Intelligent fire extinguishing manipulator for cable-supported bridge and fire extinguishing method thereof

Abstract

The application belongs to the field of mechanical hands, and discloses a cable-supported bridge intelligent fire extinguishing mechanical hand and a fire extinguishing method thereof. The cable-supported bridge intelligent fire extinguishing mechanical hand comprises a sensing unit, a decision unit, a control unit and at least one jet device. The sensing unit is used for collecting multi-dimensional data of a fire scene of the cable-supported bridge. The decision unit has a reinforcement learning model. The control unit controls the jet device to complete fire extinguishing of the fire scene of the cable-supported bridge. The method comprises the following steps: 1) the sensing unit collects multi-dimensional data; 2) the decision unit calls the reinforcement learning model to autonomously generate an operation strategy; 3) the control unit converts the operation strategy into an action instruction of the jet device; and 4) the sensing unit feeds back multi-dimensional data in real time. The application has autonomous "sensing-decision-control" capability, meets the demand of the cable-supported bridge intelligent fire extinguishing system for efficient, accurate and safe fire extinguishing, and can adapt to high-altitude and high-risk environments.

Description

Technical Field

[0001] This invention belongs to the field of robotic arms, and more specifically, relates to an intelligent fire-fighting robotic arm for cable-stayed bridges and its fire-fighting method. Background Technology

[0002] Cable-stayed bridges (such as suspension bridges and cable-stayed bridges) serve as crucial transportation hubs spanning rivers, lakes, and seas, and are a major form of long-span bridges. Their complex structural systems typically consist of core components such as main towers, main cables, suspenders, and bridge deck paving. Due to their unique geographical locations (usually at high altitudes, in windy, and humid environments), large structural height, and high flexibility, cable-stayed bridges are highly susceptible to rapid fire spread along the cable system and bridge deck in the event of a fire (such as a vehicle accident or tanker explosion).

[0003] The fire safety of cable-stayed bridges is directly related to the smooth flow of the nation's major transportation arteries and the structural safety of critical infrastructure.

[0004] Existing fire protection solutions for cable-stayed bridges primarily rely on fire hydrant systems or fixed fire monitors located on both sides of the bridge deck. These fire-fighting devices are stationary, and their range is limited by distance. When facing cable-stayed bridges several kilometers long, blind spots are easily created. When the fire source is between two fire hydrants or in the area of ​​the main cable at high altitude, it is difficult for the fire-fighting equipment to achieve effective coverage.

[0005] Most existing fire-fighting equipment for cable-stayed bridges can only perform simple horizontal rotation and limited vertical pitch, lacking sufficient degrees of freedom. For the complex spatial geometry of cable-stayed bridges (such as dense clusters of suspension cables and blind spots under the main cable), traditional fire monitors cannot bypass obstacles to deliver precise jets, resulting in a significant waste of extinguishing agents and difficulty in reaching the core fire source.

[0006] Currently, firefighting on cable-stayed bridges still largely relies on manual operation of fire monitors or firefighters entering the fire scene with handheld water hoses. Cable-stayed bridge fires are typically accompanied by high temperatures, dense smoke, and toxic gases, and due to the extreme height of the bridge deck, firefighters face the risk of falling and the threat of structural instability.

[0007] In addition, the maintenance passage and working space of the cable-stayed bridge are narrow, making it difficult for large fire trucks to directly approach the fire source. Furthermore, manually dragging fire hoses in the complex structure at high altitudes can easily lead to entanglement and obstruction, seriously delaying the opportunity to extinguish the fire.

[0008] While some semi-automatic fire monitoring systems with integrated cameras have emerged in the existing technology, they are essentially still passive monitoring rather than proactive decision-making. Most existing systems use visible light video monitoring, which cannot accurately identify the location of the fire core and the distribution of the thermal field in dense smoke, at night, or in extreme weather conditions (such as heavy fog or heavy rain).

[0009] Currently, the jet trajectory, flow rate adjustment, and robotic arm position adjustment during fire extinguishing are mostly controlled remotely by humans via joysticks. In emergency situations, operators struggle to comprehensively consider complex factors such as wind speed deflection, fluid dynamics characteristics, and bridge structure protection within a short period, resulting in a high decision-making error rate and hindering the rapid response of "extinguishing fires as soon as they are detected."

[0010] With the development of automation technology, although some firefighting robots have attempted to apply automated control algorithms, the results have been unsatisfactory. Traditional control algorithms, based on simple models, struggle to cope with the complex and ever-changing nonlinear environment of cable-stayed bridge fires (such as the severe interference of sudden gusts of wind on the jet trajectory). Because real bridge fire scenarios are unreproducible and data acquisition is extremely costly, existing automated models are often only effective in ideal laboratory environments. When deployed in real high-altitude cable-stayed bridge environments, algorithms often fail, exhibit slow response times, or suffer from severe overshoot, resulting in an inability to accurately suppress the fire. Summary of the Invention

[0011] In response to the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides a cable-supported bridge intelligent fire extinguishing robot and its fire extinguishing method, which has autonomous "perception-decision-control" capabilities, meets the needs of cable-supported bridge intelligent fire protection systems for efficient, accurate and safe fire extinguishing, and can be adapted to high-altitude and high-risk environments.

[0012] To achieve the above objectives, according to one aspect of the present invention, an intelligent fire extinguishing robot for cable-stayed bridges is provided, comprising a sensing unit, a decision-making unit, a control unit, and at least one jet device, wherein: Each of the jetting devices includes a movable base, a base drive device, a multi-axis robotic arm, a nozzle, and a hose. The movable base is movably mounted on the cable-stayed bridge. The base drive device is connected to the movable base to drive the movable base to move. The multi-axis robotic arm is mounted on the movable base. The nozzle is mounted on the end effector of the multi-axis robotic arm. The nozzle is connected to the fire water pipe on the cable-stayed bridge via a hose. The sensing unit is used to collect multi-dimensional data of the fire scene of the cable-stayed bridge and send it to the decision-making unit; The decision-making unit has a reinforcement learning model, which is used to generate a work strategy based on the multi-dimensional data and send it to the control unit; Based on the operational strategy, the control unit controls the movement of the mobile base via the base drive device, adjusts the position and orientation of the nozzles by controlling the movement of the multi-axis robotic arm, and controls the nozzle spray to extinguish the fire at the cable-stayed bridge fire scene.

[0013] Preferably, the sensing unit includes a binocular camera, an infrared imager, a wind speed sensor, a smoke sensor, a humidity sensor, and a bridge structural stress sensor to achieve real-time, multi-dimensional data acquisition of the fire scene of the cable-stayed bridge. The multi-dimensional data includes images of the fire scene captured by the binocular camera, thermal images and temperature values ​​of the fire scene detected by the infrared imager, wind speed of the fire scene detected by the wind speed sensor, smoke concentration of the fire scene detected by the smoke sensor, humidity of the fire scene detected by the humidity sensor, and structural stress of the cable-stayed bridge detected by the bridge structural stress sensor.

[0014] Preferably, the reinforcement learning model of the decision-making unit is generated using a simulation-reality transfer method, and the decision-making unit includes a virtual environment construction module, a reinforcement learning training module, and a transfer adaptation module, wherein: The virtual environment construction module is used to simulate the structural topology, fire characteristics, and environmental interference factors of cable-stayed bridges and to construct virtual training scenarios. The reinforcement learning training module is used to evaluate the fire extinguishing behavior of the agent in a virtual training scenario using a reward mechanism, so as to optimize and generate an initial jet control and path planning model, thereby establishing a mapping relationship between multi-dimensional data of the fire scene and the operation strategy. The operation strategy includes the movement parameters of the mobile base, the motion parameters of the multi-axis robotic arm, and the spray parameters of the nozzle. The migration adaptation module establishes a cross-domain mapping relationship by aligning features of multi-dimensional data from virtual training scenarios and real fire scenes, and fine-tunes the parameters of the jet control and path planning model. It introduces multi-dimensional data measured at real fire scenes to calibrate the initial jet control and path planning model, thereby optimizing the initial jet control and path planning model into a reinforcement learning model adapted to the real fire conditions of cable-stayed bridges. The reinforcement learning model then autonomously generates operational strategies based on the multi-dimensional data.

[0015] Preferably, the specific process by which the reinforcement learning training module optimizes and generates the initial jet control and path planning model through a reward mechanism is as follows: The motion parameters of the mobile base, the motion parameters of the multi-axis robotic arm, and the spray parameters of the nozzle are used as the action space, and the fire spread rate, the peak flame temperature, and the humidity of the target area are used as status feedback indicators. A reward function is constructed based on the real-time changes of the state feedback index, and the agent is iteratively trained in the virtual training scenario using a proximal policy optimization algorithm. The parameters of the action space are evaluated based on the state feedback index through a reward mechanism, and the agent is guided to optimize the selection of the action space through the reward function, thereby generating the network weights of the jet control and path planning model.

[0016] Preferably, the virtual environment construction module further includes a preset model of the fire protection zone for cable-stayed bridges; The decision-making unit determines the target fire zone corresponding to the fire scene area of ​​the cable-stayed bridge based on the preset model of the fire zone of the cable-stayed bridge, and instructs the jet device in the fire scene area to perform fire extinguishing based on the structural topology information of the target fire zone.

[0017] Preferably, the virtual environment construction module within the decision-making unit reconstructs the structural topology, including the main cable, suspension cables, and bridge deck paving, based on the 3D point cloud data and design drawings of the cable-stayed bridge structure. Based on fire dynamics simulation results, and coupled with environmental interference factors collected by the sensing unit, a dynamic fire environment model is established. These environmental interference factors include wind speed, temperature, and visibility. The module then renders the structural topology and dynamic fire environment model to generate a 1:1 scale virtual training scene.

[0018] Preferably, the transfer adaptation module performs feature alignment and difference quantification analysis based on the multi-dimensional data of the sensing unit and the corresponding data of the virtual training scene, constructs a feature mapping model using the maximum mean difference domain adaptive transfer learning algorithm, fine-tunes and optimizes the parameters of the jet control and path planning model, calibrates the jet control and path planning model using measured data from the fire scene, and corrects operational strategy deviations caused by environmental interference and structural errors through dynamic compensation factors.

[0019] Preferably, the control unit receives the operation strategy output by the decision unit, and performs inverse kinematics calculation based on the kinematic model of the mobile base drive device, the multi-axis robotic arm and the nozzle to generate an operation strategy for the mobile base drive device, the multi-axis robotic arm and the nozzle. The control unit tracks the parameters of the actual operation of the mobile base drive device, multi-axis robotic arm and nozzle based on the PID adaptive algorithm, and dynamically corrects the operation strategy to compensate for control deviations based on the multi-dimensional data fed back by the sensing unit.

[0020] Preferably, the nozzle is a universal direct-fire nozzle, and the nozzle is equipped with a flow regulating valve and a fluid pressure sensor so that the control unit can dynamically adjust the nozzle's spray parameters based on the fire scale and the distance to the fire point. The nozzle's spray parameters include spray angle, flow rate, and spray pressure.

[0021] According to another aspect of the present invention, a method for extinguishing a fire on a cable-stayed bridge using the aforementioned intelligent fire-extinguishing robot is also provided, comprising the following steps: 1) The sensing unit collects multi-dimensional data from the fire scene of the cable-stayed bridge; 2) The decision-making unit calls upon the reinforcement learning model trained through simulation-reality transfer to autonomously generate task strategies based on multi-dimensional data; 3) The control unit converts the operation strategy into action commands for the jet device, drives the mobile base to move to the target area, and the multi-axis robotic arm adjusts the nozzle to the target position and controls the nozzle to spray to extinguish the fire; 4) The sensing unit provides real-time feedback of multi-dimensional data so that the decision-making unit can obtain the fire extinguishing effect and environmental changes. The decision-making unit dynamically adjusts the operation strategy based on the multi-dimensional data fed back by the sensing unit to form a closed-loop control until the decision-making unit determines that the fire extinguishing is completed and issues a stop command.

[0022] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1) The intelligent fire-fighting manipulator for cable-stayed bridges of this invention constructs a complete closed-loop system of perception, decision-making, control, and execution. Through the organic integration of the jet device, perception unit, decision-making unit, and control unit, the system can achieve fully autonomous operation from fire information acquisition to fire-fighting strategy generation, and then to the execution of the jet device's actions. This architecture ensures that the system can respond quickly and autonomously enter combat mode within the golden time frame of a fire, without relying on a high-risk human working environment, greatly improving the real-time performance and standardization of fire response in cable-stayed bridges.

[0023] 2) The intelligent fire extinguishing robot for cable-stayed bridges of the present invention features a movable base that is movably mounted on the cable-stayed bridge and driven by a base drive device. This technical feature allows the fire extinguishing system to move flexibly along the longitudinal direction of the cable-stayed bridge or a preset path, rather than being limited to fixed fire-fighting points. For cable-stayed bridges with extremely large spans, this mobility ensures that fire extinguishing resources can be quickly delivered to the fire point on the bridge, achieving fire extinguishing coverage on the bridge deck and structural space. The stable support of the movable base also provides a reliable benchmark for the high-precision movements of the upper multi-axis robotic arm, avoiding interference from jet reaction forces on spraying accuracy.

[0024] 3) The intelligent fire extinguishing robot for cable-stayed bridges of this invention features a multi-axis robotic arm design that grants the system exceptional adjustability in three-dimensional space. Cable-stayed bridges have complex structures, including main cables, suspenders, and intricate bridge deck paving. The multi-axis robotic arm can navigate around obstacles such as suspenders and guardrails, delivering the nozzle to the optimal jet position at the fire source through the coordinated rotation and extension of multiple joints. This flexible end-effector positioning capability ensures that the extinguishing medium accurately covers the core area of ​​the fire source, especially in blind spots that are difficult for traditional firefighting methods to reach, such as at high altitudes or on the sides of bridges, demonstrating excellent spatial adaptability.

[0025] 4) The intelligent fire extinguishing robot of the cable-stayed bridge of the present invention has its nozzle connected to the fire water pipe on the cable-stayed bridge via a flexible hose. This water supply method not only utilizes the existing municipal fire protection infrastructure of the cable-stayed bridge to ensure a continuous supply of fire extinguishing medium, but also overcomes the problems of pipe entanglement and restriction during the large-scale movement of the multi-axis robot arm through the design of the flexible hose, ensuring the stability of fluid transmission in the dynamic fire extinguishing process.

[0026] 5) The cable-stayed bridge intelligent fire extinguishing robot of this invention provides the decision-making unit with a wealth of information sources through the multi-dimensional data collected by the sensing unit, including but not limited to heat source distribution, smoke diffusion trajectory, environmental wind speed vector, and the stress state of the structure itself. This comprehensive and precise data sensing capability reduces the information uncertainty caused by environmental interference, enabling the system to grasp the dynamic evolution characteristics of the fire in real time, laying a solid digital foundation for the subsequent generation of scientific operation strategies.

[0027] 6) The cable-stayed bridge intelligent fire extinguishing robot of this invention uses a reinforcement learning model that can automatically establish a complex nonlinear mapping relationship between multi-dimensional fire data and operational strategies (such as the moving speed of the mobile base, the robot arm's posture, and the spray pressure). This means that the operational strategy generated by the decision-making unit is not a fixed script, but an optimal solution that can be dynamically adjusted according to feedback indicators such as the fire spread rate and temperature changes, demonstrating a strong autonomous evolution capability. Attached Figure Description

[0028] Figure 1 This is a structural schematic diagram of the intelligent fire extinguishing robot for cable-stayed bridges of the present invention; Figure 2 This is a schematic diagram showing the arrangement of the intelligent fire extinguishing robot on the cable-stayed bridge according to the present invention. Figure 3 This is a flowchart of the spraying method of the intelligent fire extinguishing robot for cable-stayed bridges according to the present invention; In all the accompanying drawings, the same reference numerals denote the same technical features, specifically: 100. Cable-supported intelligent fire extinguishing robot; 1. Mobile base; 2. Multi-axis robotic arm; 3. Nozzle; 4. Hose; 200. Fire hose; 300. Water pump; 400. Cable-supported bridge. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0030] Reference Figure 1, Figure 2 The cable bridge intelligent fire extinguishing robot 100 includes a sensing unit, a decision-making unit, a control unit, and at least one jet device, wherein: Each jet spray device includes a movable base 1, a base drive device, a multi-axis robotic arm 2, a nozzle 3, and a hose 4. The movable base 1 is movably mounted on the cable-stayed bridge 400. The base drive device is connected to the movable base 1 to drive its movement. The movable base 1 can be directly or indirectly mounted on the cable-stayed bridge 400 via rails, wheels, or cableways, and moves under the drive of the base drive device. The multi-axis robotic arm 2 is mounted on the movable base 1, and the nozzle 3 is mounted on the end effector of the multi-axis robotic arm 2. The nozzle 3 is connected to a fire hose 200 on the cable-stayed bridge 400 via the hose 4. The fire hose 200 is connected to a water pump 300, and the water pump 300 is connected to a water source. A certain amount of slack is allowed during the installation of the hose 4 to prevent the multi-axis robotic arm 2 from becoming tangled during movement.

[0031] The sensing unit is used to collect multi-dimensional data of the fire scene of the cable-stayed bridge 400 and send it to the decision-making unit; The decision-making unit has a reinforcement learning model, which is used to generate a work strategy based on the multi-dimensional data and send it to the control unit; The control unit, based on the operational strategy, controls the movement of the mobile base 1 via a base drive device (the movement of the mobile base 1 is mainly to adjust the position of the nozzle 3, allowing the nozzle 3 to quickly reach the vicinity of the fire), adjusts the posture (position and orientation) of the nozzle 3 by controlling the movement of the multi-axis robotic arm 2, and controls the spraying of the nozzle 3 (the spraying parameters of the nozzle 3 are adjustable) to complete the fire extinguishing at the fire scene of the cable-stayed bridge 400. The operational strategy includes the movement parameters (speed, displacement) of the mobile base 1, the motion parameters (rotation parameters of each joint motor of the multi-axis robotic arm 2, flexion and extension of each arm body, etc.), and the spraying parameters (spraying angle, flow rate, and spraying pressure of the nozzle 3, etc.). Therefore, the mobile base 1, multi-axis robotic arm 2, and nozzle 3 perform operations based on the operational strategy.

[0032] The multi-axis robotic arm 2 adopts a multi-section rotatable connection arm design. It includes joint motors, harmonic reducers, rotating shafts, and sensors, which are the core components for power transmission and control of each arm section. The sensors provide real-time feedback of rotational speed and angle information to meet the adjustment requirements of firefighting operations. Through the flexion, extension, and rotation of each arm section, the nozzle 3 can move and adjust its position in three-dimensional space. The flexible hose 4 is laid along the multi-axis robotic arm 2, connecting to the cable-stayed bridge 400 fire water pipe at the inlet and to the nozzle 3 at the outlet, allowing the delivery of extinguishing media during the movement of the multi-axis robotic arm 2. The outer surface of the jet device is coated with a heat-insulating layer to ensure normal operation near fire sources or at high temperatures.

[0033] This invention, through the flexible architecture of the jet device, the all-round information acquisition of the sensing unit, the intelligent brain construction of the decision-making unit, and the precise power output of the control unit, together constitutes an intelligent and autonomous fire-fighting operation system that can adapt to the structure and complex environment of the cable-stayed bridge 400. It is not only reflected in the improvement of fire-fighting speed and accuracy, but also in the in-depth maintenance of bridge structural safety and the transformation of traditional high-risk operation mode.

[0034] Furthermore, the sensing unit includes a binocular camera, an infrared imager, a wind speed sensor, a smoke sensor, a humidity sensor, and a bridge structural stress sensor to achieve real-time, multi-dimensional data acquisition of the fire scene of the cable-stayed bridge 400. The multi-dimensional data includes images of the fire scene captured by the binocular camera, thermal images and temperature values ​​of the fire scene detected by the infrared imager, wind speed of the fire scene detected by the wind speed sensor, smoke concentration of the fire scene detected by the smoke sensor, humidity of the fire scene detected by the humidity sensor, and structural stress of the cable-stayed bridge 400 at the fire scene detected by the bridge structural stress sensor. The binocular camera can be a high-definition camera.

[0035] A comprehensive, three-dimensional monitoring network was constructed using binocular cameras, infrared imagers, wind speed sensors, smoke sensors, humidity sensors, and bridge structural stress sensors. This demonstrates the in-depth analysis of the fire scene environment through the fusion of multi-source data.

[0036] Binocular cameras provide high-resolution visible light information to identify the specific shape of flames, their direction of spread, and visual details of the fire scene. This enables the system to establish intuitive visual feedback, assisting the decision-making unit in accurately identifying the spatial relationship between the fire ignition point and surrounding obstacles.

[0037] Infrared imagers have the ability to penetrate dense smoke, enabling them to locate high-temperature heat sources and obtain real-time, accurate temperature values ​​in fire scenes with extremely low visibility. This thermal perception ensures that the system can accurately pinpoint the fire core even in the early stages of a fire or when smoke is abundant, avoiding misjudgments caused by visual obstruction and significantly improving the robustness of fire source location.

[0038] Cable-stayed bridges 400 are typically located at high altitudes or across the sea, where wind is a key variable affecting the jet trajectory of nozzle 3. Wind speed sensors collect real-time wind speed data from the fire scene, providing crucial compensation parameters for the decision-making unit. This allows the system to predict the deviation of water flow or extinguishing agent, thereby dynamically adjusting the posture of the multi-axis robotic arm 2 or the spray parameters of nozzle 3 to ensure that the extinguishing agent accurately hits the fire point, rather than being dispersed by strong winds.

[0039] By monitoring the smoke concentration at the fire scene, the system can quantitatively assess the intensity and spread stage of the fire, providing important environmental state feedback indicators for reinforcement learning models and helping the system select the optimal path planning strategy under different visibility and smoke conditions.

[0040] When the humidity in the target area reaches a preset saturation value and the temperature drops, the decision-making unit can more accurately determine that fire extinguishing is complete, thereby issuing a stop command and avoiding water waste. Fires on cable-stayed bridges are typically accompanied by high temperatures and strong winds. Humidity sensors can monitor the distribution of water vapor around the spray path and target point, helping the system determine whether the extinguishing medium has vaporized significantly due to high temperatures before reaching the fire source, thus guiding the control unit to adjust the spray pressure or flow rate. By monitoring the air humidity around the fire scene, the system can assess the risk of fire re-spreading. High humidity environments can effectively suppress secondary ignition caused by sparks.

[0041] The safety of the cable-stayed bridge 400 is highly dependent on the mechanical state of the main cable and suspenders; high temperatures can easily cause structural instability. Bridge structural stress sensors can monitor stress changes in the cable-stayed bridge 400 structure in real time during a fire. When abnormal stress is detected in a specific area, the system can prioritize the use of jet spray devices to cool and protect critical load-bearing components. This targeted firefighting strategy effectively prevents the risk of bridge structural collapse due to heat, elevating the firefighting task from simply extinguishing the fire to ensuring structural integrity.

[0042] Multi-dimensional data provides decision-making units with comprehensive digital input. This rich data dimension is key to achieving simulation-to-reality transfer training because it allows the system to recreate a dynamic fire environment model in a virtual environment, including temperature, wind force, and stress, thereby reducing the performance differences between virtual and reality.

[0043] The real-time data collected by the sensing unit is transmitted to the control unit through a feedback loop, enabling the system to dynamically correct the movement parameters of the mobile base 1 and the motion parameters of the multi-axis robotic arm 2 at the millisecond level using PID adaptive algorithms or other control logic. This closed-loop adjustment based on real-world perception greatly improves the system's adaptability in the face of complex and ever-changing bridge fire scenarios.

[0044] In summary, by integrating five key sensors, the cable-driven bridge intelligent fire-fighting robot 100 possesses perception capabilities that surpass traditional detection methods. This not only ensures the accuracy of fire source location but also achieves dual optimization of fire-fighting efficiency and bridge structural safety protection through real-time monitoring of environmental wind speed and structural stress. This provides comprehensive logical support for subsequent decision-making units to generate scientific and efficient operational strategies.

[0045] Furthermore, the reinforcement learning model of the decision-making unit is generated using the Sim-to-Real method in reinforcement learning, and the decision-making unit includes a virtual environment construction module, a reinforcement learning training module, and a transfer adaptation module, wherein: The virtual environment construction module is used to simulate the structural topology, fire characteristics, and environmental interference factors of the cable-stayed bridge 400 and to construct a virtual training scenario. The reinforcement learning training module is used to evaluate the fire extinguishing behavior of the agent in a virtual training scenario using a reward mechanism, so as to optimize and generate an initial jet control and path planning model, thereby establishing a mapping relationship between multi-dimensional data of the fire scene and the operation strategy. The operation strategy includes the movement parameters of the mobile base 1, the motion parameters of the multi-axis robotic arm 2, and the spray parameters of the nozzle 3. The migration adaptation module establishes a cross-domain mapping relationship by aligning features of multi-dimensional data from virtual training scenarios and real fire scenes, and fine-tunes the parameters of the jet control and path planning model. It introduces multi-dimensional data measured at real fire scenes to calibrate the initial jet control and path planning model, thereby optimizing the initial jet control and path planning model into a reinforcement learning model adapted to the real fire conditions of cable-stayed bridges. The reinforcement learning model then autonomously generates the optimal operation strategy based on the multi-dimensional data to achieve precise execution of fire extinguishing actions in the real environment.

[0046] By specifically defining the internal structure of the decision-making unit, a complete intellectual closed loop is constructed, from virtual simulation to algorithm optimization and then to real-world implementation.

[0047] The virtual environment construction module is used to simulate the structural topology, fire characteristics, and environmental interference factors of the cable-stayed bridge 400 and construct a virtual training scenario. By recreating the complex structure including the main cable, suspenders, and bridge deck paving, the system can simulate various extreme fire characteristics without depleting resources on the actual cable-stayed bridge 400. This high-fidelity modeling capability ensures that subsequent algorithm training has a sufficiently broad sample space to cover rare disaster conditions that are extremely difficult to occur in reality. By integrating environmental interference factors such as wind speed, temperature, and visibility into the virtual scene, the system can learn correction logic under severe weather conditions in advance, significantly improving the system's resilience against interference in real-world scenarios.

[0048] The reinforcement learning training module can automatically optimize and generate initial jet control and path planning models. Through a reward mechanism, the system can autonomously explore and lock onto the operational path that can most quickly reduce the heat load of the fire, thereby establishing a precise mapping relationship between multi-dimensional data and operational strategies (including the movement parameters of the mobile base 1, the motion parameters of the multi-axis robotic arm 2, and the spray parameters of the nozzle 3). This learning-based control method enables the system to generate more predictive response strategies when facing nonlinear fire spread patterns, ensuring that various operational strategies are always in the optimal combination when dealing with changing fire conditions.

[0049] The transfer adaptation module performs feature mapping and parameter fine-tuning based on multi-dimensional data from virtual training scenarios and real-world fire scene data. This module optimizes the original jet control and path planning model into a reinforcement learning model trained through simulation-reality transfer. Through feature mapping, the system can identify subtle differences between the simulated environment and the real scene and perform targeted parameter fine-tuning, effectively solving the problem of algorithm incompatibility under complex real-world conditions. By using dynamic compensation factors to correct deviations caused by environmental interference and structural errors, the model maintains the high accuracy and high response speed exhibited in the virtual experiment when deployed on a real cable-stayed bridge 400, truly achieving robust transfer of embodied intelligence from cloud algorithms to physical equipment.

[0050] By leveraging the collaborative work of virtual environment construction modules, reinforcement learning training modules, and transfer adaptation modules, not only are training challenges solved, but also the reliability challenge of transitioning from simulation to real-world application. This architecture endows the cable-stayed bridge intelligent fire-fighting robot 100 with exceptional adaptive evolution capabilities. It can generate customized and field-tested operational strategies for each cable-stayed bridge 400 with a specific structure, achieving efficient, intelligent, and precise handling of fires on large-span bridges while ensuring personnel safety.

[0051] Furthermore, the specific process by which the reinforcement learning training module optimizes and generates the initial jet control and path planning model through a reward mechanism is as follows: The motion parameters of the mobile base 1, the motion parameters of the multi-axis robotic arm 2, and the spray parameters of the nozzle 3 are used as the action space, and the fire spread rate, the peak flame temperature, and the humidity of the target area are used as the status feedback indicators. A reward function is constructed based on the real-time changes of the state feedback index, and the agent is iteratively trained in the virtual training scenario using a proximal policy optimization algorithm. The reward function guides the agent to optimize the selection of the action space, thereby generating the network weights of the jet control and path planning model.

[0052] By combining the movement parameters of the mobile base 1, the motion parameters of the multi-axis robotic arm 2, and the spray parameters of the nozzle 3 into a single motion space, the jet control and path planning model is no longer a simple mechanical control, but rather achieves deep coupling from the macroscopic deployment of the bridge (mobile base 1) to the microscopic end-effector execution (multi-axis robotic arm 2 and nozzle 3). This simultaneous optimization of multi-dimensional parameters ensures that the system can calculate the optimal combination of movement, posture, and jet flow when facing dynamic fire situations, enabling firefighting operations to achieve extremely high flexibility and coverage efficiency in three-dimensional space. Through motion simulation of the nozzle 3's spray parameters, the system can learn how to adjust flow rate and pressure to counteract environmental interference at different distances and angles, thereby significantly improving the accuracy of the extinguishing medium hitting the fire core.

[0053] The fire spread rate, used as a feedback indicator, enables the model to predict the fire's trajectory, resulting in proactive operational strategies that can block fire spread paths in advance. Peak flame temperature is the most direct indicator for assessing the threat level to the cable-stayed bridge 400 structure. Incorporating peak flame temperature into the feedback system guides the model to prioritize high-temperature areas threatening key load-bearing components such as the main cable and suspenders during training, optimizing fire suppression logic from a safety perspective. The addition of target area humidity provides a quantitative standard for evaluating fire suppression success, allowing the reward mechanism to more objectively assess the agent's fire suppression effectiveness, avoiding indiscriminate spraying, and achieving efficient utilization of fire-fighting resources.

[0054] The near-end policy optimization algorithm, by limiting the policy update magnitude, avoids drastic oscillations in network weights during training, ensuring that the jet control and path planning model can steadily converge to the optimal solution. This is particularly important in highly dynamic and nonlinear training tasks such as fire simulation, guaranteeing the reliability of the final model under various complex logics. Through millions of iterations of the agent in the virtual training scenario, the system can uncover extremely subtle nonlinear relationships between multi-dimensional data and complex action spaces. The learned network weights contain fire-fighting patterns that are difficult for human experts to summarize, especially demonstrating an adaptive capability far exceeding traditional preset scripts when dealing with extreme conditions such as high-altitude strong winds and narrow working surfaces.

[0055] By strictly limiting the training parameters and algorithms, the Sochengqiao Intelligent Firefighting Robot 100 shortens the response delay of the jet control and path planning model. Furthermore, through a scientific reward mechanism and near-end strategy optimization algorithm, it ensures that the system can execute scientific firefighting decisions with robust actions in real Sochengqiao 400 fire scenes, greatly improving the safety margin and effectiveness of Sochengqiao 400 firefighting operations.

[0056] Furthermore, the virtual environment construction module integrates a preset model of the 400 fire-resistant compartment of the cable-stayed bridge; The decision-making unit determines the target fire zone corresponding to the fire scene area of ​​the cable-stayed bridge 400 based on the preset model of the fire zone of the cable-stayed bridge 400, and enables the multi-jet device in the fire scene area to perform fire extinguishing based on the structural topology information of the target fire zone.

[0057] By integrating the Suochengqiao 400 fire compartment preset model into the virtual environment construction module, the system establishes a predefined spatial logical benchmark. When the sensing unit collects data on the fire scene area, the decision-making unit can immediately identify the corresponding target fire compartment based on the Suochengqiao 400 fire compartment preset model. This compartment-based indexing mechanism significantly shortens the computational delay from fire identification to spatial positioning, ensuring that the system can enter the correct operational logic plane in the early stages of a fire. The Suochengqiao 400 fire compartment preset model defines clear boundaries for firefighting operations, preventing jets from causing unnecessary loads or water damage to structures in non-fire areas, achieving efficient fire control through compartmentalized management and precise containment.

[0058] Multiple jet sprayers are preferably installed, allowing the decision-making unit to dispatch them for saturation spraying or multi-point containment based on the size of the target fire zone. This cluster coordination capability significantly improves the coverage intensity and cooling rate of the extinguishing medium when facing a large-area fire on the 400-meter-long cable-stayed bridge. By sharing the same set of structural topology information, multiple jet sprayers can achieve complementary spatial positioning; for example, some sprayers can suppress the core fire source while others can cool and protect surrounding components, forming a three-dimensional integrated operational system.

[0059] The structural topology information includes the detailed spatial relationships of the main cables, suspenders, and bridge deck paving within the zone. When generating operational strategies, the decision-making unit can effectively guide the multi-axis robotic arm 2 to automatically avoid complex cable structures during movement, preventing mechanical collisions or pipe entanglement and ensuring system reliability. Because the system possesses the structural topology information, it can prioritize heat-sensitive components (such as high-tension suspenders) during operations, automatically optimizing jet landing points during firefighting missions. Through precise multi-point jet coverage within the zone, it maximizes the protection of the bridge's main structural mechanical stability.

[0060] Because the virtual environment construction module integrates a pre-set model of the 400 cable-stayed bridge fire compartments that is consistent with the actual bridge, the training results generated in the virtual training scenario can be seamlessly applied to each compartment in reality. This compartmentalized training and application mode significantly reduces errors in cross-scenario operations. The compartmentalized decision-making mode enables the system to examine the evolution of the fire scene area from a global perspective. By uniformly allocating tasks to multiple jet devices, it avoids jet blind spots that may occur in single-machine operations, achieving a highly deterministic handling effect for fires on the 400 cable-stayed bridge.

[0061] By introducing a pre-set model of the fire compartments of the cable-stayed bridge 400, the previously isolated and disorganized firefighting actions have been upgraded into a systematic defense system based on spatial topology logic and multi-machine collaboration. This not only greatly improves the efficiency of the cable-stayed bridge intelligent firefighting robot 100 in handling fires in large transportation infrastructures, but also achieves proactive safety protection for complex bridge structures while ensuring firefighting effectiveness.

[0062] Furthermore, the virtual environment construction module within the decision-making unit, based on the 3D point cloud data and design drawings of the cable-stayed bridge 400 structure, reconstructs the structural topology including the main cable, suspension cables, and bridge deck paving. Based on fire dynamics simulation results, and after coupling environmental interference factors collected by the sensing unit, a dynamic fire environment model is established. The environmental interference factors include wind speed, temperature, and visibility. The structural topology and dynamic fire environment model are then rendered to generate a 1:1 scale virtual training scene.

[0063] By combining the measured accuracy of 3D point cloud data with the logical constraints of design drawings, the system can create a digital twin that perfectly matches the dimensions of the actual cable-stayed bridge 400. This precise reconstruction of the structural topology ensures that the multi-axis robotic arm 2 can identify minute spatial blind spots and obstacles during virtual training. Because the system accurately grasps the spatial coordinates of the main cable and suspenders, the path planning generated by the decision unit can effectively avoid the cable-stayed structure, effectively preventing collisions between the mechanical structure and critical load-bearing components of the bridge in extreme fire conditions with poor visibility, thus ensuring the safety of equipment operation.

[0064] The fire dynamics simulation results not only simulate the visual appearance of flames, but more importantly, they simulate the conduction and radiation of heat between the metal and concrete structures of the cable-stayed bridge 400. This allows the reinforcement learning model to learn to take cooling strategies in advance before the fire spreads to critical components (such as the suspension cables). The dynamic fire environment model provides realistic interactive feedback for the extinguishing jet, and the system can simulate the vaporization or displacement process of the extinguishing medium under different heat flux intensities, thereby optimizing the duration and intensity distribution of the jet.

[0065] The cable-stayed bridge 400 is typically located in environments with strong winds and high humidity, factors that drastically alter the parabolic trajectory of the nozzle 3 jet. By introducing these environmental disturbances in real-time into the virtual scene, the agent can learn how to adjust the angular offset of the multi-axis robotic arm 2 under different wind speeds and how to rely on multi-dimensional data for positioning in low-visibility environments. The system couples environmental disturbances collected by the sensing unit, including wind speed, temperature, and visibility, into the model. This coupling mechanism makes the virtual training scene no longer static but a dynamic system that changes with fluctuations in the real environment, ensuring that the final generated operation strategy has strong environmental adaptability.

[0066] The 1:1 scale ensures that feedback in the virtual world (such as the recoil caused by a water pump operating at 300 pressure, and the gravitational torque of the robotic arm) is perfectly aligned with the real world. This seamless simulation environment allows the network weights generated by the reinforcement learning model to directly act on real hardware, achieving a highly closed loop from software definition to hardware execution.

[0067] The deep coupling of multi-source data and high-fidelity rendering significantly improve the decision-making accuracy of the intelligent agent. More importantly, the accurate simulation of key variables such as wind speed and structural topology ensures the reliability of the system in the real extreme fire scenario of a cable-stayed bridge.

[0068] Furthermore, the transfer adaptation module performs feature alignment and difference quantification analysis based on the multi-dimensional data of the perception unit and the corresponding data of the virtual training scene. It constructs a feature mapping model using the maximum mean difference domain adaptive transfer learning algorithm to fine-tune and optimize the parameters of the jet control and path planning model. It also calibrates the jet control and path planning model using measured data from the fire scene and corrects operational strategy deviations caused by environmental interference and structural errors through dynamic compensation factors.

[0069] By comparing physical quantities in the virtual simulation with the raw signals collected by the real-world sensing units, the system can accurately identify subtle differences in data distribution between the simulated environment and the actual fire scene. This quantitative analysis enables the decision-making unit to understand the characteristic relationships under different environments, ensuring that advanced firefighting strategies trained in the virtual environment can be losslessly transferred to real hardware, avoiding algorithm failures in real-world applications. This alignment mechanism effectively filters out environmental noise, ensuring that the system's understanding of the fire scene remains consistent with the logical dimension of the virtual training scenario, thereby greatly improving the determinism of firefighting operations.

[0070] A feature mapping model was constructed using a maximum mean difference domain adaptive transfer learning algorithm. This algorithm minimizes the distribution distance between the virtual and real domains through mathematical means, enabling the jet control and path planning model to quickly adapt to the characteristics of the real cable-stayed bridge 400 while retaining the training intelligence. This gives the system a strong generalization ability; even if the actual fire scene conditions are not exactly the same as the simulated scenario, the system can still output efficient operational strategies through the feature mapping model. By fine-tuning and optimizing the parameters of the jet control and path planning model, the system can be personalized for specific bridge structural characteristics (such as specific vibration frequencies or obstruction features). This fine-tuning ensures that the motion control of the multi-axis robotic arm 2 and the jet parameters of the nozzle 3 achieve the optimal solution under the specific working conditions, significantly improving the accuracy of the fire extinguishing response.

[0071] The introduction of on-site measured data from the fire scene enables the system to perform on-site corrections. It can perform closed-loop calibration for minor displacements or sensor zero-point drift caused by long-term bridge operation, ensuring the jet control and path planning model remains in a high-precision operating state. Dynamic compensation factors correct operational strategy deviations caused by environmental disturbances (such as extreme instantaneous gusts of wind and water pressure fluctuations due to sudden temperature changes) and structural errors (such as deviations in actual bridge construction and minor differences in cable installation). The technical effect is that the system can adjust the joint compensation angle of the multi-axis robotic arm 2 or the spray pressure of the nozzle 3 in real time, ensuring that the jet landing point accurately covers the fire source even under complex disturbances, achieving extremely high environmental robustness. The corrected operational strategy generates smoother, more predictable control commands, reducing mechanical wear on the base drive device and multi-axis robotic arm 2, and extending the equipment's service life under harsh conditions.

[0072] Furthermore, the control unit receives the operation strategy output by the decision unit, and performs inverse kinematics calculation based on the kinematic model of the mobile base drive device, the multi-axis robotic arm and the nozzle to generate an operation strategy for the mobile base drive device, the multi-axis robotic arm and the nozzle. The control unit tracks the parameters of the actual operation of the mobile base drive device, multi-axis robotic arm and nozzle based on the PID adaptive algorithm, and dynamically corrects the operation strategy to compensate for control deviations based on the multi-dimensional data fed back by the sensing unit.

[0073] The application of the PID adaptive algorithm enables the system to simultaneously process parameters such as the displacement of the moving base drive device, the motion of the multi-axis robotic arm 2, and the flow rate and pressure of the nozzle 3. It can automatically adjust proportional, integral, and derivative parameters to cope with the complex mechanical load fluctuations during high-altitude operations on the cable-stayed bridge 400m high-altitude, ensuring that each actuator maintains smooth motion while responding at high speeds, thus avoiding secondary impacts on the bridge structure caused by mechanical vibrations. This feature enables the precise mapping of complex operational strategies generated by the decision-making unit (such as obstacle avoidance paths and jet parabolic compensation) into specific electrical signals. This conversion capability ensures that the superior performance of the intelligent algorithm is fully reflected in the operation of the moving base drive device, multi-axis robotic arm 2, and nozzle 3, achieving strict constraints and drive from the control level to the execution level output.

[0074] The sensing unit provides multi-dimensional data including real-time wind speed, temperature, image data, and stress information. The control unit uses this data for dynamic correction, enabling real-time mitigation of environmental interference. For example, when the wind speed sensor detects a sudden increase in lateral wind force causing the nozzle 3 jet to deviate from the center of the fire source, the control unit immediately adjusts the end effector position of the multi-axis robotic arm 2 or increases the spray pressure of the nozzle 3, thereby maintaining the effective hit rate of the extinguishing medium at the fire scene. By comparing the real-time status feedback of the actuators with the original operating strategy, the system forms a complete closed-loop control logic. This dynamic correction mechanism can compensate for control deviations caused by mechanical wear, sensor measurement drift, or minor structural deformations, ensuring that the system maintains centimeter-level positioning accuracy and jet control precision even within a 400-meter-deep operating range on a cable-stayed bridge spanning hundreds of meters.

[0075] In the complex operating environment of the cable-stayed bridge 400, even minor deviations in a single action, if not compensated for in a timely manner, can accumulate during the continuous movements of the multi-axis robotic arm 2. By continuously and dynamically correcting the operational strategy, the endpoint state of each action command is ensured to highly coincide with the preset operational strategy, greatly improving the system's execution certainty under extreme fire conditions. For the sensitive main cable and suspenders of the cable-stayed bridge 400, this deviation compensation capability of the control unit ensures that the spraying action will not accidentally damage the supported components due to mechanical delays or calculation errors, but rather can always accurately guide the extinguishing medium to the target area, achieving comprehensive intelligent protection for the bridge's main structure.

[0076] Because the control unit can precisely control the spray parameters of the nozzle 3 and correct deviations in real time, the extinguishing medium (such as fire water or foam) can act on the core of the fire point with maximum efficiency, reducing ineffective spraying. The PID adaptive algorithm reduces the system's settling time, which significantly shortens the time it takes for the multi-axis robotic arm 2 to reach the target spraying position from a stationary state.

[0077] Through PID adaptive algorithm and dynamic correction based on multi-dimensional data, the cable-driven bridge intelligent fire extinguishing robot 100 is provided with a precise, stable and self-healing dynamic control base. This not only ensures the implementation of decision commands, but also enables the equipment to exhibit excellent execution reliability and environmental resilience in the extremely complex bridge fire environment through real-time control deviation compensation.

[0078] Furthermore, the nozzle 3 is a universal direct-fire nozzle, and a flow regulating valve and a fluid pressure sensor are installed on the nozzle so that the control unit can dynamically adjust the spray parameters of the nozzle 3 based on the fire scale and the distance to the fire point.

[0079] The omnidirectional direct-fire nozzle's built-in power mechanism enables 360° rotation and pitch angle adjustment. In a space like the cable-stayed bridge 400, which contains numerous slender components such as main cables and slings, the omnidirectional direct-fire nozzle, in conjunction with the multi-axis robotic arm 2, can achieve millisecond-level fine-tuning of the fire source's impact point. This ensures the jet can penetrate complex structural gaps and precisely cover fire points hidden behind the components. By adjusting the spray angle, the system can compensate in real time for fluid deflection caused by crosswinds on the bridge deck, ensuring the actual impact point of the extinguishing medium remains locked in the core area of ​​the fire source.

[0080] By integrating a flow control valve and a fluid pressure sensor, the end effector achieves a technological leap from indiscriminate spraying to quantitative jet delivery. The fluid pressure sensor monitors the fluid dynamics in real time, providing precise feedback data to the control unit. The system can calculate the accurate parabolic trajectory of the jet based on real-time pressure, thereby driving the flow control valve to ensure that the output extinguishing medium has sufficient momentum to penetrate the rising airflow of the flames. The introduction of the fluid pressure sensor enables the system to sense fluctuations in the fire water supply system in real time and maintain a constant spray intensity through adaptive adjustments, avoiding jets that are too close due to insufficient pressure or media wastage due to excessive pressure.

[0081] The control unit dynamically adjusts the spray parameters of nozzle 3 based on the fire scale and distance to the fire point, enabling the system to automatically switch extinguishing modes according to the fire scale. For small initial fire points, the flow rate is reduced through the flow regulating valve to save water and cool the flames; for large-scale and intense fires, the maximum flow rate and spray pressure are activated instantly to suppress the spread of flames with high-intensity impact. The control unit calculates the optimal spray pressure in real time based on the distance to the fire point. This dynamic matching ensures that the jet maintains the best convergence shape regardless of whether the fire source is located near or far from the cable-stayed bridge 400, greatly improving the effective utilization rate of the extinguishing medium and the speed of fire control.

[0082] By dynamically adjusting the jet pressure, the system can prevent the high-pressure jet from directly impacting the outer protective layer of the main cable or suspenders, which has weakened after heating. This achieves cooling and fire suppression while maximizing the protection of the coating and structural integrity of the components. Combined with precise flow control, the system can implement differentiated cooling strategies for different heated areas, effectively preventing thermal stress caused by excessively rapid local cooling from adversely affecting the bridge structure.

[0083] By combining a universal direct-fire nozzle, a flow regulating valve, and a fluid pressure sensor, along with an algorithm based on the fire scale and the distance to the fire point, the system is endowed with high execution capabilities. This is not only reflected in the high efficiency and precision of firefighting operations, but also in the precise fire protection of the large and special transportation facility, Cable-stayed Bridge 400, through micro-control of the spray parameters.

[0084] Utilizing the multi-joint linkage and long aspect ratio characteristics of the multi-axis robotic arm 2, it can traverse the complex cable clusters, guardrails, and maintenance obstacles of the cable-stayed bridge, delivering the nozzle 3 to the ideal jet space near the fire source. This wide-range positioning process solves the problems of limited range and blind spots of traditional fire monitors, achieving full spatial accessibility of the fire-fighting medium within the three-dimensional structure of the bridge. The omnidirectional direct-fire nozzle, through 360° rotation and pitch adjustment, forms a flexible execution terminal at the end of the multi-axis robotic arm 2. When the multi-axis robotic arm 2 cannot rotate further due to space constraints (such as being extremely close to the main cable), the omnidirectional direct-fire nozzle can adjust its own posture to ensure that the jet accurately penetrates structural gaps and locks onto the fire core.

[0085] Reference Figure 3 According to another aspect of the present invention, a method for extinguishing fires on cable-stayed bridges using the aforementioned intelligent fire-extinguishing robotic arm is also provided, comprising the following steps: 1) The sensing unit collects multi-dimensional data from the fire scene of the cable-stayed bridge 400; 2) The decision-making unit calls upon the reinforcement learning model trained by Sim-to-Real transfer training to autonomously generate task strategies based on multi-dimensional data; 3) The control unit converts the operation strategy into action commands for the jet device, drives the mobile base 1 to move to the target area, and the multi-axis robotic arm adjusts the nozzle 3 to the target position and controls the nozzle 3 to spray to extinguish the fire; 4) The sensing unit provides real-time feedback of multi-dimensional data so that the decision-making unit can obtain the fire extinguishing effect and environmental changes. The decision-making unit dynamically adjusts the operation strategy based on the multi-dimensional data fed back by the sensing unit to form a closed-loop control until the decision-making unit determines that the fire extinguishing is completed and issues a stop command.

[0086] Step 1) Collect multi-dimensional data of the fire scene at Cable-Ascending Bridge 400 using the sensing unit. This step enables a comprehensive digital reconstruction of the fire scene environment. Because the sensing unit integrates multiple sensors, the collected multi-dimensional data simultaneously covers visual images, infrared thermal fields, wind speed vectors, and structural stress. This real-time input of multi-source information ensures that subsequent decisions no longer rely on a single visual judgment; even in environments with dense smoke or at night, the location of the fire source and the trajectory of fire spread can still be accurately pinpointed.

[0087] Step 2) The decision-making unit invokes a reinforcement learning model trained through simulation-reality transfer learning to autonomously generate operational strategies based on multi-dimensional data. This step demonstrates the core intelligent advantage of this method. By using a model pre-trained in a virtual environment with numerous extreme conditions, the system can extrapolate complex multi-dimensional data within milliseconds and directly generate operational strategies including path planning and jet parameters. This autonomous generation capability effectively eliminates the delays and misjudgments associated with manual remote control, enabling the system to enter operational status with the optimal path in the critical early stages of a fire.

[0088] Step 3) The control unit translates the operational strategy into action commands for the jet device, driving the mobile base 1 to move to the target area, and the multi-axis robotic arm 2 to adjust the nozzle 3 to the target position and control the nozzle 3 to spray. This step achieves a precise transition from software logic to action. By refining the commands into specific parameters for the mobile base 1 and the multi-axis robotic arm, the system can ensure that the actuators achieve dynamic and precise positioning on the extremely wide cable-stayed bridge 400 structure, overcome the mechanical inertia and vibration during high-altitude operations, deliver the nozzle 3 to the optimal spray position in the fire scene, and ensure that the initial momentum of the extinguishing medium can effectively penetrate the hot airflow and cover the core of the fire.

[0089] Step 4) The sensing unit provides real-time feedback of multi-dimensional data, enabling the decision-making unit to obtain information on the fire extinguishing effect and environmental changes, and dynamically adjust the operational strategy to form a closed-loop control until the fire is deemed extinguished. This is the most practically valuable technical feature of this method. The fire environment at the cable-stayed bridge 400 is constantly changing; sudden changes in wind direction or structural displacement can affect fire extinguishing efficiency. Through this continuous closed-loop control, the system can continuously correct its operational strategy, significantly improving its robustness. It can compensate for deviations caused by wind speed fluctuations or fire migration in real time, ensuring the continuity of the fire extinguishing process and preventing the fire from reigniting on the verge of being extinguished. The setting in Step 4) regarding the determination of fire extinguishing completion by the decision-making unit introduces a data-driven task termination judgment criterion. By continuously monitoring the temperature and stress feedback at the fire scene, the system can objectively assess whether the threat of the fire to the key load-bearing components of the cable-stayed bridge 400 (such as the main cable and suspenders) has been completely eliminated. This stop command based on measured data avoids excessive water resource consumption and prevents safety hazards caused by premature evacuation.

[0090] The fire extinguishing method of this invention improves the automation level and response accuracy of fire extinguishing operations by constructing a technical closed loop of panoramic perception, intelligent simulation, collaborative execution, and real-time evolution. It not only greatly shortens the fire extinguishing response cycle, but also achieves safe protection and efficient cooling of the main structure of the bridge through closed-loop control in the extremely harsh high-altitude fire environment.

[0091] The intelligent fire-fighting robot for cable-stayed bridges of this invention can autonomously complete the entire closed-loop operation of "perception-decision-control," adapting to high-altitude and high-risk fire-fighting scenarios on cable-stayed bridges 400 without human intervention. Through high-fidelity training in virtual scenarios and domain-adaptive transfer learning algorithms, the spray control accuracy and fire-fighting response speed are significantly improved compared to traditional fire-fighting equipment, enabling rapid and precise containment of fire spread while balancing operational efficiency and environmental adaptability.

[0092] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cable-stayed bridge intelligent fire extinguishing robot, comprising a sensing unit, a decision-making unit, a control unit, and at least one jet device. Each jet device includes a movable base, a base drive device, a multi-axis robotic arm, a nozzle, and a hose. The movable base is movably mounted on the cable-stayed bridge. The base drive device is connected to the movable base to drive the movable base to move. The multi-axis robotic arm is mounted on the movable base, and the nozzle is mounted on the end effector of the multi-axis robotic arm. The nozzle is connected to a fire hose on the cable-stayed bridge. The sensing unit is used to collect multi-dimensional data of the fire scene on the cable-stayed bridge and send it to the decision-making unit. The decision-making unit has a reinforcement learning model for generating an operational strategy based on the multi-dimensional data and sending it to the control unit. The control unit, based on the operational strategy, controls the movement of the mobile base via a base drive device, adjusts the nozzle pose by controlling the movement of a multi-axis robotic arm, and controls the nozzle spray to extinguish the fire at the cable-stayed bridge fire scene. The characteristic of this system is that... The reinforcement learning model of the decision-making unit is generated using the simulation-reality transfer method, and the decision-making unit includes a virtual environment construction module, a reinforcement learning training module, and a transfer adaptation module. The virtual environment construction module is used to simulate the structural topology, fire characteristics, and environmental interference factors of cable-stayed bridges and to construct virtual training scenarios. The reinforcement learning training module is used to evaluate the fire extinguishing behavior of the agent in a virtual training scenario using a reward mechanism, so as to optimize and generate an initial jet control and path planning model, thereby establishing a mapping relationship between multi-dimensional data of the fire scene and the operation strategy. The operation strategy includes the movement parameters of the mobile base, the motion parameters of the multi-axis robotic arm, and the spray parameters of the nozzle. The migration adaptation module establishes a cross-domain mapping relationship by aligning features of multi-dimensional data from virtual training scenarios and real fire scenes, and fine-tunes the parameters of the jet control and path planning model. It introduces multi-dimensional data measured at real fire scenes to calibrate the initial jet control and path planning model, thereby optimizing the initial jet control and path planning model into a reinforcement learning model adapted to the real fire conditions of cable-stayed bridges. The reinforcement learning model then autonomously generates operational strategies based on the multi-dimensional data.

2. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The sensing unit includes a binocular camera, an infrared imager, a wind speed sensor, a smoke sensor, a humidity sensor, and a bridge structural stress sensor to achieve real-time, multi-dimensional data acquisition of the fire scene of the cable-stayed bridge. The multi-dimensional data includes images of the fire scene captured by the binocular camera, thermal images and temperature values ​​of the fire scene detected by the infrared imager, wind speed of the fire scene detected by the wind speed sensor, smoke concentration of the fire scene detected by the smoke sensor, humidity of the fire scene detected by the humidity sensor, and structural stress of the cable-stayed bridge detected by the bridge structural stress sensor.

3. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The specific process by which the reinforcement learning training module optimizes and generates the initial jet control and path planning model through a reward mechanism is as follows: The motion parameters of the mobile base, the motion parameters of the multi-axis robotic arm, and the spray parameters of the nozzle are used as the action space, and the fire spread rate, the peak flame temperature, and the humidity of the target area are used as status feedback indicators. A reward function is constructed based on the real-time changes of the state feedback index, and the agent is iteratively trained in the virtual training scenario using a proximal policy optimization algorithm. The parameters of the action space are evaluated based on the state feedback index through a reward mechanism, and the agent is guided to optimize the selection of the action space through the reward function, thereby generating the network weights of the jet control and path planning model.

4. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The virtual environment construction module also includes a preset model of fire protection zones for cable-stayed bridges; The decision-making unit determines the target fire zone corresponding to the fire scene area of ​​the cable-stayed bridge based on the preset model of the fire zone of the cable-stayed bridge, and instructs the jet device in the fire scene area to perform fire extinguishing based on the structural topology information of the target fire zone.

5. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The virtual environment construction module within the decision-making unit reconstructs the structural topology, including the main cable, suspension cables, and bridge deck paving, based on the 3D point cloud data and design drawings of the cable-stayed bridge structure. It then establishes a dynamic fire environment model based on fire dynamics simulation results, coupled with environmental interference factors collected by the sensing unit. These environmental interference factors include wind speed, temperature, and visibility. The module also renders the structural topology and dynamic fire environment model to generate a 1:1 scale virtual training scene.

6. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The transfer adaptation module performs feature alignment and difference quantification analysis based on the multi-dimensional data of the sensing unit and the corresponding data of the virtual training scene. It constructs a feature mapping model using the maximum mean difference domain adaptive transfer learning algorithm to fine-tune and optimize the parameters of the jet control and path planning model. It also calibrates the jet control and path planning model using measured data from the fire scene and corrects operational strategy deviations caused by environmental interference and structural errors through dynamic compensation factors.

7. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The control unit receives the operation strategy output by the decision unit and performs inverse kinematics calculation based on the kinematic model of the mobile base drive device, the multi-axis robotic arm and the nozzle to generate an operation strategy for the mobile base drive device, the multi-axis robotic arm and the nozzle. The control unit tracks the parameters of the actual operation of the mobile base drive device, multi-axis robotic arm and nozzle based on the PID adaptive algorithm, and dynamically corrects the operation strategy to compensate for control deviations based on the multi-dimensional data fed back by the sensing unit.

8. The intelligent fire-fighting robot for cable-stayed bridges according to claim 1, characterized in that, The nozzle is a universal direct-fire nozzle, and a flow regulating valve and a fluid pressure sensor are installed on the nozzle so that the control unit can dynamically adjust the nozzle's spray parameters based on the fire scale and the distance to the fire point. The nozzle's spray parameters include spray angle, flow rate and spray pressure.

9. A method for extinguishing fires on a cable-stayed bridge using the intelligent fire-extinguishing robot described in any one of claims 1 to 8, characterized in that, Includes the following steps: 1) The sensing unit collects multi-dimensional data from the fire scene of the cable-stayed bridge; 2) The decision-making unit calls upon the reinforcement learning model trained through simulation-reality transfer to autonomously generate task strategies based on multi-dimensional data; 3) The control unit converts the operation strategy into action commands for the jet device, drives the mobile base to move to the target area, and the multi-axis robotic arm adjusts the nozzle to the target position and controls the nozzle to spray to extinguish the fire; 4) The sensing unit provides real-time feedback of multi-dimensional data so that the decision-making unit can obtain the fire extinguishing effect and environmental changes. The decision-making unit dynamically adjusts the operation strategy based on the multi-dimensional data fed back by the sensing unit to form a closed-loop control until the decision-making unit determines that the fire extinguishing is completed and issues a stop command.

Images