seconds (90 + 10 + 60). If the window is broken... =90 seconds, since 90 < (90 + 10), enter emergency mode. The instruction is: do not fire dry powder bombs, immediately dispatch foam drones to start spraying within 100 seconds.

[0052] Optionally, when the flashover prediction time is less than or equal to the sum of the fire extinguishing preparation time and the safety time margin, an activation command is issued to the foam fire extinguishing drone within the emergency time window from the moment the window is broken until the end of the preset maximum reaction time, controlling it to establish a foam spraying link within the preset reaction time, wherein the preset reaction time is less than the fire extinguishing preparation time.

[0053] By dynamically compressing the preparation time for firefighting operations through preset reaction times, the system can decisively abandon the conventional logic of multi-stage coordination and directly trigger the highest-priority intervention of foam firefighting drones under the pressure of insufficient predicted safety margins. This subtractive scheduling strategy ensures that the rescue operation can forcibly establish a foam spray cooling link before the fire becomes out of control, demonstrating the system's adaptive survivability and rapid risk avoidance capabilities under extreme loads.

[0054] Meanwhile, this defined emergency time window provides clear tactical time constraints for the rapid deployment of foam firefighting drones in complex airflow and high-temperature environments. By placing the process of establishing the spray link within a preset (maximum) reaction time, command redundancy or timing delays caused by multi-drone coordination are effectively avoided, achieving extreme scheduling of core cooling resources. This embodiment not only significantly enhances the fault tolerance of drone swarms in handling ultra-high-rise fires, but also forms an effective countermeasure against sudden flashovers at the tactical level, maximizing the effectiveness and certainty of firefighting operations within extreme window periods.

[0055] Optionally, the control actions for each UAV are generated by the intelligent scheduling center based on a multi-agent deep reinforcement learning model.

[0056] In some implementations, the intelligent dispatch center treats reconnaissance drones, glass-breaking drones, dry powder fire extinguishing drones, and foam fire extinguishing drones in the drone swarm as independent agents in a partially observable Markov decision process. Each agent inputs its local observation state into a pre-set distributed actor network and independently outputs a sequence of action instructions, coordinating to execute joint actions under spatial constraints and operational timing constraints.

[0057] By employing a Multi-Agent Deep Reinforcement Learning (MADRL) model, the complex firefighting decision-making process is transformed into a partially observable Markov decision process, fundamentally solving the real-time and robustness challenges of multi-drone collaboration in ultra-high-rise fire environments. At fire scenes, communication delays, high-temperature smoke interference, and dynamically evolving fire conditions create a highly uncertain environment. Traditional centralized scheduling logic is prone to decision-making delays due to data transmission interruptions or computing power bottlenecks. This embodiment defines various functional UAVs as independent intelligent agents and adopts a distributed actor network, enabling each UAV to generate action sequences autonomously and in real-time based on its own local observation state, achieving a technological leap from command dependence to intelligent autonomy.

[0058] Meanwhile, this mechanism achieves a high degree of coordination among multiple drone types under spatial and operational timing constraints. Through pre-set reinforcement learning training logic, reconnaissance, glass-breaking, dry powder, and foam fire extinguishing drones can spontaneously form tactical coordination, completing precise coordination of window-breaking, fire suppression, and cooling actions within an extremely short window period. This distributed execution architecture enhances the system's fault tolerance in complex electromagnetic environments at ultra-high-rise buildings and ensures the cluster's response speed when dealing with changing fire situations, providing solid algorithmic support for seizing the golden opportunity for fire extinguishing.

[0059] In some alternative embodiments, the training process of the distributed actor network is completed in a digital twin simulation environment; the digital twin environment is constructed by integrating fire evolution data generated by fire dynamics simulation software (FDS) with UAV dynamics models to simulate training conditions covering different fire source locations, fire intensities and building facade wind field environments.

[0060] By training a distributed actor network in a digital twin simulation environment, the technical challenges of scarce on-site measured data and extremely high training costs in ultra-high-rise fires are addressed. Through deep integration of fire evolution data generated by Fire Dynamics Simulation Software (FDS) with the UAV's own dynamic model, the system can pre-simulate extreme conditions in digital space, covering different fire source locations, fire intensities, and variable facade wind fields. This training method based on high-fidelity physical simulation data not only endows each UAV agent with strong generalization capabilities and environmental adaptability but also ensures that the algorithm model maintains robust collaborative performance and precise action execution when actually deployed in complex building scenarios.

[0061] In some embodiments, the reward function of the multi-agent deep reinforcement learning model used in the distributed actor network is set based on the principle of multi-objective optimization, specifically including: positive reward items include successful window breaking action, ground radiant heat flux reduction ratio, indoor temperature reduction rate and final fire source extinguishing state; negative reward items include total task execution time, safe distance conflict between drone platforms, and ineffective loss of fire extinguishing agent payload.

[0062] By designing a multi-objective optimized reward function, the behavioral guidelines and value orientation of the UAV swarm in firefighting operations were established. Key firefighting physical parameters such as the decrease rate of ground radiant heat flux and the rate of indoor temperature decrease were set as positive reward items, forcing the agents to pursue the optimal balance of firefighting efficiency from the algorithmic level, rather than simply completing a single action. Conversely, task time, inter-platform collision risk, and payload loss were set as negative reward items, effectively constraining the swarm's operational safety and resource utilization efficiency under limited spatial and temporal resources. This refined reward feedback mechanism enables the multi-agent system to automatically seek the optimal decision path between fire suppression effect, cooling rate, and operational safety in highly dynamic firefighting tasks.

[0063] For example, the entire firefighting operation scenario is modeled as a multi-agent partially observable Markov decision process (Dec-POMDP). Its core elements are defined as follows: Intelligent Agent: Each functional drone in the swarm (reconnaissance, glass breaking, dry powder, foam) is considered an intelligent agent, and the set is as follows: A； state( ):time t Below, the true and complete state of the environment is obtained, including the precise state of the fire source, the building structure, and the precise poses and internal states of all drones. This state cannot be fully obtained by any single intelligent agent.

[0064] Observation ( ): Each intelligent agent a At any moment t Local environmental information is obtained through the agents' own sensors. For example, a reconnaissance drone acquires thermal images and gas concentrations, while a dry powder drone only knows its own position and remaining payload. The observations of all agents collectively constitute joint observation. .

[0065] action( ): Each intelligent agent a Executable actions include moving to coordinates, firing ammunition, and activating spray. The actions of all agents constitute a joint action. .

[0066] State transition probability ( This describes the probability distribution of the transition from one state to the next under joint actions. It encapsulates uncertainties such as fire development and drone dynamics.

[0067] Reward function ( This is the core of the system design, used to quantitatively evaluate the immediate effects of joint actions in a specific state. Its design follows multi-objective optimization principles, aiming to guide the agent to learn the optimal cooperative strategy. Typical reward signals include: Positive rewards include: successfully breaking a window, using dry powder to suppress flames (proportional to the decrease in ground radiant heat flux), using foam to reduce temperature (proportional to the rate of temperature drop), and successfully extinguishing a fire.

[0068] Negative rewards: mission timeout, drone collision.

[0069] The goal of each agent is to learn a policy. This strategy determines actions based on its local observation history in order to maximize the team's expected cumulative discount reward from the environment. ).

[0070] To handle the aforementioned multi-agent partially observable Markov decision processes, a multi-agent deep reinforcement learning model is employed. During the training of this deep reinforcement learning model, a system with access to the global state is utilized. A centralized commentator for all agent observations and actions. This commentator learns a joint action-value function. The parameters are This is used to evaluate the long-term value of taking joint actions in a global context. Each agent has a "decentralized actor." The parameters are It selects actions solely based on its own observations. The gradient update strategy is as follows: ; The global gradient signal provided by the centralized commentator guides each actor to learn how to contribute to the global optimal goal, thereby solving the problem of credit allocation among agents.

[0071] A digital twin simulation environment was constructed based on high-fidelity fire dynamics simulators (such as FDS) and UAV dynamics models. This environment safely and efficiently generates massive amounts of training data covering various fire scenarios (different locations and intensities), building structures, and failure modes. Simultaneously, certain domain knowledge (e.g., "window breaking must occur before dry powder delivery") is embedded into the training process as constraints or intrinsic rewards to accelerate convergence and ensure the strategy conforms to physical and safety common sense. The ultimately learned strategy can generalize to complex fire scenarios not encountered during training.

[0072] After training, the distributed actor network is deployed as the decision-making kernel of the intelligent scheduling center. In practical tasks: Real-time observation provided by reconnaissance drones and other sensing units (e.g., coordinates of the fire source location) are used as network input; The network outputs the optimal actions for each drone. (e.g., immediately to spatial coordinates) x, y, z (Fire dry powder bombs). These action commands are translated into specific flight control and fire control commands and issued.

[0073] When applying this deep learning model, only a pre-trained distributed actor network is needed. Each agent makes decisions independently based on its own real-time local observations, without the need for real-time communication and coordination with other agents or a central node. This ensures the system's high reliability and rapid response capability even under conditions where communication in a real fire scene may be limited.

[0074] In some optional embodiments, the flame suppression effect of dry powder extinguishing agent is characterized in real time by ground radiant heat flux; after the dry powder extinguishing agent is delivered and acts, the evolution trend of ground radiant heat flux at the fire site is monitored by reconnaissance drones; if the decrease in ground radiant heat flux after a certain amount of powder acts reaches a preset threshold, it is determined that the tactical objective of flashover suppression has been achieved, and the command to intervene by foam fire extinguishing drones is triggered.

[0075] By utilizing the key physical parameter of ground radiant heat flux, the system enables quantitative assessment and feedback control of the effectiveness of dry powder fire suppression. In the extreme environment of high-rise fires, traditional visual reconnaissance is insufficient to accurately determine the degree of chemical suppression within the building. Ground radiant heat flux, however, directly reflects the intensity of energy released from the fire source to the surrounding environment and is a core indicator for assessing the risk of flashover. By monitoring the evolution trend of this value, the system can transform the assessment of dry powder fire suppression effectiveness from vague empirical judgments into precise numerical logic, providing an objective triggering basis for subsequent firefighting intervention.

[0076] Meanwhile, this tactical judgment mechanism based on a threshold of heat flux reduction ensures efficient coordination between dry powder and foam extinguishing agents in actual combat. Only when the heat flux reduction after the application of dry powder reaches a preset percentage (e.g., 61.37%), proving that the instantaneous energy of the fire has been effectively suppressed and the flashover threshold has been successfully avoided, will the system issue an intervention command for the foam fire-fighting drone. This closed-loop control logic not only avoids the waste of extinguishing agents caused by blind intervention when the fire energy is still too high, but also accurately captures the optimal timing for the transition from chemical suppression to physical cooling, significantly improving the certainty and efficiency of clustered collaborative operations.

[0077] In practical applications, after the dry powder extinguishing agent is delivered, it instantly reduces the ground radiant heat flux at the fire scene through chemical inhibition. The system monitors the evolution curve of the ground radiant heat flux in real time using reconnaissance drones. When the heat flux decreases by a preset percentage (e.g., more than 60%) after several powder bombs have been applied, the dry powder inhibition step is considered complete, providing an 80 to 120 second window for subsequent foam intervention.

[0078] In some embodiments, when the foam fire extinguishing drone performs fire extinguishing tasks in super high-rise building scenarios with a height of 100 to 200 meters, it uses fluid control logic to suppress the water hammer effect in the mooring pipeline; the foam fire extinguishing drone system uses a ground high-pressure pump group for pressure compensation and uses special lightweight pipe materials to reduce the drone's load pressure, so as to maintain the stability of foam spray flow and flight attitude anti-interference at super high-rise heights.

[0079] By introducing fluid control logic into the foam fire extinguishing drone system, the water hammer effect caused by valve opening and closing or pressure fluctuations during long-distance vertical transport of liquid is effectively suppressed, preventing damage to pipeline connection points and the drone's precision flight platform from instantaneous shock waves. This proactive control based on fluid characteristics ensures that the foam extinguishing agent maintains a constant spray flow rate and pressure even at extreme altitudes.

[0080] Meanwhile, the pressure compensation mechanism of the ground-based high-pressure pump set, combined with special lightweight pipes, significantly reduces the vertical pipe load that the UAV needs to overcome, lowers motor power consumption, and extends effective operating time. The lightweight pipes, while reducing physical weight, also reduce the swaying inertia of the pipeline under wind interference, thus significantly improving the UAV's flight attitude resistance in strong gusts at ultra-high altitudes. This embodiment ensures that the firefighting platform can accurately lock its spraying posture during high-altitude operations, achieving a technical balance between efficient agent delivery and high stability of the flight platform.

[0081] This disclosure also provides a drone swarm collaborative firefighting system for implementing the methods described in the above embodiments, comprising: an intelligent dispatch center for making firefighting operation mode decisions and issuing operation instructions to the drone swarm; a reconnaissance drone for collecting fire environment information and transmitting flame image information and smoke concentration information to the intelligent dispatch center; a glass-breaking drone for breaking the exterior windows of the target building under the control of the intelligent dispatch center; a dry powder fire extinguishing drone for delivering dry powder fire extinguishing bombs to the indoor fire source area in standard operation mode; and a foam fire extinguishing drone for spraying foam fire extinguishing agent to the fire source area.

[0082] The drone swarm collaborative firefighting system provided in this disclosure achieves modular collaboration across the entire process of handling fires in ultra-high-rise buildings through functional decomposition of hardware entities and logical integration with an intelligent dispatch center. Unlike traditional single-unit or homogeneous swarm systems, this system distributes the four core functions of reconnaissance, obstacle breaching, suppression, and cooling to specialized drone platforms, with each node forming a dynamic chain under the unified command of the intelligent dispatch center. This architecture not only distributes the load pressure on individual drones and solves the limitation of a single device struggling to handle multiple tasks in ultra-high-rise environments, but also ensures efficient connection between command issuance and data feedback through standardized interfaces and control protocols.

[0083] In terms of practical coordination, the system establishes a closed-loop response mode encompassing sensing, transmission, calculation, and execution. Multi-source environmental information provided by reconnaissance drones serves as the data foundation for the dispatch center's predictive models, while glass-breaking, dry powder, and foam firefighting drones, guided by decision-making commands, execute precise strikes according to pre-set firefighting operation modes. This highly integrated system configuration ensures tight coordination in spatial deployment and action sequence during firefighting operations. Particularly concerning is the risk of flashover, which is highly prevalent in high-rise fires. Through cross-type tactical coordination, the system can achieve rapid suppression and long-term control of the fire situation, significantly improving the overall practical effectiveness of high-altitude fire rescue.

[0084] In this embodiment, the foam fire extinguishing drone includes a ground high-pressure pump set, special lightweight anti-corrosion pipes, and a flight platform with an anti-interference configuration. When the operating altitude exceeds a preset threshold (such as 100 meters), the continuous and stable spraying of foam fire extinguishing agent is ensured through fluid control logic that suppresses water hammer effect and pressure compensation of the high-pressure pump set.

[0085] Optionally, the reconnaissance drone is equipped with an infrared thermal imager, a camera, and a methane laser remote sensing detector.

[0086] By integrating infrared thermal imagers, high-definition cameras, and methane laser remote sensing detectors onto reconnaissance drones, a multi-dimensional sensing hardware foundation for ultra-high-rise fires has been established. This payload configuration encompasses comprehensive data acquisition capabilities, ranging from thermal field distribution and visual images to characteristic gas concentrations. It solves the problem of incomplete information acquisition under dense smoke or obscured conditions using traditional single reconnaissance methods, providing high-fidelity underlying raw data for subsequent fire source location and flashover early warning.

[0087] In some embodiments, the intelligent dispatch center incorporates a flashover prediction model to calculate the flashover prediction time.

[0088] The intelligent dispatch center's built-in flashover prediction model transforms discrete environmental monitoring data into predictive tactical indicators. By calculating flashover prediction time in real time, the system can quantify the intensity of fire development and critical points of qualitative change, thus providing scientific decision-making references for command and ensuring that firefighting strategies can be formulated ahead of sudden changes in the fire's evolution.

[0089] In some embodiments, the intelligent dispatch center has a built-in fire extinguishing efficiency optimization model to determine the ratio of dry powder fire extinguishing agent to foam fire extinguishing agent.

[0090] The optimal ratio of dry powder extinguishing agent to foam extinguishing agent was established through mathematical modeling. Under the physical conditions of limited UAV payload, the model achieved an efficient combination of chemical inhibition and physical cooling extinguishing mechanisms, avoiding the indiscriminate delivery of extinguishing agents and maximizing the extinguishing efficiency and reignition prevention level of a single flight operation.

[0091] In some embodiments, the intelligent dispatch center incorporates an intelligent dispatch deep learning model to generate task dispatch strategies based on fire status information and drone status information.

[0092] The built-in intelligent scheduling deep learning model achieves automated generation of task scheduling strategies through real-time offset analysis of fire scene status and drone status. This model can handle massive dynamic variables and quickly extract the optimal execution path from the complex fire scene evolution logic, significantly reducing the response latency from information acquisition to command issuance and improving the overall coordination efficiency of the cluster system under extreme pressure.

[0093] In some embodiments, the intelligent scheduling deep learning model is trained using a multi-agent reinforcement learning model.

[0094] The system is trained using a multi-agent reinforcement learning model, which endows it with self-evolution and collaborative capabilities in highly uncertain environments. Through simulated training on large-scale fire scene data, each UAV agent can learn deep collaborative patterns under spatial constraints and temporal coordination, ensuring robustness and decision continuity even when facing complex facade environments and variable airflow in actual combat.

[0095] In some embodiments, the intelligent dispatch center performs task planning based on the drone's location, speed, remaining payload, and equipment status.

[0096] The system performs refined task planning based on the drone's location, speed, remaining payload, and equipment status, achieving Pareto optimal resource allocation. This resource scheduling logic based on real-time operating conditions ensures that task allocation is highly matched with the drone's physical capabilities and remaining flight time, effectively preventing rescue interruptions due to single drones running out of payload or lacking power, and guaranteeing the integrity of the firefighting chain.

[0097] In some embodiments, the job instructions generated by the intelligent scheduling center include target spatial coordinates, job action type, and execution time window.

[0098] The operational instructions generated by the intelligent dispatch center clearly define the target spatial coordinates, operational action types, and execution time windows, establishing a standardized and digital expression for combat instructions. This instruction system, constrained by four dimensions of time and space, enables UAV swarms to accurately align with predetermined points and perform specific actions at predetermined times, resolving the issues of time and space conflicts and action overlaps that easily occur during multi-UAV operations.

[0099] In some embodiments, reconnaissance drones are used to acquire infrared thermal images of the building facade before the window is broken in order to determine the location of the indoor fire source area.

[0100] Before breaking the window, the reconnaissance drone collected infrared thermal images of the exterior facade, enabling non-contact pre-location of the indoor fire source. By detecting abnormal heat radiation distribution behind the curtain wall, the system can perceive the core of the internal fire source through physical shielding, providing prior information for the window-breaking drone to select the best attack point and avoiding the ineffective risk of oxygen rushing in due to blindly breaking the window.

[0101] In some embodiments, the foam fire extinguishing drone is connected to a ground foam supply system via a tethered line.

[0102] Foam fire extinguishing drones connect to a ground-based foam supply system via tethered pipelines, breaking the capacity limitations of traditional airborne agent tanks. The tethered supply method ensures a continuous and high-intensity output of the foam extinguishing agent, enabling the drone to perform long-term spray cooling tasks and completely solving the rigid requirement for continuous cooling capabilities in high-rise building fires.

[0103] In this embodiment, at the initial stage of the mission, the reconnaissance drone first flies to the exterior of the target building and hovers at a predetermined safe distance from the target windows. It then uses an airborne infrared thermal imager to acquire infrared thermal images of the unexposed side of the window glass and inputs this image data in real time into an image recognition model pre-trained based on a residual network. The model extracts the thermal distribution features from the thermal image to identify the corresponding fire source location number, thereby accurately pinpointing the spatial coordinates of the indoor fire source area and providing a spatial reference for subsequent operations.

[0104] After acquiring the spatial coordinates, the intelligent dispatch center controls a glass-breaking drone to perform the demolition operation. Once the drone shatters the exterior window, a reconnaissance drone immediately uses its camera to assess the extent of the glass detachment. If the glass is determined to be completely detached, the reconnaissance drone directly captures real-time images of the indoor flames. If large pieces of remaining glass obstruct the view, the dispatch system automatically switches to the building's existing fire monitoring network to obtain unobstructed real-time footage. Simultaneously, the methane laser remote sensing detector on the reconnaissance drone monitors the methane concentration in the smoke escaping from the broken window.

[0105] Subsequently, the intelligent dispatch center extracts the real-time heat release rate and flue gas methane concentration as sequence data and inputs them into the Long Short-Term Memory (LSTM) prediction model to calculate in real time the estimated time for flashover to occur from the current moment. The intelligent dispatch center uses this dynamic forecast value and the preset firefighting preparation time as a basis. ) and safety time margin ( Logical comparison is performed.

[0106] In the first working condition of this embodiment, if the calculation yields... (For example, if a flashover is predicted to occur in 120 seconds, and the preparation time is 60 seconds), the system determines to enter standard operating mode. At this time, the intelligent dispatch center issues an instruction, and a dry powder fire extinguishing drone quickly flies to the window position and delivers 3kg of dry powder fire extinguishing bombs to the core area of ​​the fire source. Within the first time window after the window is broken, the dry powder fire extinguishing agent generated by the detonation of the dry powder bombs rapidly performs chemical suppression, and the monitored ground radiant heat flux instantly decreases by about 60%, effectively delaying the arrival of the flashover. Subsequently, after the preset fixed delay time, a tethered foam fire extinguishing drone intervenes, continuously spraying foam fire extinguishing agent into the room at a flow rate of 2L / s for physical cooling until the indoor temperature drops from the high temperature peak to below the preset safety threshold.

[0107] In the second working condition of this embodiment, if the calculation yields... (For example, if the fire evolves extremely rapidly, with a predicted flashover occurring within 40 seconds), the system immediately switches to emergency operation mode. At this point, the intelligent dispatch center cancels the dry powder delivery step and directly issues a start command to the foam fire extinguishing drone. The foam fire extinguishing drone quickly establishes a foam spraying link within a very short preset reaction time, obtaining a continuous supply of foam from the ground through tethered pipelines, directly and powerfully cooling the fire site, suppressing the fire to the maximum extent within the emergency time window before the flashover occurs, and ensuring timely rescue in high-risk environments.

[0108] Throughout the entire operation, the specific control actions of each UAV are generated in real time by the intelligent dispatch center based on a distributed actor network. The intelligent dispatch center, through a multi-agent deep reinforcement learning model, comprehensively considers the UAV's real-time location, remaining flight time, and payload status, and coordinates the independent output of action commands by reconnaissance, glass-breaking, dry powder, and foam UAVs under spatial and operational timing constraints. This decentralized execution logic ensures that, even in the complex and uncertain communication environment of ultra-high-rise fire sites, the execution units can maintain a high degree of tactical coordination and response accuracy.

[0109] For example, an indoor fire occurred on the 60-meter-high floor of a high-rise building.

[0110] Initial reconnaissance and location: After the fire broke out, a reconnaissance drone was the first to arrive at the scene, hovering approximately 20 meters from the exterior of the building where the fire originated; this moment is defined as T0. The drone activated its infrared thermal imager, continuously acquiring and transmitting infrared thermal images of the unexposed side of the target window glass. The image data was transmitted in real time to the intelligent dispatch center, where the fire source location module analyzed the data to preliminarily determine the spatial location of the fire area within the building.

[0111] Breaking the window to create a passage: After obtaining the initial location of the fire source, the glass-breaking drone, under dispatch command, flew to the outside of the target window. It aimed and launched a breaching bomb, successfully shattering and clearing the target glass curtain wall at T0+53 seconds, thus creating a passage for the delivery of fire extinguishing agents.

[0112] Real-time risk monitoring and prediction: After the window was broken, a large amount of high-temperature smoke escaped from the room. The reconnaissance drone simultaneously activated its methane laser remote sensing detector and high-definition camera to monitor the methane concentration in the escaped smoke and capture real-time images of the indoor flames, respectively. Both types of data were continuously transmitted back to the intelligent dispatch center. The flashover prediction module within the intelligent dispatch center analyzed the data stream in real time and issued a warning 62 seconds after the window was broken (i.e., T0+115 seconds): predicting that a flashover was highly likely to occur within the next 20 seconds.

[0113] Intelligent Decision-Making and Dry Powder Suppression: Based on the aforementioned flashover warning time, the intelligent dispatch center quickly determined that the current time window was still greater than the sum of the dry powder bomb preparation time and safety margin, therefore the system determined to enter standard operation mode. The intelligent dispatch center then issued precise instructions to the already positioned dry powder bomb drone. At T0+127 seconds (74 seconds after the window was broken), the dry powder bomb drone completed its final aiming and launch, accurately delivering a 3kg dry powder fire extinguishing bomb directly above the fire area, achieving rapid flame suppression and reducing ground radiant heat flux by 61.37%. Subsequently, continuous monitoring by reconnaissance drones showed that the flashover risk warning was not triggered again.

[0114] Foam Continuous Cooling and Extinguishing: After the dry powder suppression action is completed, the intelligent dispatch center instructs the foam fire extinguishing drone to start according to the preset sequence. At T0+403 seconds, the foam extinguishing agent is delivered to the floor height of the fire via the mooring pipe and continuously sprayed into the room at a flow rate of 2L / s for about 30 seconds, ultimately achieving complete extinguishing of the fire.

[0115] Throughout the firefighting process, the ground radiant heat flux (a key parameter directly reflecting fire intensity), recorded by heat flux measuring points deployed inside the burning floor, showed a sharp decrease in heat flux after the dry powder bomb was applied. During the continuous foam spraying phase, the heat flux further decreased below the safe threshold. This data objectively confirms that the collaborative combat mode of "instantaneous dry powder suppression - long-term foam cooling" of the UAV swarm in this embodiment achieved significant firefighting effectiveness, successfully preventing flashover and achieving rapid and thorough fire extinguishing.

[0116] Example 1: Standard Operating Procedure with Ample Time Window In this embodiment, an indoor fire occurring on the 30th floor of a high-rise building is used as the application scenario. In the initial stage of the fire, a reconnaissance drone arrives at the exterior facade of the target floor and uses its onboard infrared thermal imager to collect infrared thermal images of the unexposed side of the exterior window glass. This thermal image is input into an image recognition model pre-trained based on a residual network, and the model quickly locates the spatial coordinates of the indoor fire source area. Subsequently, the dispatch center controls a glass-breaking drone to fly to the target exterior window location and launch breaching bombs to shatter the glass and create a fire extinguishing path.

[0117] The instant the window breaking was completed (denoted as time T0), the fire environment began to change drastically due to the influx of fresh air. At this moment, the reconnaissance drone used its camera to confirm that the outer window glass had completely detached, and then used its onboard camera to directly acquire real-time images of the flames inside to extract the heat release rate. Simultaneously, it used a methane laser remote sensing instrument to monitor the concentration of methane in the smoke escaping from the broken window. The intelligent dispatch center used these two data points as time series inputs and, through its built-in Long Short-Term Memory (LSTM) prediction model, calculated that the time from the current moment to the expected flashover was 120 seconds.

[0118] The system's preset preparation time for dry powder extinguishing operations is 90 seconds, with a safety margin of 10 seconds. Since the predicted flashover time (120 seconds) is greater than the sum of the preparation time and the safety margin (100 seconds), the intelligent dispatch center determines that there is a sufficient operational window at the current fire scene and automatically switches to standard operation mode. Within the first time window after the window is broken, the dry powder fire extinguishing drone is controlled to fly above the fire source and, based on a pre-set fire extinguishing efficiency optimization model, precisely delivers 3 kg of dry powder extinguishing agent for instantaneous chemical suppression. After a fixed delay of 60 seconds following the action of the dry powder extinguishing agent, the system enters the second time window. At this time, the tethered foam fire extinguishing drone is activated, drawing ground supply through the tethered pipeline and continuously spraying foam extinguishing agent into the fire source area at a preset optimal flow rate of 2 liters / second until the fire is completely extinguished.

[0119] Example 2: Emergency Operation Mode for Extremely Rapid Fire Spread In this embodiment, a high-risk application scenario is presented, where a fire, due to severe heat accumulation in the early stages, is highly susceptible to rapid loss of control after a window is broken. The non-contact positioning and window-breaking actions before breaking the window are the same as in Embodiment 1. However, after the window-breaking drone shatters the outer window glass, data transmitted back by the reconnaissance drone shows an exponential increase in the indoor heat release rate and a sharp rise in the methane concentration in the overflowing smoke. The Long Short-Term Memory (LSTM) prediction model, after rapid calculation, estimates that only 90 seconds remain before a flashover occurs.

[0120] At this point, the flashover prediction time (90 seconds) was less than the sum of the system's preset preparation time and safety margin (100 seconds). To prevent the fire from spiraling out of control during the dry powder drone's flight and targeting, the intelligent dispatch center decisively switched to emergency operation mode. In this mode, the system directly skipped the dry powder suppression step, closed the first time window, and issued the highest priority start command to the tethered foam fire extinguishing drone. The foam fire extinguishing drone was required to forcibly establish a foam spray link within the emergency time window, from the moment the window was broken until the end of the preset maximum reaction time (100 seconds), to use a large flow of foam to powerfully cool and cover the fire source before the flashover occurred, thus ensuring the effectiveness of the fire extinguishing action under extreme conditions.

[0121] Example 3: Adaptive Sensing and Collaboration under Obstructed Demolition Conditions This embodiment focuses on complex working conditions where fire scene reconnaissance is limited by environmental constraints. When the glass-breaking drone performs a demolition mission, because the target building uses a special laminated glass material, although the demolition projectile shatters the glass, the glass mesh does not completely detach after breaking, resulting in a large amount of residue still obstructing the window frame.

[0122] At this point, the reconnaissance drone, through its onboard camera, detected that the exterior window glass was not completely detached, preventing external visibility from accurately capturing the complete shape of the indoor fire source. Based on this localized observation, the dispatch system triggered an adaptive image data source selection mechanism, automatically attempting to access the fire monitoring network inside the target building. The system successfully retrieved footage from a monitoring camera installed in a corner of the burning room, thus obtaining an unobstructed real-time image of the indoor flames. This internal image data was then fused with methane concentration data measured externally by the reconnaissance drone and jointly input into the LSTM prediction model. This adaptive perceptual redundancy design ensures that the system can still accurately calculate the flashover prediction time under extreme demolition conditions with limited external visibility, thereby supporting the multi-agent deep reinforcement learning model to make correct decisions for the next collaborative operation.

[0123] The foregoing description and accompanying drawings fully illustrate embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included or substituted for parts and features of other embodiments. Embodiments of the present disclosure are not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from its scope. The scope of the present disclosure is limited only by the appended claims, and the foregoing embodiments should be considered exemplary and non-limiting.

Claims

1. A method for coordinated firefighting by drone swarms in building fires, characterized in that, include: Control a glass-breaking drone at the location of the target building's exterior window to shatter the window glass and create a fire extinguishing path; After breaking the window, obtain fire environment information and calculate the fire flashover prediction time; The step of obtaining fire environment information and calculating the fire flashover prediction time after breaking the window includes: using a methane laser remote sensing detector carried by a reconnaissance drone to monitor the concentration of methane smoke escaping from the broken window, and obtaining real-time indoor flame images to extract the heat release rate; using the methane smoke concentration and the heat release rate as input sequences, inputting them into a pre-trained long short-term memory network prediction model, and calculating the time from the current moment until the expected flashover occurs, as the fire flashover prediction time; The fire extinguishing operation mode is determined by comparing the predicted flashover time with the preset fire extinguishing preparation time. When the predicted flashover time is greater than the sum of the fire extinguishing preparation time and the safety time margin, the dry powder fire extinguishing drone is controlled to deliver dry powder fire extinguishing agent to the indoor fire source area to suppress the flame, and after the dry powder fire extinguishing, the foam fire extinguishing drone is controlled to spray foam fire extinguishing agent to the fire source area for continuous cooling. When the predicted flashover time is less than or equal to the sum of the fire extinguishing preparation time and the safety time margin, the foam fire extinguishing drone is controlled to spray foam extinguishing agent onto the fire source area to extinguish the fire.

2. The method according to claim 1, characterized in that, Before obtaining the heat release rate from the real-time indoor flame image, the method further includes an adaptive image data source selection step: The camera of the reconnaissance drone was used to determine the detachment status of the exterior window glass after it was broken. If it is determined that the exterior window glass has completely detached, the camera is invoked to capture real-time images of the indoor flames. If it is determined that the exterior window glass has not completely detached, the dispatch system will automatically connect to the target building's internal fire monitoring network to obtain real-time indoor flame images.

3. The method according to claim 1, characterized in that, Before controlling the glass-breaking drone to shatter the window glass at the target building's exterior window location, a non-contact fire source positioning step is also included: Infrared thermal images of the unexposed side of the exterior window glass were collected using a reconnaissance drone. The infrared thermal image is input into an image recognition model pre-trained based on a residual network to extract thermal image features and output the fire source location number, thereby locking the spatial coordinates of the indoor fire source area.

4. The method according to claim 1, characterized in that, The delivery quality of dry powder extinguishing agent and the spray flow rate of foam extinguishing agent to indoor fire sources are determined by the fire extinguishing efficiency optimization model.

5. The method according to claim 1, characterized in that, When the predicted flashover time is greater than the sum of the fire extinguishing preparation time and the safety margin, Within the first time window from the moment the window is broken to the end of the fire extinguishing preparation time, the dry powder fire extinguishing drone is controlled to deliver and detonate the dry powder fire extinguishing agent. Within a second time window, starting from a fixed delay after the firefighting preparation time has ended, the foam firefighting drone is controlled to start and spray foam fire extinguishing agent.

6. The method according to claim 1, characterized in that, When the predicted flashover time is less than or equal to the sum of the fire extinguishing preparation time and the safety margin, Within the emergency time window from the moment the window is broken until the end of the preset reaction time, a start command is issued to the foam fire extinguishing drone to control it to establish a foam spraying link within the preset reaction time. The preset reaction time is shorter than the fire extinguishing preparation time.

7. The method according to any one of claims 1 to 6, characterized in that, The control actions for each drone are generated by the intelligent dispatch center based on a multi-agent deep reinforcement learning model.

8. A drone swarm collaborative firefighting system for building fires, used to implement the method of any one of claims 1 to 7, characterized in that, include: The intelligent dispatch center is used to make decisions on firefighting operation modes and issue operation instructions to the drone swarm; The reconnaissance drone is used to collect fire environment information and transmit flame image information and smoke concentration information to the intelligent dispatch center. Glass-breaking drones are used to shatter the exterior windows of target buildings under the control of the intelligent dispatch center. Dry powder fire extinguishing drones are used to deliver dry powder fire extinguishing bombs to indoor fire sources in standard operating mode. A foam fire extinguishing drone is used to spray foam fire extinguishing agent onto the fire source area.

9. The system according to claim 8, characterized in that, The reconnaissance drone is equipped with an infrared thermal imager, a camera, and a methane laser remote sensing detector.

Images