ICU aerosol mobile disinfection robot control method based on space-time situation map

Abstract

The application discloses an ICU aerosol mobile disinfection robot control method based on a space-time situation map and belongs to the disinfection robot field. The method comprises the following steps: constructing an initial space-time situation map; completing repositioning through visual monitoring and a sensor, creating an obstacle object and marking a character type; generating a plurality of disinfection target points, deciding on a target based on disinfection necessity scoring, and adopting an improved environmental risk weighting algorithm to complete preliminary path planning; combining a local obstacle avoidance strategy and a regional dynamic speed limit during the advancing process, setting a hard real-time event preemption interrupt mechanism, automatically changing the region into an emergency operation area when detecting that medical staff quickly gather, controlling the robot to immediately drive away and reselect a safe target; after reaching the target, quantitatively estimating the staying time according to aerosol clusters and implementing closed-loop confirmation; and cyclically executing a disinfection task and optimizing the prediction accuracy by means of gradient self-adaptive correction parameters of a reinforcement learning reward function.

Description

Technical Field

[0001] This invention belongs to the field of disinfection robots, and in particular relates to a control method for a mobile disinfection robot for ICU aerosols based on a spatiotemporal situation map. Background Technology

[0002] With the widespread adoption of automated and intelligent equipment in the field of medical infection control, mobile disinfection robots have become an important means of preventing aerosol transmission and cross-infection within hospitals. Existing mobile disinfection robots typically carry lidar or ultrasonic sensors, constructing a static geometric grid map of the hospital environment through Simultaneous Localization and Mapping (SLAM) algorithms. During operation, they perform timed and targeted air disinfection operations in wards or corridors based on preset navigation paths and basic local obstacle avoidance algorithms (such as Dynamic Window Method (DWA)). However, in the Intensive Care Unit (ICU), a special medical scenario characterized by the most complex environment, the smallest space, high dynamism, and high population density within a hospital, existing mobile disinfection robot control methods have revealed the following technical shortcomings in practical applications:

[0003] First, existing mapping and environmental cognition methods lack deep semantic perception, resulting in insufficient scene recognition capabilities. Traditional spatial grid maps only contain simple geometric obstacle boundary information, which cannot accurately distinguish between movable obstacles such as hospital beds and trolleys and immovable obstacles such as ward doors and walls in complex ICU environments. They also cannot perform intelligent topological area functional division for corridors, areas with concentrated medical equipment, and medical staff operation areas. This leads to robots' inability to understand environmental attributes and their difficulty in implementing dynamic and refined behavior and speed control based on the actual collision risks in different medical areas.

[0004] Second, the decision-making regarding disinfection target points lacks consideration of the dynamic evolution of space and time, resulting in poor targeting. Most existing robots adopt fixed-track cruising or blindly uniform coverage strategies, failing to project the real-time diffusion status of aerosols in space onto the map, nor considering the global path cost of robot movement or the semantic priority gain around critically ill patients. This prevents disinfection resources from being dynamically allocated to high-risk areas, easily leading to disinfection blind spots, and making it impossible to achieve closed-loop confirmation based on dynamic changes in aerosol concentration.

[0005] Third, existing obstacle avoidance response mechanisms are inadequate to handle real-time emergency medical events within the ICU, posing serious safety risks. Emergency rescue events are frequent in the ICU; in the event of an emergency, medical staff and a large number of rescue equipment will rapidly gather around a specific patient's bed within a short period. Existing disinfection robots, when faced with such highly dynamic gatherings of people, generally only possess passive obstacle avoidance capabilities, i.e., they remain stationary and wait, or attempt to calculate local detours nearby. In the extremely confined core passageways and operating areas of the ICU, this behavior not only fails to effectively avoid obstacles but also easily leads to collisions, potentially directly blocking core treatment channels and delaying rescue opportunities.

[0006] Therefore, how to construct a mobile disinfection robot control method that can integrate multi-source heterogeneous data and dynamically evolve aerosol spatiotemporal situation observation objects in real time, and achieve hard real-time interruption preemption and control the robot to actively retreat at full speed to avoid the core channel when facing sudden rescue events in the ICU, is a key technical problem that urgently needs to be solved in the field of intelligent medical equipment. Summary of the Invention

[0007] To address the technical deficiencies in the existing technology, the present invention aims to provide a control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map, specifically in the following aspects:

[0008] The purpose of this invention is to achieve accurate semantic obstacle and regional risk factor classification through the Fast-lio2 algorithm, extrinsic parameter calibration of depth cameras, and multi-view Bayesian fusion, and to improve environmental risk weighting. The algorithm incorporates movement risk factors and dynamic speed limiting mechanisms in high-risk areas to improve the robot's movement safety and collision avoidance robustness in complex and high-risk medical environments.

[0009] The purpose of this invention is to dynamically distinguish personnel types by combining visual monitoring with spatiotemporal situation maps, and to introduce a patient semantic association gain factor into the calculation of the necessity of disinfection, so as to eliminate the blindness of target point decision-making and achieve the purpose of priority, timely and accurate protection for special high-risk patient groups.

[0010] The purpose of this invention is to set up a hard real-time event preemption and interruption mechanism, which automatically changes the emergency work area and controls the robot to leave immediately when a rapid gathering of medical personnel is detected, and reassesses the safety disinfection target point, so as to improve the emergency response capability for sudden rescue events and avoid blocking traffic by staying in the core channel.

[0011] The purpose of this invention is to dynamically and quantitatively calculate the estimated residence time based on the aerosol cluster status, implement a closed-loop confirmation procedure based on situation map judgment, and use gradient descent to adaptively correct system parameters by introducing a reinforcement learning reward function, so as to minimize the error between the predicted disinfection time and the actual time and maximize the overall cyclic disinfection scheduling efficiency of the system.

[0012] The specific technical solution is as follows:

[0013] A control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map includes the following steps:

[0014] Step S101: The robot moves within the scene and scans with multi-line radar. It uses the SLAM algorithm to build a scene map, selects machine learning methods to segment semantic targets, and constructs an initial spatiotemporal situation map. The spatiotemporal situation map is a dynamic perception network that dynamically overlays obstacle attributes, functional area divisions, personnel category labels, and real-time aerosol concentration and distribution radius on a grid map.

[0015] Step S102: Fit the spatiotemporal situation map by combining visual monitoring with the robot's sensors, complete the relocalization, create scene obstacle objects, label the human type and divide the area;

[0016] Step S103: Generate multiple disinfection target points based on the aerosol monitoring distribution data map. Based on the distance of each target point, the characteristics of the area and the aerosol concentration, and combined with the process scheduling algorithm, select the next disinfection target point to complete the preliminary path planning.

[0017] Step S104: Combining the spatiotemporal situation map and multi-source sensor perception information, a local obstacle avoidance strategy is adopted while moving forward, and different upper limits for movement speed are set according to the area where the robot is located.

[0018] Step S105: After arriving near the disinfection target, synchronize aerosol monitoring distribution data, carry out fixed-point disinfection under the system's estimated time, and determine whether the aerosol level has dropped below the threshold through the spatiotemporal situation map until the disinfection task is confirmed to be completed.

[0019] Step S106: Select the next disinfection target point and repeat steps S104 and S105 in a loop. During the process, reinforcement learning is used to improve the rationality of the disinfection time prediction and optimize the target point selection.

[0020] In step S104, a hard real-time event preemption interruption mechanism is set. If medical staff in a certain area of ​​the spatiotemporal situation map gather rapidly in a short period of time and the number exceeds a set threshold, the area is temporarily changed to an emergency work area. If the robot is currently working in the emergency work area, it immediately leaves and re-triggers step S103, and selects another safe target point that is far away from the emergency work area and has the lowest reachability path cost to perform the disinfection task.

[0021] The present invention has the following beneficial effects:

[0022] First, through multi-view Bayesian fusion of multi-line radar and depth cameras, precise semantic segmentation of movable and immovable obstacles was achieved, and specific risk functional areas in the ICU were automatically identified. This, combined with improvements... The regional movement risk factor and regional dynamic speed limiting mechanism introduced in the algorithm technically achieve proactive avoidance of physical operation risks of core life support equipment, significantly improving the robot's collision avoidance robustness and physical safety in high-risk medical environments.

[0023] Meanwhile, by performing multi-dimensional feature recognition and attribute labeling on personnel within the scene, a patient semantic association gain factor based on the reciprocal of spatial distance is innovatively introduced into the disinfection necessity scoring model, thereby effectively eliminating the blindness of traditional path planning and realizing high-priority adaptive, immediate and precise protection for special high-risk patient groups.

[0024] Furthermore, in response to the high frequency of sudden emergency events in the ICU, this invention sets up a hard real-time event preemption and interruption mechanism. When external monitoring or situation map detects that medical staff are rapidly gathering in a short period of time, the area can be automatically changed into an emergency work zone. The robot is controlled to switch directly from the conventional passive obstacle avoidance procedure to the highest priority immediate active departure procedure. Technically, this greatly improves the response speed and active avoidance capability of sudden emergency events, and avoids the technical defects of robots blocking traffic by staying in the core rescue channel.

[0025] Finally, the system quantitatively derives the disinfection time based on the real-time concentration and cluster status of aerosols, and uses a reinforcement learning reward function that includes the total time consumption and prediction error to perform gradient descent iteration. In this way, the absolute error of time prediction is minimized through adaptive closed-loop means, thereby maximizing the scheduling efficiency of the system's overall multi-target cyclic disinfection and the quality of air purification technology while preventing over-disinfection and incomplete disinfection. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention.

[0027] Figure 2 Flowchart for constructing a spatiotemporal situation map.

[0028] Figure 3 Flowchart for handling ICU emergency interruptions.

[0029] Figure 4 Flowchart for updating parameters for disinfection prediction time. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0031] This invention provides a control method for a mobile aerosol disinfection robot in an ICU based on a spatiotemporal situation map, such as... Figure 1 As shown, it includes the following steps:

[0032] Step S101: The robot moves within the scene and scans using multi-line radar. A scene map is built using the SLAM (Simultaneous Localization and Mapping) algorithm. By applying SLAM, the robot can autonomously locate itself and construct a global map of its surroundings based on its sensor data while moving in an unknown environment and location. Then, a suitable machine learning method is selected to segment semantic targets based on features, constructing an initial spatiotemporal situation map. This spatiotemporal situation map refers to a dynamic perception network that dynamically overlays obstacle attributes, specific functional area divisions, real-time personnel category annotations, and real-time concentration and distribution radius of three-dimensional aerosols onto a grid map projected onto a two-dimensional 3D space in the ICU.

[0033] The disinfection robot is equipped with a mobile remote controller, allowing it to move around the accessible area within the intensive care unit. It uses the Fast-lio2 algorithm to acquire point cloud maps and captures images with a depth camera during movement. Fast-lio2 is a highly efficient and robust tightly coupled lidar inertial odometry calculation method. It utilizes data from a multi-line lidar and an IMU (Inertial Measurement Unit), achieving high-frequency 3D state estimation through Kalman filtering on a manifold. Therefore, extrinsic parameter calibration of the IMU and lidar is required. The calibration method uses manually coarsely measured spatial relative positions as initial values, combined with a joint optimization algorithm based on its own motion for dynamic fine calibration. The joint optimization algorithm is a nonlinear least squares-based calibration-board-less dynamic fine calibration method. It uses the manually coarsely measured relative positions as initial values ​​and iteratively solves the problem by comparing the motion trajectory calculated by the lidar during robot movement with the motion trajectory pre-integrated by the IMU, aiming to minimize the physical error between the two. This allows for a one-time high-precision correction and calculation of the 3D rotation or translation extrinsic parameters between the lidar and IMU, as well as the IMU's zero bias.

[0034] By performing frame-by-frame 2D semantic segmentation and multi-view Bayesian fusion, the 2D semantic segmentation probability results of multiple frames are accurately fused with the 3D point cloud map in terms of space and probability. This ultimately completes the semantic target segmentation of the entire 3D point cloud map, and based on this, a 2D raster map is created, named a spatiotemporal situation map. The original obstacles in the intensive care unit scene are divided into two types: movable obstacles and immovable obstacles. Movable obstacles include hospital beds, medical equipment, and mobile shelves, while immovable obstacles include ward doors, walls, and pillars.

[0035] Based on the generated initial spatiotemporal situation map, the intensive care unit is automatically divided into different areas. The rectangular area perpendicular to the beds and with a length close to the total length of the ward is designated as the aisle area; the area with a large number of medical devices concentrated between the beds is the medical equipment concentration area; and the remaining space between the beds is the medical staff operation area.

[0036] This process requires extrinsic parameter calibration of the radar and depth camera. The FAST-Calib method is used to obtain a fixed rigid body transformation, which is used to reliably project camera pixel observations onto the radar point cloud coordinate system. The FAST-Calib method is used to solve for the fixed extrinsic parameter rigid body transformation matrix from the camera coordinate system to the lidar coordinate system, achieving reliable projection of image pixels onto the radar point cloud. The formula for the fixed extrinsic parameter matrix obtained through FAST-Calib calibration is as follows:

[0037] ;

[0038] in, These are the coordinates of a 3D point in the depth camera coordinate system. These are the three-dimensional point coordinates in the lidar coordinate system. Let be the rigid body transformation matrix. for The rotation matrix, for The translation vector. Using this formula, the camera's pixel observations can be reliably projected into the radar point cloud coordinate system.

[0039] Step S102: By combining visual monitoring with a spatiotemporal situation map for fitting, and using sensors such as radar or ultrasonic waves carried by the collaborative robot for relocalization, the creation of scene obstacle objects, character labeling, and region division under visual monitoring are completed.

[0040] The system uses the video footage received from the monitoring system to help it relocate itself. Based on the existing spatiotemporal situation map in the initial state, it completes real-time target identification and creates observation objects for movable targets in the spatiotemporal situation map. If a possible target movement is detected in the monitoring, this information is fed back to the disinfection control platform, and the platform waits for the disinfection robot to arrive near the target area for verification. Once the change is confirmed, it is synchronized to the updated spatiotemporal situation map.

[0041] At the same time, by identifying human targets appearing in the monitoring, the human information is synchronized to the spatiotemporal situation map. The human types are comprehensively distinguished based on the multi-dimensional visual features and real-time behavioral features of the identified human, and the human types are divided into patients, medical staff and other personnel.

[0042] The relocation of the monitoring system employs the control point method. The control point method involves manually specifying or extracting several sets of one-to-one corresponding feature points (control points) in a known space and image, and then using geometric transformation equations to solve for the spatial mapping relationship using a matrix solution method. In this invention, the control point method is used to calculate the alignment relationship between the fixed monitoring screen plane and the physical plane of the spatiotemporal situation map. The main method for relocation of the disinfection robot is to match the point cloud obtained by multi-line radar scanning, geometrically aligning the local point cloud obtained from the current frame scan with the global point cloud map. The alignment formula between the fixed monitoring screen and the map is:

[0043] ;

[0044] in, To fix the pixel coordinates under the monitoring screen, ( , () represents the two-dimensional physical coordinates in the spatiotemporal situation diagram. The solution obtained by the control point method Homography matrix.

[0045] Step S103: Based on the aerosol monitoring distribution data map, generate multiple disinfection target points under a certain set threshold. Based on the current distance of each target point, the characteristics of the area where it is located, and the aerosol concentration, and combined with the next disinfection target point selected by the process scheduling algorithm, complete the preliminary path planning.

[0046] Aerosol distribution data is acquired through an aerosol monitoring device, which is a sensor hardware and supporting system used for real-time detection and quantitative analysis of aerosol concentration and distribution range in space. This data map displays real-time monitoring data of aerosol distribution in three-dimensional space. The aerosol distribution in this data map is then projected two-dimensionally onto the initial spatiotemporal situation map to form a complete spatiotemporal situation map. The data sources include multi-line radar data from the robot, images captured by the robot's depth camera, information from external fixed visual monitoring screens, multi-dimensional and real-time behavioral characteristics of identified individuals, and real-time distribution data provided by the aerosol monitoring device, such as... Figure 2 As shown. The target point and duration of disinfection are determined by the aerosol concentration at that location in the spatiotemporal situation map. With distribution radius Decide.

[0047] Each disinfection task requires selecting the current disinfection target, and the next target point is selected only after the disinfection task is completed. The selection of disinfection targets must be based on the necessity of disinfection. Factors influencing the necessity of disinfection include the concentration at the target point, the area where the target is located, the predicted duration of the disinfection task, the concentration of people, the movement cost required for preliminary path planning calculated from the obstacle distribution on the spatiotemporal situation map, and the concentration change trend near the target point. Simultaneously, because patients in the ward environment possess the combined characteristics of being both high-risk susceptible individuals and potential sources of dynamic bioaerosols, the priority of aerosol disinfection around patients needs to be increased. Therefore, the specific formula for calculating the necessity of disinfection is as follows:

[0048] ;

[0049] in, For target point The comprehensive score for the necessity of disinfection is a comprehensive mathematical score used to quantitatively assess the urgency and importance of disinfection for a candidate target point. It comprehensively considers six dimensions, including concentration, radius, path cost, regional risk, concentration trend, and patient benefit. This represents the real-time aerosol concentration at that point. The radius of distribution of aerosol clusters; The cost of the robot's planned path to the target point; To determine the movement risk factor for the area where this point is located, the risk factor for the corridor area is set to low, while the risk factor for the medical equipment concentration area is set to extremely high. This shows the trend of aerosol concentration changes near this point; This is a semantic association gain factor for patients, used to forcibly increase the disinfection priority around patients based on spatial distance; to These are the preset weighting coefficients for each indicator. Among them, the patient semantic association gain factor... The calculation formula is:

[0050] ;

[0051] in, The physical distance from the target point to the nearest patient. The smallest positive number is set to prevent the denominator from being zero.

[0052] After deciding on the target points for disinfection, the preliminary global path planning involved adopts an improved environmental risk-weighted approach. Algorithm complete. The algorithm is used to calculate the actual cost from the starting point to the current point and the estimated cost to the destination. This invention... The algorithm's cost evaluation function directly incorporates the product of the regional movement risk factor and the path search inflation factor to form an improved environmental risk weighting. The algorithm enables the generated global path to adaptively and proactively bypass high-risk areas with dense medical equipment in terms of geometry. The cost evaluation function formula is as follows:

[0053] ;

[0054] in, The actual physical distance traveled from the starting point to the current node n is the cost. This represents the estimated cost from the current node n to the target point.

[0055] Step S104: Combining the spatiotemporal situation map and multi-source sensor perception information, adopt a local obstacle avoidance strategy to prevent collisions while moving forward.

[0056] Based on the existing preliminary path planning, the robot moves forward. If the sensors detect an obstacle ahead, the movement speed is reduced for assessment. If the obstacle has left the planned path within a short time, the original movement speed is resumed. If the obstacle still obstructs the path after this time, a new path to the target is planned based on the current information from the spatiotemporal situation map. The local obstacle avoidance strategy employs the Dynamic Window Approach (DWA) algorithm, a classic algorithm for local path planning and real-time obstacle avoidance in mobile robots. It calculates feasible speed combinations within a short time based on the robot's current acceleration constraints and selects the path that is furthest from the obstacle and fastest in approaching the target point after predicting several short trajectories. An evaluation function is involved. The specific formula is as follows:

[0057] ;

[0058] in, This represents the linear velocity of the robot. Let ω be the robot's angular velocity. This is the azimuth evaluation function, used to evaluate the degree of alignment between the end of the trajectory and the target point after the robot travels at the predicted speed. These are the weighting coefficients of the azimuth evaluation function, used to adjust the robot's preference for approaching the target; Here is the obstacle distance evaluation function, which represents the minimum physical distance between the robot and the nearest obstacle on the predicted trajectory at this speed; These are the weighting coefficients of the obstacle distance evaluation function, used to adjust the robot's obstacle avoidance safety redundancy. This is the velocity evaluation function, representing the magnitude of the currently sampled linear velocity, used to encourage the robot to maintain high driving efficiency. The higher the velocity, the higher the score. These are the weighting coefficients of the speed evaluation function, used to adjust the robot's pursuit of mobility efficiency.

[0059] Additionally, based on the robot's current location, set upper limits for movement speed in different areas. The aisle area has a normal upper limit. The operating area for medical staff is 0.75. The maximum value for the medical equipment concentration area is 0.2Vmax.

[0060] Due to potential emergencies in intensive care unit scenarios, a hard real-time event preemption and interruption mechanism is required. If medical personnel rapidly gather in a certain area of ​​the spatiotemporal situation map within a short period and the number exceeds a set threshold, that area is temporarily designated as an emergency work zone. If the disinfection robot is currently working in the emergency work zone, it immediately leaves the area and re-triggers the third step of the disinfection necessity assessment. It prioritizes other non-contaminated, safe target points far from the emergency work zone with the lowest reachability path cost to perform disinfection tasks, avoiding lingering in core passageways and obstructing traffic. The complete process of the hard real-time event preemption and interruption mechanism is as follows: Figure 3 As shown.

[0061] Step S105: After arriving near the target area, carry out targeted disinfection at the system's estimated time by synchronously monitoring aerosol distribution data, and ensure that the disinfection task near the target area is completed.

[0062] Upon reaching the target disinfection point, the disinfection function begins operation. The estimated disinfection time at this location is determined based on the cluster size and concentration. The specific calculation formula is as follows:

[0063] ;

[0064] in, The release rate of the disinfectant agent. To adjust parameters for the environment, R represents the aerosol concentration, and R is the aerosol distribution radius at the target point.

[0065] After the scheduled disinfection period, the system uses a spatiotemporal situation map to determine whether the aerosols have been completely eliminated or whether the concentration has been reduced below the threshold. If not, the required disinfection duration is reassessed until the system confirms that the disinfection task is complete.

[0066] Step S106: Select the next disinfection target point and repeat steps four and five to start the loop. During the process, reinforcement learning is used to improve the rationality of the disinfection time prediction, thereby optimizing the selection of disinfection target points.

[0067] The reinforcement learning agent control platform first outputs a predicted disinfection time based on the current policy parameters and executes the ICU environment disinfection task. After the task is completed, it collects and records the actual operating indicators. Then, the system measures the total amount of remaining aerosols after disinfection and calculates the absolute error of the disinfection time in parallel. Next, the tiered reward function calculation component summarizes and calculates the tiered reward. Finally, the parameter optimization and update component, aiming to maximize the reward, uses gradient descent to correct the parameters frame by frame and feeds the corrected parameters back to the agent control platform, thereby achieving continuous iteration and optimization of the policy parameters. The disinfection prediction time parameter update process is as follows: Figure 4 As shown.

[0068] The reward function for reinforcement learning is:

[0069] ;

[0070] in, The system's tiered rewards; The total time required to complete this round of disinfection in the affected area; This represents the total amount of remaining aerosols at each monitoring point. To minimize the absolute error between the predicted eradication time and the actual eradication time, the system aims to minimize the system loss, i.e., maximize the step reward, and uses gradient descent to correct the parameters frame by frame.

[0071] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map, characterized in that, Includes the following steps: Step S101: The robot moves within the scene and scans with multi-line radar. It uses the SLAM algorithm to build a scene map, selects machine learning methods to segment semantic targets, and constructs an initial spatiotemporal situation map. The spatiotemporal situation map is a dynamic perception network that dynamically overlays obstacle attributes, functional area divisions, personnel category labels, and real-time aerosol concentration and distribution radius on a grid map. Step S102: Fit the spatiotemporal situation map by combining visual monitoring with the robot's sensors, complete the relocalization, create scene obstacle objects, label the human type and divide the area; Step S103: Generate multiple disinfection target points based on the aerosol monitoring distribution data map. Based on the distance of each target point, the characteristics of the area and the aerosol concentration, and combined with the process scheduling algorithm, select the next disinfection target point to complete the preliminary path planning. Step S104: Combining the spatiotemporal situation map and multi-source sensor perception information, a local obstacle avoidance strategy is adopted while moving forward, and different upper limits for movement speed are set according to the area where the robot is located. Step S105: After arriving near the disinfection target, synchronize aerosol monitoring distribution data, carry out fixed-point disinfection under the system's estimated time, and determine whether the aerosol level has dropped below the threshold through the spatiotemporal situation map until the disinfection task is confirmed to be completed. Step S106: Select the next disinfection target point and repeat steps S104 and S105 in a loop. During the process, reinforcement learning is used to improve the rationality of the disinfection time prediction and optimize the target point selection. In step S104, a hard real-time event preemption interruption mechanism is set. If medical staff in a certain area of ​​the spatiotemporal situation map gather rapidly in a short period of time and the number exceeds a set threshold, the area is temporarily changed to an emergency work area. If the robot is currently working in the emergency work area, it immediately leaves and re-triggers step S103, and selects another safe target point that is far away from the emergency work area and has the lowest reachability path cost to perform the disinfection task.

2. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S101, the Fast-lio2 algorithm is used to obtain a point cloud map, and a depth camera is used to capture images. Multi-view Bayesian fusion is performed through frame-by-frame 2D semantic segmentation. The 2D semantic segmentation probability results of multiple frames of images are spatially and probabilistically fused with the 3D point cloud map to complete the semantic target division of the entire 3D point cloud map. The original obstacles are divided into movable obstacles and immovable obstacles, and a 2D grid map is created based on this as the spatiotemporal situation map. According to the generated initial spatiotemporal situation map, different areas in the intensive care unit are automatically divided, including the corridor area, the medical equipment concentration area, and the medical staff operation area.

3. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S101, extrinsic parameter calibration of the radar and depth camera is required. A fixed rigid body transformation matrix is ​​obtained using the FAST-Calib method, and the camera pixel observations are projected onto the radar point cloud coordinate system. The rigid body transformation formula is: ； in, These are the coordinates of a 3D point in the depth camera coordinate system. These are the three-dimensional point coordinates in the lidar coordinate system. Let be the rigid body transformation matrix. for The rotation matrix, for The translation vector.

4. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S102, the fixed monitoring screen information assists the robot in repositioning itself, and real-time target recognition is completed based on the initial spatiotemporal situation map to create observation objects for movable targets; human targets appearing in the monitoring are identified, and human information is synchronized to the spatiotemporal situation map. Human types are comprehensively distinguished based on the multidimensional visual features and real-time behavioral features of the identified human, which are categorized as patients, medical staff, and other personnel; the repositioning uses the control point method to calculate the homography matrix between the fixed monitoring screen plane and the physical plane of the spatiotemporal situation map to achieve alignment. The alignment formula is: ； in, To fix the pixel coordinates under the monitoring screen, ( , () represents the two-dimensional physical coordinates in the spatiotemporal situation diagram. The solution obtained by the control point method Homography matrix.

5. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S103, the selection of the disinfection target point is based on the necessity of disinfection. Factors influencing the necessity of disinfection include the concentration of the target point, the area where the target is located, the predicted duration of the disinfection task, the concentration of people, the movement cost of the preliminary route planning, and the concentration change trend. The formula for calculating the necessity of disinfection is: ； in, For target point The comprehensive score for the necessity of disinfection. This represents the real-time aerosol concentration at that point. The radius of distribution of aerosol clusters; The cost of the robot's planned path to the target point; This is the movement risk factor for the area where this point is located. This shows the trend of aerosol concentration changes near this point; For patients' semantic relevance gain factor, to These are the preset weight allocation coefficients for each indicator.

6. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 5, characterized in that, In step S103, the preliminary global path planning adopts an improved environmental risk weighting method. The algorithm is complete, in The algorithm's cost evaluation function directly incorporates the product of the regional movement risk factor and the path search inflation factor, enabling the generated global path to proactively bypass high-risk areas densely populated with medical equipment; the cost evaluation function formula is: ； in, The actual physical distance traveled from the starting point to the current node n is the cost. This represents the estimated cost from the current node n to the target point.

7. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S104, the local obstacle avoidance strategy uses the dynamic window algorithm to avoid obstacles. Based on the robot's current acceleration limit, it calculates the feasible speed combination in a short time, predicts several short trajectories, and selects the path that is farthest from the obstacle and fastest to approach the target point. Its evaluation function formula is: ； in, This represents the linear velocity of the robot. Let ω be the robot's angular velocity. For azimuth evaluation function, These are the weighting coefficients of the azimuth evaluation function. The obstacle distance evaluation function is... These are the weighting coefficients of the obstacle distance evaluation function. For the velocity evaluation function, These are the weighting coefficients of the velocity evaluation function.

8. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 7, characterized in that, In step S104, the upper limit of movement speed in different areas is set as follows: the corridor area is the normal upper limit. The operating area for medical staff is 0.

75. The medical equipment concentration area is 0.

2. .

9. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S105, after reaching the disinfection target point, the estimated residence time for disinfection is determined based on the size and concentration of aerosol clusters. The specific calculation formula is as follows: ; in The release rate of the disinfectant agent. To adjust parameters for the environment, R represents the aerosol concentration, and R is the aerosol distribution radius at the target point.

10. The control method for an ICU aerosol mobile disinfection robot based on a spatiotemporal situation map according to claim 1, characterized in that, In step S106, the reinforcement learning agent control platform outputs a predicted disinfection time based on the current policy parameters and executes the disinfection task. After the task is completed, actual operating indicators are collected, the total amount of remaining aerosols after disinfection is measured, and the absolute error of the disinfection time is calculated. The reward is calculated using a step reward function. With the goal of maximizing the reward, gradient descent is used to correct the parameters frame by frame and feed them back to the agent control platform to achieve continuous iteration and optimization of the policy parameters. The step reward function is: ， in, The system's tiered rewards; The total time required to complete this round of disinfection in the affected area; This represents the total amount of remaining aerosols at each monitoring point. This is to determine the absolute error between the predicted disinfection time and the actual disinfection time.

Images