Hospital logistics zero inventory intelligent scheduling method and system based on AI

Abstract

The invention discloses an AI-based hospital logistics zero inventory intelligent scheduling method and system, and relates to the technical field of intelligent medical treatment and resource optimization, and the method comprises the steps: collecting multi-dimensional data to generate a multi-dimensional data set, carrying out the feature extraction through a GNN, dynamically adjusting the edge weight through an RL, carrying out the updating through a Q-learning formula, and predicting the material demand and carbon emission. Synthesizing the 3D model and the map into a virtual environment, loading a scheduling instruction set and a hospital logistics real-time state data set, generating a digital twin environment, constructing a DRL model to predict arrival time, operating the digital twin environment to update a state vector, adjusting the predicted arrival time, and updating the scheduling instruction set. According to the invention, accurate data is provided for zero inventory, the response speed of scheduling is improved, the adaptability of the system to complex scenes is enhanced, the high efficiency of scheduling execution is ensured, the stability and intelligent level of hospital logistics are improved, and the high-efficiency zero inventory management of hospital logistics is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent medical care and resource optimization, and particularly to an AI-based intelligent zero-inventory scheduling method and system for hospital logistics. Background Art

[0002] With the rapid development of artificial intelligence technology, hospital logistics management has gradually transformed from traditional manual scheduling to an intelligent and data-driven direction. Prediction models based on machine learning and GNNs have shown significant potential in the field of medical resource optimization, achieving accurate prediction of material requirements through feature extraction of multi-dimensional data. As an emerging means, digital twin technology provides a new perspective for real-time status monitoring and decision support by combining the mapping of the physical world and the virtual environment. In the field of hospital logistics, related technologies have initially realized material demand prediction, distribution scheduling, and environmental impact assessment.

[0003] However, there are still deficiencies in existing hospital logistics management. Existing prediction methods based on machine learning mostly rely on static data sets, resulting in prediction results that cannot timely reflect sudden demand changes. When using RL methods for optimization scheduling, only a single target is focused on, and the adjustment of edge weights lacks self-adaptability. Most scheduling systems mainly process two-dimensional data and fail to effectively integrate three-dimensional spatial information and digital twin environments, unable to accurately predict arrival times and dynamically adjust the instruction set. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an AI-based intelligent zero-inventory scheduling method and system for hospital logistics, which solves the problems that existing prediction methods based on machine learning mostly rely on static data sets, resulting in prediction results that cannot timely reflect sudden demand changes, RL methods only focus on a single target when optimizing scheduling, the adjustment of edge weights lacks self-adaptability, most scheduling systems mainly process two-dimensional data, fail to effectively integrate three-dimensional spatial information and digital twin environments, and are unable to accurately predict arrival times and dynamically adjust the instruction set.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an AI-based intelligent zero-inventory scheduling method for hospital logistics, which includes collecting multi-dimensional data to generate a multi-dimensional data set, defining department, material, and environmental characteristics, defining nodes and edges for the characteristics, generating a graph data format, using a GNN for feature extraction, dynamically adjusting edge weights through RL, using the Q-learning formula for updating, predicting material demand and carbon emissions; generating a zero-inventory scheduling plan through a DQN network, and integrating it into a scheduling instruction set; synthesizing a 3D model and a map into a virtual environment, loading the scheduling instruction set and the hospital logistics real-time status data set, generating a digital twin environment, constructing a DRL model to predict the arrival time, running the digital twin environment to update the state vector, adjusting the estimated arrival time, and triggering an optimization mechanism when the inventory material is less than the zero-inventory critical value to update the scheduling instruction set; executing the scheduling instruction set, and storing the multi-modal data generated by collection and analysis.

[0008] As a preferred solution of the AI-based intelligent zero-inventory scheduling method for hospital logistics according to the present invention, wherein: the collecting of multi-dimensional data to generate a multi-dimensional data set and predicting material demand and carbon emissions means using intelligent sensors to collect multi-dimensional data, preprocessing the multi-dimensional data to generate a multi-dimensional data set, and segmenting it into time period windows according to time series through an overlapping sliding window;

[0009] The multi-dimensional data includes the material number and the remaining amount of the material, the medical staff demand instruction, the image, the device power during standby and operation of the device;

[0010] Using the APL interface to obtain the hospital department configuration table, the amount and delivery time of materials that can be delivered by suppliers, the delivery path parameters, and the material demand times from the hospital information system;

[0011] Using Z-score normalization for the material status data and the carbon emission data, using binary coding for the sudden demand signal, combining the material status data and the sudden demand signal to generate department characteristics, defining the remaining amount of the material as material characteristics, and defining the carbon emission as environmental characteristics;

[0012] Using the feature splicing method to integrate the department characteristics, material characteristics, and environmental characteristics into an initial feature set, extracting the features in the initial feature set, and defining them as nodes, assigning IDs to each node to generate a dynamic node set;

[0013] Vertically stacking the initial feature set in the order of node IDs to form a feature matrix;

[0014] If a department consumes materials, a directed edge is established between the department node and the material node. If two departments have a demand for the same material, an undirected edge is established between the two department nodes. If carbon emissions are generated during material transportation, a directed edge is established between the material node and the carbon emission node, and the weights of the three edges are calculated;

[0015] Integrate the three edges and their corresponding weights to form a dynamic edge set, and use an adjacency matrix to integrate the dynamic node set and the dynamic edge set into a graph structure, perform sparsification processing, and store it in a graph data format;

[0016] Construct a GNN model, and define the action space as the operation of adjusting the increase and decrease of all edge weights;

[0017] Use fixed-step discretization to define a discrete action set for each edge, calculate the feature variance of each node through the analysis of variance method, sort them in descending order of variance, use the k-selection method to select the nodes with the largest variance, extract the associated edges of the nodes, and output the discrete action set of the associated edges;

[0018] Discretize the state vector through the K-means state clustering method to generate a discrete state set, and generate action indices for the discrete action set through the dynamic combination technology method;

[0019] Set the discrete state set as the number of rows and the action index as the number of columns to construct a Q-table, initialize the value of the actions executed in all states of the Q-table to 0, use the ϵ-greedy strategy to select the initial action for the current state, adjust the edge weights according to the initial action, update the adjacency matrix, and generate an optimized feature matrix through the second-layer convolution operation;

[0020] Perform global pooling on the optimized feature matrix to generate a global feature vector, output the predicted material demand and carbon emissions per hour through a two-layer fully connected layer, obtain the material demand error and carbon emission error through the loss function, and calculate the RL reward value according to the two errors;

[0021] Update using the Q-learning formula through the RL reward value, output the updated Q-table, feedback it to the RL module, adjust the ϵ-greedy strategy, update the adjacency matrix, perform the next round of GNN prediction, set the maximum number of iterations using the rule of thumb, and stop the iteration when the maximum iteration is reached;

[0022] Collect historical data through sensors, integrate it into a historical graph data format, input it into the GNN model for training, and optimize the model parameters using the loss function and the Adam optimizer;

[0023] Input the graph data format into the GNN model to predict the material demand and carbon emissions, and convert it into a prediction data table through structured data mapping.

[0024] As a preferred solution of the AI-based zero-inventory intelligent scheduling method for hospital logistics described in the present invention, wherein: generating a zero-inventory scheduling plan through the DQN network and integrating it into a scheduling instruction set means converting the prediction data table into a JSON format through structured data serialization;

[0025] Integrate the demand quantity, time window, remaining material quantity, and the quantity and delivery time of materials that can be delivered by suppliers in the prediction data table into a supply-demand status vector through the direct splicing method;

[0026] Define the actions as accepting supplier deliveries and rejecting supplier deliveries;

[0027] Define the DQN network, output the Q values of the defined actions through the hidden layer, and determine the actions through the ϵ-greedy strategy;

[0028] Calculate the time difference between the end time of the time window and the delivery time of the supplier, which is defined as the shipping time. Use the path planning API call method to obtain the optimal delivery path, and calculate the arrival time and delivery cost;

[0029] Merge the shipping time, optimal path, arrival time, delivery cost, and carbon emissions into a scheduling plan, and convert the scheduling plan into a scheduling instruction set through XML formatting;

[0030] As a preferred solution of the AI-based hospital logistics zero-inventory intelligent scheduling method described in the present invention, wherein: synthesizing the 3D model and the map into a virtual environment and updating the scheduling instruction set means that the digital twin system accepts the scheduling instruction set;

[0031] Extract the remaining material quantity, material consumption speed, and delivery path parameters, and generate a hospital logistics real-time status data set through ELT;

[0032] Use voxel gridding to construct a 3D model of the material storage point, use the rapidly-exploring random tree to generate a 3D virtual map of the delivery path, and synthesize a virtual environment based on the 3D model and the 3D virtual map;

[0033] Load the scheduling instruction set and the hospital logistics real-time status data set into the virtual environment to generate a digital twin environment. Set the simulation duration of the digital twin environment through real-time granularity simulation, and start real-time simulation;

[0034] Integrate the remaining material quantity, delivery path parameters, and remaining shipping time to generate an initial state vector;

[0035] The remaining shipping time is obtained by calculating the shipping time and the current time;

[0036] Construct a DRL model, extract the delivery path parameters, and calculate and output the estimated arrival time;

[0037] Run the digital twin environment to the next time point, update the current remaining material quantity according to the consumption speed, obtain the remaining time according to the current time and the shipping time, update the state vector, input it into the DRL model, adjust the estimated arrival time, and use the simple time difference calculation method to calculate the time deviation between the adjusted estimated arrival time and the arrival time;

[0038] Extract the remaining quantity of current supplies through the digital twin environment, combine the time deviation, set the critical value of zero inventory through the inventory critical threshold comparison method, compare the remaining quantity of current supplies, and if the remaining quantity of supplies is less than the zero inventory critical value, trigger the DRL model to optimize the current estimated arrival time and update the scheduling instruction set.

[0039] As a preferred solution of the AI-based hospital logistics zero-inventory intelligent scheduling method described in the present invention, wherein: the generation of the multi-dimensional dataset includes:

[0040] Calculate the material consumption speed, and integrate the material number and the remaining quantity of materials into material status data;

[0041] Detect the number of patients entering the area per minute and the equipment usage status in the image through the YOLOv5 model, calculate the equipment usage rate through the usage rate percentage formula, calculate the sudden demand index, and integrate it with the medical staff demand instruction to form a sudden demand signal;

[0042] Calculate the power consumption and carbon emissions through the equipment power, and integrate them into environmental data;

[0043] Synchronize the time of the material status data, sudden demand signal, and environmental data through the time stamp, and integrate them into a multi-dimensional dataset.

[0044] As a preferred solution of the AI-based hospital logistics zero-inventory intelligent scheduling method described in the present invention, wherein: the execution of the scheduling instruction set refers to sending the scheduling instruction set to the scheduling system of the supplier through communication for the execution of the scheduling instruction.

[0045] As a preferred solution of the AI-based hospital logistics zero-inventory intelligent scheduling method described in the present invention, wherein: the storage of the multi-modal data generated by collection and analysis refers to recording and storing the collected multi-modal data and the scheduling instruction set generated by analysis in the database. The database marks the time stamp for the stored data and uploads the stored data to the cloud for backup.

[0046] In a second aspect, the present invention provides an AI-based hospital logistics zero-inventory intelligent scheduling system, including,

[0047] A collection and prediction module, which is used to collect multi-dimensional data to generate a multi-dimensional dataset, define department, material, and environmental characteristics, define nodes and edges for the characteristics, generate a graph data format, use GNN for feature extraction, dynamically adjust the edge weights through RL, and use the Q-learning formula for update to predict the material demand and carbon emissions;

[0048] A scheduling integration module, which is used to generate a zero-inventory scheduling plan through the DQN network and integrate it into a scheduling instruction set;

[0049] The verification and update module is used to synthesize a virtual environment by combining a 3D model and a map, load a scheduling instruction set and a real-time hospital logistics status data set, generate a digital twin environment, construct a DRL model to predict the arrival time, run the digital twin environment to update the state vector, adjust the predicted arrival time, trigger an optimization mechanism when the inventory is less than the zero-inventory critical value, and update the scheduling instruction set;

[0050] The execution and storage module is used to execute the scheduling instruction set and store the multi-modal data generated by collection, analysis.

[0051] In a third aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the AI-based zero-inventory intelligent scheduling method for hospital logistics as described in the first aspect of the present invention is implemented.

[0052] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of the AI-based zero-inventory intelligent scheduling method for hospital logistics as described in the first aspect of the present invention is implemented.

[0053] The beneficial effects of the present invention are as follows: The present invention collects multi-dimensional data to generate a multi-dimensional data set, uses GNN for feature extraction, dynamically adjusts the edge weights through RL, and uses the Q-learning formula for updating to predict the material demand and carbon emissions; combines a 3D model and a map to synthesize a virtual environment, loads a scheduling instruction set and a real-time hospital logistics status data set, generates a digital twin environment, constructs a DRL model to predict the arrival time, runs the digital twin environment to update the state vector, adjusts the predicted arrival time, and updates the scheduling instruction set. The present invention provides accurate data for zero inventory, improves the response speed of scheduling, enhances the adaptability of the system to complex scenarios, ensures the efficiency of scheduling execution, improves the stability and intelligent level of hospital logistics, and realizes the efficient zero-inventory management of hospital logistics. Description of the Drawings

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0055] Figure 1 It is a flowchart of the AI-based zero-inventory intelligent scheduling method for hospital logistics in Embodiment 1.

[0056] Figure 2It is a structural diagram of the AI-based zero-inventory intelligent scheduling system for hospital logistics in Embodiment 1.

[0057] Figure 3 It is a flowchart for generating a zero-inventory scheduling plan through a DQN network and integrating it into a scheduling instruction set in Embodiment 1.

[0058] Figure 4 It is a flowchart for combining a digital twin environment with a DRL model in Embodiment 1. Specific implementation manners

[0059] To make the above objects, features, and advantages of the present invention more obvious and understandable, the specific implementation manners of the present invention will be described in detail below with reference to the accompanying drawings of the specification.

[0060] In the following description, many specific details are set forth to facilitate a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0061] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0062] Embodiment 1, referring to Figures 1 to 4 , is the first embodiment of the present invention. This embodiment provides an AI-based zero-inventory intelligent scheduling method for hospital logistics, including the following steps:

[0063] S1. Collect multi-dimensional data to generate a multi-dimensional data set, define department, material, and environmental characteristics, define nodes and edges for the characteristics, generate a graph data format, use a GNN for feature extraction, dynamically adjust the edge weights through RL, use the Q-learning formula for update, and predict the material demand and carbon emissions;

[0064] Specifically, collecting multi-dimensional data to generate a multi-dimensional data set means using intelligent sensors to collect multi-dimensional data, preprocessing the multi-dimensional data, and generating a multi-dimensional data set;

[0065] The intelligent sensors include RFID sensors based on the LoRaWAN protocol, voice recognition modules, high-definition cameras, and smart meters;

[0066] The multi-dimensional data includes the material number and remaining material quantity, medical staff demand instructions, images, equipment standby, and equipment power during operation;

[0067] Calculate the material consumption rate, with the formula:

[0068] ,

[0069] where is the material consumption rate, and are the initial quantity and the current quantity of the material respectively, is the time interval between the initial quantity and the current quantity of the material;

[0070] Integrate the material number, the remaining quantity of the material, and the material consumption rate into material status data;

[0071] Detect the number of patients entering the area per minute and the equipment usage status in the image through the YOLOv5 model, calculate the equipment utilization rate through the utilization rate percentage formula, and obtain the sudden demand index through calculation. The formula is:

[0072] ,

[0073] where is the sudden demand index, and are the weights of the number of patients and the equipment utilization rate respectively (determined through historical data regression analysis), and are the number of patients and the equipment utilization rate respectively;

[0074] Use the feature splicing method to splice the sudden demand index with the medical staff demand instruction to generate a sudden demand signal;

[0075] Calculate the power consumption and carbon emissions through the equipment power. The formula is:

[0076] , ,

[0077] where is the equipment power, is the time interval to obtain the power consumption, is the carbon emission factor (dynamically updated through an external database), is the carbon emissions, is the power consumption;

[0078] Use the data integration method to integrate the power consumption and carbon emissions into environmental data;

[0079] Synchronize the time of the material status data, the sudden demand signal, and the environmental data through timestamps and integrate them into a multi-dimensional data set.

[0080] By using intelligent sensors to collect multi-dimensional data, generating a multi-dimensional data set, enhancing the data dimension and real-time performance, which is superior to existing static modeling. By generating status data, providing accurate input for GNN feature extraction, ensuring that the prediction is close to real-time requirements. By combining the YOLOv5 model with the medical staff's demand instructions, it is superior to the traditional response mechanism that relies on manual judgment, ensuring timeliness under zero inventory.

[0081] Furthermore, define department, material, and environmental characteristics, define nodes and edges for the characteristics, generate a graph data format, use GNN for feature extraction, dynamically adjust the edge weights through RL, and use the Q-learning formula for updating to predict material demand and carbon emissions;

[0082] Divide the multi-dimensional data set into time period windows in time series through overlapping sliding windows;

[0083] Use the APL interface to obtain the hospital department configuration table, the amount of materials that suppliers can deliver, delivery time, delivery path parameters, and the number of material demands from the hospital information system;

[0084] Use Z-score normalization for material status data and carbon emission data, use binary coding for sudden demand signals, combine material status data and sudden demand signals to generate department characteristics, define the remaining amount of materials as material characteristics, and define carbon emissions as environmental characteristics;

[0085] Use the feature splicing method to integrate department characteristics, material characteristics, and environmental characteristics into an initial feature set;

[0086] Extract the features in the initial feature set and define them as nodes, assign IDs to each node, and generate a dynamic node set;

[0087] Stack the initial feature set vertically in the order of node IDs to form a feature matrix;

[0088] If a department consumes materials, a directed edge is established between the department node and the material node. If two departments have a demand for the same material, an undirected edge is established between the two department nodes;

[0089] Calculate the weights of the two edges respectively, and the formula is:

[0090] , ,

[0091] Among them, is the weight of the directed edge, is the weight of the undirected edge, is the consumption amount of the department node for the material node, is the number of times the two department nodes compete for materials, is the maximum consumption rate of the material in all departments, is the total number of demand times for competitive materials and are represented as department nodes and material nodes and are two department nodes

[0092] If carbon emissions are generated during material transportation, a directed edge is established between the material node and the carbon emission node, and the weight of the directed edge is calculated. The formula is:

[0093] ,

[0094] where is the weight of the directed edge between the material node and the carbon emission node is the standardized carbon emission value

[0095] Integrate the three edges and the corresponding weights to form a dynamic edge set

[0096] Use the adjacency matrix to integrate the dynamic node set and the dynamic edge set into a graph structure, perform sparsification processing, and store it in the graph data format

[0097] The graph data format includes an adjacency matrix and a feature matrix

[0098] Construct a GNN model, including an input layer, two hidden layers, two fully connected layers, and an output layer

[0099] Define the input layer as the graph data format. Through the first-layer graph convolution operation, calculate the initial feature matrix. The formula is:

[0100] ,

[0101] where is the initial feature matrix of the first layer is the ReLU activation function is the adjacency matrix is the input feature matrix is the weight matrix of the first layer is the bias vector of the first layer

[0102] The input initial feature matrix is flattened into a state vector through the matrix

[0103] Define a discrete action set for each edge using fixed-step discretization. Calculate the feature variance of each node through analysis of variance and sort it in descending order of variance. The formula is:

[0104] ,

[0105] where is the The characteristic variance of the nodes, is the eigenvector of the th node, is the eigenvalue of the th node at the th dimension, is the average value of the eigenvector of the th node, and n is the total number of dimensions of the th node;

[0106] Use the k-selection method to select the node with the largest variance, extract the associated edges of the node, and output the discrete action set of the associated edges;

[0107] Discretize the state vector through the K-means state clustering method to generate a discrete state set, and generate action indices from the discrete action set through the dynamic combination technology method;

[0108] Set the discrete state set as the number of rows and the action index as the number of columns to construct the Q-table;

[0109] Initialize the value of the action executed in all states of the Q-table to 0, and use the ϵ-greedy strategy to select the initial action for the current state;

[0110] Adjust the edge weights according to the initial action and update the adjacency matrix;

[0111] Generate an optimized feature matrix through the second-layer convolution operation;

[0112] Perform global pooling on the optimized feature matrix to generate a global feature vector, output the predicted hourly material demand and carbon emissions through a two-layer fully connected layer, obtain the material demand error and carbon emission error through the loss function, and calculate the RL reward value based on the two errors. The formula is:

[0113] ,

[0114] ,

[0115] where, is the material demand error, is the carbon emission error, and are the true material demand and carbon emissions, and are the predicted material demand and carbon emissions, and are the sample numbers of the material demand and carbon emissions;

[0116] ,

[0117] where, The RL reward value at time t;

[0118] Using the RL reward value, update the Q-value using the Q-learning formula, output the updated Q-table, feedback it to the RL module, adjust the ϵ-greedy policy, update the adjacency matrix, perform the next round of GNN prediction, set the maximum number of iterations using the rule of thumb, and stop the iteration when the maximum iteration is reached;

[0119] Collect the historical data of material status data, carbon emission data, sudden demand signals, and hospital department configuration tables through sensors, integrate them into the historical graph data format, input them into the GNN model for training, and optimize the model parameters using the loss function and Adam optimizer;

[0120] Input the graph data format into the GNN model to predict the material demand and carbon emissions, and convert the predicted material demand and carbon emissions into a prediction data table through structured data mapping transformation.

[0121] By splitting the multi-dimensional data set through overlapping sliding windows, reduce the missed detection of sudden demands, provide real-time input for the dynamic graph structure, ensure immediate response under zero inventory, reduce the scheduling conflict rate by integrating the initial feature set using the feature splicing method and generating dynamic node sets and edge sets, provide multi-dimensional input for the GNN, support the joint optimization of materials and the environment, screen key edges through feature variance and dynamically adjust the edge weights, enhance the sensitivity of the GNN to key relationships, improve the prediction accuracy, update the Q-table through Q-learning and iteratively optimize the prediction, realize the comprehensive prediction of materials and carbon emissions, and support green zero-inventory scheduling.

[0122] S2. Generate a zero-inventory scheduling plan through the DQN network and integrate it into a scheduling instruction set;

[0123] Specifically, convert the prediction data table into JSON format through structured data serialization;

[0124] Integrate the demand and time window, remaining material quantity, and the quantity and delivery time of materials that can be delivered by suppliers in the prediction data table into a supply and demand status vector through direct splicing;

[0125] Define the actions as accepting supplier deliveries and rejecting supplier deliveries;

[0126] Parse each element of the supply and demand status vector and generate a normalized status vector through normalization processing;

[0127] Define the DQN network, including 4 neurons in the input layer, a hidden layer, and 2 neurons in the output layer;

[0128] Among them, the 4 neurons in the input layer are the normalized status vector, and the 2 neurons in the output layer are the defined actions;

[0129] Collect historical data on demand, time window, remaining supplies, amount of supplies that suppliers can deliver, and delivery time in the prediction data table through sensors, generate a historical normalized state vector, train the DQN network, and optimize the model parameters using the Adam optimizer and loss function;

[0130] Input the normalized state vector into the trained DQN network, output the Q value defining the action through the hidden layer, and determine the action through the ε-greedy strategy;

[0131] Calculate the time difference between the end time of the time window and the delivery time of the supplier, defined as the shipping time, use the path planning API call method to obtain the optimal delivery path, and calculate the arrival time and delivery cost;

[0132] Combine the shipping time, optimal path, arrival time, delivery cost, and carbon emissions into a scheduling plan;

[0133] Convert the scheduling plan into a scheduling instruction set through XML formatting.

[0134] By converting the prediction data table into JSON format through structured data serialization, solve the problems of inconsistent data formats and low parsing efficiency in the existing technology. Generate a normalized state vector through normalization processing to ensure input consistency. Train the DQN network to dynamically optimize action selection and ensure decision robustness under zero-inventory conditions. Generate a scheduling plan by using the path planning API call method to optimize green scheduling and reduce environmental impact.

[0135] S3. Synthesize a virtual environment from the 3D model and the map, load the scheduling instruction set and the hospital logistics real-time status data set to generate a digital twin environment, build a DRL model to predict the arrival time, run the digital twin environment to update the state vector, adjust the predicted arrival time, and trigger the optimization mechanism when the inventory supplies are less than the zero-inventory critical value to update the scheduling instruction set;

[0136] The digital twin system receives the scheduling instruction set;

[0137] Extract the remaining supplies, material consumption speed, and delivery path parameters, and generate the hospital logistics real-time status data set through ELT;

[0138] Use voxel gridding to build a 3D model of the material storage point, and use the rapidly-exploring random tree to generate a 3D virtual map of the delivery path;

[0139] Based on the 3D model and the 3D virtual map, synthesize a virtual environment;

[0140] Load the scheduling instruction set and the hospital logistics real-time status data set into the virtual environment to generate a digital twin environment;

[0141] Set the simulation duration of the digital twin environment through real-time granularity simulation and start the real-time simulation;

[0142] Integrate the remaining material quantity, distribution path parameters, and remaining delivery time to generate an initial state vector;

[0143] The remaining delivery time is obtained by calculating the delivery time and the current time;

[0144] Construct a DRL model, including an input layer, a convolutional layer, a flattening layer, a fully connected layer, and an output layer;

[0145] Collect historical data and integrate it to generate a historical state vector, and train the DRL model;

[0146] The historical data includes the remaining material quantity, the material consumption rate, the distribution path parameters, and the remaining delivery time;

[0147] Input the initial state vector into the trained DRL model, extract the distribution path parameters, and calculate the estimated arrival time through the following formula:

[0148] ,

[0149] where, is the estimated arrival time, is the delivery time, and is the distribution distance, and is the vehicle speed of the delivery party, is the adjustment time;

[0150] Run the digital twin environment to the next time point, update the current remaining material quantity according to the consumption rate, obtain the remaining time according to the current time and the delivery time, and update the state vector;

[0151] Input the updated state vector into the DRL model to adjust the estimated arrival time;

[0152] Use the simple time difference calculation method for the adjusted estimated arrival time and the arrival time to obtain the time deviation;

[0153] Extract the current remaining material quantity through the digital twin environment, combine it with the time deviation, set the critical value of zero inventory through the inventory critical threshold comparison method, compare it with the current remaining material quantity. If the remaining material quantity is less than the zero inventory critical value, trigger the DRL model to optimize the current estimated arrival time and update the scheduling instruction set.

[0154] By using voxel gridding to construct a 3D model of material storage points, generating a 3D virtual map of the distribution path using Rapidly-Exploring Random Trees (RRT), synthesizing a virtual environment based on the 3D model and the 3D virtual map, the simulation accuracy is improved, providing reliable spatial data support for zero-inventory scheduling, reducing path planning errors, predicting the arrival time by constructing a Deep Reinforcement Learning (DRL) model, inputting the initial state vector into the DRL model, and calculating the estimated arrival time through formulas, significantly improving the adaptability of the prediction, ensuring the timely arrival of materials under zero inventory, updating the state vector by running the digital twin environment, triggering the optimization mechanism when the inventory is less than the zero-inventory critical value, updating the scheduling instruction set, reducing the risk of material interruption, and enhancing the emergency response ability of hospital logistics.

[0155] S4. Execute the scheduling instruction set, store and analyze the generated multi-modal data;

[0156] Executing the scheduling instruction set means sending the scheduling instruction set to the supplier's scheduling system through communication for the execution of scheduling instructions.

[0157] Through the instruction set fields and communication protocols, accurate and real-time instruction execution is achieved in zero-inventory scheduling, solving the problems of low traditional communication efficiency and response lag, and significantly enhancing the reliability and timeliness of material supply.

[0158] Storing the generated multi-modal data instruction set of collection and analysis means recording and storing the collected multi-modal data and the generated scheduling instruction set in the database. The database marks the stored data with timestamps and uploads the stored data to the cloud for backup.

[0159] Through database storage and timestamp marking, the dynamic integration of multi-modal data and instruction sets is achieved, solving the problems of temporality and traceability. Through cloud backup, the data security and accessibility are guaranteed, solving the problems of disaster recovery and collaboration.

[0160] This embodiment also provides an AI-based zero-inventory intelligent scheduling system for hospital logistics, including:

[0161] A collection and prediction module for collecting multi-dimensional data to generate a multi-dimensional data set, defining department, material, and environmental characteristics, defining nodes and edges for the characteristics, generating a graph data format, using a Graph Neural Network (GNN) for feature extraction, dynamically adjusting the edge weights through Reinforcement Learning (RL), and using the Q-learning formula for updating to predict material demand and carbon emissions;

[0162] A scheduling and integration module for generating a zero-inventory scheduling plan through a Deep Q-Network (DQN) and integrating it into a scheduling instruction set;

[0163] The verification and update module is used to synthesize a virtual environment from a 3D model and a map, load a scheduling instruction set and a real-time hospital logistics status data set, generate a digital twin environment, construct a DRL model to predict the arrival time, run the digital twin environment to update the state vector, adjust the estimated arrival time, trigger an optimization mechanism when the inventory materials are less than the zero-inventory critical value, and update the scheduling instruction set;

[0164] The execution and storage module is used to execute the scheduling instruction set and store the multi-modal data generated by collection and analysis.

[0165] This embodiment also provides a computer device applicable to the case of the AI-based hospital logistics zero-inventory intelligent scheduling method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI-based hospital logistics zero-inventory intelligent scheduling method as proposed in the above embodiment.

[0166] This computer device can be a terminal, and this computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of this computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0167] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the method for realizing AI-based intelligent scheduling of zero inventory in hospital logistics proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM for short), electrically erasable programmable read-only memory (EEPROM for short), erasable programmable read-only memory (EPROM for short), programmable read-only memory (PROM for short), read-only memory (ROM for short), magnetic memory, flash memory, magnetic disk or optical disc.

[0168] In summary, the present invention generates a multi-dimensional data set by collecting multi-dimensional data, uses GNN for feature extraction, dynamically adjusts edge weights through RL, updates using the Q-learning formula, and predicts material demand and carbon emissions; synthesizes a 3D model and a map into a virtual environment, loads a scheduling instruction set and a real-time status data set of hospital logistics, generates a digital twin environment, constructs a DRL model to predict the arrival time, runs the digital twin environment to update the state vector, adjusts the predicted arrival time, and updates the scheduling instruction set. The present invention provides accurate data for zero inventory, improves the response speed of scheduling, enhances the adaptability of the system to complex scenarios, ensures the efficiency of scheduling execution, improves the stability and intelligence level of hospital logistics, and realizes the efficient zero-inventory management of hospital logistics.

[0169] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An AI-based intelligent scheduling method for zero-inventory in hospital logistics, characterized in that: including, collecting multi-dimensional data to generate a multi-dimensional dataset, defining department, material, and environmental characteristics, defining nodes and edges for the characteristics, generating a graph data format, using a GNN for feature extraction, dynamically adjusting edge weights through RL, using the Q-learning formula for updating, and predicting material demand and carbon emissions; generating a zero-inventory scheduling plan through a DQN network and integrating it into a scheduling instruction set; synthesizing a 3D model and a map into a virtual environment, loading the scheduling instruction set and the hospital logistics real-time status dataset, generating a digital twin environment, constructing a DRL model to predict the arrival time, running the digital twin environment to update the state vector, adjusting the predicted arrival time, and triggering an optimization mechanism when the inventory material is less than the zero-inventory critical value to update the scheduling instruction set; executing the scheduling instruction set and storing the multi-modal data generated by collection and analysis.

2. The AI-based intelligent zero-inventory scheduling method for hospital logistics according to claim 1, wherein: The collecting multi-dimensional data to generate a multi-dimensional dataset and predicting material demand and carbon emissions means using intelligent sensors to collect multi-dimensional data, preprocessing the multi-dimensional data to generate a multi-dimensional dataset, and segmenting it into time period windows by overlapping sliding windows according to the time series; The multi-dimensional data includes the material number and the remaining amount of the material, the medical staff demand instruction, the image, the device standby, and the device power during operation; using the APL interface to obtain the hospital department configuration table, the deliverable material quantity and delivery time of the supplier, the delivery path parameters, and the material demand times from the hospital information system; normalizing the material status data and the carbon emission data using Z-score, binary encoding the sudden demand signal, combining the material status data and the sudden demand signal to generate department characteristics, defining the remaining amount of the material as the material characteristic, and defining the carbon emission as the environmental characteristic; using the feature splicing method to integrate the department characteristics, the material characteristics, and the environmental characteristics into an initial feature set, extracting the features in the initial feature set and defining them as nodes, assigning IDs to each node, and generating a dynamic node set; vertically stacking the initial feature set in the order of node IDs to form a feature matrix; if a department consumes materials, a directed edge is established between the department node and the material node. If two departments have a demand for the same material, an undirected edge is established between the two department nodes. If carbon emissions are generated during material transportation, a directed edge is established between the material node and the carbon emission node, and the weights of the three edges are calculated; integrating the three edges and the corresponding weights to form a dynamic edge set, using an adjacency matrix to integrate the dynamic node set and the dynamic edge set into a graph structure, performing sparsification processing, and storing it in the graph data format; constructing a GNN model and defining the action space as the increase and decrease operations for adjusting all edge weights; defining a discrete action set for each edge using fixed-step discretization, calculating the feature variance of each node through analysis of variance, sorting in descending order of variance, using the k-selection method to select the node with the largest variance, extracting the associated edges of the node, and outputting the discrete action set of the associated edges; discretizing the state vector through the K-means state clustering method to generate a discrete state set, and generating an action index through the dynamic combination technology method for the discrete action set; Set the discrete state set as the number of rows and the action index as the number of columns to construct a Q-table. Initialize the value of all state-action executions in the Q-table to 0. Use the ϵ-greedy strategy to select an initial action for the current state, adjust the edge weights according to the initial action, update the adjacency matrix, and generate an optimized feature matrix through a second-layer convolutional operation; Perform global pooling on the optimized feature matrix to generate a global feature vector. Output the predicted material demand and carbon emissions per hour through a two-layer fully connected layer. Obtain the material demand error and carbon emission error through a loss function, and calculate the RL reward value based on these two errors; Update using the Q-learning formula through the RL reward value, output the updated Q-table, feedback it to the RL module, adjust the ϵ-greedy strategy, update the adjacency matrix, and perform the next round of GNN prediction. Set the maximum number of iterations using the rule of thumb and stop the iteration when the maximum iteration is reached; Collect historical data through sensors, integrate it into the historical graph data format, input it into the GNN model for training, and optimize the model parameters using a loss function and the Adam optimizer; Input the graph data format into the GNN model to predict the material demand and carbon emissions, and convert it into a prediction data table through structured data mapping transformation.

3. The AI-based intelligent zero-inventory scheduling method for hospital logistics according to claim 2, wherein: The generation of the zero-inventory scheduling plan through the DQN network and the integration into the scheduling instruction set means converting the prediction data table into the JSON format through structured data serialization; Integrate the demand, time window, remaining material quantity, and the quantity and delivery time of materials that can be delivered by suppliers in the prediction data table into a supply-demand status vector through the direct splicing method; Define the actions as accepting and rejecting supplier deliveries; Define the DQN network, output the Q-values of the defined actions through the hidden layer, and determine the actions through the ϵ-greedy strategy; Calculate the time difference between the end time of the time window and the delivery time of the supplier, which is defined as the shipping time. Use the path planning API call method to obtain the optimal delivery path, and calculate the arrival time and delivery cost; Merge the shipping time, optimal path, arrival time, delivery cost, and carbon emissions into a scheduling plan, and convert the scheduling plan into a scheduling instruction set through XML formatting; 4. The AI-based intelligent scheduling method for zero-inventory hospital logistics according to claim 3, characterized in that: The synthesis of the 3D model and the map into a virtual environment and the update of the scheduling instruction set mean that the digital twin system accepts the scheduling instruction set; Extract the remaining material quantity, material consumption speed, and delivery path parameters, and generate a hospital logistics real-time status data set through ELT; Use voxel gridding to construct a 3D model of the material storage point, use the rapidly exploring random tree to generate a 3D virtual map of the delivery path, and synthesize a virtual environment based on the 3D model and the 3D virtual map; Load the scheduling instruction set and the hospital logistics real-time status data set into the virtual environment to generate a digital twin environment, set the simulation duration of the digital twin environment through real-time granularity simulation, and start the real-time simulation; Integrate the remaining material quantity, delivery path parameters, and remaining shipping time to generate an initial state vector; The remaining shipping time is obtained by calculating the shipping time and the current time; Construct a DRL model, extract the delivery path parameters, and calculate and output the estimated arrival time; Run the digital twin environment to the next time point, update the remaining quantity of current supplies according to the consumption rate, obtain the remaining time based on the current time and the delivery time, update the state vector, input it into the DRL model, adjust the estimated arrival time, and use the simple time difference calculation method for the adjusted estimated arrival time and the arrival time to obtain the time deviation; Extract the remaining quantity of current supplies through the digital twin environment, combine it with the time deviation, set the critical value of zero inventory through the inventory critical threshold comparison method, compare it with the remaining quantity of current supplies, and if the remaining quantity of supplies is less than the zero inventory critical value, trigger the DRL model to optimize the current estimated arrival time and update the scheduling instruction set.

5. The AI-based intelligent scheduling method for zero-inventory hospital logistics according to claim 1, characterized in that: The generation of the multi-dimensional dataset includes: Calculate the material consumption rate, and integrate it with the material number and the remaining quantity of materials into the material status data; Detect the number of patients entering the area per minute and the equipment usage status in the image through the YOLOv5 model, calculate the equipment usage rate through the usage rate percentage formula, obtain the sudden demand index through calculation, and integrate it with the medical staff demand instruction into the sudden demand signal; Calculate the power consumption and carbon emissions through the equipment power, and integrate them into the environmental data; Synchronize the time of the material status data, the sudden demand signal, and the environmental data through the timestamp, and integrate them into a multi-dimensional dataset.

6. The AI-based intelligent scheduling method for zero-inventory hospital logistics according to claim 4, characterized in that: The execution of the scheduling instruction set refers to sending the scheduling instruction set to the scheduling system of the supplier through communication for execution.

7. The AI-based intelligent scheduling method for zero-inventory hospital logistics according to claim 4, characterized in that: The storage of the multi-modal data generated by collection and analysis refers to recording and storing the collected multi-modal data and the generated scheduling instruction set in the database. The database marks the timestamp for the stored data and uploads the stored data to the cloud for backup.

8. An AI-based intelligent zero-inventory scheduling system for hospital logistics, based on the AI-based intelligent zero-inventory scheduling method according to any one of claims 1 to 7, characterized in that: Including, A collection and prediction module for collecting multi-dimensional data to generate a multi-dimensional dataset, defining department, material, and environmental characteristics, defining nodes and edges for the characteristics, generating a graph data format, using GNN for feature extraction, dynamically adjusting the edge weights through RL, and using the Q-learning formula for update to predict the material demand and carbon emissions; A scheduling and integration module for generating a zero-inventory scheduling plan through the DQN network and integrating it into a scheduling instruction set; A verification and update module for synthesizing a virtual environment by combining the 3D model and the map, loading the scheduling instruction set and the hospital logistics real-time status dataset, generating a digital twin environment, constructing a DRL model to predict the arrival time, running the digital twin environment to update the state vector, adjusting the estimated arrival time, and triggering the optimization mechanism when the inventory materials are less than the zero-inventory critical value to update the scheduling instruction set; An execution and storage module for executing the scheduling instruction set and storing the multi-modal data generated by collection and analysis.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the AI-based hospital logistics zero-inventory intelligent scheduling method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AI-based hospital logistics zero-inventory intelligent scheduling method according to any one of claims 1 to 7.

Images