A control optimization method based on a self-propelled unmanned operation fire extinguishing robot

Abstract

The application discloses a control optimization method of a self-walking unmanned operation fire extinguishing robot, relates to the technical field of intelligent control, receives fire source coordinate information, performs target bias sampling and adaptive step expansion in a global map, generates an initial discrete path, and smoothes a global reference trajectory; drives the robot to move, receives local ultrasonic detection data, performs collision prediction and speed space evaluation, generates a local avoidance control instruction to drive the robot to operate, acquires three-dimensional point cloud data of the fire source, extracts and outputs a three-dimensional feature set of the fire source based on the three-dimensional point cloud data and an AI algorithm, maps the size of the fire source into a fluid control parameter set, calculates a spatial motion trajectory and a gun distance parameter of fire extinguishing spraying, adjusts a regulating valve of a fire extinguishing spray gun in real time according to the fluid control parameter set, and drives the fire extinguishing spray gun to perform fire extinguishing operation according to the spatial motion trajectory and the gun distance parameter until it is determined that the fire source is extinguished, so that the utilization rate of fire extinguishing agent and the fire extinguishing efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a control optimization method based on a self-propelled unmanned firefighting robot. Background Technology

[0002] With the acceleration of industrialization and urbanization, the complexity and danger of various fire scenarios are increasing, making the demand for unmanned and intelligent firefighting equipment more urgent. Although existing self-propelled firefighting robots can replace firefighters in entering high-risk areas, they still have significant shortcomings in autonomous decision-making, precise firefighting, and efficient operation.

[0003] Currently, Chinese patent CN120765435A discloses a cloud management method, system, equipment, and medium for fire extinguishing robots. It combines real-time robot status information with historical operation and maintenance data, and dynamically adjusts scheduling priorities through a secondary evaluation mechanism. First, it obtains robot status information, including GPS location, remaining battery power, and type of extinguishing medium, for preliminary priority determination. Then, it queries the equipment's operation and maintenance history to correct the priority. Finally, it allocates tasks and coordinates scheduling based on the corrected priority. This achieves real-time monitoring and scientific scheduling of equipment status, improves the coordination between robots, and enhances the stability and resource utilization efficiency of rescue missions.

[0004] The relevant technologies are insufficient to enable a single robot to successfully carry out autonomous firefighting operations after arriving at the core area of ​​a fire, and it is difficult to adapt to complex fire environments. Summary of the Invention

[0005] The technical problem solved by this invention is that existing technologies are difficult to adapt to complex fire environments, lack real-time safety obstacle avoidance capabilities in dynamically changing environments, and cannot perform three-dimensional accurate modeling of fire sources, which can easily lead to premature cessation of operations causing reignition or excessive spraying causing secondary disasters.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a control optimization method based on a self-propelled unmanned firefighting robot, comprising the following steps: Step S1: Receive the fire source coordinate information, and based on the improved RRT algorithm, perform target bias sampling and adaptive step size expansion in the global map to generate an initial discrete path and obtain the global reference trajectory. Step S2: Drive the robot to move according to the global reference trajectory, receive local ultrasonic detection data in real time, perform collision prediction and velocity space evaluation, and generate local avoidance control commands. Step S3: The local avoidance control command drives the machine to operate, starts 3D vision scanning to obtain the three-dimensional point cloud data of the fire source, and extracts and outputs the three-dimensional feature set of the fire source based on AI algorithm. Step S4: Based on the three-dimensional feature set of the fire source, map the size of the fire source to a set of fluid control parameters, and calculate the spatial motion trajectory and gun distance parameters of the fire extinguishing spray. Step S5: Adjust the regulating valve of the fire extinguishing spray gun in real time according to the fluid control parameter set, and drive the fire extinguishing spray gun to perform fire extinguishing operation according to the spatial movement trajectory and gun distance parameters until the fire source is determined to be extinguished.

[0007] Preferably, step S1 includes the following sub-steps: Step S11: Receive fire alarm command, obtain global grid map and initial coordinate matrix of fire source, set random sampling probability threshold, generate pseudo-random numbers when generating random tree nodes, and determine the pseudo-random numbers according to the random sampling probability threshold to obtain the random sampling points for the current round. When the pseudo-random number is less than the random sampling probability threshold, the initial coordinates of the fire source are used as the sampling point for target bias sampling. When the pseudo-random number is greater than or equal to the random sampling probability threshold, uniform random sampling is performed in the free space of the global grid map. Step S12: Based on the random sampling points of the current round, find the node with the smallest Euclidean distance to the current random sampling point in the current random tree as the nearest neighbor node. Calculate the minimum Euclidean distance between the nearest neighbor node and local obstacles using environmental perception data, construct an adaptive step size function, and dynamically calculate the current expansion step size. Extract the preset maximum base step size. When the minimum Euclidean distance is greater than or equal to the safety setting value, use the maximum base step size as the current expansion step size. When the minimum Euclidean distance is less than the safety setting value, calculate the ratio of the minimum Euclidean distance to the safety setting value as the base attenuation coefficient. At the same time, extract the collision penalty value of the new expansion node in the Gaussian expansion layer. Dynamically correct the base attenuation coefficient by combining it with the collision penalty value to obtain the final attenuation coefficient. Multiply the maximum base step size by the final attenuation coefficient to obtain the reduced current expansion step size. The current expansion step is extended from the nearest neighbor node along the direction pointing to the random sampling point to generate a new expansion node. At the same time, a Gaussian dilation layer is superimposed on the outer edge of the obstacle boundary. Collision penalty detection is performed on the new expansion node. If the detection is successful, the new expansion node is added to the random tree. The process is iterated until the random tree reaches the initial coordinate area of ​​the fire source to generate the initial discrete path. Step S13: Connect the non-adjacent nodes in the initial discrete path sequentially using the line-of-sight judgment method. If the connection does not cross the Gaussian dilation layer, remove the non-adjacent nodes as pruned path points to obtain a control point cluster. Construct cubic B-spline basis functions, map the control point cluster to a continuous parameterized equation, and output the global reference trajectory.

[0008] Preferably, step S2 includes the following sub-steps: Step S21: Control the robot to travel along the global reference trajectory, acquire the echo timestamp signals of multiple local ultrasonic detectors, calculate the flight time by combining the sound speed after environmental stabilization correction, obtain the local distance vector, map the local distance vector to the local coordinate system, the local coordinate system is established with the robot's centroid as the origin, and construct the environmental cost grid matrix of local dynamic obstacles. Step S22: Obtain the current dynamic state of the robot. Based on the robot's motion boundary, sample the feasible velocity space set for the next time step. Input the environmental cost grid matrix, the current dynamic state, and the global reference trajectory into the evaluation model of the dynamic window obstacle avoidance algorithm. Construct a cost evaluation function for scoring and evaluation, and output the optimal local avoidance control command.

[0009] Preferably, step S21 further includes the following sub-steps: Step S211: Extract the local distance vector obtained by calculating the round-trip propagation time of the sound wave, and perform Kalman filtering on the local distance vector to obtain a smoothed distance vector; Step S212: Based on the installation offset angle and coordinate offset matrix on the robot chassis, the smooth distance vector is transformed into a set of point cloud feature points in two-dimensional Euclidean space through a rotation and translation transformation matrix. Step S213: The point cloud feature point set is superimposed onto the local scrolling window at the current moment. The occupancy probability value of each grid is iteratively updated using a probability occupancy model. Grids with occupancy probability values ​​higher than a set threshold are marked as expansion collision risk zones. The output environment cost grid matrix is ​​then updated.

[0010] Preferably, step S22 further includes the following sub-steps: Step S221: Extract each candidate velocity pair from the feasible velocity space set. The candidate velocity pair includes the expected linear velocity and the expected angular velocity. Based on the kinematic derivation model, simulate the local predicted trajectory of each candidate velocity pair within the prediction time window. Step S222: Map each local predicted trajectory to the environmental cost grid matrix, combine it with the global reference trajectory, and substitute it into the cost evaluation function for weighted evaluation calculation to obtain the comprehensive cost value corresponding to each group of candidate velocities. Step S223: Select the candidate speed pair with the lowest overall cost as the optimal speed combination, convert the optimal speed combination into the expected rotational speed of the left and right wheels, generate the differential PWM signal of the left and right wheels as a local avoidance control command.

[0011] Preferably, step S3 includes the following sub-steps: Step S31: After the robot reaches the initial coordinate area of ​​the fire source and stops according to the local avoidance control command, it drives the 3D vision scanner to project structured light into the target area, captures the spatiotemporal deformation data of the reflected light spot, generates the original three-dimensional point cloud data, and converts the original three-dimensional point cloud data into the base coordinate system of the fire extinguishing spray gun through the coordinate system calibration matrix. Step S32: Input the converted 3D point cloud data into a 3D convolutional neural network model for high-dimensional spatial feature extraction and semantic segmentation to obtain a 3D feature set of the fire source.

[0012] Preferably, steps S31 and S32 further include the following sub-steps: Step S311: Divide the original three-dimensional point cloud data into a voxel grid of a set size, and replace all points in the grid with the three-dimensional centroid of the point cloud in each voxel grid to achieve point cloud downsampling; and use a statistical outlier removal algorithm based on K nearest neighbors to calculate the average distance from each point to its K neighboring points, and remove points with an average distance greater than the global distance threshold as smoke refraction noise. Step S312: Extract local and global geometric features of the filtered point cloud through the multi-scale pooling layer of the three-dimensional convolutional neural network; use a fully connected layer to map the geometric features into a probability vector for point-by-point classification, and segment out the flame core point cloud cluster with a probability greater than 0.8. Step S313: Calculate the covariance matrix of the flame core point cloud cluster, solve the eigenvalues ​​and eigenvectors of the covariance matrix through principal component analysis, and construct a directed bounding box along the principal component direction. Step S321: Extract the major axis, width axis and height axis of the directed bounding box, and calculate the maximum projected area of ​​the directed bounding box on the plane orthogonal to the line of sight of the fire extinguishing spray gun, as the fire source area; Step S322: Using the three-dimensional density distribution of the flame core point cloud cluster as a weight matrix, calculate the weighted centroid of the fire source as the initial three-dimensional coordinates of the fire source, and perform gravity fall compensation on the vertical component of the initial three-dimensional coordinates of the fire source to obtain the corrected three-dimensional coordinates of the fire source. Step S323: Combine the directed bounding box parameters, the fire source area, and the corrected three-dimensional coordinates of the fire source to obtain a three-dimensional feature set of the fire source.

[0013] Preferably, step S4 includes the following sub-steps: Step S41: Construct a fluid parameter mapping neural network based on radial basis functions, using the fire source area and fire source morphology feature vector as network input layer node data, and output corresponding virtual process parameters through nonlinear hidden layer mapping. The virtual process parameters include target air pressure value, target jet flow rate value and target jet cone angle. Step S42: Based on the atomization characteristic curve of the fire extinguishing spray nozzle and the target spray cone angle, calculate the range distance covering the projected cross section of the oriented bounding box based on the isosceles triangular projection model, and use it as the optimal gun distance parameter. Step S43: Using the corrected three-dimensional coordinates of the fire source as the center, and combining the optimal gun distance parameters, establish an operational reference plane. On the operational reference plane, according to the set trajectory overlap rate, generate an equally spaced reciprocating S-shaped smooth scanning path as the spatial motion trajectory of the fire extinguishing spray.

[0014] Preferably, step S5 specifically includes: The underlying microcontroller module analyzes the virtual process parameters and spatial motion trajectory, maps the target air pressure value and target jet flow rate value into multi-channel PWM drive duty cycle signals, and drives the electric proportional valve at the front end of the fire extinguishing nozzle to perform opening modulation in real time. The spatial motion trajectory is analyzed into an incremental position sequence of each joint of the mechanical gimbal, and a PID control strategy is used to drive the mechanical arm gimbal to move in sequence. The infrared flame sensor array is used to capture the time-domain decay curve of the infrared radiation intensity at the fire scene. When it is determined that the time-domain decay curve is within the range of the ambient noise threshold, an interruption request is triggered to stop the fire extinguishing action.

[0015] Preferably, step S5 further includes the following sub-steps: Step S51: Extract the target air pressure value and target injection flow rate value from the virtual process parameters, call the preset duty cycle fluid dynamics lookup table, and convert the target air pressure value and target injection flow rate value into a nonlinear PWM level sequence that controls the opening cross-sectional area of ​​the proportional solenoid valve. Step S52: Map the Cartesian coordinates of the spatial motion trajectory to the target rotation angle of the mechanical gimbal yaw motor and pitch motor, calculate the deviation between the current encoder feedback angle and the target rotation angle, substitute it into the incremental PID equation, and calculate the motor compensation output for each adjustment cycle. Step S53: The ADC sampling voltage output by the infrared flame sensor array is buffered over time. The moving average voltage value and variance of M consecutive sampling periods are calculated. When the moving average voltage value is lower than the extinguishing level threshold, it is determined that the fire source is completely extinguished. All solenoid valves are closed, and the global reference trajectory is reloaded to drive the robot to retreat in the opposite direction.

[0016] The beneficial effects of this invention are as follows: By introducing target bias sampling and an adaptive step size mechanism based on obstacle distance, the RRT path grows rapidly toward the target and automatically avoids obstacles; a Gaussian dilation layer is superimposed for collision penalty detection to ensure that the path retains sufficient safety margin; finally, redundant nodes are pruned using a line-of-sight criterion and smoothed using a cubic B-spline curve to output a continuous global reference trajectory without abrupt changes, significantly improving the smoothness of robot movement and passage efficiency; a local rolling occupancy grid map is constructed using an ultrasonic array, and the dynamic window method (DWA) is combined to perform multi-dimensional cost evaluation on multiple candidate trajectories in the velocity space, which can generate optimal local avoidance commands in real time that take into account target tendency, obstacle safety distance, and driving speed, effectively avoiding the oscillation and deadlock of traditional methods and ensuring driving safety in dynamic environments; high-density point clouds are obtained by 3D structured light scanning, and smoke refraction noise is effectively filtered out by voxel downsampling and K-nearest neighbor-based statistical filtering, and then semantic segmentation is performed by a three-dimensional convolutional neural network to accurately separate the core point cloud of the flame. A directed bounding box constructed based on principal component analysis accurately quantifies the shape, area, and three-dimensional centroid coordinates of the fire source, providing crucial data support for precise fire suppression. A nonlinear mapping model from fire source area and shape to optimal injection pressure, flow rate, and spray cone angle is established using a radial basis function neural network. Based on isosceles triangular projection relationships, the optimal range covering the entire fire source surface is calculated. On this basis, an equally spaced reciprocating S-shaped spatial motion trajectory is planned to ensure precise coverage of the entire cross-section of the fire source by the extinguishing agent, significantly improving agent utilization and fire suppression efficiency. Process parameters and spatial trajectories are directly mapped to valve PWM signals and gimbal motor control quantities, achieving hardware-level decoupling control of flow rate and pressure and high-precision trajectory tracking. Finally, physical extinguishing is determined by the statistical characteristics of the time-domain decay curve of the infrared flame sensor array. Automatic shutdown and retraction are performed after ensuring the fire source is completely extinguished without risk of reignition, achieving a fully unmanned autonomous operation closed loop. Attached Figure Description

[0017] Figure 1 The flowchart illustrates the steps of a control optimization method based on a self-propelled unmanned firefighting robot, as provided in one embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example, refer to Figure 1 This paper presents a control optimization method based on a self-propelled unmanned firefighting robot, which includes the following steps: Step S1: Receive the fire source coordinate information, and based on the improved RRT algorithm, perform target bias sampling and adaptive step size expansion in the global map to generate an initial discrete path and obtain the global reference trajectory. Step S2: Drive the robot to move according to the global reference trajectory, receive local ultrasonic detection data in real time, perform collision prediction and velocity space evaluation, and generate local avoidance control commands. Step S3: The local avoidance control command drives the machine to operate, starts 3D vision scanning to obtain the three-dimensional point cloud data of the fire source, and extracts and outputs the three-dimensional feature set of the fire source. Step S4: Based on the three-dimensional feature set of the fire source, map the size of the fire source to a set of fluid control parameters, and calculate the spatial motion trajectory and gun distance parameters of the fire extinguishing spray. Step S5: Adjust the regulating valve of the fire extinguishing spray gun in real time according to the fluid control parameter set, and drive the fire extinguishing spray gun to perform fire extinguishing operation according to the spatial movement trajectory and gun distance parameters until the fire source is determined to be extinguished.

[0020] Step S1 includes the following sub-steps: Step S11: Receive fire alarm command, obtain global grid map and initial coordinate matrix of fire source, set random sampling probability threshold, generate pseudo-random numbers when generating random tree nodes, and determine the pseudo-random numbers according to the random sampling probability threshold to obtain the random sampling points for the current round. When the pseudo-random number is less than the random sampling probability threshold, the initial coordinates of the fire source are used as the sampling point for target bias sampling. When the pseudo-random number is greater than or equal to the random sampling probability threshold, uniform random sampling is performed in the free space of the global grid map. Step S12: Based on the random sampling points of the current round, find the node with the smallest Euclidean distance to the current random sampling point in the current random tree as the nearest neighbor node. Calculate the minimum Euclidean distance between the nearest neighbor node and local obstacles using environmental perception data, construct an adaptive step size function, and dynamically calculate the current expansion step size. Extract the preset maximum base step size. When the minimum Euclidean distance is greater than or equal to the safety setting value, use the maximum base step size as the current expansion step size. When the minimum Euclidean distance is less than the safety setting value, calculate the ratio of the minimum Euclidean distance to the safety setting value as the base attenuation coefficient. At the same time, extract the collision penalty value of the new expansion node in the Gaussian expansion layer. Dynamically correct the base attenuation coefficient by combining it with the collision penalty value to obtain the final attenuation coefficient. Multiply the maximum base step size by the final attenuation coefficient to obtain the reduced current expansion step size. The current expansion step is extended from the nearest neighbor node along the direction pointing to the random sampling point to generate a new expansion node. At the same time, a Gaussian dilation layer is superimposed on the outer edge of the obstacle boundary. Collision penalty detection is performed on the new expansion node. If the detection is successful, the new expansion node is added to the random tree. The process is iterated until the random tree reaches the initial coordinate area of ​​the fire source to generate the initial discrete path. Step S13: Connect the non-adjacent nodes in the initial discrete path sequentially using the line-of-sight judgment method. If the connection does not cross the Gaussian dilation layer, remove the non-adjacent nodes as pruned path points to obtain a control point cluster. Construct cubic B-spline basis functions, map the control point cluster to a continuous parameterized equation, and output the global reference trajectory.

[0021] S11: The robot receives a fire alarm command via a wireless communication module. The command includes the initial coordinate matrix of the fire source (e.g., the coordinates of the fire source center in the global map coordinate system are (12.5m, 8.2m, 1.0m)) and the current global grid map of the warehouse or factory. The random sampling probability threshold is set to 0.15. In the node expansion loop of the Rapidly Expanding Random Tree (RRT), a uniformly distributed pseudo-random number rand between 0 and 1 is generated. When rand < 0.15, the initial coordinates of the fire source are directly used as the sampling point to achieve target-biased sampling; otherwise, sampling points are randomly generated in the free space of the global grid map.

[0022] S12: In each iteration, find the node on the random tree that is closest to the sampling point. Using the robot's onboard LiDAR or pre-acquired environmental point cloud data, calculate the minimum Euclidean distance between the nearest node and the nearest obstacle, and construct an adaptive step size function; Simultaneously, a Gaussian expansion layer is constructed around the boundary of the obstacle, with an expansion radius r_gauss = 0.3m. The collision cost at any point within the expansion layer is... ,in =0.1. When a new node x_new falls into this expansion layer and cost > 0.7, a collision penalty is considered to occur, and this expansion is abandoned. After multiple iterations, an initial discrete path point sequence from the starting point to the vicinity of the fire source is generated.

[0023] S13: The initial discrete path is pruned using a line-of-sight method. Starting from the origin, lines are sequentially connected to non-adjacent nodes. If a line does not cross any Gaussian dilation layer region (i.e., all points on the line are not within the dilation range), all redundant bypass nodes between these two nodes are removed. The pruned path points are extracted as a control point cluster P={P0,P1,…,Pn}. A cubic B-spline basis function is constructed, with order k=3. The node vectors are parameterized according to the cumulative chord length, mapping the control point cluster to a continuous parameterized curve C(u), u∈[0,1]. This curve is the smoothed global reference trajectory, which is sent to the chassis motion controller via the CAN bus.

[0024] Step S2 includes the following sub-steps: Step S21: Control the robot to travel along the global reference trajectory, acquire the echo timestamp signals of multiple local ultrasonic detectors, calculate the flight time by combining the sound speed after environmental stabilization correction, obtain the local distance vector, map the local distance vector to the local coordinate system, the local coordinate system is established with the robot's centroid as the origin, and construct the environmental cost grid matrix of local dynamic obstacles. Step S21 further includes the following sub-steps: Step S211: Extract the local distance vector obtained by calculating the round-trip propagation time of the sound wave, and perform Kalman filtering on the local distance vector to obtain a smoothed distance vector; Step S212: Based on the installation offset angle and coordinate offset matrix on the robot chassis, the smooth distance vector is transformed into a set of point cloud feature points in two-dimensional Euclidean space through a rotation and translation transformation matrix. Step S213: The point cloud feature point set is superimposed onto the local scrolling window at the current moment. The occupancy probability value of each grid is iteratively updated using a probability occupancy model. Grids with occupancy probability values ​​higher than a set threshold are marked as expansion collision risk zones. The output environment cost grid matrix is ​​then updated.

[0025] Step S22: Obtain the current dynamic state of the robot. Based on the robot's motion boundary, sample the feasible velocity space set for the next time step. Input the environmental cost grid matrix, the current dynamic state, and the global reference trajectory into the evaluation model of the dynamic window obstacle avoidance algorithm. Construct a cost evaluation function for scoring and evaluation, and output the optimal local avoidance control command.

[0026] Step S22 further includes the following sub-steps: Step S221: Extract each candidate velocity pair from the feasible velocity space set. The candidate velocity pair includes the expected linear velocity and the expected angular velocity. Based on the kinematic derivation model, simulate the local predicted trajectory of each candidate velocity pair within the prediction time window. Step S222: Map each local predicted trajectory to the environmental cost grid matrix, combine it with the global reference trajectory, and substitute it into the cost evaluation function for weighted evaluation calculation to obtain the comprehensive cost value corresponding to each group of candidate velocities. Step S223: Select the candidate speed pair with the lowest overall cost as the optimal speed combination, convert the optimal speed combination into the expected rotational speed of the left and right wheels, generate the differential PWM signal of the left and right wheels as a local avoidance control command.

[0027] S21: During the robot chassis's movement along the global reference trajectory, six ultrasonic sensors emit pulses at a frequency of 40kHz. Kalman filtering is applied to each echo timestamp signal, resulting in a process noise covariance matrix Q=0.01 and a measurement noise covariance R=0.1, yielding a smoothed distance vector d=[d1,d2,…,d6]. Based on the sensor's mounting offset angle on the chassis (e.g., the three front sensors are offset at -45°, 0°, and 45° respectively) and the coordinate offset matrix T_offset, the smoothed distance vector is transformed into a two-dimensional point cloud feature point set in the robot's centroid local coordinate system through rotation and translation transformations.

[0028] A local scrolling occupancy grid map centered on the robot, with a resolution of 0.1m, is generated. After overlaying feature points from the point cloud, a binarized occupancy update formula is used: if a point exists within the grid, the occupancy probability p_occ = 0.9; otherwise, it decays exponentially. Grids with an occupancy probability higher than 0.8 are marked as expansion collision risk zones, forming an environmental cost grid matrix.

[0029] S22: Input the environmental cost grid matrix and the robot's current linear velocity v = 0.5 m / s and angular velocity ω = 0 rad / s into the Dynamic Window (DWA) evaluation model. Sample in the velocity space with a linear velocity resolution of 0.05 m / s and an angular velocity resolution of 0.1 rad / s to obtain a set of feasible velocity pairs {(v_i, ω_j)}. For each candidate velocity pair, simulate the forward trajectory based on the kinematic model (assuming a prediction time dt = 2 s). The multidimensional cost evaluation function is: ; Where h is the alignment degree between the trajectory end and the target point, d is the closest distance between the trajectory and the obstacle (calculated using the cost grid matrix), ve is the velocity magnitude, and the weights are... =0.3, =0.5, =0.2, select the speed pair with the highest total value, and parse it into the PWM duty cycle instruction of the left and right drive wheels, as the local avoidance control instruction, so that the robot can safely reach the fire source operation point.

[0030] Step S3 includes the following sub-steps: Step S31: After the robot reaches the initial coordinate area of ​​the fire source and stops according to the local avoidance control command, it drives the 3D vision scanner to project structured light into the target area, captures the spatiotemporal deformation data of the reflected light spot, generates the original three-dimensional point cloud data, and converts the original three-dimensional point cloud data into the base coordinate system of the fire extinguishing spray gun through the coordinate system calibration matrix. Step S32: Input the converted 3D point cloud data into a 3D convolutional neural network model for high-dimensional spatial feature extraction and semantic segmentation to obtain a 3D feature set of the fire source.

[0031] Steps S31 and S32 further include the following sub-steps: Step S311: Divide the original three-dimensional point cloud data into a voxel grid of a set size, and replace all points in the grid with the three-dimensional centroid of the point cloud in each voxel grid to achieve point cloud downsampling; and use a statistical outlier removal algorithm based on K nearest neighbors to calculate the average distance from each point to its K neighboring points, and remove points with an average distance greater than the global distance threshold as smoke refraction noise. Step S312: Extract local and global geometric features of the filtered point cloud through the multi-scale pooling layer of the three-dimensional convolutional neural network; use a fully connected layer to map the geometric features into a probability vector for point-by-point classification, and segment out the flame core point cloud cluster with a probability greater than 0.8. Step S313: Calculate the covariance matrix of the flame core point cloud cluster, solve the eigenvalues ​​and eigenvectors of the covariance matrix through principal component analysis, and construct a directed bounding box along the principal component direction. Step S321: Extract the major axis, width axis and height axis of the directed bounding box, and calculate the maximum projected area of ​​the directed bounding box on the plane orthogonal to the line of sight of the fire extinguishing spray gun, as the fire source area; Step S322: Using the three-dimensional density distribution of the flame core point cloud cluster as a weight matrix, calculate the weighted centroid of the fire source as the initial three-dimensional coordinates of the fire source, and perform gravity fall compensation on the vertical component of the initial three-dimensional coordinates of the fire source to obtain the corrected three-dimensional coordinates of the fire source. Step S323: Combine the directed bounding box parameters, the fire source area, and the corrected three-dimensional coordinates of the fire source to obtain a three-dimensional feature set of the fire source.

[0032] S31: After the robot reaches the target point, its chassis locks, and the 3D structured light scanner projects an infrared coded grating onto the fire source area. The camera captures the deformation of the reflected light spot and calculates the original 3D point cloud, with approximately 150,000 points. The point cloud is then transformed to the fire extinguishing spray gun's base coordinate system using a calibrated hand-eye matrix. Next, downsampling is performed: the space is divided into a voxel grid with sides of 0.02m, and the original points are replaced with the 3D centroids of all points within each voxel, reducing the number of points to approximately 30,000. Then, a statistical outlier removal algorithm based on K-nearest neighbors is used, with K=50, to calculate the average distance from each point to its 50 neighboring points. and global average distance and standard deviation .like Then, the noise from smoke refraction is removed, resulting in the filtered point cloud.

[0033] S32: The filtered point cloud is input into a pre-trained 3D convolutional neural network (3D CNN). The network structure includes two convolutional layers (3×3×3 kernels, 32 and 64 channels), one multi-scale pooling layer (2×2×2 and 3×3×3 pooling kernels in parallel), a fully connected layer, and a Softmax output layer. After point-by-point classification by the network, the core flame point cloud clusters with a probability greater than 0.8 are extracted, approximately 8000 points.

[0034] S33: Calculate the covariance matrix of the flame core point cloud cluster, perform principal component analysis, and obtain three eigenvalues. and corresponding feature vectors .along The direction is the major axis. For wide axis, A directed bounding box (OBB) is constructed with the high axis as the reference point. The projected cross-sectional area of ​​the bounding box, S_fire, is calculated as the length of the major axis multiplied by the length of the minor axis. Using point cloud density as the weight, the weighted centroid coordinates of the fire source are calculated and verified by the line of sight. Coordinate offsets caused by occlusion are corrected to obtain the corrected 3D coordinates of the fire source. The directed bounding box parameters (length, width, height, and orientation angle), the fire source area S_fire, and the corrected 3D coordinates of the fire source are combined into a 3D feature set of the fire source.

[0035] Step S4 includes the following sub-steps: Step S41: Construct a fluid parameter mapping neural network based on radial basis functions, using the fire source area and fire source morphology feature vector as network input layer node data, and output corresponding virtual process parameters through nonlinear hidden layer mapping. The virtual process parameters include target air pressure value, target jet flow rate value and target jet cone angle. Step S42: Based on the atomization characteristic curve of the fire extinguishing spray nozzle and the target spray cone angle, calculate the range distance covering the projected cross section of the oriented bounding box based on the isosceles triangular projection model, and use it as the optimal gun distance parameter. Step S43: Using the corrected three-dimensional coordinates of the fire source as the center, and combining the optimal gun distance parameters, establish an operational reference plane. On the operational reference plane, according to the set trajectory overlap rate, generate an equally spaced reciprocating S-shaped smooth scanning path as the spatial motion trajectory of the fire extinguishing spray.

[0036] S41: Construct a fluid parameter mapping neural network based on radial basis functions (RBF). The network input layer consists of the fire source area S_fire and the fire source morphology feature vector (the proportions of the three axes of the OBB). The hidden layer uses 10 Gaussian radial basis function neurons, and the output layer is the target air pressure value. Target jet flow rate and target spray cone angle The network has been pre-trained with experimental data, and the root mean square error is less than 3%.

[0037] S42: Extract the atomization characteristic curve of the extinguishing agent nozzle (flow-pressure-cone angle relationship table). Based on the isosceles triangular projection model, when the nozzle axis is perpendicular to the fire source plane, the range L covering the OBB projection cross-section (i.e., the optimal nozzle distance) satisfies... ,in This represents the length of the major axis of the oriented bounding box projection. If the fire source is not perpendicular, coordinate transformation is used for correction.

[0038] S43: Using the corrected 3D coordinates of the fire source as a reference point, offset along its surface normal by the optimal gun distance L to establish a working reference plane perpendicular to the main direction of the fire source. Plan an equally spaced reciprocating S-shaped scanning path on the reference plane, with an overlap rate set to 30%. The distance between path points is equal to the effective spray width × (1 - overlap rate). Generate a series of pose points as the spatial motion trajectory of the fire extinguishing spray, including position and attitude quaternions.

[0039] Step S5 specifically includes: The underlying microcontroller module analyzes the virtual process parameters and spatial motion trajectory, maps the target air pressure value and target jet flow rate value into multi-channel PWM drive duty cycle signals, and drives the electric proportional valve at the front end of the fire extinguishing nozzle to perform opening modulation in real time. The spatial motion trajectory is analyzed into an incremental position sequence of each joint of the mechanical gimbal, and a PID control strategy is used to drive the mechanical arm gimbal to move in sequence. The infrared flame sensor array is used to capture the time-domain decay curve of the infrared radiation intensity at the fire scene. When it is determined that the time-domain decay curve is within the range of the ambient noise threshold, an interruption request is triggered to stop the fire extinguishing action.

[0040] Step S5 further includes the following sub-steps: Step S51: Extract the target air pressure value and target injection flow rate value from the virtual process parameters, call the preset duty cycle fluid dynamics lookup table, and convert the target air pressure value and target injection flow rate value into a nonlinear PWM level sequence that controls the opening cross-sectional area of ​​the proportional solenoid valve. Step S52: Map the Cartesian coordinates of the spatial motion trajectory to the target rotation angle of the mechanical gimbal yaw motor and pitch motor, calculate the deviation between the current encoder feedback angle and the target rotation angle, substitute it into the incremental PID equation, and calculate the motor compensation output for each adjustment cycle. Step S53: The ADC sampling voltage output by the infrared flame sensor array is buffered over time. The moving average voltage value and variance of M consecutive sampling periods are calculated. When the moving average voltage value is lower than the extinguishing level threshold, it is determined that the fire source is completely extinguished. All solenoid valves are closed, and the global reference trajectory is reloaded to drive the robot to retreat in the opposite direction.

[0041] S51: The underlying microcontroller outputs drive signals through the PWM module. Based on the target pressure P_target and flow rate Q_target, it queries a pre-set duty cycle-fluid dynamics lookup table in Flash memory. This table is experimentally calibrated to obtain the required PWM duty cycle for the proportional solenoid valve under different pressure and flow rates, achieving decoupled feedforward control of flow and pressure. Simultaneously, PID closed-loop fine-tuning ensures that the actual pressure sensor feedback value approaches the target value.

[0042] S52: The Cartesian pose points of the S-shaped scanning trajectory are converted into the target angle sequence of the mechanical gimbal's yaw and pitch motors through inverse kinematics. The main controller reads the current angle of the motor encoder at a frequency of 100Hz, calculates the deviation, and inputs it into the incremental PID controller (Kp=0.8, Ki=0.05, Kd=0.1) to output and adjust the PWM compensation amount, driving the gimbal to move along the trajectory.

[0043] S53: During firefighting, an infrared flame sensor array (detection wavelength 4.3μm) positioned at the front of the robot collects radiation intensity at a sampling rate of 1kHz. The ADC voltage values ​​are buffered over time, and the average voltage across a sliding window is calculated every 50ms. and variance Preset off level threshold =0.15V (corresponding to radiation value in a flameless environment). Over 10 consecutive cycles... and Upon determining that the fire source has been physically extinguished, the main control chip immediately triggers an interrupt, closes all solenoid valves, stops the spraying of extinguishing agent, and reloads the global reference trajectory generated in step S1, driving the robot to retreat back along the original path to a safe area, thus completing the operation.

[0044] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A control optimization method based on a self-propelled unmanned firefighting robot, characterized in that, Includes the following steps: Step S1: Receive the fire source coordinate information, and based on the improved RRT algorithm, perform target bias sampling and adaptive step size expansion in the global map to generate an initial discrete path and obtain the global reference trajectory. Step S2: Drive the robot to move according to the global reference trajectory, receive local ultrasonic detection data in real time, perform collision prediction and velocity space evaluation, and generate local avoidance control commands. Step S3: The local avoidance control command drives the machine to operate, starts 3D vision scanning to obtain the three-dimensional point cloud data of the fire source, and extracts and outputs the three-dimensional feature set of the fire source. Step S4: Based on the three-dimensional feature set of the fire source, map the size of the fire source to a set of fluid control parameters, and calculate the spatial motion trajectory and gun distance parameters of the fire extinguishing spray. Step S5: Adjust the regulating valve of the fire extinguishing spray gun in real time according to the fluid control parameter set, and drive the fire extinguishing spray gun to perform fire extinguishing operation according to the spatial movement trajectory and gun distance parameters until the fire source is determined to be extinguished.

2. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: Receive fire alarm command, obtain global grid map and initial coordinate matrix of fire source, set random sampling probability threshold, generate pseudo-random numbers when generating random tree nodes, and determine the pseudo-random numbers according to the random sampling probability threshold to obtain the random sampling points for the current round. When the pseudo-random number is less than the random sampling probability threshold, the initial coordinates of the fire source are used as the sampling point for target bias sampling. When the pseudo-random number is greater than or equal to the random sampling probability threshold, uniform random sampling is performed in the free space of the global grid map. Step S12: Based on the random sampling points of the current round, find the node with the smallest Euclidean distance to the current random sampling point in the current random tree as the nearest neighbor node. Calculate the minimum Euclidean distance between the nearest neighbor node and local obstacles using environmental perception data, construct an adaptive step size function, and dynamically calculate the current expansion step size. Extract the preset maximum base step size. When the minimum Euclidean distance is greater than or equal to the safety setting value, use the maximum base step size as the current expansion step size. When the minimum Euclidean distance is less than the safety setting value, calculate the ratio of the minimum Euclidean distance to the safety setting value as the base attenuation coefficient. At the same time, extract the collision penalty value of the new expansion node in the Gaussian expansion layer. Combine the base attenuation coefficient with the collision penalty value for dynamic correction to obtain the final attenuation coefficient. Multiply the maximum base step size by the final attenuation coefficient to obtain the reduced current expansion step size. The current expansion step is extended from the nearest neighbor node along the direction pointing to the random sampling point to generate a new expansion node. At the same time, a Gaussian dilation layer is superimposed on the outer edge of the obstacle boundary. Collision penalty detection is performed on the new expansion node. If the detection is successful, the new expansion node is added to the random tree. The process is iterated until the random tree reaches the initial coordinate area of ​​the fire source to generate the initial discrete path. Step S13: Connect the non-adjacent nodes in the initial discrete path sequentially using the line-of-sight judgment method. If the connection does not cross the Gaussian dilation layer, remove the non-adjacent nodes as pruned path points to obtain a control point cluster. Construct cubic B-spline basis functions, map the control point cluster to a continuous parameterized equation, and output the global reference trajectory.

3. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S21: Control the robot to travel along the global reference trajectory, acquire the echo timestamp signals of multiple local ultrasonic detectors, calculate the flight time by combining the sound speed after environmental stabilization correction, obtain the local distance vector, map the local distance vector to the local coordinate system, the local coordinate system is established with the robot's centroid as the origin, and construct the environmental cost grid matrix of local dynamic obstacles. Step S22: Obtain the current dynamic state of the robot. Based on the robot's motion boundary, sample the feasible velocity space set for the next time step. Input the environmental cost grid matrix, the current dynamic state, and the global reference trajectory into the evaluation model of the dynamic window obstacle avoidance algorithm. Construct a cost evaluation function for scoring and evaluation, and output the optimal local avoidance control command.

4. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 3, characterized in that, Step S21 further includes the following sub-steps: Step S211: Extract the local distance vector obtained by calculating the round-trip propagation time of the sound wave, and perform Kalman filtering on the local distance vector to obtain a smoothed distance vector; Step S212: Based on the installation offset angle and coordinate offset matrix on the robot chassis, the smooth distance vector is transformed into a set of point cloud feature points in two-dimensional Euclidean space through a rotation and translation transformation matrix. Step S213: The point cloud feature point set is superimposed onto the local scrolling window at the current moment. The occupancy probability value of each grid is iteratively updated using a probability occupancy model. Grids with occupancy probability values ​​higher than a set threshold are marked as expansion collision risk zones. The output environment cost grid matrix is ​​then updated.

5. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 4, characterized in that, Step S22 further includes the following sub-steps: Step S221: Extract each candidate velocity pair from the feasible velocity space set. The candidate velocity pair includes the expected linear velocity and the expected angular velocity. Based on the kinematic derivation model, simulate the local predicted trajectory of each candidate velocity pair within the prediction time window. Step S222: Map each local predicted trajectory to the environmental cost grid matrix, combine it with the global reference trajectory, and substitute it into the cost evaluation function for weighted evaluation calculation to obtain the comprehensive cost value corresponding to each group of candidate velocities. Step S223: Select the candidate speed pair with the lowest overall cost as the optimal speed combination, convert the optimal speed combination into the expected rotational speed of the left and right wheels, generate the differential PWM signal of the left and right wheels as a local avoidance control command.

6. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 5, characterized in that, Step S3 includes the following sub-steps: Step S31: After the robot reaches the initial coordinate area of ​​the fire source and stops according to the local avoidance control command, it drives the 3D vision scanner to project structured light into the target area, captures the spatiotemporal deformation data of the reflected light spot, generates the original three-dimensional point cloud data, and converts the original three-dimensional point cloud data into the base coordinate system of the fire extinguishing spray gun through the coordinate system calibration matrix. Step S32: Input the converted 3D point cloud data into a 3D convolutional neural network model for high-dimensional spatial feature extraction and semantic segmentation to obtain a 3D feature set of the fire source.

7. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 6, characterized in that, Steps S31 and S32 further include the following sub-steps: Step S311: Divide the original three-dimensional point cloud data into a voxel grid of a set size, and replace all points in the grid with the three-dimensional centroid of the point cloud in each voxel grid to achieve point cloud downsampling; and use a statistical outlier removal algorithm based on K nearest neighbors to calculate the average distance from each point to its K neighboring points, and remove points with an average distance greater than the global distance threshold as smoke refraction noise. Step S312: Extract local and global geometric features of the filtered point cloud through the multi-scale pooling layer of the three-dimensional convolutional neural network; By using fully connected layers to map geometric features into point-by-point classification probability vectors, flame core point cloud clusters with a probability greater than 0.8 are segmented out; Step S313: Calculate the covariance matrix of the flame core point cloud cluster, solve the eigenvalues ​​and eigenvectors of the covariance matrix through principal component analysis, and construct a directed bounding box along the principal component direction. Step S321: Extract the major axis, width axis and height axis of the directed bounding box, and calculate the maximum projected area of ​​the directed bounding box on the plane orthogonal to the line of sight of the fire extinguishing spray gun, as the fire source area; Step S322: Using the three-dimensional density distribution of the flame core point cloud cluster as a weight matrix, calculate the weighted centroid of the fire source as the initial three-dimensional coordinates of the fire source, and perform gravity fall compensation on the vertical component of the initial three-dimensional coordinates of the fire source to obtain the corrected three-dimensional coordinates of the fire source. Step S323: Combine the directed bounding box parameters, the fire source area, and the corrected three-dimensional coordinates of the fire source to obtain a three-dimensional feature set of the fire source.

8. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 7, characterized in that, Step S4 includes the following sub-steps: Step S41: Construct a fluid parameter mapping neural network based on radial basis functions, using the fire source area and fire source morphology feature vector as network input layer node data, and output corresponding virtual process parameters through nonlinear hidden layer mapping. The virtual process parameters include target air pressure value, target jet flow rate value and target jet cone angle. Step S42: Based on the atomization characteristic curve of the fire extinguishing spray nozzle and the target spray cone angle, calculate the range distance covering the projected cross section of the oriented bounding box based on the isosceles triangular projection model, and use it as the optimal gun distance parameter. Step S43: Using the corrected three-dimensional coordinates of the fire source as the center, and combining the optimal gun distance parameters, establish an operational reference plane. On the operational reference plane, according to the set trajectory overlap rate, generate an equally spaced reciprocating S-shaped smooth scanning path as the spatial motion trajectory of the fire extinguishing spray.

9. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 8, characterized in that, Step S5 specifically includes: The underlying microcontroller module analyzes the virtual process parameters and spatial motion trajectory, maps the target air pressure value and target jet flow rate value into multi-channel PWM drive duty cycle signals, and drives the electric proportional valve at the front end of the fire extinguishing nozzle to perform opening modulation in real time. The spatial motion trajectory is analyzed into an incremental position sequence of each joint of the mechanical gimbal, and a PID control strategy is used to drive the mechanical arm gimbal to move in sequence. The infrared flame sensor array is used to capture the time-domain decay curve of the infrared radiation intensity at the fire scene. When it is determined that the time-domain decay curve is within the range of the ambient noise threshold, an interruption request is triggered to stop the fire extinguishing action.

10. The control optimization method based on a self-propelled unmanned firefighting robot as described in claim 9, characterized in that, Step S5 further includes the following sub-steps: Step S51: Extract the target air pressure value and target injection flow rate value from the virtual process parameters, call the preset duty cycle fluid dynamics lookup table, and convert the target air pressure value and target injection flow rate value into a nonlinear PWM level sequence that controls the opening cross-sectional area of ​​the proportional solenoid valve. Step S52: Map the Cartesian coordinates of the spatial motion trajectory to the target rotation angle of the mechanical gimbal yaw motor and pitch motor, calculate the deviation between the current encoder feedback angle and the target rotation angle, substitute it into the incremental PID equation, and calculate the motor compensation output for each adjustment cycle. Step S53: The ADC sampling voltage output by the infrared flame sensor array is buffered over time. The moving average voltage value and variance of M consecutive sampling periods are calculated. When the moving average voltage value is lower than the extinguishing level threshold, it is determined that the fire source is completely extinguished. All solenoid valves are closed, and the global reference trajectory is reloaded to drive the robot to retreat in the opposite direction.

Images