Hospital logistics task distribution method, device and equipment and storage medium

Abstract

Embodiments of the present application provide a hospital logistics task distribution method, device, equipment and storage medium, comprising: determining whether the goods need to be supplemented according to the first preset model, the current operation scheduling information and the remaining quantity of the goods, and generating a goods distribution task when it is determined that the goods need to be supplemented. The object in the monitoring video image is identified and the object position in the monitoring video image is determined according to the dynamic prompt library; the object position is mapped into an instant traffic cost matrix according to the identification result of the object; the people flow congestion coefficient is determined according to the current number of each department of the hospital, the number of appointments in the future time period, the operation scheduling information and the second preset model; the traffic path cost is determined according to the people flow congestion coefficient and the instant traffic cost matrix, and the candidate path is determined; the optimal departure time and the distribution path are determined by simulating and deducing according to the candidate path, and the goods distribution task is executed according to the distribution mode.

Description

Technical Field

[0001] This application relates to the field of medical Internet of Things, and more particularly to a method, apparatus, equipment, and storage medium for logistics task delivery in hospitals. Background Technology

[0002] The hospital logistics system is the lifeblood of a medical institution, its efficiency directly impacting clinical treatment quality, patient safety, and hospital operating costs. With increasing demands for refined medical management, the demand for internal hospital supply distribution has exploded, encompassing a wide range of items including pharmaceuticals, high-value consumables, laboratory specimens, surgical instruments, linens, and medical waste. These supplies are characterized by their complexity, small batch sizes, extremely high time sensitivity, and significant biosafety risks. Currently, hospital logistics distribution relies primarily on traditional manual delivery or single-function automated equipment. However, facing increasingly complex medical scenarios, existing technological solutions have revealed serious limitations in practical applications, failing to meet the high-efficiency operational needs of modern hospitals. Summary of the Invention

[0003] This application provides a method, apparatus, equipment, and storage medium for the delivery of logistics tasks in a hospital.

[0004] This application provides a hospital logistics task delivery method, including: acquiring current surgical schedule information and remaining quantity of items from different departments; determining whether the items need to be replenished based on a first preset model, the current surgical schedule information, and the remaining quantity of the items; if replenishment is required, generating an item delivery task to complete the replenishment; placing the item delivery task into a dynamic pool and determining a delivery mode based on the delivery time and delivery location corresponding to the item delivery task; acquiring surveillance video images and identifying and determining the positions of objects in the surveillance video images based on a dynamic prompt library, wherein the dynamic prompt library is used to record... The system records text information about objects in the hospital environment; maps the object recognition results and object locations to an instantaneous access cost matrix for a digital twin map; determines the congestion coefficient of a target area in the hospital within a future time period based on the current number of patients in each department, the number of appointments in a future time period, surgical scheduling information, and a second preset model, whereby the congestion coefficient represents the congestion situation in the future time period; determines the access path cost based on the congestion coefficient and the instantaneous access cost matrix, and determines candidate paths based on the access path cost; performs simulations based on the candidate paths to determine the optimal departure time and delivery path, and executes the item delivery task according to the delivery mode.

[0005] This application provides a hospital logistics task delivery device, comprising: an acquisition module for acquiring current surgical schedule information of different departments and the remaining quantity of the items; a first generation module for determining whether the items need to be replenished based on a first preset model, the current surgical schedule information, and the remaining quantity of the items, and generating an item delivery task to complete the replenishment when replenishment is determined; a first determination module for placing the item delivery task into a dynamic pool and determining a delivery mode based on the delivery time and delivery location corresponding to the item delivery task; and a second determination module for acquiring surveillance video images and identifying and determining the positions of objects in the surveillance video images based on a dynamic prompt library, wherein the dynamic prompt library is used to record medical... The system comprises: a first module for generating textual information about objects in the hospital environment; a second module for generating an instantaneous access cost matrix for mapping object recognition results and object locations into a digital twin map; a third module for determining the congestion coefficient of a target area in the hospital within a future time period based on the current number of patients in each department, the number of appointments in a future time period, surgical scheduling information, and a second preset model, wherein the congestion coefficient represents the congestion situation in the future time period; a fourth module for determining the access path cost based on the congestion coefficient and the instantaneous access cost matrix, and determining candidate paths based on the access path cost; and an execution module for performing simulations based on the candidate paths to determine the optimal departure time and delivery path, and executing the item delivery task according to the delivery mode.

[0006] This application provides a device, including: a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program for: acquiring historical consumption rates of items in different departments, current surgical schedule information, and the remaining quantity of the items; determining whether the items need to be replenished based on a first preset model, the historical consumption rate of the items, the current surgical schedule information, and the remaining quantity of the items; when replenishment is determined, generating an item delivery task to complete the replenishment; placing the item delivery task into a dynamic pool and determining a delivery mode based on the delivery time and delivery location corresponding to the item delivery task; acquiring surveillance video images and identifying objects in the surveillance video images according to a dynamic prompt library and determining the surveillance video images. The system identifies the location of objects in the hospital environment. A dynamic prompting library records textual information about objects within the hospital environment. Based on object recognition results and their locations mapped to a digital twin map, an initial real-time access cost matrix is ​​generated. Using the current number of patients in each hospital department, the number of appointments scheduled for a future time period, surgical scheduling information, and a second preset model, a congestion coefficient for the target area within the hospital is determined for the future time period. This congestion coefficient represents the congestion situation within the future time period. Based on the congestion coefficient and the initial real-time access cost matrix, the system determines the access path cost and the corresponding real-time access cost matrix, and then determines candidate paths based on the real-time access cost matrix. Simulations are performed based on the candidate paths to determine the optimal departure time and delivery route, and the item delivery task is executed according to the delivery mode.

[0007] This application provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by one or more processors, it causes the one or more processors to implement the steps in the above-described method.

[0008] The technical solution disclosed herein obtains the current surgical schedule information and the remaining quantity of items for different departments; determines whether items need to be replenished based on a first preset model, the current surgical schedule information, and the remaining quantity of items; if replenishment is required, generates an item delivery task to complete the replenishment; places the item delivery task into a dynamic pool and determines the delivery mode based on the delivery time and delivery location corresponding to the item delivery task; acquires surveillance video images and identifies objects in the surveillance video images and determines the positions of objects in the surveillance video images based on a dynamic prompt library, the dynamic prompt library being used to record the text of objects in the hospital environment. Information; Based on the object recognition results and the object location mapping, an instant access cost matrix is ​​generated for the digital twin map; Based on the current number of people in each department of the hospital, the number of appointments in the future time period, the surgical scheduling information, and the second preset model, the congestion coefficient of the target area in the hospital in the future time period is determined, and the congestion coefficient represents the congestion situation in the future time period; Based on the congestion coefficient and the instant access cost matrix, the access path cost is determined, and candidate paths are determined based on the access path cost; Based on the candidate paths, simulation is performed to determine the optimal departure time and delivery route, and the item delivery task is executed according to the delivery mode.

[0009] This system utilizes data from various hospital departments, including current staff numbers, future appointment numbers, surgical scheduling information, and a second pre-set model, to determine the congestion coefficient of a target area within the hospital in the future. This coefficient represents the level of congestion in the future. Based on the congestion coefficient and an instantaneous traffic cost matrix, the system determines the path cost and selects candidate paths. This multi-source data-driven spatiotemporal congestion prediction not only solves problems virtually but also predicts future congestion, allowing for advance adjustments to task execution timing and paths to avoid deadlocks and delays during actual operation. Furthermore, the selection of candidate paths based on the instantaneous traffic cost matrix enables globally optimal path and time decisions, maximizing robot utilization. This data-driven dynamic rule-based scheduling replaces manual experience-based scheduling, achieving intelligent and adaptive operation of hospital logistics. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating an exemplary hospital logistics task delivery method according to this application; Figure 2 This is a schematic diagram illustrating the process of determining whether the item needs to be replenished, as an exemplary embodiment of this application. Figure 3A schematic diagram of a hospital logistics task delivery method apparatus as an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a device that is an exemplary embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In established hospitals, especially in many older campuses, building structures are fixed, departmental spaces are cramped, corridors are narrow with huge traffic flow, and elevator resources are extremely congested. For many older hospitals that cannot be renovated or have difficult-to-wire systems, heavy-duty automation solutions are simply not feasible. Furthermore, conventional mobile robots relying on related technologies often depend on static maps and preset rules, lacking deep understanding and adaptive capabilities to narrow, dynamically changing environments, making them prone to congestion or deadlocks in complex human-robot mixed environments. Delivery demands are heterogeneous and complex, making it difficult for a single model to address the supply shortage.

[0013] Furthermore, the delivery needs of hospital clinical departments are highly heterogeneous. For example, drug delivery requires accuracy and cold chain support. High-value consumables require full traceability. Laboratory specimens have extremely high requirements for timeliness and biosafety (e.g., leak-proof and shock-proof). Surgical instruments need to respond instantly in sync with the surgical procedure. Existing manual or single-robot solutions often lack flexible task scheduling capabilities, making it difficult to simultaneously optimize delivery strategies for multiple product categories within the same system. For instance, existing systems struggle to intelligently combine along-route drug delivery with empty container collection tasks, resulting in high robot idle rates and an inability to meet the complex supply demands of diverse product types, small batches, and high timeliness.

[0014] Due to its high reliance on manual labor, management suffers from deep blind spots, yet most hospitals still heavily depend on manual delivery teams. This model inherently suffers from high labor intensity, high labor costs, and the inability to achieve continuous, efficient 24-hour operation. More seriously, manual scheduling lacks data support, relying primarily on the dispatcher's personal experience for route planning and resource allocation, resulting in significant information silos. Managers cannot grasp the overall status of material flow in real time, leading to uneven resource allocation, arbitrary delivery routes, long transit times, and even delays in critical supplies impacting treatment efficiency. This black-box management model keeps hospital logistics in a state of inefficiency and high resource consumption for a long time.

[0015] Hospital environments are highly dynamic and uncertain. Peak-hour congestion, random fluctuations in elevator wait times, unexpected medical equipment occupancy, and unforeseen equipment failures all alter access conditions in real time. Existing scheduling systems are mostly based on static rules or simple heuristic algorithms, relying solely on historical experience for path planning. They cannot perceive and respond to dynamic environmental changes in real time. When encountering sudden congestion, they often can only passively wait or perform local obstacle avoidance, lacking dynamic replanning capabilities based on global real-time data and pre-emptive verification mechanisms. This results in the inability to calculate the globally optimal solution, inefficient material flow, and severely impacts the continuity of clinical care. This creates a bottleneck due to poor dynamic adaptability and a lack of accurate calculation capabilities for the globally optimal solution.

[0016] In summary, this embodiment provides a method for automatic smart logistics task orchestration that does not require large-scale infrastructure transformation, can adapt to complex heterogeneous needs, eliminates reliance on human experience, and has holographic environmental perception and dynamic self-healing capabilities, in order to solve the severe challenges faced by hospitals in four dimensions: inability to make changes, insufficient supply, poor management, and inaccurate calculations.

[0017] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic flowchart illustrating a hospital logistics task delivery method as an exemplary embodiment of this application. The method 100 provided in this embodiment can be deployed on a server, such as a server, service cluster, or cloud platform. The method 100 includes the following steps: Step 101: Obtain the historical consumption rate of items, current surgical schedule information, and remaining quantity of items for different departments.

[0019] Step 102: Determine whether the items need to be replenished based on the first preset model, the historical consumption rate of the items, the current surgical schedule information, and the remaining quantity of the items. If it is determined that replenishment is needed, generate an item delivery task to complete the replenishment.

[0020] Step 103: Place the item delivery task into the dynamic pool and determine the delivery mode based on the delivery time and location corresponding to the item delivery task.

[0021] Step 104: Acquire surveillance video images, identify objects in the surveillance video images based on the dynamic prompt library, and determine the positions of objects in the surveillance video images.

[0022] The dynamic prompting library is used to record text information about objects in the hospital environment.

[0023] Step 105: Based on the object recognition results and the object location mapping into a digital twin map, generate an initial real-time access cost matrix.

[0024] Step 106: Based on the current number of patients in each department of the hospital, the number of appointments in the future time period, the surgical scheduling information, and the second preset model, determine the congestion coefficient of the target area in the hospital in the future time period.

[0025] Among them, the crowd congestion coefficient represents the crowd congestion situation in the future time period.

[0026] Step 107: Determine the travel path cost and the corresponding instant travel cost matrix based on the pedestrian congestion coefficient and the initial instant travel cost matrix, and determine the candidate path based on the instant travel cost matrix.

[0027] Step 108: Perform simulations based on candidate routes to determine the optimal departure time and delivery route, and execute the item delivery task according to the delivery mode.

[0028] This embodiment addresses the problems of existing hospital logistics robots in unstructured environments, such as low efficiency in passive obstacle avoidance, lack of prediction of peak traffic trends, and limited scheduling strategies when facing unknown obstacles.

[0029] This embodiment constructs a fully automated closed-loop system of perception, prediction, inference, and self-healing through visual perception, digital twin parallel extrapolation mechanism, and spatiotemporal congestion prediction model based on multi-source data. It can not only identify unknown obstacles but also predict road clearing times by combining real-time patient numbers in departments. Intelligent decision-making identifies the optimal path, achieving a leap from passive response to proactive spatiotemporal coordination, ensuring zero-interruption, low-energy consumption, and safe and efficient operation of hospital logistics in complex and dynamic environments.

[0030] The above steps are explained in detail below: Step 101: Obtain the historical consumption rate of items, current surgical schedule information, and remaining quantity of items for different departments.

[0031] This information is obtained by building a multi-dimensional demand parsing engine on the server side. Specifically, the multi-dimensional demand parsing engine can connect to the Hospital Information System (HIS) and the Medical Consumables Management System (SPD) in real time via standard API interfaces. Both systems can be deployed on different service nodes. Thus, the current surgical schedule information for different departments can be obtained through the HIS, along with the SPD data, and historical consumption rates and remaining quantities of items can be obtained through the SPD system.

[0032] For example, the engine automatically polls the system's data and caches historical consumption rate sequences of departmental items, current surgical schedules including surgery type, estimated duration, surgeon's preferences, and real-time inventory levels, such as quantity or expiration date. This enables structured data capture and multi-source heterogeneous data fusion processing.

[0033] In other words, in addition to obtaining the remaining quantity of items, you can also obtain the expiration date of the remaining items.

[0034] Step 102: Determine whether the items need to be replenished based on the first preset model, the historical consumption rate of the items, the current surgical schedule information, and the remaining quantity of the items. If it is determined that replenishment is needed, generate an item delivery task to complete the replenishment.

[0035] The first preset model refers to the time series prediction model, which is a dual-drive spatiotemporal demand prediction model. This model is based on an improved time series analysis framework, such as using LSTM long short-term memory network or seasonal ARIMA model as a base, and achieves accurate prediction of material demand through a specific multi-source feature fusion mechanism and scenario-based training process.

[0036] The LSTM (Long Short-Term Memory) network is a chain-like network unfolded along a timeline, composed of multiple structurally identical memory units connected in series. Furthermore, the output of the LSTM network is connected to a fully connected layer. Each unit updates its internal long-term memory additively through three gating mechanisms, generating a hidden state that is output and passed to the next time step. Specifically, this includes a timeline, a forget gate, an input gate, and an output gate on that timeline, as well as cell state updates. The forget gate, input gate, output gate, and cell state updates constitute an LSTM unit, and there are multiple LSTM units. The cell state update is achieved by adding the old experience filtered by the forget gate and the new knowledge filtered by the input gate, and then using the output gate to generate the hidden state for the current time step based on the updated cell state, thus obtaining the final output.

[0037] The LSTM (Long Short-Term Memory) network employs a stacked two-layer LSTM structure, followed by a fully connected layer to output predicted values. The first LSTM layer captures short-term fluctuations, while the second LSTM layer captures long-term dependencies. Through hyperparameter grid search on the hospital's monthly consumption dataset, the number of hidden units in a single layer is limited to the range [64, 128]. For example, 128 is sufficient for adequate fitting, avoiding overfitting due to excessive numbers and underfitting due to insufficient numbers. To prevent overfitting caused by the uneven distribution of surgical and non-surgical days in the training data, Dropout is applied between LSTM layers and before the fully connected layer, with a ratio limited to 0.2 to 0.5, preferably 0.3. Mean Squared Error (MSE) is used as the loss function, with the Adam optimizer; the initial learning rate is limited to 0.001 and adjusted using cosine annealing; the batch size is dynamically set between 16 and 64 based on the sample size of the department's historical data. The input time window N should not exceed 72 hours to avoid training difficulties due to excessively long gradients, while ensuring that the model is not sensitive to historical information exceeding three days, better reflecting the short-cycle response characteristics of clinical material replenishment. To achieve spatiotemporal demand prediction for materials in different departments across the hospital, a learnable embedding vector for each department is introduced into the LSTM input. During training, the consumption data of all departments are shuffled and fed into the same LSTM core, sharing the temporal variation patterns, but the department embeddings distinguish the consumption baseline and surgical modes of different departments, thus enabling reasonable predictions even when there are sudden changes in the surgical volume of a particular department.

[0038] Each time step input of the LSTM consists of three concatenated features to inject planned spatial knowledge into the time series. These include: historical consumption features (as described above), representing the consumption of goods over N consecutive time units (e.g., N=24 hours, sampled at hourly granularity), forming a time series of length N; surgical scheduling-driven features, extracting planned consumption demands for the next M time units (e.g., M=2 hours) from the current surgical scheduling information. This feature can be quantified as a surgical consumption index, for example, calculating the expected increase in item consumption within each future time interval based on the type of pre-scheduled surgery, its estimated duration, and the standard item consumption corresponding to that procedure, and then aligning this future sequence to the corresponding input time step; and static supplementary features, such as the current remaining quantity of items, whether it's a working day, and the department type, which can be represented through one-hot encoding or embedding vectors. This input design allows the LSTM to simultaneously learn the inertial trend of item consumption and sudden increases in surgical planning demands, forming a dual-driven feature fusion mechanism.

[0039] A seasonal ARIMA model is a combination of a traditional ARIMA model and a seasonal component. The model includes a time series pattern that can be decomposed into a non-seasonal component and a purely seasonal component, both of which work together. Specifically, a seasonal ARIMA model is explicitly constructed from differencing operators, autoregressive and moving average polynomials. Through statistical fitting, such as maximum likelihood estimation, the equation coefficients are calculated in one step. Data is preprocessed through differencing, and short-term and long-term periodic memories are modeled simultaneously using a linear combination. The model determines the parameters p, d, q, P, D, Q, s based on historical data and estimates the parameters in all polynomials. Difference is applied to the historical data to obtain the differencing results. The known differencing results and unknown random error terms are substituted into the known model equations. The equations predict future stationary values. Then, by performing a reverse differencing operation, the trend and periodicity are added back to obtain the predicted result for the original series.

[0040] This application instantiates it as a SARIMAX(p,d,q)(P,D,Q,s) model, incorporating manual scheduling as an exogenous variable. For the non-seasonal component (p,d,q), the difference order d is limited to 0 or 1. For example, after confirming stationarity through a unit root test, d=0 is preferred to prevent information loss due to multiple differences. The autoregressive order p and the moving average order q are limited to 0≤p≤3 and 0≤q≤3, respectively, and optimized within the limited set using the AIC criterion. For the seasonal component (P,D,Q,s), the seasonal cycle s is set to a 24-hour intraday cycle or a 7-day weekly cycle based on the hospital's material consumption pattern. The seasonal difference order D is limited to 0 or 1, and the seasonal autoregressive order P and the seasonal moving average order Q are both limited to 0≤P and Q≤2. This strict constraint avoids spurious periodic fitting caused by high-order seasonal parameters, and is particularly suitable for a 7-day weekly cycle pattern with alternating surgical and non-surgical days. Introducing surgical scheduling as an exogenous variable explicitly extends the model to a SARIMAX model with an exogenous regression term. The exogenous variable X(t) is defined as the number of surgeries requiring this item that are currently in progress or about to begin at future time t, expressed in surgical impact equivalents. This exogenous variable directly affects the model's conditional mean equation: Y(t) = constant term. The model incorporates a combination of partial, seasonal, and error components. Demand spikes caused by temporary surgeries, which cannot be explained by the autocorrelation of historical trends alone, are captured by the exogenous variable coefficient β, achieving a dual-driven linear effect. Model fitting and data preprocessing constraints are implemented: a Box-Cox transformation is applied to the item consumption sequence, limiting the transformation parameter λ to the [-1,2] interval to stabilize variance. The maximum likelihood estimation iteration count is limited to 100 times, with a convergence tolerance of 1e-5 to ensure controllable computational latency. When the historical data sample size is less than two complete seasonal cycles (e.g., <14 days of hourly data), the seasonal component is forcibly disabled, meaning the model reverts to ARIMAX(p,d,q) to reduce uncertainty. An adaptive switching mechanism prioritizes the LTM model if a department has sufficient historical consumption data (e.g., sequence length ≥30 days) and high surgical variability. If it is a newly established department or a newly introduced item, and consumption exhibits a clear cyclical pattern, the SARIMAX model is invoked. This selection method is written into the control code with executable threshold conditions to ensure clear decision-making logic. The supplementary judgment condition is to obtain the predicted total consumption in the next H hours, for example, H=2. Then, combined with the current remaining quantity of the item and the safety stock threshold, if the remaining quantity - predicted consumption < safety stock, it is determined that replenishment is needed. This judgment logic is specific and the parameters are definite, avoiding ambiguity.

[0041] The training process of the aforementioned first time series prediction model, such as LSTM, includes: acquiring the historical consumption rate of an item, the corresponding historical surgical schedule information, and the historical consumption quantity of the item; the historical consumption rate of an item corresponds to a historical time earlier than and adjacent to the historical consumption quantity of the adjacent item; the historical surgical schedule information corresponds to a historical time earlier than and adjacent to the historical consumption quantity of the adjacent item; according to a sliding window, acquiring the fluctuation value within a preset time period in the historical consumption rate sequence and generating a corresponding consumption rate feature vector; training an initial time series prediction model based on the consumption rate feature vector corresponding to the historical consumption rate sequence, the quantization feature vector corresponding to the historical surgical schedule information, and the historical consumption quantity of the item to obtain the first time series prediction model.

[0042] In this context, the historical surgical schedule information corresponds to a time earlier than the historical consumption quantity of adjacent items, and the historical consumption rate of an item corresponds to a time earlier than the historical consumption quantity of adjacent items. Examples of this can be in hourly or shift units. The historical consumption rate of an item corresponds to the historical surgical schedule information, and both are earlier than the historical consumption quantity. For example, if the historical time 16:00 is earlier than and adjacent to the historical time 17:00, the historical consumption rate of the item is data before 16:00 on day A, and the historical consumption quantity of the item is data after 17:00 on day A. The historical surgical schedule information is the schedule information before 16:00 in the historical schedule information for day A. The same applies to shifts, so further explanation is unnecessary. It should be noted that the historical data mentioned above can be data corresponding to Day A, such as data from 00:00 to 16:00 on Day A, data from 17:00 to 24:00 on Day A, and the surgical schedule information before 16:00 on Day A. It can also include data corresponding to days prior to Day A, such as data before 16:00 on Day A, data from other days prior to Day A, and the surgical schedule information before 16:00 on Day A, as well as schedule information from days prior to Day A. The selection of corresponding data based on different prediction time periods is not elaborated further.

[0043] Unlike traditional univariate prediction that relies solely on historical sales figures, this embodiment constructs a composite input vector Xt that includes both static periodic features and dynamic event features, resulting in a multi-source feature vector. The aforementioned first time series prediction model is then trained based on this multi-source feature vector.

[0044] The static periodic feature extraction includes: obtaining fluctuation values ​​within a preset time period from the historical consumption rate sequence using a sliding window, and generating a corresponding consumption rate feature vector. For example, extracting the historical consumption rate sequence of items from different departments over the past N historical periods, processing it with a sliding window, and capturing the periodic fluctuation patterns or fluctuation values ​​of days, weeks, and months to form a basic material consumption rate vector Vbase. For example, the daily fluctuation value of A is 60, and the monthly fluctuation value of A is 2000.

[0045] Dynamic event feature extraction includes parsing the current surgical schedule information and transforming the unstructured surgical schedule information into a quantified surgical schedule feature vector. For example, parsing the current surgical schedule table and transforming the unstructured surgical information into a quantified feature vector Vsurg.

[0046] The process involves transforming unstructured surgical scheduling information into quantified surgical scheduling feature vectors. This includes: mapping different surgical types to consumable consumption coefficients based on the correspondence between different surgical types and consumable consumption coefficients; constructing a high-demand probability distribution curve for a predetermined timeframe based on the start time and expected duration of each surgical type; statistically analyzing the consumption habits of different surgeons for different items based on historical consumption data; and generating a quantified surgical scheduling feature vector based on the consumable consumption coefficients corresponding to surgical types, the demand probability distribution curve, and the surgeons' consumption habits. For example, surgical types are encoded, and different surgical types are mapped to consumable consumption coefficients based on the predetermined relationship between different surgical types and coefficients. For instance, coronary artery bypass surgery corresponds to a high consumable consumption coefficient, while pacemaker implantation corresponds to a low consumable consumption coefficient. A high-demand probability distribution curve for the next T minutes is constructed based on the start time and expected duration of different types of surgeries or surgical procedures, where high demand refers to demand exceeding a demand threshold. Based on the historical consumption of items, specific doctors' additional usage habits for specific consumables are statistically analyzed. For example, if it is determined that surgeon A uses more than a threshold for item A, the usage habit is considered strong, and this quantity can be mapped to a corresponding coefficient. This mapping relationship can be based on a pre-defined relationship between quantity and coefficient. Vbase and Vsurg are aligned and concatenated on the time axis to form an enhanced time-series input vector rich in business semantics. This vector serves as the input to the model, for example, as a training sample for the initial time-series prediction model. The model is then trained based on the historical consumption quantities of the items described above, resulting in the trained model.

[0047] For example, to ensure the model adapts to the complex fluctuations in hospital operations, model training can include a phase one: offline pre-training. This involves cleaning the data before training, collecting all material outbound records and surgical scheduling history data from the hospital over the past 3-5 years, and removing outliers caused by pandemics, system failures, etc. Training samples are then constructed, using hours or shifts as the time granularity, creating supervised learning sample pairs: (historical consumption rate sequence feature vector + historical surgical scheduling feature) -> (actual historical consumption at the next moment). It should be understood that while sample pairs can be constructed based on hours in the historical timeframe, the historical data acquisition still involves acquiring long-term data and decomposing it according to time to form different sample pairs. Examples include 1-hour data decomposition, 5-hour data decomposition, 24-hour data decomposition, 96-hour data decomposition, etc. Further details are omitted.

[0048] The LSTM was trained using the above samples to learn the nonlinear mapping relationship between surgical scheduling features and material consumption increment features, thus obtaining a general benchmark model.

[0049] For the seasonal ARIMA model, a corresponding prediction relationship is obtained based on the supervised learning sample pairs constructed above (historical time consumption rate sequence feature vector + historical time surgery scheduling features) -> (actual historical consumption at the next moment), so as to determine the predicted value. Other details will not be repeated here. Finally, both models obtain a nonlinear mapping relationship between the surgery scheduling features and the material consumption increment features, to obtain a general benchmark model.

[0050] Model training may include the second stage: online fine-tuning. In the incremental data learning step, the consumption quantity of items in the latest preset time in the future, such as the actual consumption data of the latest week, is automatically captured every day, and surgery scheduling execution deviation data, such as data of delayed surgeries and temporarily added operating tables, is also captured. The residual between the predicted value and the actual value is calculated. If the residual exceeds the threshold for K consecutive time steps, a lightweight fine-tuning mechanism is triggered, which only updates the weights of the last layer of the first time series prediction model, such as the weights of the fully connected layer, so that the model can quickly adapt to the latest consumption habits or seasonal changes of the department and prevent model aging.

[0051] Specifically, as Figure 2 shown, determining whether an item needs to be replenished according to the first preset model, current surgery scheduling information and the remaining quantity of the item includes: Step 201: Parse the current surgery scheduling information, and convert the unstructured surgery scheduling information into a quantized surgery scheduling feature vector. Step 202: Input the quantized surgery scheduling feature vector into the first time series prediction model, and output the future item consumption quantity of the corresponding department. Step 203: Determine whether the item needs to be replenished according to the future material consumption quantity and the remaining quantity of the item.

[0052] As described above, using the trained model and inputting the corresponding current real-time features, the estimated material consumption Qpred in the future time window, such as the next 30-60 minutes, is predicted in a rolling manner. The current item inventory level Qcurr of the SPD system is read in real time. The dynamic safety threshold Threshold = Qpred + SafetyBuffer is calculated, where SafetyBuffer is a dynamic buffer determined based on the distribution duration, and may also have a preset proportional relationship with the duration. If Qcurr < Threshold, it is immediately determined that the replenishment condition is triggered, a standard purchase order containing the recommended replenishment quantity is automatically generated, and converted into an initial distribution task unit.

[0053] In addition to the above methods, the corresponding task may also be generated through other methods such as instructions. Specifically, the method 100 further includes: acquiring a task instruction; extracting the intent and item-related information in the task instruction; generating an item distribution task according to the intent and the item-related information to complete the distribution of the item.

[0054] The methods for obtaining task instructions can include obtaining voice instructions or text instructions. For example, a server can receive voice instructions from the terminal.

[0055] Extracting intent and item-related information from task instructions can be achieved by: loading a pre-built medical logistics-specific corpus, which covers commonly used instructions by medical staff, such as colloquial instructions: "Station 3 is out of hemostats," "Emergency room needs two boxes of gloves," and "Put away the empty boxes." Each instruction also includes corresponding intents and entities. A pre-trained general-purpose natural language processing model is invoked, and scenario adaptation is performed using prompt engineering and domain entity linking technology. For example, by invoking a pre-trained general-purpose natural language processing model and using prompt engineering, intent and item-related information can be extracted from task instructions. Item-related information can include the name, quantity, target location, and urgency of the supplies or items. Intents can include emergency replenishment, routine requisition, and recycling requests.

[0056] The aforementioned pre-trained general natural language processing model can be a pre-trained language model based on the Transformer architecture, without changing its underlying network architecture. It includes an input layer, encoder, decoder, and output layer. The encoder layer can include multi-head self-attention and feedforward neural networks, and the decoder can include masked multi-head attention and encoder-decoder attention. The output layer can include linear layers and softmax. Each layer in the encoder and decoder of this architecture is wrapped with residual connections and layer normalization.

[0057] The received voice commands are converted to text using ASR (Automatic Speech Recognition) or the received text commands are concatenated with the current context information to construct an enhanced input vector rich in contextual semantics. The current context information refers to the information carried by the command, such as the speaker's department, current time, and associated operating room number. This information is sent along with the command. It could be a command sent by a medical professional via an instant messaging system, which would carry the user's department, time, and currently associated operating room number. The model learns the intent and entities corresponding to each command in the corpus based on the aforementioned corpus and the corresponding intents and entities. It then determines the corresponding intent and entities through the input vector and outputs structured data, such as JSON data, containing intent categories and key entity slots. The entities corresponding to the key entity slots can include the name, quantity, target location, and urgency of the supplies or items.

[0058] In addition, a medical supplies knowledge graph can be set as the post-processing rule layer to perform standardized mapping on the entities extracted by the above model. The graph may include the association relationships between entities and standardized codes. For example, mapping the "large box" in the expression "take a large box" to a specific SKU code, verifying the rationality of the quantity logic, and eliminating ambiguity in colloquial expressions.

[0059] It should be noted that the urgency level of a task can be determined through the above embodiments. For example, the urgency level of the task can be determined according to the above instruction, and the entity data can be determined through JSON data, such as the entity data of "immediately" or the entity data of "quickly". Meanwhile, in the case of generating a supplementary delivery task for items as described above, the urgency level of the task can also be determined. Specifically, as described above, it is determined whether the items need to be supplemented according to the relationship between the future material consumption quantity and the remaining quantity of the items. For example, the dynamic safety threshold Threshold = Qpred + SafetyBuffer is calculated, where SafetyBuffer is the dynamic buffer amount based on the delivery duration. If the remaining quantity of the item is less than Threshold and the quantity by which it is less reaches a preset ratio, for example, a preset threshold, the task is determined as an urgent task; if it does not reach the preset ratio, it can be set as a normal task, etc.

[0060] Step 103: Put the item delivery task into a dynamic pool, and determine the delivery mode according to the delivery time and delivery position corresponding to the item delivery task.

[0061] The delivery modes include a point-to-point direct delivery mode, a carpooling order combining mode, a round-trip empty box collection mode, and a continuous connection mode. The point-to-point direct delivery mode is single-line transportation, which is applicable to first-aid supplies with high urgency. The carpooling order combining mode is applicable to conventional supplies for multiple departments that share the same path and have overlapping time windows. The round-trip empty box collection mode refers to collecting waste packaging or empty boxes along the way after delivery. The continuous connection mode refers to the chained execution of multiple tasks.

[0062] Specifically, determining the delivery mode according to the delivery time and delivery position corresponding to the item delivery task includes: determining the delivery mode of the item delivery task according to the delivery time and delivery position corresponding to the item delivery task, as well as a genetic algorithm or an ant colony optimization algorithm.

[0063] The server-side scheduling center aggregates discrete initial task units into a dynamic demand pool. Based on operations research optimization algorithms, such as mature improved genetic algorithms or ant colony optimization algorithms, the optimal delivery mode is automatically selected and matched according to the spatiotemporal attributes of the tasks. A standard genetic algorithm framework is adopted, including core modules such as population initialization, fitness function calculation, selection, crossover, and mutation. The spatiotemporal attributes include task time, spatial location, and regional coverage as algorithm input parameters. Task time can include delivery time, which can be preset according to the urgency of the task, such as the expected delivery time. Spatial location and regional coverage can be determined by the delivery location.

[0064] Extract task data from the dynamic demand pool, extract the spatiotemporal attributes of the task data, and optimize the output of the optimal delivery mode matching result through the above genetic algorithm.

[0065] The above-mentioned genetic algorithm training or optimization process includes: based on historical delivery task data, dividing the historical delivery task data into a tuning set D_tune and a test set D_test, and setting reasonable upper and lower bounds for each parameter to be optimized, such as population size, crossover rate, and mutation rate. The optimal parameter configuration, such as population size, crossover rate, and mutation rate, is found through iterative optimization. Parameter adaptation is then performed according to the hospital's aforementioned spatiotemporal attribute constraints.

[0066] To avoid a large number of invalid solutions due to completely random initialization, semi-regular constraints are introduced during population initialization. These constraints include temporal proximity constraints, prioritizing tasks with overlapping delivery time windows (expected time ± allowable deviation) to be assigned to the same delivery batch (chromosome segment). Tasks with expected times exceeding a preset threshold (e.g., 30 minutes) are forcibly assigned to different batches, reducing the proportion of invalid solutions in the search space. Spatial clustering constraints use hospital floors or ward units as the smallest spatial grid. Multiple tasks located on the same floor and with interconnected routes are prioritized and assigned to the same pickup / delivery path segment. Spatial connectivity is determined using the hospital's road network topology. If there is no direct path between two task nodes (e.g., located in clean and contaminated areas with no accessible buffer zones), they are prohibited from appearing in the same continuous delivery sequence. The fitness function is designed with a spatiotemporal penalty term. It consists of three weighted components that directly reflect the spatiotemporal constraints of hospital delivery: Fitness = ω1 Total Delivery Efficiency Score + ω2 Time Penalty Term + ω3 Spatial Violation Penalty Term, where ω1, ω2, and ω3 are weighting coefficients. The total delivery efficiency score is the reciprocal sum of the completion times of all tasks, encouraging early completion. The time penalty term, for each task j, accumulates a timeout penalty λ·max(0, actual time - expected time) if the actual execution time is later than the expected time. For vehicles such as robots, additional penalties are added if the total trip time exceeds their endurance or disinfection cycle threshold. The spatial violation penalty term examines the number of floor jumps between adjacent task nodes in the path sequence; each non-connected elevator floor jump is included in the penalty value. If the path needs to cross the clean-contaminated physical isolation boundary and no buffer time is reserved in the planning, a larger fixed penalty is accumulated. Through these penalties, the genetic algorithm automatically eliminates solutions that violate the hard spatiotemporal constraints and evolves towards a better solution. The genetic operator operations are constrained by spatial coherence. Boundary control is implemented for the crossover operation, allowing crossover points to be set only after the task batch separator on the chromosome, or at the natural boundary between tasks on different floors or in different wards. Crossover cannot be arbitrarily truncated within densely populated task segments to prevent disruption of verified spatial coherence. Local search restrictions are imposed on the mutation operation. If a task is re-inserted into another location during mutation, the candidate insertion point must be within the spatially adjacent physical node range of the task's original location, such as within the same floor and the floors above and below it. After insertion, path validity checks are required, such as the existence of the passageway and whether access permissions are satisfied. This restriction reduces the mutation search radius and improves the algorithm's convergence speed. The genetic algorithm parameters themselves have a spatiotemporal self-adaptation mechanism. Based on the spatiotemporal characteristics of the task set in the dynamic demand pool, the core parameters of the genetic algorithm are adjusted online. Population size: The default value is 50. If the number of tasks N < 20, the population = 20. If N ∈ [20, 60], the population = N. If N > 60, the population = 60, preventing insufficient population size for high-dimensional chromosomes.The task spatial distribution dispersion, i.e., the average Euclidean distance between all task locations, is multiplied by 1.3 if it exceeds half the diagonal of the floor. The default crossover rate is 0.8. If the standard deviation of the expected task time is greater than 20 minutes, indicating a large difference in time urgency, the crossover rate is increased to 0.9 to accelerate combinatorial optimization. If the distribution is uniform, it is reduced to 0.7 to maintain stability. The default mutation rate is 0.1. If the highest time-sensitive tasks in the dynamic pool, such as emergency surgical supplies, account for ≥30%, the mutation rate is increased to 0.2 to strengthen global search and quickly find response paths, preventing getting stuck in local optima and delaying emergency tasks. The default termination generation is 200. When there is no improvement in the optimal solution for 30 consecutive generations, or when the solution is feasible after 100 generations, there is no time penalty, and the algorithm is forcibly terminated. The overall time limit is set to 2 seconds during online operation to ensure real-time scheduling. This self-adaptive mechanism uses statistics extracted from task data as trigger conditions and computational inputs to ensure the stable and efficient operation of the genetic algorithm under different time periods and task densities in the hospital. Formal definition of constraints: In a hospital setting, the spatiotemporal constraints are as follows: Hard time window: Task j must be executed within [expected time - Δt1, expected time + Δt2], where Δt1 = 15 min and Δt2 = 10 min. For emergency tasks, Δt2 = 5 min. Spatial connectivity: There must be a physically passable path between adjacent points in the task sequence. Access control systems and elevator operation status are included in the digital twin map. Cross-zone restrictions: Consecutive non-connected area pairs are not allowed in the task sequence, such as no buffer between clean and contaminated areas. If crossing zones is necessary, intermediate nodes must be permissible transition zones. Residual capacity: The total travel distance + time of a single vehicle in a single sequence must be ≤ 85% of the battery's safe operating range / disinfection cycle.

[0067] Based on the matching results, the system automatically pre-assigns suitable container types and robot models. Depending on the delivery method of the task, it selects different sized containers and robots capable of carrying multiple or large containers. For example, in ride-sharing and continuous shuttle modes, the system can select large containers or multiple different containers, as well as robots capable of carrying multiple or large containers.

[0068] Step 104: Acquire surveillance video images, identify objects in the surveillance video images based on the dynamic prompt library, and determine the positions of objects in the surveillance video images.

[0069] The dynamic cueing library records textual information about objects in the hospital environment, including object names specific to the hospital setting, such as IV stands, wheelchairs, and medical waste bins. These objects can be considered obstacles.

[0070] The method of acquiring surveillance video images is through image acquisition devices, such as cameras, set up in different spatial locations within the hospital.

[0071] Specifically, the process involves acquiring surveillance video images and identifying and determining the locations of objects in the images based on a dynamic cue library. This includes: extracting image features at multiple scales and levels from the surveillance video images using an image encoder; extracting text features from the dynamic cue library using a text encoder; calculating the similarity between the image features and the text features; fusing the image features and the corresponding text features based on the calculated similarity to generate fused features; and determining the coordinates of the objects and the confidence level of the object category based on the fused features.

[0072] In this context, image features at multiple scales and levels refer to a multi-scale visual feature pyramid. For example, in a bottom-up path, the feature map size gradually decreases as the network deepens, while the semantic information becomes stronger. In a top-down path, the deepest and most abstract high-level features are upsampled, increasing their size. Lateral connections fuse shallow features of the same size from the bottom-up path with high-level features from the top-down path. Finally, each fused feature is used independently for prediction. Large-scale feature maps are suitable for detecting small objects, while small-scale feature maps are suitable for detecting large objects.

[0073] This embodiment uses mature YOLO-E or an open-set detection model with the same architecture, such as YOLO-World, as the core vision engine. This model overcomes the limitations of traditional closed-set detection, supporting real-time definition of the detection category via text prompts, eliminating the need to retrain the model for new categories. This embodiment can directly call pre-trained YOLO-E model weights, and its standard architecture includes the following core components: The image encoder, or backbone, employs an efficient CSPDarknet or similar backbone network to extract the aforementioned multi-scale visual feature maps from the input surveillance video frames. These maps are responsible for capturing texture, shape, and spatial structure information within the image. Specifically, the original surveillance video frames are downsampled and convolutionally processed by the encoder to output a multi-scale visual feature pyramid.

[0074] The text encoder, integrating a pre-trained language model such as CLIP Text Encoder, encodes the user-defined dynamic prompt library into high-dimensional text feature vectors. Specifically, the encoder segments and embeds the dynamic prompt word list from the library to generate a sequence of text feature vectors.

[0075] The reparameterized fusion module RepVL-PAN or Feature Fusion performs cross-modal interaction between the aforementioned visual features and textual features, for example, through attention mechanisms or dot product operations, to calculate a similarity heatmap between visual regions and textual descriptions. Specifically, the visual feature map and textual feature vectors are fused across modally to generate an enhanced feature map containing semantic information.

[0076] The decoupled detection head, based on the fused features described above, outputs in parallel the bounding box regression coordinates and the class confidence score for the item category. Since the category is dynamically determined by the text prompt, the detection head no longer outputs a fixed softmax classification probability, but instead outputs a similarity score matching the prompt word. Specifically, the enhanced feature map is decoupled through convolutional layers to obtain the output coordinates and confidence score (x, y, w, h, score), and non-maximum suppression is used to obtain the final detection result.

[0077] Furthermore, this embodiment pre-constructs a text-based dynamic cue library, PromptLibrary, containing frequently used dynamic objects in the hospital, such as: ["temporary trolley", "scattered medical record folder", "attendant chair", "isolation block", "unknown package", etc.]. Therefore, there is no need to collect a large amount of labeled data to train a new model; simply updating the text list is sufficient for the model to learn to recognize new objects.

[0078] As mentioned earlier, real-time monitoring video frames and the currently active text-based dynamic cue library are input into the aforementioned architecture. The video frames are fed into the image encoder, and the corresponding cue words from the dynamic cue library are input into the text encoder. The model automatically calculates the matching degree between the image regions and the cue words and directly outputs obstacle bounding boxes with category labels, which are derived from the dynamic cue library and confidence scores.

[0079] In addition, the object type includes unknown categories. This embodiment can identify unknown obstacles that exist in the dynamic prompt library mentioned above. Based on the above content, unrecognizable objects can be output as unknown categories.

[0080] When the confidence level of an object is too high, the area where the object is located can be considered an impassable area. Specifically, the method 100 also includes: if the confidence level of an object category is higher than a high threshold, it belongs to a high-confidence object category, and the location corresponding to the object category is marked as an impassable or high-cost area in the digital twin map. Impassable means that the path is blocked or the path is unreachable. Then, the dynamic cost coefficient corresponding to the object category is infinite. The dynamic cost coefficient is determined based on the object recognition result and the degree of occupancy determined by the location.

[0081] If a high-confidence obstacle is detected, such as one with a confidence level exceeding a high threshold of 90%, the area is immediately marked as impassable or high-cost. The identification results, such as location coordinates, type or category, and the occupied area determined by the location coordinates, are mapped in real time to an instantaneous access cost matrix of the digital twin map, with the corresponding areas marked as impassable or high-cost zones in this matrix.

[0082] It should be noted that areas marked as high-cost can also be considered as impassable, or their cost can be marked as infinite. A digital twin map is a 3D map model built based on a digital twin platform and the hospital's environmental data. Environmental data can include large-scale 3D point clouds and textures of the ground surface, quickly acquired through oblique photogrammetry and LiDAR scanning, forming a high-precision 3D real-world model. Importing traditional GIS vector data from the hospital's BIM model provides accurate geometric dimensions, component attributes, and spatial relationships. Real-time status information of people, objects, and the environment, such as temperature, is continuously collected through IoT sensors, cameras, GPS, etc.

[0083] Step 105: Based on the object recognition results and the object location, map the instant access cost matrix into a digital twin map.

[0084] As described above, the identification results, such as location, type or category, and the occupied area determined by the location coordinates, are mapped in real time to an instantaneous access cost matrix of the digital twin map. This matrix records the identification results.

[0085] Step 106: Based on the current number of patients in each department of the hospital, the number of appointments in the future time period, the surgical scheduling information, and the second preset model, determine the congestion coefficient of the target area in the hospital in the future time period.

[0086] Among them, the crowd congestion coefficient represents the crowd congestion situation in the future time period.

[0087] The second preset model refers to a time series prediction model, which is trained based on the number of patients at the current historical time for each department, the number of appointments adjacent to but later than the current historical time, historical surgical scheduling information adjacent to but later than the current historical time, and patient flow data adjacent to but later than the current historical time. For example, using hours or minutes as the dimension, if the current historical time is 16:00 on day A, then later and adjacent times could be 17:00 or 16:10 on day A, which will not be elaborated further.

[0088] The time series forecasting model architecture employs either the standard Prophet time series forecasting model or the aforementioned LSTM model. The standard Prophet time series forecasting model includes components such as a trend term, seasonal term, holiday effect, and error term. The trend term captures the overall upward, downward, or stationary trend of the series. The trend term can employ a piecewise linear trend, suitable for data with a clear linear trend. The trend term incorporates a change point detection function, automatically identifying points in the time series where the trend changes significantly and allowing the trend to adjust at these points, enabling the model to adapt to trend inflection. The seasonal term models periodic variations. Prophet uses Fourier series to construct the seasonal model, simultaneously modeling multiple periods, such as annual, weekly, and daily periods. Users can control the smoothing of seasonality by adjusting the order of the Fourier series. The holiday term incorporates the impact of known events occurring on specific dates. A list of holidays is required from the user. The model assigns separate parameters to each holiday to estimate the magnitude of the event's impact on the series. Prophet not only provides the predicted values, but also quantifies the error term of the prediction using the Markov chain Monte Carlo method or maximum a posteriori estimation, generating the prediction interval.

[0089] For Prophet, the input includes date, time, and observations. The model is optimized using a fitting algorithm to estimate all parameters for trend, seasonality, and holiday terms. Based on the fitted model, the estimated trend, seasonality, and holiday terms for future times are extrapolated and summed to obtain the final predicted value.

[0090] By introducing multi-source business data as exogenous regression terms, in the hospital scenario of this implementation, relying solely on historical patient flow data cannot accurately predict sudden changes in patient flow caused by surgery and examination scheduling. Therefore, the multi-source business data, such as the number of patients currently examined, the number of patients scheduled for future examinations, and the surgical scheduling plan in each department of the HIS system, are quantified as exogenous regression terms and integrated into the additive equation of the Prophet model: y(t)=g(t)+s(t)+h(t)+ + Error term ε. Here, X(t) is the business volume driving index for future time t extracted from multi-source business data, and β is its regression coefficient. The construction and alignment of exogenous variables include feature quantification, where the number of appointments for examinations and pre-scheduled surgeries in each department are uniformly converted into expected flow equivalents based on the average time taken for each surgery or examination type and the degree of population aggregation. For example, a major surgery at its start time can be converted into 3 standard flow units, including medical staff and waiting family members. Time alignment involves pre-writing known business data that has already been scheduled into the X(t) vector at the corresponding future time point when the model predicts future periods, achieving accurate injection of future knowledge. The concretization of model fitting and extrapolation involves using historical data from the past 3-6 months to train the model during the fitting phase. y represents the historical actual flow observation value, and X represents the business volume driving index converted from historical business data within the same time period. Prophet's L-BFGS optimizer simultaneously estimates all parameters of the trend term, seasonal term, holiday term, and exogenous regression coefficient β. The extrapolation stage involves summing the results to obtain the final predicted value. The extrapolation of the trend term g(t) extends the piecewise linear growth rate or logical growth trend line of the last segment to the future, based on the change points detected in the fitting stage. The extrapolation of the seasonal term s(t) utilizes the fitted Fourier series coefficients, naturally extrapolating future periodic fluctuations based on the mathematical properties of periodic functions. The extrapolation of the holiday term h(t) uses a pre-entered list of future holidays, such as statutory holidays and specific hospital academic weeks, combined with the fitted parameters for the impact magnitude of each holiday, to accurately generate the holiday effect offset for future dates. (Business exogenous terms...) The construction of this method involves transforming the known future surgical schedules and appointment numbers for examinations in the hospital information system into an X(t) matrix for future time points, and directly calculating the values ​​for this part. This differs from traditional extrapolation; this part is not speculative but directly incorporates known future business volume into the prediction. The final predicted values ​​are then summed, adding the four time-series components related to the future one-to-one along the same time axis. (t)=g_future(t)+s_future(t) +h_future(t)+ The final predicted curve for the crowd congestion coefficient is obtained. The adaptive expansion of the uncertainty interval fully considers the inherent uncertainty of the exogenous business data when generating the prediction interval using the Markov Chain Monte Carlo (MCMC) method. For example, a prior probability distribution is set for the outpatient appointment no-show rate. During MCMC sampling, each sampling not only samples the model parameters but also samples an actual number of appointments from the no-show distribution, calculating the prediction fluctuation under different actual values. This results in an expansion coefficient on the final prediction interval, providing a more robust and reliable estimate that better reflects the actual situation at the hospital. For example, predicting the crowd congestion in a certain examination area at 10:00 AM tomorrow: the trend component, based on the growth trend fitted from the past six months of crowd flow data, extrapolates and predicts a trend value of 120 people at that time. The seasonal component, based on the daily cycle and intraday hourly cycle characteristics, Fourier series extrapolation shows that this time is a regular low period, with a component of -20 people. The holiday component, after checking, is 0 because tomorrow is not a holiday. The business-driven component, extracted from the HIS system, shows 15 people scheduled for CT scans and 10 for ultrasound scans in this area at 10:00 AM tomorrow, as well as approximately 5 families waiting for an orthopedic surgery scheduled for the same time. Based on the model regression coefficient β, with a fitted value of 1.05 over the past month, this component is... People. Final predicted value = 120 - 20 + 0 + 31.5 = 131.5 people. This generates a real-time, dynamic, and plan-aware high-precision prediction of pedestrian congestion.

[0091] For example, the model can be trained using patient flow data from the hospital's historical data spanning the previous 3-6 months, with actual patient flow serving as the supervision label. Alternatively, multi-source business data, along with actual patient flow, can be used as training samples or observations to train the model, resulting in a trained model.

[0092] This includes real-time acquisition of multi-source business data, such as the number of patients currently examined, the number of patients scheduled for future examinations, and surgical schedules in each department of the HIS system. It may also include outpatient data, such as inpatient mobility records and outpatient registration data.

[0093] Instead of improving the prediction algorithm itself, the approach is to combine mature time-series prediction models with hospital business data to build hospital-specific spatiotemporal congestion prediction application scenarios.

[0094] Specifically, based on the current number of patients in each department of the hospital, the number of appointments in the future time period, surgical scheduling information, and the second preset model, the congestion coefficient of the target area in the hospital in the future time period is determined. This includes: extracting the time-series features corresponding to the current number of patients in each department, the number of appointments in the future time period, and surgical scheduling information; inputting the time-series features into the second time series prediction model to obtain the population density of the target area in the future time period; and determining the congestion coefficient of the target area in the future time period based on the population density.

[0095] For example, the model performs predictions by acquiring the aforementioned multi-source business data. It then cleans and fuses this data, extracts its temporal features, inputs them into a Prophet or LSTM model, and outputs a predicted pedestrian density value P. e (t), calculate the congestion coefficient ρ based on the predicted pedestrian flow density. e (t).

[0096] This embodiment utilizes a time-series prediction algorithm, employing mature Prophet or LSTM models to perform short-term predictions of pedestrian flow density in key or target areas of a hospital over the next 15-30 minutes, and maps this prediction to a pedestrian congestion coefficient. The formula for calculating the pedestrian congestion coefficient is as follows:

[0097] Among them, P e (t) represents the predicted pedestrian density, in person / square meter, Ce is the channel capacity threshold, and κ is the adjustment factor.

[0098] Meanwhile, the prediction results are pushed to the digital twin map in real time, providing future status information for path planning.

[0099] This embodiment integrates multimodal open-set visual detection and hospital business data at the perception and prediction layers. On one hand, it utilizes a semantic dynamic cue library to identify undefined temporary obstacles in real time. On the other hand, it collects real-time data on the number of patients examined in each department, waiting queues, and service rates to construct a spatiotemporal business flow prediction engine, quantifying and predicting the congestion dissipation time for specific road segments.

[0100] Step 107: Determine the travel path cost based on the pedestrian congestion coefficient and the real-time travel cost matrix, and determine the candidate path based on the travel path cost.

[0101] Specifically, the cost of the travel path is determined based on the congestion coefficient and the real-time travel cost matrix, and candidate paths are determined based on the travel path cost, including: determining the travel path cost function based on the congestion coefficient and the real-time travel cost matrix in the future time period; and determining candidate departure times and candidate delivery paths based on the travel path cost function and the path planning algorithm.

[0102] As mentioned above, the instantaneous passage cost matrix contains the above identification results, and the passage path cost function can be determined based on these identification results.

[0103] The process involves determining the travel path cost function based on the pedestrian congestion coefficient and the real-time travel cost matrix over a future time period. This includes: determining the degree of occupancy of the current area by the object's location; determining the dynamic cost coefficient based on the object's identification result and occupancy degree, where the identification result represents the object category; and determining the travel time of the target area based on the pedestrian congestion coefficient of the target area over a future time period, the corresponding dynamic cost coefficient, and the side length of the target area's passage segment, which serves as the travel path cost function.

[0104] The method involves determining the degree of occupancy of an object within its current area based on its location. This occupancy can be determined by the area or size of the region it occupies. Larger areas or sizes result in greater occupancy. A correlation is established between area or size and occupancy coefficients, and these correlations are used to determine the object's occupancy coefficient.

[0105] The dynamic cost coefficient δe(t) is determined based on the object's recognition result and occupancy level, including: determining it according to the object's category and occupancy level. The dynamic cost coefficient can be calculated using a weighted summation formula, assigning different weight coefficients to different categories and the same weight to corresponding occupancy levels, thus determining the final dynamic cost coefficient. For example, one weight is assigned to small-sized objects, one weight to medium-sized objects, and one weight to large-sized objects, thereby determining the final coefficient. Alternatively, the coefficient can be determined directly based solely on the occupancy level. The coefficient should fall within the following range: for example, coefficient = 0 when there are no obstacles, coefficient = 0.5-1.0 when there are temporary obstacles, and coefficient →∞ approaching infinity when there is severe obstruction. When the confidence level is too high, the coefficient can be directly set to infinity.

[0106] The path cost function is determined by the following formula:

[0107] Where Te(t) is the travel time of the channel segment in the target region, and Le is the side length of the channel segment. δ e ( t () represents the dynamic cost coefficient. ρ e ( t ) is the congestion coefficient for human traffic. α , β , γ The weighting coefficient is adjustable; the default is... α =1.0, β =0.5, γ =0.3. For normal tasks, increase... α Save energy for mission delivery; increase efficiency for urgent missions. γ To reduce delays.

[0108] Before the physical robot is deployed to the task, the server initiates a parallel intelligent route simulation in the digital twin environment. Based on the real-time perception cost δe(t) and the predicted pedestrian congestion coefficient ρe(t) from the previous steps, a multimodal cost decision is made using the aforementioned cost function. In this function, for each edge e of the passage segment in the digital twin map, its estimated travel time at time t is defined as Te(t). The edge length Le can be directly obtained from the digital twin map. Adjustable weight coefficients... α , β , γ This can be determined through backtracking simulations using historical delivery data. For example, for emergency supplies missions, γ can be increased to minimize delays. For regular clothing missions, α can be increased to improve the energy efficiency of the delivery robot.

[0109] Specifically, candidate departure times and candidate delivery routes are determined based on the travel path cost function and the path planning algorithm. This includes: searching multiple candidate delivery routes and their corresponding travel times based on the path planning algorithm, delivery origin, delivery destination, and travel path cost function; when congestion exists on a candidate delivery route, determining the total travel time cost for each candidate delivery route based on its corresponding travel time, waiting time, expected task time, and time penalty coefficient, where the time penalty coefficient is positively correlated with the urgency of the task; and selecting the final candidate delivery route, departure time, waiting time, and corresponding waiting position from the multiple candidate delivery routes based on the total travel time cost.

[0110] Among them, when there is congestion in the candidate delivery route, it can mean that the congestion coefficient is greater than or equal to the coefficient threshold, then it is determined as a congestion situation or congestion event.

[0111] The time penalty coefficient is positively correlated with the urgency of the task, meaning that the higher the urgency of the task, the greater the time penalty coefficient.

[0112] Based on the aforementioned travel cost path function, a time-dependent approach is adopted. The algorithm, or dynamic programming algorithm, searches for the optimal path from the delivery origin to the delivery destination. The algorithm also considers the possibility of waiting at nodes on the digital twin map to avoid future congestion peaks. The total cost is calculated, including travel time, waiting time, and task time window penalties. The total cost is determined by the following formula:

[0113] in, ω For the waiting time, TW start TW endThis represents the expected time window for the task, i.e., the expected time of the task. λ represents the penalty weight per unit of waiting time, with a default value of 0.8. Its function is to prevent the algorithm from choosing excessively long detours to avoid extremely short periods of congestion. The value of λ can be dynamically adjusted according to the urgency of the task; the higher the urgency, the larger λ, and the more the algorithm tends to depart immediately rather than wait in place. t end and t start T represents the start and arrival times of the candidate paths and their corresponding candidate times, respectively. ei (t) i The sum of the travel costs of all segments in the candidate path is represented by . The optimal path is determined based on the total time cost, and the path with the lowest total cost is selected as the final path. That is, if congestion exists, the waiting time can be determined after the congestion is cleared, and the departure time can be postponed. It should be understood that this waiting time can also be negative, meaning that by removing the departure point and avoiding congestion, the total cost can still be determined, which can also be used as a path selection method. Congestion clearing occurs when the congestion coefficient is less than the coefficient threshold. When it is finally determined whether to wait or advance the delivery time, the start and end times of the delivery also need to be adjusted to determine the final delivery time.

[0114] It should be understood that the above route planning can include routes that avoid congestion and detours, non-detour routes that start early to avoid congestion, normal routes that allow waiting for congestion to dissipate, or a combination of the above routes.

[0115] This embodiment's path planning algorithm incorporates a penalty for item timeliness decay. In a hospital setting, indefinitely waiting for congestion to dissipate is not feasible; the timeliness of items, such as the viability of test specimens and the sterilization expiration date of surgical instruments, must be considered. A maximum allowable delay time is preset for each item delivery task; for example, 60 minutes for routine specimens and 15 minutes for intraoperative urgent specimens. Cost adjustment is applied when calculating the total cost, taking the travel time cost T into account. ei (t) iBuilding upon this foundation, a time-sensitive penalty is introduced. An early arrival penalty is implemented, where early departure avoids congestion but may incur additional costs. For example, if the delivery robot arrives at the target department before surgery is finished, instruments cannot be received, and the robot enters an ineffective waiting area. A limit setting is implemented, assigning an earliest acceptable time to each task. This is typically defined as the surgery completion time minus a buffer period. A cost correction is applied, where robot resources used during the early arrival waiting period are counted as costs if the expected arrival time (tend) of a candidate path is earlier than the earliest acceptable time. Restrictions on in-hospital access areas and infection control zones are introduced. Unlike ordinary logistics operations, detours cannot be haphazard and must comply with infection control requirements. Spatial restrictions are imposed, marking nodes in the digital twin map as different area types, such as clean areas, contaminated areas, and public areas. An absolute prohibition is enforced, strictly forbidding pathways from directly crossing physical access control points without using buffer zones during route planning. The detour penalty allows for planning detours to avoid congestion, but choosing a long detour through a sterile area to avoid crowds in public corridors incurs a very high cost penalty, effectively prohibiting the operation. The revised decision cost calculation function, considering the above restrictions, changes the total cost formula for path selection from a simple time cost to a comprehensive cost including spatiotemporal penalties. The final cost function is: Total Cost = Walking Time Cost + Waiting Cost + Early Arrival Ineffective Occupancy Cost + Item Time-Decline Penalty + Infection Control Violation Detour Penalty. Dynamic Decision Logic: If the congestion coefficient indicates that the upcoming passage will exceed a threshold in the future, causing a sharp increase in total cost, the algorithm will perform the following comparison: Detour Option A: Calculate the alternative route to a longer but unobstructed clean corridor or external connecting corridor, assessing the time cost of the increased distance. Early Departure Option B: Extract the avoidance of congestion and calculate the lead time (tend) for the expected arrival time. Early arrival is represented by negative waiting time, but the ineffective waiting cost of early arrival is calculated. Option C (Wait in Place): Wait for congestion to dissipate, i.e., for the waiting coefficient to drop below the threshold. Calculate the sum of the pure waiting time cost and the time-delay penalty. Finally, compare the comprehensive costs of options A, B, and C, and forcibly select the departure time and path corresponding to the minimum total cost. If option B or C triggers the item's time-delay limit or violates infection control regulations, its total cost is directly set to infinity, thus automatically eliminating it from the selection process.

[0116] In this embodiment, at the decision-making and extrapolation layer, a multimodal dynamic passage cost function is used to weight and fuse physical distance, predicted waiting time, detour cost, task urgency, and multimodal risk coefficients. Multiple paths are extrapolated in parallel within a digital twin environment, automatically comparing the combined costs of waiting for passage versus detours, and outputting the globally optimal strategy.

[0117] Step 108: Perform simulations based on candidate routes to determine the optimal departure time and delivery route, and execute the item delivery task according to the delivery mode.

[0118] In a digital twin environment, path simulation is performed to determine the optimal departure time and delivery route, and the item delivery task is executed according to the delivery mode.

[0119] Specifically, simulations are performed based on candidate paths to determine the optimal departure time and delivery route. This includes: discretizing candidate paths into multiple state points according to preset time segments; simulating the position and state of the robot performing the delivery task in each time segment in a digital twin map based on dynamic cost coefficients and pedestrian congestion coefficients; detecting whether there are any object categories with infinite dynamic cost coefficients corresponding to the state points, resulting in uneliminable object categories, or whether there are any congestion events caused by pedestrian congestion coefficients exceeding the coefficient threshold; monitoring whether the placement space at the delivery destination is accessible; if all state points are passable and the placement space at the delivery destination is accessible, the simulation result of the candidate path is "passable"; otherwise, the simulation result is "path blocked" or "path unreachable".

[0120] For example, a virtual simulation of candidate paths can be performed to verify whether each edge of the path is unreachable at the predicted passage time. Candidate paths can be discretized into multiple state points in time segments, such as discretizing the path based on 1 minute. In a digital twin environment, based on the real-time perception cost δe(t) and the predicted pedestrian congestion coefficient ρe(t) from the above steps, the robot's position state in each time segment is simulated. Each state point is checked for unavoidable obstacles, such as infinite δe(t) or congestion events. The accessibility of the destination placement space is checked; for example, the operating room buffer can be monitored in real-time via a visual perception module using acquired video to determine if it is full and accessible. If all state points are passable, the simulation passes; if blockages are detected, or the path is impassable, an early warning mechanism can be triggered.

[0121] The early warning mechanism can be implemented as follows: Specifically, the method 100 further includes: when the simulation result of the candidate path is that the path is blocked or the path is unreachable, then according to the identification result that caused the blockage or the path is unreachable and the corresponding future crowd congestion coefficient, calling AI voice call to notify relevant personnel; suspending the issuance of the item delivery task and putting the item delivery task into the waiting queue; repeatedly executing the step of obtaining the monitoring video image until the simulation result is passed, which is used to indicate that it is reachable, resuming the issuance of the item delivery task, and executing the steps of determining the optimal departure time and delivery route, and executing the item delivery task according to the delivery mode.

[0122] As mentioned above, if the simulation detects a blockage or inaccessibility, an early warning mechanism is automatically triggered. An AI voice call is made to notify the nearest logistics, security, or operational personnel, reporting the specific type of blockage. For example, if the identification results from the above steps and the future pedestrian congestion coefficient determine that a temporary cart is obstructing the corridor in Area A, and congestion is expected to worsen in 15 minutes, please clear it immediately. The server suspends task issuance and places tasks in a waiting queue. The environmental perception agent is periodically invoked again for visual re-inspection. After confirming that the spatial adjustments are complete, a virtual simulation is performed again to ensure the path is clear before proceeding, thus ensuring that the simulation is truly feasible and accessible.

[0123] This embodiment establishes a closed loop in the execution and self-healing layers: virtual verification, obstruction warning, intelligent intervention, and visual re-inspection. If the simulation determines that the path is unreachable, it automatically calls relevant personnel to clear the obstacle based on the obstacle's semantics and sets an intelligent wake-up timer based on the predicted time. Physical commands are only issued to execute the task after both the virtual environment and visual re-inspection confirm that the path is passable.

[0124] During the execution of tasks, such as delivery tasks, robots can avoid obstacles in the following ways.

[0125] When the robot's sensors detect an obstruction ahead, the robot's controller enters the decision-making phase. These sensors, including LiDAR or depth cameras, output raw physical data such as point cloud coordinates, pixel brightness, and depth values. The sensors input this raw physical data into a visual model, which extracts features from the data and compares them with text information from a dynamic cue library to ultimately provide semantic labels.

[0126] The semantic labels include both objects and crowds. Objects can include labels such as trolleys, wheelchairs, and IV stands. When the semantic label type is "object," the system enters anchor point locking mode. When the semantic label type is "crowd," the system enters dynamic perturbation mode. Simultaneously, utilizing the depth information from the digital twin map, the discrete point cloud collected by the sensors is fitted to a minimum convex hull using the Quickhull algorithm. For trolleys, the convex hull solves the problem of incomplete point clouds; for crowds, it solves the problem of blurred edges.

[0127] For example, as mentioned above, the sensor collects a set of depth points. The model identifies these points as conforming to the geometric features and texture of a cart, and then labels them with the corresponding cart tag. This model can be the mature YOLO-E or an open-set detection model or vision engine with the same architecture, as described above. It can also be other mature image recognition models; there are no restrictions here.

[0128] Because the objects seen by sensors, including LiDAR or depth cameras, are not perfect, regular shapes, but rather scattered, fragmented point clouds, and because the surface materials of objects can reflect light or obstruct it (e.g., the material of a cart can reflect light or obstruct it), the outline of the object is fragmented in the digital twin map. This can create gaps within the object, causing the robot to attempt to pass through these gaps and collide with it. For example, a spurious sampling point might emerge from the gap between the cart's handle and the cart body. This embodiment automatically connects these fragmented point clouds into a continuous, closed geometry, ensuring that the robot does not attempt to pass through the cart due to perception errors. The controller receives the obstacle information with semantic labels. Based on the labels, it determines the current mode. As mentioned earlier, if the label is a crowd, the controller switches the obstacle avoidance algorithm to dynamic perturbation mode; if it is an object, it switches to anchor point locking mode.

[0129] This embodiment simplifies an object into a finite set of polygonal faces using an anchor-point locking mode. The controller can quickly calculate the unit normal of each face facing the empty side of the path. Regardless of the shape and form of the obstacle, the controller avoids obstacles by using the space pointed to by the unit normal. Crowds are handled in a similar way.

[0130] For objects parked at an angle, the pattern in this embodiment can accurately depict their deflection angle. This can free up 20%-30% more passable area in hospital corridors. The specific implementation method is as follows: As mentioned above, the controller can acquire point cloud data sets Praw, semantic labels Lsem, and ground elevation reference Z from the digital twin map. ground And the depth information matrix D(u,v) from the depth camera.

[0131] The controller extracts the projection points of the original point cloud dataset Praw onto the XY plane of the digital twin map, and uses the Quickhull algorithm to fit the minimum convex hull Sobs, which represent the area of ​​the enclosed closed polygon. For each sampling grid (x,y) within the convex hull Sobs, its effective height h(x,y) in the vertical direction is obtained.

[0132]

[0133] Where zi represents the data values ​​of all point clouds above the coordinates of the sampling raster, and max is the maximum height value.

[0134] The dynamic blocking volume Vactive is determined by the following formula:

[0135] In the discretized system where the controller actually executes, the formula for Vactive becomes:

[0136] λ is the volume correction factor; when Lsem represents an object, such as a cart, λ = 1.1; when Lsem represents a crowd, λ = 1.5. N is the number of data values ​​in the point cloud, and ΔA is the unit grid area of ​​the digital twin map. Using the results from Vactive, a virtual entity blocking pillar is generated in the digital twin map, including pillars with a Sobs base and a height of [missing information]. The virtual entity blocking pillar is used. The Dynamic Spatiotemporal Pose Sampling (DSTPS) algorithm corresponding to this mode automatically determines the boundary values ​​of the virtual entity blocking pillar. Furthermore, the robot's Rrobot area within the Vactive's interior and its edge normal direction is designated as a sampling no-go zone. This ultimately generates a spatial elastic anchor matrix (SEAM) on the safe normal side of the Vactive's edge, guiding the robot to perform precise avoidance maneuvers.

[0137] By converting the point cloud obtained from sensors into Vactive mode through λ correction, a dynamic defensive boundary was successfully established for the robot. This boundary is fitted to objects such as carts, and looser for crowds. This addresses the issues of overly aggressive or insufficient obstacle avoidance in medical delivery.

[0138] Specifically, the controller initiates sampling along the normal direction of the convex hull envelope of the Vactive. The sampling constraint is a uniform random distribution with hard constraints. A spatial elastic anchor point matrix (SEAM) is generated around the Vactive using the robot's physical radius superimposed with a dynamic safety margin as the minimum spacing threshold. This ensures that each anchor point in the matrix represents a physically legal pose that the robot can traverse.

[0139] The controller extracts the optimal pose from SEAM based on the pattern and makes path decisions using the following anchor point Pi cost function:

[0140] Where Pnow is the current pose vector, which refers to the robot's real-time coordinates (x, y) and heading angle θ in the digital twin map; Dist represents the distance; Jitter(nenv) is the normal cost jitter parameter, which is determined by the variance of the point cloud coordinate fluctuation at the convex hull edge within a short period, such as 500ms, reflecting the looseness or turbulence of the crowd boundary; α and β are weighting coefficients, where α is the dynamic gain and β is the obstacle avoidance redundancy gain.

[0141] For anchor-point locking modes, such as those for objects like trolleys or wheelchairs, the object's center of mass and motion vector are calculated. Using the direction of the normal n to its convex hull, three equally spaced fixed anchor points Pfixed are selected or locked from the SEAM. The robot then cuts straight in and locks onto the accompanying object using the anchor point with the lowest Cost value obtained from the above formula. This state continues until the object's convex hull moves out of the detection boundary, thus avoiding the obstacle.

[0142] For dynamic perturbation modes, such as those affecting clustered crowds, the controller refreshes the volume gradient of the crowd's convex hull at a fixed frequency, such as 10Hz. V. The jitter parameter Jitter(nenv) based on the normal cost is used to handle the real-time expansion and contraction of crowd edges. A velocity command Vcmd, synthesized from a weighted displacement vector, is output. The robot is driven to move smoothly and cautiously along the dynamic ridges formed by the SEAM pose clusters.

[0143] The controller continuously monitors the overlap between the original path region and the Vactive convex hull. Once the overlap disappears, indicating an open path, the controller issues a tangential regression command starting from the current anchor point, smoothly returning to the main task path.

[0144] To generate a smooth vector field, the aforementioned cost (Cost) needs to be transformed into the aforementioned probability weights (wi), and normalized using a Softmax structure. wi is determined by the following formula:

[0145] Here, the anchor points are from j to k, thus the lower the cost of the anchor point, the greater its contribution to the robot's final motion trend. The formula for the velocity vector field command Vcmd is as follows:

[0146] Where vmax is the robot's maximum permissible speed, and anchors i to k are defined. The above-mentioned di is the maximum velocity vector pointing from the i-th anchor point to that anchor point. Traditional navigation output is a single target point. If the target point is occupied during robot movement, an abrupt stop will occur. However, this embodiment outputs Vcmd. The robot is affected by the combined action of all high-value anchors in SEAM. Even if a gap disappears, the direction of the resultant force vector will automatically and smoothly shift without triggering a logic interruption. When the convex hull of the crowd contracts, i.e., a gap appears, the Jitter(nenv) value of the corresponding anchor point decreases, and its weight wi increases instantaneously, thereby guiding the robot to accelerate and cut into the gap. Conversely, when the convex hull expands, it guides the robot to decelerate and avoid the gap.

[0147] For example, as mentioned earlier, the controller continuously acquires the latest Jitter(n_env) parameters from the perception layer at a frequency of 10Hz-20Hz. Within a 120-degree sector in front of the robot's direction of motion, a local SEAM subset containing k dynamic anchor points is constructed and maintained in real time. Using the aforementioned cost function and weight wi formula, the globally optimal motion vector Vcmd guiding the robot's motion at the current moment is calculated. The calculated resultant velocity vector Vcmd is decomposed and sent to the linear velocity v and angular velocity ω controllers of the robot chassis for execution. Throughout the avoidance process, the robot exhibits a continuous fine-tuning state with small steps and high frequency. If the path is completely blocked, the weights of all anchor points approach the mean and their directions contradict each other, the resultant force Vcmd will approach zero, and the robot will stop naturally and smoothly.

[0148] For example, in the anchor-locked mode, as mentioned earlier, the geometric center of the convex hull (Sobs) is obtained as the centroid (xc, yc), and the object's displacement vector is calculated. The normal n is extracted from the edge of the convex hull on the side overlapping with the robot path. The baseline safety distance Dsafe = 1.2 * robot radius is calculated. The center anchor point P2 = (xc, yc) + Anchor points P1 and P3 on both sides lie on tangents perpendicular to the normal n and are symmetrically arranged at a preset step size Δs distance, for example, the robot radius. {P1, P2, P3} is encapsulated as a static SEAM and assigned extremely high stability weights. The selected optimal fixed anchor point coordinates are output and used for anchor point locking and triggering the State_wait command.

[0149] For example, for dynamic perturbation modes, the above volume gradient is calculated. V is used to determine whether the population is currently in a state of relaxation or contraction. Volume gradient. V is determined by the following formula:

[0150] Within the non-blocking fan-shaped region between the robot's current position Pnow and the dynamic blocking volume Vactive, the above sampling is initiated. That is, the distance between any two sampling points must satisfy dist(si, sj) ≥ robot radius Rrobot + Δr, where Δr is the dynamic safety margin. The cost values ​​for each sampling point are calculated, and the weights are normalized using the Softmax method. Based on the weight wi of each anchor point, the resultant force vector Vcmd guiding the robot's movement is generated and output in real time. Vcmd(v, ω), i.e., the resultant force vector command of linear velocity and angular velocity, is also output, along with the sampling refresh rate, which is recommended to be 10Hz - 20Hz.

[0151] At the same time, it can also combine the real-time maneuverability coefficient of small spaces such as operating rooms to dynamically splice continuous tasks of delivery and empty box retrieval.

[0152] Specifically, the method 100 also includes: obtaining the department's maneuvering space coefficient based on the surveillance video image; obtaining the on-duty personnel coefficient; when the maneuvering space coefficient is greater than the threshold and / or the on-duty personnel coefficient is 1, reducing the passage time of the corresponding passage segment in the corresponding target area, so that the corresponding passage segment is prioritized in the candidate path, so as to complete the item retrieval task when passing through the corresponding passage segment.

[0153] For confined areas such as operating rooms, dynamic rule weights are preset to obtain the department's maneuvering space coefficient in real time. This can be determined by monitoring the available area ratio of buffer zones through a visual perception module, and by obtaining the on-duty personnel coefficient, for example, through a scheduling system. When the maneuvering space coefficient is less than 0.3 or the on-duty personnel coefficient is 0 (meaning no one is on duty), the cost weight of path planning for that area is automatically increased (e.g., by 1.5-2.0 times), increasing time costs. Otherwise, the weight can be decreased to reduce time costs. This prioritizes the corresponding passageway segment in the candidate path, allowing the item retrieval task to be completed when passing through that passageway segment. Furthermore, by combining the aforementioned pedestrian flow prediction information, time window offset optimization can be applied to tasks in congested periods and areas, advancing or delaying them by 5-15 minutes to avoid peak travel times.

[0154] As mentioned earlier, the splicing of multiple tasks can include optimizing the splicing of consecutive tasks by time windows, such as splicing the tasks of delivery and empty box retrieval, calculating the optimal task execution order, and maximizing robot utilization.

[0155] Specifically, the method 100 further includes: when the expected times of multiple item delivery tasks are in a continuous time series, splicing the multiple item delivery tasks together; based on at least one delivery origin and multiple delivery destinations in the splicing result, for the multiple delivery tasks, performing the step of determining candidate paths based on the cost of the travel path, until the optimal departure time and delivery path are determined; wherein, some delivery destinations of multiple delivery destinations are used as intermediate travel points in the candidate paths, and when at least one delivery origin has multiple delivery origins, some delivery origins are used as intermediate travel points.

[0156] When the expected times of multiple tasks are in a continuous time sequence, and this continuous time is triggered, the multiple tasks can be merged and concatenated. Based on all the starting and ending points of the multiple tasks, path planning is performed as described above. At least one starting point serves as the starting point for multiple tasks, and one ending point serves as the ending point for multiple tasks. Other starting and ending points can serve as intermediate points along the planned path, thus achieving the goal of completing the delivery of all multiple tasks according to the optimal planned path, rather than taking detours or back-and-forth routes. Furthermore, multiple tasks can have the same starting point but different ending points, or the same ending point but different starting points.

[0157] This embodiment also supports a rescheduling mechanism for task cancellation or adjustment. Once the environment or requirements change, a rescheduling is immediately triggered, which will not be elaborated further.

[0158] This embodiment implements task execution and closed-loop feedback, sending verified task instructions to the robot task scheduling platform. During robot execution, an environmental perception agent continuously monitors the environment, forming a closed loop of perception-decision-execution-feedback. The robot reports real-time status information such as current location, task progress, and remaining battery power, providing specific feedback. When encountering new obstacles or sudden congestion during execution, it immediately reports to the scheduling center, providing abnormal event feedback. After task completion, it reports completion time, actual path, and time consumed, providing task completion feedback. The actual execution data, such as actual travel time and actual congestion, is compared with the predicted data to form feedback data for subsequent model parameter calibration. Without retraining the model, only parameter fine-tuning is performed, achieving model optimization feedback. Through these feedback mechanisms, prediction accuracy and scheduling strategies can be continuously optimized, achieving adaptive operation.

[0159] Example 1: Operating room scenario in a top-tier hospital. The system detects that high-value consumables in an operating room are below the minimum requirement and automatically generates an order. The multimodal open set perception module discovers that the corridor leading to the operating room is occupied by temporary equipment, indicated by a text message: "Temporary cart." The spatiotemporal prediction module, based on HIS data, predicts that the corridor will become increasingly congested in 15 minutes due to peak outpatient traffic. The parallel simulation layer calculates that the current route requires a detour and may delay surgery; therefore, the simulated path is blocked. The system immediately calls the logistics supervisor: "There is temporary obstruction in corridor A. Peak traffic is expected in this area in 15 minutes. Priority clearing is recommended." Five minutes later, visual re-inspection confirms the passage is clear, the simulation passes, the robot departs, and after delivering the goods, it picks up the empty box from the previous surgery along the way. Example 2: Multi-task concurrent scenario. The system collects the needs of three adjacent departments and automatically matches them using a ride-sharing order mode. During parallel simulations, the system predicted that congestion would occur at the entrance of a certain department in 15 minutes due to the peak of outpatient examinations. Therefore, the departure time of the robot was slightly adjusted to be brought forward, or the stopping order was optimized to ensure that congestion was avoided. Finally, the optimal stopping order and time window were simulated in the digital twin to avoid duplicate paths and delays.

[0160] like Figure 3 As shown, this application also proposes a hospital logistics task delivery device, applied to service equipment, the device 300 comprising: The acquisition module 301 is used to acquire the current surgical schedule information of different departments and the remaining quantity of the items.

[0161] The first generation module 302 is used to determine whether the items need to be replenished based on the first preset model, the current surgical schedule information, and the remaining quantity of the items. When it is determined that replenishment is needed, an item delivery task is generated to complete the replenishment of the items.

[0162] The first determining module 303 is used to put the item delivery task into a dynamic pool and determine the delivery mode according to the delivery time and delivery location corresponding to the item delivery task.

[0163] The second determining module 304 is used to acquire surveillance video images, identify objects in the surveillance video images according to a dynamic prompt library, and determine the position of objects in the surveillance video images. The dynamic prompt library is used to record text information of objects in the hospital environment.

[0164] The second generation module 305 is used to generate an instant access cost matrix for a digital twin map based on the object recognition results and the object position mapping.

[0165] The third determining module 306 is used to determine the crowd congestion coefficient of a target area in the hospital in a future time period based on the current number of people in each department of the hospital, the number of appointments in a future time period, surgical scheduling information, and the second preset model. The crowd congestion coefficient represents the crowd congestion situation in the future time period.

[0166] The fourth determining module 307 is used to determine the passage path cost and the corresponding instant passage cost matrix based on the pedestrian congestion coefficient and the instant passage cost matrix, and to determine the candidate path based on the instant passage cost matrix.

[0167] The execution module 308 is used to perform simulations based on the candidate paths, determine the optimal departure time and delivery route, and execute the item delivery task according to the delivery mode.

[0168] In addition, the device 300 also includes: a calling module, used to call an AI voice call to notify relevant personnel when the simulation result of the candidate path is blocked or the path is unreachable, based on the identification result of the blockage or the unreachable path and the corresponding future crowd congestion coefficient; a pause module, used to pause the issuance of the item delivery task and put the item delivery task into a waiting queue; and a repeat module, used to repeatedly execute the step of obtaining the monitoring video image until the simulation result is passed, indicating that the path is reachable, resume the issuance of the item delivery task, and execute the steps of determining the optimal departure time and delivery route, and executing the item delivery task according to the delivery mode.

[0169] Specifically, the first generation module 302 includes: a parsing unit, used to parse the current surgical schedule information and convert the unstructured surgical schedule information into a quantified surgical schedule feature vector; and an input unit, used to input the quantified surgical schedule feature vector into the first time series prediction model and output the future consumption quantity of the corresponding department. In addition, the acquisition module 301 is also used to acquire the historical consumption rate of the items, the corresponding historical surgical schedule information, and the historical consumption quantity of the items. The historical consumption rate of an item corresponds to a time earlier than the historical consumption quantity of adjacent items, and the historical surgical schedule information corresponds to a time earlier than the historical consumption quantity of adjacent items. The device 300 includes: a third generation module, used to acquire the fluctuation value within a preset time period in the historical consumption rate sequence according to a sliding window and generate a corresponding consumption rate feature vector; a training module, used to train an initial time series prediction model to obtain a first time series prediction model based on the consumption rate feature vector corresponding to the historical consumption rate sequence, the quantified feature vector corresponding to the historical surgical schedule information, and the historical consumption quantity of the items; and a fifth determination module, used to determine whether the items need to be replenished based on the future consumption quantity and the remaining quantity of the items.

[0170] The parsing unit is specifically used to map the corresponding surgical type to a consumable consumption coefficient that characterizes the severity of item consumption, based on the correspondence between different surgical types and consumable consumption coefficients; to construct a high-demand probability distribution curve for a future preset time period based on the start time and expected duration of the corresponding surgical type, where the high-demand probability distribution curve refers to the high-demand probability distribution curve where the demand exceeds a threshold; to statistically analyze the consumption habits of different surgeons for different items based on the historical consumption quantity of items; and to generate a quantitative surgical scheduling feature vector based on the consumable consumption coefficient corresponding to the surgical type, the demand probability distribution curve, and the surgeon's consumption habits.

[0171] In addition, the acquisition module 301 is also used to acquire task instructions; the device 300 also includes: an extraction module, used to extract the intent and item-related information from the task instructions; and a first generation module 302, used to generate an item delivery task based on the intent and item-related information to complete the delivery of the item.

[0172] Specifically, the first determining module 303 includes: a first determining unit, used to determine the delivery mode of the item delivery task based on the delivery time and location corresponding to the item delivery task, and a genetic algorithm or ant colony optimization algorithm. The delivery modes include point-to-point direct delivery mode, ride-sharing mode, round-trip empty box pickup mode, and continuous connection mode.

[0173] Specifically, the second determining module 304 includes: a first extraction unit, used to extract image features of multiple different scales and multiple different levels from the surveillance video image through an image encoder; and to extract text features from the text information in the dynamic prompt library through a text encoder; a fusion unit, used to calculate the similarity between the image features and the text features, and to fuse the image features and the corresponding text features according to the calculated similarity to generate fused features; and a second determining unit, used to determine the coordinates of the object and the confidence level of the object category based on the fused features.

[0174] In addition, the device 300 also includes a marking module, which is used to mark the location corresponding to the object category as an inaccessible or high-cost area in the digital twin map if the confidence level of the object category is higher than a high threshold. If the object category belongs to a high-confidence object category, the module marks the location corresponding to the object category as an inaccessible or high-cost area. If the object category is inaccessible, the dynamic cost coefficient corresponding to the object category is infinite. The dynamic cost coefficient is determined based on the object recognition result and the degree of occupancy determined by the location.

[0175] Specifically, the third determining module 306 includes: a second extraction unit, used to extract the time-series features corresponding to the current number of people in each department, the number of appointments in the future time period, and the surgical scheduling information; an input unit, used to input the time-series features into the second time series prediction model to obtain the population density of the target area in the future time period; and a third determining unit, used to determine the population congestion coefficient of the target area in the future time period based on the population density; wherein, the second time series prediction model is trained based on the current historical number of people in each department in the historical time period, the number of appointments adjacent to the current historical time but later than the current historical time, the historical surgical scheduling information adjacent to the current historical time but later than the current historical time, and the population data adjacent to the current historical time but later than the current historical time.

[0176] Specifically, the fourth determining module 307 includes: a fourth determining unit, used to determine the travel path cost function based on the passenger flow congestion coefficient and the real-time travel cost matrix within a future time period; and to determine the candidate departure time and candidate delivery path based on the travel path cost function and the path planning algorithm.

[0177] The fourth determining unit is used to determine the degree of occupancy of the current area by the object based on the object's position; to determine a dynamic cost coefficient based on the object's identification result and the degree of occupancy, wherein the identification result represents the object category; and to determine the passage time of the target area based on the pedestrian congestion coefficient of the target area in the future time period, the corresponding dynamic cost coefficient, and the side length of the target area channel segment, as a passage path cost function.

[0178] The path cost function is determined by the following formula:

[0179] Where Te(t) is the travel time of the channel segment in the target region, and Le is the side length of the channel segment. δ e ( t () represents the dynamic cost coefficient. ρ e ( t ) is the congestion coefficient for human traffic. α , β , γ These are adjustable weighting coefficients; For regular tasks, by increasing... α Save energy for mission delivery; increase efficiency for urgent missions. γ To reduce delays.

[0180] Specifically, the fourth determining unit is used to search for multiple candidate delivery routes and their corresponding travel times based on the route planning algorithm, delivery origin, delivery destination, and travel path cost function; when congestion exists in the candidate delivery routes, for the multiple candidate delivery routes, the total travel time cost of the corresponding candidate delivery routes is determined based on the corresponding travel time, waiting time, expected task time, and time penalty coefficient, where the time penalty coefficient is positively correlated with the urgency of the task; and the final candidate delivery route, departure time, waiting time, and corresponding waiting position are selected from the multiple candidate delivery routes based on the total travel time cost.

[0181] Specifically, the execution module 308 includes: a discretization unit, used to discretize the candidate path into multiple state points according to a preset time segment; a simulation unit, used to simulate the position state of the robot performing the delivery task in each time segment in the digital twin map according to the dynamic cost coefficient and the pedestrian congestion coefficient; a detection unit, used to detect whether there is an object category corresponding to the state point with an infinite dynamic cost coefficient, resulting in an uneliminable object category, or whether there is a congestion event caused by the pedestrian congestion coefficient being greater than the coefficient threshold; a monitoring unit, used to monitor whether the placement space of the delivery destination of the delivery task is accessible; and a deduction unit, used to determine whether the deduction result of the candidate path is passable if all the state points are passable and the placement space of the delivery destination is accessible, otherwise the deduction result is path blockage or path unreachable.

[0182] In addition, the device 300 also includes: a splicing module, used to splice multiple item delivery tasks when the expected time of multiple item delivery tasks is in a continuous time sequence; based on at least one delivery starting point and multiple delivery endpoints in the splicing result, for the multiple delivery tasks, the step of determining candidate paths based on the travel path cost in the fourth determination module 307 is executed until the optimal departure time and delivery path are determined; wherein, some of the multiple delivery endpoints are used as intermediate travel points in the candidate paths, and when the at least one delivery starting point has multiple delivery starting points, some delivery starting points are used as intermediate travel points.

[0183] In addition, the acquisition module 301 is also used to acquire the department's maneuvering space coefficient and the on-duty personnel coefficient based on the monitoring video images. The device 300 also includes a reduction module, which is used to reduce the passage time of the corresponding channel segment in the corresponding target area when the maneuvering space coefficient is greater than a threshold and / or the on-duty personnel coefficient is 1, so that the corresponding channel segment is given priority in the candidate path, so as to complete the item retrieval task when passing through the corresponding channel segment.

[0184] This will not be elaborated upon further here; for any content that has not been covered in detail, please refer to the content mentioned above.

[0185] like Figure 4 As shown, this application also proposes a device 400, including: a memory 401 and a processor 402; the memory 401 is used to store a computer program; the processor 402 is used to execute the computer program for: Obtain the current surgical schedule information for different departments and the remaining quantity of the items; Based on the first preset model, the current surgical schedule information, and the remaining quantity of the items, it is determined whether the items need to be replenished. If it is determined that replenishment is needed, an item delivery task is generated to complete the replenishment of the items. The item delivery tasks are placed in a dynamic pool, and the delivery mode is determined based on the delivery time and delivery location corresponding to the item delivery tasks. Acquire surveillance video images, identify objects in the surveillance video images according to a dynamic prompt library, and determine the position of objects in the surveillance video images. The dynamic prompt library is used to record text information of objects in the hospital environment. Based on the object recognition results and the object location mapping to a digital twin map, an initial real-time access cost matrix is ​​generated. Based on the current number of patients in each department of the hospital, the number of appointments in the future time period, the surgical scheduling information, and the second preset model, the congestion coefficient of the target area in the hospital in the future time period is determined. The congestion coefficient represents the congestion situation in the future time period. The travel path cost is determined based on the pedestrian congestion coefficient and the real-time travel cost matrix, and candidate paths are determined based on the travel path cost. The optimal departure time and delivery route are determined by simulation based on the candidate routes, and the item delivery task is executed according to the delivery mode.

[0186] This will not be elaborated upon further here; for any content that has not been covered in detail, please refer to the content mentioned above.

[0187] This application also proposes a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by one or more processors, it causes the one or more processors to implement the steps in method 100.

[0188] This will not be elaborated upon further here; for any content that has not been covered in detail, please refer to the content mentioned above.

[0189] Furthermore, in some of the processes described in the above embodiments and accompanying drawings, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 201, 202, etc., are merely used to distinguish different operations and do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0190] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0191] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of a necessary general-purpose hardware platform, or by a combination of hardware and software. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a computer product. The present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable multimedia data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable multimedia data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0193] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable multimedia data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions may also be loaded onto a computer or other programmable multimedia data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0195] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0196] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0197] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for delivering logistics tasks in a hospital, characterized in that, include: Obtain current surgical schedule information and remaining quantities of supplies for different departments; Based on the first preset model, the current surgical schedule information, and the remaining quantity of the items, it is determined whether the items need to be replenished. If it is determined that replenishment is needed, an item delivery task is generated to complete the replenishment of the items. The item delivery tasks are placed in a dynamic pool, and the delivery mode is determined based on the delivery time and delivery location corresponding to the item delivery tasks. Acquire surveillance video images, identify objects in the surveillance video images according to a dynamic prompt library, and determine the position of objects in the surveillance video images. The dynamic prompt library is used to record text information of objects in the hospital environment. Based on the object recognition results and the object location mapping, an instant access cost matrix is ​​generated for the digital twin map; Based on the current number of patients in each department of the hospital, the number of appointments in the future time period, the surgical scheduling information, and the second preset model, the congestion coefficient of the target area in the hospital in the future time period is determined. The congestion coefficient represents the congestion situation in the future time period. The travel path cost is determined based on the pedestrian congestion coefficient and the real-time travel cost matrix, and candidate paths are determined based on the travel path cost. The optimal departure time and delivery route are determined by simulation based on the candidate routes, and the item delivery task is executed according to the delivery mode. If the simulation of the candidate path fails, an AI voice call is made to notify relevant personnel, the factors that caused the failure are eliminated, and the step of obtaining the surveillance video image is repeated.

2. The method according to claim 1, characterized in that, When the simulation of the candidate path fails, the step of invoking an AI voice call to notify relevant personnel, eliminating factors that led to the failure, and repeating the step of acquiring surveillance video images includes: If the simulation result of the candidate path is that the path is blocked or the path is unreachable, then based on the identification result that caused the blockage or the unreachable path and the corresponding future crowd congestion coefficient, an AI voice call is made to notify the relevant personnel. Pause the issuance of the item delivery task and place the item delivery task into the waiting queue; Repeat the steps of acquiring surveillance video images until the simulation result is passed, indicating that the item is reachable. Then, resume issuing the item delivery task and execute the steps of determining the optimal departure time and delivery route, and executing the item delivery task according to the delivery mode.

3. The method according to claim 1, characterized in that, The step of determining whether to replenish the items based on the current surgical schedule information and the remaining quantity of the items in the first preset model includes: The current surgical schedule information is analyzed, and the unstructured surgical schedule information is transformed into a quantified surgical schedule feature vector. The quantified surgical scheduling feature vector is input into the first time series prediction model, which outputs the future consumption quantity of items for the corresponding department. The training process of the first time series prediction model includes: acquiring the historical consumption rate of an item, the corresponding historical surgical schedule information, and the historical consumption quantity of the item, wherein the historical consumption rate of the item corresponds to a time earlier than and adjacent to the historical consumption quantity of the item, and the historical surgical schedule information corresponds to a time earlier than and adjacent to the historical consumption quantity of the item; acquiring the fluctuation value within a preset time period in the historical consumption rate sequence according to a sliding window, and generating a corresponding consumption rate feature vector; training an initial time series prediction model based on the consumption rate feature vector corresponding to the historical consumption rate sequence, the quantization feature vector corresponding to the historical surgical schedule information, and the historical consumption quantity of the item to obtain the first time series prediction model; Based on the future consumption of resources and the remaining quantity of the items, determine whether it is necessary to replenish the items.

4. The method according to claim 3, characterized in that, The process of converting unstructured surgical scheduling information into quantified surgical scheduling feature vectors includes: Based on the correspondence between different surgical types and consumable consumption coefficients, the corresponding surgical types are mapped to consumable consumption coefficients that characterize the severity of item consumption. Based on the start time and expected duration of the corresponding surgical type, construct a high demand probability distribution curve for the future preset time period. The high demand probability distribution curve refers to the high demand probability distribution curve where the demand is greater than the threshold. Based on the historical consumption statistics of items, different surgeons have different consumption habits for different items; A quantitative surgical scheduling feature vector is generated based on the consumable consumption coefficient, demand probability distribution curve, and the surgeon's consumption habits corresponding to the surgical type.

5. The method according to claim 1, characterized in that, The method further includes: Get task instructions; Extract the intent and item-related information from the task instructions; Based on the stated intent and relevant item information, an item delivery task is generated to complete the item delivery.

6. The method according to claim 1, characterized in that, The step of determining the delivery mode based on the delivery time and location corresponding to the item delivery task includes: The delivery mode of the item delivery task is determined based on the delivery time and location corresponding to the item delivery task, as well as the genetic algorithm or ant colony optimization algorithm; the delivery mode includes point-to-point direct delivery mode, ride-sharing mode, round-trip empty box pickup mode, and continuous connection mode.

7. The method according to claim 1 or 2, characterized in that, The step of acquiring surveillance video images and identifying and determining the positions of objects in the surveillance video images based on a dynamic prompting library includes: The image encoder extracts image features at multiple scales and at multiple levels from the surveillance video images; The text features of the text information in the dynamic prompt library are extracted using a text encoder; The image features and the text features are similar to each other. Based on the calculated similarity, the image features and the corresponding text features are fused to generate fused features. Based on the fusion features, the coordinates of the object and the confidence level of the object category are determined.

8. The method according to claim 7, characterized in that, The method further includes: If the confidence level of the object category is higher than the high threshold, it belongs to the high-confidence object category. The location corresponding to the object category is marked as an impassable or high-cost area in the digital twin map. The impassable means that the path is blocked or the path is unreachable. The dynamic cost coefficient corresponding to the object category is infinite. The dynamic cost coefficient is determined based on the object recognition result and the degree of occupancy determined by the location.

9. The method according to claim 1, characterized in that, The method involves determining the congestion coefficient of a target area within the hospital in a future time period based on the current number of patients in each department, the number of appointments scheduled for a future time period, surgical scheduling information, and a second preset model. This includes: Extract the temporal features corresponding to the current number of patients in each department, the number of appointments in the future time period, and the surgical scheduling information; The time-series features are input into the second time-series prediction model to obtain the pedestrian density of the target area in the future time period. The pedestrian congestion coefficient of the target area in the future time period is determined based on the pedestrian flow density. The second time series prediction model is trained based on the number of people in each department at the current historical time, the number of appointments adjacent to and later than the current historical time, the historical surgical scheduling information adjacent to and later than the current historical time, and the flow of people adjacent to and later than the current historical time.

10. The method according to claim 1, characterized in that, The step of determining the travel path cost based on the pedestrian congestion coefficient and the real-time travel cost matrix, and determining candidate paths based on the travel path cost, includes: The travel path cost function is determined based on the pedestrian congestion coefficient in the future time period and the real-time travel cost matrix. Candidate departure times and candidate delivery routes are determined based on the travel path cost function and the path planning algorithm.

11. The method according to claim 10, characterized in that, The step of determining the travel path cost function based on the congestion coefficient of pedestrian flow in the future time period and the real-time travel cost matrix includes: The extent to which the object occupies the current area is determined based on the object's position. A dynamic cost coefficient is determined based on the object identification result and the occupancy level, wherein the identification result represents the object category; Based on the congestion coefficient of the target area in the future time period, the corresponding dynamic cost coefficient, and the side length of the channel segment in the target area, the travel time of the target area is determined as the travel path cost function.

12. The method according to claim 11, characterized in that, The travel path cost function is determined by the following formula: Where Te(t) is the travel time of the channel segment in the target region, and Le is the side length of the channel segment. δ e ( t () represents the dynamic cost coefficient. ρ e ( t ) is the congestion coefficient for human traffic. α , β , γ These are adjustable weighting coefficients; For regular tasks, by increasing... α Save energy for mission delivery; increase efficiency for urgent missions. γ To reduce delays.

13. The method according to claim 10, characterized in that, The step of determining candidate departure times and candidate delivery routes based on the travel path cost function and the path planning algorithm includes: Based on the route planning algorithm, delivery origin, delivery destination, and travel path cost function, multiple candidate delivery routes and their corresponding travel times are searched. When there is congestion in the candidate delivery route, the total travel time cost of the corresponding candidate delivery route is determined based on the corresponding travel time, waiting time, expected task time and time penalty coefficient. The time penalty coefficient is positively correlated with the urgency of the task. The final candidate delivery route, departure time, waiting time, and corresponding waiting position are selected from multiple candidate delivery routes based on the total travel time cost.

14. The method according to claim 1, characterized in that, The step of simulating and extrapolating based on the candidate paths to determine the optimal departure time and delivery route includes: Based on a preset time segment, the candidate path is discretized into multiple state points; In the digital twin map, the location status of the robot performing the delivery task is simulated in each time segment based on the dynamic cost coefficient and the crowd congestion coefficient; The system detects whether there is an object category whose dynamic cost coefficient is infinitely large, resulting in an object category that cannot be eliminated, or whether there is a congestion event caused by the pedestrian congestion coefficient being greater than the coefficient threshold. Monitor whether the placement space at the delivery destination of the delivery task is accessible; If all the status points are passable and the delivery destination is accessible, the candidate path is considered to be passable; otherwise, the path is considered to be blocked or unreachable.

15. The method according to claim 1, characterized in that, The method further includes: When the expected times of multiple item delivery tasks are in a continuous time sequence, the multiple item delivery tasks are spliced ​​together; Based on at least one delivery origin and multiple delivery destinations in the splicing results, for the multiple delivery tasks, the step of determining candidate routes based on the travel path cost is performed until the optimal departure time and delivery route are determined; Among them, some of the multiple delivery endpoints are used as intermediate transit points in the candidate path. When the at least one delivery origin has multiple delivery origins, some of the delivery origins are used as intermediate transit points.

16. The method according to any one of claims 1 or 10-13, characterized in that, The method further includes: Based on the surveillance video images, obtain the department's maneuverability coefficient; Obtain the on-duty personnel attendance coefficient; When the maneuvering space coefficient is greater than the threshold and / or the on-duty personnel coefficient is 1, the passage time of the corresponding channel segment in the corresponding target area is reduced so that the corresponding channel segment is prioritized in the candidate path so that the item retrieval task can be completed when passing through the corresponding channel segment.

17. A hospital logistics task delivery device, characterized in that, include: The acquisition module is used to acquire the current surgical schedule information of different departments and the remaining quantity of the items; The first generation module is used to determine whether the items need to be replenished based on the first preset model, the current surgical schedule information, and the remaining quantity of the items. When it is determined that replenishment is needed, an item delivery task is generated to complete the replenishment of the items. The first determining module is used to put the item delivery task into a dynamic pool and determine the delivery mode according to the delivery time and delivery location corresponding to the item delivery task. The second determining module is used to acquire surveillance video images, identify objects in the surveillance video images according to a dynamic prompting library, and determine the position of objects in the surveillance video images. The dynamic prompting library is used to record text information of objects in the hospital environment. The second generation module is used to generate an instant access cost matrix for a digital twin map based on the object recognition results and the object's location. The third determining module is used to determine the crowd congestion coefficient of a target area in the hospital in a future time period based on the current number of people in each department of the hospital, the number of appointments in a future time period, surgical scheduling information, and the second preset model. The crowd congestion coefficient represents the crowd congestion situation in a future time period. The fourth determining module is used to determine the travel path cost based on the pedestrian congestion coefficient and the real-time travel cost matrix, and to determine the candidate path based on the travel path cost. The execution module is used to perform simulations based on the candidate paths, determine the optimal departure time and delivery route, and execute the item delivery task according to the delivery mode.

18. A device, characterized in that, include: Memory, processor; The memory is used to store computer programs; The processor is configured to execute the computer program for: Obtain the current surgical schedule information for different departments and the remaining quantity of the items; Based on the first preset model, the current surgical schedule information, and the remaining quantity of the items, it is determined whether the items need to be replenished. If it is determined that replenishment is needed, an item delivery task is generated to complete the replenishment of the items. The item delivery tasks are placed in a dynamic pool, and the delivery mode is determined based on the delivery time and delivery location corresponding to the item delivery tasks. Acquire surveillance video images, identify objects in the surveillance video images according to a dynamic prompt library, and determine the position of objects in the surveillance video images. The dynamic prompt library is used to record text information of objects in the hospital environment. Based on the object recognition results and the object location mapping to a digital twin map, an initial real-time access cost matrix is ​​generated. Based on the current number of patients in each department of the hospital, the number of appointments in the future time period, the surgical scheduling information, and the second preset model, the congestion coefficient of the target area in the hospital in the future time period is determined. The congestion coefficient represents the congestion situation in the future time period. The travel path cost is determined based on the pedestrian congestion coefficient and the real-time travel cost matrix, and candidate paths are determined based on the travel path cost. The optimal departure time and delivery route are determined by simulation based on the candidate routes, and the item delivery task is executed according to the delivery mode.

19. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by one or more processors, it causes the one or more processors to perform the steps of the method according to claims 1-16.

Images