Bridge fire extinguishing method and system based on multi-field perception and trajectory prediction, and storage medium

Abstract

The invention relates to the technical field of bridges, provides a bridge fire extinguishing method and system based on multi-field perception and trajectory prediction, and a storage medium, realizes jet flow control in a complex bridge fire scene, and remarkably improves the fire protection capability of a bridge. By estimating the heat release rate and the two-dimensional velocity field distribution, the problem that fire source dynamics and flow field interference cannot be sensed in real time in the prior art is solved. By establishing the thermal plume prior model and fusing the thermal plume prior model with actually measured data, the defects that the thermal plume is ignored and actually measured and theoretical advantages cannot be combined in the prior art are overcome. A reinforcement learning model taking the fire source heat release rate, the environment wind speed, the historical deviation and the current model prediction control weight as input is constructed, and the problem that a traditional control strategy cannot adapt to environment changes is solved. By adopting a hybrid optimization strategy combining reinforcement learning and model prediction control, the defect that a traditional model prediction control parameter is fixed and cannot adapt to a dynamic environment is overcome.

Description

Technical Field

[0001] This invention relates to the field of bridge technology, and in particular to a bridge fire extinguishing method, system, and storage medium based on multi-field perception and trajectory prediction. Background Technology

[0002] Fire protection for bridge structures (such as suspension bridges and cable-stayed bridges) is becoming increasingly challenging. The widespread adoption of new energy vehicles further complicates the fire risks associated with bridges. Currently, passive fire-retardant coatings are mainly used, and effective active fire suppression methods are lacking.

[0003] Existing automatic jet fire extinguishing systems face severe challenges in open bridge environments: (1) Complex environmental interference: strong crosswinds, Coriolis force generated by the Earth's rotation, and heat plume from the fire source all act on the jet trajectory, resulting in serious deviations; (2) Inaccurate models: traditional jet models only consider gravity and simple wind resistance, ignoring key physical effects, and the error can reach several meters at a distance of hundreds of meters; (3) Single control strategy: existing PID control based on fixed parameters cannot adapt to the complex environment of dynamic changes.

[0004] For example, existing technologies have proposed an automatic aiming device for fire monitors, but have not considered the coupling effect of thermal plume and Coriolis force; vision-based firefighting robots have been proposed, but lack accurate physical modeling of the jet trajectory. These solutions are all difficult to implement for precise fire suppression in real fire environments.

[0005] Therefore, it is necessary to propose a bridge fire extinguishing method, system, and storage medium based on multi-field perception and trajectory prediction to solve or at least alleviate the above-mentioned defects. Summary of the Invention

[0006] The main objective of this invention is to provide a bridge fire extinguishing method, system, and storage medium based on multi-field perception and trajectory prediction, in order to solve the technical problem that existing automatic jet fire extinguishing systems are difficult to achieve accurate fire extinguishing in real bridge fire environments and have serious deviations.

[0007] To achieve the above objectives, this invention provides a bridge fire extinguishing method based on multi-field sensing and trajectory prediction, comprising the following steps: S1, acquire the target location of the fire source area and the infrared radiation data of the fire source area, and estimate the heat release rate of the fire source based on the infrared radiation data; acquire the two-dimensional velocity field distribution in the jet path plane, and acquire the ambient wind speed vector and ambient temperature in the nozzle area of ​​the jet device; S2, based on the target location, heat release rate of the fire source, ambient wind speed vector and ambient temperature, establish a priori model of the thermal plume velocity field to obtain a theoretical estimate of the thermal plume velocity field; then fuse the two-dimensional velocity field distribution with the theoretical estimate to obtain the total velocity field of the jet path. S3, Construct a pre-trained reinforcement learning model; wherein, the reinforcement learning model takes the current heat release rate of the fire source, the current environmental wind speed vector, the historical landing point deviation sequence, and the weight parameters of the current model prediction control as input, and outputs the adjustment amount of the weight parameters; S4. Based on the adjustment of the weight parameters, update the weight matrix of the model predictive control. Then, using the total velocity field of the jet path as input, and based on the pre-established droplet dynamics model, adopt a hybrid optimization strategy that combines the reinforcement learning model and model predictive control to solve the optimal control command sequence in the future time domain. The optimal control command sequence includes pitch angle correction, yaw angle correction, and jet pressure correction. S5, the optimal control command sequence is output to the jet device to drive the nozzle and water pump to operate; then the actual landing point deviation data is acquired and fed back to the reinforcement learning model to update the parameters of the reinforcement learning model online. The updated parameters of the reinforcement learning model are used to solve the optimal control command sequence in subsequent steps.

[0008] Preferably, obtaining the two-dimensional velocity field distribution in the jet path plane in step S1 includes the following steps: A laser sheet light source and a high-speed camera are coaxially mounted at the nozzle of the jet device; The laser sheet light source is controlled to emit pulsed laser light within the jet path plane to illuminate the tracer particles within the jet path plane; The high-speed camera is controlled to continuously acquire image sequences of tracer particles at a frequency synchronized with the pulsed laser. Cross-correlation calculations are performed on two consecutive frames of images to obtain the two-dimensional velocity field distribution in the jet path plane.

[0009] Preferably, the prior model for the thermal plume velocity field established in step S2 based on the target location, the heat release rate of the fire source, the ambient wind speed vector, and the ambient temperature includes the following steps: S21, Based on the heat release rate of the fire source and the ambient temperature, the distribution of the rising velocity of the center line of the heat plume above the fire source area with height is calculated using an axisymmetric plume model under windless conditions. S22, based on the ambient wind speed vector and the rising velocity of the thermal plume centerline under windless conditions, as well as the target position, calculate the horizontal offset of the thermal plume centerline caused by the ambient wind, and obtain the position of the tilted thermal plume centerline. S23. Based on the rising velocity of the thermal plume centerline under windless conditions and the position of the thermal plume centerline after tilting, the three-dimensional thermal plume velocity field is reconstructed using the Gaussian distribution assumption.

[0010] Preferably, step S2, which involves fusing the two-dimensional velocity field distribution with the theoretical estimate to obtain the total velocity field of the jet path, includes the following steps: Based on the image signal-to-noise ratio of particle image velocimetry, the confidence level of particle image velocimetry under the current environmental conditions is calculated in real time; wherein, the value range of the confidence level of particle image velocimetry is [0,1]. Based on the confidence level of the particle image velocimetry technology, the two-dimensional velocity field distribution and the theoretical estimate are weighted and fused to obtain the total velocity field of the jet path.

[0011] Preferably, step S3 includes the following steps: S31, Define the state space of the reinforcement learning model; wherein, the state space includes the current heat release rate of the fire source, the current environmental wind speed vector, the historical landing point deviation sequence of N control cycles before the current moment, the landing point deviation weight matrix of the current model prediction control, and the control increment weight matrix. S32, Define the action space of the reinforcement learning model; wherein, the action space includes the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix; S33, Define the reward function of the reinforcement learning model; wherein, the reward function is calculated based on the actual landing point deviation and control increment of the current control cycle; S34, Based on the state space, action space and reward function, a deep reinforcement learning algorithm is used to pre-train and update the reinforcement learning model online, so that the reinforcement learning model learns the mapping strategy from the state space to the action space to maximize the cumulative reward; S35, in each control cycle, the state space at the current moment is input to the reinforcement learning model, and the action space at the current moment is output by the reinforcement learning model to obtain the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix.

[0012] Preferably, step S4 includes the following steps: S41, based on the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix, update the weight matrix of the model prediction control. S42, the model prediction controller, which takes the total velocity field of the jet path as input and the updated weight matrix as the weight of the cost function, performs control. S43, establish a droplet dynamics model that considers the coupling of gravity, aerodynamic drag, Coriolis force, pressure gradient force and virtual mass force; S44, in each control cycle, using the current droplet state as the initial condition, the model predicts the controller to solve for the optimal control command sequence in the future time domain based on the droplet dynamics model.

[0013] Preferably, step S43 includes the following steps: S431, based on Newton's second law, establish the vector form dynamic equation of a single droplet in an inertial coordinate system; S432 introduces gravity terms, aerodynamic drag terms, Coriolis force terms, pressure gradient force terms, and virtual mass force terms; S433, by adding the vector form dynamic equations with the gravity term, aerodynamic drag term, Coriolis force term, pressure gradient force term, and virtual mass force term, we obtain the continuous dynamic equations; S434, Discretize the continuous dynamics equation and establish a droplet dynamics discrete prediction model for use in the rolling optimization solution of step S44.

[0014] The present invention also provides a bridge fire extinguishing system based on multi-field perception and trajectory prediction, including a jet device and a control system. The jet device includes a support, a nozzle and a water pump. The control system includes a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the bridge fire extinguishing method based on multi-field perception and trajectory prediction as described above.

[0015] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the bridge fire extinguishing method based on multi-field perception and trajectory prediction as described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This application achieves high-precision, adaptive intelligent jet control in complex bridge fire scenarios, significantly improving the fire protection capabilities of bridges. Specifically, by acquiring infrared radiation data of the fire source to estimate the heat release rate and obtaining the two-dimensional velocity field distribution, it solves the problem that existing technologies cannot perceive the dynamics of the fire source and flow field interference in real time. By establishing a prior model of the thermal plume and fusing it with measured data, it overcomes the shortcomings of existing technologies that ignore the thermal plume and cannot combine the advantages of measured and theoretical data. By constructing a reinforcement learning model with the heat release rate of the fire source, ambient wind speed, historical deviation, and current model predictive control weights as inputs, it solves the problem that traditional control strategies cannot adapt to environmental changes. By adopting a hybrid optimization strategy that combines reinforcement learning and model predictive control, it overcomes the shortcomings of traditional model predictive control parameters being fixed and unable to adapt to dynamic environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of one embodiment of the present invention; Figure 2 This is a schematic diagram illustrating an application scenario of the jet device in one embodiment of the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings.

[0020] Explanation of icon numbers: 10. Support frame; 20. Nozzle; 30. Bridge; 40. High-speed CMOS camera; 50. Dual-band infrared thermal imager; 60. Anemometer. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] In this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. If the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0024] Please refer to Figures 1 to 2 The present invention provides a bridge fire extinguishing method based on multi-field perception and trajectory prediction, comprising the following steps: S1, acquire the target location of the fire source area and the infrared radiation data of the fire source area, and estimate the heat release rate of the fire source based on the infrared radiation data; acquire the two-dimensional velocity field distribution in the jet path plane, and acquire the ambient wind speed vector and ambient temperature in the nozzle 20 area of ​​the jet device; This application utilizes multi-source data acquisition to achieve real-time acquisition of environmental, fire source, and flow field information, providing a data foundation for subsequent modeling and control. Specifically, the fire source target location can be obtained by combining the fire source coordinates with a dual-band infrared thermal imager 50 installed on the top of the bridge tower using existing technology. An anemometer 60 can be used to collect the ambient wind speed vector in real time, and a temperature sensor can be used to obtain the ambient temperature.

[0025] The jet path plane refers to the vertical plane from the jet nozzle 20 to the fire source target. The two-dimensional velocity field consists of the horizontal (range direction) and vertical directions, and is based on measured data, containing two velocity components within the plane. However, in the subsequent droplet dynamics model, a complete three-dimensional velocity field is required to accurately calculate the jet trajectory. Therefore, through the data fusion step S2, the two-dimensional measured data is fused with the three-dimensional theoretical model to obtain the three-dimensional total velocity field.

[0026] S2, based on the target location, heat release rate of the fire source, ambient wind speed vector and ambient temperature, establish a priori model of the thermal plume velocity field to obtain a theoretical estimate of the thermal plume velocity field; then fuse the two-dimensional velocity field distribution with the theoretical estimate to obtain the total velocity field of the jet path. S3, Construct a pre-trained reinforcement learning model; wherein, the reinforcement learning model takes the current heat release rate of the fire source, the current environmental wind speed vector, the historical landing point deviation sequence, and the weight parameters of the current model prediction control as input, and outputs the adjustment amount of the weight parameters; S4, based on the adjustment of the weight parameters, update the weight matrix of the model predictive control, and then, using the total velocity field of the jet path as input, based on the pre-established droplet dynamics model, adopt a hybrid optimization strategy combining the reinforcement learning model and model predictive control to solve the optimal control command sequence in the future time domain. The optimal control command sequence includes pitch angle correction, yaw angle correction, and jet pressure correction. The trajectory prediction in this application refers to the pre-calculation and dynamic updating of the entire flight trajectory of the jet from nozzle 20 to the fire source target based on the droplet dynamics model coupled with multi-physics fields.

[0027] S5, the optimal control command sequence is output to the jet device to drive the nozzle 20 and the water pump to operate; then the actual landing point deviation data is obtained and fed back to the reinforcement learning model to update the parameters of the reinforcement learning model online. The updated parameters of the reinforcement learning model are used to solve the optimal control command sequence in subsequent steps.

[0028] In a preferred embodiment, obtaining the two-dimensional velocity field distribution in the jet path plane in step S1 includes the following steps: A laser sheet light source and a high-speed camera are coaxially mounted at the nozzle 20 of the jet device. Specifically, a low-power, eye-safe pulsed laser and a high-speed CMOS camera 40 are mounted on the housing of the movable nozzle 20 of the jet device. The pulsed laser, together with a cylindrical lens group, constitutes the laser sheet light source. The laser sheet light source and the high-speed camera are coaxially configured, that is, their optical axes are parallel and consistent with the initial aiming direction of the jet nozzle 20, ensuring that the plane covered by the laser sheet light is the plane of the expected jet path from the nozzle 20 to the fire source target.

[0029] The laser sheet light source is controlled to emit pulsed laser light within the jet path plane, illuminating the tracer particles within the jet path plane. Simultaneously with the start of jet ejection, the control system triggers the laser to emit pulsed laser light at a preset frequency (e.g., 100Hz). The laser light is shaped by a cylindrical lens group into a thin, sheet-like light plane approximately 1-2 mm thick, precisely covering the expected jet path from nozzle 20 to the fire source target. Naturally occurring airborne particles (such as water mist, dust particles, and smoke) or trace tracer particles actively released by the system within the light plane are illuminated by the laser, forming scattered light spots that can be captured by a camera.

[0030] The high-speed camera is controlled to continuously acquire image sequences of tracer particles at a frequency synchronized with the pulsed laser. The high-speed CMOS camera 40 and the laser are synchronously triggered using the same clock source, ensuring that each frame corresponds to one pulsed laser illumination. The camera continuously captures images at the same frequency as the laser, with each exposure time being extremely short (microseconds) to freeze the high-speed moving particles. In the acquired image sequence, the time interval between two adjacent frames, and the displacement of the particles within that time interval, are completely recorded.

[0031] Cross-correlation calculations are performed on two consecutive image frames to obtain the two-dimensional velocity field distribution within the jet path plane. The two consecutively acquired image frames are divided into several small query windows (e.g., 32×32 pixels, 50% overlap), and a two-dimensional cross-correlation operation is performed on the grayscale distribution within each query window. The peak position of the cross-correlation corresponds to the displacement vector of the particle swarm between the two image frames. The cross-correlation algorithm is the core algorithm of particle image velocimetry (PIV) technology and is a mature algorithm in this field; it will not be elaborated upon here. Based on the known time interval, the velocity vector at the location of the window can be calculated.

[0032] As a preferred embodiment, the prior model of the thermal plume velocity field established in step S2 based on the target location, the heat release rate of the fire source, the ambient wind speed vector, and the ambient temperature includes the following steps: S21, based on the heat release rate of the fire source and the ambient temperature, the distribution of the rising velocity of the centerline of the thermal plume above the fire source area with height under windless conditions is calculated using an axisymmetric plume model; the heat release rate of the fire source (unit: W) and the ambient temperature (unit: K) at the current moment are obtained from step S1. The heat release rate of the fire source can be obtained by collecting radiation intensity data of the fire source area by a dual-band infrared thermal imager 50, combined with visible light image analysis and a deep learning-based algorithm for real-time estimation. This part can use existing mature technologies, such as referring to Chinese patent, patent number: ZL202311226491.7. The heat release rate of the fire source reflects the actual combustion intensity of the fire source.

[0033] Establish a physical prior model based on the characteristics of the fire source. In actual fire scenarios (such as a vehicle fire on bridge 30), the fire source has certain physical dimensions. If this dimension is ignored in the calculation and the actual height z is directly substituted into the original formula, it will lead to significant errors in the calculation near the fire source region.

[0034] To address this issue, the concept of a "virtual point source" is commonly used in engineering to modify the Zukoski model. The core idea is to represent a real fire source of a certain size as an equivalent point source located at a certain depth below the fire source. The ideal point source. In this correction, the virtual point source height... It is a function of the characteristic diameter D of the fire source.

[0035] According to Heskestad's research, virtual point source height The diameter of the ignition source can be correlated with the following empirical formula: ; Or a simpler form can be used (suitable for most engineering scenarios): ; in: The characteristic diameter of the ignition source (m). For non-circular ignition sources, the effective diameter is used. A represents the area of ​​the fire source.

[0036] Therefore, for axisymmetric plumes, the change in centerline velocity with height can be described by the Zukoski model: ;in, The rising velocity of the thermal plume centerline (unit: m / s). This is an empirical constant, typically taken as 0.96 for axisymmetric plumes. Ambient air density (unit: kg / m³). Specific heat capacity of air at constant pressure (unit: J / (kg·K)) Ambient temperature (unit: K). Heat release rate from the heat source (unit: W). Acceleration due to gravity (unit: m / s²) 2 ), For height variables.

[0037] S22, based on the environmental wind speed vector, the rising velocity of the thermal plume centerline under windless conditions, and the target position, calculate the horizontal offset of the thermal plume centerline caused by the environmental wind to obtain the tilted thermal plume centerline position; obtain the environmental wind speed vector at the current moment from step S1. Since the thermal plume is mainly affected by horizontal wind, take the projection of the environmental wind speed vector on the horizontal plane. Based on the principle of "plumes being tilted in crosswinds" in fluid mechanics, calculate the horizontal offset of the thermal plume centerline caused by the environmental wind. Assume that at each height, the time required for the plume micro-cluster to rise from the fire source to that height, during this time, the distance the environmental wind pushes it horizontally is the projection of the environmental wind speed vector on the horizontal plane multiplied by the time required for the plume micro-cluster to rise from the fire source to that height. Therefore, the tilted thermal plume centerline position can be obtained.

[0038] S23. Based on the rising velocity of the thermal plume centerline under windless conditions and the position of the thermal plume centerline after tilting, the three-dimensional thermal plume velocity field is reconstructed using the Gaussian distribution assumption. Based on the classical theory of turbulent jets, the velocity distribution of the thermal plume approximately follows a Gaussian distribution (normal distribution) on a cross-section perpendicular to the axis. The Gaussian distribution assumption is a classic assumption in turbulent jet and plume modeling and is a mature technical approach. While the Gaussian distribution assumption itself is conventional, its application in this embodiment has the following special characteristics: This embodiment couples the Gaussian distribution with real-time fire source parameters, enabling the model to dynamically respond to changes in fire intensity; it couples the Gaussian distribution with ambient wind, allowing the axis to tilt with height; and it fuses the Gaussian distribution with PIV measured data, achieving complementary advantages between theory and measurement.

[0039] As an example, the process of reconstructing the functional expression of the thermal plume velocity field in three-dimensional space using the Gaussian distribution assumption in this embodiment is as follows: (1) Environmental wind field: For the bridge environment at a scale of 30 meters, the environmental wind field can be approximated as a uniform field, or its variation with height can be described using a logarithmic wind profile: Uniform field approximate: ; For bridge 30, the vertical component of the ambient wind is relatively small and has little impact on the jet action and flame. Therefore, to simplify the calculation, the vertical component of the ambient wind is... The velocity component in the axial direction is set to 0; , The horizontal levels measured by the anemometer 60 are respectively Axial direction, Wind speed component in the axial direction.

[0040] To accurately describe the spatial positions of each component of the system and the motion state of the jet, this invention defines the following right-handed Cartesian coordinate system: Origin of coordinate system O: The center point of the bottom of the bridge tower where this system is installed; Positive direction of X-axis: along the longitudinal direction of bridge 30, pointing to the other end of the bridge tower; Positive direction of Y-axis: along the transverse direction of bridge 30, pointing to the downstream side of bridge 30; Positive direction of Z-axis: vertically upward.

[0041] In this coordinate system, the coordinates of the nozzle 20 outlet center are: Coordinates of the fire source center The coordinates of any point in space, as determined by the dual-band infrared thermal imager 50, are expressed as follows: .

[0042] (2) Thermal plume flow field: The Gaussian distribution assumption is used to describe the upward velocity component of the thermal plume. : ; The main direction of motion of the thermal plume is vertically upward, therefore its vector form only includes... The directional component and the horizontal component are negligible. This is the plume spread coefficient, typically taken as 0.11.

[0043] (3) Correction of the influence of ambient wind on the morphology of thermal plumes: When ambient winds are present, thermal plumes will tilt and deflect, which is an important effect that must be considered. This invention addresses this by employing the following method: 1. Plume axis shift: Under the influence of horizontal winds, the plume axis will bend towards the leeward direction. The coordinates of the shifted fire source axis become: ; ; in, , These are the corrected coordinates of the fire source axis.

[0044] 2. Synthetic plume field under the influence of wind: After considering wind-induced displacement, the complete plume field expression is corrected to: ; (4) Complete expression for the composite flow field function: Integrating the above, we obtain the three-dimensional synthetic flow field function for jet trajectory calculation according to this invention: ; The function expression described above It is the theoretical flow field in three-dimensional space, i.e., the theoretical estimate.

[0045] In a preferred embodiment, step S2, which involves fusing the two-dimensional velocity field distribution with the theoretical estimate to obtain the total velocity field of the jet path, includes the following steps: Based on the image signal-to-noise ratio (SNR) of particle image velocimetry (PEV) technology, the confidence level of PPV technology under the current environmental conditions is calculated in real time. The confidence level ranges from [0,1]. Specifically, environmental visibility sensor data (in meters) or the image SNR (in dB) of PPV technology under the current environmental conditions is obtained. The confidence level of PPV technology under the current environmental conditions is then calculated in real time. Confidence of particle image velocimetry technology The calculation method is as follows: map environmental visibility or image signal-to-noise ratio to the [0,1] interval. For example, the following linear mapping function can be used: ; in, Environmental visibility, in meters (m). The visibility threshold for a clean environment (e.g., 100m) is used. When visibility is above this value, PIV measurements are completely reliable. This is the visibility threshold for severe environments (e.g., 10m). When visibility is below this value, PIV measurements are completely unreliable.

[0046] Confidence of particle image velocimetry The value range is [0,1]. =1 indicates that the current environmental conditions are very suitable for PIV measurement (clean environment, high signal-to-noise ratio), and the measured data are completely reliable; =0 indicates that the current environmental conditions are extremely unsuitable for PIV measurement (dense smoke environment, low signal-to-noise ratio), and the measured data is completely unreliable; 0 < <1 indicates that the environmental conditions are between the two, and the reliability of the measured data increases monotonically with the confidence level.

[0047] Based on the confidence level of the particle image velocimetry technology, the two-dimensional velocity field distribution and the theoretical estimate are weighted and fused to obtain the total velocity field of the jet path.

[0048] Using formula Weighted fusion is performed to obtain the total velocity field of the jet path. (Unit: m / s); where, The theoretical estimate (i.e.) (Unit: m / s) This represents the two-dimensional velocity field distribution (unit: m / s).

[0049] In a preferred embodiment, step S3 includes the following steps: S31, define the state space of the reinforcement learning model; wherein, the state space includes the current heat release rate of the fire source, the current environmental wind speed vector, the historical landing point deviation sequence of N control cycles before the current time, the landing point deviation weight matrix of the current model predictive control, and the control increment weight matrix; the state space is the set of environmental information observed by the reinforcement learning model in the current control cycle, and is the input basis for the model decision. In this embodiment, the state space is composed of the following four types of parameters: (1) The current heat release rate of the fire source reflects the combustion intensity of the fire source. The larger the fire, the stronger the heat plume, and the more significant the interference with the jet. Therefore, it is necessary to adjust the control strategy to pursue the accuracy of the hit more aggressively.

[0050] (2) The current ambient wind speed vector reflects the interference of the ambient wind on the jet. The greater the wind speed and the more drastic the change in wind direction, the greater the uncertainty of the jet trajectory, and the more necessary it is to make corresponding adjustments to the control system.

[0051] (3) The historical landing point deviation sequence of N control cycles before the current moment. The landing point deviation is the vector difference between the actual landing point position and the target position of the fire source. The historical landing point deviation sequence reflects the recent control effect of the system. If the historical deviation is large, it means that the current control strategy needs to be adjusted; if the deviation gradually converges, it means that the strategy is effective.

[0052] (4) The current model predictive control weight parameters, including the landing deviation weight matrix and the control increment weight matrix, are the core parameters in the MPC (Model Predictive Control) cost function. They determine the controller's trade-off between hit accuracy and motion stability: the larger the landing deviation weight matrix, the more the controller values ​​landing accuracy, even at the cost of drastic motion; the larger the control increment weight matrix, the more the controller values ​​motion stability, even at the cost of some accuracy, to avoid drastic adjustments. The current values ​​of these two weight matrices reflect the system's current control preferences and are the objects that the reinforcement learning model needs to adjust.

[0053] The control cycle refers to the time interval between two consecutive decisions made by a control system. Within each control cycle, the system completes one full perception-decision-execution cycle. Collect sensor data at the current moment; run control algorithms (including reinforcement learning inference and model predictive control optimization); output control commands to the actuators; and wait for the next cycle to begin.

[0054] Landing point deviation weight matrix It is a 3×3 diagonal matrix. ;in, This represents the weight of the deviation in the X-direction (along the longitudinal direction of the bridge). The weight of the landing point deviation in the Y direction (lateral bridge direction) is... This matrix represents the weight of the landing point deviation in the Z-direction (vertical direction). It is used as the state deviation penalty term in the cost function of model predictive control; a larger weight indicates that the system prioritizes accuracy in that direction.

[0055] Control Incremental Weight Matrix It is a 3×3 diagonal matrix. ;in, The weight of the pitch angle correction. The weight of the yaw angle correction. This represents the weight of the jet pressure correction. This matrix is ​​used in the control increment penalty term of the cost function of model predictive control. The larger the weight, the more the system values ​​the smoothness of action in that direction, i.e., it does not want the control variable to change too drastically.

[0056] S32, define the action space of the reinforcement learning model; wherein, the action space includes the adjustment amount of the landing deviation weight matrix and the adjustment amount of the control increment weight matrix; the action space is the decision result output by the reinforcement learning model in the current control cycle, used to adjust the weight parameters of MPC. In this embodiment, the action space consists of two adjustment amounts: The adjustment amount for the landing deviation weight matrix is ​​a matrix with the same dimensions as the landing deviation weight matrix, representing the amount of correction to the current landing deviation weight matrix. The adjusted new weight matrix is: ;in, This is the adjustment amount for the landing deviation weight matrix. The landing point deviation weight matrix is... This is the adjusted landing point deviation weight matrix.

[0057] The adjustment amount for the control increment weight matrix is ​​a matrix of the same dimension as the control increment weight matrix, representing the correction amount to the current control increment weight matrix. The adjusted new weight matrix is: ;in, To control the adjustment amount of the incremental weight matrix, To control the incremental weight matrix, This is the adjusted control increment weight matrix.

[0058] Each component in the action space is a continuous real number, and its value range is set according to actual needs (e.g., [-0.1, 0.1]) to ensure the smoothness of weight adjustment.

[0059] S33, define the reward function of the reinforcement learning model; wherein, the reward function is calculated based on the actual landing point deviation and control increment of the current control cycle; the reward function is the learning signal of the reinforcement learning model, used to evaluate the performance of the current control cycle. In this embodiment, the reward function is defined as follows: ; in, For the reward function; These are the actual jet impact points, in meters (m). The coordinates of the fire source are in meters. , , These are all reward weight coefficients, dimensionless, and need to be adjusted during training; The time from system detection of the fire source to initial target hit, measured in seconds (s). For control vector ,in, This is the pitch angle correction amount. This is the yaw angle correction amount. This is the correction amount for the jet pressure.

[0060] S34, Based on the state space, action space and reward function, a deep reinforcement learning algorithm is used to pre-train and update the reinforcement learning model online, so that the reinforcement learning model learns the mapping strategy from the state space to the action space to maximize the cumulative reward; This process consists of two phases: offline pre-training and online updates.

[0061] (1) Offline pre-training stage: Before actual system deployment, large-scale pre-training is conducted using a high-fidelity simulation environment. The simulation environment can simulate various fire scenarios (fire source power varies randomly from 1 to 10 MW), various wind field conditions (wind speed 0-15 m / s, wind direction varies randomly), and various fire source locations (distance 50-200 m, height 0-100 m). The reinforcement learning model continuously tries and fails in interaction with the simulation environment, learning preliminary control strategies. In this embodiment, the soft actor-critic algorithm is used for training. This algorithm is based on the maximum entropy reinforcement learning framework, which maximizes the entropy of the policy while maximizing the cumulative reward, so as to balance utilization and exploration.

[0062] (2) After the system is actually deployed, new empirical data will be generated in each control cycle. This data is stored in the empirical replay buffer and used to update the reinforcement learning model online. The online update uses the same soft actor-critic algorithm as the pre-training algorithm, but the learning rate is appropriately reduced (e.g., from 10). -3 Reduced to 10 -4 This is to ensure stable fine-tuning rather than overturning existing strategies.

[0063] This step combines the sample efficiency of offline learning with the adaptive capabilities of online learning through pre-training and online updates. Pre-training gives the model initial experience, while online updates make it more accurate with repeated use.

[0064] S35, in each control cycle, the state space at the current moment is input to the reinforcement learning model, and the action space at the current moment is output by the reinforcement learning model to obtain the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix.

[0065] The current state space is input into the actor network of the trained reinforcement learning model. For a continuous action space, the actor network typically outputs the mean and log-standard deviation of the action distribution. In actual execution, the mean is usually taken directly as the deterministic action. The adjustment amounts of the landing deviation weight matrix and the control increment weight matrix output by the model are passed to step S4 to update the weight matrix of the model's predictive control.

[0066] In a preferred embodiment, step S4 includes the following steps: S41, based on the adjustment of the landing deviation weight matrix and the adjustment of the control increment weight matrix, update the weight matrix of the model predictive control; using an additive update strategy, calculate the weight matrix to be used in the next control cycle (i.e., this control cycle): the new landing deviation weight matrix equals the current value plus the adjustment; the new control increment weight matrix equals the current value plus the adjustment. The updated weight matrix will be used for solving the model predictive control in this control cycle. To ensure system stability, each element of the weight matrix is ​​limited to a preset safety range (e.g., between 0.1 and 10.0). If the adjustment causes an element to exceed the range, its boundary value is taken to avoid excessive single adjustment leading to control instability.

[0067] S42, the total velocity field of the jet path is used as input, and the updated weight matrix is ​​used as the cost function weights for control by a model predictive controller; the controller uses the droplet dynamics model established in step S43. This model can predict the jet trajectory over a future period based on the current droplet state and control commands. One of the core inputs of the model is the total velocity field of the jet path obtained in step S2, which includes the combined effects of ambient wind and thermal plume, enabling the predictive model to perceive the influence of the disturbance field.

[0068] S43, establish a droplet dynamics model that considers the coupling of gravity, aerodynamic drag, Coriolis force, pressure gradient force and virtual mass force; S44, in each control cycle, using the current droplet state as the initial condition, based on the droplet dynamics model, the model predicts the optimal control command sequence for the future time domain in a rolling solution. In each control cycle (e.g., 50 milliseconds), the following operation is performed once: the droplet state at the current moment is obtained from the sensor or state estimator, including but not limited to the droplet's current position coordinates and current velocity vector, thus forming the initial conditions for prediction. A prediction time domain is set, and its length is defined, for example, 20 seconds, meaning the controller will predict the jet trajectory for the next 20 control cycles (i.e., the next second). The choice of prediction time domain requires a trade-off between computational burden and prediction accuracy: a longer time domain results in greater computation, but may lead to better control performance. Based on the droplet dynamics model, a finite-time optimization problem is constructed. The goal of this problem is: under the premise of satisfying constraints, to select the control command sequence for the next 20 cycles that minimizes the cost function. Preferably: Objective function: ; Constraints: ; ; in, Let be the objective function. For the first The predicted jet landing point, in meters; For prediction in the time domain; Let be the system state vector. , This represents the position of the droplet in the global coordinate system, in meters (m). The velocity components of the droplet in the three directions are expressed in m / s. For control vector ,in, This is the pitch angle correction amount. This is the yaw angle correction amount. This is a correction for the jet pressure. For the first One control cycle; +1 is the first +1 control cycle; function The system dynamics function is obtained by numerical discretization of the droplet dynamics equation described in this invention. The specific discretization method has been disclosed in a large number of documents and will not be elaborated in this invention. To minimize the control vector, This is the maximum value of the control vector. Weighted fusion is performed to obtain the total velocity field of the jet path. (Unit: m / s); where, The theoretical estimate (i.e.) (Unit: m / s) This represents the two-dimensional velocity field distribution (unit: m / s).

[0069] A numerical optimization algorithm is used to solve the above problem, obtaining the optimal control command sequence for the next 20 cycles. Each optimal control command sequence includes pitch angle correction, yaw angle correction, and jet pressure correction. The first command in the optimal sequence is executed and output to the jet device in step S5 for execution. The remaining commands are not executed and are re-optimized and solved in the next control cycle. When the next control cycle arrives, the new current state is obtained, and the above process is repeated to form a continuous control loop.

[0070] In a preferred embodiment, step S43 includes the following steps: S431, based on Newton's second law, establish the vector form dynamic equation of a single droplet in an inertial coordinate system; S432 introduces gravity terms, aerodynamic drag terms, Coriolis force terms, pressure gradient force terms, and virtual mass force terms; S433, by adding the vector form dynamic equations with the gravity term, aerodynamic drag term, Coriolis force term, pressure gradient force term, and virtual mass force term, we obtain the continuous dynamic equations; S434, Discretize the continuous dynamics equation and establish a droplet dynamics discrete prediction model for use in the rolling optimization solution of step S44.

[0071] Traditional jet trajectory prediction uses a simple parabolic model, considering only gravity and empirical wind resistance, which is feasible for short-distance indoor fire fighting. However, in the open environment of Bridge 30, the jet needs to cross a distance of hundreds of meters and last for several seconds. At this time, more physical effects that are ignored in the short distance must be considered. These effects include: (1) the Coriolis effect caused by the Earth's rotation; (2) the strong thermal plume generated by the fire source: above the fire source, due to the density difference between the high-temperature exhaust gas and the surrounding cold air, a strong buoyancy is generated, forming a continuous updraft. This thermal plume has two key characteristics: velocity field: at a certain height above the flame, an upward velocity distribution will be formed, with the highest velocity at the center. temperature field: the air density in the high-temperature area is low, and the viscosity changes, which will produce additional aerodynamic effects on the jet droplets that pass through. Ignoring the thermal plume will cause the jet to be "pushed away" from the target trajectory when it approaches the fire source, thus deviating significantly from the fire point; (3) the complex interaction between the jet and the air.

[0072] To fully consider the impact of these effects on bridge fire suppression using 30mm jets and improve the fire suppression effect, this invention attempts to incorporate the influence of these effects into the key parameters of the fire suppression water jet (pitch angle (vertical angle)). Yaw angle (horizontal angle) Jet water pressure Calculations of (etc.).

[0073] Considering that in the far-field region of the jet (i.e., most of the jet's flight path), the water column has completely broken into a large number of discrete droplets, a discrete droplet model is used to describe the jet motion.

[0074] This invention starts from the most basic Newton's second law and establishes the vector form of the dynamic equations for a droplet. In a preferred example, to simplify calculations, this invention focuses on introducing gravity and aerodynamic drag. (For extremely high positioning accuracy, in another preferred example, the Coriolis force, pressure gradient force, and virtual mass force can also be introduced. All three forces can be obtained using mature technologies, such as reference patent number: ZL202210630410.9.) In an inertial coordinate system, the dynamic equation of a single droplet is: ; in: The mass of the droplet, expressed in kg, can be calculated based on the dynamic breakup model of the KH / RT unsteady wave theory. The suggested value in this invention is 5.2 × 10⁻⁶. -7 ~1.4×10 -5 kg; This is the droplet velocity vector, with units of m / s; Time, in seconds; This is the gravity vector, with units of N. ,in This is the vector of gravitational acceleration, with units of m / s². This represents aerodynamic drag, measured in N (N). ; drag coefficient As the Reynolds number varies, an empirical formula is used within the intermediate Reynolds number range: Applicable to ; The Reynolds number is defined as: ; This refers to air density, expressed in kg / m³. This refers to the droplet diameter, in meters (m). A typical value is 1.0 mm or can be obtained from the parameters of the jet extinguishing device. The area of ​​the droplet facing the wind is expressed in meters (m²). 2 ; This is the aerodynamic viscosity, measured in Pa·s, typically taken as 1.8 × 10⁻⁶. -5 Pa·s; This is a relative velocity vector, with units of m / s. ; in Weighted fusion is performed to obtain the total velocity field of the jet path. (Unit: m / s); where, The theoretical estimate (i.e.) (Unit: m / s) This represents the two-dimensional velocity field distribution (unit: m / s).

[0075] This invention also provides a bridge fire extinguishing system 30 based on multi-field perception and trajectory prediction, including a jet device and a control system. The jet device includes a support 10, a nozzle 20, and a water pump. The control system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the bridge fire extinguishing method 30 based on multi-field perception and trajectory prediction as described above. For example, the jet device of this application can be installed on both sides of the bridge 30 road, such as... Figure 2 As shown, the spray angle of nozzle 20 is adjustable.

[0076] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the bridge 30 fire extinguishing method based on multi-field perception and trajectory prediction as described above.

[0077] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0078] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A bridge fire extinguishing method based on multi-field perception and trajectory prediction, characterized in that, Includes the following steps: S1, acquire the target location of the fire source area and the infrared radiation data of the fire source area, and estimate the heat release rate of the fire source based on the infrared radiation data; The two-dimensional velocity field distribution in the jet path plane is obtained, as well as the ambient wind speed vector and ambient temperature in the nozzle area of ​​the jet device; S2, Based on the target location, heat release rate of the fire source, ambient wind speed vector and ambient temperature, establish a priori model of the thermal plume velocity field and obtain a theoretical estimate of the thermal plume velocity field. The two-dimensional velocity field distribution is then fused with the theoretical estimate to obtain the total velocity field of the jet path. S3, Construct a pre-trained reinforcement learning model; wherein, the reinforcement learning model takes the current heat release rate of the fire source, the current environmental wind speed vector, the historical landing point deviation sequence, and the weight parameters of the current model prediction control as input, and outputs the adjustment amount of the weight parameters; S4. Based on the adjustment of the weight parameters, update the weight matrix of the model predictive control. Then, using the total velocity field of the jet path as input, and based on the pre-established droplet dynamics model, adopt a hybrid optimization strategy that combines the reinforcement learning model and model predictive control to solve the optimal control command sequence in the future time domain. The optimal control command sequence includes pitch angle correction, yaw angle correction, and jet pressure correction. S5, the optimal control command sequence is output to the jet device to drive the nozzle and water pump to operate; then the actual landing point deviation data is acquired and fed back to the reinforcement learning model to update the parameters of the reinforcement learning model online. The updated parameters of the reinforcement learning model are used to solve the optimal control command sequence in subsequent steps.

2. The bridge fire extinguishing method based on multi-field perception and trajectory prediction according to claim 1, characterized in that, The step S1 of obtaining the two-dimensional velocity field distribution in the jet path plane includes the following steps: A laser sheet light source and a high-speed camera are coaxially mounted at the nozzle of the jet device; The laser sheet light source is controlled to emit pulsed laser light within the jet path plane to illuminate the tracer particles within the jet path plane; The high-speed camera is controlled to continuously acquire image sequences of tracer particles at a frequency synchronized with the pulsed laser. Cross-correlation calculations are performed on two consecutive frames of images to obtain the two-dimensional velocity field distribution in the jet path plane.

3. The bridge fire extinguishing method based on multi-field perception and trajectory prediction according to claim 1, characterized in that, The prior model for the thermal plume velocity field established in step S2 based on the target location, heat release rate of the fire source, ambient wind speed vector, and ambient temperature includes the following steps: S21, Based on the heat release rate of the fire source and the ambient temperature, the distribution of the rising velocity of the center line of the heat plume above the fire source area with height is calculated using an axisymmetric plume model under windless conditions. S22, based on the ambient wind speed vector and the rising velocity of the thermal plume centerline under windless conditions, as well as the target position, calculate the horizontal offset of the thermal plume centerline caused by the ambient wind, and obtain the position of the tilted thermal plume centerline. S23. Based on the rising velocity of the thermal plume centerline under windless conditions and the position of the thermal plume centerline after tilting, the three-dimensional thermal plume velocity field is reconstructed using the Gaussian distribution assumption.

4. The bridge fire extinguishing method based on multi-field perception and trajectory prediction according to claim 3, characterized in that, Step S2, which involves fusing the two-dimensional velocity field distribution with the theoretical estimate to obtain the total velocity field of the jet path, includes the following steps: Based on the image signal-to-noise ratio of particle image velocimetry, the confidence level of particle image velocimetry under the current environmental conditions is calculated in real time; wherein, the value range of the confidence level of particle image velocimetry is [0,1]. Based on the confidence level of the particle image velocimetry technology, the two-dimensional velocity field distribution and the theoretical estimate are weighted and fused to obtain the total velocity field of the jet path.

5. The bridge fire extinguishing method based on multi-field perception and trajectory prediction according to claim 1, characterized in that, Step S3 includes the following steps: S31, Define the state space of the reinforcement learning model; wherein, the state space includes the current heat release rate of the fire source, the current environmental wind speed vector, the historical landing point deviation sequence of N control cycles before the current moment, the landing point deviation weight matrix of the current model prediction control, and the control increment weight matrix. S32, Define the action space of the reinforcement learning model; wherein, the action space includes the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix; S33, Define the reward function of the reinforcement learning model; wherein, the reward function is calculated based on the actual landing point deviation and control increment of the current control cycle; S34, Based on the state space, action space and reward function, a deep reinforcement learning algorithm is used to pre-train and update the reinforcement learning model online, so that the reinforcement learning model learns the mapping strategy from the state space to the action space to maximize the cumulative reward; S35, in each control cycle, the state space at the current moment is input to the reinforcement learning model, and the action space at the current moment is output by the reinforcement learning model to obtain the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix.

6. The bridge fire extinguishing method based on multi-field perception and trajectory prediction according to claim 5, characterized in that, Step S4 includes the following steps: S41, based on the adjustment amount of the landing point deviation weight matrix and the adjustment amount of the control increment weight matrix, update the weight matrix of the model prediction control. S42, the model prediction controller, which takes the total velocity field of the jet path as input and the updated weight matrix as the weight of the cost function, performs control. S43, establish a droplet dynamics model that considers the coupling of gravity, aerodynamic drag, Coriolis force, pressure gradient force and virtual mass force; S44, in each control cycle, using the current droplet state as the initial condition, the model predicts the controller to solve for the optimal control command sequence in the future time domain based on the droplet dynamics model.

7. The bridge fire extinguishing method based on multi-field perception and trajectory prediction according to claim 6, characterized in that, Step S43 includes the following steps: S431, based on Newton's second law, establish the vector form dynamic equation of a single droplet in an inertial coordinate system; S432 introduces gravity terms, aerodynamic drag terms, Coriolis force terms, pressure gradient force terms, and virtual mass force terms; S433, by adding the vector form dynamic equations with the gravity term, aerodynamic drag term, Coriolis force term, pressure gradient force term, and virtual mass force term, we obtain the continuous dynamic equations; S434, Discretize the continuous dynamics equation and establish a droplet dynamics discrete prediction model for use in the rolling optimization solution of step S44.

8. A bridge fire extinguishing system based on multi-field perception and trajectory prediction, characterized in that, The method includes a jetting device and a control system. The jetting device includes a support, a nozzle, and a water pump. The control system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the bridge fire extinguishing method based on multi-field sensing and trajectory prediction as described in any one of claims 1 to 7.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the bridge fire extinguishing method based on multi-field perception and trajectory prediction as described in any one of claims 1 to 7.

Images