A control system for an intelligent vaccine restocking robot that integrates multi-sensor perception.

Abstract

This invention relates to the field of medical warehousing automation and intelligent logistics technology, and discloses a vaccine intelligent replenishment robot control system integrating multi-sensor perception. The system includes a mobile replenishment robot, a vaccine distribution terminal, and a management server. The robot is equipped with layered sensor modules. An onboard computing unit utilizes a fusion of LiDAR and visual-inertial sensors for autonomous positioning, and identifies inventory through a weighted verification of visual and RFID confidence levels. Flexible suction cups are controlled to complete damage-free handling. The server integrates an LSTM model with an attention mechanism to predict replenishment needs based on historical circulation and expiration date data, and executes a task scheduling strategy based on expiration date priority. Furthermore, the system calculates a comprehensive anomaly severity index by evaluating temperature control and equipment status, triggering a tiered safety linkage mechanism. This invention achieves fully integrated vaccine perception and proactive replenishment, improving the safety and turnover efficiency of cold chain logistics.

Description

Technical Field

[0001] This invention relates to the field of medical warehousing automation and intelligent logistics technology, specifically to a vaccine intelligent replenishment robot control system that integrates multi-sensor perception. Background Technology

[0002] With the improvement of the public health system and the rapid development of smart healthcare, the accuracy, continuity, and compliance of the entire distribution process of vaccines, as special biological products, have become core requirements for ensuring vaccination efficiency. Traditional vaccine warehousing management relies heavily on manual inspection and replenishment, which not only suffers from slow response times and large data entry errors, but also struggles to meet the monitoring requirements of temperature and humidity, inventory levels, and equipment operating status. This makes it easy for human error to cause inventory disruptions or for abnormal environments to lead to reduced potency. Therefore, intelligent monitoring and replenishment solutions based on autonomous mobile robots and IoT technology are gradually becoming an industry trend.

[0003] In existing automated technologies for vaccine distribution and warehousing, multi-sensor fusion and intelligent scheduling algorithms have been initially applied. At the perception level, general-purpose warehousing robots primarily utilize equipment such as LiDAR and depth cameras for environmental modeling and obstacle avoidance; some pharmaceutical logistics systems have introduced visual recognition technology for drug traceability. At the decision-making level, warehouse management systems employ threshold-based triggering mechanisms or utilize general time-series algorithms for simple demand forecasting. At the control level, a hierarchical path planning and motion control system has been established. However, these existing technologies still have limitations when facing the specific application scenario of vaccines.

[0004] First, existing multi-sensor fusion technologies focus on robot localization and navigation or single-dimensional environmental monitoring, lacking the binding and collaborative perception of vaccine recipients' identity, location, and status. Traditional single-sensor monitoring modes (such as relying on vision or relying solely on radio frequency) are easily affected by light occlusion, tag overlap, or multipath effects in complex warehousing environments, resulting in insufficient accuracy of inventory data and a lack of real-time monitoring capabilities for microenvironmental temperature and humidity during handling.

[0005] Secondly, existing replenishment decision-making logic is based on a passive response model using inventory thresholds and fails to fully consider the expiration date characteristics of vaccines. Common prediction algorithms often ignore the decisive impact of vaccine expiration date on outbound priority, making it difficult to implement first-in-first-out (FIFO) or near-expiration-first-out (NIFO) strategies in the automatic replenishment process, easily leading to vaccine waste and losses due to expiration. Simultaneously, the lack of a physical verification step at the end of the operation results in the risk of inconsistencies between information flow and physical flow.

[0006] Finally, existing robot control systems and cold chain environment monitoring systems operate independently, lacking a deep, interconnected safety mechanism. General-purpose warehouse robots struggle to dynamically adjust their movement strategies (such as slowing down, avoiding obstacles, or emergency power cuts) based on environmental risk levels when facing cold chain environmental anomalies (e.g., drastic temperature fluctuations) or sudden risks. This fails to meet the specific requirements of high safety and rapid risk response in pharmaceutical cold chain scenarios. Therefore, there is an urgent need for an intelligent replenishment robot control system capable of achieving multi-dimensional integrated perception, proactive expiration date management, and environmental and equipment safety linkage. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a vaccine intelligent replenishment robot control system that integrates multi-sensor perception. This system solves the problems in existing vaccine warehousing and logistics, such as low accuracy due to reliance on a single sensing source for inventory counting, passive replenishment decisions that neglect expiration date management leading to high losses, and delayed emergency response due to the lack of a robot-environment safety linkage mechanism adapted to the cold chain environment.

[0008] This invention provides a multi-sensor integrated intelligent vaccine restocking robot control system, applicable to vaccine storage areas, distribution areas, and logistics channels. The system mainly includes a mobile restocking robot, a vaccine distribution terminal, a management server, and a wireless communication network. The management server, acting as the control center, establishes bidirectional communication connections with both the mobile restocking robot and the vaccine distribution terminal via the wireless communication network. The mobile restocking robot integrates an onboard computing unit, a robotic arm actuator, and a multi-dimensional sensor module. The end effector of the robotic arm actuator is equipped with a flexible suction cup for grasping vaccine boxes. The multi-dimensional sensor module is logically divided into an environmental perception group and a task perception group. The onboard computing unit uses data from the environmental perception group to perform autonomous positioning and navigation in a dynamic environment, and uses data from the task perception group to identify specific vaccine objects. Based on the identification results, it controls the robotic arm actuator to work with the flexible suction cup to complete the vaccine handling operation. The management server is equipped with a time series prediction module. This module generates restocking instructions based on historical circulation data and vaccine expiration information and sends them to the mobile restocking robot, achieving demand-driven automated restocking.

[0009] In a preferred embodiment, to achieve perception of the environment and the work object, the system's multi-dimensional sensor modules adopt a hierarchical distributed layout. The environmental perception group is responsible for the robot's global localization and obstacle avoidance, including a LiDAR mounted on the top of the robot, a depth camera mounted on the side of the body, and an inertial measurement unit mounted at the center of the chassis. The work perception group is responsible for fine-grained perception of the end effector, including an industrial RGB camera and RFID reader antenna mounted on the end of the robotic arm actuator, and a temperature and humidity sensor integrated into the flexible suction cup assembly. This layout allows the perception area of ​​the work perception group to move synchronously with the movement of the robotic arm actuator, thereby acquiring real-time visual, radio frequency, and environmental microclimate data of the work point during the grasping process.

[0010] To address the positioning stability issue in complex warehouse environments, the onboard computing unit incorporates an adaptive fusion positioning module. This module receives the scanned and matched pose from the LiDAR and the estimated pose from visual-inertial odometry in parallel. The innovation lies in the system's dynamic adjustment of the fusion weights by monitoring the matching residual value between the current LiDAR scan data and the pre-built map: when the matching residual value is less than a preset threshold, indicating high LiDAR observation quality, the system increases the weight of the scanned and matched pose; when the matching residual value is greater than the preset threshold, indicating a possible long corridor or dynamically occluded environment, the system automatically increases the weight of the estimated pose. This mechanism ensures that the robot always outputs the optimal estimated pose, improving navigation robustness.

[0011] To address the limitations of relying on a single sensor in inventory counting, the onboard computing unit executes confidence-based inventory verification logic. The system simultaneously acquires the visual inspection count from the industrial RGB camera and the radio frequency (RF) read count from the RFID reader, calculating the difference between the two. When this difference exceeds the allowable error range, the system no longer simply relies on a single sensor. Instead, it selects the higher confidence level as the final confirmed inventory quantity based on a weighted comparison of the average confidence level of the visual inspection and the RF signal strength value. This logic effectively avoids the impact of visual occlusion or RF multipath effects on inventory counting accuracy.

[0012] In terms of demand forecasting and task generation, the time series forecasting module of the management server employs a Long Short-Term Memory (LSTM) network model with an attention mechanism. This LTM network model constructs an input feature sequence using the actual consumption of vaccines, the reciprocal of the remaining expiration date, and the variance of environmental temperature fluctuations. Through the attention mechanism, the LTM network model can assign differentiated weights to historical information at different time steps based on the hidden layer states, emphasizing the impact of near-expiration vaccine data on the prediction results. This mechanism enables the prediction results to accurately capture the fluctuations in stockouts caused by approaching expiration dates, thereby outputting a more precise replenishment demand for the next time step.

[0013] Based on the above predictions, the management server executes a task scheduling strategy prioritizing expiration dates. When faced with multiple concurrent replenishment orders, the server sorts the orders by calculating a comprehensive priority score. The calculation logic for this score ensures that the priority is negatively correlated with the remaining expiration date of the vaccine, negatively correlated with the current inventory level, and positively proportional to the distance of the replenishment path. This strategy ensures that vaccine batches with the shortest remaining expiration date are prioritized for dispatch, minimizing waste caused by expired vaccines.

[0014] At the robot's motion planning level, the onboard computing unit performs path planning based on a dynamic potential field. The system overlays an obstacle repulsive potential field onto a grid map; the strength of this field increases as the distance to the obstacle decreases, thus guiding the robot away from dynamic obstacles. Simultaneously, an improved A* algorithm is used to search for the path node sequence with the minimum overall cost within the potential field, and this sequence is smoothed to generate a smooth trajectory that conforms to the robot's kinematic constraints, ensuring the stability of the vaccine transportation process.

[0015] To achieve closed-loop verification in the inbound and outbound processes, a laser counter is installed at the delivery port of the distribution area. When the mobile replenishment robot uses its flexible suction cup component to pick up vaccine boxes based on the target depth and places them at the delivery port, the laser counter records the actual number of pulses passing through. The management server compares this actual number of pulses with the planned replenishment quantity: if the data matches, the inventory information is updated; if they do not match, a replenishment task is immediately generated or an anomaly alarm is triggered to prevent missed or excessive deliveries.

[0016] In terms of system security and environmental monitoring, the management server is equipped with an anomaly severity assessment module, which constructs a multi-level linkage safety response mechanism. This anomaly severity assessment module collects real-time data from temperature and humidity sensors, robot operating status, and fire alarm signals. It calculates a comprehensive anomaly severity index by weighting and summing temperature control risk factors based on the magnitude of real-time temperature exceeding a critical threshold, equipment failure factors based on the proportion of faulty robots, and fire risk factors based on fire alarm signals. When the index exceeds the first threshold, the system determines it as a general risk and triggers environmental adjustment and avoidance commands; when the index exceeds the second threshold, the system determines it as a severe risk, immediately triggering an emergency braking command and cutting off power to non-critical equipment to ensure vaccine assets and on-site safety.

[0017] In addition, the management server includes a digital twin monitoring platform that uses real-time sensor data to build a virtual warehouse model and discretizes the warehouse area into multiple monitoring grids. The platform calculates a comprehensive risk index based on the temperature deviation and robot density within each monitoring grid and renders the risk level of each grid in the form of a heat map, providing managers with an intuitive global monitoring view to assist in decision-making.

[0018] Through the above-mentioned technical solution, this invention realizes the full-process automation and intelligence of vaccine replenishment operations. While improving logistics efficiency, it also improves the accuracy and security of inventory management through multi-sensor fusion verification and intelligent scheduling based on expiration date.

[0019] This invention provides a control system for an intelligent vaccine restocking robot that integrates multi-sensor perception. It has the following beneficial effects:

[0020] 1. This invention achieves comprehensive fusion perception of vaccine identity, location, and status by constructing a hierarchical distributed multidimensional sensor module. It utilizes an onboard computing unit to execute confidence-based inventory verification logic, and weightedly fuses visual inspection data from industrial RGB cameras with radio frequency signals from RFID readers. This solves the problem of single sensors being susceptible to occlusion or multipath interference in complex warehousing environments. Combined with temperature and humidity sensors integrated into flexible suction cups, it ensures batch traceability accuracy and environmental compliance of vaccines throughout the entire handling process, improving the level of precision in vaccine warehousing management.

[0021] 2. This invention upgrades vaccine replenishment management from passive response to proactive prediction. By configuring a long short-term memory network model with an attention mechanism, the system can deeply mine historical circulation data and vaccine expiration information to predict replenishment needs at the next moment. Combined with a task scheduling strategy based on expiration date priority, it automatically assigns higher priority to vaccine batches with shorter remaining expiration dates, which not only reduces vaccine losses due to expiration but also ensures consistency between the physical flow and information flow of replenishment operations through the closed-loop verification mechanism of the laser counter in the distribution area, thus optimizing inventory turnover efficiency.

[0022] 3. This invention establishes a layered collaborative control and safety protection mechanism for robots adapted to cold chain scenarios. At the execution level, a flexible suction cup component is used in conjunction with an adaptive fusion positioning algorithm to achieve high-precision navigation in dynamic environments while ensuring undamaged grasping of vaccine packaging. At the monitoring level, a multi-level linkage response system is constructed through an anomaly severity assessment module to calculate a comprehensive risk index covering temperature control risks, equipment failures, and fire signals in real time. When an anomaly occurs, environmental adjustment, avoidance, or emergency braking power-off commands are triggered according to the risk level, filling the gap in the adaptation of general warehousing robots to the high safety requirements of the pharmaceutical cold chain. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall architecture of the intelligent vaccine restocking robot control system that integrates multi-sensor perception provided in an embodiment of the present invention;

[0024] Figure 2 A flowchart of a confidence-based vaccine inventory fusion perception system provided in this embodiment of the invention;

[0025] Figure 3 The diagram shows the structure of a time-series demand prediction model based on an attention mechanism, as provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See attached document Figure 1 This invention provides a multi-sensor integrated intelligent vaccine replenishment robot control system, whose application environment includes a vaccine storage area, a distribution area, and a logistics channel connecting the two. The system mainly includes: a mobile replenishment robot, a vaccine distribution terminal, a management server, and a wireless communication network.

[0028] Mobile replenishment robots are deployed within the vaccine storage and distribution areas to autonomously handle, replenish, and inventory checks of vaccines between different zones. Vaccine distribution terminals are located at vaccination stations or distribution points at various levels, equipped with temporary storage units for storing vaccines awaiting administration. These terminals are configured to monitor their own inventory status in real time and send replenishment request signals to the management server. The management server establishes bidirectional data communication connections with both the mobile replenishment robots and the vaccine distribution terminals via a wireless communication network. This network collects status data, processes business logic, and issues scheduling and control commands. The wireless communication network utilizes either 5G or industrial WiFi protocols to cover the application environment. Near-field communication interfaces are also configured between the mobile replenishment robots and the vaccine distribution terminals for localized data exchange during replenishment operations.

[0029] The mobile replenishment robot includes: a mobile chassis, a robotic arm actuator, a multi-dimensional sensor module, and an onboard computing unit.

[0030] The mobile chassis serves as the robot's motion carrier and integrates a servo motor drive system and a power management module. The servo motor drive system is connected to the hub motors to control the robot's start, stop, steering, and speed adjustment, enabling the robot to move omnidirectionally or differentially in the horizontal plane.

[0031] The robotic arm actuator is mounted on the upper structure of the mobile chassis and is used to perform the grasping and placing of vaccine packaging boxes. A flexible suction cup assembly is located at the end of the robotic arm actuator, and this assembly is connected to a vacuum pump inside the machine body via an air duct. The vacuum pump is controlled by an onboard computing unit to generate negative pressure suction or positive pressure airflow at the flexible suction cup.

[0032] The multi-dimensional sensor module is used to collect information about the external environment and the status of the work object, and its signal output terminals are connected to the onboard computing unit. Specifically, the multi-dimensional sensor module includes:

[0033] LiDAR is installed on the top of the mobile replenishment robot or at a high position with a wide field of view to collect horizontal environmental contour distance data.

[0034] A depth camera is installed on the side of the mobile replenishment robot in the direction of its forward movement to collect three-dimensional depth point cloud data of the environment and obstacle features.

[0035] Industrial RGB cameras are installed at the end of the robotic arm's actuator or on the side of the machine to capture images of the appearance of vaccine packaging, as well as QR code or barcode information.

[0036] An inertial measurement unit (IMU) is rigidly fixed to the center of the mobile chassis and is used to measure the robot's three-axis acceleration and three-axis angular velocity data.

[0037] An RFID reader, equipped with a directional radio frequency antenna, is installed with the antenna facing the working side and is used to read the data of the radio frequency tag attached to the vaccine packaging;

[0038] Temperature and humidity sensors are distributed on the surface of the mobile replenishment robot and in the gripping execution area to monitor the real-time temperature and humidity values ​​of the local microenvironment.

[0039] The onboard computing unit is electrically connected to the servo motor drive system, robotic arm actuator, and multi-dimensional sensor module of the mobile chassis via an internal communication bus. The onboard computing unit is configured to receive data collected by the multi-dimensional sensor module and perform environmental perception processing, positioning calculation, inventory status monitoring, and motion planning and control.

[0040] Specifically, the onboard computing unit is configured to perform multi-source information fusion using LiDAR data, depth camera data, and inertial measurement unit data to determine the real-time pose of the mobile replenishment robot on a pre-built map. The onboard computing unit is also configured to combine visual data from an industrial RGB camera with radio frequency data from an RFID reader to identify and verify the quantity and attributes of the target vaccines. Furthermore, the onboard computing unit determines the environmental condition based on data from temperature and humidity sensors and generates an abnormal status signal when the values ​​exceed preset ranges.

[0041] The multi-dimensional sensor fusion perception system of this invention adopts a distributed hardware layout of body plus robotic arm and a three-level logical architecture of acquisition, preprocessing and fusion to achieve high-precision perception of dynamic environment and static inventory.

[0042] In terms of hardware sensor layout, the system defines a base coordinate system and an end effector coordinate system based on the physical structure of the mobile replenishment robot, and spatially deploys sensors in two groups: an environmental perception group and a task perception group. The environmental perception group, primarily used for the robot's autonomous navigation and obstacle avoidance, includes a lidar, depth cameras, and an inertial measurement unit (IMU). The lidar is rigidly mounted at the geometric center of the top of the mobile replenishment robot, with its mounting plane parallel to the ground and a 360-degree scanning field of view, used to establish a polar coordinate measurement system with the robot's center as the origin. The depth cameras include a forward-looking depth camera and a side-looking depth camera. The forward-looking depth camera is mounted on the front face of the mobile chassis, with its optical axis having a downward pitch angle relative to the horizontal plane, set to 15 to 25 degrees to cover the near-field blind zone below the lidar. The inertial measurement unit (IMU) is mounted at the center of mass of the mobile chassis, and its three-axis sensing directions are kept parallel to the three-axis directions of the base coordinate system through mechanical calibration, used to output acceleration and angular velocity data without lever arm effects.

[0043] The task sensing unit is primarily used for vaccine object identification and grasping, and includes an industrial RGB camera, an RFID reader antenna, and a temperature and humidity sensor. Both the industrial RGB camera and the RFID reader antenna are mounted on the end flange of the robotic arm actuator, allowing their sensing areas to move synchronously with the end effector's movement, thus achieving dynamic tracking and sensing of the target. The industrial RGB camera's lens optical axis is perpendicular to the end flange plane, used to acquire image data directly in front of the end effector. The RFID reader antenna is a directional polarized antenna, with its main radiation beam direction parallel to the industrial RGB camera's lens optical axis, and a metal shielding layer on the back of the antenna to block radio frequency signal interference from directions other than the task area. The temperature and humidity sensor is integrated inside the end effector's gripper, used to acquire microenvironmental data of the vaccine surface at the moment of grasping contact.

[0044] In terms of the layered sensing architecture, the system is divided into a data acquisition layer, a data preprocessing layer, and a feature fusion layer from bottom to top. Data is transferred between the layers through shared memory or message queues. The data acquisition layer is responsible for the synchronous acquisition of multi-source heterogeneous data. It is equipped with hardware triggers or uses the Network Precise Time Protocol (PTP) to map the point cloud data of LiDAR, the image frame data of the camera, the inertial data of the IMU, and the tag data of RFID to the same time reference axis, ensuring that the data of different sensors are aligned at the same time.

[0045] The data preprocessing layer is responsible for cleaning and spatial transformation of the raw data. For LiDAR data, the preprocessing layer performs motion distortion removal algorithms, using IMU data to compensate for point cloud offsets caused by robot motion during laser scanning. For visual image data, the preprocessing layer performs distortion correction using a pre-calibrated intrinsic parameter matrix and transforms the image coordinate system to the base coordinate system using an extrinsic parameter matrix calibrated by hand and eye. For RFID data, the preprocessing layer performs moving average filtering of RSSI signal strength to filter out signal fluctuations caused by multipath effects.

[0046] The feature fusion layer is responsible for generating high-dimensional perception results. In environmental perception, the feature fusion layer spatially registers the preprocessed laser point cloud with the visual depth point cloud, outputting an environmental occupancy raster map containing geometric boundaries and depth information. In job perception, the feature fusion layer spatiotemporally correlates the vaccine location bounding box output by the visual target detection algorithm with the electronic tag ID read by RFID. When the spatial deviation between the visually detected target location and the peak location of the RFID signal strength is less than a preset threshold, the two are bound as the same object, outputting fused inventory information including location, ID, and status.

[0047] The airborne computing unit, based on the feature fusion layer in the hierarchical architecture, performs state estimation and SLAM (Simultaneous Localization and Mapping) processing for environmental data. This processing aims to address the positioning drift and feature loss issues of a single sensor in the dynamic environment of a vaccine warehouse, and achieves highly robust pose estimation by jointly optimizing LiDAR, vision, and inertial data.

[0048] Specifically, the system first defines the mobile replenishment robot at time... Full state vector This full-state vector encompasses the robot's kinematic state and the intrinsic parameter biases of the sensors, and its mathematical definition is as follows:

[0049] ;

[0050] in, This represents the robot's three-dimensional position coordinates in the world coordinate system, corresponding to the robot's spatial translation. A quaternion representing the robot's pose, used to describe the robot's three-dimensional rotational state; This represents the linear velocity vector of the robot in the world coordinate system; This represents the zero-bias error term of the inertial measurement unit (IMU), which further includes the accelerometer zero bias. and gyroscope zero bias .

[0051] To obtain the optimal state estimate, the system constructs a nonlinear least squares optimization model based on a sliding window. This model minimizes the joint optimization cost function. To solve for the state vector, we need to jointly optimize the cost function. It integrates lidar observation constraints, visual feature point observation constraints, and IMU pre-integration constraints, and its specific expression is:

[0052]

[0053] in, This represents the residual term from lidar observations; This is the set of keyframes for the LiDAR within the sliding window; For the first Point cloud observation data acquired by frame lidar; For the first The robot state at the corresponding frame time; A residual function is matched for laser scanning to calculate the geometric deviation between the current scanned point cloud and the local map; It is the inverse matrix of the covariance matrix of lidar observations, i.e., the information matrix, used to characterize the credibility weights of the lidar data; This represents the summation operator; The square of the Mahalanobis distance is defined as follows: It is used to measure the magnitude of the error.

[0054] Represents the visual observation residual term; The set of visual feature points within the sliding window; For the first The observed coordinates of a visual feature point on the image plane; For the first The robot's state at the moment when each feature point is observed; This is the visual reprojection residual function, used to calculate the pixel error between the observed position of a feature point and the projected position predicted based on the state. Information matrix for visual observation; for Robust kernel functions are used to suppress the interference of outlier features generated by dynamic obstacles (such as moving people or forklifts) on the optimization results.

[0055] This represents the IMU pre-integration residual term; From time At the time Pre-integrated values ​​of IMU measurement data during the period; This is the IMU pre-integration residual function, used to constrain the states of two adjacent keyframes. and The relative motion relationship between them; This is the information matrix for inertial measurements.

[0056] By utilizing the Gauss-Newton method or the Levenberg-Marquardt algorithm to jointly optimize the cost function Through iterative solving, the system can calculate the robot's optimal pose estimate in real time. Furthermore, based on the optimized pose and laser point cloud data, the system uses an improved Cartographer algorithm to construct a grid-based occupancy map of the vaccine warehouse, where each grid cell stores the probability that the location is occupied by an obstacle, thereby achieving positioning accuracy and environmental mapping capabilities at the ±2cm level.

[0057] See attached document Figure 2 The onboard computing unit performs fusion perception processing for the target object based on the feature fusion layer in the hierarchical architecture. This processing aims to leverage the complementary characteristics of visual recognition and radio frequency identification (RFID) to solve the counting error problem of a single sensor under conditions of dense vaccine stacking or changing lighting, and to combine environmental data to ensure the safety of vaccine storage.

[0058] The onboard computing unit first controls an industrial RGB camera to acquire image data of the target shelf area and then runs a deep learning-based object detection algorithm. The object detection algorithm identifies and locates the vaccine vials in the image and outputs the first... Number of visual inspections for vaccine-like products and the corresponding average confidence level of visual detection Visual inspection average confidence level For all those identified as the first The arithmetic mean of the confidence scores of the target bounding boxes for the vaccine, with values ​​ranging from [0,1].

[0059] Simultaneously, the onboard computing unit controls the RFID reader to initiate radio frequency scanning, reading electronic tags in batches within the target area via a directional antenna. The RFID reader outputs the first... Number of radio frequency reads for vaccine and the corresponding average Received Signal Strength Indication (RSSI) value Average received signal strength. This value is used to characterize the quality of the tag signal read; a higher value indicates a more reliable read.

[0060] The onboard computing unit executes a confidence-based weighted fusion strategy to calculate the final confirmed inventory level. This strategy introduces a preset allowable error threshold. This is used to determine if there is a significant conflict between the data from the two sensors. Finally, the inventory level is confirmed. The calculation formula is as follows:

[0061]

[0062] in, Indicates the number obtained by the RFID reader / writer Number of radio frequency reads for vaccine-like vaccines; This represents the first number recognized by the visual algorithm. The number of visual inspections for vaccine-like products; It represents the absolute difference between the number of visual detections and the number of radio frequency reads, used to quantify the degree of conflict between the two sensing methods; This indicates the preset allowable error threshold, which serves as the baseline for determining whether two sets of data are consistent.

[0063] Indicates when The value is less than When the values ​​are consistent, the system determines that the visual data and the radio frequency data are the same, and directly adopts the number of radio frequency reads. .when The value is greater than or equal to When the value is reached, the system determines that there is a perceptual conflict and executes [the necessary actions]. The logic of branching; This indicates a parameter retrieval operation, i.e., from a set. Choose one from the options such that the subsequent decision function The element with the largest value is taken as the final result; This represents the confidence decision function based on signal quality.

[0064] Decision function The comparison logic used to arbitrate which sensor's data is more reliable in the event of a sensing conflict is as follows:

[0065]

[0066] in, This represents the average confidence level of the visual detection, with a value range of [0,1]. This indicates the average received signal strength, measured in decibels and milliwatts (dBm). This represents the preset visual weight coefficients, used to define the importance of visual data in arbitration; This represents the preset radio frequency weighting coefficient, used to define the importance of radio frequency data in arbitration, and satisfies... ; This represents the normalization mapping function, used to... The numerical values ​​are mapped to the [0,1] interval to eliminate dimensional differences.

[0067] This represents the weighted visual reliability score; This represents the weighted RF reliability score; The comparison operator indicates that the numeric item on its left and the numeric item on its right are two objects being compared. This indicates a numerical comparison between the two combined items: if the weighted visual reliability score is greater than the weighted radio frequency reliability score, the decision function favors the visual result, and the system selects... Conversely, the system selects... .

[0068] In addition, the onboard computing unit collects real-time state data of the vaccine storage microenvironment using temperature and humidity sensors. An environmental state vector is defined. for:

[0069] ;

[0070] in, This represents the temperature measurement value at the current moment. This indicates the humidity measurement value at the current moment; This represents the vector transpose operation. The airborne computing unit will... With respect to the preset safe temperature range for vaccines Real-time comparison is performed. Once detected... The system immediately generates an abnormal temperature alarm signal and marks the batch of vaccines as a potential risk object, restricting its dispatch from the warehouse.

[0071] See attached document Figure 3 The management server is equipped with a deep learning-based time series prediction module to predict vaccine consumption at the next moment based on historical data, thereby triggering replenishment tasks in advance. This prediction model uses a Long Short-Term Memory (LSTM) network as a feature extractor and integrates a time attention mechanism to address long-sequence dependencies and strengthen the weighting of vaccine expiration date factors.

[0072] The management server first preprocesses the historical data on vaccine distribution to construct an input feature sequence. (Time point defined) Input feature vector as follows:

[0073] ;

[0074] in, Indicates time The input feature vector; Indicates at time The actual number of vaccines consumed; Indicates at time The remaining shelf life of this batch of vaccines; This represents the reciprocal of the remaining shelf life, used to map the urgency of the expiration date to a larger numerical feature input, making the model more sensitive to vaccines nearing their expiration date; Indicates time The variance of ambient temperature fluctuations is used to introduce the characteristics of the impact of the environment on vaccine loss. This represents the vector transpose operation.

[0075] Input feature vector The information is fed into an LSTM network layer for temporal feature extraction. The LSTM unit controls the information flow through forget gates, input gates, and output gates. Its state update formula is defined as follows:

[0076]

[0077] in, Representing time respectively The forget gate activation vector, input gate activation vector, and output gate activation vector; Indicates time The hidden layer state vector; Indicates time The cell state vector; Indicates time The candidate cell state vector; Indicates time Updated cell state vector; Indicates time Updated hidden layer state vector; These correspond to the weight matrices of each gate control unit; These correspond to the bias vectors of each gate unit; This represents the Sigmoid activation function; Represents the hyperbolic tangent activation function; This represents the Hadamard product, which is the element-wise multiplication operation of a matrix.

[0078] To enhance the model's ability to perceive critical historical moments (such as surges in demand due to sudden outbreaks of epidemics or near-expiration points), a time attention mechanism layer is introduced after the LSTM layer. This layer is based on the hidden state vector. Calculate normalized attention weights and generate context vectors. The calculation process is as follows:

[0079]

[0080] in, Indicates time The energy score of the hidden state; The weight matrix representing the attention mechanism; The bias vector representing the attention mechanism; This represents the transpose of the parameter vector used to map high-dimensional features to scalar scores; Indicates time The normalized attention weights, whose magnitudes reflect the time intervals... The degree of importance of historical information to current prediction results; This indicates the total number of time points contained in the input sequence; This represents the final generated context feature vector, which is a weighted sum of the hidden states at all historical time points; This represents an exponential function with the natural constant e as its base. The summation symbol is used.

[0081] The system will use context feature vectors The input is fed into a fully connected output layer to calculate the predicted demand for the next time step. :

[0082] ;

[0083] in, This represents the predicted demand at the next time step, as output by the model. This represents the weight matrix of the output layer; This represents the bias scalar of the output layer; This represents the activation function of the linear rectifier unit, used to ensure that the predicted demand output is non-negative. The management server will calculate the... Compared with the current actual inventory level, when When the inventory level exceeds the difference between the current available inventory and the safety stock threshold, a replenishment scheduling instruction will be automatically generated.

[0084] The management server is equipped with a dynamic demand prediction engine based on an improved LSTM (Long Short-Term Memory) network. To capture the impact of vaccine expiration dates on outbound demand, the model introduces an attention mechanism, assigning different weights to the input time-series data, with a focus on enhancing the feature representation of near-expiration data.

[0085] The management server first calculates the replenishment demand for each type of vaccine in the warehouse. Define the... The demand for replenishing similar vaccines Its calculation formula encompasses both the predicted consumption based on the attention mechanism and the safety stock constraint:

[0086] ;

[0087] in, Indicates the first The basic safety stock threshold for vaccines; This represents the output function of the prediction model that incorporates an attention mechanism; This represents a historical data distribution sequence; This represents the attention weight vector, used to automatically focus on the changing trends of recent high-frequency outbound periods or batches with specific expiration dates; This indicates the current actual inventory level in the warehouse. When... When a replenishment task is triggered, the system executes a scheduling order based on the expiration date priority principle. The management server uses a multi-dimensional urgency function to calculate the comprehensive priority score for each task. This function explicitly assigns the highest scheduling weight to vaccines nearing their expiration date, and its mathematical expression is as follows:

[0088] ;

[0089] in, This is an item related to the urgency of the expiration date; Indicates the first The expiration date of the earliest batch of this type of vaccine; Indicates the current system time; To prevent the use of tiny constants with a denominator of zero, this term indicates that the shorter the remaining shelf life, the larger the reciprocal value, and the higher the priority, thus implementing a management strategy of prioritizing replenishment and distribution as the expiration date approaches. This is an item related to inventory scarcity, indicating that the lower the current inventory, the more urgent the need to replenish stock. Path cost penalty term, used to optimize logistics efficiency; These are preset coefficients for effectiveness weight, inventory weight, and distance weight, respectively, and are set as follows: Ensuring the safety of the expiration date is the primary consideration.

[0090] The navigation control unit of the mobile replenishment robot performs global path planning from the current dwell point to the target shelf location based on a grid map constructed by the environmental perception component. This process employs an improved method that incorporates a dynamic safety potential field. The algorithm aims to generate an optimal navigation path that satisfies both the shortest distance requirement and the maximum safety margin for obstacle avoidance.

[0091] The navigation control unit first abstracts the grid map into a set of search nodes, and then, for each node to be traversed... Calculate the comprehensive cost This comprehensive cost value is used to assess the value passed through the nodes. Expected costs to reach the goal. Total cost. The calculation formula is as follows:

[0092] ;

[0093] in, Represents a node Based on the comprehensive cost value, the system prioritizes adding the node with the smallest value to the path sequence; This indicates moving from the starting point to a node along the generated path. The actual cumulative distance cost; Indicates from node Heuristic cost estimates for reaching the target destination are typically calculated using Manhattan distance or Euclidean distance. This represents a preset risk sensitivity factor used to adjust the aggressiveness of obstacle avoidance behavior; Represents a node The repulsive potential field strength of the obstacle at the point, this value is related to the node The Euclidean distance to the nearest obstacle is inversely proportional; This represents the maximum saturation value of the potential field strength, used for normalization. This term causes the path planning algorithm to artificially increase the estimated cost when approaching obstacle areas, thereby driving the planned path to automatically maintain a safe distance from shelves or walls.

[0094] Generate the original path point set consisting of a series of discrete raster center points. Subsequently, the navigation control unit performs trajectory smoothing processing based on nonlinear optimization. This processing aims to eliminate polygonal corners in the original path and generate a smooth curve that conforms to the robot's kinematic constraints. The system constructs a trajectory optimization objective function. as follows:

[0095] ;

[0096] in, This represents the total cost function for trajectory smoothing optimization; Indicates the total number of path points; Indicates the first in the original path The coordinate vectors of the smooth trajectory points are used as reference constraints for optimization. Indicates the generated number after optimization. The coordinate vectors of the points on the smooth trajectory are the solution variables for this optimization problem. and They represent the first The preceding and succeeding points of a smooth trajectory point; This indicates the penalty for positional deviation. This represents the positional fidelity weighting coefficient; This represents the square of the offset distance of the smoothed trajectory point relative to the original path point. This constraint ensures that the optimized trajectory does not deviate excessively from the originally planned safe passage. Indicates a smoothness penalty term; This represents the curve smoothing weighting coefficient. Mathematically, this corresponds to the second-order difference of the trajectory; physically, it approximates the robot's instantaneous acceleration or path curvature at that point. This constraint aims to minimize the rate of change of the path's curvature, preventing the robot from performing sharp turns.

[0097] The navigation control unit uses gradient descent to optimize the objective function of the trajectory. The solution is iteratively solved until the objective function converges. The final result is a series of... The coordinate points form a smooth and continuous motion trajectory. The underlying motion controller of the mobile replenishment robot is configured to track this trajectory and complete the replenishment and transportation task with a constant linear velocity and a smoothly changing angular velocity.

[0098] The control system of the mobile replenishment robot adopts a hierarchical design, strictly divided from top to bottom into a decision-making and planning layer, a collaborative control layer, and an execution and drive layer. This architecture aims to achieve decoupled control from macro-level task scheduling to micro-level motor drive, ensuring the robot's motion stability and operational accuracy in the highly dynamic vaccine storage environment.

[0099] The decision-making and planning layer is deployed on a cloud management server. It is responsible for processing replenishment requests from the upper-level business system, generating a series of discrete target waypoint sequences based on global map information, and then sending these sequences to the onboard computing unit. The collaborative control layer runs in the robot's onboard industrial control computer and acts as the core of the system. Its core function is to execute trajectory tracking control algorithms and convert the received waypoint sequences into real-time speed control commands to eliminate position tracking errors.

[0100] The collaborative control layer first calculates the tracking error between the robot's current position and the desired trajectory point. To avoid confusion with the aforementioned environment state vector, time is defined here. pose tracking error vector as follows:

[0101] ;

[0102] in, Representing time respectively The longitudinal position error components, the lateral position error components, and the heading angle error components; These represent the robot's position at time [time] as fed back by the positioning sensors. The actual world x-coordinate, actual world y-coordinate, and actual heading angle; They represent the desired trajectory points at time [time]. The reference world x-coordinate, reference world y-coordinate, and reference heading angle. (The middle part of the formula...) The matrix represents the rotation transformation matrix, which is used to project and map the position difference in the world coordinate system to the robot's own motion coordinate system.

[0103] Based on the calculated pose tracking error vector The collaborative control layer uses a nonlinear control law designed based on Lyapunov stability theory to generate the velocity command vector. :

[0104] ;

[0105] in, Indicates time Target linear velocity command; Indicates time Target angular velocity command; and These represent the preset feedforward linear velocity and feedforward angular velocity of the reference trajectory at that moment, respectively. These represent the feedback gain coefficients for the longitudinal, lateral, and directional directions, respectively, used to adjust the system's response speed and convergence. The design of this control law ensures that, with time... The pose tracking error vector tends towards infinity. It asymptotically converges to the zero vector.

[0106] The execution driver layer consists of an embedded motion controller and a servo driver, and is responsible for transmitting speed command vectors. This is converted into voltage signals for each drive wheel motor. The embedded motion controller first converts the target linear velocity command into voltage signals based on the robot's differential inverse kinematics model. and target angular velocity command Solving for the desired speed of the left wheel and the desired speed of the right wheel Subsequently, regarding the first One motor The controller uses an incremental PID algorithm to adjust the output voltage. This is to achieve closed-loop tracking of the desired rotational speed.

[0107] ;

[0108] in, Indicates time Applying at the The duty cycle value of the control voltage on each motor; This indicates the control voltage value at the previous moment; Indicates time The speed deviation, i.e., the first The difference between the expected speed of each motor and the actual speed fed back by the encoder; and Representing time respectively and time The speed deviation; They represent the first The proportional, integral, and derivative coefficients of each motor control loop are used. This three-tier architecture ensures millimeter-level control response for the robot when performing high-precision docking tasks through data flow and closed-loop feedback between layers.

[0109] The mid-level scheduling module is primarily responsible for the high-precision positioning and path execution of the mobile replenishment robot in dynamic environments. To address potential issues in warehouses such as shelf obstruction and dynamic personnel movement leading to LiDAR signal loss or feature degradation, this module integrates Visual Inertial Navigation (VINS) technology, using a multi-source fusion algorithm to maintain positioning stability across all scenarios.

[0110] The mid-level scheduling module constructs a fusion localizer based on Extended Kalman Filter (EKF). This fusion localizer receives scan-matched poses from the LiDAR in real time. and estimated pose from visual inertial odometry The system dynamically adjusts the fusion weights based on the feature matching confidence of the laser point cloud, enabling the computational robot to perform calculations at any given time. The best estimated pose after fusion :

[0111] ;

[0112] in, Indicates the robot at a certain moment The optimal estimated pose after fusion, i.e., the global pose vector. ; This represents the pose calculated using the improved Cartographer algorithm based on matching laser point clouds with a grid map. This represents the global pose calculated by superimposing the relative pose increments using the image feature stream acquired by the industrial camera and the acceleration / angular velocity data of the IMU through the VINS algorithm. This represents the real-time confidence coefficient for laser positioning, with a value range of [0,1]. Its calculation depends on the matching residual between the current laser scan point and map obstacles. :

[0113] ;

[0114] in, This represents the matching residual between the current laser scan point and the obstacles on the map; This represents the preset residual threshold. This represents the slope factor of the Sigmoid function.

[0115] The physical meaning of this fusion logic is: when the robot is in an open area and the laser matching is good (matching residuals) Hour, Approaching 1, the system primarily relies on high-precision laser positioning; however, when the robot enters a narrow passage or is obstructed by objects, causing the laser features to become blurred (matching residuals)... When (large), The system automatically and smoothly switches to a high-weight VINS positioning mode to reduce the localization risk, using visual feature points and inertial data to fill in positioning blind spots and ensure that the robot's positioning accuracy remains superior in complex environments. cm.

[0116] The drive layer controls the telescopic rods and flexible suction cups mounted on the robot chassis to perform non-contact vaccine box gripping operations, avoiding the physical crushing damage that traditional mechanical grippers might cause to fragile medicine boxes. Simultaneously, a laser counting verification mechanism is introduced to ensure absolute accuracy in replenishment quantities.

[0117] The execution drive layer first controls the servo motor of the telescopic rod, driving the end of the suction cup along... The axial direction is close to the target vaccine box. Define the target extension length of the telescopic rod. The target's extension length is based on the visually recognized target depth. offset from the robot's current base The calculation shows that:

[0118] ;

[0119] in, This indicates a reserved safety buffer distance to prevent the suction cup from making a hard impact. When the telescopic rod reaches... Once positioned, the controller activates the vacuum generator and monitors the air pressure inside the suction cup in real time. The system's absorption judgment logic is set as follows:

[0120] ;

[0121] in, This indicates the preset vacuum threshold (negative pressure value). When the detected air pressure is below this threshold and remains stable, the system determines that the vaccine cartridge has been reliably adsorbed, and then controls the telescopic rod to retract and place the vaccine cartridge on the transport tray; This indicates successful adsorption; This indicates that adsorption has failed; This represents the logical AND operation; express At the time of its establishment, The value is ; express When not valid, The value is .

[0122] After the replenishment task is completed, to form a closed-loop verification, the execution driver layer activates the laser counter located at the delivery port. When the vaccine cartridge is pushed into the dispensing equipment chute, it interrupts the laser beam to generate a pulse signal. The system records the actual number of pulses that pass through. And the planned replenishment quantity in the task instruction. Perform the comparison and define the verification function. :

[0123] ;

[0124] in, This indicates the actual number of falling vaccines detected by the laser sensor; This indicates the amount of supplementary funds required as issued by the decision-making level. Indicates a qualified verification status; Indicates a missing quantity status; This indicates a severely abnormal state. When... When the system determines that the replenishment quantity is correct, it automatically updates the inventory database and closes the task; when When the system determines that there is a shortage, it triggers a replenishment command; when If the system determines that there has been an over-supply or a false alarm from the sensor, it immediately triggers an audible and visual alarm and suspends operations until manual verification is completed. This mechanism eliminates discrepancies between inventory data and the physical inventory.

[0125] The data interaction and closed-loop monitoring platform, serving as the upper-level command center of the entire system, is deployed on the central server. This platform integrates a digital twin rendering engine to map in real time the motion status of all robots within the physical warehouse, the distribution of vaccine inventory, and changes in the environmental temperature and humidity field, thereby providing managers with a holistic decision support view.

[0126] The monitoring platform first establishes a high-precision 3D model of the physical warehouse and maintains real-time full-duplex communication with the underlying mobile robots and environmental sensor network via the WebSocket protocol. The platform performs spatiotemporal alignment and fusion processing on the received discrete sensor data (such as RFID read records, temperature probe values, and robot encoder coordinates) to reconstruct the warehouse's real-time operating status in virtual space. The system is configured to update dynamic primitives in the 3D view at a refresh rate of no less than 20Hz, ensuring that the synchronization delay between the virtual and physical scenes is less than 50 milliseconds.

[0127] To visually represent potential risk areas within the warehouse, the monitoring platform incorporates a dynamic heatmap generation algorithm. This algorithm divides the warehouse floor plan into numerous tiny monitoring grids and generates heatmaps for each grid. At any moment Calculate its comprehensive risk index This risk index determines the rendering color depth of the grid area on the monitoring interface, and its calculation formula is defined as follows:

[0128] ;

[0129] in, Indicates time Time The comprehensive risk index of each monitoring grid is normalized and mapped to a color wheel to generate a visual color value (e.g., a gradient from a green safe zone to a red high-risk zone). Represents the weighting coefficients of environmental anomaly factors; The weighting coefficients representing traffic congestion factors; This indicates the total number of fixed temperature sensors deployed within the warehouse; Indicates the first A temperature sensor at time The actual temperature value collected; This indicates the optimal set temperature for vaccine storage in this area; This indicates the system's allowable temperature fluctuation tolerance value; Indicates the first The absolute deviation of the temperature sensor reading from the optimal temperature; The spatial effect decay function is typically represented by a Gaussian kernel function or an inverse distance function. Indicates the first The monitoring grid and the first Euclidean distance between the physical locations of the temperature sensors; Indicates the summation symbol; Indicates at time At that time, located in the The real-time number of mobile robots within each monitoring grid area; This indicates the maximum passage capacity or safe carrying capacity of a single grid area design.

[0130] The monitoring platform is based on the calculated Numerical values ​​are rendered in the visualization interface. When When the preset warning threshold is exceeded, the platform will not only render the corresponding grid as dark red, but also automatically display detailed status parameters of the area (such as specific temperature value and the number of the stranded robot) in a pop-up window, and trigger an audible and visual alarm to alert management personnel to intervene.

[0131] The data interaction and closed-loop monitoring platform has a built-in real-time anomaly monitoring engine, which is configured to continuously scan heterogeneous data streams from environmental sensor networks, robot status feedback interfaces, and security systems. The core of this mechanism lies in building a quantitative anomaly severity assessment model, which triggers tiered hardware linkage strategies to maximize the protection of vaccine assets and personnel safety in emergency scenarios such as fires, equipment failures, or temperature control malfunctions.

[0132] The anomaly detection engine calculates the anomaly severity index of the current system in real time using a sliding time window. The anomaly severity index is a comprehensive score used to quantify the degree to which the overall operational status of the warehouse deviates from the normal baseline at any given moment. Its calculation formula is as follows:

[0133] ;

[0134] in, This represents the collection of all temperature sensors deployed within the warehouse. Represents a set The first in A temperature sensor at time The collected real-time temperature values; This indicates the highest permissible critical temperature threshold for vaccine storage; This indicates linear rectification operation, meaning that the difference is taken when the real-time temperature exceeds the critical threshold, and zero is taken otherwise. Used to extract the most severe local overheating point within the current warehouse; This represents the temperature sensitivity coefficient, used to adjust the growth rate of the exponential function; This means that the exponential function is used to non-linearly amplify the overheating amplitude, which implies that once the temperature exceeds the critical value, the severity index will increase explosively, reflecting the high sensitivity of the vaccine to temperature. The weighting coefficient representing the risk of temperature control; Indicates time The number of mobile robots that are in fault alarm state or communication interruption state; This indicates the total number of mobile robots registered in the warehouse; Weighting coefficients representing the risk of equipment failure; This indicates a fire alarm variable, representing the time when a smoke detector or thermal imaging camera is activated. When an open flame or smoke signal is detected, this variable takes the value of 1; otherwise, it takes the value of 0. The weighting coefficient representing fire risk is usually set to a very high value to ensure that the severity index immediately reaches the highest level once a fire signal is triggered.

[0135] The data interaction and closed-loop monitoring platform will calculate the results. It compares the data in real time with preset tiered handling thresholds and executes safety linkage operations based on the comparison results. When When the threshold exceeds the Level 1 warning threshold but is below the Level 2 emergency threshold, the platform automatically triggers an environmental linkage command, remotely increases the cooling power of the air conditioning system, and issues avoidance commands to mobile robots in the relevant area.

[0136] when When the emergency threshold for Level 2 is exceeded, the platform immediately initiates the highest-priority emergency shutdown procedure. This procedure includes: issuing an emergency braking broadcast command to the underlying execution drive layer, forcing all mobile robots to lock their wheels at their current positions; simultaneously triggering the closing action of fireproof roller shutters to isolate the controlled area; and automatically cutting off the power supply circuits of non-critical equipment while maintaining the emergency power supply to the cold chain protection system until the alarm is manually cleared.

[0137] This embodiment details the timing coordination and logical flow of the aforementioned functional modules in actual operation, covering the entire closed-loop process from environment initialization, demand perception, decision scheduling to execution feedback. This process is uniformly coordinated by a central management server and directs the collaborative work of distributed sensors and robot clusters through an industrial local area network.

[0138] The system begins its operation during the initialization phase. The management server first loads the pre-built high-precision raster map and establishes communication connections with all RFID readers, temperature and humidity sensors, and security cameras within the warehouse. The system then executes a full inventory count command via the RFID middleware, reading the current... Update the database with the actual inventory levels of vaccine-like products based on their in-stock label data. Meanwhile, the environmental monitoring subsystem began continuously collecting temperature data from various areas to construct an initial environmental thermal field model, providing environmental constraint parameters for subsequent path planning.

[0139] Entering the periodic monitoring and decision-making phase, the system triggers a demand forecasting process every preset time step (e.g., every 5 minutes). The management server calls the aforementioned improved LSTM model to calculate the predicted demand for the next time window based on the latest inventory depletion rate and environmental factors. Subsequently, the system determines whether there is a potential gap based on the replenishment trigger logic. If the calculated replenishment demand... If the value is greater than zero, the system generates a task to be executed that includes the vaccine type, target pickup location, and target replenishment location, and pushes the task into the global task pool.

[0140] For tasks awaiting assignment in the task pool, the system executes a dynamic allocation strategy based on utility functions to determine the best robot object to perform the task. Define the... Robot No. 1 targets the first Matching utility score for task # The calculation formula is as follows:

[0141] ;

[0142] in, Represents robots With the task Based on the matching score, the system selects the robot with the highest score as the task executor. Represents robots At any moment Real-time battery state of charge; This indicates the minimum reserve power threshold required for the robot to perform its task; Indicates the battery's full charge capacity; The normalized effective remaining power factor is used to ensure that only robots with sufficient power participate in scheduling; Represents the weighting coefficient of the power factor; Represents robots Current location arrives at the mission The estimated path length from the starting point, which is determined by... The algorithm provides a fast estimate. This represents the maximum distance constant between the warehouse's diagonals; The weighting coefficients representing the distance cost; Represents robots The idle state indicator variable takes a value of 1 when the robot is in an idle standby state, and a value of 0 when it is performing other tasks. This indicates the reward weight for idle status, which is used to prioritize the use of idle resources and avoid interrupting robots that are currently working.

[0143] Once the robot to perform the task is identified The system sends the specific waypoint sequence of the task to the robot. The robot chassis responds immediately, using the aforementioned Dynamic Window (DWA) method for local path planning and obstacle avoidance navigation. During movement, the robot uploads its pose coordinates to the monitoring platform in real time, and the platform updates its trajectory synchronously in the digital twin interface. When the robot reaches the designated vaccine storage shelf, the chassis stops moving and locks the brakes, while simultaneously sending a positioning signal to the robotic arm controller.

[0144] After receiving the signal, the robotic arm activates its end-effector depth camera for visual servo alignment, uses admittance control algorithms to flexibly grasp the target vaccine box, and places it into the robot's onboard temperature-controlled transport box. Once placed, the robot releases its brakes and moves to the replenishment station according to its planned path. At the replenishment station, the robot repeats the grasping action to place the vaccine box into the designated chute. At this point, the RFID reader at the chute entrance reads the vaccine tag again, and the system compares the outbound and inbound records. When they match, the system determines that the replenishment task is complete and automatically updates the inventory database. Numerical values, and the robot Its state is reset to idle, awaiting the next round of scheduling instructions.

Claims

1. A vaccine intelligent restocking robot control system integrating multi-sensor perception, characterized in that, The application environment includes vaccine storage areas, distribution areas, and logistics channels, including: Mobile replenishment robots, vaccine distribution terminals, management servers, and wireless communication networks; The management server communicates with the mobile replenishment robot and the vaccine distribution terminal respectively through the wireless communication network; The mobile replenishment robot includes an onboard computing unit, a robotic arm actuator, and a multi-dimensional sensor module. The robotic arm actuator is equipped with a flexible suction cup assembly at its end. The multi-dimensional sensor module includes an environmental perception group and an operational perception group. The onboard computing unit is configured to perform autonomous positioning and navigation using data from the environmental perception group, identify vaccine objects using data from the task perception group, and control the robotic arm actuator to cooperate with the flexible suction cup assembly to complete vaccine handling. The management server is equipped with a time series prediction module, which is used to generate replenishment instructions based on historical circulation data and vaccine expiration information, and send them to the mobile replenishment robot. The multi-dimensional sensor module adopts a hierarchical distributed layout: The environmental perception group includes a lidar installed on the top of the mobile replenishment robot, a depth camera installed on the side of the body, and an inertial measurement unit installed at the center of the chassis. The operation sensing group includes an industrial RGB camera and RFID reader antenna installed at the end of the robotic arm actuator, as well as a temperature and humidity sensor integrated into the flexible suction cup assembly. The sensing area of ​​the operation sensing group moves synchronously with the movement of the robotic arm actuator; The airborne computing unit has a built-in adaptive fusion positioning module, which is configured to receive the scanning and matching pose of the lidar and the estimated pose based on the visual inertial odometry in real time. The adaptive fusion positioning module dynamically adjusts the fusion weights based on the matching residual value between the current laser scanning data and the pre-built map: When the matching residual value is less than a preset threshold, the weight of the scanning matching pose is increased; When the matching residual value is greater than the preset threshold, the weight of the inferred pose is increased to output the robot's optimal estimated pose. The onboard computing unit performs path planning based on a dynamic potential field: An obstacle repulsive potential field is superimposed on a grid map, and the intensity of the obstacle repulsive potential field increases as the distance to the obstacle decreases. An improved A* algorithm is used to search for the path node sequence with the minimum overall cost, and the path node sequence is smoothed to generate a smooth trajectory that conforms to the kinematic constraints of the mobile replenishment robot.

2. The intelligent vaccine restocking robot control system integrating multi-sensor perception as described in claim 1, characterized in that, The onboard computing unit executes confidence-based inventory verification logic: The number of visual detections output by the industrial RGB camera and the number of radio frequency reads output by the RFID reader are acquired simultaneously. Calculate the difference between the number of visual detections and the number of radio frequency reads; When the difference exceeds the allowable error range, the airborne computing unit selects the one with higher confidence as the final confirmed inventory quantity based on the weighted comparison result of the average confidence level of visual inspection and the radio frequency signal strength value.

3. The intelligent vaccine restocking robot control system integrating multi-sensor perception as described in claim 1, characterized in that, The time series prediction module employs a long short-term memory network model that incorporates an attention mechanism; The Long Short-Term Memory Network Model uses the actual consumption of vaccines, the reciprocal of the remaining shelf life, and the variance of environmental temperature fluctuations as input feature sequences. The attention mechanism assigns weights to historical information at different time steps based on the hidden layer state, with a focus on enhancing the impact of near-expiration vaccine data on the prediction results, thereby outputting the predicted replenishment demand at the next moment.

4. The intelligent vaccine restocking robot control system integrating multi-sensor perception as described in claim 1, characterized in that, The management server executes a task scheduling strategy based on expiration date priority. When multiple replenishment orders are generated, the management server calculates the comprehensive priority score for each replenishment order; The overall priority score is negatively correlated with the remaining shelf life of the vaccine, negatively correlated with the current inventory level, and positively proportional to the distance of the restocking path; The management server sorts the replenishment instructions according to the comprehensive priority score, and prioritizes the release of vaccine batches with the shortest remaining expiration date.

5. The vaccine intelligent replenishment robot control system integrating multi-sensor perception according to claim 1, characterized in that, The intelligent vaccine replenishment robot control system also includes a laser counter located at the delivery port of the distribution area; When the mobile replenishment robot performs a replenishment task, the onboard computing unit controls the flexible suction cup assembly to pick up the vaccine box and place it at the delivery port based on the target depth. The management server compares the actual number of pulses reported by the laser counter with the planned replenishment quantity: If they match, the inventory is updated; if they do not match, a supplementary task is generated or an anomaly alarm is triggered, forming a closed-loop verification.

6. The intelligent vaccine restocking robot control system integrating multi-sensor perception according to claim 1, characterized in that, The management server is configured with an anomaly severity assessment module; The anomaly severity assessment module collects data from the temperature and humidity sensor, the operating status of the mobile replenishment robot, and fire alarm signals in real time. The anomaly severity assessment module calculates a comprehensive anomaly severity index by weighting and summing temperature control risk factors based on the magnitude of real-time temperature exceeding the critical threshold, equipment failure factors based on the proportion of faulty robots, and fire risk factors based on fire alarm signals. When the comprehensive anomaly severity index exceeds a first threshold, an environmental adjustment and avoidance command is triggered. When the comprehensive anomaly severity index exceeds the second threshold, an emergency braking command is triggered and the power supply to non-critical equipment is cut off.

7. The intelligent vaccine restocking robot control system integrating multi-sensor perception according to claim 1, characterized in that, The management server includes a digital twin monitoring platform; The digital twin monitoring platform uses real-time sensor data to construct a virtual warehouse model and divides the warehouse area into multiple monitoring grids; The digital twin monitoring platform calculates a comprehensive risk index based on the temperature deviation and robot density within each monitoring grid, and renders the risk level of each monitoring grid in the form of a heat map.

Images